Direct3D ShaderX

Vertex and Pixel Shader

Tips and Tricks
Edited by Wolfgang F. Engel
Wordware Publishing, Inc.
Team LRN
Library of Congress Cataloging-in-Publication Data
Direct3d ShaderX : vertex and pixel shader tips and tricks / edited by
Wolfgang F. Engel.
p. cm.
Includes bibliographical references and index.
ISBN 1-55622-041-3
1. Computer games--Programming. 2. Three-dimensional display systems.
3. Direct3D. I. Engel, Wolfgang F.
QA76.76.C672 D57 2002
794.8'15265--dc21
2002006571
2002, Wordware Publishing, Inc.
All Rights Reserved
2320 Los Rios Boulevard
Plano, Texas 75074
No part of this book may be reproduced in any form or by any means
without permission in writing from Wordware Publishing, Inc.
Printed in the United States of America
ISBN 1-55622-041-3
10 9 8 7 6 5 4 3 2 1
0502
RenderMan is a registered trademark of Pixar Animation Studios.
krass engine is a trademark of Massive Development.
3D Studio Max is a registered trademark of Autodesk/Discreet in the USA and/or other countries.
GeForce is a trademark of NVIDIA Corporation in the United States and/or other countries.
Some figures in this book are copyright ATI Technologies, Inc., and are used with permission.
Other product names mentioned are used for identification purposes only and may be trademarks of their respective companies.
All inquiries for volume purchases of this book should be addressed to Wordware Publishing, Inc., at the above
address. Telephone inquiries may be made by calling:
(972) 423-0090
Team LRN
Contents
Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi
Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii
Part 1: Introduction to Shader Programming
Fundamentals of Vertex Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Wolfgang F. Engel
What You Need to Know/Equipment . . . . . . . . . . . . . . . . . . . . . . . . . 5
Vertex Shaders in the Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Why Use Vertex Shaders? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Vertex Shader Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
NVIDIA Effects Browser 2/3 . . . . . . . . . . . . . . . . . . . . . . . . . . 8
NVIDIA Shader Debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Shader City . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Vertex Shader Assembler. . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
NVIDIA NVASM Vertex and Pixel Shader Macro Assembler . . . . 10
Microsoft Vertex Shader Assembler . . . . . . . . . . . . . . . . . . . 11
Shader Studio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
NVLink 2.x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
NVIDIA Photoshop Plug-ins. . . . . . . . . . . . . . . . . . . . . . . . . . 13
Diffusion Cubemap Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
DLL Detective with Direct3D Plug-in. . . . . . . . . . . . . . . . . . . . . 15
3D Studio Max 4.x/gmax 1.1. . . . . . . . . . . . . . . . . . . . . . . . . . 16
Vertex Shader Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
High-Level View of Vertex Shader Programming . . . . . . . . . . . . . . . . . . 18
Check for Vertex Shader Support . . . . . . . . . . . . . . . . . . . . . . . 19
Vertex Shader Declaration . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Set the Vertex Shader Constant Registers . . . . . . . . . . . . . . . . . . . 22
Write and Compile a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . 22
Application Hints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Complex Instructions in the Vertex Shader. . . . . . . . . . . . . . . . 27
Putting It All Together . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Swizzling and Masking . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Guidelines for Writing Vertex Shaders . . . . . . . . . . . . . . . . . . 32
Compiling a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . 33
Create a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Set a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Free Vertex Shader Resources . . . . . . . . . . . . . . . . . . . . . . . . . 35
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
What Happens Next? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
iii
Team LRN
Programming Vertex Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Wolfgang F. Engel
RacorX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
The Common Files Framework . . . . . . . . . . . . . . . . . . . . . . . . 40
Check for Vertex Shader Support. . . . . . . . . . . . . . . . . . . . . 42
Vertex Shader Declaration . . . . . . . . . . . . . . . . . . . . . . . . 42
Set the Vertex Shader Constant Registers . . . . . . . . . . . . . . . . 43
The Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Compile a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . 45
Create a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Set a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Free Vertex Shader Resources . . . . . . . . . . . . . . . . . . . . . . 47
Non-Shader Specific Code . . . . . . . . . . . . . . . . . . . . . . . . 47
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
RacorX2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Create a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
RacorX3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Vertex Shader Declaration . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Set the Vertex Shader Constant Registers . . . . . . . . . . . . . . . . . . . 54
The Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Directional Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Diffuse Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
RacorX4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Vertex Shader Declaration . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Set the Vertex Shader Constants . . . . . . . . . . . . . . . . . . . . . . . . 60
The Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Specular Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Non-Shader Specific Code . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
RacorX5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Point Light Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Light Attenuation for Point Lights . . . . . . . . . . . . . . . . . . . . . . . 66
Set the Vertex Shader Constants . . . . . . . . . . . . . . . . . . . . . . . . 67
The Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
What Happens Next? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Fundamentals of Pixel Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Wolfgang F. Engel
Why Use Pixel Shaders? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Pixel Shaders in the Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Pixel Shader Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Microsoft Pixel Shader Assembler. . . . . . . . . . . . . . . . . . . . . . . 79
MFC Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
iv Contents
Team LRN
ATI ShadeLab. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Pixel Shader Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Constant Registers (c0-c7) . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Output and Temporary Registers (ps.1.1-ps.1.3: r0+r1; ps.1.4: r0-r5) . . . . 82
Texture Registers (ps.1.1-ps.1.3: t0-t3; ps.1.4: t0-t5) . . . . . . . . . . . . . 82
Color Registers (ps.1.1-ps.1.4: v0+v1) . . . . . . . . . . . . . . . . . . . . 83
Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
High-Level View of Pixel Shader Programming . . . . . . . . . . . . . . . . . . 84
Check for Pixel Shader Support . . . . . . . . . . . . . . . . . . . . . . . . 84
Set Texture Operation Flags (D3DTSS_* flags). . . . . . . . . . . . . . . . 85
Set Texture (with SetTexture()) . . . . . . . . . . . . . . . . . . . . . . . . 86
Define Constants (with SetPixelShaderConstant()/def ) . . . . . . . . . . . 86
Pixel Shader Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Texture Address Instructions . . . . . . . . . . . . . . . . . . . . . . . 88
Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . 101
Instruction Modifiers . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Instruction Modifiers . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Instruction Pairing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Assemble Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Create a Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Set Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Free Pixel Shader Resources . . . . . . . . . . . . . . . . . . . . . . . . . 121
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
What Happens Next? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Programming Pixel Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Wolfgang F. Engel
RacorX6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Check for Pixel Shader Support . . . . . . . . . . . . . . . . . . . . . . . 126
Set Texture Operation Flags (with D3DTSS_* flags) . . . . . . . . . . . . 126
Set Texture (with SetTexture()) . . . . . . . . . . . . . . . . . . . . . . . . 127
Define Constants (with SetPixelShaderConstant()/def) . . . . . . . . . . . 127
Pixel Shader Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Per-Pixel Lighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Assemble Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Create Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Set Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Free Pixel Shader Resources . . . . . . . . . . . . . . . . . . . . . . . . . 136
Non-Shader Specific Code . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
RacorX7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Define Constants (with SetPixelShaderConstant()/def) . . . . . . . . . . . 137
Pixel Shader Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
RacorX8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Set Texture Operation Flags (with D3DTSS_* flags) . . . . . . . . . . . . 140
Contents v
Team LRN
Set Texture (with SetTexture()) . . . . . . . . . . . . . . . . . . . . . . . . 140
Pixel Shader Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
RacorX9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Basic Shader Development with Shader Studio . . . . . . . . . . . . . . . . . . 149
John Schwab
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
What You Should Know . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Installation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Directories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Coordinate Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Default Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Create a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Step 1: Loading a Mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Step 2: Transformations. . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Step 3: Starting a Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Step 4: Editing Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Step 5: Adding Some Color. . . . . . . . . . . . . . . . . . . . . . . . . . 155
Step 6: Saving Your Work . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Step 7: Loading a Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . 156
Step 8: Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Step 9: Texture Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Create a Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Step 1: Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Step 2: Writing a Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . 159
Shaders Reference. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Workspace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Menu Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Shaders Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Declarations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Constant Properties Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Materials Dialog. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Textures Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Browser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Transformations Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
vi Contents
Team LRN
Camera. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Statistics Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Settings Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
State Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Camera Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Configure Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Assets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Further Info . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Part 2: Vertex Shader Tricks
Vertex Decompression in a Shader . . . . . . . . . . . . . . . . . . . . . . . . 172
Dean Calver
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Vertex Compression Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Vertex Shader Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Compressed Vertex Stream Declaration Example . . . . . . . . . . . . . . . . . 174
Basic Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Compression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Decompression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Scaled Offset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Compression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Decompression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Compression Transform. . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Compression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Decompression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Advanced Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Multiple Compression Transforms . . . . . . . . . . . . . . . . . . . . . . 180
Compression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Decompression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Sliding Compression Transform . . . . . . . . . . . . . . . . . . . . . . . 182
Compression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Decompression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Displacement Maps and Vector Fields . . . . . . . . . . . . . . . . . . . . 184
Basic Displacement Maps . . . . . . . . . . . . . . . . . . . . . . . . 184
Entering Hyperspace . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Contents vii
Team LRN
Shadow Volume Extrusion Using a Vertex Shader. . . . . . . . . . . . . . . . . 188
Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Creating Shadow Volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Effect File Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Using Shadow Volumes with Character Animation . . . . . . . . . . . . . . . . 192
Character Animation with Direct3D Vertex Shaders . . . . . . . . . . . . . . . . 195
David Gosselin
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Tweening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Skinning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Skinning and Tweening Together . . . . . . . . . . . . . . . . . . . . . . . 199
Animating Tangent Space for Per-Pixel Lighting . . . . . . . . . . . . . . . . . 202
Per-Pixel Lighting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Compression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Lighting a Single-Surface Object . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Greg James
Vertex Shader Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Enhanced Lighting for Thin Objects . . . . . . . . . . . . . . . . . . . . . . . . 213
About the Demo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Optimizing Software Vertex Shaders . . . . . . . . . . . . . . . . . . . . . . . 215
Kim Pallister
Introduction to Pentium 4 Processor Architecture . . . . . . . . . . . . . . . . . 216
Introduction to the Streaming SIMD Extensions . . . . . . . . . . . . . . . . . 217
Optimal Data Arrangement for SSE Instruction Usage . . . . . . . . . . . . . . 218
How the Vertex Shader Compiler Works . . . . . . . . . . . . . . . . . . . . . . 220
Performance Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Optimization Guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Write Only the Results Youll Need. . . . . . . . . . . . . . . . . . . 221
Use Macros Whenever Possible . . . . . . . . . . . . . . . . . . . . 222
Squeeze Dependency Chains . . . . . . . . . . . . . . . . . . . . . . 222
Write Final Arithmetic Instructions Directly to Output Registers . . . 223
Reuse the Same Temporary Register If Possible . . . . . . . . . . . . 223
Dont Implicitly Saturate Color and Fog Values . . . . . . . . . . . . 223
Use the Lowest Acceptable Accuracy with exp and log Functions. . . 223
Avoid Using the Address Register When Possible . . . . . . . . . . . 223
Try to Order Vertices . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Profile, Profile, Profile . . . . . . . . . . . . . . . . . . . . . . . . 224
A Detailed Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
viii Contents
Team LRN
Compendium of Vertex Shader Tricks . . . . . . . . . . . . . . . . . . . . . . . 228
Scott Le Grand
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Periodic Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
One-Shot Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Random Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Flow Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Cross Products. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Perlin Noise and Returning Results from Shader Programs . . . . . . . . . . . 232
Steven Riddle and Oliver C. Zecha
Limitations of Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Perlin Noise and Fractional Brownian Motion . . . . . . . . . . . . . . . . . . . 234
Final Thoughts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Part 3: Pixel Shader Tricks
Blending Textures for Terrain . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Alex Vlachos
Image Processing with 1.4 Pixel Shaders in Direct3D. . . . . . . . . . . . . . . 258
Jason L. Mitchell
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Simple Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Black and White Transfer Function . . . . . . . . . . . . . . . . . . . . . 260
Sepia Tone Transfer Function. . . . . . . . . . . . . . . . . . . . . . . . . 260
Heat Signature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Filter Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Texture Coordinates for Filter Kernels . . . . . . . . . . . . . . . . . . . . 261
Edge Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Roberts Cross Gradient Filters . . . . . . . . . . . . . . . . . . . . . . . . 262
Sobel Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Mathematical Morphology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
The Definition of Dilation . . . . . . . . . . . . . . . . . . . . . . . . . . 265
A Dilation Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
The Definition of Erosion. . . . . . . . . . . . . . . . . . . . . . . . . . . 267
An Erosion Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Hardware Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Sample Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Contents ix
Team LRN
Hallo World Font Smoothing with Pixel Shaders. . . . . . . . . . . . . . . . 270
Steffen Bendel
Emulating Geometry with Shaders Imposters . . . . . . . . . . . . . . . . . 273
Steffen Bendel
Smooth Lighting with ps.1.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Steffen Bendel
Per-Pixel Fresnel Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Fresnel Effects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Effect Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Diffuse Cube Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Using Diffuse Cube Maps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Generating Dynamic Diffuse Cube Maps . . . . . . . . . . . . . . . . . . . . . 288
Accurate Reflections and Refractions by Adjusting for Object Distance . . . . . 290
Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
The Artifacts of Environment Mapping . . . . . . . . . . . . . . . . . . . . . . 290
UV Flipping Technique to Avoid Repetition . . . . . . . . . . . . . . . . . . . . 295
Alex Vlachos
Photorealistic Faces with Vertex and Pixel Shaders . . . . . . . . . . . . . . . 296
Kenneth L. Hurley
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
3D Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Setup for Separation of Lighting Elements. . . . . . . . . . . . . . . . . . 298
Diffuse Lighting Component . . . . . . . . . . . . . . . . . . . . . . . . . 298
Extracting the Bump and Specular Map . . . . . . . . . . . . . . . . . . . 299
Separating Specular and Bump Maps . . . . . . . . . . . . . . . . . . . . 299
Normal Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
More on Normal Maps . . . . . . . . . . . . . . . . . . . . . . . . . 301
A Word About Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
Half Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
Vertex Shader Explanation Pass 1 (Listing 1) . . . . . . . . . . . . . . . . 303
Vertex Shader Explanation Pass 1 (Listing 2) . . . . . . . . . . . . . . . . 305
Pixel Shader Explanation (Listing 3) . . . . . . . . . . . . . . . . . . . . . 307
x Contents
Team LRN
Full Head Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Mapping the Photographs . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Creating a Single Texture Map . . . . . . . . . . . . . . . . . . . . . . . . 310
Putting It All Together . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
Whats Next? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
Environmental Lighting. . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
How to Get Diffusion Cube Maps. . . . . . . . . . . . . . . . . . . . 312
Generating the Cube Maps . . . . . . . . . . . . . . . . . . . . . . . 313
The Pixel Shader to Use All These Maps . . . . . . . . . . . . . . . . 314
Environment Mapping for Eyes. . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Final Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Facial Animation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Non-Photorealistic Rendering with Pixel and Vertex Shaders . . . . . . . . . . 319
Drew Card and Jason L. Mitchell
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Rendering Outlines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Cartoon Lighting Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
Hatching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Gooch Lighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
Image Space Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
Compositing the Edges . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
Depth Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
Alpha Test for Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
Shadow Outlines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
Thickening Outlines with Morphology. . . . . . . . . . . . . . . . . . . . 332
Summary of Image Space Technique. . . . . . . . . . . . . . . . . . . . . 332
Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Animated Grass with Pixel and Vertex Shaders . . . . . . . . . . . . . . . . . . 334
John Isidoro and Drew Card
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
Waving the Grass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
Lighting the Grass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
Texture Perturbation Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
John Isidoro, Guennadi Riguer, and Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Wispy Clouds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Perturbation-Based Fire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
Plasma Glass. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
Contents xi
Team LRN
Rendering Ocean Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
John Isidoro, Alex Vlachos, and Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Sinusoidal Perturbation in a Vertex Shader. . . . . . . . . . . . . . . . . . . . . 348
CMEMBM Pixel Shader with Fresnel Term . . . . . . . . . . . . . . . . . . . . 350
Ocean Water Shader Source Code . . . . . . . . . . . . . . . . . . . . . . . . . 352
Sample Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
Rippling Reflective and Refractive Water . . . . . . . . . . . . . . . . . . . . . 357
Alex Vlachos, John Isidoro, and Chris Oat
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
Generating Reflection and Refraction Maps . . . . . . . . . . . . . . . . . . . . 358
Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
Crystal/Candy Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
John Isidoro and David Gosselin
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
Bubble Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
John Isidoro and David Gosselin
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Vertex Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Per-Pixel Strand-Based Anisotropic Lighting . . . . . . . . . . . . . . . . . . . 376
John Isidoro and Chris Brennan
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Strand-Based Illumination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Rendering Using the Texture Lookup Table . . . . . . . . . . . . . . . . . 377
Per-Pixel Strand-Based Illumination . . . . . . . . . . . . . . . . . . . . . . . . 378
Per-Pixel Strand-Based Illumination with Colored Light and Base Map . . 379
Per-Pixel Strand-Based Illumination with Four Colored Lights and
Base Map in One Pass . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Per-Pixel Bump-Mapped Strand-Based Illumination Using Gram-Schmidt
Orthonormalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
xii Contents
Team LRN
A Non-Integer Power Function on the Pixel Shader . . . . . . . . . . . . . . . . 383
Philippe Beaudoin and Juan Guardado
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Traditional Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Texture Lookup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Successive Multiplications . . . . . . . . . . . . . . . . . . . . . . . . . . 386
The Need for a New Trick . . . . . . . . . . . . . . . . . . . . . . . . . . 386
Mathematical Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Power Function on the Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . 391
Constant Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
Per-Pixel Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
Smooth Conditional Function. . . . . . . . . . . . . . . . . . . . . . . . . 396
Volume Bounded Pixel Shader Effects . . . . . . . . . . . . . . . . . 398
Phong Shading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Phong Equation with Blinn Half-Vector . . . . . . . . . . . . . . . . 400
Expressing the Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . 400
The Pixel Shader . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
Bump Mapped BRDF Rendering . . . . . . . . . . . . . . . . . . . . . . . . . . 405
dm Moravnszky
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Bidirectional Reflectance Distribution Functions . . . . . . . . . . . . . . . . . 406
Decomposing the Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Reconstruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Adding the Bumps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
Real-Time Simulation and Rendering of Particle Flows. . . . . . . . . . . . . . 414
Daniel Weiskopf and Matthias Hopf
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
Ingredients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
How Does a Single Particle Move?. . . . . . . . . . . . . . . . . . . . . . . . . 416
Basic Texture Advection: How Do a Bunch of Particles Move? . . . . . . . . . 417
Inflow: Where are the Particles Born? . . . . . . . . . . . . . . . . . . . . . . . 421
How Can Particles Drop Out? . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Contents xiii
Team LRN
Part 4: Using 3D Textures with Shaders
3D Textures and Pixel Shaders. . . . . . . . . . . . . . . . . . . . . . . . . . . 428
Evan Hart
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
3D Textures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Storage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
Function of Three Independent Variables . . . . . . . . . . . . . . . . 430
Noise and Procedural Texturing. . . . . . . . . . . . . . . . . . . . . 431
Attenuation and Distance Measurement . . . . . . . . . . . . . . . . 432
Representation of Amorphous Volume . . . . . . . . . . . . . . . . . 435
Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
Volumetric Fog and Other Atmospherics. . . . . . . . . . . . . . . . . . . 435
Light Falloff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
Truly Volumetric Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
Martin Kraus
The Role of Volume Visualization . . . . . . . . . . . . . . . . . . . . . . . . . 438
Basic Volume Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Animation of Volume Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
Rotating Volume Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . 441
Animated Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . . 441
Blending of Volume Graphics . . . . . . . . . . . . . . . . . . . . . . . . 442
High-Quality but Fast Volume Rendering . . . . . . . . . . . . . . . . . . . . . 444
Where to Go from Here . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Part 5: Engine Design with Shaders
First Thoughts on Designing a Shader-Driven Game Engine . . . . . . . . . . . 450
Steffen Bendel
Bump Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
Real-time Lighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
Use Detail Textures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Use Anisotropic Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Split Up Rendering into Independent Passes . . . . . . . . . . . . . . . . . . . . 451
Use _x2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
xiv Contents
Team LRN
Visualization with the Krass Game Engine . . . . . . . . . . . . . . . . . . . . 453
Ingo Frick
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
General Structure of the Krass Engine . . . . . . . . . . . . . . . . . . . . . . . 453
Developmental History of the Rendering Component . . . . . . . . . . . . . . . 454
Previous Drawbacks of Hardware Development . . . . . . . . . . . . . . . . . . 455
Current Drawbacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Ordering Effects in the Krass Engine . . . . . . . . . . . . . . . . . . . . . . . . 456
Application of the IMG Concept for Terrain Rendering . . . . . . . . . . . . . . 457
Particle Rendering to Exemplify a Specialized Effect Shader . . . . . . . . . . . 460
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
Designing a Vertex Shader-Driven 3D Engine for the Quake III Format . . . . . 463
Bart Sekura
Quake III Arena Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Vertex Program Setup in the Viewer . . . . . . . . . . . . . . . . . . . . . . . . 464
Vertex Shader Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Deformable Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Texture Coordinate Generation . . . . . . . . . . . . . . . . . . . . . . . . 468
Texture Matrix Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . 469
Rendering Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
About the Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
Contents xv
Team LRN
Foreword
With the advent of Microsoft

DirectX

version 8.0, the revolution of programmable graphics

had arrived. With vertex shaders for the programmable geometry pipeline and pixel shaders for
the programmable pixel pipeline, the control over geometry and pixels was handed back to the
developer. This unprecedented level of control in the graphics pipeline means graphics devel-
opers, once they have mastered the basics of shaders, now have the tools to generate new, as
yet unheard-of, effects. Wolfgang and his contributors have selected shader topics that they
believe will help to open wide the doors of illumination on shaders and the programmable
pipeline. Read on, be illuminated, and learn how to create your own effects using the program-
mable graphics pipeline.
Phil Taylor
Program Manager
Windows Graphics & Gaming Technologies
Microsoft Corporation
xvi
Team LRN
Acknowledgments
Like any book with a number of authors, this one has its own history. In late autumn of 2000,
I was very impressed by the new capabilities that were introduced with DirectX 8.0 by
Microsoft and NVIDIA. At Meltdown 2001 in London, I met Philip Taylor for the first time
and discussed with him the idea to write an entire book dedicated to vertex and pixel shader
programming. We had a good time thinking about a name for the book, and it was this discus-
sion that led to the title: Direct3D ShaderX.
Philip was one of the driving spirits who encouraged me to start this project, so I was very
glad when he agreed to write the foreword without hesitation. Jason L. Mitchell from ATI,
Juan Guardado from Matrox, and Matthias Wloka from NVIDIA (I met Matthias at the
Unterhaltungs Software Forum in Germany) were the other driving spirits who motivated their
colleagues to contribute to the book.
During a family vacation, I had the pleasure to get to know Javier Izquierdo
([email protected]) who showed me some of his artwork. I was very impressed and asked
him if he would create a ShaderX logo and design the cover of the book. His fantastic final
draft formed the basis of the cover design, which shows in-game screen shots of AquaNox
from Massive Development. These screen shots were sent to me by Ingo Frick, the technical
director of Massive and a contributor to this book. AquaNox was one of the first games that
used vertex and pixel shaders extensively.
A number of people have enthusiastically contributed to the book:
David Callele (University of Saskatchewan), Jeffrey Kiel (NVIDIA), Jason L. Mitchell
(ATI), Bart Sekura (People Can Fly), and Matthias Wloka (NVIDIA) all proofread several of
these articles.
A big thank you goes to the people at the Microsoft Direct3D discussion group
(http://DISCUSS.MICROSOFT.COM/archives/DIRECTXDEV.html), who were very helpful
in answering my numerous questions.
Similarly, a big thank you goes to Jim Hill from Wordware Publishing, along with Wes
Beckwith, Alan McCuller, and Paula Price. Jim gave me a big boost when I first brought this
idea to him. I had numerous telephone conferences with him about the strategic positioning of
the book in the market and the books progress, and met him for the first time at GDC 2002 in
San Jose.
I have never before had the pleasure of working with so many talented people. This great
teamwork experience will be something that I will never forget, and I am very proud to have
had the chance to be part of this team.
My special thanks goes to my wife, Katja, and our infant daughter, Anna, who had to
spend most evenings and weekends during the last five months without me, and to my parents
who always helped me to believe in my strength.
Wolfgang F. Engel
xvii
Team LRN
Team LRN
Direct3D ShaderX
Vertex and Pixel Shader
Tips and Tricks
Team LRN
Part I36 Foundation
for Security
This introduction covers the fundamentals of vertex shader and pixel shader
programming. You will learn everything here necessary to start program-
ming vertex and pixel shaders from scratch for the Windows family of
operating systems. Additionally, there is a tutorial on Shader Studio, a tool
for designing vertex and pixel shaders.
Fundamentals of Vertex Shaders
This article discusses vertex shaders, vertex shader tools, and lighting and
transformation with vertex shaders.
Programming Vertex Shaders
This article outlines how to write and compile a vertex shader program.
2
Team LRN
Fundamentals of Pixel Shaders
This article explains the pixel shader architecture, describes various pixel
shader tools, and gives an overview of basic programming.
Programming Pixel Shaders
This article takes a look at writing and compiling a pixel shader program,
including texture mapping, texture effects, and per-pixel lighting.
Basic Shader Development with Shader Studio
The creator of Shader Studio explains how to use this tool for designing ver-
tex and pixel shaders.
3
Team LRN
Fundamentals of Vertex
Shaders
Wolfgang F. Engel
We have seen ever-increasing graphics performance in PCs since the release of the first 3dfx
Voodoo cards in 1995. Although this performance increase has allowed PCs to run graphics
faster, it arguably has not allowed graphics to run much better. The fundamental limitation thus
far in PC graphics accelerators has been that they are mostly fixed-function, meaning that the
silicon designers have hard-coded specific graphics algorithms into the graphics chips, and as a
result, game and application developers have been limited to using these specific fixed
algorithms.
For over a decade, a graphics technology known as RenderMan

from Pixar Animation

Studios has withstood the test of time and has been the professionals choice for high-quality
photorealistic rendering.
Pixars use of RenderMan in its development of feature films such as Toy Story and A
Bugs Life has resulted in a level of photorealistic graphics that have amazed audiences world-
wide. RenderMans programmability has allowed it to evolve as major new rendering
techniques were invented. By not imposing strict limits on computations, RenderMan allows
programmers the utmost in flexibility and creativity. However, this programmability has lim-
ited RenderMan to software implementations.
Now, for the first time, low-cost consumer hardware has reached the point where it can
begin implementing the basics of programmable shading similar to RenderMan with real-time
performance.
The principal 3D APIs (DirectX and OpenGL) have evolved alongside graphics hardware.
One of the most important new features in DirectX graphics is the addition of a programmable
pipeline that provides an assembly language interface to the transformation and lighting hard-
ware (vertex shader) and the pixel pipeline (pixel shader). This programmable pipeline gives
the developer a lot more freedom to do things that have never before been seen in real-time
applications.
Shader programming is the new and real challenge for game-coders. Face it...
4
Team LRN
What You Need to Know/Equipment
You need a basic understanding of the math typically used in a game engine, and you need a
basic to intermediate understanding of the DirectX Graphics API. It helps if you know how to
use the Transform & Lighting (T&L) pipeline and the SetTextureStageState() calls. If you need
help with these topics, I recommend working through an introductory level text first, such as
Beginning Direct3D Game Programming.
Your development system should consist of the following hardware and software:
n
DirectX 8.1 SDK
n
Windows 2000 with at least Service Pack 2 or higher or Windows XP Professional (the
NVIDIA shader debugger only runs on these operating systems)
n
Visual C/C++ 6.0 with at least Service Pack 5 (needed for the DirectX 8.1 SDK) or higher
n
More than 128 MB RAM
n
At least 500 MB of hard drive storage
n
A hardware-accelerated 3D graphics card. To be able to get the maximum visual experi-
ence of the examples, you need to own relatively new graphics hardware. The pixel shader
examples will only run properly on GeForce3/4TI or RADEON 8x00 boards at the time of
publication.
n
The newest graphics card device driver
If you are not the lucky owner of a GeForce3/4TI, RADEON 8x00, or an equivalent graphics
card (that supports shaders in hardware), the standardized assembly interface will provide
highly tuned software vertex shaders that AMD and Intel have optimized for their CPUs. These
software implementations should jump in when there is no vertex shader capable hardware
found. There is no comparable software-emulation fallback path for pixel shaders.
Vertex Shaders in the Pipeline
The diagram on the following page shows the source or polygon, vertex, and pixel operations
levels of the Direct3D pipeline in a very simplified way.
On the source data level, the vertices are assembled and tessellated. This is the high-order
primitive module, which works to tessellate high-order primitives such as N-Patches (as sup-
ported by the ATI RADEON 8500 in hardware), quintic Bziers, B-splines, and rectangular
and triangular (RT) patches.
A GPU that supports RT-Patches breaks higher-order lines and surfaces into triangles and
vertices.
Fundamentals of Vertex Shaders 5
Note: It appears that, beginning with the 21.81 drivers,
NVIDIA no longer supports RT-Patches on the GeForce3/4TI.
Team LRN
A GPU that supports N-Patches generates the
control points of a Bzier triangle for each
triangle in the input data. This control mesh
is based on the positions and normals of the
original triangle. The Bzier surface is then
tessellated and evaluated, creating more tri-
angles on chip [Vlachos01].
The next stage shown in Figure 1 covers the vertex operations in the Direct3D pipeline. There
are two different ways of processing vertices:
1. The fixed-function pipeline. This is the standard Transform & Lighting (T&L) pipeline,
where the functionality is essentially fixed. The T&L pipeline can be controlled by setting
render states, matrices, and lighting and material parameters.
2. Vertex shaders. This is the new mechanism introduced in DirectX 8. Instead of setting
parameters to control the pipeline, you write a vertex shader program that executes on the
graphics hardware.
Our focus is on vertex shaders. It is obvious from the simplified diagram in Figure 1 that face
culling, user clip planes, frustum clipping, homogenous divide, and viewport mapping operate
on pipeline stages after the vertex shader. Therefore, these stages are fixed and cant be con-
trolled by a vertex shader. A vertex shader is not capable of writing to vertices other than the
6 Part 1: Introduction to Shader Programming
Note: If you use N-Patches together with pro-
grammable vertex shaders, you have to store
the position and normal information in input
registers v0 and v3. Thats because the N-Patch
tessellator needs to know where this informa-
tion is to notify the driver.
Note: The normal interpolation order can be
set to either D3DORDER_LINEAR or
D3DORDER_ QUADRATIC. In Direct3D 8.0,
the position interpolation was hard-wired to
D3DORDER_CUBIC and the normal interpola-
tion was hard-wired to D3DORDER_LINEAR.
Figure 1: Direct3D pipeline
Note: The N-Patch functionality was
enhanced in Direct3D 8.1. There is more con-
trol over the interpolation order of the positions
and normals of the generated vertices. The
new D3DRS_POSITIONORDER and D3DRS_
NORMALORDER render states control this inter-
polation order. The position interpolation order
can be set to either D3DORDER_LINEAR or
D3DORDER_CUBIC.
Team LRN
one it currently shades. It is also not capable of creating vertices; it generates one output vertex
from each vertex it receives as input.
So what are the capabilities and benefits of using vertex shaders?
Why Use Vertex Shaders?
If you use vertex shaders, you bypass the fixed-function pipeline or T&L pipeline. But why
would you want to skip them?
The hardware of a traditional T&L pipeline doesnt support all of the popular vertex attrib-
ute calculations on its own, and processing is often job-shared between the geometry engine
and the CPU. Sometimes, this leads to redundancy.
There is also a lack of freedom. Many of the effects used in games look similar to the
hard-wired T&L pipeline. The fixed-function pipeline doesnt give the developer the freedom
he needs to develop unique and revolutionary graphical effects. The procedural model used
with vertex shaders enables a more general syntax for specifying common operations. With the
flexibility of the vertex shaders, developers are able to perform operations including:
n
Procedural geometry (cloth simulation, soap bubble [Isidoro/Gosselin])
n
Advanced vertex blending for skinning and vertex morphing (tweening) [Gosselin]
n
Texture generation [Riddle/Zecha]
n
Advanced keyframe interpolation (complex facial expression and speech)
n
Particle system rendering [Le Grand]
n
Real-time modifications of the perspective view (lens effects, underwater effects)
n
Advanced lighting models (often in cooperation with the pixel shader) [Bendel]
n
First steps to displacement mapping [Calver]
There are many more effects possible with vertex shaders, some that have not been thought of
yet. For example, a lot of SIGGRAPH papers from the last couple of years describe graphical
effects that are realized only on SGI hardware so far. It might be a great challenge to port these
effects with the help of vertex and pixel shaders to consumer hardware.
In addition to opening up creative possibilities for developers and artists, shaders also
attack the problem of constrained video memory bandwidth by executing on-chip on
shader-capable hardware. Take, for example, Bzier patches. Given two floating-point values
per vertex (plus a fixed number of values per primitive), one can design a vertex shader to gen-
erate a position, a normal, and a number of texture coordinates. Vertex shaders even give you
the ability to decompress compressed position, normal, color, matrix, and texture coordinate
data and to save a lot of valuable bandwith without any additional cost [Calver].
There is also a benefit for your future learning curve. The procedural programming model
used by vertex shaders is very scalable. Therefore, the adding of new instructions and new reg-
isters will happen in a more intuitive way for developers.
Fundamentals of Vertex Shaders 7
Team LRN
Vertex Shader Tools
As you will soon see, you are required to master a specific RISC-oriented assembly language
to program vertex shaders because using the vertex shader is taking responsibility for program-
ming the geometry processor. Therefore, it is important to get the right tools to begin to
develop shaders as quickly and productively as possible.
I would like to present the tools that I am aware of at the time of publication.
NVIDIA Effects Browser 2/3
NVIDIA provides its own DirectX 8 SDK, which encapsulates all its tools, demos, and presen-
tations on DirectX 8.0. All the demos use a consistent framework called the Effects Browser.
The Effects Browser is a wonderful tool for testing and developing vertex and pixel shaders.
You can select the effect you would like to see in the left column. The middle column gives
you the ability to see the source of the vertex and/or pixel shader. The right column displays
the effect.
Not all graphics cards will support all the effects available in the Effects Browser.
GeForce3/4TI will support all the effects. Independent of your current graphic card prefer-
ences, I recommend downloading the NVIDIA DirectX 8 SDK and trying it out. The many
examples, including detailed explanations, show you a variety of the effects possible with ver-
tex and pixel shaders. The upcoming NVIDIA Effects Browser 3 will provide automatic online
update capabilities.
8 Part 1: Introduction to Shader Programming
Figure 2: NVIDIA Effects Browser
Team LRN
NVIDIA Shader Debugger
Once you use it, you wont live without it. The NVIDIA Shader Debugger provides you with
information about the current state of the temporary registers, input streams, output registers,
and constant memory. This data changes interactively while stepping through the shaders. It is
also possible to set instruction breakpoints as well as specific breakpoints.
A user manual that explains all the possible
features is provided. You need at least Win-
dows 2000 with Service Pack 1 to run the
Shader Debugger because debug services in
DX8 and DX8.1 are only supplied in Win-
dows 2000 and higher. It is important that
your application uses software vertex pro-
cessing (or you have switched to the
reference rasterizer) at run time for the
debugging process.
Fundamentals of Vertex Shaders 9
Figure 3: NVIDIA Shader Debugger
Note: You are also able to debug pixel
shaders with this debugger, but due to a
bug in DirectX 8.0, the contents of t0 are
never displayed correctly and user-added
pixel shader breakpoints will not trigger.
DirectX 8.1 fixes these issues, and you
receive a warning message if the applica-
tion finds an installation of DirectX 8.0.
Team LRN
Shader City
You can find another vertex and
pixel shader tool, along with
source code, at
http://www.palevich.com/3d/
ShaderCity/. Designed and
implemented by Jack Palevich,
Shader City allows you to see
any modification of the vertex
and/or pixel shaders in the small
client window in the upper-left:
The results of a modification of
a vertex and/or pixel shader can
be seen after they are saved and
reloaded. In addition, you are
able to load index and vertex
buffers from a file. The source
code for this tool might help
you to encapsulate Direct3D in
an ActiveX control, so go ahead
and try it.
Vertex Shader Assembler
To compile a vertex shader ASCII file (for example, basic.vsh) into a binary file (for example,
basic.vso), you must use a vertex shader assembler. As far as I know, there are two vertex
shader assemblers: the Microsoft vertex shader assembler and the NVIDIA vertex and pixel
shader macro assembler. The latter provides all of the features of the vertex shader assembler
plus many other features, whereas the vertex shader assembler gives you the ability to also use
the D3DX effect files (as of DirectX 8.1).
NVIDIA NVASM Vertex and Pixel Shader Macro Assembler
NVIDIA provides its vertex and pixel shader macro assembler as part of its DirectX 8 SDK.
NVASM has very robust error reporting built into it. It not only tells you what line the error
was on, but it is also able to backtrack errors. Good documentation helps you get started.
NVASM was written by Direct3D ShaderX author Kenneth Hurley, who provides additional
information in his article [Hurley]. We will learn how to use this tool in one of the upcoming
examples in the next article.
10 Part 1: Introduction to Shader Programming
Figure 4: Jack Palevich Shader City
Team LRN
Microsoft Vertex Shader Assembler
The Microsoft vertex shader assembler is delivered in the DirectX 8.1 SDK in
C:\dxsdk\bin\DXUtils.
If you call vsa.exe from the command line, you will get the following options:
usage: vsa -hp012 <files>
-h : Generate .h files (instead of .vso files)
-p : Use C preprocessor (VisualC++ required)
-0 : Debug info omitted, no shader validation performed
-1 : Debug info inserted, no shader validation performed
-2 : Debug info inserted, shader validation performed. (default)
I havent found any documentation for the vertex shader assembler. It is used by the
D3DXAssembleShader*() methods or by the effect file method D3DXCreateEffectFromFile(),
which compiles the effect file.
If you want to be hardware vendor-independent, you should use the Microsoft vertex
shader assembler.
Shader Studio
John Schwab has developed a
tool that will greatly aid in your
development of vertex and
pixel shaders. Whether you are
a beginner or an advanced
Direct3D programmer, this tool
will save you a lot of time by
allowing you to get right down
to development of any shader
without actually writing any
Direct3D code. Therefore, you
can spend your precious time
working on whats important,
the shaders.
The tool encapsulates a
complete vertex and pixel
shader engine with a few nice
ideas. For a hands-on tutorial
and detailed explanation, see
[Schwab]. The newest version
should be available online at
www.shaderstudio.com.
Fundamentals of Vertex Shaders 11
Figure 5: John Schwabs Shader Studio: phong lighting
Note: The default path of the DirectX 8 SDK is c:\mssdk.
The default path of the DirectX 8.1 SDK is c:\dxsdk.
Team LRN
NVLink 2.x
NVLink is a very interesting tool that allows you to:
n
Write vertex shaders that consist of fragments with #beginfragment and #endfragment
statements. For example:
#beginfragment world_transform
dp4 r_worldpos.x, v_position, c_world0
dp3 r_worldpos.y, v_position, c_world1
dp4 r_worldpos.z, v_position, c_world2
#endfragment
n
Assemble vertex shader files with NVASM into fragments
n
Link those fragments to produce a binary vertex shader at run time
NVLink helps you to generate shaders on demand that will fit into the end users hardware
limits (registers/instructions/constants). The most attractive feature of this tool is that it will
cache and optimize your shaders on the fly. NVLink is shown in the NV Effects Browser:
You can choose the vertex shader capabilities in the dialog box and the resulting vertex shader
will be shown in output0.nvv in the middle column.
12 Part 1: Introduction to Shader Programming
Figure 6:
NVLink
Note: The default path of the DirectX 8 SDK is c:\mssdk.
The default path of DirectX 8.1 SDK is c:\dxsdk.
Team LRN
NVIDIA Photoshop Plug-ins
You will find on NVIDIAs web site two frequently updated plug-ins for Adobe Photoshop:
NVIDIAs Normal Map Generator and Photoshop Compression plug-in.
The Normal Map Generator can generate normal maps that can be used, for example, for
Dot3 lighting. The plug-in requires DirectX 8.0 or later to be installed. The dynamic preview
window, located in the upper-left corner, shows an example light that is moved with the Ctrl +
left mouse button. You are able to clamp or wrap the edges of the generated normal map by
selecting or deselecting the Wrap check box. The height values of the normal map can be
scaled by providing a height value in the Scale entry field.
There are different options for height generation:
n
Alpha Channel Use alpha channel
n
Average RGB Average R, G, B
n
Biased RGB, h = average (R, G, B) Average of whole image
n
Red Use red channel
n
Green Use green channel
n
Blue Use blue channel
n
Max Use max of R, G, B
n
Colorspace, h = 1.0 [(1.0 r) * (1.0 g) * (1.0 b)]
This plug-in also works with layers. The softwares readme.txt file provides you with more
information about its features.
Another Adobe Photoshop plug-in provided by NVIDIA is the Photoshop Compression
plug-in. It is used by choosing <Save As> in Adobe Photoshop and then the <DDS> file for-
mat. The following dialog provides a wide variety of features:
Fundamentals of Vertex Shaders 13
Figure 7: NVIDIA Normal
Map Generator
Team LRN
A 3D preview shows the different quality levels that result from different compression formats.
This tool can additionally generate mip-maps and convert a height map to a normal map. The
provided readme file is very instructive and explains the hundreds of features of this tool. As
the name implies, both tools support Adobe Photoshop 5.0 and higher.
Diffusion Cubemap Tool
Kenneth Hurley wrote a tool that helps you produce diffusion cube maps. It aids in the extrac-
tion of cube maps from digital pictures. The pictures are of a completely reflective ball. The
program also allows you to draw an exclusion rectangle to remove the picture taker from the
cube map.
To extract the reflection maps, first load in the picture and use the mouse to draw the
ellipse enclosed in a rectangle. This rectangle should be stretched and moved so that the ellipse
falls on the edges of the ball. Then set which direction is associated with the picture in the
menu options. The following screen shots use the X and Z direction:
14 Part 1: Introduction to Shader Programming
Figure 8: NVIDIA Compression plug-in
Figure 9: Negative X sphere picture Figure 10: Negative Z sphere picture
Team LRN
The cube maps are generated with the Generate menu option. The program, the source code,
and much more information can be found in [Hurley].
DLL Detective with Direct3D Plug-in
dm Moravnszky wrote a tool called DLL Detective. It is not only very useful as a perfor-
mance analysis tool but also for vertex and pixel shader programming:
It is able to intercept vertex and pixel shaders, disassemble them, and write them into a file. A
lot of different graphs show the usage of the Direct3D API under different conditions to help
find performance leaks. You can even suppress API calls to simulate other conditions. To
impede the parallelism of the CPU and GPU usage, you can lock the rendertarget buffer.
DLL Detective is especially suited for instrumenting games or any other applications
which run in full-screen mode, preventing easy access to other windows (like DLL Detective,
for example). To instrument such programs, DLL Detective can be configured to control
instrumentation via a multimonitor setup and even from another PC over a network.
The full source code and compiled binaries can be downloaded from dm
Moravnszkys web site at http://n.ethz.ch/student/adammo/DLLDetective/index.html.
Fundamentals of Vertex Shaders 15
Figure 11: dm
Moravnszkys
DLL Detective
Team LRN
3D Studio Max 4.x/gmax 1.1
The new 3D Studio Max 4.x gives a graphic artist the ability to produce vertex shader code and
pixel shader code while producing models and animations.
A WYSIWYG view of your work will appear by displaying multitextures, true transparency,
opacity mapping, and the results of custom pixel and vertex shaders.
gmax is a derivative of 3D Studio Max 4.x that supports vertex and pixel shader program-
ming. However, the gmax free product does not provide user interface to access or edit these
controls. Find more information at www.discreet.com.
Vertex Shader Architecture
Lets get deeper into vertex shader
programming by looking at a graphi-
cal representation of the vertex
shader architecture:
Figure 13: Vertex
shader architecture
16 Part 1: Introduction to Shader Programming
Figure 12: 3D
Studio Max
4.x/gmax 1.1
Team LRN
All data in a vertex shader is represented by 128-bit quad-floats (4x32-bit):
A hardware vertex shader can be seen as a typical SIMD (Single Instruction Multiple Data)
processor, as you are applying one instruction and affecting a set of up to four 32-bit variables.
This data format is very useful because most of the transformation and lighting calculations are
performed using 4x4 matrices or quaternions. The instructions are very simple and easy to
understand. The vertex shader does not allow any loops, jumps, or conditional branches, which
means that it executes the program linearly one instruction after the other. The maximum
length of a vertex shader program in DirectX 8.x is limited to 128 instructions. Combining ver-
tex shaders to have one to compute the transformation and another to compute the lighting is
impossible. Only one vertex shader can be active at a time, and the active vertex shader must
compute all required per-vertex output data.
A vertex shader uses up to 16 input registers (named v0-v15, where each register consists
of 128-bit (4x32-bit) quad-floats) to access vertex input data. The vertex input register can eas-
ily hold the data for a typical vertex: its position coordinates, normal, diffuse and specular
color, fog coordinate, and point size information with space for the coordinates of several
textures.
The constant registers (constant memory) are loaded by the CPU before the vertex shader
starts executing parameters defined by the programmer. The vertex shader is not able to write
to the constant registers. They are used to store parameters such as light position, matrices,
procedural data for special animation effects, vertex interpolation data for morphing/key frame
interpolation, and more. The constants can be applied within the program and can even be
addressed indirectly with the help of the address register a0.x, but only one constant can be
used per instruction. If an instruction needs more than one constant, it must be loaded into one
of the temporary registers before it is required. The names of the constant registers are c0-c95
or, in the case of the ATI RADEON 8500, c0-c191.
The temporary registers consist of 12 registers used to perform intermediate calculations.
They can be used to load and store data (read/write). The names of the temporary registers are
r0-r11.
There are up to 13 output registers (vertex output), depending on the underlying hardware.
The names of the output registers always start with o for output. The vertex output is available
per rasterizer, and your vertex shader program has write-only access to it. The final result is yet
another vertex, a vertex transformed to the homogenous clip space. The following table is an
overview of all available registers.
Fundamentals of Vertex Shaders 17
Figure 14: 128 bits
Team LRN
Registers Number of Registers Properties
Input (v0-v15) 16 RO1
Output (o*) GeForce 3/4TI: 9; RADEON 8500: 11 WO
Constants (c0-c95) vs.1.1 Specification: 96; RADEON 8500: 192 RO1
Temporary (r0-r11) 12 R1W3
Address (a0.x) 1 (vs.1.1 and higher) WO (W: only with mov)
An identifier of the streaming nature of this vertex shader architecture is the read-only input
registers and the write-only output registers.
High-Level View of Vertex Shader Programming
Only one vertex shader can be active at a time. It is a good idea to write vertex shaders on a
per-task basis. The overhead of switching between different vertex shaders is smaller than, for
example, a texture change. So if an object needs a special form of transformation or lighting, it
will get the proper shader for this task. Lets build an abstract example:
You are shipwrecked on a foreign planet. Dressed in your regular armor, armed only with a
jigsaw, you move through the candlelit cellars. A monster appears, and you crouch behind one
of those crates that one normally finds on other planets. While thinking about your destiny as a
hero who saves worlds with jigsaws, you start counting the number of vertex shaders for this
scene.
There is one for the monster to animate it, light it, and perhaps to reflect its environment.
Other vertex shaders will be used for the floor, the walls, the crate, the camera, the candlelight,
and your jigsaw. Perhaps the floor, the walls, the jigsaw, and the crate use the same shader, but
the candlelight and the camera might each use one of their own. It depends on your design and
the power of the underlying graphic hardware.
Every vertex shader-driven program must run through the following steps:
n
Check for vertex shader support by checking the D3DCAPS8::VertexShaderVersion field.
n
Declare the vertex shader with the D3DVSD_* macros to map vertex buffer streams to
input registers.
n
Set the vertex shader constant registers with SetVertexShaderConstant().
n
Compile previously written vertex shader with D3DXAssembleShader*() (this could be
precompiled using a shader assembler).
n
Create a vertex shader handle with CreateVertexShader().
n
Set a vertex shader with SetVertexShader() for a specific object.
n
Delete a vertex shader with DeleteVertexShader().
18 Part 1: Introduction to Shader Programming
Note: You might also use vertex shaders on a per-object or per-mesh basis. If,
for example, a *.md3 model consists of, lets say, ten meshes, you can use ten
different vertex shaders, but that might harm your game performance.
Team LRN
Check for Vertex Shader Support
It is important to check the installed vertex shader software or hardware implementation of the
end-user hardware. If there is a lack of support for specific features, then the application can
fall back to a default behavior or give the user a hint as to what he might do to enable the
required features. The following statement checks for support of vertex shader version 1.1:
if( pCaps->VertexShaderVersion < D3DVS_VERSION(1,1) )
return E_FAIL;
The following statement checks for support of vertex shader version 1.0:
if( pCaps->VertexShaderVersion < D3DVS_VERSION(1,0) )
return E_FAIL;
The D3DCAPS8 structure caps must be filled in the startup phase of the application with a call
to GetDeviceCaps(). If you use the common files framework provided with the DirectX 8.1
SDK, this is done by the framework. If your graphics hardware does not support your
requested vertex shader version, you must switch to software vertex shaders by using the
D3DCREATE_SOFTWARE_VERTEXPROCESSING flag in the CreateDevice() call. The
previously mentioned optimized software implementations made by Intel and AMD for their
respective CPUs will then process the vertex shaders.
Supported vertex shader versions are:
Version Functionality
0.0 DirectX 7
1.0 DirectX 8 without address register a0
1.1 DirectX 8 and DirectX 8.1 with one address register a0
2.0 DirectX 9
The only difference between levels 1.0 and 1.1 is the support of the a0 register. The DirectX
8.0 and DirectX 8.1 reference rasterizer and the software emulation delivered by Microsoft
and written by Intel and AMD for their respective CPUs support version 1.1. At the time
publication, only GeForce3/4TI and RADEON 8500-driven boards support version 1.1 in hard-
ware. No known graphics card supports vs.1.0-only at the time of publication, so this is a
legacy version.
Vertex Shader Declaration
You must declare a vertex shader before using it. This declaration can be called a static exter-
nal interface. An example might look like this:
float c[4] = {0.0f,0.5f,1.0f,2.0f};
DWORD dwDecl0[] = {
D3DVSD_STREAM(0),
D3DVSD_REG(0, D3DVSDT_FLOAT3 ), // input register v0
D3DVSD_REG(5, D3DVSDT_D3DCOLOR ), // input Register v5
// set a few constants
D3DVSD_CONST(0,1),*(DWORD*)&c[0],*(DWORD*)&c[1],*(DWORD*)&c[2],*(DWORD*)&c[3],
D3DVSD_END()
};
Fundamentals of Vertex Shaders 19
Team LRN
This vertex shader declaration sets data stream 0 with D3DVSD_STREAM(0). Later,
SetStreamSource() binds a vertex buffer to a device data stream by using this declaration. You
are able to feed different data streams to the Direct3D rendering engine this way.
You must declare which input vertex
properties or incoming vertex data has to be
mapped to which input register.
D3DVSD_REG binds a single vertex regis-
ter to a vertex element/property from the
vertex stream. In our example, a
D3DVSDT_FLOAT3 value should be
placed in the first input register, and a D3DVSDT_D3DCOLOR color value should be placed
in the sixth input register. For example, the position data could be processed by the input regis-
ter 0 (v0) with D3DVSD_REG(0, D3DVSDT_FLOAT3) and the normal data could be
processed by input register 3 (v3) with D3DVSD_REG(3, D3DVSDT_FLOAT3).
How a developer maps each input vertex property to a specific input register is only
important if one wants to use N-Patches because the N-Patch tessellator needs the position data
in v0 and the normal data in v3. Otherwise, the developer is free to define the mapping as he
sees fit. For example, the position data could be processed by the input register 0 (v0) with
D3DVSD_REG(0, D3DVSDT_FLOAT3), and the normal data could be processed by input
register 3 (v3) with D3DVSD_REG(3, D3DVSDT_FLOAT3).
The second parameter of D3DVSD_REG specifies the dimensionality and arithmetic data type.
The following values are defined in d3d8types.h:
// bit declarations for _Type fields
#define D3DVSDT_FLOAT1 0x00 // 1D float expanded to (value, 0., 0., 1.)
#define D3DVSDT_FLOAT2 0x01 // 2D float expanded to (value, value, 0., 1.)
#define D3DVSDT_FLOAT3 0x02 // 3D float expanded to (value, value, value, 1.)
20 Part 1: Introduction to Shader Programming
Note: For example, one data stream could hold
positions and normals, while a second held color
values and texture coordinates. This also makes
switching between single-texture rendering and
multitexture rendering trivial: Just dont enable the
stream with the second set of texture coordinates.
Note: In contrast, the mapping of the vertex data input to specific registers is fixed for the
fixed-function pipeline. d3d8types.h holds a list of #defines that predefine the vertex input
for the fixed-function pipeline. Specific vertex elements such as position or normal must be
placed in specified registers located in the vertex input memory. For example, the vertex
position is bound by D3DVSDE_POSITION to Register 0, the diffuse color is bound by
D3DVSDE_DIFFUSE to Register 5, etc. Heres the complete list from d3d8types.h:
#define D3DVSDE_POSITION 0
#define D3DVSDE_BLENDWEIGHT 1
#define D3DVSDE_BLENDINDICES 2
#define D3DVSDE_NORMAL 3
#define D3DVSDE_PSIZE 4
#define D3DVSDE_DIFFUSE 5
#define D3DVSDE_SPECULAR 6
#define D3DVSDE_TEXCOORD0 7
#define D3DVSDE_TEXCOORD1 8
#define D3DVSDE_TEXCOORD2 9
#define D3DVSDE_TEXCOORD3 10
#define D3DVSDE_TEXCOORD4 11
#define D3DVSDE_TEXCOORD5 12
#define D3DVSDE_TEXCOORD6 13
#define D3DVSDE_TEXCOORD7 14
#define D3DVSDE_POSITION2 15
#define D3DVSDE_NORMAL2 16
Team LRN
#define D3DVSDT_FLOAT4 0x03 // 4D float
// 4D packed unsigned bytes mapped to 0. to 1. range
// Input is in D3DCOLOR format (ARGB) expanded to (R, G, B, A)
#define D3DVSDT_D3DCOLOR 0x04
#define D3DVSDT_UBYTE4 0x05 // 4D unsigned byte
// 2D signed short expanded to (value, value, 0., 1.)
#define D3DVSDT_SHORT2 0x06
#define D3DVSDT_SHORT4 0x07 // 4D signed short
D3DVSD_CONST loads the constant values into the vertex shader constant memory. The first
parameter is the start address of the constant array to begin filling data. Possible values range
from 0-95 or, in the case of the RADEON 8500, from 0-191. We start at address 0. The second
number is the number of constant vectors (quad-float) to load. One vector is 128 bits long, so
we load four 32-bit FLOATs at once. If you want to load a 4x4 matrix, you would use the fol-
lowing statement to load four 128-bit quad-floats into the constant registers c0-c3:
float c[16] = (0.0f, 0.5f, 1.0f, 2.0f,
0.0f, 0.5f, 1.0f, 2.0f,
0.0f, 0.5f, 1.0f, 2.0f,
0.0f, 0.5f, 1.0f, 2.0f);
D3DVSD_CONST(0, 4), *(DWORD*)&c[0],*(DWORD*)&c[1],*(DWORD*)&c[2],*(DWORD*)&c[3],
*(DWORD*)&c[4],*(DWORD*)&c[5],*(DWORD*)&c[6],*(DWORD*)&c[7],
*(DWORD*)&c[8],*(DWORD*)&c[9],*(DWORD*)&c[10],*(DWORD*)&c[11],
*(DWORD*)&c[12],*(DWORD*)&c[13],*(DWORD*)&c[14],*(DWORD*)&c[15],
D3DVSD_END generates an END token to mark the end of the vertex shader declaration.
Another example is:
float c[4] = {0.0f,0.5f,1.0f,2.0f};
DWORD dwDecl[] = {
D3DVSD_STREAM(0),
D3DVSD_REG(0, D3DVSDT_FLOAT3 ), //input register v0
D3DVSD_REG(3, D3DVSDT_FLOAT3 ), // input register v3
D3DVSD_REG(5, D3DVSDT_D3DCOLOR ), // input register v5
D3DVSD_REG(7, D3DVSDT_FLOAT2 ), // input register v7
D3DVSD_CONST(0,1),*(DWORD*)&c[0],*(DWORD*)&c[1],*(DWORD*)&c[2],*(DWORD*)&c[3],
D3DVSD_END()
};
Data stream 0 is set with D3DVSD_STREAM(0). The position values (value, value, value,
1.0) might be bound to v0, the normal values might be bound to v3, the diffuse color might be
bound to v5, and one texture coordinate (value, value, 0.0, 1.0) might be bound to v7. The con-
stant register c0 gets one 128-bit value.
Fundamentals of Vertex Shaders 21
Note: GeForce3/4TI doesnt support D3DVSDT_UBYTE4, as
indicated by the D3DVTXPCAPS_NO_VSDT_UBYTE4 caps bit.
Team LRN
Set the Vertex Shader Constant Registers
You will fill the vertex shader constant registers with SetVertexShaderConstant() and get the
values from these registers with GetVertexShaderConstant():
// Set the vertex shader constants
m_pd3dDevice->SetVertexShaderConstant( 0, &vZero, 1 );
m_pd3dDevice->SetVertexShaderConstant( 1, &vOne, 1 );
m_pd3dDevice->SetVertexShaderConstant( 2, &vWeight, 1 );
m_pd3dDevice->SetVertexShaderConstant( 4, &matTranspose, 4 );
m_pd3dDevice->SetVertexShaderConstant( 8, &matCameraTranspose, 4 );
m_pd3dDevice->SetVertexShaderConstant( 12, &matViewTranspose, 4 );
m_pd3dDevice->SetVertexShaderConstant( 20, &fLight, 1 );
m_pd3dDevice->SetVertexShaderConstant( 21, &fDiffuse, 1 );
m_pd3dDevice->SetVertexShaderConstant( 22, &fAmbient, 1 );
m_pd3dDevice->SetVertexShaderConstant( 23, &fFog, 1 );
m_pd3dDevice->SetVertexShaderConstant( 24, &fCaustics, 1 );
m_pd3dDevice->SetVertexShaderConstant( 28, &matProjTranspose, 4 );
SetVertexShaderConstant() is declared as
HRESULT SetVertexShaderConstant(
DWORD Register,
CONST void* pConstantData,
DWORD ConstantCount);
As stated earlier, there are at least 96 constant registers (RADEON 8500 has 192) that can be
filled with four floating-point values before the vertex shader is executed. The first parameter
holds the register address at which to start loading data into the vertex constant array. The last
parameter holds the number of constants (4 x 32-bit values) to load into the vertex constant
array. So in the first row above, vZero will be loaded into register 0, matTranspose will be
loaded into registers 4 through 7, matViewTranspose will be loaded into registers 12 through
15, registers 16 through 19 are not used, fLight is loaded into register 20, and registers 25
through 27 are not used.
We have learned how to check the supported version number of the vertex shader hardware,
how to declare a vertex shader, and how to set the constants in the constant registers of a vertex
shader unit. Next we shall learn how to write and compile a vertex shader program.
Write and Compile a Vertex Shader
Before we are able to compile a vertex shader, we must write one. I would like to give you a
high-level overview of the instruction set first and then give further details of vertex shader
programming in the next article, Programming Vertex Shaders.
The syntax for every instruction is:
OpName dest, [-]s1 [,[-]s2 [,[-]s3]] ;comment
e.g.
mov r1, r2
mad r1, r2, -r3, r4 ; contents of r3 are negated
22 Part 1: Introduction to Shader Programming
Note: So whats the difference between D3DVSD_CONST used in the vertex shader
declaration and SetVertexShaderConstant()? D3DVSD_CONST can be used only once,
while SetVertexShaderConstant() can be used before every DrawPrimitive*() call.
Team LRN
There are 17 different instructions:
Instruction Parameters Action
add dest, src1, src2 add src1 to src2 (and the optional negation creates substraction)
dp3 dest, src1, src2 three-component dot product
dest.x = dest.y = dest.z = dest.w =
(src1.x * src2.x) + (src1.y * src2.y) + (src1.z * src2.z)
dp4 dest, src1, src2 four-component dot product
dest.w = (src1.x * src2.x) + (src1.y * src2.y) + (src1.z *
src2.z) + (src1.w * src2.w);
dest.x = dest.y = dest.z = the scalar result of dp4
What is the difference between dp4 and mul? dp4 produces a scalar product
and mul is a component by component vector product.
dst dest, src1, src2 The dst instruction works like this: The first source operand (src1) is assumed to
be the vector (ignored, d*d, d*d, ignored) and the second source operand
(src2) is assumed to be the vector (ignored, 1/d, ignored, 1/d).
Calculate distance vector:
dest.x = 1;
dest.y = src1.y * src2.y
dest.z = src1.z
dest.w = src2.w
dst is useful to calculate standard attenuation. Here is a code snippet that
might calculate the attenuation for a point light:
; r7.w = distance * distance = (x*x) + (y*y) + (z*z)
dp3 r7.w, VECTOR_VERTEXTOLIGHT,
VECTOR_VERTEXTOLIGHT
; VECTOR_VERTEXTOLIGHT.w = 1/sqrt(r7.w)
; = 1/||V|| = 1/distance
rsq VECTOR_VERTEXTOLIGHT.w, r7.w
...
; Get the attenuation
; d = distance
; Parameters for dst:
; src1 = (ignored, d * d, d * d, ignored)
; src2 = (ignored, 1/d, ignored, 1/d)
;
; r7.w = d * d
; VECTOR_VERTEXTOLIGHT.w = 1/d
dst r7, r7.wwww, VECTOR_VERTEXTOLIGHT.wwww
; dest.x = 1
; dest.y = src0.y * src1.y
; dest.z = src0.z
; dest.w = src1.w
; r7(1, d * d * 1 / d, d * d, 1/d)
; c[LIGHT_ATTENUATION].x = a0
; c[LIGHT_ATTENUATION].y = a1
; c[LIGHT_ATTENUATION].z = a2
; (a0 + a1*d + a2* (d * d))
dp3 r7.w, r7, c[LIGHT_ATTENUATION]
; 1 / (a0 + a1*d + a2* (d * d))
rcp ATTENUATION.w, r7.w
...
; Scale the light factors by the attenuation
mul r6, r5, ATTENUATION.w
Fundamentals of Vertex Shaders 23
Team LRN
Instruction Parameters Action
expp dest, src.w Exponential 10-bit precision
------------------------------------------
float w = src.w;
float v = (float)floor(src.w);
dest.x = (float)pow(2, v);
dest.y = w - v;
// Reduced precision exponent
float tmp = (float)pow(2, w);
DWORD tmpd = *(DWORD*)&tmp & 0xffffff00;
dest.z = *(float*)&tmpd;
dest.w = 1;
--------------------------------------------
Shortcut:
dest.x = 2 **(int) src.w
dest.y = mantissa(src.w)
dest.z = expp(src.w)
dest.w = 1.0
lit dest, src Calculates lighting coefficients from two dot products and a power.
---------------------------------------------
To calculate the lighting coefficients, set up the registers as shown:
src.x = N*L; the dot product between normal and direction to light
src.y = N*H; the dot product between normal and half vector
src.z = ignored; this value is ignored
src.w = specular power; the value must be between 128.0 and 128.0
----------------------------------------------
usage:
dp3 r0.x, rn, c[LIGHT_POSITION]
dp3 r0.y, rn, c[LIGHT_HALF_ANGLE]
mov r0.w, c[SPECULAR_POWER]
lit r0, r0
------------------------------------------------
dest.x = 1.0;
dest.y = max (src.x, 0.0, 0.0);
dest.z= 0.0;
if (src.x > 0.0 && src.w == 0.0)
dest.z = 1.0;
else if (src.x > 0.0 && src.y > 0.0)
dest.z = (src.y)src.w
dest.w = 1.0;
-----------------------------------------------------
Shortcut:
dest.x = exponent((int)src.w)
dest.y = mantissa(src.w)
dest.z = log2(src.w)
dest.w = 1.0
24 Part 1: Introduction to Shader Programming
Team LRN
Instruction Parameters Action
logp dest, src.w Logarithm 10-bit precision
---------------------------------------------------
float v = ABSF(src.w);
if (v != 0)
{
int p = (int)(*(DWORD*)&v >> 23) - 127;
dest.x = (float)p; // exponent
p = (*(DWORD*)&v & 0x7FFFFF) | 0x3f800000;
dest.y = *(float*)&p; // mantissa;
float tmp = (float)(log(v)/log(2));
DWORD tmpd = *(DWORD*)&tmp & 0xffffff00;
dest.z = *(float*)&tmpd;
dest.w = 1;
}
else
{
dest.x = MINUS_MAX();
dest.y = 1.0f;
dest.z = MINUS_MAX();
dest.w = 1.0f;
}
mad dest, src1, src2,
src3
dest = (src1 * src2) + src3
max dest, src1, src2 dest = (src1 >= src2)?src1:src2
min dest, src1, src2 dest = (src1 < src2)?src1:src2
mov dest, src move
Optimization tip: Question every use of mov, because there might be methods
that perform the desired operation directly from the source register or accept
the required output register as the destination.
mul dest, src1, src2 set dest to the component by component product of src1 and src2
; To calculate the Cross Product (r5 = r7 X r8),
; r0 used as a temp
mul r0,r7.zxyw,r8.yzxw
mad r5,r7.yzxw,r8.zxyw,r0
nop do nothing
rcp dest, src.w if(src.w == 1.0f)
{
dest.x = dest.y = dest.z = dest.w = 1.0f;
}
else if(src.w == 0)
{
dest.x = dest.y = dest.z = dest.w = PLUS_INFINITY();
}
else
{
dest.x = dest.y = dest.z = m_dest.w = 1.0f/src.w;
}
Division:
; scalar r0.x = r1.x/r2.x
RCP r0.x, r2.x
MUL r0.x, r1.x, r0.x
Fundamentals of Vertex Shaders 25
Team LRN
Instruction Parameters Action
rsq dest, src reciprocal square root of src
(much more useful than straight square root):
float v = ABSF(src.w);
if(v == 1.0f)
{
dest.x = dest.y = dest.z = dest.w = 1.0f;
}
else if(v == 0)
{
dest.x = dest.y = dest.z = dest.w = PLUS_INFINITY();
}
else
{
v = (float)(1.0f / sqrt(v));
dest.x = dest.y = dest.z = dest.w = v;
}
Square root:
; scalar r0.x = sqrt(r1.x)
RSQ r0.x, r1.x
MUL r0.x, r0.x, r1.x
sge dest, src1, src2 dest = (src1 >=src2) ? 1 : 0
useful to mimic conditional statements:
; compute r0 = (r1 >= r2) ? r3 : r4
; one if (r1 >= r2) holds, zero otherwise
SGE r0, r1, r2
ADD r1, r3, r4
; r0 = r0*(r3r4) + r4 = r0*r3 + (1r0)*r4
; effectively, LERP between extremes of r3 and r4
MAD r0, r0, r1, r4
slt dest, src1, src2 dest = (src1 < src2) ? 1 : 0
You can download this list as a Word file from www.shaderx.com. Check out the SDK for
additional information.
The Vertex Shader ALU is a multithreaded vector processor that operates on quad-float
data. It consists of two functional units. The SIMD Vector Unit is responsible for the mov, mul,
add, mad, dp3, dp4, dst, min, max, slt, and sge instructions. The Special Function Unit is
responsible for the rcp, rsq, logp, expp, and lit instructions. Most of these instructions take one
cycle to execute, although rcp and rsq take more than one cycle under specific circumstances.
They take only one slot in the vertex shader, but they actually take longer than one cycle to
execute when the result is used immediately because that leads to a register stall.
Application Hints
rsq, for example, is used in normalizing vectors to be used in lighting equations. The exponen-
tial instruction expp can be used for fog effects, procedural noise generation (see the NVIDIA
Perlin Noise example on the companion CD), behavior of particles in a particle system (see the
NVIDIA Particle System example on the companion CD), or to implement a system for how
objects in a game are damaged. You will use it when a fast-changing function is necessary.
This is contrary to the use of logarithm functions with logp, which are useful if extremely slow
growth is necessary (also they grow pretty fast at the beginning). A log function can be the
26 Part 1: Introduction to Shader Programming
Team LRN
inverse of an exponential function, meaning it undoes the operation of the exponential
function.
The lit instruction deals by default with directional lights. It calculates the diffuse and
specular factors with clamping based on N * L and N * H and the specular power. There is no
attenuation involved, but you can use an attenuation level separately with the result of lit by
using the dst instruction. This is useful for constructing attenuation factors for point and spot
lights. The min and max instructions allow for clamping and absolute value computation.
Complex Instructions in the Vertex Shader
There are also complex instructions that are supported by the vertex shader. The term macro
should not be used to refer to these instructions because they are not simply substituted like a
C-preprocessor macro. You should think carefully before using these instructions. If you use
them, you might lose control over your 128-instruction limit and possible optimization path(s).
On the other hand, the software emulation mode provided by Intel or by AMD for their proces-
sors is able to optimize a m4x4 complex instruction (and perhaps others now or in the future).
It is also possible that in the future, some graphics hardware may use gate count to optimize
the m4x4. So, if you need, for example, four dp4 calls in your vertex shader assembly source,
it might be a good idea to replace them by m4x4. If you have decided to use an m4x4 instruc-
tion in your shader, you should not use a dp4 call on the same data later because there are
slightly different transformation results. If, for example, both instructions are used for position
calculation, z-fighting could result:
Macro Parameters Action Clocks
expp dest, src1 Provides exponential with full precision to at least 1/220 12
frc dest, src1 Returns fractional portion of each input component 3
log dest, src1 Provides log2(x) with full float precision of at least 1/220 12
m3x2 dest, src1, src2 Computes the product of the input vector and a 3x2 matrix 2
m3x3 dest, src1, src2 Computes the product of the input vector and a 3x3 matrix 3
m3x4 dest, src1, src2 Computes the product of the input vector and a 3x4 matrix 4
m4x3 dest, src1, src2 Computes the product of the input vector and a 4x3 matrix 3
m4x4 dest, src1, src2 Computes the product of the input vector and a 4x4 matrix 4
You are able to perform all transform and lighting operations with these instructions. If it
seems to you that some instructions are missing, rest assured that you can achieve them
through the existing instructions. For example, the division of two numbers can be accom-
plished with a reciprocal and a multiply. You can even implement the whole fixed-function
pipeline by using these instructions in a vertex shader. This is shown in the NVIDIA NVLink
example on the companion CD.
Putting It All Together
Now lets see how these registers and instructions are typically used in the vertex shader ALU.
In vs.1.1, there are 16 input registers, 96 constant registers, 12 temporary registers, 1
address register, and up to 13 output registers per rasterizer. Each register can handle 4x32-bit
Fundamentals of Vertex Shaders 27
Team LRN
values. Each 32-bit value is accessible via an x, y, z, and w subscript. That is, a 128-bit value
consists of an x, y, z, and w value. To access these register components, you must add .x, .y, .z,
and .w at the end of the register name. Lets start with the input registers.
Using the Input Registers
The 16 input registers can be accessed by using their names, v0 - v15. Typical values provided
to the input vertex registers are:
n
Position (x,y,z,w)
n
Diffuse color (r,g,b,a) 0.0 to +1.0
n
Specular color (r,g,b,a) 0.0 to +1.0
n
Up to eight texture coordinates (each as s, t, r, q or u, v, w, q) but normally four or six,
depending on hardware support
n
Fog (f,*,*,*) value used in fog equation
n
Point size (p,*,*,*)
You can access the x-component of the position with v0.x, the y-component with v0.y, and so
on. If you need to know the green component of the RGBA diffuse color, you check v1.y. You
may set the fog value, for example, into v7.x. The other three 32-bit components, v7.y, v7.z,
and v7.w, would not be used. The input registers are read only. Each instruction may access
only one vertex input register. Unspecified components of the input register default to 0.0 for
the x, y, and z components and to 1.0 for the w component. In the following example, the
four-component dot product between each of c0-c3 and v0 is stored in oPos:
dp4 oPos.x , v0 , c0
dp4 oPos.y , v0 , c1
dp4 oPos.z , v0 , c2
dp4 oPos.w , v0 , c3
Such a code fragment is usually used to map from projection space into clip space with the
help of the already concatenated world, view, and projection matrices. The four component dot
product performs the following calculation:
oPos.x = (v0.x * c0.x) + (v0.y * c0.y) + (v0.z * c0.z) + (v0.w * c0.w)
Given that we use unit length (normalized) vectors, it is known that the dot product of two vec-
tors will always range between [1, 1]. Therefore, oPos will always get values in that range.
Alternatively, you could use:
m4x4 oPos, v0 , c0
Dont forget to use those complex instructions consistently throughout your vertex shader
because as discussed earlier, there might be slight differences between dp4 and m4x4 results.
You are restricted to using only one input register in each instruction.
All data in an input register remains persistent throughout the vertex shader execution and
even longer. That means they retain their data longer than the lifetime of a vertex shader. So it
is possible to reuse the data of the input registers in the next vertex shader.
28 Part 1: Introduction to Shader Programming
Team LRN
Using the Constant Registers
Typical uses for the constant registers include:
n
Matrix data quad-floats are typically one row of a 4x4 matrix
n
Light characteristics (position, attenuation, etc.)
n
Current time
n
Vertex interpolation data
n
Procedural data
There are 96 quad-floats (or in the case of the RADEON 8500, 192 quad-floats) for storing
constant data. This reasonably large set of matrices can be used, for example, for indexed ver-
tex blending, more commonly known as matrix palette skinning.
The constant registers are read only from the perspective of the vertex shader, whereas the
application can read and write into the constant registers. The constant registers retain their
data longer than the life-time of a vertex shader, so it is possible to reuse this data in the next
vertex shader. This allows an app to avoid making redundant SetVertexShaderConstant() calls.
Reads from out-of-range constant registers return (0.0, 0.0, 0.0, 0.0).
You can use only one constant register per instruction, but you can use it several times. For
example:
; the following instruction is legal
mul r5, c11, c11 ; The product of c11 and c11 is stored in r5
; but this is illegal
add v0, c4, c3
A more complicated-looking, but valid, example is:
; dest = (src1 * src2) + src3
mad r0, r0, c20, c20 ; multiplies r0 with c20 and adds c20 to the result
Using the Address Register
You access the address registers with a0 to an (more than one address register should be avail-
able in vertex shader versions higher than 1.1). The only use of a0 in vs.1.1 is as an indirect
addressing operator to offset constant memory.
c[a0.x + n] ; supported only in version 1.1 and higher
; n is the base address and a0.x is the address offset
Here is an example using the address register:
...
//Set 1
mov a0.x,r1.x
m4x3 r4,v0,c[a0.x + 9];
m3x3 r5,v3,c[a0.x + 9];
...
Depending on the value that is stored in temporary register r1.x, different constant registers are
used in the m4x3 and m3x3 instructions. Please note that register a0 only stores whole num-
bers and no fractions (integers only) and that a0.x is the only valid component of a0. Further, a
vertex shader may write to a0.x only via the mov instruction.
Fundamentals of Vertex Shaders 29
Team LRN
Using the Temporary Registers
You can access the 12 temporary registers using r0 to r11. Here are a few examples:
dp3 r2, r1, -c4 ; A three-component dot product: dest.x = dest.y = dest.z =
; dest.w = (r1.x * -c4.x) + (r1.y * -c4.y) + (r1.z * -c4.z)
...
mov r0.x, v0.x
mov r0.y, c4.w
mov r0.z, v0.y
mov r0.w, c4.w
Each temporary register has single write and triple read access. Therefore, an instruction could
have the same temporary register as a source three times. Vertex shaders cannot read a value
from a temporary register before writing to it. If you try to read a temporary register that was
not filled with a value, the API will give you an error message while creating the vertex shader
(== CreateVertexShader()).
Using the Output Registers
There are up to 13 write-only output registers that can be accessed using the following register
names. They are defined as the inputs to the rasterizer; the name of each register is preceded by
a lowercase o. The output registers are named to suggest their use by pixel shaders.
Name Value Description
oDn 2 quad-floats Output color data directly to the pixel shader. Required
for diffuse color (oD0) and specular color (oD1).
oPos 1 quad-float Output position in homogenous clipping space. Must be
written by the vertex shader.
oTn up to 8 quad-floatsGeforce 3:
4RADEON 8500: 6
Output texture coordinates. Required for maximum
number of textures simultaneously bound to the texture
blending stage.
oPts.x 1 scalar float Output point-size registers. Only the scalar x-component
of the point size is functional.
oFog.x 1 scalar float The fog factor to be interpolated and then routed to the
fog table. Only the scalar x-component is functional.
Here is a typical example that shows how to use the oPos, oD0, and oT0 registers:
dp4 oPos.x , v0 , c4 ; emit projected x position
dp4 oPos.y , v0 , c5 ; emit projected y position
dp4 oPos.z , v0 , c6 ; emit projected z position
dp4 oPos.w , v0 , c7 ; emit projected w position
mov oD0, v5 ; set the diffuse color
mov oT0, v2 ; outputs the texture coordinates to oT0 from input register v2
Using the four dp4 instructions to map from projection to clip space with the already concate-
nated world, view, and projection matrices was shown above. The first mov instruction moves
the content of the v5 input register into the color output register, and the second mov instruc-
tion moves the values of the v2 register into the first output texture register.
30 Part 1: Introduction to Shader Programming
Note: Beware of a0.x if there is only a software emulation mode;
performance can be significantly reduced [Pallister].
Team LRN
Using the oFog.x output register is shown in the following example:
; Scale by fog parameters :
; c5.x = fog start
; c5.y = fog end
; c5.z = 1/range
; c5.w = fog max
dp4 r2, v0, c2 ; r2 = distance to camera
sge r3, c0, c0 ; r3 = 1
add r2, r2, -c5.x ; camera space depth (z) - fog start
mad r3.x, -r2.x, c5.z, r3.x ; 1.0 - (z - fog start) * 1/range
; because fog=1.0 means no fog, and
; fog=0.0 means full fog
max oFog.x, c5.w, r3.x ; clamp the fog with our custom max value
Having a fog distance value permits more general fog effects than using the positions z or w
values. The fog distance value is interpolated before use as a distance in the standard fog equa-
tions used later in the pipeline.
Every vertex shader must write at least to one component of oPos, or you will get an error
message by the assembler.
All iterated values transferred out of the vertex shader are clamped to [0..1]. That means any
negative values are cut off to 0 and the positive values remain unchanged. If you need signed
values in the pixel shader, you must bias them in the vertex shader by multiplying them with
0.5 and adding 0.5, and then re-expand them in the pixel shader by using _bx2.
Swizzling and Masking
If you use the input, constant, and temporary registers as source registers, you can swizzle the
.x, .y, .z, and .w values independently of each other. If you use the output and temporary regis-
ters as destination registers, you can use the .x, .y, .z, and .w values as write masks. Here are
the details:
Swizzling (only source registers: vn, cn, rn)
Swizzling is very useful for efficiency, where the source registers need to be rotated (like cross
products). Another use is converting constants such as (0.5,
0.0, 1.0, 0.6) into other forms, such as (0.0, 0.0, 1.0, 0.0) or
(0.6, 1.0, 0.5, 0.6).
All registers that are used in instructions as source registers
can be swizzled. For example:
mov R1, R2.wxyz;
Fundamentals of Vertex Shaders 31
Note: Optimization tip: Emit to oPos as
early as possible to trigger parallelism in the
pixel shader. Try to reorder the vertex shader
instructions to make this happen.
Figure 15: Swizzling
Team LRN
The destination register is R1, where R could be a
write-enabled register like the output (o*) or any of the tem-
porary registers (r). The source register is R2, where R could
be an input (v), constant (c), or temporary register (source reg-
isters are located on the right side of the destination register in
the instruction syntax).
The following instruction copies the negation of R2.x into
R1.x, the negation of R2.y into R1.y and R1.z, and the nega-
tion of R2.z into R1.w. As shown, all source registers can be
negated and swizzled at the same time:
mov R1, -R2.xyyz
Masking (only destination registers: on, rn)
A destination register can mask which components are written to it. If you use R1 as the desti-
nation register (actually any write-enabled registers : o*, r), all the components are written
from R2 to R1. If you choose, for example,
mov R1.x, R2
only the x component is written to R1, whereas
mov R1.xw, R2
writes only the x and w components of R2 to R1. No swizzling or negation is supported on the
destination registers.
Here is the source for a three-vector cross-product:
; r0 = r1 x r2 (3-vector cross-product)
mul r0, r1.yzxw, r2.zxyw
mad r0, -r2.yzxw, r1.zxyw, r0
This is explained in detail in [LeGrand].
The following table summarizes swizzling and masking:
Component Modifier Description
R.[x][y][z][w] Destination mask
R.xwzy (for example) Source swizzle
R Source negation
Since any source can be negated, there is no need for a subtract instruction.
Guidelines for Writing Vertex Shaders
The most important restrictions you should remember when writing vertex shaders are the
following:
n
They must write to at least one component of the output register oPos.
n
There is a 128 instruction limit.
n
Every instruction may source no more than one constant register (e.g., add r0, c4, c3
will fail).
32 Part 1: Introduction to Shader Programming
Figure 16: Swizzling #2
Team LRN
n
Every instruction may source no more than one input register (e.g., add r0, v1, v2
will fail).
n
There are no C-like conditional statements, but you can mimic an instruction of the form
r0 = (r1 >= r2) ? r3 : r4 with the sge instruction.
n
All iterated values transferred out of the vertex shader are clamped to [0..1].
There are several ways to optimize vertex shaders. Here are a few rules of thumb:
n
See [Pallister] for information on optimizing software vertex shaders. It also helps on opti-
mizing vertex shaders running in hardware.
n
When setting vertex shader constant data, try to set all data in one SetVertexShader-
Constant() call.
n
Stop and think before using a mov instruction; you may be able to avoid it.
n
Choose instructions that perform multiple operations over instructions that perform single
operations:
mad r4,r3,c9,r4
mov oD0,r4
==
mad oD0,r3,c9,r4
n
Collapse (remove complex instructions like m4x4 or m3x3 instructions) vertex shaders
before thinking about optimizations.
n
A rule of thumb for load-balancing between the CPU/GPU: Many calculations in shaders
can be pulled outside and reformulated per-object instead of per-vertex and put into con-
stant registers. If you are doing some calculation, which is per-object rather than
per-vertex, then do it on the CPU and upload it on the vertex shader as a constant, rather
than doing it on the GPU.
Now that you have an abstract overview of how to write vertex shaders, I would like to men-
tion at least three different ways to compile one.
Compiling a Vertex Shader
Direct3D uses bytecodes, whereas OpenGL implementations parse a string. Therefore, the
Direct3D developer needs to assemble the vertex shader source with an assembler. This might
help you find bugs earlier in your development cycle, and it also reduces loadtime.
I see three different ways to compile a vertex shader:
n
Write the vertex shader source into a separate ASCII file (for example, test.vsh) and com-
pile it with a vertex shader assembler into a binary file (for example, test.vso). This file
will be opened and read at game startup. This way, not every person will be able to read
and modify your vertex shader source.
Fundamentals of Vertex Shaders 33
Note: One of the most interesting methods to optimize your applications
bandwidth usage is the usage of compressed vertex data [Calver].
Note: Dont forget that NVLink can link together already compiled shader
fragments at run time.
Team LRN
n
Write the vertex shader source into a separate ASCII file or as a char string into your *.cpp
file, and compile it on the fly while the app starts up with the D3DXAssembleShader*()
functions.
n
Write the vertex shader source in an effect file and open this effect file when the app
starts up. The vertex shader can be compiled by reading the effect files with
D3DXCreateEffectFromFile(). It is also possible to precompile an effect file. This way,
most of the handling of vertex shaders is simplified and handled by the effect file
functions.
Lets review what we have examined so far. After we checked the vertex shader support with
the D3DCAPS8::VertexShaderVersion field, we declared a vertex shader with the D3DVSD_*
macros. Then we set the constant registers with SetVertexShaderConstant() and wrote and
compiled the vertex shader.
Now we need to get a handle to call it.
Create a Vertex Shader
The CreateVertexShader() function is used to create and validate a vertex shader:
HRESULT CreateVertexShader(
CONST DWORD* pDeclaration,
CONST DWORD* pFunction,
DWORD* pHandle,
DWORD Usage);
This function takes the vertex shader declaration (which maps vertex buffer streams to differ-
ent vertex input registers) in pDeclaration as a pointer and returns the shader handle in
pHandle. The second parameter pFunction gets the vertex shader instructions compiled by
D3DXAssembleShader() / D3DXAssembleShaderFromFile() or the binary code precompiled
by a vertex shader assembler. With the fourth parameter, you can force software vertex pro-
cessing with D3DUSAGE_ SOFTWAREPROCESSING. It must be used when
D3DRS_SOFTWAREVERTEXPROCESSING is set to TRUE. By setting the software pro-
cessing path explicitly, vertex shaders are simulated by the CPU by using the software vertex
shader implementation of the CPU vendors. If a vertex shader-capable GPU is available, using
hardware vertex processing should be faster. You must use this flag or the reference rasterizer
for debugging with the NVIDIA Shader Debugger.
Set a Vertex Shader
You set a vertex shader for a specific object by using SetVertexShader() before the
DrawPrimitive*() call of this object. This function dynamically loads the vertex shader
between the primitive calls.
// set the vertex shader
m_pd3dDevice->SetVertexShader( m_dwVertexShader );
34 Part 1: Introduction to Shader Programming
Note: Another way is to use the opcodes shown in d3dtypes.h and
build your own vertex assembler/disassembler.
Team LRN
The only parameter you must provide is the handle of the vertex shader created by
CreateVertexShader(). The overhead of this call is lower than a SetTexture() call, so you are
able to use it often.
Vertex shaders are executed with SetVertexShader() as many times as there are vertices.
For example, if you try to visualize a rotating quad with four vertices implemented as an
indexed triangle list, you will see in the NVIDIA Shader Debugger that the vertex shader runs
four times before the DrawPrimitive*() function is called.
Free Vertex Shader Resources
When the game shuts down or when the device is changed, the resources used by the vertex
shader must be released. This must be done by calling DeleteVertexShader() with the vertex
shader handle:
// delete the vertex shader
if (m_pd3dDevice->m_dwVertexShader != 0xffffffff)
{
m_pd3dDevice->DeleteVertexShader( m_dwVertexShader );
m_pd3dDevice->m_dwVertexShader = 0xffffffff;
}
Summary
We have now stepped through the vertex shader creation process on a high level. Lets summa-
rize what was shown so far:
n
To use vertex shaders, you must check the vertex shader support of the hardware vertex
shader implementation installed on the end-users computer with the D3DCAPS8::Ver-
texShaderVersion field.
n
You must declare which input vertex properties or incoming vertex data have to be
mapped to which input register. This mapping is done with the D3DVSD_* macros. You
are able to fill the constant registers of the vertex shader with values by using the provided
macros or by using the SetVertexShaderConstant() function.
n
After you have prepared everything this way and you have written a vertex shader, you are
able to compile it, retrieve a handle to it by calling CreateVertexShader(), and prepare it
for execution by using SetVertexShader().
n
To release the resources that are allocated by the vertex shader you should call
DeleteVertexShader() at the end of your game.
What Happens Next?
In the next article, Programming Vertex Shaders, we will start writing our first vertex shader.
We will also discuss basic lighting algorithms and how to implement them.
Fundamentals of Vertex Shaders 35
Team LRN
References
[Bendel] Steffen Bendel, Smooth Lighting with ps.1.4, Direct3D ShaderX: Vertex and Pixel
Shader Tips and Tricks, (Wordware Publishing, Inc., 2002).
[Calver] Dean Calver, Vertex Decompression in a Shader, Direct3D ShaderX.
[Gosselin] David Gosselin, Character Animation with Direct3D Vertex Shaders, Direct3D
ShaderX.
[Hurley] Kenneth Hurley, Photorealistic Faces with Vertex and Pixel Shaders, Direct3D
ShaderX.
[Isidoro/Gosselin] John Isidoro and David Gosselin, Bubble Shader, Direct3D ShaderX.
[Le Grand] Scott Le Grand, Compendium of for Vertex Shaders, Direct3D ShaderX.
[Pallister] Kim Pallister, Optimizing Software Vertex Shaders, Direct3D ShaderX.
[Riddle/Zecha] Steven Riddle and Oliver C. Zecha, Perlin Noise and Returning Results from
Shader Programs, Direct3D ShaderX.
[Schwab] John Schwab, Basic Shader Development with Shader Studio, Direct3D ShaderX.
[Vlachos01] Alex Vlachos, Jrg Peters, Chas Boyd, and Jason L. Mitchell, Curved PN
Triangles, ACM Symposium on Interactive 3D Graphics, 2001
(http://www.ati.com/na/pages/resource_centre/dev_rel/CurvedPNTriangles.pdf).
Additional Resources
A lot of information on vertex shaders can be found at the web sites of NVIDIA (devel-
oper.nvidia.com), ATI (www.ati.com), and various others. I would like to name a few:
Author Article Published at
Richard Huddy Introduction to DX8 Vertex Shaders NVIDIA web site
Erik Lindholm, Mark J Kilgard,
Henry Moreton
SIGGRAPH 2001 A User Programma-
ble Vertex Engine
NVIDIA web site
Evan Hart, Dave Gosselin,
John Isidoro
Vertex Shading with Direct3D and
OpenGL
ATI web site
Jason L. Mitchell Advanced Vertex and Pixel Shader
Techniques
ATI web site
Philip Taylor Series of articles on shader programming http://msdn.microsoft.com/directx
Keshav B. Channa Geometry Skinning / Blending and Vertex
Lighting
http://www.flipcode.com/tutori-
als/tut_dx8shaders.shtml
Konstantin Martynenko Introduction to Shaders http://www.reactorcritical.com/
review-shadersintro/review-
shadersintro.shtml
36 Part 1: Introduction to Shader Programming
Team LRN
Acknowledgments
Id like to recognize a couple of individuals that were involved in proofreading and improving
this paper (in alphabetical order):
n
David Callele (University of Saskatchewan)
n
Jeffrey Kiel (NVIDIA)
n
Jason L. Mitchell (ATI)
Fundamentals of Vertex Shaders 37
Team LRN
Programming Vertex
Shaders
Wolfgang F. Engel
In this article, a very simple program that shows a rotating quad will evolve into a more
sophisticated application showing a Bzier patch class with a diffuse and specular reflection
model and featuring a point light source. The example applications are all built on each other
in such a way that most of the code in the previous example is reused. This way, the explana-
tion of the features stays focused on the advancements of the specific example.
RacorX
RacorX displays a green color
that is applied to the quad
evenly. This example demon-
strates the usage of the common
files framework provided with
the DirectX 8.1 SDK and shows
how to compile vertex shaders
with the D3DXAssemble-
Shader() function.
As with all the upcoming
examples, which are based on
the common files,
<Alt>+<Enter> switches
between the windowed and
full-screen mode, <F2> gives
you a selection of the usable
drivers, and <Esc> will shut
down the application.
38
Figure 1: RacorX
Team LRN
First lets take a look at the files you need to compile the program:
The source file is RacorX.cpp, the resource files are winmain.rc and resource.h, the icon file is
directx.ico, and the executable is RacorX.exe. The remaining files are for the use of the Visual
C/C++ 6 IDE.
To compile this example, you should link it with the following *.lib files:
n
d3d8.lib
n
d3dx8dt.lib
n
dxguid.lib
n
d3dxof.lib
n
winmm.lib
n
gdi32.lib
n
user32.lib
n
kernel32.lib
n
advapi32.lib
Most of these *.lib files are COM wrappers. The d3dx8dt.lib is the debug version of the
Direct3DX static link library.
All of these *.lib files have to be included in the Object/libary modules entry field. This is
located at Project->Settings on the Link tab:
Programming Vertex Shaders 39
Figure 2: Directory content
Figure 3: Project Settings dialog
Note: The release Direct3DX static link library is called d3dx8.lib. There is
also a *.dll version of the debug build called d3dx8d.dll in the system32
directory. It is used by linking to the d3dx8d.lib COM wrapper.
Team LRN
The provided Visual C/C++ 6 IDE workspace references the
common files in a folder with the same name.
They were added to the project by selecting Project->Add
To Project->Files:
The Common Files Framework
The common files framework helps you get up to speed because:
n
It helps to avoid how-tos for Direct3D in general, so the focus of this text is the real stuff.
n
Its common and is a tested foundation, which helps reduce the debug time.
n
All of the Direct3D samples in the DirectX SDK use it. Learning time is very short.
n
Its window mode makes debugging easier.
n
Self-developed production code could be based on the common files, so knowing them is
always an advantage.
A high-level view of the common files shows 14 *.cpp files in C:\DXSDK\samples\Multime-
dia\Common\src.
These files encapsulate the basic functionality you need to start programming a Direct3D
application. The most important, d3dapp.cpp, contains the class CD3DApplication. It provides
seven functions that can be overridden and used in the main *.cpp file of any project in this
introduction:
n
virtual HRESULT OneTimeSceneInit() { return S_OK; }
n
virtual HRESULT InitDeviceObjects() { return S_OK; }
n
virtual HRESULT RestoreDeviceObjects() { return S_OK; }
n
virtual HRESULT DeleteDeviceObjects() { return S_OK; }
n
virtual HRESULT Render() { return S_OK; }
n
virtual HRESULT FrameMove(FLOAT) { return S_OK; }
n
virtual HRESULT FinalCleanup() { return S_OK; }
All that has to be done to create an application based on this framework code is to create a new
project and new implementations of these overridable functions in the main source file. This is
also shown in all Direct3D examples in the DirectX SDK.
40 Part 1: Introduction to Shader Programming
Figure 4: Workspace
Figure 5: Adding files to a project
Team LRN
RacorX uses these framework functions in
RacorX.cpp. They can be called the public interface of
the common files framework.
The following functions are called in the following
order in RacorX.cpp at startup:
n
ConfirmDevice()
n
OneTimeSceneInit()
n
InitDeviceObjects()
n
RestoreDeviceObjects()
While the application is running, the framework calls:
n
FrameMove()
n
Render()
in a loop.
If the user resizes the window, the framework will call:
n
InvalidateDeviceObjects()
n
RestoreDeviceObjects()
If the user presses F2 or clicks File->Change Device and changes the device by choosing, for
example, another resolution or color quality, the framework will call:
n
InvalidateDeviceObjects()
n
DeleteDeviceObjects()
n
InitDeviceObjects()
n
RestoreDeviceObjects()
If the user quits the application, the framework will call:
n
InvalidateDeviceObjects()
n
DeleteDeviceObjects()
n
FinalCleanup()
There are matching functional pairs. InvalidateDeviceObjects() destroys what RestoreDevice-
Objects() has built up, and DeleteDeviceObjects() destroys what InitDeviceObjects() has built
up. The FinalCleanup() function destroys what OneTimeSceneInit() built up.
The idea is to give every functional pair its own tasks. The OneTimeSceneInit()/Final-
Cleanup() pair is called once at the beginning and at the end of a life cycle of the game. Both
are used to load or delete data that is not device-dependent. A good candidate might be geome-
try data. The target of the InitDeviceObjects()/DeleteDeviceObjects() pair is, like the name
implies, data that is device-dependent. If the already loaded data has to be changed when the
device changes, it should be loaded here. The following examples will load, create, or destroy
their vertex buffer and index buffers and their textures in these functions.
The InvalidateDeviceObjects()/RestoreDeviceObjects() pair has to react to the changes
of the window size. So, for example, code that handles the projection matrix might be placed
here. Additionally, the following examples will set most of the render states in
RestoreDeviceObjects().
Programming Vertex Shaders 41
Figure 6: Framework public interface
Team LRN
Now back to RacorX. As shown in Fundamentals of Vertex Shaders, the following list
tracks the life cycle of a vertex shader:
n
Check for vertex shader support by checking the D3DCAPS8::VertexShaderVersion field.
n
Declare the vertex shader with the D3DVSD_* macros to map vertex buffer streams to
input registers.
n
Set the vertex shader constant registers with SetVertexShaderConstant().
n
Compile an already written vertex shader with D3DXAssembleShader*() (alternatively:
could be precompiled using a shader assembler).
n
Create a vertex shader handle with CreateVertexShader().
n
Set a vertex shader with SetVertexShader() for a specific object.
n
Free vertex shader resources handled by the Direct3D engine with DeleteVertexShader().
We will walk step by step through this list in the following pages.
Check for Vertex Shader Support
The supported vertex shader version is checked in ConfirmDevice() in RacorX.cpp:
HRESULT CMyD3DApplication::ConfirmDevice( D3DCAPS8* pCaps,
DWORD dwBehavior,
D3DFORMAT Format )
{
if( (dwBehavior & D3DCREATE_HARDWARE_VERTEXPROCESSING ) ||
(dwBehavior & D3DCREATE_MIXED_VERTEXPROCESSING ) )
{
if( pCaps->VertexShaderVersion < D3DVS_VERSION(1,1) )
return E_FAIL;
}
return S_OK;
}
If the framework has already initialized hardware or mixed vertex processing, the vertex
shader version will be checked. If the framework initialized software vertex processing, the
software implementation provided by Intel and AMD jumps in and a check of the hardware
capabilities is not needed.
The globally available pCaps capability data structure is filled with a call to
GetDeviceCaps() by the framework. pCaps->VertexShaderVersion holds the vertex shader ver-
sion in a DWORD. The macro D3DVS_VERSION helps with checking the version number.
For example, the support of at least vs.2.0 in hardware will be checked with
D3DVS_VERSION(2,0).
After checking the hardware capabilities for vertex shader support, the vertex shader has to
be declared.
Vertex Shader Declaration
Declaring a vertex shader means mapping vertex data to specific vertex shader input registers;
therefore, the vertex shader declaration must reflect the vertex buffer layout because the vertex
buffer must transport the vertex data in the correct order. The one used in this example pro-
gram is very simple. The vertex shader will get the position data via v0.
42 Part 1: Introduction to Shader Programming
Team LRN
// shader decl
DWORD dwDecl[] =
{
D3DVSD_STREAM(0),
D3DVSD_REG(0, D3DVSDT_FLOAT3 ), // D3DVSDE_POSITION,0
D3DVSD_END()
};
The corresponding layout of the vertex buffer looks like this:
// vertex type
struct VERTEX
{
FLOAT x, y, z; // The untransformed position for the vertex
};
// Declare custom FVF macro.
#define D3DFVF_VERTEX (D3DFVF_XYZ)
The position values will be stored in the vertex buffer and bound through the SetStream-
Source() function to a device data stream port that feeds the primitive processing functions
(this is the Higher-Order Surfaces (HOS) stage or directly the vertex shader, depending on the
usage of HOS; see the Direct3D pipeline in Figure 1 in the previous article).
We do not use vertex color here, so no color values are declared.
Set the Vertex Shader Constant Registers
The vertex shader constant registers have to be filled with a call to SetVertexShaderConstant().
In this example, set the material color in RestoreDeviceObjects() in c8:
// set material color
FLOAT fMaterial[4] = {0,1,0,0};
m_pd3dDevice->SetVertexShaderConstant(8, fMaterial, 1);
SetVertexShaderConstant() is declared like:
HRESULT SetVertexShaderConstant (DWORD Register,
CONST void* pConstantData,
DWORD ConstantCount);
The first parameter provides the number of the constant register that should be used; in this
case, it is 8. The second parameter stores the 128-bit data in that constant register, and the third
parameter gives you the ability to use the following registers as well. A 4x4 matrix can be
stored with one SetVertexShaderConstant() call by providing 4 in ConstantCount. This is done
for the clipping matrix in FrameMove():
// set the clip matrix
...
m_pd3dDevice->SetVertexShaderConstant(4, matTemp, 4);
This way, the c4, c5, c6, and c7 registers are used to store the matrix.
The Vertex Shader
The vertex shader that is used by RacorX is very simple:
// reg c4-7 = WorldViewProj matrix
Programming Vertex Shaders 43
Team LRN
// reg c8 = constant color
// reg v0 = input register
const char BasicVertexShader[] =
"vs.1.1 //Shader version 1.1 \n"\
"dp4 oPos.x, v0, c4 //emit projected position \n"\
"dp4 oPos.y, v0, c5 //emit projected position \n"\
"dp4 oPos.z, v0, c6 //emit projected position \n"\
"dp4 oPos.w, v0, c7 //emit projected position \n"\
"mov oD0, c8 //material color = c8 \n";
It is used inline in a constant char array in RacorX.cpp. This vertex shader incorporates the
vs.1.1 vertex shader implementation rules. It transforms from the concatenated and transposed
world, view, and projection matrix to the clip matrix or clip space with the four dp4 instruc-
tions and kicks out into a oD0 green material color with mov.
As shown above, the values of the c4-c7 constant registers are set in FrameMove(). These
values are calculated by the following code snippet:
// rotates the object about the y-axis
D3DXMatrixRotationY( &m_matWorld, m_fTime * 1.5f );
// set the clip matrix
D3DXMATRIX matTemp;
D3DXMatrixTranspose( &matTemp , &(m_matWorld * m_matView * m_matProj) );
m_pd3dDevice->SetVertexShaderConstant(4, matTemp, 4);
First the quad is rotated around the y-axis by the D3DMatrixRotationY() call, then the concate-
nated matrix is transposed and stored in the constant registers c4-c7. The source of the
D3DMatrixRotationY() function might look like:
VOID D3DMatrixRotationY(D3DMATRIX * mat, FLOAT fRads)
{
D3DXMatrixIdentity(mat);
mat._11 = cosf(fRads);
mat._13 = -sinf(fRads);
mat._31 = sinf(fRads);
mat._33 = cosf(fRads);
}
=
cosf(fRads) 0 -sinf(fRads) 0
0 0 0 0
sinf(fRads) 0 cosf(fRads) 0
0 0 0 0
fRads equals the amount you want to rotate about the y-axis. After changing the values of the
matrix this way, we transpose the matrix by using D3DXMatrixTranspose(), so its columns are
stored as rows. Why do we have to transpose the matrix?
A 4x4 matrix looks like this:
a b c d
e f g h
i j k l
m n o p
The formula for transforming a vector (v0) through the matrix is:
dest.x = (v0.x * a) + (v0.y * e) + (v0.z * i) + (v0.w * m)
dest.y = (v0.x * b) + (v0.y * f) + (v0.z * j) + (v0.w * n)
dest.z = (v0.x * c) + (v0.y * g) + (v0.z * k) + (v0.w * o)
dest.w = (v0.x * d) + (v0.y * h) + (v0.z * l) + (v0.w * p)
44 Part 1: Introduction to Shader Programming
Team LRN
So each column of the matrix should be multiplied with each component of the vector.
Our vertex shader uses four dp4 instructions:
dest.w = (src1.x * src2.x) + (src1.y * src2.y) +
(src1.z * src2.z) + (src1.w * src2.w)
dest.x = dest.y = dest.z = unused
The dp4 instruction multiplies a row of a matrix with each component of the vector. Without
transposing, we would end up with:
dest.x = (v0.x * a) + (v0.y * b) + (v0.z * c) + (v0.w * d)
dest.y = (v0.x * e) + (v0.y * f) + (v0.z * g) + (v0.w * h)
dest.z = (v0.x * i) + (v0.y * j) + (v0.z * k) + (v0.w * l)
dest.w = (v0.x * m) + (v0.y * n) + (v0.z * o) + (v0.w * p)
which is wrong. By transposing the matrix, it looks like this in constant memory:
a e i m
b f j n
c g k o
d h l p
So the 4 dp4 operations would now yield:
dest.x = (v0.x * a) + (v0.y * e) + (v0.z * i) + (v0.w * m)
dest.y = (v0.x * b) + (v0.y * f) + (v0.z * j) + (v0.w * n)
dest.z = (v0.x * c) + (v0.y * g) + (v0.z * k) + (v0.w * o)
dest.w = (v0.x * d) + (v0.y * h) + (v0.z * l) + (v0.w * p)
or
oPos.x = (v0.x * c4.x) + (v0.y * c4.y) + (v0.z * c4.z) + (v0.w * c4.w)
oPos.y = (v0.x * c5.x) + (v0.y * c5.y) + (v0.z * c5.z) + (v0.w * c5.w)
oPos.z = (v0.x * c6.x) + (v0.y * c6.y) + (v0.z * c6.z) + (v0.w * c6.w)
oPos.w = (v0.x * c7.x) + (v0.y * c7.y) + (v0.z * c7.z) + (v0.w * c7.w)
which is exactly how the vector transformation should work.
dp4 gets the matrix values via the constant register c4-c7 and the vertex position via the
input register v0. Temporary registers are not used in this example. The dot product of the dp4
instructions is written to the oPos output register, and the value of the constant register c8 is
moved into the output register oD0, which is usually used to output diffuse color values.
Compile a Vertex Shader
The vertex shader that is stored in a char array is compiled with a call to the following code
snippet in RestoreDeviceObjects():
// Assemble the shader
rc = D3DXAssembleShader( BasicVertexShader , sizeof(BasicVertexShader) -1,
0 , NULL , &pVS , &pErrors );
if ( FAILED(rc) )
{
OutputDebugString( "Failed to assemble the vertex shader, errors:\n" );
OutputDebugString( (char*)pErrors->GetBufferPointer() );
OutputDebugString( "\n" );
}
Programming Vertex Shaders 45
Team LRN
D3DXAssembleShader() creates a binary version of the shader in a buffer object via the
ID3DXBuffer interface in pVS:
HRESULT D3DXAssembleShader(
LPCVOID pSrcData,
UINT SrcDataLen,
DWORD Flags,
LPD3DXBUFFER* ppConstants,
LPD3DXBUFFER* ppCompiledShader,
LPD3DXBUFFER* ppCompilationErrors
);
The source data is provided in the first parameter, and the size of the data length in bytes is
provided in the second parameter. There are two possible flags for the third parameter:
#define D3DXASM_DEBUG 1
#define D3DXASM_SKIPVALIDATION 2
The first one inserts debug info as comments into the shader and the second one skips valida-
tion. This flag can be set for a working shader.
Via the fourth parameter, an ID3DXBuffer interface can be exported to get a vertex shader
declaration fragment of the constants. To ignore this parameter, it is set to NULL here. In case
of an error, the error explanation would be stored in a buffer object via the ID3DXBuffer inter-
face in pErrors. To see the output of OutputDebugString(), the debug process in the Visual
C/C++ IDE must be started with F5.
Create a Vertex Shader
The vertex shader is validated and a handle for it is retrieved via a call to CreateVertexShader()
in m_dwVertexShader. The following lines of code can be found in RestoreDeviceObjects():
// Create the vertex shader
rc = m_pd3dDevice->CreateVertexShader( dwDecl, (DWORD*)pVS->GetBufferPointer(),
&m_dwVertexShader, 0 );
if ( FAILED(rc) )
{
OutputDebugString( "Failed to create the vertex shader, errors:\n" );
D3DXGetErrorStringA(rc,szBuffer,sizeof(szBuffer));
OutputDebugString( szBuffer );
OutputDebugString( "\n" );
}
CreateVertexShader() gets a pointer to the buffer with the binary version of the vertex shader
via the ID3DXBuffer interface. This function gets the vertex shader declaration via dwDecl,
which maps vertex data to specific vertex shader input registers. If an error occurs, its explana-
tion is accessible via a pointer to a buffer object that is retrieved via the ID3DXBuffer interface
in pVS->GetBufferPointer(). D3DXGetErrorStringA() interprets all Direct3D and Direct3DX
HRESULTS and returns an error message in szBuffer.
It is possible to force the usage of software vertex processing with the last parameter by
using the D3DUSAGE_SOFTWAREPROCESSING flag. It must be used when the
D3DRS_SOFTWAREVERTEXPROCESSING member of the D3DRENDERSTATETYPE
enumerated type is TRUE.
46 Part 1: Introduction to Shader Programming
Team LRN
Set a Vertex Shader
The vertex shader is set via SetVertexShader() in the Render() function:
// set the vertex shader
m_pd3dDevice->SetVertexShader( m_dwVertexShader );
The only parameter that must be provided is the handle to the vertex shader. This function exe-
cutes the vertex shader as often as there are vertices.
Free Vertex Shader Resources
Vertex shader resources must be freed with a call to:
if ( m_dwVertexShader != 0xffffffff )
{
m_pd3dDevice->DeleteVertexShader( m_dwVertexShader );
m_dwVertexShader = 0xffffffff;
}
This example frees the vertex shader resources in the InvalidateDeviceObjects() framework
function; this must happen if a window size or a device changes.
Non-Shader Specific Code
The non-shader specific code of RacorX deals with setting render states and the handling of the
vertex and index buffer. A few render states have to be set in RestoreDeviceObjects():
// z-buffer enabled
m_pd3dDevice->SetRenderState( D3DRS_ZENABLE, TRUE );
// Turn off D3D lighting, since we are providing our own vertex shader lighting
m_pd3dDevice->SetRenderState( D3DRS_LIGHTING, FALSE );
// Turn off culling, so we see the front and back of the quad
m_pd3dDevice->SetRenderState( D3DRS_CULLMODE, D3DCULL_NONE );
The first instructions enable the z-buffer (a corresponding flag has to be set in the constructor
of the Direct3D framework class so that the device is created with a z-buffer).
The fixed-function lighting is not needed, so it is switched off with the second statement.
To see both sides of the quad, backface culling is switched off with the third statement.
The vertex and index buffer is created in InitDeviceObjects():
// create and fill the vertex buffer
// Initialize vertices for rendering a quad
VERTEX Vertices[] =
{
// x y z
{ -1.0f,-1.0f, 0.0f, },
{ 1.0f,-1.0f, 0.0f, },
{ 1.0f, 1.0f, 0.0f, },
{ -1.0f, 1.0f, 0.0f, },
};
m_dwSizeofVertices = sizeof (Vertices);
// Create the vertex buffers with four vertices
if( FAILED( m_pd3dDevice->CreateVertexBuffer( 4 * sizeof(VERTEX),
D3DUSAGE_WRITEONLY , sizeof(VERTEX), D3DPOOL_MANAGED, &m_pVB ) ) )
return E_FAIL;
Programming Vertex Shaders 47
Team LRN
// lock and unlock the vertex buffer to fill it with memcpy
VOID* pVertices;
if( FAILED( m_pVB->Lock( 0, m_dwSizeofVertices, (BYTE**)&pVertices, 0 ) ) )
return E_FAIL;
memcpy( pVertices, Vertices, m_dwSizeofVertices);
m_pVB->Unlock();
// create and fill the index buffer
// indices
WORD wIndices[]={0, 1, 2, 0, 2, 3};
m_wSizeofIndices = sizeof (wIndices);
// create index buffer
if(FAILED (m_pd3dDevice->CreateIndexBuffer(m_wSizeofIndices, 0,
D3DFMT_INDEX16, D3DPOOL_MANAGED, &m_pIB)))
return E_FAIL;
// fill index buffer
VOID *pIndices;
if (FAILED(m_pIB->Lock(0, m_wSizeofIndices, (BYTE **)&pIndices, 0)))
return E_FAIL;
memcpy(pIndices, wIndices, m_wSizeofIndices);
m_pIB->Unlock();
The four vertices of the quad are stored in a VERTEX structure, which holds for each vertex
three FLOAT values for the position.
By using the flag D3DFMT_INDEX16 in CreateIndexBuffer(), 16-bit variables are used to
store the indices into the wIndices structure. So the maximum number of available indices is
64 K. Both buffers use a managed memory pool with D3DPOOL_MANAGED, so they will be
cached in the system memory.
48 Part 1: Introduction to Shader Programming
Note: D3DPOOL_MANAGED resources are read from the system memory, which is
quite fast, and are written to the system memory and then uploaded to wherever the
non-system copy has to go (AGP or VIDEO memory). This upload happens when the
resource is unlocked. There are always two copies of a resource, one in the system
and one in the AGP or VIDEO memory. This is a less efficient but bullet-proof way. It
works for any class of driver and must be used with unified memory architecture
boards. Handling resources with D3DPOOL_DEFAULT is more efficient. In this case,
the driver will choose the best place for the resource.
Note: Why do we use a vertex buffer at all? The vertex buffer can be stored in the
memory of your graphic card or AGP, where it can be accessed very quickly by 3D
hardware, so a copy between system memory and the graphic card/AGP memory
can be avoided. This is important for hardware that accelerates transformation and
lighting. Without vertex buffers, a lot of bus traffic would happen by transforming
and lighting the vertices.
Team LRN
Summary
RacorX shows a simple vertex shader together with its infrastructure. The shader is inlined in
RacorX.cpp and compiled with D3DXAssembleShader(). It uses four dp4 instructions for the
transformation of the quad and only one material color.
The upcoming examples are built on this example, and only the functional additions will
be shown on the following pages.
RacorX2
The main difference between RacorX and RacorX2 is the compilation of the vertex shader
with NVASM. Whereas the first example compiles the vertex shader with D3DXAssemble-
Shader() while the application starts up, RacorX2 uses a precompiled vertex shader.
To add the NVIDIA vertex and pixel shader assembler, you have to do the following steps:
n
Create a directory (for example, <C:\NVASM>:) and unzip nvasm.exe and the documenta-
tion into it.
n
Show your Visual C++ IDE path to this exe with Tools->Options->Directories and choose
Executable files from the Show directories for drop-down list.
n
Add the path to NVASM <C:\NVASM>.
Now the dialog box should look like this:
Programming Vertex Shaders 49
Figure 7: Integrating NVASM into Visual
C/C++
Note: Why do we use an index buffer? You will get maximum performance when
you reduce the duplication in vertices transformed and sent across the bus to the ren-
dering device. A non-indexed triangle list, for example, achieves no vertex sharing,
so its the least optimal method because DrawPrimitive*() is called several times.
Using indexed lists or strips reduces the call overhead of DrawPrimitive*() methods
(reducing DrawPrimitve*() methods is also called batching) and because of the reduc-
tion of vertices to send through the bus, it saves memory bandwidth. Indexed strips
are more hardware cache-friendly on newer hardware than indexed lists.
The performance of index processing operations depends heavily on where the
index buffer exists in memory. At the time of pubication, the only graphic cards that
support index buffers in hardware are the RADEON 8x00 series.
Team LRN
n
Additionally, you have to tell every vertex shader file in the IDE that it has to be compiled
with NVASM. The easiest way to do that is by looking into the example RacorX2. Just fire
up your Visual C++ IDE by clicking on RacorX.dsp in the RacorX2 directory. Click on the
FileView tab of the Workspace dialog and then on Shaders to view the available shader
files. A right-click on the file basic.vsh should show you a pop-up. Click on Settings. The
settings of your project might look like this:
n
The entry in the Commands field is nvasm.exe $(InputPath) shaders\$(InputName).vso.
In the Outputs field entry, I use the input name as the name for the output file with an
*.vso extension. The output directory should be the shaders directory. Direct3D ShaderX
author Kenneth Hurley is the author of NVASM. Read more in [Hurley].
The output of NVASM in the build window of your Visual C/C++ IDE should look like this:
The vertex shader is provided in its own ASCII file called basic.vsh. After compilation, a
binary object file with the name basic.vso is created in the shaders directory:
50 Part 1: Introduction to Shader Programming
Figure 8: Custom Build options
for Vertex Shader files
Figure 9: NVASM output
Figure 10: Directory content of RacorX2
Team LRN
Create a Vertex Shader
Because of the already compiled vertex shader, the creation of the vertex shader has to be done
in a different way than in the previous example. This happens in the InitDeviceObjects()
function:
//shader decl
DWORD dwDecl[] =
{
D3DVSD_STREAM(0),
D3DVSD_REG(0, D3DVSDT_FLOAT3 ), // input register 0
D3DVSD_END()
};
// loads a *.vso binary file, already compiled with NVASM and
// creates a vertex shader
if (FAILED(CreateVSFromCompiledFile (m_pd3dDevice, dwDecl,"shaders/basic.vso",
&m_dwVertexShader)))
return E_FAIL;
CreateVSFromCompiledFile() opens and reads in the binary vertex shader file and creates a
vertex shader. The source of this function can be found at the end of the RacorX.cpp file in the
RacorX2 directory:
//----------------------------------------------------------------------------
// Name: CreateVSFromBinFile
// Desc: loads a binary *.vso file that was compiled by NVASM
// and creates a vertex shader
//----------------------------------------------------------------------------
HRESULT CMyD3DApplication::CreateVSFromCompiledFile (IDirect3DDevice8*
m_pd3dDevice,
DWORD* dwDeclaration,
TCHAR* strVSPath,
DWORD* m_dwVS)
{
char szBuffer[128]; // debug output
DWORD* dwpVS; // pointer to address space of the calling process
HANDLE hFile, hMap; // handle file and handle mapped file
TCHAR tempVSPath[512]; // temporary file path
HRESULT hr; // error
if( FAILED( hr = DXUtil_FindMediaFile( tempVSPath, strVSPath ) ) )
return D3DAPPERR_MEDIANOTFOUND;
hFile = CreateFile(tempVSPath, GENERIC_READ,0,0,OPEN_EXISTING,
FILE_ATTRIBUTE_NORMAL,0);
if(hFile != INVALID_HANDLE_VALUE)
{
if(GetFileSize(hFile,0) > 0)
hMap = CreateFileMapping(hFile,0,PAGE_READONLY,0,0,0);
else
{
CloseHandle(hFile);
return E_FAIL;
}
}
else
return E_FAIL;
Programming Vertex Shaders 51
Team LRN
// maps a view of a file into the address space of the calling process
dwpVS = (DWORD *)MapViewOfFile(hMap, FILE_MAP_READ, 0, 0, 0);
// Create the vertex shader
hr = m_pd3dDevice->CreateVertexShader( dwDeclaration, dwpVS, m_dwVS, 0 );
if ( FAILED(hr) )
{
OutputDebugString( "Failed to create Vertex Shader, errors:\n" );
D3DXGetErrorStringA(hr,szBuffer,sizeof(szBuffer));
OutputDebugString( szBuffer );
OutputDebugString( "\n" );
return hr;
}
UnmapViewOfFile(dwpVS);
CloseHandle(hMap);
CloseHandle(hFile);
return S_OK;
}
DXUtil_FindMediaFile(), a helper function located in the framework file dxutil.cpp, returns
the path to the already compiled vertex shader file. CreateFile() opens and reads the existing
file:
HANDLE CreateFile(
LPCTSTR lpFileName, // file name
DWORD dwDesiredAccess, // access mode
DWORD dwShareMode, // share mode
LPSECURITY_ATTRIBUTES lpSecurityAttributes, // SD
DWORD dwCreationDisposition, // how to create
DWORD dwFlagsAndAttributes, // file attributes
HANDLE hTemplateFile // handle to template file
);
Its first parameter is the path to the file. The flag GENERIC_READ specifies read access to
the file in the second parameter. The following two parameters are not used because file shar-
ing should not happen and the file should not be inherited by a child process. The fifth
parameter is set to OPEN_EXISTING. This way, the function fails if the file does not exist.
Setting the sixth parameter to FILE_ATTRIBUTE_NORMAL indicates that the file has no
other attributes. A template file is not used here, so the last parameter is set to 0. Please consult
the Platform SDK help file for more information.
CreateFileMapping() creates or opens a named or unnamed file-mapping object for the
specified file:
HANDLE CreateFileMapping(
HANDLE hFile, // handle to file
LPSECURITY_ATTRIBUTES lpAttributes, // security
DWORD flProtect, // protection
DWORD dwMaximumSizeHigh, // high-order DWORD of size
DWORD dwMaximumSizeLow, // low-order DWORD of size
LPCTSTR lpName // object name
);
The first parameter is a handle to the file from which to create a mapping object. The file must
be opened with an access mode compatible with the protection flags specified by the flProtect
parameter. We have opened the file in CreateFile() with GENERIC_READ, therefore we use
52 Part 1: Introduction to Shader Programming
Team LRN
PAGE_READONLY here. Other features of CreateFileMapping() are not needed, so we set the
rest of the parameters to 0. MapViewOfFile()maps a view of a file into the address space of the
calling process:
LPVOID MapViewOfFile(
HANDLE hFileMappingObject, // handle to file-mapping object
DWORD dwDesiredAccess, // access mode
DWORD dwFileOffsetHigh, // high-order DWORD of offset
DWORD dwFileOffsetLow, // low-order DWORD of offset
SIZE_T dwNumberOfBytesToMap // number of bytes to map
);
This function only gets the handle to the file-mapping object from CreateFileMapping() and
the access mode FILE_MAP_READ. The access mode parameter specifies the type of access
to the file view and, therefore, the protection of the pages mapped by the file. Other features
are not needed; therefore, the rest of the parameters are set to 0.
CreateVertexShader() is used to create and validate a vertex shader. It takes the vertex
shader declaration (which maps vertex buffer streams to different vertex input registers) in its
first parameter as a pointer and returns the shader handle in the third parameter. The second
parameter gets the vertex shader instructions of the binary code precompiled by a vertex shader
assembler. With the fourth parameter, you can force software vertex processing with
D3DUSAGE_SOFTWAREPROCESSING.
As in the previous example, OutputDebugString() shows the complete error message in the
output debug window of the Visual C/C++ IDE, and D3DXGetErrorStringA() interprets all
Direct3D and Direct3DX HRESULTS and returns an error message in szBuffer.
Summary
This example showed the integration of NVASM to precompile a vertex shader and how to
open and read a binary vertex shader file.
RacorX3
The main improvement of RacorX3 over
RacorX2 is the addition of a per-vertex
diffuse reflection model in the vertex
shader. This is one of the simplest light-
ing calculations, which outputs the color
based on the dot product of the vertex
normal with the light vector.
RacorX3 uses a light positioned at
(0.0, 0.0, 1.0) and a green color.
As usual, we are tracking the life
cycle of the vertex shader.
Programming Vertex Shaders 53
Figure 11: RacorX3
Team LRN
Vertex Shader Declaration
The vertex shader declaration has to map vertex data to specific vertex shader input registers.
In addition to the previous examples, we need to map a normal vector to the input register v3:
// vertex shader declaration
DWORD dwDecl[] =
{
D3DVSD_STREAM(0),
D3DVSD_REG(0, D3DVSDT_FLOAT3 ), // input register #1
D3DVSD_REG(3, D3DVSDT_FLOAT3 ), // normal in input register #4
D3DVSD_END()
};
The corresponding layout of the vertex buffer looks like this:
struct VERTICES
{
FLOAT x, y, z; // The untransformed position for the vertex
FLOAT nx, ny, nz; // the normal
};
// Declare custom FVF macro.
#define D3DFVF_VERTEX (D3DFVF_XYZ|D3DFVF_NORMAL)
Each vertex consists of three position floating-point values and three normal floating-point val-
ues in the vertex buffer. The vertex shader gets the position and normal values from the vertex
buffer via v0 and v3.
Set the Vertex Shader Constant Registers
The vertex shader constants are set in FrameMove() and RestoreDeviceObjects(). This exam-
ple uses a more elegant way to handle the constant registers. The file const.h that is included in
RacorX.cpp and diffuse.vsh gives the constant registers names that are easier to remember:
#define CLIP_MATRIX 0
#define CLIP_MATRIX_1 1
#define CLIP_MATRIX_2 2
#define CLIP_MATRIX_3 3
#define INVERSE_WORLD_MATRIX 4
#define INVERSE_WORLD_MATRIX_1 5
#define INVERSE_WORLD_MATRIX_2 6
#define LIGHT_POSITION 11
#define DIFFUSE_COLOR 14
#define LIGHT_COLOR 15
In FrameMove(), a clipping matrix and an inverse world matrix are set into the constant
registers:
HRESULT CMyD3DApplication::FrameMove()
{
// rotates the object about the y-axis
D3DXMatrixRotationY( &m_matWorld, m_fTime * 1.5f );
// set the clip matrix
m_pd3dDevice->SetVertexShaderConstant(CLIP_MATRIX, &(m_matWorld *
m_matView * m_matProj), 4);
D3DXMATRIX matWorldInverse;
54 Part 1: Introduction to Shader Programming
Team LRN
D3DXMatrixInverse(&matWorldInverse, NULL, &m_matWorld);
m_pd3dDevice->SetVertexShaderConstant(INVERSE_WORLD_MATRIX,
&matWorldInverse,3);
return S_OK;
}
Contrary to the previous examples, the concatenated world, view, and projection matrix, which
is used to rotate the quad, is not transposed here. This is because the matrix will be transposed
in the vertex shader, as shown below.
To transform the normal, an inverse 4x3 matrix is sent to the vertex shader via c4-c6.
The Vertex Shader
The vertex shader is a bit more complex than the one used in the previous examples:
; per-vertex diffuse lighting
#include "const.h"
vs.1.1
; transpose and transform to clip space
mul r0, v0.x, c[CLIP_MATRIX]
mad r0, v0.y, c[CLIP_MATRIX_1], r0
mad r0, v0.z, c[CLIP_MATRIX_2], r0
add oPos, c[CLIP_MATRIX_3], r0
; transform normal
dp3 r1.x, v3, c[INVERSE_WORLD_MATRIX]
dp3 r1.y, v3, c[INVERSE_WORLD_MATRIX_1]
dp3 r1.z, v3, c[INVERSE_WORLD_MATRIX_2]
; renormalize it
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1, r1, r1.w
; N dot L
; we need L vector towards the light, thus negate sign
dp3 r0, r1, -c[LIGHT_POSITION]
mul r0, r0, c[LIGHT_COLOR] ; modulate against light color
mul oD0, r0, c[DIFFUSE_COLOR] ; modulate against material
The mul, mad, and add instructions transpose and transform the matrix provided in c0-c3 to
clip space. As such, they are nearly functionally equivalent to the transposition of the matrix
and the four dp4 instructions shown in the previous examples. There are two caveats to bear in
mind: The complex matrix instructions like m4x4 might be faster in software emulation mode
and v0.w is not used here. oPos.w is automatically filled with 1.
These instructions should save the CPU cycles used for transposing.
The normals are transformed in the following three dp3 instructions and then renormalized
with the dp3, rsq, and mul instructions.
You can think of a normal transform in the following way: Normal vectors (unlike position
vectors) are simply directions in space, and as such they should not get squished in magnitude,
and translation doesnt change their direction. They should simply be rotated in some fashion
to reflect the change in orientation of the surface. This change in orientation is a result of rotat-
ing and squishing the object but not moving it. The information for rotating a normal can be
extracted from the 4x4 transformation matrix by doing transpose and inversion. A more
math-related explanation is given in [Haines/Mller][Turkowski].
Programming Vertex Shaders 55
Team LRN
So the bullet-proof way to use normals is to transform the transpose of the inverse of the
matrix that is used to transform the object. If the matrix used to transform the object is called
M, then we must use the matrix N, below, to transform the normals of this object.
N = transpose( inverse(M) )
Thats exactly what the source is doing. The inverse world matrix is delivered to the vertex
shader via c4-c6. The dp3 instruction handles the matrix in a similar way as dp4.
By multiplying a matrix with a vector, each column of the matrix should be multiplied
with each component of the vector. dp3 and dp4 are only capable of multiplying each row of
the matrix with each component of the vector. In the case of the position data, the matrix is
transposed to get the right results.
In the case of the normals, no transposition is done, so dp3 calculates the dot product by
multiplying the rows of the matrix with the components of the vector. This is like using a trans-
posed matrix.
The normal is renormalized with the dp3, rsq, and mul instructions. Renormalizing a vec-
tor means to align its length to 1. Thats because we need a unit vector to calculate our diffuse
lighting effect.
To calculate a unit vector, divide the vector by its magnitude or length. The magnitude of
vectors is calculated by using the Pythagorean theorem:
x
2
+ y
2
+ z
2
= m
2
The length of the vector is retrieved by:
||A|| = sqrt(x
2
+ y
2
+ z
2
)
The magnitude of a vector has a special symbol in mathematics. It is a capital letter with two
vertical bars on each side: ||A||. So dividing the vector by its magnitude is:
UnitVector = Vector / sqrt(x
2
+ y
2
+ z
2
)
The lines of code in the vertex shader that handle the calculation of the unit vector look like
this:
; renormalize it
dp3 r1.w, r1, r1 ; (src1.x * src2.x) + (src1.y * src2.y) + (src1.z * src2.z)
rsq r1.w, r1.w ; if (v != 0 && v != 1.0) v = (float)(1.0f / sqrt(v))
mul r1, r1, r1.w ; r1 * r1.w
56 Part 1: Introduction to Shader Programming
Note: The normal can be transformed with the transformation matrix (usually the
world matrix) that is used to transform the object in the following cases:
n
Matrix formed from rotations (orthogonal matrix) because the inverse of an
orthogonal matrix is its transpose
n
Matrix formed from rotations and translation (rigid-body transforms) because
translations do not affect vector direction
n
Matrix formed from rotations and translation and uniform scalings because
such scalings affect only the length of the transformed normal, not its direc-
tion. A uniform scaling is simply a matrix, which uniformly increases or
decreases the objects size, vs. a non-uniform scaling, which can stretch or
squeeze an object. If uniform scalings are used, then the normals do have to
be renormalized.
Therefore, using the world matrix would be sufficient in this example.
Team LRN
dp3 squares the x, y, and z components of the temporary register r1, adds them, and returns the
result in r1.w. rsq divides 1 by the result in r1.w and stores the result in r1.w. mul multiplies all
components of r1 with r1.w. Afterward, the result in r1.w is no longer used in the vertex
shader.
The underlying calculation of these three instructions can be represented by the following
formula, which is mostly identical to the formula postulated above:
UnitVector = Vector * 1/sqrt(x
2
+ y
2
+ z
2
)
Lighting is calculated with the following three instructions:
dp3 r0, r1, -c[LIGHT_POSITION]
mul r0, r0, c[LIGHT_COLOR] ; modulate against light color
mul oD0, r0, c[DIFFUSE_COLOR] ; modulate against diffuse color
Nowadays, the lighting models used in current games are not based on much physical theory.
Game programmers use approximations that try to simulate the way photons are reflected from
objects in a rough but efficient manner.
One usually differentiates between different kinds of light sources and different reflection
models. The common lighting sources are called directional, point light, and spotlight. The
most common reflection models are ambient, diffuse, and specular lighting.
This example uses a directional light source with a diffuse reflection model.
Directional Light
RacorX3 uses a light source in an infinite distance. This simulates the long distance that light
beams have to travel from the sun. We treat these light beams as being parallel. This kind of
light source is called directional light source.
Diffuse Reflection
Whereas ambient light is considered to be uniform from any direction, diffuse light simulates
the emission of an object by a particular light source. Therefore, you are able to see that light
falls onto the surface of an object from a particular direction by using the diffuse lighting
model.
It is based on the assumption that light is reflected equally well in all directions, so the
appearance of the reflection does not depend on the position of the observer. The intensity of
the light reflected in any direction depends only on how much light falls onto the surface.
If the surface of the object is facing the light source, which means it is perpendicular to the
direction of the light, the density of the incidental light is the highest. If the surface is facing
the light source at an angle smaller than 90 degrees, the density is proportionally smaller.
The diffuse reflection model is based on a law of physics called Lamberts Law, which
states that for ideally diffuse (totally matte) surfaces, the reflected light is determined by the
cosine between the surface normal N and the light vector L. (See Figure 12.)
Programming Vertex Shaders 57
Team LRN
The left figure shows a geometric interpretation of Lamberts Law (see also [RTR]). The mid-
dle figure shows the light rays hitting the surface perpendicularly at a distance d apart. The
intensity of the light is related to this distance. It decreases as d becomes greater. This is shown
in the right figure. The light rays make an angle with the normal of the plane. This illustrates
that the same amount of light that passes through one side of a right-angle triangle is reflected
from the region of the surface corresponding to the triangles hypotenuse. Due to the relation-
ships in a right-angle triangle, the length of the hypotenuse is d/cos of the length of the
considered side. Thus, you can deduce that if the intensity of the incident light is Idirected, the
amount of light reflected from a unit surface is Idirected cos . Adjusting this with a coefficient
that describes reflection properties of the matter leads to the following equation (see also
[Savchenko]):
Ireflected = Cdiffuse * Idirected cos
This equation demonstrates that the reflection is at its peak for surfaces that are perpendicular
to the direction of light and diminishes for smaller angles because the cosinus value is very
large. The light is obscured by the surface if the angle is more than 180 or less than 0 degrees
because the cosinus value is small. You will obtain negative intensity of the reflected light,
which will be clamped by the output registers.
In an implementation of this model, you have to find a way to compute cos . By defini-
tion, the dot or scalar product of the light and normal vector can be expressed as
N dot L = ||N|| ||L||cos
where ||N|| and ||L|| are the lengths of the vectors. If both vectors are unit length, you can com-
pute cos as the scalar or dot product of the light and normal vector. Thus, the expression is
Ireflected = Cdiffuse * Idirected(N dot L)
So (N dot L) is the same as the cosine of the angle between N and L; therefore, as the angle
decreases, the resulting diffuse value is higher.
This is exactly what the dp3 instruction and the first mul instruction are doing. Here is the
source with the relevant part of constant.h:
#define LIGHT_POSITION 11
#define MATERIAL_COLOR 14
-----
dp3 r0, r1, -c[LIGHT_POSITION]
mul r0, r0, c[LIGHT_COLOR] ; modulate against light color
mul oD0, r0, c[DIFFUSE_COLOR] ; modulate against material
So the vertex shader registers are involved in the following way:
58 Part 1: Introduction to Shader Programming
Figure 12: Diffuse lighting
Team LRN
r0 = (r1 dot -c11) * c14
This example modulates additionally against the blue light color in c15:
r0 = (c15 * (r1 dot -c11)) * c14
Summary
RacorX3 shows the usage of an include file to give constants a name that can be remembered
more easily. It shows how to normalize vectors, and although it just touches the surface in
solving the problem of transforming normals, it shows a bullet-proof method to do it.
The example introduces an optimization technique that eliminates the need to transpose
the clip space matrix with the help of the CPU and shows the usage of a simple diffuse reflec-
tion model that lights the quad on a per-vertex basis.
RacorX4
RacorX4 has a few additional fea-
tures compared to RacorX3. First of
all, this example will not use a
plain quad anymore; it uses a
Bzier patch class instead that
shows a round surface with a tex-
ture attached to it. To simulate light
reflections, a combined diffuse and
specular reflection model is used.
RacorX4 uses the trackball
class provided with the common
files framework to rotate and move
the object. You can choose different
specular colors with the C key and
zoom in and out with the mouse
wheel.
Vertex Shader Declaration
Additionally, the texture coordinates will be mapped to input register v7.
// shader decl
DWORD dwDecl[] =
{
D3DVSD_STREAM(0),
D3DVSD_REG(0, D3DVSDT_FLOAT3 ), // input register v0
D3DVSD_REG(3, D3DVSDT_FLOAT3 ), // normal in input register v3
D3DVSD_REG(7, D3DVSDT_FLOAT2), // tex coordinates
D3DVSD_END()
};
Programming Vertex Shaders 59
Figure 13: RacorX4
Team LRN
The corresponding layout of the vertex buffer in BPatch.h looks like this:
struct VERTICES {
D3DXVECTOR3 vPosition;
D3DXVECTOR3 vNormal;
D3DXVECTOR2 uv;
};
// Declare custom FVF macro.
#define D3DFVF_VERTEX (D3DFVF_XYZ|D3DFVF_NORMAL|D3DFVF_TEX1)
The third flag used in the custom FVF macro indicates the usage of one texture coordinate pair.
This macro is provided to CreateVertexBuffer().
The vertex shader gets the position values from the vertex buffer via v0, the normal values
via v3, and the two texture coordinates via v7.
Set the Vertex Shader Constants
The vertex shader constants are set in FrameMove() and RestoreDeviceObjects(). The file
const.h holds the following defines:
#define CLIP_MATRIX 0
#define CLIP_MATRIX_1 1
#define CLIP_MATRIX_2 2
#define CLIP_MATRIX_3 3
#define INVERSE_WORLD_MATRIX 8
#define INVERSE_WORLD_MATRIX_1 9
#define INVERSE_WORLD_MATRIX_2 10
#define LIGHT_VECTOR 11
#define EYE_VECTOR 12
#define SPEC_POWER 13
#define SPEC_COLOR 14
In FrameMove() a clipping matrix, the inverse world matrix, an eye vector, and a specular
color are set:
// set the clip matrix
m_pd3dDevice->SetVertexShaderConstant(CLIP_MATRIX,&(m_matWorld * m_matView *
m_matProj),4);
// set the world inverse matrix
D3DXMATRIX matWorldInverse;
D3DXMatrixInverse(&matWorldInverse, NULL, &m_matWorld);
m_pd3dDevice->SetVertexShaderConstant(INVERSE_WORLD_MATRIX, matWorldInverse,3);
// stuff for specular lighting
// set eye vector E
m_pd3dDevice->SetVertexShaderConstant(EYE_VECTOR, vEyePt,1);
// specular color
if(m_bKey['C'])
{
m_bKey['C']=0;
++m_dwCurrentColor;
if(m_dwCurrentColor >= 3)
m_dwCurrentColor=0;
}
m_pd3dDevice->SetVertexShaderConstant(SPEC_COLOR, m_vLightColor
[m_dwCurrentColor], 1);
60 Part 1: Introduction to Shader Programming
Team LRN
Like in the previous example, the concatenated world, view, and projection matrix are set into
c0-c3 to get transposed in the vertex shader, and the inverse 4x3 world matrix is sent to the
vertex shader to transform the normal.
The eye vector (EYE_VECTOR) that is used to build up the view matrix is stored in con-
stant register c12. As shown below, this vector is helpful to build up the specular reflection
model used in the upcoming examples.
The user can pick one of the following specular colors with the C key:
m_vLightColor[0]=D3DXVECTOR4(0.3f,0.1f,0.1f,1.0f);
m_vLightColor[1]=D3DXVECTOR4(0.1f,0.5f,0.1f,1.0f);
m_vLightColor[2]=D3DXVECTOR4(0.0f,0.1f,0.4f,1.0f);
In RestoreDeviceObjects(), the specular power, the light vector, and the diffuse color are set:
// specular power
m_pd3dDevice->SetVertexShaderConstant(SPEC_POWER, D3DXVECTOR4(0,10,25,50),1);
// light direction
D3DXVECTOR3 vLight(0,0,1);
m_pd3dDevice->SetVertexShaderConstant(LIGHT_VECTOR, vLight,1);
D3DXCOLOR matDiffuse(0.9f, 0.9f, 0.9f, 1.0f);
m_pd3dDevice->SetVertexShaderConstant(DIFFUSE_COLOR, &matDiffuse, 1);
As in the previous examples, the light is positioned at (0.0, 0.0, 1.0). There are four specular
power values, from which one will be used in the vertex shader.
The Vertex Shader
The vertex shader handles a combined diffuse and specular reflection model:
; diffuse and specular vertex lighting
#include "const.h"
vs.1.1
; transpose and transform to clip space
mul r0, v0.x, c[CLIP_MATRIX]
mad r0, v0.y, c[CLIP_MATRIX_1], r0
mad r0, v0.z, c[CLIP_MATRIX_2], r0
add oPos, c[CLIP_MATRIX_3], r0
; output texture coords
mov oT0, v7
; transform normal
dp3 r1.x, v3, c[INVERSE_WORLD_MATRIX]
dp3 r1.y, v3, c[INVERSE_WORLD_MATRIX_1]
dp3 r1.z, v3, c[INVERSE_WORLD_MATRIX_2]
; renormalize it
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1, r1, r1.w
; light vector L
; we need L towards the light, thus negate sign
mov r5, -c[LIGHT_VECTOR]
; N dot L
dp3 r0.x, r1, r5
Programming Vertex Shaders 61
Note: To optimize the usage of SetVertexShaderConstant(), the specular
color could be set into the fourth component of the light vector, which only
used three of its components.
Team LRN
; compute normalized half vector H = L + V
add r2, c[EYE_VECTOR], r5 ; L + V
; renormalize H
dp3 r2.w, r2, r2
rsq r2.w, r2.w
mul r2, r2, r2.w
; N dot H
dp3 r0.y, r1, r2
; compute specular and clamp values (lit)
; r0.x - N dot L
; r0.y - N dot H
; r0.w - specular power n
mov r0.w, c[SPEC_POWER].y ; n must be between 128.0 and 128.0
lit r4, r0
; diffuse color * diffuse intensity(r4.y)
mul oD0, c[DIFFUSE_COLOR], r4.y
; specular color * specular intensity(r4.z)
mul oD1, c[SPEC_COLOR], r4.z
Compared to the vertex shader in the previous example program, this shader maps the texture
coordinates to texture stage 0 with the following instruction:
; output texture coords
mov oT0, v7
The corresponding texture stage states set for the shading of the pixels in the multitexturing
unit will be shown later in the Non-shader Specific Code section. The instructions that trans-
form the normal and calculate the diffuse reflection were discussed in the previous example.
The real new functionality in this shader is the calculation of the specular reflection, which
happens in the code lines starting with the add instruction.
Specular Reflection
Compared to the diffuse reflection model, the appearance of the reflection in the specular
reflection model depends on the position of the viewer. When the direction of the viewing
coincides, or nearly coincides, with the direction of specular reflection, a bright highlight is
observed. This simulates the reflection of a light source by a smooth, shiny, and polished
surface.
To describe reflection from shiny sur-
faces, an approximation is commonly used,
which is called the Phong illumination
model (not to be confused with Phong
shading), named after its creator Phong Bui
Tong [Foley]. According to this model, a
specular highlight is seen when the viewer
is close to the direction of reflection. The
intensity of light falls off sharply when the
viewer moves away from the direction of
the specular reflection.
62 Part 1: Introduction to Shader Programming
Figure 14: Vectors for specular reflection
Team LRN
A model describing this effect has to be aware of at least the location of the light source L, the
location of the viewer V, and the orientation of the surface N. Additionally, a vector R that
describes the direction of the reflection might be useful. The halfway vector H that halves the
angle between L and V is discussed below.
The original Phong formula approximates the falloff of the intensity. It looks like this:
kspecular cos
n
()
where k
specular
is a scalar coefficient showing the percentage of the incident light reflected.
describes the angle between R and V. The exponent n characterizes the shiny properties of the
surface and ranges from 1 to infinity. Objects that are matte require a small exponent, since
they produce a large, dim, specular highlight with a gentle falloff. Shiny surfaces should have a
sharp highlight that is modeled by a very large exponent, making the intensity falloff very
steep.
Together with the diffuse reflection model shown above, the Phong illumination model can
be expressed in the following way:
Ireflected = Idirected((N dot L) + kspecular cos
n
())
cos
n
() can be replaced by using the mechanism of the dot or scalar product of the unit vectors
R and V:
Ireflected = Idirected((N dot L) + kspecular (R dot V)
n
)
This is the generally accepted phong reflection equation. As the angle between V and R
decreases, the specularity rises.
Because it is expensive to calculate the reflection vector R (mirror of light incidence
around the surface normal), James F. Blinn [Blinn] introduced a way to do this using an imagi-
nary vector H, which is defined as halfway between L and V. H is therefore:
H = (L + V) / 2
When H coincides with N, the direction of the reflection coincides with the viewing direction
V and a specular hightlight is observed. So the original formula
(R dot V)
n
is expanded to
(N dot ((L + V) / 2))
n
or
(N dot H)
n
The complete Blinn-Phong model formula looks like:
Ireflected = Idirected((N dot L) + kspecular (N dot H)
n
)
Now back to the relevant code that calculates the specular reflection:
; compute normalized half vector H = L + V
add r2, c[EYE_VECTOR], r5 ; L + V
; renormalize H
dp3 r2.w, r2, r2
Programming Vertex Shaders 63
Team LRN
rsq r2.w, r2.w
mul r2, r2, r2.w
; N dot H
dp3 r0.y, r1, r2
; compute specular and clamp values (lit)
; r0.x - N dot L
; r0.y - N dot H
; r0.w - specular power n
mov r0.w, c[SPEC_POWER].y ; n must be between 128.0 and 128.0
lit r4, r0
The add instruction adds the L and V vectors, but compared to the Blinn-Phong formula shown
above, the divide by 2 is missing. This is compensated by altering the specular power value,
which leads to good visual results. So this code snippet handles the following equation, which
is very similar to the equation of the Blinn-Phong reflection formula:
(N dot (L+V))
n
Although L + V is not equivalent to the half vector H, the term half vector H will be used
throughout the upcoming examples.
In the vertex shader source, H is normalized in the following dp3, rsq, and mul
instructions.
The reduction of the original Blinn-Phong formula reduces the number of vertex shader
instructions and thus leads to a higher performance. This is a typical way to optimize vertex
shader code on a functional level [Taylor]. Although L + V is not equivalent to H, the term
vector H is used throughout the example source.
One of the main instructions in this piece of code is the lit instruction. It handles a lot of
work in only one instruction. The following code fragment shows the operations performed by
lit:
dest.x = 1;
dest.y = 0;
dest.z = 0;
dest.w = 1;
float power = src.w;
const float MAXPOWER = 127.9961f;
if (power < -MAXPOWER)
power = -MAXPOWER; // Fits into 8.8 fixed point format
else if (power > MAXPOWER)
power = -MAXPOWER; // Fits into 8.8 fixed point format
if (src.x > 0)
{
dest.y = src.x;
if (src.y > 0)
{
// Allowed approximation is EXP(power * LOG(src.y))
dest.z = (float)(pow(src.y, power));
}
}
lit returns the diffuse intensity in the y component of the destination register and the specular
intensity in the z component of the destination register. If there is no diffuse light, there will be
no specular light. So N dot L is only used to switch off the specular light in case there is no dif-
fuse lighting.
64 Part 1: Introduction to Shader Programming
Team LRN
The last two instructions multiply the intensity with the diffuse and specular color components
and move them into their respective output registers:
; diffuse color * diffuse intensity(r4.y)
mul oD0, c[DIFFUSE_COLOR], r4.y
; specular color * specular intensity(r4.z)
mul oD1, c[SPEC_COLOR], r4.z
Non-Shader Specific Code
This example uses the multitexturing unit to shade pixels. The proper texture stage states are
set in RestoreDeviceObjects():
// texture settings
m_pd3dDevice->SetTextureStageState(0, D3DTSS_COLOROP, D3DTOP_MODULATE);
m_pd3dDevice->SetTextureStageState(0, D3DTSS_COLORARG1, D3DTA_TEXTURE);
m_pd3dDevice->SetTextureStageState(0, D3DTSS_COLORARG2, D3DTA_DIFFUSE);
m_pd3dDevice->SetTextureStageState( 1, D3DTSS_COLOROP, D3DTOP_DISABLE );
m_pd3dDevice->SetTextureStageState( 1, D3DTSS_ALPHAOP, D3DTOP_DISABLE );
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_MINFILTER, D3DTEXF_LINEAR );
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_MAGFILTER, D3DTEXF_LINEAR );
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_MIPFILTER, D3DTEXF_LINEAR );
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_ADDRESSU, D3DTADDRESS_CLAMP );
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_ADDRESSV, D3DTADDRESS_CLAMP );
These texture stage states modulate the diffuse color with the texture color, disable texture
stage 1, and switch on a bilinear filtering with the D3DTEXF_LINEAR flags for the
D3DTSS_MINFILTER/D3DTSS_MAGFILTER texture stage states. The
D3DTSS_MIPFILTER texture stage state indicates the texture filter to use between mipmap
levels. By providing D3DTEXF_LINEAR, trilinear texture filtering, or mip map filtering, is
switched on. The clamp texture-addressing mode applies the texture only once on the polygon
and then smears the color of the edge pixels of the texture. Furthermore, all negative values
become 0 and the positive values remain unchanged.
We use a Bzier patch here to build up the surface used on the source provided by
Direct3D ShaderX author Bart Sekura. A Bzier patch is the surface extension of the Bzier
curve. Whereas a curve is a function of one variable and takes a sequence of control points, the
patch is a function of two variables with an array of control points. A Bzier patch can be
viewed as a continuous set of Bzier curves. The Bzier patch is the most commonly used sur-
face representation in the computer graphic (see [Skinner]). CD3DBPatch::Create() builds the
Bzier patch in InitDeviceObjects() from nine control points and calculates the normals.
CD3DBPatch::RestoreDeviceObjects() fills the vertex and index buffer with the data on which
the m_pvVertices and m_pwIndices pointers refer.
Programming Vertex Shaders 65
Note: A way to do specular lighting without the lit instruction would be to
raise the specular power with a few mul instructions, like this:
mul r3, r3, r3 ; 2nd
mul r3, r3, r3 ; 4th
mul r3, r3, r3 ; 8th
mul r3, r3, r3 ; 16th
Team LRN
Summary
RacorX4 shows a specular reflection model that is very similar to the Blinn-Phong reflection
model. Additionally, it shows the functionality behind the lit instruction and the basics of tex-
ture mapping with a vertex shader.
RacorX5
The only improvement of
RacorX5 over RacorX4 is the
addition of a point light source.
The light can be moved via the
Home and End keys and the Left,
Right, Down, and Up arrow keys.
As with the previous example, the
specular light can be changed with
the C key.
Point Light Source
A point light source has color and
position within a scene but no sin-
gle direction. All light rays
originate from one point and illu-
minate equally in all directions.
The intensity of the rays will
remain constant regardless of their
distance from the point source,
unless a falloff value is explicitly stated. A point light is useful to simulate a lightbulb.
Light Attenuation for Point Lights
Usually, the energy from the point light that hits a surface diminishes as the distance increases.
There are a few functions used to simulate the idea of attenuation in computer graphics. The
simplest one might be:
func
Attenuation
= 1 / d
L
The inverse linear falloff doesnt look as good as the inverse squared falloff:
func
Attenuation
= 1 / d
L
* d
L
This is the attenuation equation that might be used in conjunction with a directional light
source. The problem with this function and a point light source is that when a light is far away,
66 Part 1: Introduction to Shader Programming
Figure 15: RacorX5 with point light
Note: A spot light is a light source in which all light rays illuminate in the shape
of a cone. The falloff, spread, and dropoff of a spotlight are adjustable.
Team LRN
its intensity is too low, but when it is close, the intensity becomes high, leading to visual
artifacts.
To get a wider range of effects, a decent attenuation equation is used:
func
Attenuation
= 1 / A
0
+ A
1
* d
L
+ A
2
* d
L
* d
L
This is the same equation that is used by the Direct3D fixed-function pipeline to calculate the
falloff of point lights. A
0
, A
1
, and A
2
are known as the attenuation constants. You get a radial
distance falloff with A
1
= 0 and A
2
> 0 and a linear falloff with A
1
> 0 and A
2
= 0. If the light
is too close to the surface, then attenuation becomes very large, making the light too bright. To
avoid this, set A
0
to 1 or greater.
The attenuation at the maximum range of the light is not 0.0. To prevent lights from sud-
denly appearing when they are at the light range, an application can increase the light range.
Or, the application can set up attenuation constants so that the attenuation factor is close to 0.0
at the light range. The attenuation value is multiplied by the red, green, and blue components
of the lights color to scale the lights intensity as a factor of the distance light travels to a
vertex.
To demonstrate the code used to implement this equation, the life cycle of a vertex shader
is tracked in the following sections.
Compared to the previous examples, the same input register layout is used by RacorX5,
but to feed the vertex shader with the attenuation data, additional constant registers have to be
filled with data.
Set the Vertex Shader Constants
The three attenuation values are set in RestoreDeviceObjects():
; in const.h:
#define LIGHT_ATTENUATION 17
...
// Attenuation
D3DXVECTOR4 Attenuation(1.0f, 0.0f, 0.0f, 0.0f);
m_pd3dDevice->SetVertexShaderConstant(LIGHT_ATTENUATION, &Attenuation.x, 1);
c17.x, c17.y, and c17.z hold the A0, A1, and A2 values. The world matrix and the point light
position, which is changeable with the Left, Right, Up, and Down arrow keys, are set in
FrameMove():
// set the world matrix
m_pd3dDevice->SetVertexShaderConstant(WORLD_MATRIX,&m_matWorld,4);
...
if(m_bKey[VK_LEFT]) { m_bKey[VK_RIGHT]=0;m_vLight.x -= 1.0f * m_fElapsedTime;}
if(m_bKey[VK_RIGHT]) { m_bKey[VK_LEFT]=0;m_vLight.x += 1.0f * m_fElapsedTime;}
if(m_bKey[VK_UP]) { m_bKey[VK_DOWN]=0; m_vLight.y += 1.0f * m_fElapsedTime;}
if(m_bKey[VK_DOWN]) { m_bKey[VK_UP]=0; m_vLight.y -= 1.0f * m_fElapsedTime;}
if(m_bKey[VK_END]) { m_bKey[VK_HOME]=0; m_vLight.z += 1.0f * m_fElapsedTime;}
if(m_bKey[VK_HOME]) { m_bKey[VK_END]=0; m_vLight.z -= 1.0f * m_fElapsedTime;}
// light direction
m_pd3dDevice->SetVertexShaderConstant(LIGHT_VECTOR, m_vLight,1);
The world matrix is used to calculate a vector that has its origin at the position of the vertex in
world space and its end point at the position of the light in world space.
Programming Vertex Shaders 67
Team LRN
The Vertex Shader
The vertex shader handles a combined diffuse and specular reflection model and the point light
source. The point lights position in world space and the coordinate of the vertex in world
space is used to derive a vector for the direction of the light and the distance that the light has
traveled.
If you compare this vertex shader to the previous one, the only new elements are the calcu-
lation of a light vector and the calculation of the attenuation factor.
; diffuse and specular point light
#include "const.h"
vs.1.1
; transpose and transform to clip space
mul r0, v0.x, c[CLIP_MATRIX]
mad r0, v0.y, c[CLIP_MATRIX_1], r0
mad r0, v0.z, c[CLIP_MATRIX_2], r0
add oPos, c[CLIP_MATRIX_3], r0
; output texture coords
mov oT0, v7
; transform normal
dp3 r1.x, v3, c[INVERSE_WORLD_MATRIX]
dp3 r1.y, v3, c[INVERSE_WORLD_MATRIX_1]
dp3 r1.z, v3, c[INVERSE_WORLD_MATRIX_2]
; renormalize it
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1, r1, r1.w
; transpose and transform position to world space
mul r7, v0.x, c[WORLD_MATRIX]
mad r7, v0.y, c[WORLD_MATRIX_1], r7
mad r7, v0.z, c[WORLD_MATRIX_2], r7
add r7, c[WORLD_MATRIX_3], r7
; calculate vector from light position
; to vertex position in world space
add r10.xyz, c[LIGHT_VECTOR], -r7.xyz
; renormalize vector VERTEX_TO_LIGHT
; d = distance
dp3 r10.w, r10, r10 ; r10.w = d * d = (x*x) + (y*y) + (z*z)
rsq r11.w, r10.w ; r11.w = 1/d = 1/||V|| = 1/sqrt(r7.w)
mul r11, r10, r11.w
dp3 r0.x, r1, r11 ; N dot L
; Get the attenuation
; Parameters for dst:
; src1 = (ignored, d * d, d * d, ignored)
; src2 = (ignored, 1/d, ignored, 1/d)
; r10.w = d * d
; r11.w = 1 / d
dst r4, r10.w, r11.w
; dest.x = 1
; dest.y = src0.y * src1.y
; dest.z = src0.z
68 Part 1: Introduction to Shader Programming
Team LRN
; dest.w = src1.w
; r4(1, d * d * 1 / d, d * d, 1/d)
; c[LIGHT_ATTENUATION].x = a0
; c[LIGHT_ATTENUATION].y = a1
; c[LIGHT_ATTENUATION].z = a2
dp3 r7.x, r4, c[LIGHT_ATTENUATION] ; (a0 + a1*d + a2* (d * d))
rcp r7.x, r7.x ; simplified:
; if (src.w != 0.0 && src.w != 1.0)
; 1 / (a0 + a1*d + a2* (d * d))
; compute normalized half vector H = L + V
add r2, c[EYE_VECTOR], r11 ; L + V
; renormalize H
dp3 r2.w, r2, r2
rsq r2.w, r2.w
mul r2, r2, r2.w
; N dot H
dp3 r0.y, r1, r2
; compute specular and clamp values (lit)
; r0.x - N dot L
; r0.y - N dot H
; r0.w - specular power n
mov r0.w, c[SPEC_POWER].y ; n must be between -128.0 and 128.0
lit r4, r0
; Scale the diffuse and specular intensity by the attenuation
mul r4, r4, r7.xxxx
; diffuse color * diffuse intensity (r4.y)
mul oD0, c[DIFFUSE_COLOR], r4.y
; specular color * specular intensity (r4.z)
mul oD1, c[SPEC_COLOR], r4.z
The lines, including the renormalization of the normal, should look quite familiar. The position
of the vertex is then transformed to world space, which is used to calculate the light vector L in
the following add instruction. L is renormalized in the following three instructions. It is impor-
tant to note that this renormalization is used at the same time to get the parameters for the dst
instruction. The following dp3 instruction performs the well-known diffuse reflection.
In the next paragraph, the dst, dp3, and rcp instructions calculate the attenuation factor
with the formula already shown above:
func
Attenuation
= 1 / A
0
+ A
1
* d
L
+ A
2
* d
L
* d
L
Therefore, dst gets distance * distance (squared distance) and 1 / distance via r10.w and r11.w.
This instruction then performs the following calculation:
r4.x = 1
r4.y = d = d * d * 1 /d
r4.z = d * d
r4.w = 1 / d
In the next dp3 instruction, the dot product between the first three components of r4 and the
attenuation constants is calculated:
; c[LIGHT_ATTENUATION].x = a0
; c[LIGHT_ATTENUATION].y = a1
; c[LIGHT_ATTENUATION].z = a2
Programming Vertex Shaders 69
Team LRN
dp3 r7.x, r4, c[LIGHT_ATTENUATION] ; (a0 + a1*d + a2* (d * d))
rcp r7.x, r7.x ; 1 / (a0 + a1*d + a2* (d * d))
The rcp instruction performs the following calculation:
if(src.w == 1.0f)
{
dest.x = dest.y = dest.z = dest.w = 1.0f;
}
else if(src.w == 0)
{
dest.x = dest.y = dest.z = dest.w = PLUS_INFINITY();
}
else
{
dest.x = dest.y = dest.z = m_dest.w = 1.0f/src.w;
}
Therefore, the source value of src.w is divided through 1.0f if it is not 1.0f or 0.0f. This corre-
sponds to the attenuation function postulated above.
The only thing left to get the point light effect is the component-wise multiplication of the
attenuation factor with the diffuse and specular factors:
; Scale the diffuse and specular intensity by the attenuation
mul r4, r4, r7.xxxx
Summary
This example showed the implementation of a point light with a falloff model as used in the
Direct3D fixed function pipeline. The light uses a combined diffuse and specular reflection
model. To improve this example, an ambient light might be added and a number of point lights
with different colors might be implemented.
What Happens Next?
This article covered basic lighting models and very basic transformations with vertex shaders.
To read more on advanced transformations, see [Gosselin] and [Isidoro/Card].
The next article shows the fundamentals for using pixel shaders.
References
[Blinn] James F. Blinn and Martin E. Newell, Texture and Reflection in Computer Generated
Images, Communications of the ACM (1976), 19: 542-546.
[Foley] James D. Foley, Andries van Dam, Steven K. Feiner, and John F. Hughes, Computer
Graphics Principles and Practice, 2nd Ed. (Addison Wesley, 1996) pp. 728-731.
[Gosselin] David Gosselin, Character Animation with Direct3D Vertex Shaders, Direct3D
ShaderX.
[Haines/Mller] Eric Haines and Thomas Mller, Normal Transforms,
http://www.gignews.com/realtime020100.htm.
70 Part 1: Introduction to Shader Programming
Team LRN
[Hurley] Kenneth Hurley, Photorealistic Faces with Vertex and Pixel Shaders, Direct3D
ShaderX.
[Isidoro/Card] John Isidoro and Drew Card, Animated Grass with Pixel and Vertex Shaders,
Direct3D ShaderX.
[RTR] Thomas Mller and Eric Haines, Real-Time Rendering (A. K. Peters, Ltd., 1999), p. 71.
[Savchenko] Sergei Savchenko, 3D Graphics Programming Games and Beyond (SAMS,
2000), p. 266.
[Skinner] Michael Skinner, Bzier Patches, http://www.gamedev.net/reference/articles/
article1584.asp.
[Taylor] Philip Taylor, Shader Optimization Techniques to Write More Efficient Shaders,
http://msdn.microsoft.com/directx.
[Turkowski] Ken Turkowski, Properties of Surface-Normal Transformations, in Andrew
Glassner (ed.), Graphics Gems (Academic Press, Inc., 1990), pp. 539-547, or
http:/www.worldserver.com/turk/computergraphics/index.html.
Acknowledgments
Id like to recognize the following individuals who were involved in proofreading and improv-
ing this article (in alphabetical order): David Callele (University of Saskatchewan) and Bart
Sekura (People Can Fly).
Programming Vertex Shaders 71
Team LRN
Fundamentals of Pixel
Shaders
Wolfgang F. Engel
The final output of any 3D graphics hardware consists of pixels. Depending on the resolution,
in excess of 2 million pixels may need to be rendered, lit, shaded, and colored. Prior to DirectX
8.0, Direct3D used a fixed-function multitexture cascade for pixel processing. The effects pos-
sible with this approach were very limited, depending on the implementation of the graphics
card device driver and the specific underlying hardware. A programmer was restricted on the
graphic algorithms implemented by these.
With the introduction of shaders in DirectX 8.0 and the improvements of pixel shaders in
DirectX 8.1, a whole new programming universe can be explored by game/demo coders. Pixel
shaders are small programs that are executed on individual pixels. This is an unprecedented
level of hardware control for their users.
This article shows you:
n
The place and the tasks of pixel shaders in the Direct3D pipeline
n
The architecture of the pixel shader unit
n
Tools that help to program pixel shaders
n
The ingredients for a pixel shader-driven application
The following article, Programming Pixel Shaders, will present many pixel shader concepts
and algorithms with example code.
Why Use Pixel Shaders?
Pixel shaders are, at the time of publication, supported by GeForce 3/4TI and RADEON
8500-based cards. Unlike vertex shaders, however, there is no feasible way to emulate pixel
shaders using software.
The best argument for using pixel shaders is to take a look at a few demos that use them or
one phrase: per-pixel non-standardized lighting. The gain in visual experience is enormous.
You can use membrane shaders for balloons, skins, Kubelka-Munk shaders for translucency
effects, fur/hair shaders, and new lighting formulas that lead to a totally new lighting experi-
ence. Think of a foreign planet with three moons moving faster than the two suns or with a
planet surface consisting of different crystalline substances, reflecting light in different ways.
72
Team LRN
The following list should give you a glimpse of the many types of effects that are possible by
using pixel shaders now:
n
Single pass, per-pixel lighting (see the next section)
n
True Phong shading [Beaudoin/Guardado]
n
Anisotropic lighting [Isidoro/Brennan]
n
Non-photorealistic rendering: cartoon shading, hatching, gooch lighting, image space tech-
niques [Card/Mitchell]
n
Per-pixel Fresnel term [Brennan]
n
Volumetric effects [Kraus][Hart]
n
Advanced bump mapping (self-shadowing bump maps, also known as horizon mapping)
n
Procedural textures [Zecha] and texture perturbation [Isidoro/Riguer]
n
Bidirectional reflectance distribution functions [Moravnsky]
There are also effects that have not yet been discovered or that are discovered but used only in
scientific journals. These visual effects are waiting until they get implemented by you.
One of the biggest catches of pixel shaders
is that they often have to be driven by the
vertex shader. For example, to calculate
per-pixel lighting, the pixel shader needs the
orientation of the triangle, the orientation of
the light vector, and, in some cases, the orien-
tation of the view vector. The graphics pipeline
shows the relationship between vertex and
pixel shaders.
The diagram at right shows the DX6/7
multitexturing unit and the new pixel shader
unit. On pixel shader-capable hardware, the
pixel shader heightens the number of transis-
tors on the graphic chip because it is added to
the already existing DX 6/7 multitexturing
hardware. Therefore, it is also functionally an
independent unit that the developer can choose
instead of the DX6/7 multitexturing unit.
But what happens in the 3D pipeline
before the pixel shader?
A vertex leaves the vertex shader as a
transformed and colored vertex. The so-called
backface culling removes all triangles that are
facing away from the viewer or camera. These
are by default the vertices that are grouped
counterclockwise. On average, half of your
game world triangles will be facing away from
the camera at any given time, so this helps
reduce rendering time. A critical point is the
usage of translucent or transparent front
Fundamentals of Pixel Shaders 73
Figure1: Direct3D pipeline
Team LRN
triangles. Depending on what is going to be rendered, backface culling can be switched off
with the D3DCULL_NONE flag.
Pixel Shaders in the Pipeline
User clip planes can be set by the developer, with the help of the graphics card, to clip trian-
gles that are outside of these planes and thus reduce the number of calculations. How to set
user clip planes is shown in the Direct3D 8.1 clip mirror example and in the ATI RADEON
8500 nature demo. The RADEON and RADEON 8500 support six independent user clip
planes in hardware. The clip planes on the GeForce 1/2/3/4TI chips are implemented using tex-
ture stages. That means that using two user clip planes eats up one texture stage.
One alternative to clip planes is shown in the TexKillClipPlanes example in the NVIDIA
Effects Browser, where the pixel shader instruction texkill is used to get a functionality similar
to clip planes.
Another alternative to clip planes is guard band clipping as supported by GeForce hard-
ware [Dietrich01]. The basic idea of guard band clipping is that hardware with a guard band
region can accept triangles that are partially or totally off-screen, thus avoiding expensive clip-
ping work. There are four cap fields for guard band support in the D3DCAPS8 structure:
GuardBandLeft, GuardBandRight, GuardBandTop, and GuardBandBottom. These represent
the boundaries of the guard band. These values have to be used in the clipping code. Guard
band clipping is handled automatically by switching on clipping in Direct3D.
Frustum clipping is performed on the viewing frustum. Primitives that lie partially or
totally off-screen must be clipped to the screen or viewport boundary, which is represented by
a 3D viewing frustum. The viewing frustum can be visualized as a pyramid, with the camera
positioned at the tip. The view volume defines what the camera will see and wont see. The
entire scene must fit between the new and far clipping planes and also be bounded by the sides,
bottom, and top of the frustum. A given object in relation to the viewing frustum can fall into
one of three categories [Watt92]:
n
The object lies completely outside the viewing area, in which case it is discarded.
n
The object lies completely inside the viewing frustum, in which case it is passed on to the
homogenous divide.
n
The object intersects the viewing frustum, in which case it is clipped against it and then
passed on to the homogenous divide.
74 Part 1: Introduction to Shader Programming
Note: Backface culling uses the crossproduct of two sides of a triangle to calculate a
vector that is perpendicular to the plane that is formed by these two sides. This vector is
the face normal. The direction of this normal determines whether the triangle is front-
or backfacing. Because the Direct3D API always uses the same vertex order to calculate
the crossproduct, it is known whether a triangles vertices are wound clockwise or
counterclockwise.
Note: It looks like NVIDIA no longer exposes the cap bits for these. The DirectX Caps
Viewer reports MaxUserClipPlanes == 0 since the release of the 12.41 drivers.
Team LRN
Translating the definition of the viewing frustrum above into homogenous coordinates gives us
the clipping limits:
-w <= x <= w
-w <= y <= w
0 <= z <= w
This would be a slow process if it had to be done by the CPU because each edge of each trian-
gle that crosses the viewport boundary must have an intersection point calculated, and each
parameter of the vertex (x,y,z diffuse r,g,b, specular r,g,b, alpha, fog, u, and v) must be interpo-
lated accordingly. Therefore, frustum clipping is done by modern graphics hardware on the
GPU essentially for free with clip planes and guard band clipping.
After frustum clipping, the homogenous or perspective divide happens. This means that
the x-, y-, and z-coordinates of each vertex of the homogenous coordinates are divided by w.
The perspective divide makes nearer objects larger and farther objects smaller, as you would
expect when viewing a scene in reality. After this division, the coordinates are in a normalized
space:
-1 <= x/w <= 1
-1 <= y/w <= 1
0 <= z/w <= 1
Clip coordinates are also referred to as normalized device coordinates (NDC). These coordi-
nates are now mapped to the screen by transforming into screen space via the so-called
viewport mapping. To map these NDCs to screen space, the following formula is used for a
resolution of 1024x768:
ScreenX(max) = NDC(X) * 512 + 512
ScreenY(max) = NDC(Y) * 384 + 384
The minimum and maximum value of X and Y is (1, 1) to (1, 1). When the NDCs is (1,
1), the screen coordinates are:
ScreenX = -1 * 512 + 512
ScreenY = -1 * 384 + 384
This point lies in the upper-left corner. For example, the lower-right corner will be reached
with NDCs of (1, 1). Although z and w values are retained for depth buffering tests, screen
space is essentially a 2D coordinate system, so only x and y values need to be mapped to
screen resolution.
Now comes the triangle setup, where the life of the vertices end and the life of the pixel
begins. It computes, triangle for triangle, the parameters required for the rasterization of the tri-
angles; among other things, it defines the pixel-coordinates of the triangle outline. This means
that it defines the first and the last pixel of the triangle scan line by scan line (see Figure 2):
Fundamentals of Pixel Shaders 75
Note: Why do we need this divide through w? By definition, every point gets a fourth
component w that measures distance along an imaginary fourth-dimensional axis called
w. For example, (4, 2, 8, 2) represents the same point as (2, 1, 4, 1). Such a w-value is
useful, for example, to translate an object because a 3x3 matrix is not able to translate an
object without changing its orientation. The fourth coordinate, w, can be thought of as
carrying the perspective information necessary to represent a perspective transformation.
Team LRN
Then the rasterizer interpolates color and depth values for each pixel from the color and depth
values of the vertices. These values are interpolated using a weighted average of the color and
depth values of the edges vertex values, where the color and depth data of edge pixels closer
to a given vertex more closely approximate values for that vertex. Then the rasterizer fills in
pixels for each line.
In addition, the texture coordinates are interpolated for use during the multitexturing/pixel
shader stage in a similar way.
The rasterizer is also responsible for Direct3D multisampling. Because it is done in the
rasterizer stage, multisampling only affects triangles and groups of triangles, not lines. It
increases the resolution of polygon edges and, therefore, depth and stencil tests as well. The
RADEON 7x00/GeForce 3 supports this kind of spatial anti-aliasing by setting the D3DRAS-
TERCAPS_STRECHBLTMULTISAMPLE flag, but both ignore the D3DRS_MULTI-
SAMPLEMASK render state to control rendering into the sub-pixel samples, whereas the
RADEON 8500 is able to mask the sub-samples with a different bit pattern by using this ren-
der state. By affecting only specific sub-samples, effects like motion blur or depth of field and
others can be realized.
76 Part 1: Introduction to Shader Programming
Figure 3: Filled triangle Figure 2: Triangle setup
Note: Alternatively, motion blur and depth of field effects are possible with the help of
vertex and pixel shaders; see the NVIDIA examples in the Effects Browser.
Note: GeForce 3/4TI, RADEON 7x00, and RADEON 8500 chips have another stage
in-between the triangle setup and the pixel shader. Both chips have a visibility subsystem
that is responsible for z occlusion culling. In the case of ATI cards, it is called hierarchi-
cal-Z. It determines if a pixel will be hidden behind a pixel rendered earlier or not. If the
z occlusion culling should determine that the pixel is hidden, it will directly be discarded
and it wont enter the pixel shader, thus saving the initial z-buffer read of the pixel-ren-
dering pipeline. During scan line conversion, the screen is divided into 8x8 pixel blocks
(RADEON 7x00) or 4x4 pixel blocks (RADEON 8500), and each block is assigned a bit
that determines whether that block is not cleared, meaning visible, or cleared,
meaning occluded. A quick z-check is performed on each of these blocks to determine
whether any pixels in them are visible. If a pixel in a block is visible, the status bit is
flipped to visible and the new z-value for that pixel is written to the z-buffer. If the block
is not visible, its status bit remains set to occluded and the z-buffer isnt touched. The
whole block will not be sent to the pixel shader.
Sorting objects or object groups from front to back when rendering helps z-culling to
kick out pixels from the graphics pipeline very early, saving not only a write into the
z-buffer (as with texkill) but also a read of the z-buffer.
Team LRN
The pixel shader is not involved on the sub-pixel level. It gets the already multisampled pixels
along with z, color values, and texture information. The already Gouraud-shaded or flat-shaded
pixel might be combined in the pixel shader with the specular color and the texture values
fetched from the texture map. For this task, the pixel shader provides instructions that affect
the texture addressing and instructions to combine the texture values in different ways with
each other.
There are five different pixel shader standards supported by Direct3D 8.1. Currently, no
pixel shader-capable graphics hardware is restricted on the support of ps.1.0. Therefore, all
available hardware supports at least ps.1.1. I will not mention the legacy ps.1.0 again. ps.1.1 is
the pixel shader version that is supported by GeForce 3. GeForce 4TI additionally supports
ps.1.2 and ps.1.3. The RADEON 8x00 supports all of these pixel shader versions, plus ps.1.4.
Whereas the ps.1.1, ps.1.2, and ps.1.3 are, from a syntactical point of view, built on each
other, ps.1.4 is a new and more flexible approach. ps.1.1-ps.1.3 differentiate between texture
address operation instructions and texture blending instructions. As the name implies, the first
kind of instructions is specialized and only usable for address calculations, and the second kind
of instructions is specialized and only usable for color shading operations. ps.1.4 simplifies the
pixel shader language by allowing the texture blending (color shader) instruction set to be used
for texture address (address shader) operations as well. It differentiates between arithmetic
instructions that modify color data and texture address instructions that process texture coordi-
nate data and, in most cases, sample a texture.
One might say that the usage of the instructions in ps.1.4 happens in a more RISC-like manner,
whereas the ps.1.1-ps.1.3 instruction sets are only usable in a CISC-like manner.
This RISC-like approach will be used in ps.2.0 that will appear in DirectX 9. Syntactically,
ps.1.4 is compared to ps.1.1-ps.1.3, an evolutionary step into the direction of ps.2.0.
What are the other benefits of using ps.1.4?
n
Unlimited texture address instructions: Direct3D 8.0 with ps.1.1 is restricted to a small
number of meaningful ways in which textures can be addressed. Direct3D 8.1 with ps.1.4
allows textures to be addressed and sampled in a virtually unlimited number of ways
because the arithmetic operation instructions are also usable as texture address operation
instructions.
n
Flexible dependent texture reads.
n
RADEON 8x00 performs ps.1.4 faster than ps.1.1-ps.1.3 because the latter is emulated.
n
All of the upcoming graphics hardware that will support ps.2.0 will support ps.1.4 for
backward compatibility.
The next stage in the Direct3D pipeline is fog. A fog factor is computed and applied to the
pixel using a blending operation to combine the fog amount (color) and the already shaded
pixel color, depending on how far away an object is. The distance to an object is determined by
its z- or w-value or by using a separate attenuation value that measures the distance between
the camera and the object in a vertex shader. If fog is computed per-vertex, it is interpolated
across each triangle using Gouraud shading. The MFC Fog example in the DirectX 8.1 SDK
shows linear and exponential fog calculated per-vertex and per-pixel. A layered range-based
Fundamentals of Pixel Shaders 77
Note: Sampling means looking up a color value in a texture at the specified coordi-
nates (up to four) (u, v, w, q) while taking into account the texture stage state attributes.
Team LRN
fog is shown in the height fog example in the NVIDIA Effects Browser. The volume fog exam-
ple in the DirectX 8.1 SDK shows volumetric fog produced with a vertex and a pixel shader
and the alpha blending pipeline. As shown in these examples, fog effects can be driven by the
vertex and/or pixel shader.
The alpha test stage kicks out pixels with a specific alpha value because these shouldnt be
visible. This is, for example, one way to map decals with an alpha-mapped texture. The alpha
test is switched on with D3DRS_ALPHATESTENABLE. The incoming alpha value is com-
pared with the reference alpha value, using the comparison function provided by the
D3DRS_ALPHAFUNC render state. If the pixel passes the test, it will be processed by the
subsequent pixel operation; otherwise, it will be discarded.
Alpha testing does not incur any extra overhead, even if all the pixels pass the test. By tun-
ing the reference value of alpha test appropriately to reject the transparent or almost
transparent pixels, one can improve the application performance significantly if the application
is memory bandwidth fill-bound. Basically, varying the reference value acts like a threshold,
setting up how many pixels are going to be evicted. The more pixels being discarded, the faster
the application will run.
There is a trick to drive the alpha test with a pixel shader, which is shown in the section on
swizzling. The image space outlining technique in [Card/Mitchell] uses alpha testing to reject
pixels that are not outlines and consequently improve performance.
The stencil test masks the pixel in the render target with the contents of the stencil buffer.
This is useful for dissolves, decaling, outlining, or building shadow volumes [Brennan]. See
the RadeonShadowShader example on the ATI web site.
The depth test determines whether a pixel is visible by comparing its depth value to the
stored depth value. An application can set up its z-buffer, z-min, and z-max with positive z
going away from the view camera in a left-handed coordinate system. The depth test is a
pixel-by-pixel logical test that asks Is this pixel in back of another pixel at this location? If
the answer is yes, that pixel gets discarded; if the answer is no, that pixel will travel further
through the pipeline, and the z-buffer is updated. There is also a more precise and band-
width-saving depth buffer form called w-buffer. It saves bandwidth by having only to send x-,
y-, and z-coordinates over the bus, while the z-buffer in certain circumstances has to send all
that plus z.
The pixel shader is able to drive the depth test with the texdepth (ps.1.4 only) or the
texm3x2depth (ps.1.3 only) instructions. These instructions can calculate the depth value used
in the depth buffer comparison test on a per-pixel basis.
The alpha blending stage blends the pixels data with the pixel data already in the render
target. The blending happens with the following formula:
FinalColor = SourcePixelColor * SourceBlendFactor +
DestPixelColor * DestBlendFactor
78 Part 1: Introduction to Shader Programming
Note: This is the only depth buffering method used by the Savage 2000-based boards.
These cards emulate the z-buffer with the help of the hardware w-buffer.
When using vertex shaders with w-buffering on a GeForce 3/4TI card, make sure the
projection matrix is set in the traditional way (using SetTransform()); otherwise, w-buffer-
ing wont work correctly. (Read more in the GeForce 3 FAQ, which can be found at
www.nvidia.com/developer.)
Team LRN
There are different flags that can be set with the D3DRS_SRCBLEND (SourceBlendFactor)
and D3DRS_DESTBLEND (DestBlendFactor) parameters in a SetRenderState() call. Alpha
blending was formerly used to blend different texture values together, but today it is more effi-
cient to do that with the multitexturing or pixel shader unit, depending on hardware support.
Alpha blending is used on newer hardware to simulate different levels of transparency.
Dithering tries to fool the eye into seeing more colors than are actually present by placing
different colored pixels next to one another to create a composite color that the eye will see.
For example, using a blue pixel next to a yellow pixel would lead to a green appearance. That
was a common technique in the days of the 8-bit and 4-bit color systems. You switch on the
dithering stage globally with D3DRS_DITHERENABLE.
The render target is normally the backbuffer in which you render the finished scene, but it
could also be the surface of a texture. This is useful for the creation of procedural textures or
for reusing results of former pixel shader executions. The render target is read by the stencil
test, depth test, and alpha blending stages.
To summarize the tasks of the pixel shader in the Direct3D pipeline:
n
The pixel shader is fed with rasterized pixels, which means, in most cases, z-occlusion
culled, multisampled, and already Gouraud-shaded pixels.
n
The pixel shader receives an interpolated texture coordinate and interpolated diffuse and
specular colors.
n
There are mainly four flavors of pixel shaders. The ps.1.1-ps.1.3 standards are, from a syn-
tactical point of view, a more CISC-like approach, whereas the ps.1.4 standard uses a
RISC-like approach that will also be used in the upcoming ps.2.0.
n
The pixel shader delivers the shaded pixel to the fog stage.
n
The pixel shader is able to support or drive the consecutive Direct3D stages.
Before getting our feet wet with a first look at the pixel shader architecture, lets take a look at
the currently available tools.
Pixel Shader Tools
I already introduced Shader Studio, Shader City, DLL Detective, 3D Studio Max 4.x/gmax,
NVASM, the Effects Browser, the Shader Debugger, and the Photoshop plug-ins from NVIDIA
in Fundamentals of Vertex Shaders. There is one thing to remember specifically for pixel
shaders: Because the GeForce 4TI supports ps.1.1-ps.1.3, it is possible that a few of the
NVIDIA tools wont support ps.1.4. Additionally, there are the following pixel shader-only
tools.
Microsoft Pixel Shader Assembler
The pixel shader assembler is provided with the DirectX 8.1 SDK. Like with the vertex shader
assembler, it does not come with any documentation. Its output look like Figure 4 on the
following page.
The pixel shader assembler is used by the Direct3DX functions that compile pixel shaders
and can also be used to precompile pixel shaders.
Fundamentals of Pixel Shaders 79
Team LRN
MFC Pixel Shader
The MFC pixel shader example
provided with the DirectX 8.1 SDK
comes with source. It is very useful
for quickly trying out pixel shader
effects and debugging them. Just
type in the pixel shader syntax you
want to test, and it will compile at
once. Debugging information is
provided in the window at the bot-
tom. If your graphics card doesnt
support a particular pixel shader
version, you can always choose the
reference rasterizer and test all
desired pixel shader versions. In the
figure shown at right, the reference
rasterizer was chosen on a GeForce
3 to simulate ps.1.3.
ATI ShadeLab
The ATI ShadeLab helps
with designing pixel
shaders. After writing
the pixel shader source
into the entry field in the
middle, the compilation
process starts immedi-
ately. To be able to load
the pixel shader later, it
has to be saved with the
Save button and loaded
with the Load button.
You may set up to
six textures with specific
texture coordinates and
the eight constant regis-
ters. The main advantage
80 Part 1: Introduction to Shader Programming
Figure 5: MFC pixel shader
Figure 4: Pixel shader
assembler output
Figure 6: ATI ShadeLab
Team LRN
over the MFC pixel shader tool is the ability to load constant registers and textures on your
own. This tool is provided on the companion CD in the Tools directory.
With that overview on the available tools in mind, we can go one step further by examin-
ing a diagram with the pixel shader workflow.
Pixel Shader Architecture
The diagram shows the
logical pixel shader data
workflow. The shaded
field at the top left
marks functionality spe-
cific for ps.1.1-ps.1.3.
The shaded field at the
bottom left marks func-
tionality that is specific
to ps.1.4.
On the right half of
the diagram, the pixel
shader arithmetic logic
unit (ALU) is sur-
rounded by four kinds of
registers. The color reg-
isters stream iterates
vertex color data from a
vertex shader or a
fixed-function vertex pipeline to a pixel shader. The constant registers provide constants to the
shader that are loaded by using the SetPixelShaderConstant() function or in the pixel shader
with the def instruction. The temporary registers rn are able to store temporary data. The r0
register also serves as the output register of the pixel shader.
The texture coordinates can be supplied as part of the vertex format or can be read from
certain kinds of texture maps. Texture coordinates are full precision and range, as well as per-
spective correct, when used in a pixel shader. There are D3DTSS_* texture operations that are
not replaced by the pixel shader functionality; they can be used on the textures (up to four for
ps.1.1-ps.1.3 or up to six for ps.1.4). The texture stages hold a reference to the texture data that
might be a one-dimensional (for example, in a cartoon shader), two-dimensional, or
three-dimensional texture (volume textures or cube map). Each value in a texture is called a
texel. These texels are most commonly used to store color values, but they can contain any
kind of data desired including normal vectors for bump maps, shadow values, or general
loopup tables.
Sampling occurs when a texture coordinate is used to address the texel data at a particular
location with the help of the texture registers. The usage of the texture registers tn differ
between the ps.1.1-ps.1.3 (t0-t3) and the ps.1.4 (t0-t5) implementations.
With ps.1.1-ps.1.3, the association between the texture coordinate set and the texture data
is a one-to-one mapping, which is not changeable in the pixel shader. Instead this association
Fundamentals of Pixel Shaders 81
Figure 7: Pixel shader logical data flow
Team LRN
can be changed by using the oTn registers in a vertex shader or by using the texture stage state
flag TSS_TEXCOORDINDEX together with SetTextureStageState(), in case the fixed-func-
tion pipeline is used.
In ps.1.4, the texture data and the texture coordinate set can be used independent of each
other in a pixel shader. The texture stage from which to sample the texture data is determined
by the register number rn, and the texture coordinate set that should be used for the texture data
is determined by the number of the tn register specified.
Lets take a closer look at the different registers shown in Figure 7:
Type Name ps.1.1 ps.1.2 ps.1.3 ps.1.4 Read/Write Caps
Constant Registers c
n
8 8 8 8 RO
Texture Registers t
n
4 4 4 6 RW / ps.1.4: RO
Temporary Registers r
n
2 2 2 6 RW
Color Registers v
n
2 2 2 2 in Phase 2 RO
Constant Registers (c0-c7)
There are eight constant registers in every pixel shader specification. Every constant register
contains four floating-point values or channels. They are read only from the perspective of the
pixel shader, so they could be used as a source register but never as destination registers in the
pixel shader. The application can write and read constant registers with calls to
SetPixelShaderContant() and GetPixelShaderConstant(). A def instruction used in the pixel
shader to load a constant register is effectively translated into a SetPixelShaderConstant() call
by executing SetPixelShader().
The range of the constant registers goes from 1 to +1. If you pass anything outside of this
range, it just gets clamped. Constant registers are not usable by ps.1.1-ps.1.3 texture address
instructions except for texm3x3spec, which uses a constant register to get an eye-ray vector.
Output and Temporary Registers (ps.1.1-ps.1.3:
r0+r1; ps.1.4: r0-r5)
The temporary registers r0-rn are used to store intermediate results. The output register r0 is
the destination argument for the pixel shader instruction, so r0 is able to serve as a temporary
and output register. In ps.1.4, r0-r5 are also used to sample texture data from the texture stages
0-5 in conjunction with the texture registers. In ps.1.1-ps.1.3, the temporary registers are not
usable by texture address instructions.
CreatePixelShader() will fail in shader preprocessing if a shader attempts to read from a
temporary register that has not been written to by a previous instruction. All shaders have to
write to r0.rgba the final result, or the shader will not assemble or validate.
Texture Registers (ps.1.1-ps.1.3: t0-t3; ps.1.4: t0-t5)
The texture registers are used in different ways in ps.1.1-ps.1.3 and in ps.1.4. In ps.1.1-ps.1.3,
the usage of one of the t0-t3 texture registers determines the usage of a specific pair of texture
data and texture coordinates. You cant change this one-to-one mapping in the pixel shader:
82 Part 1: Introduction to Shader Programming
Team LRN
ps.1.1 // version instruction
tex t0 // samples the texture at stage 0
// using texture coordinates from stage 0
mov r0, t0 // copies the color in t0 to output register r0
tex samples the texture data from the texture stage 0 and uses the texture coordinates set that is
set in the vertex shader with the oTn registers. In ps.1.4, having texture coordinates in their
own registers means that the texture coordinate set and the texture data are independent from
each other. The texture stage number with the texture data from which to sample is determined
by the destination register number (r0-r5) and the texture coordinate set is determined by the
source register (t0-t5) specified in phase 1.
ps.1.4 // version instruction
texld r4, t5
mov r0, r4
The texld instruction samples the map set via SetTexture (4, lpTexture) using the sixth set of
texture coordinates (set in the vertex shader with oT5) and puts the result into the fifth tempo-
rary register r4.
Texture registers that dont hold any values will be sampled to opaque black (0.0, 0.0, 0.0,
1.0). They can be used as temporary registers in ps.1.1-ps.1.3. The texture coordinate registers
in ps.1.4 are read only and, therefore, not usable as temporary registers. The maximum number
of textures is the same as the maximum number of simultaneous textures supported
(MaxSimultaneousTextures flag in D3D8CAPS).
Color Registers (ps.1.1-ps.1.4: v0+v1)
The color registers can contain per-vertex color values in the range 0 through 1 (saturated). It is
common to load v0 with the vertex diffuse color and v1 with the specular color.
Pixel shaders have read-only access to color registers. In ps.1.4, color registers are only avail-
able during the second phase, which is the default phase. All of the other registers are available
in each of the two phases of ps.1.4.
Range
One reason for using pixel shaders is that, compared to the multitexturing unit, its higher preci-
sion is used by the pixel shader arithmetic logic unit.
Register Name Range Versions
cn 1 to +1 all versions
rn D3DCAPS8.MaxPixelShaderValue to
D3DCAPS8.MaxPixelShaderValue
all versions
Fundamentals of Pixel Shaders 83
Note: Using a constant color (flat shading) is more efficient than using a
per-pixel Gouraud-shaded vertex color. If the shade mode is set to
D3DSHADE_FLAT, the application iteration of both vertex colors (diffuse and
specular) is disabled. Regardless of the shade mode, fog will still be iterated
later in the pipeline.
Team LRN
Register Name Range Versions
tn D3DCAPS8.MaxPixelShaderValue to
D3DCAPS8.MaxPixelShaderValue
ps.1.1-ps.1.3
D3DCAPS8.MaxTextureRepeat to
D3DCAPS8.MaxTextureRepeat
ps.1.4
vn 0 to +1 all versions
The color register vn is 8-bit precision per channel (i.e., 8-bit red, 8-bit green, etc.). For ps.1.1
to ps.1.3, D3DCAPS8.MaxPixelShaderValue is a minimum of one, whereas in ps.1.4,
D3DCAPS8.MaxPixelShaderValue is a minimum of eight. The texture coordinate registers
provided by ps.1.4 use high-precision signed interpolators. The DirectX caps viewer reports a
MaxTextureRepeat value of 2048 for the RADEON 8500. This value will be clamped to
MaxPixelShaderValue when used with texcrd because of the usage of an rn register as the
destination register. In this case, it is safest to stick with source coordinates within the Max-
PixelShaderValue range. However, if tn registers are used for straight texture lookups (i.e.,
texld r0, t3), then the MaxTextureRepeat range should be expected to be honored by hardware.
Using textures to store color values leads to much higher color precision with ps.1.4.
High-Level View of Pixel Shader Programming
Pixel shading takes place on a per-pixel, per-object basis during a rendering pass.
Lets start by focusing on the steps required to build a pixel shader-driven application. The
following list, in the sequence of execution, shows the necessary steps to build a pixel
shader-driven application:
n
Check for pixel shader support.
n
Set texture operation flags (with D3DTSS_* flags).
n
Set texture (with SetTexture()).
n
Define constants (with SetPixelShaderConstant()/def).
n
Pixel shader instructions.
n
Texture address instructions
n
Arithmetic instructions
n
Assemble pixel shader.
n
Create pixel shader.
n
Set pixel shader.
n
Free pixel shader resources.
The following text will work through this list step by step.
Check for Pixel Shader Support
It is important to check for the proper pixel shader support because there is no feasible way to
emulate pixel shaders. If there is no pixel shader support or the required pixel shader version is
84 Part 1: Introduction to Shader Programming
Team LRN
not supported, there have to be fallback methods to a default behavior (i.e., the multitexturing
unit or ps.1.1). The following statement checks the supported pixel shader version:
if( pCaps->PixelShaderVersion < D3DPS_VERSION(1,1) )
return E_FAIL;
This example checks the support of the pixel shader version 1.1. The support of at least ps.1.4
in hardware might be checked with D3DPS_VERSION(1,4). The D3DCAPS structure has to
be filled in the startup phase of the application with a call to GetDeviceCaps(). If the common
files framework, which is provided with the DirectX 8.1 SDK, is used, this is done by the
framework. If your graphics card does not support the requested pixel shader version and there
is no fallback mechanism that switches to the multitexturing unit, the reference rasterizer will
jump in. This is the default behavior of the common files framework, but it is not useful in a
game because the REF is too slow.
Set Texture Operation Flags (D3DTSS_* flags)
The pixel shader functionality replaces the D3DTSS_COLOROP and D3DTSS_ALPHAOP
operations and their associated arguments and modifiers that were used with the fixed-function
pipeline. For example, the following four SetTextureStageState() calls could be handled now
by the pixel shader:
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_COLORARG1, D3DTA_TEXTURE );
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_COLORARG2, D3DTA_DIFFUSE);
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_COLOROP, D3DTOP_MODULATE);
m_pd3dDevice->SetTexture( 0, m_pWallTexture);
The following texture stage states are still observed:
D3DTSS_ADDRESSU
D3DTSS_ADDRESSV
D3DTSS_ADDRESSW
D3DTSS_BUMPENVMAT00
D3DTSS_BUMPENVMAT01
D3DTSS_BUMPENVMAT10
D3DTSS_BUMPENVMAT11
D3DTSS_BORDERCOLOR
D3DTSS_MAGFILTER
D3DTSS_MINFILTER
D3DTSS_MIPFILTER
D3DTSS_MIPMAPLODBIAS
D3DTSS_MAXMIPLEVEL
D3DTSS_MAXANISOTROPY
D3DTSS_BUMPENVLSCALE
D3DTSS_BUMPENVLOFFSET
D3DTSS_TEXCOORDINDEX
D3DTSS_TEXTURETRANSFORMFLAGS
The D3DTSS_BUMP* states are used with the bem, texbem, and texbeml instructions.
In ps.1.1-ps.1.3, all D3DTSS_TEXTURETRANSFORMFLAGS flags are available and
have to be property set for a projective divide, whereas in ps.1.4, the texture transform flag
D3DTTFF_PROJECTED is ignored. It is accomplished by using source register modifiers
with the texld and texcrd registers.
Fundamentals of Pixel Shaders 85
Team LRN
The D3DTSS_TEXCOORDINDEX flag is valid only for fixed-function vertex processing.
When rendering with vertex shaders, each stages texture coordinate index must be set to its
default value. The default index for each stage is equal to the stage index.
ps.1.4 gives you the ability to change the association of the texture coordinates and the tex-
tures in the pixel shader. The texture wrapping, filtering, color border, and mip mapping flags
are fully functional in conjunction with pixel shaders.
Set Texture (with SetTexture())
After checking the pixel shader support and setting the proper texture operation flags, all tex-
tures have to be set by SetTexture(), as with the DX6/7 multitexturing unit. The prototype of
this call is:
HRESULT SetTexture(DWORD Stage, IDirect3DBaseTexture8* pTexture);
The texture stage that should be used by the texture is provided in the first parameter, and the
pointer to the texture interface is provided in the second parameter. A typical call might look
like this:
m_pd3dDevice->SetTexture( 0, m_pWallTexture);
This call sets the already loaded and created wall texture to texture stage 0.
Define Constants (with SetPixelShaderConstant()/def )
The constant registers can be filled with SetPixelShaderConstant() or the def instruction in the
pixel shader. Similar to the SetVertexShaderConstant() call, the prototype of the pixel shader
equivalent looks like this:
HRESULT SetPixelShaderConstant(
DWORD Register,
CONST void* pConstantData,
DWORD ConstantCount
);
First the constant register must be specified in Register. The data to transfer into the constant
register is provided in the second argument as a pointer. The number of constant registers that
have to be filled is provided in the last parameter. For example, to fill c0-c4, you might provide
c0 as the Register and 4 as the ConstantCount.
The def instruction is an alternative to SetPixelShaderConstant(). When SetPixelShader()
is called, it is effectively translated into a SetPixelShaderConstant() call. Using the def instruc-
tion makes the pixel shader easier to read. A def instruction in the pixel shader might look like
this:
def c0, 0.30, 0.59, 0.11, 1.0
Each value of the constant source registers has to be in the range [1.0..1.0].
86 Part 1: Introduction to Shader Programming
Note: A change to these texture stage states doesnt require the regeneration of
the currently bound shader because they are not available to the shader at com-
pile time, and the driver can therefore make no assumption about them.
Team LRN
Pixel Shader Instructions
Using vertex shaders, the programmer is free to
choose the order of the used instructions in any
way that makes sense, whereas pixel shaders need
a specific arrangement of the used instructions.
This specific instruction flow differs between
ps.1.1-ps.1.3 and ps.1.4.
ps.1.1-ps.1.3 allow four types of instructions
that must appear in the order shown in Figure 8.
This example shows a per-pixel specular
lighting model that evaluates the specular power
with a lookup table. Every pixel shader starts with
the version instruction. It is used by the assembler
to validate the instructions that follow. Below the
version instruction, a constant definition could be
placed with def. Such a def instruction is trans-
lated into a SetPixelShaderConstant() call when
SetPixelShader() is executed.
The next group of instructions are the so-called texture address instructions. They are used
to load data from the tn registers and, in ps.1.1-ps.1.3, to modify texture coordinates. Up to
four texture address instructions could be used in a ps.1.1-ps.1.3 pixel shader.
In this example, the tex instruction is used to
sample the normal map that holds the normal
data. There must always be at least two texm*
instructions:
texm3x2pad t1, t0_bx2
texm3x2tex t2, t0_bx2
Both instructions calculate the proper u/v texture
coordinate pair with the help of the normal map
in t0 and sample the texture data from t2 with it.
This texture register holds the light map data
with the specular power values. The last texture
addressing instruction samples the color map
into t3.
The next type of instructions are the arithme-
tic instructions. There could be up to eight
arithmetic instructions in a pixel shader.
mad adds t2 and c0, the ambient light, multi-
plies the result with t3, and stores it into the
output register r0.
Instructions in a ps.1.4 pixel shader must
appear in the order shown in Figure 9.
This is a simple transfer function, which
could be useful for sepia or heat signature
effects. It is explained in detail in [Mitchell]. The
Fundamentals of Pixel Shaders 87
Figure 8: ps.1.1-ps.1.3 pixel shader instruc-
tion flow (specular lighting with lookup
table)
Figure 9: ps.1.4 pixel shader instruction flow
(simple transfer function for sepia or heat
signature effects)
Team LRN
ps.1.4 pixel shader instruction flow has to start with the version instruction ps.1.4. After that,
as many def instructions as needed might be placed into the pixel shader code. This example
sets a luminance constant value with one def.
There could be up to six texture addressing instructions used after the constants. The texld
instruction loads a texture from texture stage 0 with the help of the texture coordinate pair 0,
which is chosen by using t0. In the following arithmetic instructions (up to eight), color, tex-
ture, or vector data might be modified. This shader uses only the arithmetic instruction to
convert the texture map values to luminance values.
So far, a ps.1.4 pixel shader has the same instruction flow as a ps.1.1-ps.1.3 pixel shader,
but the phase instruction allows it to double the number of texture addressing and arithmetic
instructions. It divides the pixel shader into two phases: phase 1 and phase 2. That means that
as of ps.1.4, a second pass through the pixel shader hardware can be done.
The additional six texture addressing instruction slots after the phase instruction are only used
by the texld r5, r0 instruction. This instruction uses the color in r0, which was converted to
luminance values before as a texture coordinate to sample a 1D texture (sepia or heat signature
map), which is referenced by r5. The result is moved with a mov instruction into the output
register r0.
Adding up the number of arithmetic and addressing instructions shown in the pixel shader
instruction flow in Figure 9 leads to 28 instructions. If no phase marker is specified, the default
phase 2 allows up to 14 addressing and arithmetic instructions.
Both of the preceding examples show the ability to use dependent reads. A dependent read
is a read from a texture map using a texture coordinate that was calculated earlier in the pixel
shader. More details on dependent reads will be presented in the next section.
Texture Address Instructions
Texture address instructions are used on texture coordinates. The texture coordinate address is
used to sample data from a texture map. Controlling the u/v pair, u/v/w triplet, or u/v/w/q qua-
druplet of texture coordinates with address operations gives the ability to choose different areas
of a texture map. Texture coordinate data storage space can also be used for other reasons
than sampling texture data. The registers that reference texture coordinate data are useful to
transport any kind of data from the vertex shader to the pixel shader via the oTn registers of
a vertex shader. For example, the light or half-angle vector or a 3x2, 3x3, or 4x4 matrix might
be provided to the pixel shader this way.
ps.1.1-ps.1.3 Texture Addressing
The following diagram shows the ways that texture address instructions work in ps.1.1-ps.1.3
for texture addressing:
88 Part 1: Introduction to Shader Programming
Note: Another way to reuse the result of a former pixel shader pass is to render
into a texture and use this texture in the next pixel shader pass. This is accom-
plished by rendering into a separate render target.
Team LRN
All of the texture addressing happens encapsulated in the texture address instructions
masked with a gray field. That means results of texture coordinate calculations are not accessi-
ble in the pixel shader. The texture address instruction uses these results internally to sample
the texture. The only way to get access to texture coordinates in the pixel shader is to use the
texcoord instruction. This instruction converts texture coordinate data to color values so that
they can be manipulated by texture addressing or arithmetic instructions. These color values
could be used as texture coordinates to sample a texture with the help of the texreg2*
instructions.
The following instructions are texture address instructions in ps.1.1-ps.1.3. The d and s in
the Para column are the destination and source parameters of the instruction. The usage of tex-
ture coordinates is shown by parentheses around the texture register, (t0)for example.
Instruction 1.1 1.2 1.3 Para Action
tex x x x d Loads tn with color data (RGBA) sampled from the texture.
texbem x x x d, s Transforms red and green components as du, dv signed values
of the source register using a 2D bump mapping matrix to mod-
ify the texture address of the destination register.
Can be used for a variety of techniques based on address per-
turbation such as fake per-pixel environment mapping, diffuse
lighting (bump mapping), environment matting, etc.
There is a difference in the usage of the texture stages between
environment mapping with a pixel shader (texbem, texbeml, or
bem) and environment mapping with the DX 6/7 multitexturing
unit. texbem (texbeml or bem) needs the matrix data connected
to the texture stage that is sampled. This is the environment map.
Environment mapping with the DX 6/7 multitexturing unit needs
the matrix data connected to the texture stage used by the bump
map (see also the example program Earth Bump on the com-
panion CD):
------------------
// texture stages for environment mapping
// with a pixel shader:
SetRenderState(D3DRS_WRAP0,D3DWRAP_U|D3DWRAP_V);
// color mapSetTexture( 0, m_pEarthTexture );
SetTextureStageState(0, D3DTSS_TEXCOORDINDEX, 1);
Fundamentals of Pixel Shaders 89
Figure 10 - Texture addressing
in ps.1.1-ps.1.3
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texbem (cont.)
// bump map
SetTexture(1, m_psBumpMap);
SetTextureStageState(1, D3DTSS_TEXCOORDINDEX, 1);
// environment map
SetTexture(2, m_pEnvMapTexture);
SetTextureStageState(2, D3DTSS_TEXCOORDINDEX, 0);
SetTextureStageState(2, D3DTSS_BUMPENVMAT00,
F2DW(0.5f));
SetTextureStageState(2, D3DTSS_BUMPENVMAT01,
F2DW(0.0f));
SetTextureStageState(2, D3DTSS_BUMPENVMAT10,
F2DW(0.0f));
SetTextureStageState(2, D3DTSS_BUMPENVMAT11,
F2DW(0.5f));
texbem performs the following operations:
u += 2x2 matrix(du)
v += 2x2 matrix(dv)
Then sample (u, v)
Read more in the Bump Mapping Formulas section in the
DirectX 8.1 SDK documentation.
Some rules for texbem/texbeml:
The s register cannot be read by any arithmetic instruction until it
is overwritten again by another instruction:
...
texbem/l t1, t0
mov r0, t0 ; not legal
...
texbem/l t1, t0
mov t0, c0
mov r0, t0 ; legal
...
The s register of texbem/l cannot be read by any texture instruc-
tion except for the texbem/l instruction:
...
texbem/l t1, t0
texreg2ar t2, t0 ; not legal
...
texbem/l t1, t0
texbem/l t2, t0 ; legal
...
---------------------------------
; t2 environment map
; (t2) texture coordinates environment map
; t1 du/dv perturbation data
ps.1.1
tex t0 ; color map
tex t1 ; bump map
texbem t2, t1 ; Perturb and then sample the
; environment map.
add r0, t2, t0
See the bem instruction for the ps.1.4 equivalent. See [Weiskopf]
for an interesting use of texbem.
90 Part 1: Introduction to Shader Programming
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texbeml x x x d,s Same as above, but applies additional luminance. Now three
components of the source register are used: red, green, and
blue as du, dv, and l for luminance.
u += 2x2 matrix(du)
v += 2x2 matrix(dv)
Then sample (u, v) and apply luminance.
See the texbem/l rules in the texbem section.
--------------------------------
; t1 holds the color map
; bump matrix set with the
; D3DTSS_BUMPENVMAT* flags
ps.1.1
tex t0 ; bump map with du, dv, l data
texbeml t1, t0 ; compute u, v
; sample t1 using u, v
; apply luminance correction
mov r0, t1 ; output result
texcoord x x x d Clamps the texture coordinate to the range [0..1.0] and outputs
it as color. If the texture coordinate set contains fewer than three
components, the missing components are set to 0. The fourth
component is always 1.
Provides a way to pass vertex data interpolated at high precision
directly into the pixel shader.
--------------------------------
ps.1.1
texcoord t0 ; convert texture coordinates
; to color
mov r0, t0 ; move color into output
; register r0
texdp3 x x d, s Performs a three-component dot product between the texture
coordinate set corresponding to the d register number and the
texture data in s and replicates clamped values to all four color
channels of d.
--------------------------------
; t0 holds color map
; (t1) hold texture coordinates
ps.1.2
tex t0 ; color map
texdp3 t1, t0 ; t1 = (t1) dot t0
mov r0, t1 ; output result
texdp3tex x x d, s Performs a three-component dot product between the texture
coordinate set corresponding to the d register number and the
texture data in s. Uses the result to do a 1D texture lookup in d
and places the result of the lookup into d. A common application
is to look up into a function table stored in a 1D texture for pro-
cedural specular lighting terms.
---------------------------------
Fundamentals of Pixel Shaders 91
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texdp3tex
(cont.)
x x d, s ; t1 holds 1D color map
ps.1.2
tex t0 ; vector data (x, y, z, w)
texdp3tex t1, t0 ; u = (t1) dot t0
; lookup data in t1
; result in t1
mov r0, t1 ; output result
texkill x x x s Cancels rendering of the current pixel if any of the first three
components of the texture coordinates in s is less than zero.
When using vertex shaders, the application is responsible for
applying the perspective transform. If the arbitrary clipping
planes contain anisomorphic scale factors, you have to apply the
perspective transform to these clip planes as well.
---------------------------------
ps.1.1
texkill t0 ; mask out pixel using
; uvw texture coordinates < 0.0
mov r0, v0
texm3x2depth x d, s Calculates together with a texm3x2pad instruction the depth
value to be used in depth testing for this pixel. Performs a
three-component dot product between the texture coordinate set
corresponding to the d register number and the second row of a
3x2 matrix in the s register and stores the resulting w into d.
After execution, the d register is no longer available for use in
the shader.
The benefit of a higher resolution of the depth buffer resulting
from multisampling is eliminated because texm3x2depth (same
with ps.1.4: texdepth) will output the single depth value to each
of the sub-pixel depth comparison tests.
Needs to be clamped to [0..1] w and z values or the result
stored in the depth buffer will be undefined.
---------------------------------
; (t1) holds row #1 of the 3x2 matrix
; (t2) holds row #2 of the 3x2 matrix
; t0 holds normal map
ps.1.3
tex t0 ; normal map
texm3x2pad t1, t0 ; calculates z from row #1
; calculates w from row #2
; stores a result in t2 depending on
; if (w == 0)
; t2 = 1.0;
; else
; t2 = z/w;
texm3x2depth t2, t0
texm3x2pad x x x d, s This instruction cannot be used by itself. It must be combined
with either texm3x2tex or texm3x2depth. It performs a
three-component dot product between the texture coordinate
set corresponding to the d register number and the data of the
s register and stores the result in d.
92 Part 1: Introduction to Shader Programming
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texm3x2pad
(cont.)
x x x d, s See the example shown for the texm3x2depth and texm3x2tex
instructions.
texm3x2tex x x x d, s It calculates the second row of a 3x2 matrix by performing a
three-component dot product between the texture coordinate set
corresponding to the d register number and the data of the s
register to get a v value, which is used to sample a 2D texture.
This instruction is used in conjunction with texm3x2pad, which
calculates the u value.
--------------------------------
; Dot3 specular lighting with a lookup table
ps.1.1
; t0 holds normal map
; (t1) holds row #1 of the 3x2 matrix (light vector)
; (t2) holds row #2 of the 3x2 matrix (half vector)
; t2 holds a 2D texture (lookup table)
; t3 holds color map
tex t0 ; sample normal
texm3x2pad t1, t0_bx2 ; calculates u from first row
texm3x2tex t2, t0_bx2 ; calculates v from second row
; samples texel with u,v
; from t2 (lookup table)
tex t3 ; sample base color
mul r0,t2,t3 ; blend terms
-------
; A ps.1.4 equivalent to the
; texm3x2pad/texm3x2tex pair could be
; specular power from a lookup table
ps.1.4
; r0 holds normal map
; t1 holds light vector
; t2 holds half vector
; r2 holds 2D texture (lookup table)
; r3 holds color map
texld r0, t0
texcrd r1.rgb, t1
texcrd r4.rgb, t2
dp3 r1.r, r1, r0_bx2 ; calculates u
dp3 r1.g, r4, r0_bx2 ; calculates v
phase
texld r3, t3
texld r2, r1 ; samples texel with u,v
; from r2 (lookup table)
mul r0, r2, r3
texm3x3 x x d, s Performs a 3x3 matrix multiply similar to texm3x3tex, except that
it does not automatically perform a lookup into a texture. The
returned result vector is placed in d with no dependent read.
The .a value in d is set to 1.0. The 3x3 matrix is comprised of
the texture coordinates of the third texture stage and by the two
preceding texture stages. Any texture assigned to any of the
three texture stages is ignored. This instruction must be used with
two texm3x3pad instructions.
Fundamentals of Pixel Shaders 93
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texm3x3
(cont.)
x x d, s -----------------------------------
; (t1) holds row #1 of the 3x3 matrix
; (t2) holds row #2 of the 3x3 matrix
; (t3) holds row #3 of the 3x3 matrix
ps.1.2
tex t0 ; normal map
texm3x3pad t1, t0 ; calculates u from row #1
texm3x3pad t2, t0 ; calculates v from row #2
texm3x3 t3, t0 ; calculates w from row #3
; store u, v , w in t3
mov r0, t3
; ps.1.4 equivalent
; r1 holds row #1 of the 3x3 matrix
; r2 holds row #2 of the 3x3 matrix
; r3 holds row #3 of the 3x3 matrix
ps.1.4
def c0, 1.0, 1.0, 1.0, 1.0
texld r0, t0 ; r0 normal map
texcrd r1.rgb, t1 ; calculates u from row #1
texcrd r2.rgb, t2 ; calculates v from row #2
texcrd r3.rgb, t3 ; calculates w from row #3
dp3 r4.r, r1, r0
dp3 r4.g, r2, r0
dp3 r4.b, r3, r0 ; store u, w, w in r4.rgb
mov r0.a, c0
+mov r0.rgb, r4
texm3x3pad x x x d, s Performs the first or second row of a 3x3 matrix multiply. The
instruction cannot be used by itself and must be used with
texm3x3, texm3x3spec, texm3x3vspec, or texm3x3tex.
texm3x3spec x x x d, s1,
s2
Performs, together with two texm3x3pad instructions, a 3x3
matrix multiply. The resulting vector is used as a normal vector
to reflect the eye-ray vector from a constant register in s2 and
then uses the reflected vector as a texture address for a texture
lookup in d.
The 3x3 matrix is typically useful for orienting a normal vector
of the correct tangent space for the surface being rendered. The
3x3 matrix is comprised of the texture coordinates of the third
texture stage and the results in the two preceding texture stages.
Any texture assigned to the two preceding texture stages is
ignored.
This can be used for specular reflection and environment
mapping.
---------------------------------------
; (t1) holds row #1 of the 3x3 matrix
; (t2) holds row #2 of the 3x3 matrix
; (t3) holds row #3 of the 3x3 matrix
; t3 is assigned a cube or volume map with
; color data (RGBA)
; t0 holds a normal map
; c0 holds the eye-ray vector E
94 Part 1: Introduction to Shader Programming
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texm3x3spec
(cont.)
ps.1.1
tex t0
texm3x3pad t1, t0 ; calculates u from row #1
texm3x3pad t2, t0 ; calculates v from row #2
; calculates w from row #3
; reflect u, v and w by the
; eye-ray vector in c0
; use reflected vector to lookup texture in t3
texm3x3spec t3, t0, c0
mov r0, t3 ; output result
; A similar effect is possible with the following
; ps.1.4 pixel shader.
; The eye vector is stored as a normalized
; vector in a cube map
ps.1.4
texld r0, t0 ; Look up normal map.
texld r1, t4 ; Eye vector through normalizer cube map
texcrd r4.rgb, t1 ; 1st row of environment matrix
texcrd r2.rgb, t2 ; 2st row of environment matrix
texcrd r3.rgb, t3 ; 3rd row of environment matrix
dp3 r4.r, r4, r0_bx2 ; 1st row of matrix multiply
dp3 r4.g, r2, r0_bx2 ; 2nd row of matrix multiply
dp3 r4.b, r3, r0_bx2 ; 3rd row of matrix multiply
dp3 r5, r4, r1_bx2 ; (N.Eye)
mov r0, r5
texm3x3tex x x x d, s Performs, together with two texm3x3pad instructions, a 3x3
matrix multiply and uses the result to look up the texture in d.
The 3x3 matrix is typically useful for orienting a normal vector
to the correct tangent space for the surface being rendered. The
3x3 matrix is comprised of the texture coordinates of the third
texture stage and the two preceding texture stages. The resulting
u, v, and w is used to sample the texture in stage 3. Any textures
assigned to the preceding textures are ignored.
------------------------------------
; (t1) holds row #1 of the 3x3 matrix
; (t2) holds row #2 of the 3x3 matrix
; (t3) holds row #3 of the 3x3 matrix
; t3 is assigned a cube or volume map with
; color data (RGBA)
ps.1.1
tex t0 ; normal map
texm3x3pad t1, t0 ; calculates u from row #1
texm3x3pad t2, t0 ; calculates v from row #2
; calculates w from row #3
; uses u, v and w to sample t3
texm3x3tex t3, t0
mov r0, t3 ; output result
; ps.1.4 equivalent
; r1 holds row #1 of the 3x3 matrix
; r2 holds row #2 of the 3x3 matrix
; r3 holds row #3 of the 3x3 matrix
Fundamentals of Pixel Shaders 95
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texm3x3tex
(cont.)
; r3 is assigned a cube or volume map with
; color data (RGBA)
ps.1.4
texld r0, t0
texcrd r1.rgb, t1
texcrd r2.rgb, t2
texcrd r3.rgb, t3
dp3 r4.r, r1, r0 ; calculates u from row #1
dp3 r4.g, r2, r0 ; calculates v from row #2
dp3 r4.b, r3, r0 ; calculates w from row #3
phase
texld r3, r4
mov r0, r3
texm3x3vspec x x x d, s Performs, together with two texm3x3pad instructions, a 3x3
matrix multiply. The resulting vector is used as a normal vector
to reflect the eye-ray vector and then uses the reflected vector as
a texture address for a texture lookup. It works just like
texm3x3spec, except that the eye-vector is taken from the q
coordinates of the three sets of 4D textures. The 3x3 matrix is
typically useful for orienting a normal vector of the correct tan-
gent space for the surface being rendered. The 3x3 matrix is
comprised of the texture coordinates of the third texture stage
and the results in the two preceding texture stages. Any texture
assigned to the two preceding texture stages is ignored.
This can be used for specular reflection and environment map-
ping, where the eye-vector is not constant.
-----------------------------------------
; (t1) holds row #1 of the 3x3 matrix
; (t2) holds row #2 of the 3x3 matrix
; (t3) holds row #3 of the 3x3 matrix
; t3 is assigned a cube or volume map with
; color data (RGBA)
; t0 holds a normal map
; used for cubic bump mapping
; the per-vertex eye vector is derived using
; the camera position in the vertex shader
ps.1.1
tex t0
texm3x3pad t1, t0 ; calculates u from row #1
texm3x3pad t2, t0 ; calculates v from row #2
; calculates w from row #3
; calculates eye-ray vector from the q texture
; coordinate values of t1 - t3
; reflect u, v and w by the eye-ray vector
; use reflected vector to lookup texture in t3
texm3x3vspec t3, t0
mov r0, t3 ; output result
96 Part 1: Introduction to Shader Programming
Team LRN
Instruction 1.1 1.2 1.3 Para Action
texreg2ar x x x d, s General dependent texture read operation that takes the alpha
and red color components of s as texture address data (u, v)
consisting of unsigned values to sample a texture at d.
---------------------------------------
ps.1.1
tex t0 ; color map
texreg2ar t1, t0
mov r0, t1
texreg2gb x x x d, s General dependent texture read operation that takes the green
and blue color components of s as texture address data (u, v)
consisting of unsigned values to sample a texture at d.
---------------------------------------
ps.1.1
tex t0 ; color map
texreg2gb t1, t0
mov r0, t1
texreg2rgb x x d, s General dependent texture read operation that takes the red,
green, and blue color components of s as texture address data
(u, v, w) consisting of unsigned values to sample a texture at d.
This is useful for color-space remapping operations.
---------------------------------------
ps.1.2
tex t0 ; color map
texreg2rgb t1, t0
mov r0, t1
All of these texture address instructions use only the tn registers with the exception of
tex3x3spec, which uses a constant register for the eye-ray vector. In a ps.1.1-ps.1.3 pixel
shader, the destination register numbers for texture addressing instructions had to be in increas-
ing order.
In ps.1.1-ps.1.3, the ability to reuse a texture coordinate after modifying it in the pixel
shader is available through specific texture address instructions that are able to modify the tex-
ture coordinates and sample a texture with these afterward. The following diagram shows this
reliance:
Fundamentals of Pixel Shaders 97
Figure 11: Dependent
read in ps.1.1-ps.1.3
Team LRN
The texture address operations that sample a texture after modifying the texture coordinates
are:
n
texbem/texbeml
n
texdp3tex
n
texm3x2tex
n
texm3x3spec
n
texm3x3tex
n
texm3x3vspec
The following instructions sample a texture with the help of color values as texture coordi-
nates. If one of these color values has been changed by an instruction, the sampling happens to
be a dependent read.
n
texreg2ar
n
texreg2gb
n
texreg2rgb
Therefore, these instructions are called general dependent texture read instructions.
As already stated above, each ps.1.1-ps.1.3 pixel shader has a maximum of eight arithme-
tic instructions and four texture address instructions. All texture address instructions use one of
the supplied slots, with the exception of texbeml, which uses one texture address slot plus one
arithmetic slot.
ps.1.4 Texture Addressing
To use texels or texture coordinates in ps.1.4, you have to load them first with texld or texcrd.
These instructions are the only way to get access to texels or texture coordinates. Texture coor-
dinates can be modified after a conversion to color data via texcrd, with all available arithmetic
instructions. As a result, texture addressing is more straightforward with ps.1.4.
The following instructions are texture address instructions in ps.1.4:
Instruction Para Action
texcrd d, s Copies the texture coordinate set corresponding to s into d as color data (RGBA). No
texture is sampled. Clamps the texture coordinates in tn with a range of
[MaxTextureRepeat, MaxTextureRepeat] (RADEON 8500: 2048) to the range of rn
[8, 8] (MaxPixelShaderValue). This clamp might behave differently on different hard-
ware. To be safe, provide data in the range of [8, 8].
A .rgb or .rg modifier should be provided to d. The fourth channel of d is unset/unde-
fined in all cases. The third channel is unset/undefined for a projective divide with
_dw.xyz (D3DTFF_PROJECTED is ignored in ps.1.4). The allowed syntax, taking into
account all valid source modifier/selector and destination write mask combinations, is
shown below:
texcrd rn.rgb, tn.xyz
texcrd rn.rgb, tn
texcrd rn.rgb, tn.xyw
texcrd rn.rg, tn_dw.xyw
98 Part 1: Introduction to Shader Programming
Team LRN
Instruction Para Action
texdepth d Calculates the depth value used in the depth buffer comparison test for the pixel by
using the r5 register. The r5 register is then unavailable for any further use in the pixel
shader. texdepth updates the depth buffer with the value of r5.r and r5.g. The .r chan-
nel is treated as the z-value and the .g channel is treated as the w-value. The value in
the depth buffer is replaced by the result of the .r channel divided by the .g channel
== z/w. If the value in .g channel is zero, then the depth buffer is updated with 1.0.
texdepth is only available in phase 2.
Using this instruction eliminates the benefit of a higher resolution of the depth buffer
resulting from multisampling, because texdepth (and texm3x2depth) will output the
single depth value to each of the sub-pixel depth comparison tests.
----------------------------------------
ps.1.4
texld r0, t0 ; samples from texture stage 0 with
; texture coordinates set t0
texcrd r1.rgb, t1 ; load texture coordinates from
; t1 into r1 as color values
add 5.rg, r0, r1 ; add both values
phase
texdepth r5 ; calculate pixel depth as r5.r/r5.g
// do other color calculation here and output it to r0
texkill s Cancels rendering of the current pixel if any of the first three components of the texture
data (ps.1.1-ps.1.3: texture coordinates) in s is less than zero. You can use this instruc-
tion to implement arbitrary clipping planes in the rasterizer.
When using vertex shaders, the application is responsible for applying the perspective
transform. If the arbitrary clipping planes contain anisomorphic scale factors, you have
to apply the perspective transform to the clip planes as well.
texkill is only available in phase 2 and sources rn or tn registers.
---------------------------------
ps.1.4
... ; include other shader instructions here
phase
texkill t0 ; mask out pixel using
; uvw texture coordinates < 0.0
mov r0, v0 ; move diffuse into r0
texld d, s Loads d with the color data (RGBA) sampled using the texture coordinates from s. The
texture stage number with the texture data from which to sample is determined by the
number of d (r0-r5), and the texture coordinate set is determined by the number of src
(t0-t5) in phase 1. texld is able to additionally use rn as s in phase 2.
The allowed syntax is:
texld rn, tn
texld rn, tn.xyz ; same as previous
texld rn, tn.xyw
texld rn, tn_dw.xyw
texld rn, rn
texld rn, rn.xyz ; same as previous
texld rn, rn_dz ; only valid on rn
; no more than two times per shader
texld rn, rn_dz.xyz ; same as previous
Fundamentals of Pixel Shaders 99
Team LRN
Instruction Para Action
texld (cont.) ----------------------------------
; Simple transfer function for sepia or
; heat signature effects [Mitchell]
; c0 holds the luminance value
; t0 holds the texture coordinates
; r0 holds the original image
; r5 holds the 1D sepia or heat signature map
ps.1.4
def c0, 0.30, 0.59, 0.11, 1.0
texld r0, t0
dp3 r0, r0, c0
phase
texld r5, r0 ; dependent read
mov r0, r5
; ps.1.2 equivalent
; t0 holds the original image
; t1 holds the 1D sepia or heat signature map
; (t1) holds 0.30, 0.59, 0.11, 1.0
ps.1.2
tex t0 ; color map
texdp3tex t1, t0 ; u = (t1) dot t0
; lookup data in t1
; result in t1
mov r0, t1 ; output result
In ps.1.4, there are only four texture address instructions, but, as mentioned before, all the
arithmetic instructions can be used to manipulate texture address information. So there are
plenty of instruments to solve texture addressing tasks.
Valid source registers for first phase texture address instructions are tn. Valid source regis-
ters for second phase texture address instructions are tn and also rn. Each rn register may be
specified as a destination to a texture instruction only once per phase. Aside from this, destina-
tion register numbers for texture instructions do not have to be in any particular order (as
opposed to previous pixel shader versions in which destination register numbers for texture
instructions had to be in increasing order).
No dependencies are allowed in a block of tex* instructions. The destination register of a
texture address instruction cannot be used as a source in a subsequent texture address instruc-
tion in the same block of texture address instruction (same phase).
Dependent reads with ps.1.4 are not difficult to locate in the source. Pseudocode of the two
possible dependent read scenarios in ps.1.4 might look like:
; transfer function
texld ; load first texture
modify color data here
phase
texld ; sample second texture with changed color data as address
or
texcrd ; load texture coordinates
modify texture coordinates here
phase
texld ; sample texture with changed address
100 Part 1: Introduction to Shader Programming
Team LRN
Another way to think of it is that if the second argument to a texld after the phase marker is rn
(not tn), then its a dependent read because the texture coordinates are in a temp register so
they must have been computed:
.....
phase
texld rn, rn
Set the first three channels of an rn register, which is used as a source register. This must be set
before it is used as a source parameter. Otherwise, the shader will fail.
To manipulate texture coordinates with arithmetic instructions, they have to be loaded into
texture data registers (ps.1.1-ps.1.3: tn; ps.1.4: rn) via texcoord or texcrd. There is one impor-
tant difference between these two instructions; texcoord clamps to [0..1] and texcrd does not
clamp at all.
If you compare texture addressing used in ps.1.1-ps.1.3 and texture addressing used in
ps.1.4, it is obvious that the more CISC-like approach uses much more powerful instructions to
address textures compared to the more RISC-like ps.1.4 approach. On the other hand, ps.1.4
offers a greater flexibility in implementing different texture addressing algorithms by using all
of the arithmetic instructions compared to ps.1.1-ps.1.3.
Arithmetic Instructions
The arithmetic instructions are used by ps.1.1-ps.1.3 and ps.1.4 in a similar way to manipulate
texture or color data. Here is an overview of all available instructions in these
implementations:
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
dest
x x ps.1.1-ps.1.3
x ps.1.4
add src0, src1 x x x x ps.1.1-ps.1.3
x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Performs a component-wise addition of register src0 and src1:
dest.r = src0.r + src1.r
dest.g = src0.g + src1.g
dest.b = src0.b + src1.b
dest.a = src0.a + src1.a
--------------------------------------
; glow mapping
ps.1.1
tex t0 ; color map
tex t1 ; glow map
add r0, t0, t1 ; add the color values
; increase brightness lead to a glow effect
; glow mapping
ps.1.4
texld r0, t0 ; color map
texld r1, t1 ; glow map
Fundamentals of Pixel Shaders 101
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
add r0, r0, r1 ; add the color values
; increase brightness lead to a glow effect
---------
; detail mapping
ps.1.1
tex t0 ; color map
tex t1 ; detail map
add r0, t0, t1_bias ; detail map is add-signed to the color map
; watch out for the used texture coordinates of
; the detail map
; detail mapping
ps.1.4
texld r0, t0 ; color map
texld r1, t0 ; sample detail map with the texture coords of the color map
add r0, r0, r1_bias ; detail map is add-signed to the color map
You may increase the detail map effect by using _bx2 as a modifier in the add instruction.
dest x ps.1.4 phase 1
bem src0 x x ps.1.4 phase 1
src1 x ps.1.4 phase 1
Apply a fake bump environment transform.
There is a difference in the usage of the texture stages between environment mapping with a pixel shader
(meaning texbem, texbeml, or bem) and environment mapping with the DX 6/7 multitexturing unit. bem (texbeml
or texbem) needs the matrix data connected to the texture stage that is sampled. This is the environment map.
Environment mapping with the DX 6/7 multitexturing unit needs the matrix data connected to the texture stage
used by the bump map (see the example code for the texbem instruction).
bem has a lot of restrictions when used in a pixel shader:
n
bem must appear in the first phase of a shader (that is, before a phase marker).
n
bem consumes two arithmetic instruction slots.
n
Only one use of this instruction is allowed per shader.
n
Destination writemask must be .rg /.xy.
n
This instruction cannot be co-issued.
n
Aside from the restriction that destination write mask be .rg, modifiers on source src0, src1, and instruc-
tion modifiers are unconstrained.
bem performs the following calculation:
(Given n == dest register #)
dest.r = src0.r + D3DTSS_BUMENVMAT00(stage n) * src1.r
+ D3DTSS_BUMPENVMAT10(stage n) * src1.g
dest.g = src0.g + D3DTSS_BUMENVMAT01(stage n) * src1.r
+ D3DTSS_BUMPENVMAT11(stage n) * src1.g
------------------------------------
102 Part 1: Introduction to Shader Programming
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
ps.1.4
; r2 environment map texture coordinates
; r1 du/dv perturbation data
texld r1, t1 ; bump map
texcrd r2.rgb, t2
bem r2.rg, r2, r1 ; perturb
; r2.rg = tex coordinates to sample environment map
phase
texld r0, t0 ; color map
texld r2, r2 ; environment map
add r0, r0, r2
See the BumpEarth example program on the companion CD. See also the articles on improved environment
mapping techniques, such as cube mapping [Brennan2][Brennan3] and per-fresnel term [Brennan].
dest
x x ps.1.2, ps.1.3
x ps.1.4
cmp x x x x ps.1.2, ps.1.3
src0, src1, src2 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Conditionally chooses between src1 and src2 based on a per-channel comparison src0 >= 0.
For ps.1.2 and ps.1.3, cmp uses two arithmetic instruction slots. CreatePixelShader() erroneously assumes that
this instruction consumes only one instruction slot. So you have to check the instruction count of a pixel shader
that uses this instruction manually. Another validation problem is that in ps.1.2 and ps.1.3, the destination regis-
ter of cmp cannot be the same as any of the source registers.
// Compares all four components.
ps.1.2
... fill t1, t2 and t3
// t1 holds -0.6, 0.6, 0, 0.6
// t2 holds 0, 0, 0, 0
// t3 holds 1, 1, 1, 1
cmp r0, t1, t2, t3 // r0 is assigned 1,0,0,0 based on the following:
// r0.x = t3.x because t1.x < 0
// r0.y = t2.y because t1.y >= 0
// r0.z = t2.z because t1.z >= 0
// r0.w = t2.w because t1.w >= 0
----------
// Compares all four components.
ps.1.4
texcrd r1, t1
texcrd r2, t2
texcrd r3, t3
// r1 holds -0.6, 0.6, 0, 0.6
// r2 holds 0, 0, 0, 0
// r3 holds 1, 1, 1, 1
cmp r0, r1, r2, r3 // r0 is assigned 1,0,0,0 based on the following:
// r0.x = r3.x because r1.x < 0
// r0.y = r2.y because r1.y >= 0
Fundamentals of Pixel Shaders 103
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
// r0.z = r2.z because r1.z >= 0
// r0.w = r2.w because r1.w >= 0
----------
; Cartoon pixel shader
; explained in detail in [Card/Mitchell]
; c0 holds falloff 1
; c1 holds falloff 2
; c2 holds dark
; c3 holds average
; c4 holds bright
; t0 holds normal information
; t1 holds the light vector
ps.1.4
def c0, 0.1f, 0.1f, 0.1f, 0.1f
def c1, 0.8f, 0.8f, 0.8f, 0.8f
def c2, 0.2f, 0.2f, 0.2f, 1.0f
def c3, 0.6f, 0.6f, 0.6f, 1.0f
def c4, 0.9f, 0.9f, 1.0f, 1.0f
texcrd r0.xyz, t0 ; place normal vector in r0
texcrd r1.xyz, t1 ; place light vector in r1
dp3 r3, r0, r1 ; n.l
sub r4, r3, c0 ; subtract falloff #1 from n.l
cmp_sat r0, r4, c3, c2 ; check if n.l is > zero
; if yes use average color
; otherwise darker color
sub r4, r3, c1 ; subtract falloff #2 from n.l
cmp_sat r0, r4 c4, r0 ; check if n.l is > zero
; if yes use bright color
; otherwise use what is there
; ps.1.2 equivalent with less precision
ps.1.2
def c0, 0.1f, 0.1f, 0.1f, 0.1f
def c1, 0.8f, 0.8f, 0.8f, 0.8f
def c2, 0.2f, 0.2f, 0.2f, 1.0f
def c3, 0.6f, 0.6f, 0.6f, 1.0f
def c4, 0.9f, 0.9f, 1.0f, 1.0f
texcoord t0 ; place normal vector in t0
texcoord t1 ; place light vector in t1
dp3 r1, t0, t1 ; n.l
sub t3, r1, c0 ; subtract falloff #1 from n.l
cmp_sat r0, t3, c3, c2 ; check if n.l is > zero
; if yes use average color
; otherwise darker color
sub t3, r1, c1 ; subtract falloff #2 from n.l
cmp_sat r0, t3, c4, r0 ; check if n.l is > zero
; if yes use bright color
; otherwise use what is there
The ps.1.2 version is not able to provide the same precision as the ps.1.4 version because of texcoord, which
clamps to [0..1]. texcrd does not clamp at all. It is able to handle values in the range of its source registers rn
[8..+8].
104 Part 1: Introduction to Shader Programming
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
dest
x x ps.1.1-ps.1.3
x ps.1.4
cnd src0 r0.a ps.1.1-ps.1.3
src1, src2 x x x x ps.1.1-ps.1.3
src0, src1, src2 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Conditionally chooses between src1 and src2 based on the comparison r0.a > 0.5, whereas ps.1.4 condition-
ally chooses between src1 and src2 based on the comparison src0 > 0.5 by comparing all channels of src0.
// Version 1.1 to 1.3
if (r0.a > 0.5)
dest = src1
else
dest = src2
// Version 1.4 compares each channel separately.
for each channel in src0
{
if (src0.channel > 0.5)
dest.channel = src1.channel
else
dest.channel = src2.channel
}
------------------------------------------
// Compares r0.a > 0.5
ps.1.1
def c0, -0.5, 0.5, 0, 0.6
def c1, 0, 0, 0, 0
def c2, 1, 1, 1, 1
mov r0, c0
cnd r1, r0.a, c1, c2 // r1 is assigned 0,0,0,0 based on the following:
// r0.a > 0.5, therefore r1.xyzw = c1.xyzw
-----------
// Compares all four components.
ps.1.4
// r1 holds -0.6, 0.5, 0, 0.6
// r2 holds 0, 0, 0, 0
// r3 holds 1, 1, 1, 1
texcrd r1, t1
texcrd r2, t2
texcrd r3, t3
cnd r0, r1, r2, r3 // r0 is assigned 1,1,1,0 based on the following:
// r0.x = r3.x because r1.x < 0.5
// r0.y = r3.y because r1.y = 0.5
// r0.z = r3.z because r1.z < 0.5
// r0.w = r2.w because r1.w > 0.5
See [Bendel] for an interesting usage of cnd to smooth fonts.
Fundamentals of Pixel Shaders 105
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
dest
x x ps.1.1-ps.1.3
x ps.1.4
dp3 x x x x ps.1.1-ps.1.3
src0, src1 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Calculates a three-component dot product. The scalar result is replicated to all four channels:
dest.w = (src0.x * src1.x) + (src0.y * src1.y) + (src0.z * src1.z);
dest.x = dest.y = dest.z = dest.w = the scalar result of dp3
It does not automatically clamp the result to [0..1]. This instruction executes in the vector pipeline so it can be
paired or co-issued with an instruction that executes in the alpha pipeline. (More on co-issuing below.)
dp3 r0.rgb, t0, v0
+mov r2.a, t0
In ps.1.1-ps.1.3, dp3 always writes out to .rgb. In ps.1.4, you are free to choose three channels of the four rgba
channels in any combination by masking the destination register.
----------------------------------------------
; Dot3 per-pixel specular lighting
; specular power comes from a lookup table
ps.1.4
; r0 holds normal map
; t1 holds light vector
; t2 holds half vector
; r2 holds a 2D texture (lookup table)
; r3 holds color map
texld r0, t0 ; normal map
texcrd r1.rgb, t1
texcrd r4.rgb, t2
dp3 r1.r, r1, r0_bx2 ; calculates u
dp3 r1.g, r4, r0_bx2 ; calculates v
phase
texld r3, t3
texld r2, r1 ; samples texel with u,v from t2 (lookup table)
mul r0, r2, r3
You will find the ps.1.1 equivalent as an example for the texm3x2tex instruction. See the RacorX8 and RacorX9
examples in Programming Pixel Shaders.
dest
x ps.1.2, ps.1.3
x ps.1.4
dp4 x x x x ps.1.2, ps.1.3
src0, src1 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Calculates a four-component dot product. It does not automatically clamp the result to [0..1]. This instruction
executes in the vector and alpha pipeline, so it cannot be co-issued. Unfortunately, CreatePixelShader() assumes
that this instruction consumes only one instruction slot, although it really consumes two. So the instruction count
in a pixel shader that uses dp4 must be checked manually.
106 Part 1: Introduction to Shader Programming
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
Additionally, in ps.1.2 and ps.1.3, the destination register for dp4 cannot be the same as any of the source reg-
isters. CreatePixelShader() will not catch a wrong usage.
A maximum of four dp4 commands are allowed in a single pixel shader.
dp4 is useful for handling 4x4 matrices or quaternions in a pixel shader. dp4 does not seem to be useful in con-
junction with most of the texture address instructions of ps.1.2 and ps.1.3 because these instructions support
only matrices with three columns.
dest
x x ps.1.1-ps.1.3
x ps.1.4
lrp x x x x ps.1.1-ps.1.3
src0, src1, src2 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Performs a linear interpolation based on the following formula:
dest = src2 + src0 * (src1 src2)
src0 determines the amount for the blend of src1 and src2.
-------------------------------------------
ps.1.1
def c0, 0.4, 0.2, 0.5, 1.0
tex t0
tex t3
lrp r0, c0, t3, t0 ; the texture values of t3 and t0 are
; interpolated depending on c0
-----------------------
ps.1.4
def c0, 0.4, 0.2, 0.5, 1.0
texld r0, t0
texld r3, t3
lrp r0, c0, r3, r0 ; the texture values of t3 and t0 are
; interpolated depending on c0
[Vlachos] shows how to programmatically interpolate with lrp between two textures based on their normal.
dest
x x ps.1.1-ps.1.3
x ps.1.4
mad x x x x ps.1.1-ps.1.3
src0, src1, src2 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Performs a multiply accumulate operation based on the following formula:
dest = src0 * src1 + src2
This might be useful, for example, for dark mapping with diffuse color.
-------------------------------------------
; The following examples show a modulation of a light map with a color map.
; This technique is often used to darken parts of a color map. In this case
; it is called dark mapping. Additionally, a diffuse color is added.
ps.1.1
tex t0 ; color map
tex t3 ; light map
mad r0, t3, t0, v0
Fundamentals of Pixel Shaders 107
Team LRN
Instruction Arguments Registers Version
v
n
c
n
t
n
r
n
; Dark Mapping + diffuse color
ps.1.4
texld r0, t0 ; color map
texld r3, t3 ; light map
mad r0, r3, r0, v0
dest
x x ps.1.1-ps.1.3
x ps.1.4
mov x x x x ps.1.1-ps.1.3
src x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Copies the content of the source to the destination register. Question every use of move because there might be
more suitable instructions.
dest
x x ps.1.1-ps.1.3
x ps.1.4
mul x x x x ps.1.1-ps.1.3
src0, src1 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Performs the following operation:
dest = src0 * src1
---------------------------------------
; The following examples show a modulation of a light map with a color map.
; This technique is often used to darken parts of a color map. In this case
; it is called dark mapping
ps.1.1
tex t0 ; color map
tex t3 ; light map
mul r0, t0, t3
; Dark Mapping
ps.1.4
texld r0, t0 ; color map
texld r3, t3 ; light map
mul r0, r0, r3
nop
Performs no operation in ps.1.1-ps.1.4.
dest
x x ps.1.1-ps.1.3
x ps.1.4
sub x x x x ps.1.1-ps.1.3
src0, src1 x x ps.1.4 phase 1
x x x ps.1.4 phase 2
Performs the following operation:
dest = src0 - src1
---------------------------------
ps.1.1
tex t0 ; color map #1
tex t1 ; color map #2
sub r0, t1, t0 ; subtract t0 from t1
108 Part 1: Introduction to Shader Programming
Team LRN
All arithmetic instructions can use the temporary registers rn. The rn registers are initially
unset and cannot be used as source operands until they are written. This requirement is
enforced independently per each channel of each rn register. In ps.1.4, the tn registers cannot
be used with any arithmetic instruction, so they are restricted on texture addressing instructions
(exception: texdepth).
Valid source registers for first phase arithmetic instructions are rn and cn. Valid source reg-
isters for second phase arithmetic instructions are rn, vn, and cn.
The comparison of ps.1.1-ps.1.3 to ps.1.4 shows only a few differences. The ps.1.4-only
instruction is bem. It substitutes the texbem and texbeml capabilities in an arithmetic operation
in ps.1.4. Furthermore, the cmd and cnd instructions are more powerful in ps.1.4. The scope of
the arithmetic instructions is much bigger in ps.1.4 than in ps.1.1-ps.1.3 because they are used
for all texture addressing and blending tasks in ps.1.4.
As with the vertex shader, the pixel shader arithmetic instructions provide no If statement,
but this functionality can be emulated with cmp or cnd.
All of the rn.a channels are marked as unset at the end of the first phase and cannot be
used as a source operand until written. As a result, the fourth channel of color data will be lost
during the phase transition. This problem can be partly solved by reordering the instructions.
For example, the following code snippet will lose the alpha value in r3.
ps.1.4
...
texld r3, t3
phase
...
mul r0, r2, r3
The next code snippet will not loose the alpha value:
ps.1.4
...
phase
texld r3, t3
...
mul r0, r2, r3
If no phase marker is present, then the entire shader is validated as being in the second phase.
All four channels of the shader result r0 must be written.
ps.1.1-ps.1.3 and ps.1.4 are limited in different ways regarding the maximum number of
source registers of the same type that can be read.
Read Port Limit
The read port limit gives you the maximum number of registers of the same register type that
can be used as a source register in a single instruction.
Register Name Version
ps.1.1 ps.1.2 ps.1.3 ps.1.4
cn 2 2 2 2
rn 2 2 2 3
tn 2 3 3 1
vn 2 2 2 2
Fundamentals of Pixel Shaders 109
Team LRN
The color registers have a read port limit of 2 in all versions. In the following code snippet,
mad uses v0 and v1 as a source register:
ps.1.1 // Version instruction
tex t0 // Declare texture
mad r0, v0, t0, v1
This example exposes a readport limit of 2. The following example exposes a readport limit of
1 because v0 is used twice:
ps.1.1 // Version instruction
tex t0 // Declare texture
mad r0, v0, t0, v0
The following pixel shader fails in ps.1.1:
ps.1.1
tex t0
tex t1
tex t2
mad r0, t0, t1, t2
It exceeds the readport limit of 2 for the texture registers. This shader wont fail with ps.1.2
and ps.1.3 because these versions have a readport limit of 3 for the tn registers. The functional
equivalent in ps.1.4 wont fail either:
ps.1.4
texld r0, t0
texld r1, t1
texld r2, t2
mad r0, r0, r1, r2
Another example for the usage of three temporary registers in ps.1.4 in the same instruction is
shown in the examples for the cmp and cnd instructions. In ps.1.4 the tn registers cannot be
used with arithmetic instructions and none of the texture address instructions can use more
than one tn register as a source; therefore, it is not possible to cross the readport limit of the tn
registers in ps.1.4.
There is no write port limit because every instruction has only one destination register.
Instruction Modifiers
How instructions can be modified is best illustrated in the following diagram:
110 Part 1: Introduction to Shader Programming
Team LRN
This diagram shows the parallel pipeline structure of the pixel shader ALU. The vector, or
color, pipeline handles the color values and the scalar, or alpha, pipeline handles the alpha
value of a 32-bit value. There are extensions for instructions that enable the programmer to
change the way data is read and/or written by the instruction. These are called swizzling,
source register modifiers, instruction modifiers, and destination register modifiers.
We will work through all instruction extensions shown in Figure 12 in the following
sections.
Swizzling/Using Selectors
In contrast to the more powerful swizzles that can be used in vertex shaders, the swizzling sup-
ported in pixel shader is only able to replicate a single channel of a source register to all
channels. This is done by so-called source register selectors.
Source Register Selectors Syntax ps.1.1 ps.1.2 ps.1.3 ps.1.4
Red replicate source.r x
Green replicate source.g x
Blue replicate source.b x x x x
Alpha replicate source.a x x x x
The .r and .g selectors are only available in ps.1.4. The following instruction replicates the red
channel to all channels of the source register:
r1.r ; r1.rgba = r1.r
Fundamentals of Pixel Shaders 111
Figure 12: Pixel shader ALU
Team LRN
As shown in Figure 12, selectors are applied first in the pixel shader ALU. They are only valid
on source registers of arithmetic instructions.
The .b replicate functionality is available in ps.1.1-ps.1.3 since the release of DirectX 8.1,
but this swizzle is only valid together with an alpha write mask in the destination register of an
instruction like this:
mul r0.a, r0, r1.b
This means that the .b source swizzle cannot be used with dp3 in ps.1.1-ps.1.3 because the only
valid write destination masks for dp3 are .rgb or .rgba (write masks will be presented later):
dp3 r0.a, r0, c0.b ; fails
The ability to replicate the blue channel to alpha opens the door to a bandwidth optimization
method, described in an NVIDIA OpenGL paper titled Alpha Test Tricks [Domin01]. A
pixel shader allows a dot product operation at the pixel level between two RGB vectors. There-
fore, one can set one of the vectors to be (1.0, 1.0, 1.0), turning the dot product into a
summation of the others vectors components:
(R, G, B) dot (1.0, 1.0, 1.0) = R + G + B
In the following code snippet, the pixel shader instruction dp3 calculates the sum of the RGB
color by replicating the scalar result into all four channels of r1.
ps.1.1
def c0, 1.0, 1.0, 1.0, 1.0
tex t0
dp3 r1, t0, c0
mov r0.a, r1.b
+mov r0.rgb, t0
An appropriate blending function might look like this:
dev->SetRenderState(D3DRS_ALPHAREF, (DWORD)0x00000001);
dev->SetRenderState(D3DRS_ALPHATESTENABLE, TRUE);
dev->SetRenderState(D3DRS_ALPHAFUNC, D3DCMP_GREATEREQUAL);
If the color data being rasterized is more opaque than the color at a given pixel
(D3DPCMPCAPS_GREATEREQUAL), the pixel is written. Otherwise, the rasterizer ignores
the pixel altogether, saving the processing required to, for example, blend the two colors.
An even more clever method to utilize the alpha test for fillrate optimization was shown by
Direct 3D ShaderX author Dean Calver on the public Microsoft DirectX discussion forum
[Calver]. He uses an alpha map to lerp three textures this way:
; Dean Calver
; 2 lerps from a combined alpha texture
ps.1.1
tex t0 ; combined alpha map
tex t1 ; texture0
tex t2 ; texture1
tex t3 ; texture2
mov r1.a, t0.b ; copy blue to r1.a
lrp r0, r1.a, t2, t1 ; lerp between t1 and t2
lrp r0, t0.a, t3, r0 ; lerp between result and t3
112 Part 1: Introduction to Shader Programming
Note: ps.1.1 does not support the .b replicate in DirectX 8.0.
Team LRN
The .a replicate is analogous to the D3DTA_ALPHAREPLICATE flag in the DX6/7
multitexturing unit.
In ps.1.4, there are specific source register selectors for texld and texcrd:
Source Register Selectors Description Syntax
.xyz/.rgb maps x, y, z to x, y, z ,z source.xyz/source.rgb
.xyw/.rga maps x, y, w to x, y, w, w source.xyw/source.rga
texld and texcrd are only able to use three channels in their source registers, so these selectors
provide the option of taking the third component from either the third or the fourth component
of the source register. Here are two examples of how to use these selectors:
texld r0, t1.rgb
...
texcrd r1.rgb, t1.rga
texcrd r4.rgb, t2.rgb
An overview of all possible source register selectors, modifiers, and destination write masks is
provided with the description of the texcrd and texld instructions above.
Source Register Modifiers
Source register modifiers are useful for adjusting the range of register data in preparation for
the instruction or to scale the value.
Modifier Syntax ps.1.1 ps.1.2 ps.1.3 ps.1.4
Bias r0_bias x x x x
Invert 1r0 x x x x
Negate r0 x x x x
Scale x2 r0_x2 x
Signed Scaling r0_bx2 x x x x
All modifiers can be used on arithmetic instructions. In ps.1.1, you can use the signed scale
modifier _bx2 on the source register of any texm3x2* and texm3x3* instruction. In ps.1.2 and
ps.1.3, it can be used on the source register of any texture address instruction.
Fundamentals of Pixel Shaders 113
Note: To move any channel to any other channel, use dp3 to replicate the channel across,
and then mask it out with the help of def instructions. The following pixel shader moves the
blue channel to the alpha channel:
; move the red to blue and output combined
ps.1.1
def c0, 1.f, 0.f, 0.f, 0.f ; select red channel
def c1, 0.f, 0.f, 1.f, 0.f ; mask for blue channel
def c2, 1.f, 1.f, 0.f, 1.f ; mask for all channels but blue
tex t0
dp3 r0, t0, c0 ; copy red to all channels
mul r0, r0, c1 ; mask so only blue is left
mad r0, t0, c2, r0 ; remove blue from original texture and
; add red shifted into blue
This trick was shown by Direct3D ShaderX author Dean Calver on the Microsoft DirectX
discussion forum [Calver].
Team LRN
bias subtracts 0.5 from all components. It allows the same operation as D3DTOP_ADD-
SIGNED in the DX6/7 multitexturing unit. It is used to change the range of data from 0 to 1 to
0.5 to 0.5. Applying bias to data outside this range may produce undefined results. Data out-
side this range might be saturated with the _sat instruction modifier to the range [0..1] before
being used with a biased source register (more on instruction modifiers in the next section).
A typical example for this modifier is detail mapping, shown in the add example.
Invert inverts (1 value) the value for each channel of the specified source register. The
following code snippet uses inversion to complement the source register r1:
mul r0, r0, 1-r1 ; multiply by (1.0 - r1)
Negate negates all source register components by using a subtract sign before a register. This
modifier is mutually exclusive with the invert modifier, so it cannot be applied to the same
register.
Scale with the _x2 modifier is only available in ps.1.4. It multiplies a value by two before
using it in the instruction, and it is mutually exclusive to the invert modifier.
Signed scaling with the _bx2 modifier is a combination of bias and scale, so it subtracts
0.5 from each channel and scales the result by 2. It remaps input data from unsigned to signed
values. As with bias, using data outside of the range 0 to 1 may produce undefined results. This
modifier is typically used in dp3 instructions. An example for this is presented with the
description of the dp3 instruction above. Signed scaling is mutually exclusive with the invert
modifier.
None of these modifiers change the content of the source register they are applied to.
These modifiers are applied only to the data read from the register, so the value stored in the
source register is not changed.
Modifiers and selectors may be combined freely. In the following example, r1 uses the
negative, bias, and signed scaling modifiers as well as a red selector:
-r1_bx2.r
With the help of the source register modifiers, per-pixel normalization is possible in the pixel
shader. Per-pixel normalization can be used instead of a cubemap:
; Assuming v0 contains the unnormalized biased & scaled vector ( just
; like a normal map), r0 will end up with a very close to normalized
; result.
; This trick is also useful to do 'detail normal maps' by adding
; a vector to a normal map, and then renormalizing it.
dp3 r0, v0_bx2, v0_bx2 ; r0 = N . N
mad r0, v0_bias, 1-r0, v0_bx2 ; (v0_bias * (1-r0)) + v0_bx2
; ((N - 0.5) * (1 - N.N)) + (N - 0.5) * 2
Normalization requires calculating 1/sqrt(N). This code snippet normalizes the normal vector
with a standard Newton-Raphson iteration for approximating a square root. This trick was
shown by Sim Dietrich in the Microsoft DirectX forum [Dietrich-DXDev].
In the first line, the normal vector N is biased and scaled by _bx2 and then multiplied with
itself via a dot product. The result is stored in r0. In the next line, the normal is again biased
and scaled and added to the inverse of r0. The result of this addition is multiplied with the
biased normal.
114 Part 1: Introduction to Shader Programming
Team LRN
There are additional modifiers specific to the texld and texcrd instructions in the ps.1.4
implementation. These modifiers provide projective divide functionality by dividing the x and
y values by either the z or w values; therefore, projective dependent reads are possible in
ps.1.4.
Source Register
Modifiers
Description Syntax
_dz/_db divide x, y components by z (pseudocode):
if (z==0)
x' = 1.0;
else
x' = x/z;
if (z==0)
y' = 1.0;
else y' = y/z;
// z' and w' are undefined
source_dz/source_db
_dw/_da divide x, y components by w (pseudocode):
if (w==0)
x' = 1.0;
else
x' = x/w;
if (w==0)
y' = 1.0;
else
y' = y/w;
// z and w are undefined
source_dw/source_da
These modifiers provide a functional replacement of the D3DTFF_PROJECTED flag of the
D3DTSS_TEXTURETRANSFORMFLAGS texture stage state flag in the pixel shader. A typi-
cal instruction would look like this:
texcrd r2.rg, t1_dw.xyw ; third channel unset
The modifier copies x/w and y/w from t1 into the first two channels of r2. The third and fourth
channels of r2 are uninitialized. Any previous data written to these channels will be lost. The
per-pixel perspective divide is useful, for example, for projective textures.
The restriction for the two texture addressing registers are:
texld texcrd
_dz/_db only valid with rn (second phase) Does not support_dz/_db
_dz/_db may be used no more that two times per
shader
The fourth channel result of texcrd is unset/
undefined in all cases.
The third channel is unset/undefined for the .xyw case.
The same .xyz or .xyw modifier must be applied to every read of an individual tn register within a
shader. If .xyw is being used on tn register read(s), this can be mixed with other read(s) of the same
tn register using .xyw.
Fundamentals of Pixel Shaders 115
Team LRN
Instruction Modifiers
After the swizzling of the source register channels and the modification of the values read out
from a source register with source register modifiers, the instruction starts executing. As
shown in Figure 12, now the instruction modifiers are applied. These are indicated as an
appendix to the instruction connected via an underscore. Instruction modifiers are used to
change the output of an instruction. They can multiply or divide a result or clamp a result to
[0..1]:
Source Register
Modifiers
Description Syntax ps.1.1-ps.1.3 ps.1.4
_x2 multiply by 2 instruction_x2 x x
_x4 multiply by 4 instruction_x4 x x
_x8 multiply by 8 instruction_x8 x
_d2 divide by 2 instruction_d2 x x
_d4 divide by 4 instruction_d4 x
_d8 divide by 8 instruction_d8 x
_sat clamp to [0..1] instruction_sat x x
Instruction modifiers can be used only on arithmetic instructions. The _x8, _d4, and _d8 modi-
fiers are new to the 1.4 shading model. _sat may be used alone or combined with one of the
other modifiers (i.e., mad_d8_sat).
Multiplier modifiers are useful for scaling a value. Note that any such scaling reduces
accuracy of results. The following examples scale the results by using _x2 or _x4:
ps.1.1
tex t0
tex t1
mul_x2 r0, t1, t0 ; (t1 * t0) * 2
...
mul_x2 r0, 1-t1, t0 ; t0 * inverse(t1) * 2
...
mad_x2 r0, t1, v0, t0 ; (t1 + ( v0 * t0)) * 2
...
mul_x4 r0, v0, t0 ; (v0 * t0) * 4
mul_x4 r1, v1, t1 ; (v1 * t1) * 4
add r0, r0, r1 ; (v0*t0 * 4)+(v1*t1 * 4)
The _x2 modifer does the same as a shift left in C/C++.
The _d2 modifer does the same as a right shift in C/C++. Here is a more complex
example:
; Here is an example for per-pixel area lighting
ps.1.1
def c1, 1.0, 1.0, 1.0, 1.0 ; sky color
def c2, 0.15, 0.15, 0.15, 1.0 ; ground color
def c5, 0.5, 0.5, 0.5, 1.0
tex t0 ; normal map
tex t1 ; base texture
dp3_d2 r0, v0_bx2, t0_bx2 ; v0.rgb is hemi axis in tangent space
; dot normal with hemi axis
116 Part 1: Introduction to Shader Programming
Team LRN
add r0, r0, c5 ; map into range
lrp r0, r0, c1, c2
mul r0, r0, t1 ; modulate base texture
This pixel shader biases the hemisphere axis in v0 and scales it by 2. The same is done to the
values of the normal map. dp3_bx2 divides the result through 2. The add instruction adds 0.5
to the vector of r0. lrp uses r0 as the proportion to linearly interpolate between sky color in c1
and ground color in c2.
The saturation modifer _sat clamps each component or channel of the result to the range
[0..1]. It is most often used to clamp dot products, like in the code snippet:
dp3_sat r0, t1_bx2, r0 ; N.H
dp3_sat r1.rgb, t1_bx2, r1 ; N.L
The result of the dot product operation of the normal vector with the half angle vector and the
result of the dot product operation of the normal and the light vector are saturated. That means
the values in r0 and r1.rgb are clamped to [0..1].
Destination Register Modifiers/Masking
A destination register modifer, or write mask, controls which channel in a register is updated,
so it only alters the value of the channel it is applied to.
Write masks are supported for arithmetic instructions only. The following destination write
masks are available for all arithmetic instructions:
Write Mask Syntax ps.1.1-ps.1.3 ps.1.4
color destination register.rgb x x
alpha destination register.a x x
red destination register.r x
green destination register.g x
blue destination register.b x
arbitrary x
In ps.1.1-ps.1.3, a pixel shader can only use the .rgb or .a write masks. The arbitrary write
mask in ps.1.4 allows any set of channels in the order r, g, b, a to be combined. It is possible to
choose, for example:
mov r0.ra, r1
If no destination write mask is specified, the destination write mask defaults to the .rgba case,
which updates all channels in the destination register. An alternate syntax for the r, g, b, a
channels is x, y, z, w.
As with the source register selectors and source register modifiers, the texld and texcrd
instructions have additional write masks and write mask rules. texcrd can write only to the .rgb
channels. It also supports a write mask that masks the first two channels with .rg or .xy. texld
uses all four channels of the destination register. There is no alternative write mask available.
The usage of write masks is shown in the following ps.1.4 pixel shader that handles diffuse
bump mapping with two spotlights (taken from the file 14_bumpspot.sha of the ATI Treasure
Chest example program on the companion CD):
Fundamentals of Pixel Shaders 117
Team LRN
ps.1.4
def c0, 1.0f, 1.0f, 1.0f, 1.0f ; Light 1 Color
def c1, 1.0f, -0.72f, 1.0f, 1.0f ; Light 1 Angle scale(x) and bias(Y)
def c2, 1.0f, 1.0f, 1.0f, 1.0f ; Light 2 Color
def c3, 0.25f, 0.03f, 1.0f, 1.0f ; Light 2 Angle scale(x) and bias(Y)
texcrd r0.rgb, t2 ; Spot light 1 direction
texcrd r1.rgb, t4 ; Spot light 2 direction
texld r2, t1 ; Light 1 to Point vector
texld r3, t3 ; Light 2 to Point vector
texcrd r4.rgb, t1 ; Light 1 space position for attenuation
texcrd r5.rgb, t3 ; Light 2 space position for attenuation
dp3_sat r4.x, r4, r4 ; Light 1 Distance^2
dp3_sat r5.x, r5, r5 ; Light 2 Distance^2
dp3_sat r4.y, r0, r2_bx2 ; Light 1 Angle from center of spotlight
dp3_sat r5.y, r1, r3_bx2 ; Light 2 Angle from center of spotlight
mad_x4 r4.y, r4.y, c1.x, c1.y ; Light 1 scale and bias for angle
mad_x4 r5.y, r5.y, c3.x, c3.y ; Light 2 scale and bias for angle
phase
texld r0, t0 ; Base Map
texld r1, t0 ; Normal Map
texld r4, r4 ; Distance/Angle lookup map
texld r5, r5 ; Distance/Angle lookup map
dp3_sat r2.rgb, r1_bx2, r2_bx2 ; *= (N.L1)
mul_x2 r2.rgb, r2, r4.r ; Attenuation from distance and angle
mad r2.rgb, r2, c0, c7 ; * Light Color + Ambient
dp3_sat r3.rgb, r1_bx2, r3_bx2 ; *= (N.L2)
mul_x2 r3.rgb, r3, r5.r ; Attenuation from distance and angle
mad r3.rgb, r3, c2, r2 ; * Light 2 Color + Light 1 Color + Ambient
mul r0.rgb, r3, r0 ; Modulate by base map
+mov r0.a, c0
There are four different write masks used throughout this shader. These are the .rgb, .x, .y, and
.a write masks. The first write mask used for the texcrd instructions are imperative. texld cant
handle write masks other than .rgba, which is the same as applying no explicit write mask. The
first four dp3 and the next two mad instructions write to the first component x of the r4 register
and the second component y of the r5 register. These write masks are not supported by
ps.1.1-ps.1.3. The usage of the .rgb write mask in the second phase of this shader is supported
by all implementations. The last two lines of this shader show the pairing of two instructions
using co-issue. We will discuss instruction pairing, or co-issuing, in the next section.
118 Part 1: Introduction to Shader Programming
Team LRN
Instruction Pairing
As shown in Figure 12, there are two pipelines: one for the color data and one for the alpha
data. Because of the parallel nature of these pipelines, the instructions that write color data and
instructions that write only alpha data can be paired. This helps reduce the fill-rate. Only arith-
metic instructions can be co-issued, with the exception of dp4. Pairing, or co-issuing, is
indicated by a plus sign (+) preceding the second instruction of the pair. The following shader
fragment shows three pairs of co-issued instructions:
dp3_sat r1.rgb, t1_bx2, r1
+mul r1.a, r0, r0
mul r1.rgb, r1, v0
+mul r1.a, r1, r1
add r1.rgb, r1, c7
+mul_d2 r1.a, r1, r1
First, a dp3 instruction is paired with a mul, then a mul instruction with a mul, and finally an
add instruction with a mul. Pairing happens in ps.1.1-ps.1.3 always with the help of a pair of
.rgb and .a write masks. In ps.1.4, a pairing of the .r, .g., or .b write masks together with an .a
masked destination register is possible. The line:
mul r1.a, r0, r0
only writes the alpha value of the result of the multiplication of r0 with itself into r1.a.
Co-issued instructions are considered a single entity, with the result from the first instruc-
tion unavailable until both instructions are finished and vice versa. The following shader will
fail shader validation:
ps.1.1
def c0, 1, 1, 1, 1
mov r0, v0
dp3 r1.rgb, r0, c0
+mov r0.a, r1.b
Fundamentals of Pixel Shaders 119
Note: GeForce3/4TI has a problem with co-issuing instructions in the eighth arithmetic
instruction slot. It stops showing the results when a co-issue happens in the eighth arithme-
tic instruction, whereas the REF works as expected. The following meaningless pixel shader
doesnt show a result with the driver version 28.32:
ps.1.1
tex t0 ; color map
tex t1 ; normal map
dp3 r0,t1_bx2,v1_bx2; ; dot(normal,half)
mul r1,r0,r0;
mul r0,r1,r1;
mul r1,r0,r0;
mul r0,r1,r1;
mul r1, r0, r0
mul r0, r1, r1
; assemble final color
mul r0.rgb,t0,r0
+mov r0.a, r1
If you use the REF or remove the last line, the results are as expected.
Team LRN
mov tries to read r1.b, but dp3 did not write to r1.b at that time. The shader will fail because
r1.b was not initialized before. This could be troublesome when r1.b is initialized before by
any instruction. The validator will not catch the bug, and the results will not look as expected.
Another restriction to be aware of is the maximum number of three different register types
that can be used across two co-issued instructions.
Assemble Pixel Shader
After checking for pixel shader support, setting the proper textures with SetTexture(), writing a
pixel shader, and setting the needed constant values, the pixel shader has to be assembled. This
is needed because Direct3D uses pixel shaders as byte-code.
Assembling the shader is helpful in finding bugs earlier in the development cycle.
At the time of publication, there are three different ways to compile a pixel shader:
n
Precompiled Shaders Use the pixel shader in a separate ASCII file, such as test.psh,
and compile it with a pixel shader assembler (Microsoft Pixel Shader Assembler or
NVASM) to produce a byte-code file which could be named test.pso. This way, not every
person will be able to read and modify your source.
n
On-the-Fly Compiled Shaders Write the pixel shader in a separate ASCII file or as a
char string into your *.cpp file and compile it on-the-fly while the application starts up
with the D3DXAssembleShader*() functions.
n
Shaders in Effect Files Write the pixel shader source in an effect file and open this
effect file when the application starts up. The pixel shader will be compiled by reading the
effect file with D3DXCreateEffectFromFile(). It is also possible to precompile an effect
file. This way, most of the handling of pixel shaders is simplified and handled by the
effects file functions.
The precompiled shader should be the preferred way of compiling shaders, since compilation
happens during development of the code (i.e., at the same time that the *.cpp files are
compiled).
Create a Pixel Shader
The CreatePixelShader() function is used to create and validate a pixel shader:
HRESULT CreatePixelShader(
CONST DWORD* pFunction,
DWORD* pHandle
);
This function takes the pointer to the pixel shader byte-code in pFunction and returns a handle
to the pixel shader in pHandle. A typical piece of source might look like this:
TCHAR Shad[255];
LPD3DXBUFFER pCode = NULL;
120 Part 1: Introduction to Shader Programming
Team LRN
DXUtil_FindMediaFile(Shad,_T("environment.psh"));
if(FAILED(D3DXAssembleShaderFromFile(Shad,0,NULL, &pCode,NULL) ) )
return E_FAIL;
if( FAILED(m_pd3dDevice->CreatePixelShader((DWORD*)pCode->GetBufferPointer(),
&m_dwPixShader) ) )
return E_FAIL;
DXUtil_FindMediaFile() helps you find the ASCII file. D3DXAssembleShaderFromFile()
compiles it before CreatePixelShader() returns the handle in m_dwPixShader.
The pointer pCode to the ID3DXBuffer interface is used to store the object code and return
a pointer to this object code with GetBufferPointer().
Set Pixel Shader
You set a pixel shader for a specific amount of vertices by using the SetPixelShader() function
before the DrawPrimitive*() call for these vertices:
m_pd3dDevice->SetPixelShader(m_dwPixShader);
The only parameter that has to be provided is the handle of the pixel shader created by
CreatePixelShader(). The pixel shader is executed for every pixel that is covered by the verti-
ces in the DrawPrimitive*() call.
Free Pixel Shader Resources
While the game shuts down or before a device change, the resources used by the pixel shader
have to be freed. This must be done by calling DeletePixelShader() with the pixel shader han-
dle like this:
if(m_dwPixShader)
m_pd3dDevice->DeletePixelShader(m_dwPixShader);
Summary
We have walked step by step through a vertex shader creation process. Lets summarize what
was shown so far:
n
First, the pixel shader support of end-user hardware has to be checked with the caps bit
PixelShaderVersion.
n
All textures have to be set with SetTexture(), like with the multitexturing unit.
n
The constant values for a pixel shader have to be set afterward in the application code with
the SetPixelShaderConstant() or in the pixel shader code with def.
n
There are texture address and arithmetic instructions. The scope of the texture address
instruction in ps.1.1-ps.1.3 enfolds loading and changing of texture data. The scope of the
texture address instructions in ps.1.4 enfolds only the loading of texture data. Changing of
texture data is done in ps.1.4 and ps.1.1-ps.1.3 with the arithmetic instructions.
Fundamentals of Pixel Shaders 121
Team LRN
n
After a pixel shader is written into a source file, it has to be compiled.
n
To get a handle to a pixel shader, it has to be created with CreatePixelShader().
n
To use a pixel shader, it has to be set with a call to SetPixelShader().
n
At the end of a pixel shader-driven application, the resources occupied by the pixel shader
must be freed with a call to DeletePixelShader().
What Happens Next?
In the next article, Programming Pixel Shaders, we will start with a basic pixel shader pro-
gram and discuss a few basic algorithms and how to implement them with pixel shaders.
References
[Bendel] Steffen Bendel, Hallo World Font Smoothing with Pixel Shaders, Direct3D
ShaderX.
[Beaudoin/Guardado] Philippe Beaudoin and Juan Guardado, A Non-Integer Power Function
on the Pixel Shader, Direct3D ShaderX.
[Brennan] Chris Brennan, Per-Pixel Fresnel Term, Direct3D ShaderX.
[Brennan2] Chris Brennan, Diffuse Cube Mapping, Direct3D ShaderX.
[Brennan3] Chris Brennan, Accurate Reflections and Refractions by Adjusting for Object
Distance, Direct3D ShaderX.
[Card/Mitchell] Drew Card and Jason L. Mitchell, Non-Photorealistic Rendering with Pixel
and Vertex Shaders, Direct3D ShaderX.
[Calver] Dean Calver, Microsoft DirectX discussion forum, mail from Fri, 21 Sep 2001
10:00:55, http://discuss.microsoft.com/SCRIPTS/WA-MSD.EXE?A2=ind0109C&L=
DIRECTXDEV&P=R25479.
[Dietrich01] Sim Dietrich, Guard Band Clipping in Direct3D, NVIDIA web site.
[Dietrich-DXDev] Sim Dietrich, Microsoft DirectX discussion forum, mail from Tue, 14 Aug
2001 20:36:02, http://discuss.microsoft.com/SCRIPTS/WA-MSD.EXE?A2=ind0108B&L
=DIRECTXDEV&P=R13431.
[Domin01] Sbastien Domin, Alpha Test Tricks, NVIDIA web site.
[Hart] Evan Hart, 3D Textures and Pixel Shaders, Direct3D ShaderX.
[Isidoro/Brennan] John Isidoro and Chris Brennan, Per-Pixel Strand-Based Anisotropic
Lighting, Direct3D ShaderX.
[Isidoro/Riguer] John Isidoro and Guennadi Riguer, Texture Perturbation Effects, Direct3D
ShaderX.
122 Part 1: Introduction to Shader Programming
Team LRN
[Kraus] Martin Kraus, Truly Volumetric Effects, Direct3D ShaderX.
[Mitchell] Jason L. Mitchell, Image Processing with Pixel Shaders in Direct3D, Direct3D
ShaderX.
[Moravnsky] dm Moravnszky, Bump Mapped BRDF Rendering, Direct3D ShaderX.
[Vlachos] Alex Vlachos, Blending Textures for Terrain, Direct3D ShaderX.
[Watt92] Alan Watt and Mark Watt, Advanced Animation and Rendering Techniques (Addison
Wesley, 1992).
[Weiskopf] Daniel Weiskopf and Matthias Hopf, Real-Time Simulation and Rendering of
Particle Flows, Direct3D ShaderX.
[Zecha] Oliver Zecha, Perlin Noise and Returning Results for Shader Programs, Direct3D
ShaderX.
Additional Resources
The best resource to accompany this article is the pixel shader assembler reference in the
Direct3D 8.1 documentation at DirectX Graphics->Reference->Pixel Shader Assembler
Reference.
A lot of valuable information on pixel shaders can be found at the NVIDIA (devel-
oper.nvidia.com) and ATI (http://www.ati.com/developer/) web sites, as well as other sites.
Author Article Published at
Philip Taylor Per-Pixel Lighting http://msdn.microsoft.com/directx
Miscellaneous Meltdown Power Point Slides http://www.microsoft.com/mscorp/corpevents/ \
meltdown2001/presentations.asp
Sim Dietrich Intro to Pixel Shading in DX8 NVIDIA web site (courseware)
Sim Dietrich DX8 Pixel Shader Details NVIDIA web site (courseware)
Sim Dietrich DX8 Pixel Shaders NVIDIA web site (courseware)
Sim Dietrich AGDC Per-Pixel Shading NVIDIA web site (courseware)
Jason L. Mitchell 1.4 Pixel Shaders ATI web site
Jason L. Mitchell Advanced Vertex and Pixel Shader
Techniques
ATI web site
Alex Vlachos Preparing Sushi How Hardware
Guys Write a 3D Graphics Engine
ATI web site
Rich Direct3D 8.0 pipeline http://www.xmission.com/~legalize/book/
Dave Salvator 3D Pipeline Part I-III http://www.extremetech.com and thier <3D Graph-
ics, Gaming & Audio> -> <Analysis & Tutorials>
Fundamentals of Pixel Shaders 123
Team LRN
Acknowledgments
Id like to recognize the following individuals who were involved in proofreading and improv-
ing this article (in alphabetical order): Jason L. Mitchell (ATI), dm Moravnszky (Swiss
Federal Institute of Technology), and Matthias Wloka (NVIDIA).
124 Part 1: Introduction to Shader Programming
Team LRN
Programming Pixel
Shaders
Wolfgang F. Engel
This article covers the basics of pixel shader programming. You will learn in the following
pages how to code a pixel shader-driven application by using the same lighting reflection mod-
els as in Programming Vertex Shaders. There is one big difference: This time, we will
calculate light reflection on a per-pixel basis, which leads to a bumpy impression of the
surface.
Most of the effects that modify the appearance of a surface in many games are calculated
on a per-vertex basis. This means that the lighting/shading calculations are done for each ver-
tex of a triangle, as opposed to each pixel that gets rendered or per-pixel. In some cases,
per-vertex lighting produces noticeable artifacts. Think of a large triangle with a light source
close to the surface. As long as the light is close to one of the vertices of the triangle, you can
see the diffuse and specular reflection on the triangle. When the light moves toward the center
of the triangle, the rendering gradually loses these lighting effects. In the worst case, the light
is directly in the middle of the triangle and there is almost no effect visible on the triangle,
when there should be a triangle with a bright spot in the middle.
That means that a smooth-looking object needs a lot of vertices or a high level of
tesselation; otherwise, the coarseness of the underlying geometry is visible.
Thats the reason the examples in Programming Vertex Shaders used a highly tessellated
Bzier patch class to make the lighting effect look smooth.
RacorX6
Our first example shows a directional light source with a diffuse reflection model.
Like all the previous and upcoming examples, the example on the following page is based
on the common files framework provided with the DirectX 8.1 SDK (read more in Pro-
gramming Vertex Shaders). Therefore, Alt+Enter switches between the windowed and
full-screen mode, F2 gives you a selection of the usable drivers, and Esc will shut down the
application. Additionally, the B key toggles diffuse lighting, the W and S keys allow zooming
in and out, the left mouse button rotates the globe, the right mouse button moves the direc-
tional light, and the P key toggles between ps.1.1 and ps.1.4, if available.
125
Team LRN
The following examples wont
run on graphics hardware that
doesnt support pixel shaders.
On hardware that is not capable
of running ps.1.4, the message
ps.1.4 Not Supported is
displayed.
As in the previous three arti-
cles, we will track the life cycle
of the pixel shader in the follow-
ing pages.
Check for Pixel Shader Support
To be able to fall back on a different pixel shader version or the multitexturing unit, the pixel
shader support of the graphics hardware is checked with the following piece of code in
ConfirmDevice():
if( pCaps->PixelShaderVersion < D3DPS_VERSION(1,4) )
m_bPS14Available = FALSE;
if( pCaps->PixelShaderVersion < D3DPS_VERSION(1,1) )
return E_FAIL;
The value in the D3DCAPS structure, which is filled by GetDeviceCaps(), corresponds to the
ps.x.x instruction at the beginning of the pixel shader source. Therefore, the macro
D3DPS_VERSION(1,4) checks for support of ps.1.4. If ps.1.4 is not available, the BOOL vari-
able m_bPS14Available is set to false and the user is prevented from switching between the
pixel shader versions.
Set Texture Operation Flags (with D3DTSS_* flags)
The pixel shader functionality replaces the D3DTSS_COLOROP and D3DTSS_ALPHAOP
operations and their associated arguments and modifiers, but the texture addressing, bump
environment, texture filtering, texture border color, mip map, and texture transform flags
(except in ps.1.4: D3DTFF_PROJECTED) are still valid. The texture coordinate index is still
valid together with the fixed-function T&L pipeline (read more in Fundamentals of Pixel
Shaders). Using the D3DTEXF_LINEAR flag for the D3DTSS_MINFILTER and
D3DTSS_MAGFILTER texture stage states indicates the usage of bilinear filtering of textures
(linear filtering in Direct3D terminology). This is done in RestoreDeviceObjects() for the
color map:
126 Part 1: Introduction to Shader Programming
Figure 1: Per-pixel diffuse reflection
Team LRN
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_MINFILTER, D3DTEXF_LINEAR);
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_MAGFILTER, D3DTEXF_LINEAR);
Set Texture (with SetTexture())
With proper pixel shader support and the texture stage states set, the textures are set with the
following calls in Render():
// diffuse lighting
if(m_bBump)
{
m_pd3dDevice->SetTexture(0,m_pColorTexture);
m_pd3dDevice->SetTexture(1,m_pNormalMap);
m_pd3dDevice->SetPixelShader(m_dwPixShaderDot3);
}
else
{
//no lighting, just base color texture
m_pd3dDevice->SetTexture(0,m_pColorTexture);
m_pd3dDevice->SetPixelShader(m_dwPixShader);
}
If the user switched the diffuse lighting off, only the color texture is set. Otherwise, the color
texture and the normal or bump map texture are set.
Define Constants (with SetPixelShaderConstant()/def)
We set four constant values in c33 in RestoreDeviceObjects(). These constants are used to bias
the values that should be sent via the vertex shader color output register to v0 of the pixel
shader:
// constant values
D3DXVECTOR4 half(0.5f,0.5f,0.5f,0.5f);
m_pd3dDevice->SetVertexShaderConstant(33, &half, 1);
In FrameMove(), the light direction, the concatenated world, view, and projection matrix, and
the world matrix is set in c12, c8, and c0:
// light direction
D3DXVec4Normalize(&m_LightDir,&m_LightDir);
m_pd3dDevice->SetVertexShaderConstant(12, &m_LightDir, 1 ); // light direction
// world * view * proj matrix
D3DXMATRIX matTemp;
D3DXMatrixTranspose(&matTemp,&(m_matWorld * m_matView * m_matProj));
m_pd3dDevice->SetVertexShaderConstant(8, &matTemp, 4);
// world matrix
D3DXMatrixTranspose(&matTemp,&m_matWorld);
m_pd3dDevice->SetVertexShaderConstant(0, &matTemp, 4);
Programming Pixel Shaders 127
Note: Switching on mip map filtering or trilinear filtering with the following
statement would be a good idea in production code:
m_pd3dDevice->SetTextureStageState( 0, D3DTSS_MIPFILTER,
D3DTEXF_LINEAR);
Team LRN
This example uses a directional light that can be moved with the right mouse button. The
mouse movement is tracked via WM_MOUSEMOVE in the Windows message procedure
function MsgProc().
Pixel Shader Instructions
Provided in all the upcoming examples is a very simple ps.1.1 pixel shader in diff.psh that only
displays the color map. Additionally, a ps.1.1 pixel shader in diffDot3.psh and a ps.1.4 pixel
shader in diffDot314.psh that are specific for the respective examples can be found in the
example directory on the companion CD.
ps.1.1
tex t0 //sample texture
mov r0,t0
ps.1.1
tex t0 ; color map
tex t1 ; normal map
dp3 r1, t1_bx2, v0_bx2 ; dot(normal,light)
mul r0,t0, r1 ; modulate against base color
ps.1.4
texld r0, t0 ; color map
texld r1, t0 ; normal map
dp3 r2, r1_bx2, v0_bx2 ; dot(normal, light)
mul r0, r0, r2
Unlike vertex shaders, the instructions in a pixel shader need a specific ordering, which is
called the instruction flow. This instruction flow differs between ps.1.1-ps.1.3 and ps.1.4.
ps.1.1-ps.1.3 allow four types of instructions, which must appear in this order:
n
Version instruction: ps.1.1
n
Constant instruction: def c0, 1.0, 1.0, 1.0, 1.0
n
Texture address instructions: tex*
n
Arithmetic instructions: mul, mad, dp3, etc.
Every pixel shader starts with the version instruction. It is used by the assembler to validate the
instructions that follow. After the version instruction, a constant definition could be placed
with def. Such a def instruction is translated into a SetPixelShaderConstant() call when
SetPixelShader() is executed.
The next group of instructions are the texture address instructions. They are used to load
data into the tn registers and additionally in ps.1.1-ps.1.3 to modify texture coordinates. Up to
four texture address instructions could be used in a ps.1.1-ps.1.3 pixel shader. The ps.1.1 pixel
shader uses the tex instruction to load a color map and a normal map.
Until ps.1.4, it was not possible to use tex* instructions after an arithmetic instruction.
Therefore, dp3 and mul had to come after the tex* instructions. There could be up to eight
arithmetic instructions in a ps.1.1 shader.
The ps.1.4 pixel shader instruction flow is a little bit more complex:
n
Version instruction: ps.1.4
n
Constant instruction: def c0, 1.0, 1.0, 1.0, 1.0
n
Texture address instructions: tex*
128 Part 1: Introduction to Shader Programming
Team LRN
n
Arithmetic instructions: mul, mad, dp3, etc.
n
Phase marker
n
Texture address instruction
n
Arithmetic instructions
A ps.1.4 pixel shader must start with the version instruction. Then as many def instructions as
needed may be placed into the pixel shader. This example doesnt use constants. There can be
up to six texture addressing instructions after the constants. The diffuse reflection model
shader only uses two texld instructions to load a color map and a normal map.
After the tex* instructions, up to eight arithmetic instructions can modify color, texture, or
vector data. This shader only uses two arithmetic instructions: dp3 and mul.
Up to this point, a ps.1.4 pixel shader has the same instruction flow as a ps.1.1-ps.1.3 pixel
shader. The phase marker allows it to double the number of texture addressing and arithmetic
instructions. It divides the pixel shader into two phases: phase 1 and phase 2. That means as of
ps.1.4, a second pass through the pixel shader hardware can be done. Adding up the number of
arithmetic and addressing instructions shown in the pixel shader instruction flow above leads
to 28 instructions. If no phase marker is specified, the default phase 2 allows up to 14 address-
ing and arithmetic instructions.
This pixel shader doesnt need more tex* or arithmetic instructions; therefore, a phase
marker is not used.
In this simple example, the main difference between the ps.1.1 and the ps.1.4 pixel shader
is the usage of the tex instructions in ps.1.1 to load the texture data into t0 and t1 and the usage
of the texld instruction in ps.1.4 to load the texture data into r0 and r1. Both instructions are
able to load four components of a texture. Valid registers for tex are tn registers only, whereas
texld accepts only rn registers as the destination registers, tn in both phases as source registers,
and rn only in the second phase.
The number of the temporary destination register of texld is the number of the texture
stage. The source register always holds the texture coordinates. If the source register is a tex-
ture coordinate register, the number of the tn register is the number of the texture coordinate
pair. For example, texld r0, t4 samples a texture from texture stage 0 with the help of the tex-
ture coordinate set 4. In this pixel shader, the texture with the color map is sampled from
texture stage 0, with the texture coor-
dinate set 0 and the texture with the
normal map is sampled from texture
stage 1 with the texture coordinate
set 1.
The dp3 instruction calculates
the diffuse reflection with a dot prod-
uct of the light and the normal vector
on a per-pixel basis. This instruction
replicates the result to all four chan-
nels. dp3 does not automatically
clamp the result to [0..1]. For exam-
ple, the following line of code needs
a _sat modifier:
Programming Pixel Shaders 129
Note: The dot product instruction executes in the vector
portion of the pixel pipeline; therefore, it can be co-issued
with an instruction that executes only in the alpha pipeline:
dp3 r0.rgb, t0, r0
+ mov r.a, t0, r0
Because of the parallel nature of these pipelines, the
instructions that write color data and the instructions that
write only alpha data can be paired. This helps reduce the
fill-rate. Co-issued instructions are considered a single
entity, since the first instruction is not available until both
instructions are finished and vice versa. Pairing happens in
ps.1.1-ps.1.3 always with the help of a pair of .rgb and .a
write masks. In ps.1.4, a pairing of the .r, .g., or .b write
masks together with an .a masked destination register is
possible. Therefore, for example, dp3.r is not possible in
ps.1.1-ps.1.3.
Team LRN
dp3_sat r0, t1_bx2, r0
The calculation of the dp3 instruction in both pixel shaders is similar to the calculation of the
diffuse reflection on a per-vertex basis in RacorX3 in Programming Vertex Shaders,
although the values provided to dp3 for the per-pixel reflection model are generated in differ-
ent ways.
Per-Pixel Lighting
Per-pixel lighting needs per-pixel data. High-resolution information about the normal vector is
stored in a two-dimensional array of three-dimensional vectors called a bump map or normal
map. Each vector in such a normal map represents the direction in which the normal vector
points. A normal map is typically constructed by extracting normal vectors from a height map
whose contents represent the height of a flat surface at each pixel (read more in
[Dietrich][Lengyel]). This is done in the following code snippet:
if(FAILED(D3DXCreateTextureFromFile(m_pd3dDevice, m_cColorMap,
&m_pColorTexture)))
return E_FAIL;
LPDIRECT3DTEXTURE8 pHeightMap = NULL;
if(FAILED(D3DXCreateTextureFromFile(m_pd3dDevice,m_cHeightMap,&pHeightMap)))
return E_FAIL;
D3DSURFACE_DESC desc;
pHeightMap->GetLevelDesc(0,&desc);
if(FAILED(D3DXCreateTexture(m_pd3dDevice, desc.Width, desc.Height, 0, 0,
D3DFMT_A8R8G8B8,D3DPOOL_MANAGED, &m_pNormalMap)))
return E_FAIL;
D3DXComputeNormalMap(m_pNormalMap,pHeightMap,NULL,0,D3DX_CHANNEL_RED,10);
SAFE_RELEASE(pHeightMap);
D3DXCreateTextureFromFile() reads in the height map file earthbump.bmp from the media
directory and provides a handle to this texture. A new and empty texture is built with D3DX-
CreateTexture() with the same width and height as the height map. D3DXComputeNormal-
Map() converts the height field to a normal map and stores this map in the texture map created
with D3DXCreateTexture().
The most important function is D3DXComputeNormalMap(), which was introduced in the
DirectX 8.1 SDK. It maps the (x,y,z) components of each normal to the (r,g,b) channels of the
output texture. Therefore, the height map is provided in the second parameter, and the normal
map is retrieved via the first parameter. This example doesnt use a paletized texture, so no pal-
ette has to be provided in the third parameter. The flags in the fourth entry field allow the user
to mirror or invert the normal map or store an occlusion term in its alpha channel. The last
parameter is the amplitude parameter, which multiplies the height map information. This
example scales the height map data by 10.
With the help of D3DXComputeNormalMap(), we create a map with normals on a
per-pixel basis, but there is still one problem left: The normals in the normal map were defined
130 Part 1: Introduction to Shader Programming
Note: The CreateFileBasedNormalMap() function of the DotProduct3
example from the DirectX 8.1 SDK shows how to convert a height map into
a normal map with source code.
Team LRN
in one space based on how the textures were applied to the geometry. This is called texture
space. In contrast, the light is defined in world space. The L dot N product between vectors in
two different spaces will lead to unintentional results.
There are two solutions for this problem: Generate the normal maps to always be defined
relative to world space, or move the light into texture space.
The second solution is the most common. On a very abstract level, it can be divided into
two steps:
n
A texture space coordinate system is established at each vertex.
n
The direction to light vector L is calculated at each vertex and transformed into the texture
space.
This is shown in the following two sections.
Establish a Texture Space Coordinate System at Each Vertex
The texture coordinates at each vertex form a coordinate sys-
tem with a U (tangent), V (binormal = u x v), and W (normal)
axis.
V and U are also called tangent vectors. These three vec-
tors form a rotation/shear matrix that transforms or maps from
world to texture space.
This example uses a sphere on which the textures are
mapped. At different vertices of this sphere, these vectors will
point in an entirely different direction. Therefore, you have to
calculate these vectors for every vertex.
To retrieve the U and V vectors, the partial derivatives of
U and V relative to X, Y, and Z in world space are calculated
(read more in [Dietrich][Lengyel]). This is done with the following piece of code in the
LoadXFile() function in RacorX.cpp for the whole mesh:
// compute the normals
hr = D3DXComputeNormals(pMeshSysMem,NULL);
if(FAILED(hr))
return E_FAIL;
// compute texture
hr = D3DXComputeTangent(pMeshSysMem,0,pMeshSysMem2,1,
D3DX_COMP_TANGENT_NONE,TRUE,NULL);
if(FAILED(hr))
return E_FAIL;
D3DXComputeNormals() computes normals for each vertex in a mesh. It uses in its first
parameter a pointer to the mesh. The second parameter can be used to specify the three neigh-
bors for each face in the created progressive mesh, which is not used here.
D3DXComputeTangent() (new in DirectX 8.1) is used to compute the tangent U vector
based on the texture coordinate gradients of the first texture coordinate set in the input mesh.
HRESULT D3DXComputeTangent(
LPD3DXMESH InMesh,
DWORD TexStage,
LPD3DXMESH OutMesh,
DWORD TexStageUVec,
DWORD TexStageVVec,
Programming Pixel Shaders 131
Figure 2: Texture space
Team LRN
DWORD Wrap,
DWORD* pAdjacency
);
The first parameter must be a pointer to an ID3DXMESH interface, representing the input
mesh, and the third parameter will return a mesh with the one or two vectors added as a texture
coodinate set. The texture coordinate set is specified in the TexStageUVec and TexStageVVec
parameters. The flag D3DX_COMP_TANGENT_NONE used in one of these parameters pre-
vents the generation of a vector. TexStage chooses the texture coordinate set in the input mesh
that will be used for the calculation of the tangent vectors.
Wrap wraps the vectors in the U and V direction if this value is set to 1, as in this example.
Otherwise, wrapping doesnt happen. With the last parameter, one can get a pointer to an array
of three DWORDs per face that specify the three neighbors for each face in the created mesh.
Both functions store the vectors in the mesh. Therefore, the mesh will transport the vec-
tors to the vertex shader via the vertex buffer as a texture coordinate set, consisting of the three
vector values.
The normal from D3DXComputeNormals() and the tangent from
D3DXComputeTangent() form the two axes necessary to build the per-vertex texture space
coordinate system to transform L.
Transforming L into Texture Space
After building up the U and W vectors, L is transformed into texture space in the following
lines in the vertex shader:
; Input Registers
; v0 - Position
; v3 - Normal
; v7 - Texture
; v8 - Tangent
...
m3x3 r5, v8, c0 ; generate tangent U
m3x3 r7, v3, c0 ; generate normal W
; Cross product
; generate binormal V
mul r0, r5.zxyw, r7.yzxw;
mad r6, r5.yzxw, r7.zxyw,r0;
;transform the light vector with U, V, W
dp3 r8.x, r5, c12
dp3 r8.y, r6, c12
dp3 r8.z, r7, c12
// light -> oD0
mad oD0.xyz, r8.xyz, c33, c33 ; multiply by a half to bias, then add half
...
The tangent U produced by D3DXComputeTangent() is delivered to the vertex shader via v8
and the normal W is provided in v3. To save bandwidth, the binormal is calculated via the
cross product of the tangent and the normal in the vertex shader. The transform happens with
the following formula:
132 Part 1: Introduction to Shader Programming
Team LRN
L.x' = U dot L
L.y' = V dot L
L.z' = W dot L
At the end of this code snippet, L is clipped by the output register oD0 to the range [0..1]. That
means any negative values are set to 0 and the positive values remain unchanged. To prevent
the cutting off of negative values in the range of [1..1] of v8 to [0..1], the values have to be
shifted into the [0..1] range. This is done by multiplying by 0.5 and adding 0.5 in the last
instruction of the vertex shader. This is not necessary for texture coordinates because they are
usually in the range [0..1].
To get the values back into the [1..1] range, the _bx2 source register modifiers subtract by
0.5 and multiply by 2 in the pixel shader:
ps.1.1
tex t0 ; color map
tex t1 ; normal map
dp3 r1, t1_bx2, v0_bx2 ; dot(normal,light)
mul r0,t0, r1 ; modulate against base color
ps.1.4
texld r0, t0 ; color map
texld r1, t0 ; normal map
dp3 r2, r1_bx2, v0_bx2 ; dot(normal, light)
mul r0, r0, r2
Pixel shader instructions can be modified by an instruction modifier, a destination register
modifier, a source register modifier, or a selector (swizzle modifier).
mov_IM dest_DM, src_SM || src_SEL
The instruction modifiers (IM) _x2, _x8, _d2, _d8, and _sat multiply or divide the result or, in
the case of the _sat modifier, saturate it to [0..1].
The destination register modifiers (DM) .rgb, .a, and in ps.1.4 .r, .g, and .b, control which
channel in a register is updated. They only alter the value of the channel to which they are
applied.
The source register modifiers (SM) _bias (0.5), 1 (invert), (negate), _x2 (scale), and
_bx2 (signed scaling) adjust the range of register data. Alternatively, a selector or swizzle mod-
ifier .r, .g, .b, and .a replicates a single channel of a source register to all channels (read more in
Fundamentals of Pixel Shaders). Both pixel shaders use _bx2 as a source register modifier
that doesnt change the value in the source register. For example, the value of v0 in the mov
instruction has not changed after the execution of the following instruction:
mul r1, v0_bx2
The values that are delivered to mul for the multiplication are modified by _bx2 but not the
content of the registers itself.
Both pixel shaders shown above produce a valid Lambertian reflection term with the dot
product between L, which is now in texture space, and a sample from the normal map (read
more in [Lengyel]).
The last two lines of the vertex shader store the texture coordinate values of the color tex-
ture in two texture coordinate output registers:
mov oT0.xy, v7.xy
mov oT1.xy, v7.xy
Programming Pixel Shaders 133
Team LRN
Sending the same texture coordinates via two texture output registers is redundant; therefore,
using only one of the output registers would be an improvement. The first idea that comes to
mind is setting D3DTSS_TEXCOORDINDEX as a texture stage state to use the texture coor-
dinates of the first texture additionally for the second texture. Unfortunately, this flag is only
valid for usage with the fixed-function T&L pipeline, but not for the usage with vertex shaders.
With ps.1.1, there is no way to use the texture coordinates of the first texture for the second
texture without sending them via oT1 to the pixel shader a second time. The texture coordi-
nates must be sent via one of the texture coordinate registers with the same number as the
texture stage.
In a ps.1.4 pixel shader, the programmer is able to choose the texture coordinates that
should be used to sample a texture in the texld instruction, as shown in the ps.1.4 pixel shader:
...
texld r1, t0 ; normal map
...
This way, the remaining texture coordinate output registers in the vertex shader can be used for
other data, such as vector data.
Assemble Pixel Shader
Similar to the examples used in Programming Vertex Shaders, the pixel shaders are assem-
bled with NVASM or, in the case of the ps.1.4 pixel shaders, with Microsofts psa.exe since
NVASM doesnt support ps.1.4. The integration of NVASM into the Visual C/C++ IDE is done
in the same way as for vertex shaders (read more in Programming Vertex Shaders).
Create Pixel Shader
The pixel shader binary files are opened and read, and the pixel shader is created by
CreatePSFromCompiledFile() in InitDeviceObjects():
CreatePSFromCompiledFile (m_pd3dDevice, "shaders/diffdot3.pso",
&m_dwPixShaderDot3);
CreatePSFromCompiledFile (m_pd3dDevice, "shaders/diff.pso", &m_dwPixShader);
CreatePSFromCompiledFile (m_pd3dDevice, "diffdot314.pso",
&m_dwPixShaderDot314);
The CreatePSFromCompiledFile() function is located at the end of RacorX.cpp:
//---------------------------------------------------------
// Name: CreatePSFromBinFile
// Desc: loads a binary *.pso file
// and creates a pixel shader
//---------------------------------------------------------
HRESULT CMyD3DApplication::CreatePSFromCompiledFile (IDirect3DDevice8*
m_pd3dDevice,
TCHAR* strPSPath,
DWORD* dwPS)
{
char szBuffer[128]; // debug output
DWORD* pdwPS; // pointer to address space of the calling process
HANDLE hFile, hMap; // handle file and handle mapped file
TCHAR tchTempVSPath[512]; // temporary file path
HRESULT hr; // error
134 Part 1: Introduction to Shader Programming
Team LRN
if( FAILED( hr = DXUtil_FindMediaFile( tchTempVSPath, strPSPath ) ) )
return D3DAPPERR_MEDIANOTFOUND;
hFile = CreateFile(tchTempVSPath, GENERIC_READ,0,0,OPEN_EXISTING,
FILE_ATTRIBUTE_NORMAL,0);
if(hFile != INVALID_HANDLE_VALUE)
if(GetFileSize(hFile,0) > 0)
hMap = CreateFileMapping(hFile,0,PAGE_READONLY,0,0,0);
else
{
CloseHandle(hFile);
return E_FAIL;
}
}
else
return E_FAIL;
// maps a view of a file into the address space of the calling process
pdwPS = (DWORD *)MapViewOfFile(hMap, FILE_MAP_READ, 0, 0, 0);
// Create the pixel shader
hr = m_pd3dDevice->CreatePixelShader(pdwPS, dwPS);
if ( FAILED(hr) )
{
OutputDebugString( "Failed to create Pixel Shader, errors:\n" );
D3DXGetErrorStringA(hr,szBuffer,sizeof(szBuffer));
OutputDebugString( szBuffer );
OutputDebugString( "\n" );
return hr;
}
UnmapViewOfFile(pdwPS);
CloseHandle(hMap);
CloseHandle(hFile);
return S_OK;
}
This function is nearly identical to the CreateVSFromCompiledFile() function shown in Pro-
gramming Vertex Shaders. Please consult the Create a Vertex Shader section in the
RacorX2 example section of that article for more information. The main difference is the usage
of CreatePixelShader() instead of CreateVertexShader().
CreatePixelShader() is used to create and validate a pixel shader. It takes a pointer to the
pixel shader byte-code in its first parameter and returns a handle in the second parameter.
OutputDebugString() shows the complete error message in the output debug window of the
Visual C/C++ IDE, and D3DXGetErrorStringA() interprets all Direct3D and Direct3DX
HRESULTS and returns an error message in szBuffer.
Programming Pixel Shaders 135
Note: CreatePixelShader() fails on hardware that does not support ps.1.4 pixel
shaders. You can track that in the debug window after pressing F5. Therefore, the
following examples indicate this with a warning message that says ps.1.4 Not Sup-
ported. To be able to see a ps.1.4 shader running on the Reference Rasterizer (REF),
the function CreatePixelShader() has to be called once again after switching to the REF.
This functionality is not supported by the examples used throughout Part One.
Team LRN
Set Pixel Shader
Depending on the choice of the user, three different pixel shaders might be set with
SetPixelShader() in Render():
//diffuse lighting.
if(m_bBump)
{
m_pd3dDevice->SetTexture(0,m_pColorTexture);
m_pd3dDevice->SetTexture(1,m_pNormalMap);
if (m_bPixelShader)
m_pd3dDevice->SetPixelShader(m_dwPixShaderDot314);
else
m_pd3dDevice->SetPixelShader(m_dwPixShaderDot3);
}
else
{
//no lighting, just base color texture
m_pd3dDevice->SetTexture(0,m_pColorTexture);
m_pd3dDevice->SetPixelShader(m_dwPixShader);
}
If diffuse lighting is switched on, the user may select with the P key between the ps.1.1 and
ps.1.4 pixel shader, if supported by hardware. If diffuse lighting is switched off, the simple
ps.1.1 pixel shader that only uses the color texture is set.
Free Pixel Shader Resources
All pixel shader resources are freed in DeleteDeviceObjects():
if(m_dwPixShaderDot3)
m_pd3dDevice->DeletePixelShader(m_dwPixShaderDot3);
if(m_dwPixShaderDot314)
m_pd3dDevice->DeletePixelShader(m_dwPixShaderDot314);
if(m_dwPixShader)
m_pd3dDevice->DeletePixelShader(m_dwPixShader);
Non-Shader Specific Code
I used source code published by SGI as open source some years ago to emulate a virtual
trackball. The file trackball.h holds an easy to use interface to the SGI code in
SGITrackball.cpp.
Summary
RacorX6 shows how the pixel shader is driven by the vertex shader. All the data that is needed
by the pixel shader is calculated and provided by the vertex shader. Calculating the U and W
vectors, which is necessary to build up a texture space coordinate system, is done one time,
while the *.x file is loaded with the D3DXComputeNormals() and D3DXComputeTangent()
functions. The V vector as the binormal is calculated for every vertex in the vertex shader
because it saves bandwidth. These three vectors that form a texture space coordinate system
are used to transform the light vector L to texture space. L is sent through the color output
136 Part 1: Introduction to Shader Programming
Team LRN
register oD0 to the pixel shader. The usage of a texture space coordinate system is the basis of
per-pixel lighting and will be used throughout the upcoming example.
RacorX7
The main improvement of RacorX7 over RacorX6 is the usage of a specular reflection model
instead of the diffuse reflection model used by the previous example.
RacorX7 sets the same textures and texture stage states as RacorX6, but with only one addi-
tional constant.
Define Constants (with SetPixelShaderConstant()/def)
Because this example uses a specular reflection model, an eye vector must be set into c24 in
FrameMove(), as already shown in the RacorX4 example:
// eye vector
m_pd3dDevice->SetVertexShaderConstant(24, &vEyePt, 1);
This eye vector helps to build up the V vector, which describes the location of the viewer. This
is shown in the next paragraph.
Pixel Shader Instructions
The pixel shader in this example calculates the specular reflection on the basis of a modified
Blinn-Phong reflection model that was already used in RacorX4 in Programming Vertex
Shaders.
The following diagram visualizes the vectors involved in the common specular reflection
models.
Programming Pixel Shaders 137
Figure 3: RacorX7 specular lighting
Team LRN
A model describing a specular reflec-
tion has to be aware of at least the
location of the light source L, the
location of the viewer V, and the ori-
entation of the surface normal N.
Additionally, a vector R that
describes the direction of the reflec-
tion might be useful. The half way
vector H was introduced by Jim Blinn
to bypass the expensive calculation of
the R vector in the original Phong
formula.
The original Blinn-Phong formula for the specular reflection looks like this:
kspecular (N dot H)
n
)
whereas:
H = (L + V)/2
The simplified formula that was used in RacorX4 and used here is
kspecular (N dot (L + V))
n
)
Compared to the Blinn-Phong formula, the specular reflection formula used by the examples in
Part 1 does not divide through L + V through 2. This is compensated by a higher specular
power value, which leads to good visual results. Although L + V is not equivalent to H, the
term half vector is used throughout the example source for this.
The calculation of the vector H works in a similar way as the calculation of the light vector
L in the previous example in the vertex shader:
vs.1.1
...
; position in world space
m4x4 r2, v0, c0
; get a vector toward the camera/eye
add r2, r2, c24
; normalize eye vector
dp3 r11.x, r2.xyz, r2.xyz
rsq r11.xyz, r11.x
mul r2.xyz, r2.xyz, r11.xyz
add r2.xyz, r2.xyz, c12 ; get half angle
; normalize half angle vector
dp3 r11.x, r2.xyz, r2.xyz
rsq r11.xyz, r11.x
mul r2.xyz, r2.xyz, r11.xyz
; transform the half angle vector into texture space
dp3 r8.x,r3,r2
dp3 r8.y,r4,r2
dp3 r8.z,r5,r2
; half vector -> oD1
mad oD1.xyz, r8.xyz, c33, c33 ; multiply by a half to bias, then add half
138 Part 1: Introduction to Shader Programming
Figure 4: Vectors for specular reflection
Team LRN
The first three code blocks calculate V. The next two code blocks generate the half vector H. H
is transformed into texture space with the three dp3 instructions like the light vector in the pre-
vious example, and it is stored biased in oD1 in the same way as the light vector was transfered
in the previous example.
Like in the previous example, this pixel shader is driven by the vertex shader:
ps.1.1
tex t0 ; color map
tex t1 ; normal map
dp3 r0,t1_bx2,v1_bx2; ; dot(normal,half)
mul r1,r0,r0; ; raise it to 32nd power
mul r0,r1,r1;
mul r1,r0,r0;
mul r0,r1,r1;
; assemble final color
mul r0,t0,r0
ps.1.4
texld r0, t0 ; color map
texld r1, t1 ; normal map
dp3 r2, r1_bx2, v1_bx2 ; dot(normal, half)
mul r3,r2,r2 ; raise it to 32nd power
mul r2,r3,r3
mul r3,r2,r2
mul r2,r3,r3
mul r0, r0, r2
The pixel shader gets H via v1 and N, as in the previous example, via a normal map. The spec-
ular power is caculated via four mul instructions. This was done in RacorX4 in Programming
Vertex Shaders via a lit instruction in the vertex shader.
Using the mul instructions to
perform the specular color leads
to visible banding artifacts as
shown in Figure 5.
These artifacts are a result of
the small precision of the pixel
shader registers and the precision
loss of the H vector by shifting
this vector from the range [1..1]
to [0..1] in the vertex shader and
back in the pixel shader, which is
necessary because the vertex
shader output registers deliver
only values in the range [0..1].
The only way to provide higher
precision is by delivering a tex-
ture map with the specular values
encoded, which is loaded in the
pixel shader. This is shown below
in the RacorX8 section.
Programming Pixel Shaders 139
Figure 5: RacorX7 visual artifacts
Team LRN
Summary
RacorX7 showed one way to implement a specular reflection model. The drawback to this
example is its low precision when it comes to specular power values. RacorX8 will show a
way to get a higher precision specular reflection.
RacorX8
The main difference between
RacorX8 and RacorX7 is the use
of a lookup table to store the
specular power values for the
specular reflection instead of
using a few mul instructions in
the pixel shader. The advantage
of using a lookup table is the
banding is reduced. This is due to
the higher value range of the
lookup table compared to the
solution with multiple mul
instructions.
The drawback of using a
lookup table as shown in this
example is the need for an addi-
tional texture stage.
Set Texture Operation Flags (with D3DTSS_* flags)
RacorX8 sets RestoreDeviceObjects() in these two additional texture stage states for the tex-
ture map that holds the specular power values (specular map):
// additionally for specular table
m_pd3dDevice->SetTextureStageState( 2, D3DTSS_ADDRESSU, D3DTADDRESS_CLAMP );
m_pd3dDevice->SetTextureStageState( 2, D3DTSS_ADDRESSV, D3DTADDRESS_CLAMP );
With D3DTADDRESS_CLAMP flag, the texture is applied once and then the color of the edge
pixel is smeared. Clamping sets all negative values to 0, whereas all positive values remain
unchanged. Without the clamping, a white ring around the earth would be visible.
Set Texture (with SetTexture())
This example sets the specular power lookup table or specular map with the handle
m_pLightMap16 in Render():
140 Part 1: Introduction to Shader Programming
Figure 6: RacorX8 specular lighting with specular power from
a lookup table
Team LRN
...
m_pd3dDevice->SetTexture(2, m_pLightMap16);
...
This lookup table is created as a texture map and filled with specular power values in the fol-
lowing lines:
// specular light lookup table
void LightEval(D3DXVECTOR4 *col, D3DXVECTOR2 *input,
D3DXVECTOR2 *sampSize, void *pfPower)
{
float fPower = (float) pow(input->y,*((float*)pfPower));
col->x = fPower;
col->y = fPower;
col->z = fPower;
col->w = input->x;
}
...
//
// create light texture
//
if (FAILED(D3DXCreateTexture(m_pd3dDevice, 256, 256, 0, 0,
D3DFMT_A8R8G8B8, D3DPOOL_MANAGED, &m_pLightMap16)))
return S_FALSE;
FLOAT fPower = 16;
if (FAILED(D3DXFillTexture(m_pLightMap16, LightEval, &fPower)))
return S_FALSE;
D3DXFillTexture() (new in DirectX 8.1) uses the user-provided function LightEval() in its sec-
ond parameter to fill each texel of each mip level of the lookup table texture that is returned in
the first parameter:
HRESULT D3DXFillTexture(
LPDIRECT3DTEXTURE8 pTexture,
LPD3DXFILL2D pFunction,
LPVOID pData
);
This function is useful to build all kind of procedural textures that might be used in the pixel
shader as a lookup table.
LightEval(), which is provided in pFunction, uses this declaration:
VOID (*LPD3DXFILL2D)(
D3DXVECTOR4* pOut,
D3DXVECTOR2* pTexCoord,
D3DXVECTOR2* pTexelSize,
LPVOID pData
);
The first parameter returns the result in a pointer to a vector. The second parameter gets a vec-
tor containing the coordinates of the texel currently being processed. In our case, this is a
pointer to a 2D vector named input. The third parameter is unused in LightEval() and might be
useful for providing the texel size. The fourth parameter is a pointer to user data. Here,
LightEval() gets a pointer to the pfPower variable. This value is transferred via the third
parameter of the D3DXFillTexture() function.
This example sets the same constants as the previous example, so we can proceed with the
pixel shader source.
Programming Pixel Shaders 141
Team LRN
Pixel Shader Instructions
The vertex shader that drives the pixel shader differs from the vertex shader in the previous
examples only in the last four lines:
; oT0 coordinates for normal map
; oT1 half angle
; oT2 half angle
; oT3 coordinates for color map
mov oT0.xy, v7.xy
mov oT1.xyz, r8
mov oT2.xyz, r8
mov oT3.xy, v7.xy
The texture coordinates for the normal map and the color map are stored in oT0 and oT3. The
half angle vector is stored as a texture coordinate in oT1 and oT2. This example uses a 3x2
table of exponents, stored in the specular map in texture stage 2.
The two pixel shaders TableSpec.psh and TableSpecps14.psh calculate the u and v position
and sample a texel from the specular map. After the color texture is sampled, the color value
and the value from the specular map is modulated:
ps.1.1
; t0 holds normal map
; (t1) holds row #1 of the 3x2 matrix
; (t2) holds row #2 of the 3x2 matrix
; t2 holds the lookup table
; t3 holds color map
tex t0 ; sample normal
texm3x2pad t1, t0_bx2 ; calculates u from first row
texm3x2tex t2, t0_bx2 ; calculates v from second row
; samples texel with u,v
tex t3 ; sample base color
mul r0,t2,t3 ; blend terms
; specular power from a lookup table
ps.1.4
; r0 holds normal map
; t1 holds the half vector
; r2 holds the lookup table
; r3 holds color map
texld r0, t0
texcrd r1.rgb, t1
dp3 r1.rg, r1, r0_bx2 ; (N dot H)
phase
texld r3, t0
texld r2, r1 ; samples specular value with u, v
mul r0, r2, r3
In the ps.1.1 pixel shader, texm3x2pad performs a three-component dot product between the
texture coordinate set corresponding to the destination register number and the data of the
source register and stores the result in the destination register. The texm3x2tex instruction cal-
culates the second row of a 3x2 matrix by performing a three-component dot product between
the texture coordinate set corresponding to the destination register number and the data of the
source register.
142 Part 1: Introduction to Shader Programming
Team LRN
texcrd in the ps.1.4 shader copies the texture coordinate set corresponding to the source
register into the destination register as color data. It clamps the texture coordinates in the desti-
nation register with a range of [MaxTextureRepeat, MaxTextureRepeat] (RADEON 8500:
2048) to the range of the source register [8, 8] (MaxPixelShaderValue). This clamp might
behave differently on different hardware. To be safe, provide data in the range of [8, 8].
A .rgb or .rg modifier should be provided to the destination register of texcrd because the
fourth channel of the destination register is unset/undefined in all cases.
The arithmetic instruction dp3 performs a three-component dot product between the two
source registers. Its result is stored in r and g of r1.
Both shaders perform a dependent read. A dependent read is a read from a texture map
using a texture coordinate that was calculated earlier in the pixel shader. The texm3x2pad/
texm3x2tex instruction pair calculate the texture coordinate that is later used to sample a texel
by the texm3x2tex instruction. In the ps.1.4 shader, the second texld instruction uses the tex-
ture coordinate that was calculated earlier in the two dp3 instructions.
It is interesting to note that the first texld instruction after the phase marker uses the same
texture coordinate pair as the normal map. This reusage of texture coordinates is only possible
in ps.1.4. It is also important to note that using the texm3x2pad/texm3x2tex pair to load a
value from a specular map is inefficient, because both instructions calculate the same value and
get the same half vector via two texture coordinate registers. Using only the texm3x2tex
instruction is not possible, because this instruction can only be used together with a
texm3x2pad instruction.
A more elegant solution comparable to the ps.1.4 shader is possible by using the texdp3tex
instruction together with a 1D specular map, but this instruction needs ps.1.2 or ps.1.3 capable
hardware.
Summary
This example improved the specular power precision by using a specular power lookup table.
The drawback of this technique is the usage of an additional texture, which may be overcome
by using the alpha channel of the normal map.
Programming Pixel Shaders 143
Note: Values from the output registers of the vertex shader are clamped to [0..1]; that
means the negative values are set to 0, while the positive values remain unchanged. To
bypass the problem of clamping, the data can be loaded in a texture into the pixel
shader directly.
In a ps.1.1-ps.1.3 pixel shader, the rn, tn, and cn registers can handle a range of
[1..1]. The color registers can only handle a range of [0..1]. To load data in the range
[1..1] via a texture in ps.1.1-ps.1.3 the tex tn instruction can be used. In ps.1.4, the
rn registers can handle a range of [8..8] and the tn registers can handle, in the case of
the RADEON 8500, a range of [2048..2048]. So data from a texture in the range of
[8..8] can be loaded via texcrd rn, tn; via texld rn, tn; or texld rn, rn (only phase 2) in a
ps.1.4 pixel shader.
Note: You cannot change the order of the t0-t3 registers in a ps.1.1-ps.1.3 pixel
shader. These registers must be arranged in this pixel shader version in numerical order.
For example, setting the color map in texture stage 1 in the above ps.1.1 shader wont
work. In ps.1.4, it is not necessary to order the r0-r5 or t0 and t1 registers in any way.
Team LRN
RacorX9
RacorX9 combines a diffuse reflection model with the specular reflection model. It is based on
RacorX6, RacorX7, and a further improved ps.1.4 shader, which gets the specular value from
the alpha value of the normal map.
The main point about this
example is that it handles both
reflection models with the help of
only two textures, so there is
some room left to use additional
textures for other tasks or for
more per-pixel lights (see
[Gosselin] for a ps.1.4 shader
with three lights with falloff).
This example uses two dif-
ferent vertex shaders: one for the
ps.1.1 and one for the ps.1.4
pixel shader. The vertex shader
that feeds the ps.1.1 pixel shader
named SpecDot3Pix.vsh stores
the half vector in oD1 and the
light vector in oD0. The two tex-
ture coordinates are stored in oT0
and oT1.
; half vector -> oD1 ps.1.1
mad oD1.xyz, r8, c33, c33 ; multiply by a half to bias, then add half
...
; light -> oD0
mad oD0.xyz, r8.xyz, c33, c33 ; multiply a half to bias, then add half
mov oT0.xy, v7.xy
mov oT1.xy, v7.xy
The only difference compared to the vertex shader used in RacorX7 is the storage of an addi-
tional light vector in oD0. The light vector is necessary to calculate the diffuse reflection in the
pixel shader in the same way as shown in RacorX6:
ps.1.1
tex t0 ; color map
tex t1 ; normal map
dp3 r0,t1_bx2,v1_bx2 ; dot(normal,half)
mul r1,r0,r0; ; raise it to 32nd power
mul r0,r1,r1;
mul r1,r0,r0;
mul r0,r1,r1;
dp3 r1, t1_bx2, v0_bx2 ; dot(normal,light)
mad r0, t0, r1, r0
Heightening the specular power value with four mul instructions in the pixel shader is a very
efficient method. The drawback of visible banding effects is reduced by combining the
144 Part 1: Introduction to Shader Programming
Figure 7: RacorX9 diffuse and specular lighting
Team LRN
specular reflection model with a diffuse reflection model. The light vector in the dp3 instruc-
tion is used in the same way as in RacorX7.
Compared to the vertex shader above, the second vertex shader named SpecDot314.psh
stores the half vector in oT2 and oT3 instead of oD1 and the texture coordinates, which are
used later in the pixel shader for both textures in oT0:
; half vector -> oT2/oT3 ps.1.4
mad oT2.xyz, r8, c33, c33 ; multiply by a half to bias, then add half
mad oT3.xyz, r8, c33, c33 ; multiply by a half to bias, then add half
...
; light -> oD0
mad oD0.xyz, r8.xyz, c33, c33 ; multiply a half to bias, then add half
mov oT0.xy, v7.xy
--------
; specular power from a lookup table
ps.1.4
; r1 holds normal map
; t0 holds texture coordinates for normal and color map
; t2 holds half angle vector
; r0 holds color map
texld r1, t0
texcrd r4.rgb, t2
dp3 r4.rg, r4_bx2, r1_bx2 ; (N dot H)
mov r2, r1 ; save normal map data to r2
phase
texld r0, t0
texld r1, r4 ; samples specular value from normal map with u,v
dp3 r3, r2_bx2, v0_bx2 ; dot(normal,light)
mad r0, r0, r3, r1.a
The new thing in this pixel shader is the storage of the specular power value in the alpha value
of the normal map. This lookup table is accessed like the lookup table in RacorX8. Therefore,
the normal map is sampled a second time in phase 2.
The lookup table is built up with the following piece of code:
//specular light lookup table
void LightEval(D3DXVECTOR4 *col,D3DXVECTOR2 *input,
D3DXVECTOR2 *sampSize,void *pfPower)
{
float fPower = (float) pow(input->y,*((float*)pfPower));
col->x = fPower;
col->y = fPower;
col->z = fPower;
col->w = input->x;
}
...
Programming Pixel Shaders 145
Note: If this pixel shader tried to use the v0 color register, the two dp3 instructions
would have to be moved in phase 2, but then the necessary dependent texture read
done in the second texld instruction in phase 2 would not be possible. Therefore, the
ps.1.4 shader wouldnt work with the half vector in v0 at all.
Team LRN
//
// create light texture
//
if (FAILED(D3DXCreateTexture(m_pd3dDevice, desc.Width, desc.Height, 0, 0,
D3DFMT_A8R8G8B8, D3DPOOL_MANAGED, &m_pLightMap16)))
return S_FALSE;
FLOAT fPower = 16;
if (FAILED(D3DXFillTexture(m_pLightMap16,LightEval,&fPower)))
return S_FALSE;
// copy specular power from m_pLightMap16 into the alpha
// channel of the normal map
D3DLOCKED_RECT d3dlr;
m_pNormalMap->LockRect( 0, &d3dlr, 0, 0 );
BYTE* pDst = (BYTE*)d3dlr.pBits;
D3DLOCKED_RECT d3dlr2;
m_pLightMap16->LockRect( 0, &d3dlr2, 0, 0 );
BYTE* pDst2 = (BYTE*)d3dlr2.pBits;
for( DWORD y = 0; y < desc.Height; y++ )
{
BYTE* pPixel = pDst;
BYTE* pPixel2 = pDst2;
for( DWORD x = 0; x < desc.Width; x++ )
{
*pPixel++;
*pPixel++;
*pPixel++;
*pPixel++ = *pPixel2++;
*pPixel2++;
*pPixel2++;
*pPixel2++;
}
pDst += d3dlr.Pitch;
pDst2 += d3dlr2.Pitch;
}
m_pNormalMap->UnlockRect(0);
m_pLightMap16->UnlockRect(0);
SAFE_RELEASE(m_pLightMap16);
A specular map in m_pLightMap16 is created, as already shown in RacorX8, with the help of
the LightEval() function. The values of this map are stored in the alpha values of the normal
map after retrieving a pointer to the memory of both maps. The specular map is then released.
This way, the ps.1.4 pixel shader only uses two texture stages, but there is a weak point in this
example: The ps.1.4 pixel shader is slow compared to the ps.1.1 pixel shader. The higher preci-
sion of the specular value has its price. Using the normal map with 2048x1024 pixels for
storage of the specular power slows down the graphics card. Using a smaller normal map
speeds up the frame rate substantially, but on the other hand, it reduces the precision of the nor-
mals. Using specular power in an additional texture would eat up one texture stage. Using an
equivalent to the ps.1.1 shader, which is shown below, wont allow us to use more than one or
two lights:
ps.1.4
texld r0, t0 ; color map
texld r1, t1 ; normal map
dp3 r2, r1_bx2, v1_bx2 ; dot(normal, half)
146 Part 1: Introduction to Shader Programming
Team LRN
mul r3,r2,r2 ; raise it to 32nd power
mul r2,r3,r3
mul r3,r2,r2
mul r2,r3,r3
dp3 r1, r1_bx2, v0_bx2 ; dot(normal,light)
mad r0, r0, r1, r2
The best way to improve the ps.1.1 and ps.1.4 pixel shaders in this example is to store the
specular power value in the alpha value of an additional smaller texture, which might add new
functionality to this example. This is shown by Kenneth L. Hurley [Hurley] for a ps.1.1 shader
and by Steffen Bendel [Bendel] and other Direct3D ShaderX authors for a ps.1.4 pixel shader.
Summary
This example has shown the usage of a combined diffuse and specular reflection model while
using only two texture stages. It also demonstrates the trade-off that has to be made by using
the specular power in the alpha value of a texture, but there is also an advantage: There are
more instruction slots left for using more than one per-pixel light. A rule of thumb might be
using up to three per-pixel lights in a scene to highlight the main objects and lighting the rest
of the scene with the help of per-vertex lights.
These examples might be improved by adding an attenuation factor calculated on a
per-vertex basis like in RacorX5 or by adding an attenuation map in one of the texture stages
[Dietrich00][Hart].
Further Reading
I recommend reading A Non-Integer Power Function on the Pixel Shader by Philippe
Beaudoin and Juan Guardado [Beaudoin/Guardado] to see a power function in a pixel shader
that calculates a high-precision specular power value. David Gosselin implements three lights
with a light falloff in [Gosselin]. Kenneth Hurley [Hurley] describes how to produce diffuse
and specular maps in an elegant way with Paint Shop Pro. Additionally, he describes a ps.1.1
pixel shader that uses a diffuse and specular texture map to produce better looking diffuse and
specular reflections. This pixel shader is an evolutionary step forward compared to the ps.1.1
shaders shown here. Steffen Bendel [Bendel] describes a way to produce a combined diffuse
and specular reflection model in a ps.1.4 pixel shader that uses a much higher precision and
leads to a better visual experience. Thats the reason he called it smooth lighting.
References
[Beaudoin/Guardado] Philippe Beaudoin and Juan Guardado, A Non-Integer Power Function
on the Pixel Shader, Direct3D ShaderX.
[Bendel] Steffen Bendel, Smooth Lighting with ps.1.4, Direct3D ShaderX.
[Dietrich] Sim Dietrich, Per-Pixel Lighting, NVIDIA developer web site.
[Dietrich00] Sim Dietrich, Attenuation Maps, in Game Programming Gems (Charles River
Media, 2000), pp. 543-548.
Programming Pixel Shaders 147
Team LRN
[Gosselin] David Gosselin, Character Animation with Direct3D Vertex Shaders, Direct3D
ShaderX.
[Hart] Evan Hart, 3DTextures and Pixel Shaders, Direct3D ShaderX.
[Hurley] Kenneth L. Hurley, Photorealistic Faces with Vertex and Pixel Shaders, Direct3D
ShaderX.
[Lengyel] Eric Lengyel, Mathematics for 3D Game Programming & Computer Graphics
(Charles River Media, Inc., 2002), pp. 150-157.
Acknowledgments
I would like to thank Philip Taylor for permission to use the earth textures from the Shader
workshop held at Meltdown 2001. Additionally, I would like to thank Jeffrey Kiel from
NVIDIA for proofreading this article.
Epilogue
Improving this introduction is a constant effort. Therefore, I appreciate any comments and sug-
gestions. Improved versions of this text will be published on http://www.direct3d.net and
http://www.gamedev.net.
148 Part 1: Introduction to Shader Programming
Team LRN
Basic Shader
Development with
Shader Studio
John Schwab
Introduction
Shader Studio is a tool that lets Microsoft Direct3D Version 8.1 developers experiment and
design vertex and pixel shaders. By presenting most of the internal operations of the Direct3D
API in a unique and intuitive form, you can spend more time designing and less time program-
ming. When using Shader Studio, the only programming you will have to do is in the shaders
themselves; geometry, transformations, textures, and states are all managed by Shader Studio.
What You Should Know
It is assumed that the user is very familiar with the architecture, structure, and especially Ver-
tex and Pixel Shader Instruction Set of the Direct3D 8.1 API. Though this is not a requirement,
it will allow you to understand the terminology and detailed aspects of Shader Studio much
easier. Beginners may want to review the introductory sections on writing shaders, and then
come back and explore the possibilities of shaders without having to do any Direct3D
programming.
Installation
Locate the Shader Studio directory on the companion CD and run Setup.exe to install.
Directories
Shader Studio/
Data/
Shaders/
Workspaces/
149
Team LRN
Coordinate Systems
Shader Studio uses a right-handed coordinate system; future versions will allow left and right
configuration. When you reset the translation and rotation, all transformation modes (see the
Transformations section later in this article) default to the identity matrix for orientation and
then the camera and light are translated to the coordinates. The following diagram shows how
the light and camera are aligned within the coordinate system.
Features
Shader Studio has the following features:
Full DirectX 8.1 API shader support: Every state, texture stage, and shader capability is
exposed and can be changed.
.x File mesh loading: Currently you can only load meshes using the .x file format. If your data
is in a different format, there are many conversion tools and exporters available. You can find
most of these tools in the extras directory of the DirectX 8.1 SDK before installation.
Automatic constants calculations: Data in the constant registers (for example, matrices,
materials, and lights) can be manually edited. As this data is changed in the Shader Studio
interface, it is automatically transferred to the appropriate constant registers.
Full state control: Every Direct3D render state, texture state, and all eight texture stages can
be directly manipulated in the Shader Studio interface.
All capable texture formats supported: Shader Studio can load all the texture formats that
Direct3D supports.
Alternate texture loading and rebinding to any stage: Any textures referenced in an .x file
will be loaded automatically. You have the option to load and apply your own textures to any
of the eight texture stages.
Shader declaration remapping: The declarator maps mesh format to the data entering the
vertex shader. You can control how the data is formatted and to which vertex the data goes.
150 Part 1: Introduction to Shader Programming
Team LRN
Dynamic vertex and pixel shader programming: The rendered view immediately displays
the results of successfully compiled code. Shader Studio also displays errors from your code as
you work.
Flexible shader pipeline: Shader Studio supports all combinations of shaders. You can use the
fixed-function pipeline, vertex shaders only, pixel shaders only, or both vertex shaders and
pixel shaders. The choice is yours.
Mesh material listing: Shader Studio provides a complete list of all material properties loaded
with the mesh.
Real-time rendering preview with interactive controls: Shader Studio provides a limited set
of camera, light, and object controls to change how you view the effect of your shaders. This
data is exposed to the shader through the automatic constants.
Vector browser (a great debugging tool): This is a display window that converts the color
values under the cursor to vector values very useful for pixel shaders.
Effect browsing: Workspace files can be associated with the application and quickly loaded.
Code generation: To quickly start your development, the automatic code generator can create
the minimal shader code for you based on the mesh format and declaration. The automatic
code generator is a great way to minimize errors and eliminate the effort of repeatedly entering
the same data.
Limitations
There are several items that you should be aware of before developing in Shader Studio.
n
The current version of Shader Studio only allows one vertex and pixel shader to be devel-
oped at the same time. Therefore, you should be as certain as possible that you can
accomplish your desired effect within this constraint.
n
You may be tempted to put some def instructions within your vertex shader code. Cur-
rently, def instructions will have no effect due to the way Direct3D handles them. Instead,
you must directly enter the data into the constant registers for it to be accessible by your
shaders. You can, however, enter def instructions in your code for use by your exported
shader. Pixel shaders dont have this problem.
n
The current version of Shader Studio implements single pass rendering only. Shader Stu-
dio does not (at the time of publication) handle multiple rendering pass techniques.
n
States in Shader Studio are considered a setting. This is because you can only set one at a
time and dont have programmatic control when they are set. You may have to manually
play with the states to get the effect youre looking for.
n
No alternate buffer formats are exposed. Shader Studio does not, for example, support
stencil operations (such as shadows) beyond what the states can do.
n
The list of all material properties loaded with the mesh is read only.
Basic Shader Development with Shader Studio 151
Team LRN
Default Light
Shader Studio has one directly controllable light source. You can access its properties through
the automatic constants, but there are several properties that are fixed internally. Below is a list
of fixed properties in the D3DLIGHT8 structure:
n
Type = spot light
n
All white colors
n
Maximum range
n
0 Falloff
n
Constant attenuation
n
180 inner and outer cone
Controls
The primary means of controlling the objects is by mouse manipulation.
Left Button: Rotate (polar orbit) around the current center. The x-axis controls azimuth, and
y-axis controls elevation.
Right Button: Translates (pans) the center relative to the camera view. With Shift down you
move along the depth axis (z-camera axis).
Middle Button: Translates (dollies) the object relative to the center and controls the orbit
distance.
Wheel: If you have a mouse wheel, rolling it will dolly the object in one unit increments.
Create a Vertex Shader
This section introduces the features of Shader Studio in the context of building a very simple
vertex shader. The step-by-step instructions illustrate how to use Shader Studio as intended:
what to do and what to avoid.
When you first start Shader Studio, you wont see anything; there is no geometry available
for Shader Studio to render. Therefore, your first step is to load a mesh.
152 Part 1: Introduction to Shader Programming
Note: Shader Studio only supports .x files. If necessary, the DirectX 8.1
SDK contains utilities to convert or export files from your favorite modeling
tool to the .x file format.
Team LRN
Step 1: Loading a Mesh
In this exercise, you are going to use the cornell.x mesh, available on the companion CD, as a
starting point. To load the file, go to the File menu and select the Load Mesh option. Change to
the Data subdirectory and select the file cornell.x. Press OK to load the file. You should now
see a room with a sphere and a cube suspended in midair. During the load process, the texture
for the walls is automatically loaded. If you are interested, you can inspect the materials and
textures in the Materials or Textures views.
Step 2: Transformations
Shader Studio allows you to easily change the position and orientation of the camera, the light,
and the object in the scene via the commands within the Transform menu (please refer to the
Transformations section under Shaders Reference later in this article for detailed informa-
tion about these controls and how they are used).
Begin with controlling the position and orientation of the camera. You must ensure that the
camera transform is selected; either press Alt+C or select the Transform/Camera menu item.
Use the left mouse button to orbit the scene, the right mouse button to pan, and Shift and the
right mouse button or rolling wheel to dolly.
If you lose your orientation and you cant find the objects in your scene, then you can use
the reset options from the Transform menu or press Alt+R to reset all positions and orienta-
tions to their default values. Experiment with the controls to become familiar with them, then
set the camera view a little to the right and above the center of the scene (see the following
figure).
Switch to the light with either Alt+L or select the Transform/Light menu item. The light is rep-
resented as a blue arrow. You can adjust the position and the orientation of the light using the
same controls you just used with the camera.
Basic Shader Development with Shader Studio 153
Note: You can use a mesh other than the suggested cornell.x mesh.
Team LRN
If you dont see the light, dolly it in with the middle mouse button or wheel. Now move it
to a location to the right of the camera. You should see a highlight off all the surfaces.
For the purposes of this tutorial, you do not need to alter the object transformation. For
your reference, altering the object transformation modifies the world transformation matrix.
Step 3: Starting a Shader
In Steps 1 and 2, Shader Studio is rendering the scene using the Direct3D fixed-function
graphics pipeline. Here is your chance to take control and create your first vertex shader.
To begin, select the View/Shaders menu item to open the Shaders dialog; you may have to
move or resize your windows to make items fit. The Shaders dialog may look complicated
now, but it will become very easy to use with some experience. The Shaders Reference
section later in this article contains detailed explanations of each screen element.
Begin by selecting the Vertex tab to enable vertex shader operations. Shader Studio pro-
tects you from trying to enable defective vertex shaders; you cannot enable vertex shaders at
this time because there is no vertex shader available or the current vertex shader contains
invalid code.
You can enter a vertex shader by typing in the Code box or by loading a vertex shader
from disk. If you are creating a new vertex shader, you can use the Default button to automati-
cally generate some initial code for you. The initial code is based on Shader Studios analysis
of the current declarations.
You can now enable the default vertex shader to see what effect it has on the scene. The
default code works, but the lighting model is very simple and most of you will want to go
much farther.
Step 4: Editing Code
You can extend the vertex shader by editing the code in the Code box. Any valid changes that
you make are immediately updated in the render view; errors are reported in the Status
window.
To continue with the tutorial, you can now add a directional light to the scene. Change the
third line of code from:
dp3 oD0, v3, c2
to:
dp3 oD0, v3, c21
The scene has disappeared! Actually, everything in the scene just turned black. The new code
instructs the vertex processor to take the dot product of the (negation of the) vector stored in
constant register 21 with the vector stored in v3 (the current vertex normal see the Declara-
tions section) and store the result in output register oD0 (the diffuse color for this vertex).
154 Part 1: Introduction to Shader Programming
Note: Shader Studio often refers to a specific constant register (for exam-
ple, constant register 21) as a constant (for example, constant 21 or c21).
Team LRN
However, constant 21 is currently undefined and contains the default value of zero. To set the
value of constant 21, locate row 21 in the ID column of the Constants box. Double-click the
line at 21 to bring up the Constant Properties dialog box so that you can change its properties
(please refer to the Constant Properties section for detailed information about this dialog box).
For now, just check the Auto Set box to enable the Auto Constant which defaults to Light
Direction for this constant register. Select OK and you will see that the scene is now lit by the
direction of the light source. Try moving the light around to verify that the lighting equations
are using only the direction of the light vector and not the position of the light.
Step 5: Adding Some Color
You have now added a directional light to your scene, but what happened to the materials that
are a part of the mesh? You can inspect the material definitions in the Materials view (select
the View/Materials menu item). Assuming that you loaded cornell.x as directed, you will see
eight materials listed along with their properties (see the Materials section for a detailed
explanation of the material properties). In many cases, you dont need to know the properties
of the materials, only that some materials have textures and some dont.
You can extend your vertex shader to include the effects of the diffuse material. Make the
following changes from the third line:
dp3 oD0, v3, -c21
to:
dp3 r0, v3, -c21 // calculate the contribution from the light
mul oD0, r0, c23 // calculate the interaction of the light and material
Once again, the scene turns black. As you have probably guessed, the material isnt set at con-
stant 23. Repeat the operations of step 4 to enable the constant. Go to ID 23 and enable the
Auto Set check box (you can see that it defaults to Diffuse Material), and then select OK. The
sphere and cube now display the proper colors.
Step 6: Saving Your Work
You have made a lot of progress at this point, so you should save what you have done to disk.
Since Shader Studio exposes all of Direct3Ds states, it is possible to set them to an unsup-
ported or invalid state if youre not careful, thereby producing unpredictable results. Therefore,
it is always wise to save your work often.
Basic Shader Development with Shader Studio 155
Note: You can add comments to your vertex shader code using the //
as illustrated above.
Note: The next step in the tutorial will overwrite all of your work, so you
are strongly advised to save your workspace.
Note: Auto Set places a comment beside constant 21 that documents the
purpose of the register.
Team LRN
To save everything to disk, select the File/Save Workspace menu item, select the location for
the saved workspace (for convenience, it is best to put your workspace in the Workspaces
directory), and give the workspace a meaningful name (use tutorial.sws for the current
workspace).
Step 7: Loading a Vertex Shader
You are now going to load a prebuilt vertex shader into your workspace. You can continue with
the current workspace or create a new workspace (select the File/New Workspace menu item).
If you create a new workspace, then you must reload cornell.x.
Open the Shaders view and select the Vertex tab. Press the Load button and select
Shaders/PhongPnt.vsh. Once the shader is loaded, your existing code is overwritten. Remem-
ber this when you are loading shaders!
A quick review of the code illustrates that it is significantly more complex than the code
used earlier in the tutorial. This vertex shader implements the functionality of almost all fea-
tures of the default fixed-function pipeline.
If you started with a new workspace, you will have a black screen again when you enable
this shader. From steps 1 through 5, you will remember that you need to set the constants to
have the vertex shader operate properly. Since this code is well documented, you can just
assign the proper constants to their values. You can start by enabling those constants identified
in the comments at the beginning of the code (C20, C23, C24, C25, C27); each of these will
default to the proper auto constant.
Constant registers C0 through C3 should already be set by default to the transform auto
constant. The final constant register is C4; you must manually assign the auto constant for this
register. In the Constant Properties (double-click the line at C4), check the Auto Set check box,
select World from the drop-down list, and close the dialog.
If you have the shader enabled, it is likely that a large portion of the scene is white. This is
due to the fact that the default specular power is 0, causing most colors to approach infinity. To
make the specular power more reasonable, set C26.w to 50 by manually entering this value in
the Constant Properties; you can also enter a comment like specular power while the dialog
is open.
Now, when you toggle between your vertex shader and the fixed-function pipeline (by
enabling and disabling the vertex shader via the check box), you should see little difference.
The difference between this vertex shader and the fixed-function pipeline is that every feature
(such as attenuation) isnt implemented in this vertex shader. Of course, you can extend this
shader and add support for attenuation; that is the power of programmable shaders in Shader
Studio.
156 Part 1: Introduction to Shader Programming
Note: Shader Studio uses absolute file references by default. Therefore, if
you change any file locations at a later date, you may lose the links to files
referenced in your workspace. You can change Shader Studio to use
relative file references by checking the Relative File References box when
saving your workspace, but a side effect of doing this is your referenced
files must be on the same drive as the workspace.
Team LRN
Step 8: Settings
You can do a lot with a shader, but you still need control over Direct3Ds state engine. Shader
Studio controls Direct3D states via the Settings dialog (select the Options/Settings menu item).
The details of all the settings are beyond the scope of this article (see the Direct3D documenta-
tion for further information). Simply put, all of Direct3Ds render and texture states are here
and can be easily changed.
To illustrate, you can add some fog to the current scene. Start by finding the
D3DRS_FOGENABLE state and open the branch to determine the current state. You should
see that it is FALSE; right-click on the state information and select TRUE to enable this state.
Nothing will happen yet because fog also requires several other states to be set. Next, find the
D3DRS_FOGTABLEMODE state and set it to D3DFOG_EXP. The scene will go black
because the two remaining states that affect the fog, the color and the density, default to unus-
able values and need to be changed.
First, you can get the scene back by changing the default density state from 1 to 0.05. Find
the D3DRS_FOGDENSITY state and right-click, enter the value of 0.05 via the keyboard, and
press Enter (or select elsewhere with the mouse). The scene is now visible, properly fogged.
You can also change the color of the fog. Locate the D3DRS_FOGCOLOR state and
right-click to change the value. Notice that this state is a hex value; all of Direct3Ds color val-
ues are in ARGB order, so if you want blue fog, you must enter 0x000000ff (or just FF) into
the field and press Enter. The scene is now rendered with blue fog.
At this point, you have covered three of the four possible ways to change a state; the only one
left is the multiple bit flags state. To see an example of this control mechanism, find the
D3DRS_COLORWRITEENABLE state. This state uses a bit mask to control which colors are
used. If you only want to write to the blue color component, then mask out the color compo-
nents that are not blue. Right-click and remove the check marks by Alpha, Red, and Green, and
then select somewhere off of the menu to confirm your changes. The scene is now rendered
with only the information in the blue channel. However, the blue fog may be interfering with
your perception, so try disabling the fog to verify that only the blue values are being written.
Before you continue, return the mask back to its default state (all color components enabled).
Step 9: Texture Control
This step illustrates how Shader Studio controls textures. You will now add multitexturing to
the scene.
Basic Shader Development with Shader Studio 157
Note: Fog is controlled by and implemented by the fixed-function pipeline,
not the vertex shader. Therefore, it is possible to mix a vertex shader with
some fixed functionality by setting the appropriate states.
Note: This step assumes that your hardware has the ability to perform
multitexturing. To verify your hardware capabilities, check Options/Statistics
and see how many max simultaneous textures your hardware can support.
If your hardware can only support one texture you may have to switch to the
(software) Reference Rasterizer (see Options/Configure).
Team LRN
To begin, open the View/Texture dialog. The dialog identifies the texture that was loaded with
the mesh and the Set by Materials option at stage #0 (for a detailed explanation of this dialog,
see the Textures section).
Load a new texture called Spot.bmp. Press the Load button and find the texture in the Data
directory, and then click OK to load it. Once the texture is loaded, you will see most of its
properties in the main area. If you want to use this new texture, you have to assign it to a tex-
ture stage.
To verify that Spot.bmp is loaded, manually set it at stage #0, thereby overriding Set by
Materials and forcing this texture to be applied to all geometry with texture coordinates.
Right-click on the Set by Materials option to display the list of the available options. Select the
new texture (Spot.bmp). Since the Textures dialog is a modal dialog, you will have to close it
(press OK) to see its changes. You should see the spot texture on all surfaces. Open the Tex-
tures dialog again and set stage #0 back to Set By Materials and assign the Spot.bmp texture to
stage #1. Then close the dialog to see the change. The scene returns to its original form. Since
you have set the new texture to a previously unused texture stage, you must make several
changes to see it.
Open the Shaders dialog and verify that the vertex shader is enabled. Begin by adding a
second set of texture coordinates so that the new texture can be mapped to the geometry. You
dont need to change the geometry to do this; simply use the same data as the base texture by
adding this line to the code in the vertex shader:
mov oT1, v7
You can place the new line right below this line:
mov oT0, v7
Not much will happen yet since all texturing is currently handled in the fixed-function pipeline
and we havent added a pixel shader yet.
Return to the Settings dialog, and this time select the Texture States tab; this will allow
you to change the texture states. To enable Stage 1, look at the Texture settings under stage #1.
The first state is D3D-
TSS_COLOROP (note that
stages begin at 0). Change the
state to D3DTOP_MOD-
ULATE, and the two textures
will be blended together. A
subtle property that you may
not notice is that the texture
filtering for the remaining
stages defaults to D3D-
TEXF_POINT. You might
want to change the
D3DTSS_MAGFILTER and
D3DTSS_MINFILTER to
D3DTEXF_LINEAR to make
the scene look much nicer.
158 Part 1: Introduction to Shader Programming
Team LRN
That concludes your introduction to the basic functionality of creating a vertex shader. The
next section explains how to create more advanced effects using a pixel shader. See the follow-
ing figure for the completed vertex shader.
Create a Pixel Shader
In this tutorial, you will expand on the previous tutorial by writing a pixel shader with func-
tionality equivalent to that of the fixed-function pipeline, in the sense that a pixel shader can
only handle a single material passed from the vertex shader. Even though a pixel shader fol-
lows a vertex shader in the pipeline, it is not necessarily dependent on it. In this tutorial,
however, the pixel shader does depend on the vertex shader, both to reduce complexity and to
illustrate Shader Studios ability to simultaneously develop a vertex shader and a pixel shader.
Step 1: Configuration
The pixel shader you are about to construct works with a single material. The last tutorial
enabled multi-texturing; this configuration must now be modified. Select the View/Textures
menu option. In the Textures dialog, configure the text stages as follows:
n
Set stage #0 to the base texture env2.bmp.
n
Set the remaining stages to NULL and close the Textures dialog.
The sphere and the cube are now textured because their meshes have texture coordinates and
stage #0 has been configured to apply the texture env2.bmp.
You can now begin writing your pixel shader.
Step 2: Writing a Pixel Shader
To begin, open the Shaders dialog and select the Pixel tab to configure for pixel shader devel-
opment. Begin with the simplest pixel shader possible; use the default code (press the Default
button) and enable the pixel shader. The scene is now rendered using diffuse lighting only.
The default pixel shader code is always implemented using version 1.0 of the pixel shader
specification. If you wish to develop using other versions of the pixel shader specification, you
must manually change this value. To enable the functionality equivalent to the fixed-function
pipeline, you must replace the default pixel shader code with the following:
ps.1.0
tex t0
mov r1,v1
mad r0,t0,v0,r1
Basic Shader Development with Shader Studio 159
Note: If you are continuing from the Vertex Shader tutorial, proceed to
Step 1. If you skipped the tutorial on vertex shaders, please load the
workspace vPhong.sws.
Team LRN
You can verify that there is no difference between the results of the fixed-function pipeline ren-
dering and the pixel shader rendering by enabling and disabling the pixel shader. That means
both approaches are processing the same information.
Pixel shaders support local definitions for constant values because the constant values are
compiled with the shader and, therefore, override any external constants defined in the Con-
stants view. Of course, you could enter the constant value in the Constants view and the pixel
shader would behave the same way. To illustrate, replace the current pixel shader with the fol-
lowing code:
ps.1.0
def c0, 1.0f, 1.0f, 1.0f, 0.0f
tex t0
mul r1,v1,c0
mad r0,t0,v0,r1
The local definition of c0 is used to define a new specular color that the pixel shader uses to
override the specular color from the vertex shader. This is just a trivial example and is probably
not practical, but you can manually adjust the constant to see the changes in the specular
highlight.
Shaders Reference
This section explains all the functionality of Shader Studios controls.
Workspace
Menu Layout
160 Part 1: Introduction to Shader Programming
Team LRN
Files
n
New Workspace: Resets all properties in Shader Studio to their default settings.
n
Open Workspace: Loads an existing workspace so you can continue your work where
you left off.
n
Save Workspace: Saves the current workspace. All states, shaders, and settings will be
saved in this file. Texture files and the mesh are referenced but not saved in the workspace
file itself. If you plan on sharing these workspaces, be sure to enable Relative File Refer-
encing before saving. All referenced file paths are then saved relative to the workspace
file.
n
Load Mesh: Loads a mesh into the workspace. If a mesh is already loaded, you will be
asked to confirm before the loaded mesh is overwritten.
Shaders Dialog
The principal interface for developing shader code in Shader Studio, this modeless dialog
allows you to modify your code and immediately see the results in the rendering window. This
dialog has many options (please refer to the following diagram); each option is discussed
below.
Along the right-hand side of the dialog are several buttons (Load, Save, Default, and Refresh
Constants) whose use depends on the identity of the currently active tab.
Basic Shader Development with Shader Studio 161
Note: Relative paths to referenced files must be local to the same logical drive as
the workspace. It is also efficient to have the files relatively close to the workspace.
Team LRN
n
Load and Save buttons: These are context-dependent Load and Save functions. For
example, if youre currently working on a vertex shader (the Vertex tab is active), you can
only load or save a vertex shade rwhen you click the Load button or the Save button.
n
Default button: The Default button takes the current mesh declarations and automatically
creates the minimal code required for a working shader. The default code is generated
using version 1.0 of the vertex shader and pixel shader specification.
n
Refresh Constants: Refresh Constants updates the values displayed in the constants pane.
Note that the values displayed are correct at the time that they are requested. However,
some values are used many times during the rendering process, so the values displayed
may not represent the final values.
n
Shader Code: The Shader Code portion of the dialog is where you place your shader
instructions. It is immediately compiled and is updated in the rendered view when
successful.
Vertex Shader
A functional vertex shader requires vertex shader assembly language code, configuration of the
constant registers (constants), and declarations. Shader Studio does not support the def instruc-
tion within vertex shader code. Instead, constants must be defined and set in the Constants
view. Auto constants allows you to quickly configure the constants that are required.
The only property not within this tab is the declaration, which is processed separately (see
Declarations).
Pixel Shader
For the most part, there is little difference between the Vertex Shader and Pixel Shader tabs.
Pixel shaders use the pixel shader instruction set and definitions (def instruction) that are sup-
ported in the code. Pixel shader constants can also be controlled in the Constants view, but auto
constants are not supported.
n
Constants View: The Constants view lists all available constants, their values, and an
optional comment. To access a constant, double-click to edit its properties. When editing
group constants, be sure to select the base item of a group to change it or you will set the
base ID of the group to the new base ID location and possibly overwrite a constant that
you manually configured.
n
Vertex Constants (Auto Constants): A highlighted constant is controlled by the auto con-
stants. See the Constant Properties section for more information.
n
Pixel Constants: Pixel constants are identical in operation to the vertex constants except
that auto constants are not supported.
n
Error Reporting: Errors reported by the shader compiler (as you write your shader) are
reported in the status bar at the bottom of the dialog.
162 Part 1: Introduction to Shader Programming
Team LRN
Declarations
Declarations map the internal vertex data format of a mesh to the registers used when the mesh
data enters the vertex shader. You can define any mapping that you require from the vertex
data to the registers.
To change the map for a register, right-click on the register, and then select the source and
data type. In most cases, the data type should remain constant or you may get unpredictable
results. Once you change the map for a register, you will most likely get an error status of
Vertex shader doesnt match declaration, unless you have already made the corresponding
changes to your vertex shader code. Similarly, if you change the registers in your vertex shader
code, you must also change the declarations.
If you are unsure of the internal mesh format, you can check Mesh Format in Options/Sta-
tistics to see what vertex data is present.
Constant Properties Dialog
The Constant Properties dialog allows you to change
the behavior of a constant; you can either set the val-
ues manually or use the auto constant sets. When you
enter this dialog (by double-clicking on a constant in
the Shaders dialog), two values are displayed: ID and
size. The size is determined by the intended purpose
of the constant. Some constants, such as matrices,
are composed of multiple constant registers. Other
constants, such as positions, require only a single
constant register. Constants that are not auto con-
stants are assumed to require a single constant
register.
n
Values: When the Auto Set check box is disabled, you can manually enter values in the X,
Y, Z, and W data fields. You also have the option to type in a comment for this constant.
n
Auto Constant: When the Auto Set check box is enabled, you can only select from the
predefined constant sets.
n
Constant Sets: The most common uses for the constant registers are predefined as the fol-
lowing constant sets.
P = Projection matrix (Calculated from FOV, Near, and Far Clip)
V = View matrix (Calculated from the camera, center, and dolly)
W = World matrix (Calculated from the object; can also be an auxiliary matrix)
T = Transformation matrix (concatenation of several matrices)
C = Center of orbit
D = Z dolly vector from C as (0,0,D)
R
X
= X rotation matrix
R
Y
= Y rotation matrix
L
P
= Light position
Basic Shader Development with Shader Studio 163
Note: Only stream 0 is supported, and you cant add or delete registers.
Team LRN
L
C
= Center of light orbit
L
D
= Z dolly vector from L
C
as (0,0,L
D
)
L
N
= Normal direction of light
E
P
= Eye (camera) position
E
N
= Eye (camera) normalized facing direction
Name Default ID Equation
Projection unassigned
Inverse projection unassigned
View unassigned
Inverse view 4
World unassigned
Inverse world 8
Transform 0
Inverse transform 12
World view unassigned
Inverse world view 16
Light position 20
Light direction 21
Ambient material 22 n/a
Diffuse material 23 n/a
Specular material 24 n/a
Emissive material 25 n/a
Material power 26 n/a
Eye position 27
Eye direction 28
Materials Dialog
The materials loaded with the current mesh are inspected
via this dialog.
The material number and its properties are dis-
played. The properties directly correspond to Direct3Ds
D3DMATERIAL8 structure.
164 Part 1: Introduction to Shader Programming
Note: All matrices are pre-transposed before being passed to the constant registers.
Note: You cannot change the properties of a material;
they are presented here for reference only.
Team LRN
Textures Dialog
Textures are managed via this dia-
log. You can load and remove any
number (and type) of texture(s)
that Direct3D supports, with the
exception of textures that are
loaded by your mesh.
The Textures dialog lists the
texture files and some parameters.
The Texture Stages box shows
which texture is applied to each
texture stage. You can edit the tex-
ture stages by right-clicking the
name and selecting the new tex-
ture. Texture stage #0 also allows
you to set it to the default material
with the Set by Material option.
Choosing this option will automat-
ically select the texture that was
loaded with the mesh.
Browser
The browser is a modeless dialog that determines the color
underneath the cursor position in the render view and con-
verts the color to a signed biased value vector for display.
One of the best uses for this is to visualize normal maps
for bump mapping. Here is an equation that converts the
color to a vector.
Transformations Dialog
The positions and orientations of the light, the camera, and the geometry are represented by
transformations. To adjust any transformation, choose the mode (the target for the transforma-
tion) and use the mouse to set the new position and orientation.
Basic Shader Development with Shader Studio 165
Team LRN
Object
The object mode controls the transformation information for an object. The following Auto
Constants (see the Constant Properties section) are affected:
n
World
n
Inverse world
n
Transform
n
Inverse transform
n
World view
n
Inverse World view
Light
The light mode controls the transformation information for the light. The following Auto Con-
stants are affected:
n
Light position
n
Light direction
Camera
The camera mode controls the transformation information for the scene viewpoint. The follow-
ing Auto Constants are affected:
n
Transform
n
Inverse transform
n
World view
n
Inverse world view
n
View
n
Inverse view
n
Eye position
n
Eye direction
There are three ways to reset the transformations to a known state.
n
Reset: Reset both the origin and orientation of the current transformation mode.
n
Reset Origin: Set the translation (position) to 0,0,0.
n
Reset Rotation: Set the orientation (rotation) to the identity matrix.
166 Part 1: Introduction to Shader Programming
Team LRN
Statistics Dialog
This dialog lists various internal statistics
about the current workspace.
n
Mesh: Displays how many faces and
vertices are in the mesh that is cur-
rently loaded.
n
Vertex Shader: Displays how many
vertices are processed by the vertex
shader.
n
Mesh Format: Displays the Flexible
Vertex Format (FVF) definition for the
mesh.
n
Capabilities: Displays the capabilities
of your video card relevant to Shader
Studio.
Settings Dialog
The Settings dialog presents all of Direct3Ds internal states
in a tree structure. Most of these internal states will not be
used when you have shaders enabled, but they are all used
when rendering is performed by the fixed-function pipeline.
The viewport settings (the Projection block) defines the
rendering window for both the fixed-function and shader
pipelines (see the Camera Projection section for further
information).
The settings are separated into render states and texture
states. The texture states are further organized by stage.
The names of the states directly correspond to their
respective property in Direct3D; this is also true for the states
that you set. For both tabs, the mechanism for changing
states is the same and will follow one of the patterns illus-
trated by the following state methods. In all cases, select for
edit with the right mouse button or by pressing the Enter key
when the state is highlighted.
State Methods
n
Single State: This method allows selection of a single state. You are pre-
sented with a pop-up menu of possible options. Once you have made the
selection, select anywhere off the pop-up menu to set the selection.
n
Bit Flags: Bit Flags allows selection of multiple states. You are pre-
sented with a pop-up menu of possible options with a check mark
Basic Shader Development with Shader Studio 167
Team LRN
beside states that are enabled. Once you have made the selection(s),
select anywhere off the pop-up menu to set the selection(s).
n
Hex Value: This allows entry of colors and numbers as a 32-bit value
as an unsigned hexadecimal number. It is not necessary to type the
0X before a number.
n
Float Value: This method allows entry of floating point numbers. Be
careful when entering these numbers; make sure they are within the
states respective capabilities.
Camera Projection
n
Adjusts the FOV, near, and far clipping planes.
n
FOV: Field Of View in degrees, can be from 0-180.
n
Far must be greater than near and can be any positive number.
Configure Dialog
The Configure dialog allows you to set various internal proper-
ties for Shader Studio.
The Rendering Device selects between hardware accelera-
tion (HAL) and the internal (Direct3D) reference software
rasterizer (REF). Be careful when changing rendering devices
most drivers dont implement all the features of the software
rasterizer, so you could end up with errors such as invalid states or unsupported capabilities.
You should always save your work before changing the rendering device.
168 Part 1: Introduction to Shader Programming
Team LRN
Assets
The following files are available on the companion CD:
Vertex Shaders
n
PhongPnt.vsh
Pixel Shaders
n
PhongPnt.psh
Workspaces
n
vPhong.sws
n
pPhong.sws
n
Tutorial.sws
Meshes
n
Cornell.x
n
Light.x
n
teapot.x
Textures
n
Env2.bmp
n
Spot.bmp
Further Info
To get the latest version and any other up-to-date information, please go to the Shader Studio
home page at http://www.shaderstudio.com/.
Basic Shader Development with Shader Studio 169
Team LRN
170
Vertex Decompression in a Shader
by Dean Calver
Shadow Volume Extrusion Using a Vertex Shader
by Chris Brennan
Character Animation with Direct3D Vertex Shaders
by David Gosselin
Lighting a Single-Surface Object
by Greg James
Team LRN
171
Optimizing Software Vertex Shaders
by Kim Pallister
Compendium of Vertex Shader Tricks
by Scott Le Grand
Perlin Noise and Returning Results from Shader Programs
by Steven Riddle and Oliver C. Zecha
Team LRN
Vertex Decompression
in a Shader
Dean Calver
Introduction
Vertex compression can help save two valuable resources, memory and bandwidth. As you can
never have enough of either, compression is almost always a good thing. The main reasons to
not use compression are if the processing time to decompress is too high or if the quality is
noticeably lower than the original data (for lossy compression schemes). The methods
described here are designed to avoid both problems with reasonable compression ratios.
Graphics literature has a wide range of vertex compression schemes, but most of these
require a complex decompression step that cannot be executed inside a vertex shader.
Vertex Compression Overview
Our aim is to get the entire decompression step in the vertex shader itself. This rules out many
of the standard compressors that require either large compression tables or access to other ver-
tices (predictor methods). The amount of data we can use for the decompression algorithm
(almost all vertex shader constant space may already be used, i.e., skinning matrices, etc.) is
very limited. We generally have a few per-object constants to upload into vertex constant
memory.
A number of techniques can be used inside a vertex shader. These include:
n
Quantization The simplest form of compression; we simply lose the least significant
bits.
172
Figure 1: Vertex compression
block diagram
Team LRN
n
Scaled offset Actually a special case of linear equations method, but limiting the equa-
tion to a scaled offset from a point.
n
Compression transform A method that calculates linear equations to transform by
where the variables (the vertex values) need less precision.
n
Multiple compression transforms Each vertex selects from a palette of compression
transforms.
n
Sliding compression transform Extends the precision of the compression transforms by
putting spare padding bytes to good use.
n
Displacement mapping and vector fields Only for special kinds of data, but by storing
geometry like a texture, we can save space.
All these compression methods rely on the same basic technique storing the vertex compo-
nents in fewer bits. Floating-point is designed with a huge dynamic range; usually we dont
need much so we can save space. DirectX helps us by having a number of non-floating-point
data types that are converted to floating-point before entry into the vertex shader.
Vertex Shader Data Types
As of DirectX 8.1, there are four different data types for us to work with (excluding the float
data types):
n
D3DVSDT_UBYTE4 Four unsigned 8-bit integers that map in the vertex shader in the
range 0.0 to 255.0.
n
D3DVSDT_D3DCOLOR Four unsigned 8-bit integers that map in the vertex shader in
the range 0.0 to 1.0. Also, the mapping of each component into the vertex shader register
is different from D3DVSDT_UBYTE4.
n
D3DVSDT_SHORT2 Two signed 16-bit integers that map in the vertex shader in the
range 32767.0 to +32768.0.
n
D3DVSDT_SHORT4 Four signed 16-bit integers that map in the vertex shader in the
range 32767.0 to +32768.0.
Some video cards dont support D3DVSDT_UBYTE4, but this isnt a problem in practice
since D3DVSDT_D3DCOLOR can be used as a replacement with a few minor modifications.
The two changes needed are accounting for the change in range (the vertex shader code has to
multiply the incoming values by 255.0) and swapping the vertex stream components so they
appear in the correct places when used. This can be achieved by either changing the input data
(i.e., swapping the data in the vertex buffer and multiplying a transformation matrix by 255.0)
or adding the following vertex shader code:
; v1 = D3DVSDT_D3DCOLOR4 to be used a D3DVSDT_UBYTE4
; c0 = <255.0, 255.0, 255.0, 255.0>
mul r0.zyxw, v1, c0 ; multiply color4 by 255 and swizzle
If we are going to use these types instead of standard float vector, then we are going to have to
insert dummy pad values for odd vector sizes. The pad values show up a lot with vertex data
because most components (positions, normals, etc.) are 3D vectors. Even with pad spaces, the
new types are much smaller than floats, and often we can use these pad values to store other
data.
Vertex Decompression in a Shader 173
Team LRN
As different vertex components have different numerical ranges and requirements, it is
quite common to have varying compression methods in a single stream. We set up our stream
declarations with the new types and fill vertex buffers with the packed values.
Compressed Vertex Stream Declaration Example
A lit vertex with a single uv channel format before compression:
D3DVSD_STREAM(0),
D3DVSD_REG( 0, D3DVSDT_FLOAT3), // position
D3DVSD_REG( 1, D3DVSDT_FLOAT3), // normal
D3DVSD_REG( 2, D3DVSDT_FLOAT2), // texture coordinate
D3DVSD_END()
After compression, the position is stored in three signed 16-bit integers, the normal is stored in
three unsigned 8-bit integers, and the texture coordinate is stored in two signed 16-bit integers.
We have a signed 16-bit integer and an unsigned 8-bit integer that are unused (padding). The
original size is 32 bytes, and the compressed size is 16 bytes, so we have a 50% savings.
D3DVSD_STREAM(0),
D3DVSD_REG( 0, D3DVSDT_SHORT4), // position
D3DVSD_REG( 1, D3DVSDT_UBYTE4), // normal
D3DVSD_REG( 2, D3DVSDT_SHORT2), // texture coordinate
D3DVSD_END()
Basic Compression
Several simple forms of vertex compression are discussed in this section.
Quantization
The simplest form of vertex compression is to store fewer bits; as there is no floating-point for-
mat of less than 32 bits, this also implies a shift to integer or fixed-point data representation.
This has problems with almost all data sets but is a reasonable solution for data with a very
limited dynamic range. The obvious candidates among standard vertex components are nor-
malized values (normals and tangents are usually normalized for lighting calculations).
Compression
I choose a policy of clamping to a minimum and maximum range of the data type, but if this
happens, we shouldnt use quantization as a compression method.
We also have to choose how many bits to use for the integer part and how many for the
fractional part. This is very data-sensitive because there is no reasonable default with so few
bits in total.
fracScale = number of bits for fractional values
Quantise( originalVal )
{
return ClampAndFloatToInteger ( originalVal * fracScale );
}
174 Part 2: Vertex Shader Tricks
Team LRN
Decompression
Decompression just reverses the compression process, taking the compressed values and
restoring any fractional part.
Constants
{
fracScale = number of bits for fractional vales
}
Decompress( compressedVal )
{
return IntegerToFloat( compressedVal ) / fracScale
}
Practical Example
Normalized normals or tangents are usually stored as three floats (D3DVSDT_FLOAT3), but
this precision is not needed. Each component has a range of 1.0 to 1.0 (due to normalization),
ideal for quantization. We dont need any integer components, so we can devote all bits to the
fractional scale and a sign bit. Biasing the floats makes them fit nicely into an unsigned byte.
For most normals, 8 bits will be enough (thats 16.7 million values over the surface of a
unit sphere). Using D3DVSDT_UBYTE4, we have to change the vertex shader to multiply by
1.0/127.5 and subtract 1.0 to return to signed floats. Ill come back to the optimizations that
you may have noticed, but for now, Ill accept the one cycle cost for a reduction in memory
and bandwidth saving of one-fourth for normals and tangents.
Quantization vertex shader example:
; v1 = normal in range 0 to 255 (integer only)
; c0 = <1.0/127.5, -1.0, ???? , ????>
mad r0, v1, c0.xxxx, c0.yyyy ; multiply compressed normal by 1/127.5, subtract 1
Scaled Offset
This method produces better results than quantization by redistributing the bits to a more rea-
sonable range for this object. By choosing a scale that covers the entire object, we ensure the
object is correct to the accuracy of the quantization level.
A good candidate is to use the axis-aligned bounding box (AABB) as scale and translation
vector. Each vector is defined in object space, so we use the entire compressed range to cover
the AABB. This requires us to translate and then scale the input vectors. You can use either the
center of the AABB or a corner. I usually use a corner, which gives each compressed vertex a
range of 0.0 to 1.0.
Compression
For static objects, we determine the minimum and maximum values for each axis and use these
to calculate the offset point and scales. For dynamic objects, we have to choose the maximum
and minimum that will encompass all dynamic changes. We maximize our dynamic range by
allowing separate scale and offset per axis.
Vertex Decompression in a Shader 175
Team LRN
DetermineScaleAndOffset()
{
For Every Vertex
LowerRange = Minimum (LowerRange, Vertex)
UpperRange = Maximum (UpperRange, Vertex)
}
offset = LowerRange
scale = (UpperRange LowerRange)
ScaledOffset( originalVal )
{
scaledOffset = (originalVal offset) / scale;
return FloatToInteger (scaledOffset );
}
Decompression
The incoming values are in the range 0.0 to 1.0. We simply have to multiply by the scale value
and add the offset to get the original vertex coordinate back.
Constants
{
scale = scale used in the compression
offset = corner used in the compression
}
Decompress( compressedVal )
{
return (IntegerToFloat( compressedVal ) * scale) + centroid
}
Practical Example
Texture coordinate data is a good candidate for this form of compression. 8 bits is usually too
low for texture data (only enough for 1 texel accuracy with a 256x256 texture), but 16 bits is
often enough (unless tiling textures often). We apply the compression and then upload the scale
and offset constants into vertex constants; the scale is itself scaled by 65,535, and the offset has
an additional 0.5 to convert the short back into the 0.0 to 1.0 range. We use a mad instruction
to perform the decompression, so it costs us one cycle but saves us 50% in size.
Scaled offset vertex shader example:
; v1 = uv texture coordinate in the range -32767.0 to 32768.0 (integer only)
; c0 = <u scale / 65535.0, v scale / 65535.0, u offset + 0.5, v offset + 0.5, >
mad r0.xy, v1.xy, c0.xy, c0.zw ; multiply uv by scale and add offset
Compression Transform
This is the superset of the previous two compression methods. We transform our input data
into a space that compresses better. With the scaled offset method, we took our uncompressed
data and applied a scale and offset. This can be represented by applying a scale matrix and a
translation matrix to our input vector. This method further extends the idea by applying another
linear transform (i.e., a matrix).
176 Part 2: Vertex Shader Tricks
Team LRN
To see why this method can produce better results than the other two, look at the case in
Figure 2. We have a simple two-polygon square that is oriented at 45 degrees (A). After scaled
offset compression, we might be left with (B) which has clearly been distorted by the loss of
precision. While higher precision would reduce the error, we may not wish/be able to spend
more precision on it. Looking at (C), it is clear that if the object was rotated by 45 degrees, the
existing precision would produce an exact result.
The compression stage mainly consists of finding a transformation that will minimize the error
when quantization occurs. The method I use is based on finding an optimal oriented bounding
box for the vertices and using the basis matrix of this as my compression space. While it is
possible to find better compression matrices for some objects (this method doesnt consider
transforms such as shears), for most objects it finds a near-optimal linear quantization space.
The compression space I usually use consists of a rotate, a scale, and a translation (a standard
affine transformation), but in theory, any linear transformations could be used. A similar idea is
used in image compression systems, such as MPEG, that use the DCT transform (a DCT is
essentially an 8D rotation matrix).
Compression
Given a set of vertices, we want to find a compression space that minimizes the artifacts
caused by the quantization. The problem has parallels in the collision detection field of trying
to find an optimal oriented bounding box, so I use algorithms developed there to compute the
compression transformation. It uses the vertex data from all triangles to set up a covariance
matrix that we then extract the eigenvectors from.
This gives us a basis system that is aligned along the dominant axes of the data. Combined
with the correct scale and translation, this gives excellent results for almost all data. From a
geometrical point of view, compression transforms each input vector by the OBB basis space
and then translates it so the minimum vertex in AABB space becomes the vector <0.0, 0.0,
0.0> and scales it so the maximum vertex becomes <1.0, 1.0, 1.0>.
CalculateRotationMatrix()
{
n = Number of triangles
m = vector size of data to be compressed
p
i
= 1
st
vertex of triangle i
q
i
= 2nd vertex of triangle i
r
i
= 3rd vertex of triangle i

1
3
0
n
p q r
i i i
i
n
+ +
=

Vertex Decompression in a Shader 177

Figure 2: A square
quantized in both the
rotated and non-rotated
case
Team LRN
p p
i i

q q
i
i

r r
i
i

R
n
p p q q r r
jk
i
n
j
i
k
i
j
i
k
i
j
i
k
i

1
3
0
1 j k m ,
ExtractNormalizedEigenvectors(R)
}
DetermineCompressionMatrix()
{
CalculateRotationMatrix();
For Every Vertex
v' = R
1
* v
LowerRange = Minimum (LowerRange, v')
UpperRange = Maximum (UpperRange, v')
O = TranslationMatrix( LowerRange )
S = ScaleMatrix( UpperRange LowerRange )
C = R
1
* O
1
* S
1
}
CompressionTransform( originalVal )
{
v = originalVal
v
"
= C*v
return FloatToInteger(v
"
)
}
Decompression
We transform the incoming values by the decompression matrix. This has the effect of trans-
forming the compressed vector from compression space back into normal space.
Constants
{
D = C
1
}
Decompress( compressedVal )
{
return (IntegerToFloat( compressedVal ) * D)
}
Practical Example
This compression method is ideal for position data, except for huge objects. Using 16-bit data
types gives 65,535 values across the longest distance of the object; in most cases, it will pro-
vide no visual loss in quality from the original float version. We have to compensate for the
data type range by changing the scale and translation matrix (same as scaled offset) and then
178 Part 2: Vertex Shader Tricks
Team LRN
transpose the matrix (as usual for vertex shader matrices). This gives us a 25% savings in ver-
tex size for four cycles of vertex code.
Compression transform shader example:
; v1 = position in the range 32767.0 to 32768.0 (integer only)
; c0 c3 = Transpose of the decompression matrix
m4x4 r0, v1, c0 ; decompress input vertex
Using the same data stream from the earlier example, we can now see what the vertex shader
might look like.
Compressed vertex example:
D3DVSD_STREAM(0),
D3DVSD_REG( 0, D3DVSDT_SHORT4), // position
D3DVSD_REG( 1, D3DVSDT_UBYTE4), // normal
D3DVSD_REG( 2, D3DVSDT_SHORT2), // texture coordinate
D3DVSD_END()
Compressed vertex shader example:
; v1.xyz = position in the range 32767.0 to 32768.0 (integer only)
; v2.xyz = normal in the range 0.0 to 255.0 (integer only)
; v3.xy = uv in the range 32767.0 to 32768.0 (integer only)
; c0 c3 = Transpose of the world view projection matrix
; c4 c7 = Transpose of the decompression matrix
; c8 = <1.0/255.0, 1.0/255.0, 1.0/255.0, 1.0/255.0>
; c9 = <u scale / 65535.0, v scale / 65535.0, u offset + 0.5, v offset + 0.5>
m4x4 r0, v1, c4 ; decompress compress position
mul r1, v2, c8 ; multiply compressed normal by 1/255
mad r2.xy, v3.xy, c9.xy, c9.zw ; multiply uv by scale and add offset
; now the normal vertex shader code, this example just transforms
; the position and copies the normal and uv coordinate to the rasterizer
m4x4 oPos, r0, c0 ; transform position into HCLIP space
mov oT0, r2 ; copy uv into texture coordinate set 0
mov oT1, r1 ; copy normal into texture coordinate set 1
For an extra six cycles of vertex shader code, we have reduced vertex data size by 50%.
Optimizations
We can usually eliminate most of the decompression instructions, reducing the shader execu-
tion time to roughly the same as the uncompressed version.
The first optimization is noting that anywhere we are already doing a 4x4 matrix trans-
form, we can get any of the previous compressions for free. By incorporating the
decompression step into the transform matrix, the decompression will occur when the matrix
vector multiply is performed (this works because all the compressions involve linear trans-
forms). This is particularly important for position; we always do a 4x4 matrix transform (to
take us from local space into HCLIP space) so we can decompress any of the above for free!
Another optimization is the usual vertex shader rule of never moving a temporary register
into an output register. Wherever you are, simply output directly into the output register.
Anywhere you use quantized D3DVSDT_UBYTE4, you might as well use D3D-
VSDT_D3DCOLOR instead, as it automatically does the divide by 255.0 and swizzling is
free in the vertex shader. This also applies the other way; if you are scaling by a
Vertex Decompression in a Shader 179
Team LRN
D3DVSDT_D3DCOLOR by 255.0 (and your device supports both), use
D3DVSDT_UBYTE4.
Practical Example
By applying these three rules to the same data as before, we achieve 50% savings of memory
and bandwidth for zero cycles.
Optimized compressed vertex example:
D3DVSD_STREAM(0),
D3DVSD_REG( D3DVSDE_POSITION, D3DVSDT_SHORT4),
D3DVSD_REG( D3DVSDE_NORMAL, D3DVSDT_D3DCOLOR),
D3DVSD_REG( D3DVSDE_TEXCOORD0, D3DVSDT_SHORT2),
D3DVSD_END()
Optimized compressed vertex shader example:
; v1.xyz = position in the range 32767.0 to 32768.0 (integer only)
; v2.xyz = normal in the range 0.0 to 255.0 (integer only) (swizzled)
; v3.xy = uv in the range 32767.0 to 32768.0 (integer only)
; c0 c3 = Transpose of the decompression world view projection matrix
; c4 = <u scale / 65535.0, v scale / 65535.0, u offset + 0.5, v offset + 0.5>
m4x4 oPos, v1, c0 ; decompress and transform position
mad oT0.xy, v3, c4.xy, c4.zw ; multiply uv by scale and add offset
mov oT1.xyz, v2.zyx ; swizzle and output normal
Advanced Compression
There are several methods that extend and improve on the basic techniques described in the
previous section for special situations.
Multiple Compression Transforms
The main reason that you may not be able to use a compression transform for position is if the
object is huge but still has fine details (these are classic objects like large buildings, huge space
ships, or terrain). This technique trades vertex constant space for an increase in precision
across the object. It doesnt mix with palette skinning very well, but for static data it can
increase precision significantly.
Essentially, it breaks the object up into separate compressed areas and each vertex picks
which areas it belongs to. The main issue is making sure the compressor doesnt cause gaps in
the object.
We load the address register with the matrix to use, which we store by using a spare com-
ponent (position w is usually spare with compressed data).
180 Part 2: Vertex Shader Tricks
Note: This is the same technique used to speed up rendering of lots of small
objects (things like cubes): Every vertex is treated as if it were skinned with a sin-
gle index and a weight of 1.0. This allows a single render call to have multiple
local matrices. We just use it to select compression matrices rather than local
space matrices.
Team LRN
Compression
The object is broken into chunks beforehand and the standard compressor is applied to each
chunk, storing the chunk number in the vertex.
The exact method of breaking the object into chunks will be very data-dependent. A sim-
ple method might be to subdivide the object into equally spaced chunks; another might be to
choose an area of high detail.
MultipleCompressionTransforms
{
BreakObjectIntoChunks()
For Each Chunk
DetermineCompressionMatrix()
compressionMatrixArray[chunkNum] = C
For Each Vertex v in Chunk
{
v" = C * v
Store FloatToInteger(v") && chunkNum
}
}
Decompression
The only change from the standard compression transform decompression is to use the stored
chunkNum to index into the compression matrix first.
Constants
{
D[MAX_CHUNKS] = C
1
[MAX_CHUNKS]
}
Decompress( compressedVal, chunkNum )
{
return (IntegerToFloat( compressedVal ) * D[chunkNum] )
}
Practical Example
This compression method is good for position data, even for large objects, but isnt usable for
skinned or multi-matrix objects. Each matrix takes four constants, so for equally spaced com-
pression matrices, it quickly fills up constant space. The only extra cost is a single instruction
to fill the index register.
Multiple compression transform shader example:
; v1.xyz = position in the range 32767.0 to 32768.0 (integer only)
; v1.w = (compression matrix index * 4) in the range 0 to MAX_CHUNKS*4
; c0 = < 1.f, ?, ?, ?>
; array of decompression matrices from c1 up to MAX_CHUNKS*4
; i.e., c0 c3 = Transpose of the 1st decompression matrix
; c4 c7 = Transpose of the 2nd decompression matrix
mov a0.x, v1.w ; choose which decompression matrix to use
mov r0.xyz, v1.xyz ; for replacing w
Vertex Decompression in a Shader 181
Team LRN
mov r0.w, c0.x ; set w =1
m4x4 r0, r0 c[a0.x + 0] ; decompress input vertex using selected matrix
Sliding Compression Transform
Rather than pack more data into the already used vector components, this time we put the extra
dummy padding to good use. By encoding displacement values in the padding, we displace
the entire compression matrix, allowing each vertex to slide the compression space to a better
place for it.
The number and size of the spare padding ultimately decides how much extra precision is
needed. Vertex shaders are not very good at extracting multiple numbers from a single pad-
ding, so it is best not to, but we can split values in the compressor and reconstruct the original
value in the vertex shader.
Vertices store offset from local grid origin and store the grid number in the padding. This
example has six times more precision along both world axes than the usual compression trans-
form. The grid is aligned with the compression space, thus preserving maximum precision.
Another way to look at it is that you break the range into discrete chunks and they are treated
separately until decompression.
Compression
The first calculation is to decide how much extra precision you can get out of the padding.
There is a different execution cost depending on how or what you have to extract, and this is
likely to be the most important factor. The ideal situation is three padding values with enough
range. The worst is to extract three values from a single padding. A typical case is one short
pad and one byte pad (from a position and a normal); for this case, we can rearrange the data
so we have three 8-bit values for extra precision.
In all cases, there is an extra precision that has to be factored into the compressor. This can
be done by dividing the range that produces values greater than 1.f and then taking the integer
components as the displacement scalars and the remainder as the compression values. The
other method is to treat the slide as adding directly to the precision.
182 Part 2: Vertex Shader Tricks
Figure 3: Sliding compression
space example
Team LRN
Sliding Compression Transform
{
CalculateRotationMatrix()
For Each Vertex v
{
v' = R
1
* v
LowerRange = Minimum (LowerRange, v')
UpperRange = Maximum (UpperRange, v')
O = TranslationMatrix( LowerRange )
S = ScaleMatrix( UpperRange LowerRange )
}
S = S / size of slide data-type
C = R
1
* O
1
* S
1
For Each Vertex v
{
v" = C * v
c = frac(v")
g = floor(v")
store FloatToInteger( c )
store PackPadding ( g )
}
}
The packing into a spare short and a byte is done by changing the four shorts into two lots of 4
bytes. The space used is the same, but we can access it differently.
Original stream definition:
D3DVSD_REG(0, D3DVSDT_SHORT4),
D3DVSD_REG(1, D3DVSDT_UBYTE4),
New stream definition:
D3DVSD_REG( 0, D3DVSDT_UBYTE4),
D3DVSD_REG( 1, D3DVSDT_UBYTE4),
D3DVSD_REG( 2, D3DVSDT_UBYTE4),
We place the lower 8 bits of each position component in register 0 and the upper 8 bits in regis-
ter 1. This leaves the w component of each register for our slide value. One thing to watch for
is the change from signed 16-bit values to unsigned 16-bit values caused by the split; this
means the decompression matrix may have to change a little.
Decompression
To decompress, we recombine the position, get the slide values, and scale this by the grid size,
which is then added to the compressed value.
Constants
{
D = C
1
GridSize = size of data-type
S = <255.0, 255.0, 255.0, 0.0>
}
Vertex Decompression in a Shader 183
Team LRN
Decompress( lowerVal, upperVal, slideVal )
{
compressedVal = (IntegerToFloat(upperVal) * S) + IntegerToFloat( lowerVal )
return ( compressedVal + (slideVal * GridSize)) * D)
}
; v1.xyz = lower half of position in the range 0 to 255.0 (integer only)
; v1.w = x slide
; v2.xyz = upper half of position in the range 0 to 255.0 (integer only)
; v2.w = y slide
; v3.xyz = normal in the range 0 to 255.0 (integer only)
; v3.w = z slide
; c1 = <255.0, 255.0, 255, 0.0>
; c2 = <65535.0, 1.0, 0.0, 0.0>
mov r1.xyz, v1.xyz ; due to 1 vertex register per instruction
mad r0.xyz, v2.xyz, c1.xyz, r1.xyz ; r0.xyz = position (upper * 255.0) + lower
mad r0.x, v1.w, c2.x, r0.x ; (x slide * 65535.0) + position x
mad r0.y, v2.w, c2.x, r0.y ; (y slide * 65535.0) + position y
mad r0.z, v3.w, c2.x, r0.z ; (z slide * 65535.0) + position z
mov r0.w, c2.y ; set w = 1
; now decompress using a compression matrix
A cost of six cycles has improved our precision 255 times in all axes by reusing the padding
bytes. This effectively gives us 24 bits across the entire object, matching the precision (but not
the range) of the original floating-point version.
Displacement Maps and Vector Fields
Displacements maps (and to a lesser degree their superset vector fields) are generally consid-
ered to be a good thing. Its likely that the future hardware and APIs will support this primitive
directly, but by using vertex shaders, we can do a lot today.
First some background. What are displacement maps and vector fields? A displacement
map is a texture with each texel representing distance along the surface normal. If that sounds a
lot like a height field, its because displacement maps are the superset of height fields (height
fields always displace along the world up direction). Vector fields store a vector at each texel,
with each component representing the distance along a basis vector.
Basic Displacement Maps
184 Part 2: Vertex Shader Tricks
Figure 4: Displacement map example
Team LRN
We have a supporting plane which is constant for the entire displacement map and the dis-
placement along the plane normal at every vertex. For true hardware support, that would be all
that is needed, but we need more information.
To do displacement maps in a vertex shader requires an index list, whereas true displace-
ment map hardware could deduce the index list from the rectilinear shape. Since the index data
is constant for all displacement maps (only one index list per displacement map size is
needed), the overhead is very small.
We also have to store the displacement along the u and v direction, as we cant carry infor-
mation from one vertex to another. This is constant across all displacement maps of the same
size, and can be passed in through a different stream. This makes the vector fields and dis-
placement maps the same for a vertex shader program; the only difference is whether the u and
v displacement is constant and stored in a separate stream (displacement maps) or arbitrarily
changes per vertex (vector fields).
We cant automatically compute the normal and tangent (we would need per-primitive
information), so we have to send them along with a displacement value (if needed; world space
per-pixel lighting removes the need).
The easiest way (conceptually) is to store in vertex constants the three basis vectors repre-
senting the vector field. One basis vector is along the u edge, another the v edge, and the last is
the normal. We also need to store maximum distance along each vector and a translation vector
to the <0.0, 0.0, 0.0> corner of the displacement map.
For each vertex, we take the displacement along each vector and multiply it by the appropriate
scale and basis vector. We then add them to the translation vector to get our displaced point.
Basis displacement map/vector field shader example:
; v1.x = displacement w basis vector (normal) (dW)
; v2.xy = displacements along u and v basis vectors (dU and dV)
; c0.xyz = U basis vector (normalized) (U)
; c1.xyz = V basis vector (normalized) (V)
; c2.xyz = W basis vector (normalized) (W)
; c3.xyz = translation to <0.0, 0.0, 0.0> corner of map (T)
; c4.xyz = U V W scale
; c5 c8 = world view projection matrix
; multiple normalized vectors
mul r0.xyz, c0.xyz, c4.x ; scale U vector
mul r1.xyz, c1.xyz, c4.y ; scale V vector
mul r2.xyz, c2.xyz, c4.z ; scale W vector
; generate displaced point
mul r0.xyz, v2.x, r0.xyz ; r0 = dU * U
Vertex Decompression in a Shader 185
Figure 5: Displacement vector diagram
Team LRN
mad r0.xyz, v2.y, r1.xyz, r0.xyz ; r0 = dV * V + dU * U
mad r0.xyz, v1.x, r2.xyz, r0.xyz ; r0 = dW * W + dV * V + dU * U
add r0.xyz, r0.xyz, c3.xyz ; r0 = T + dW * W + dV * V + dU * U
; standard vector shader code
m4x4 oPos, r0, c5 ; transform position into HCLIP space
Hopefully, you can see this is just an alternative formulation of the compression matrix method
(a rotate, scale, and translation transform). Rather than optimize it in the same way
(pre-multiply the transform matrix by the compression transform), we follow it through in the
other way, moving the displacement vector into HCLIP space.
Entering Hyperspace
Moving the displacement vector into HCLIP space involves understanding four-dimensional
space. By using this 4D hyperspace, we can calculate a displacement vector that works directly
in HCLIP space.
It sounds more difficult than it really is. HCLIP space is the space before perspective divi-
sion. The non-linear transform of perspective is stored in the fourth dimension, and after the
vertex shader is finished (and clipping has finished), the usual three dimensions will be divided
by w.
The most important rule is that everything works as it does in 3D, just with an extra com-
ponent. We treat it as we would any other coordinate space; it just has an extra axis.
All we have to do is transform (before entry into the vertex shader) the basis and transla-
tion vectors into HCLIP space and we use them as before (only we now use four components),
and the displaced point is generated in HCLIP so we dont need to transform them again. We
also pre-multiply the basis scales before transforming the basis vector into HCLIP space.
HCLIP displacement map/vector field shader example:
; v1.x = displacement w basis vector (normal) (dW)
; v2.xy = displacements along u and v basis vectors (dU and dV)
; c0.xyzw = Scaled U basis vector in HCLIP space (U)
; c1.xyzw = Scaled V basis vector in HCLIP space (V)
; c2.xyzw = Scaled W basis vector in HCLIP space (W)
; c3.xyzw = translation to <0.0, 0.0, 0.0> corner of map in HCLIP space (T)
; generate displaced point in HCLIP space
mul r0.xyzw, v2.x, c0.xyzw ; r0 = dU * U
mad r0.xyzw, v2.y, c1.xyzw, r0.xyzw ; r0 = dV * V + dU * U
mad r0.xyzw, v1.x, c2.xyzw, r0.xyzw ; r0 = dW * W + dV * V + dU * U
add oPos, r0.xyzw, c3.xyzw ; r0 = T + dW * W + dV * V + dU * U
This is a matrix transform. The only difference from the m4x4 (4 dp4) or the mul/mad methods
is that the m4x4 is using the transpose of a matrix, whereas mul/mad uses the matrix directly.
Clearly there is no instruction saving (if we could have two constants per instruction we could
get down to three instructions) using vertex shaders for displacement mapping over compres-
sion transform, but we can reuse our index buffers and our u v displacements and make savings
if we have lots of displacement maps.
This allows us to get down to only 4 bytes (either D3DVSDT_UBYTE4 or D3DVSDT_
SHORT2) for every displacement position past the first map. For vector fields, the only sav-
ings we make over compression transform is reusing the index buffers.
186 Part 2: Vertex Shader Tricks
Team LRN
Conclusion
Ive demonstrated several methods for reducing the size of vertex data. The best method obvi-
ously depends on your particular situation, but you can expect a 50% reduction in size quite
easily. Thats good in many ways, as models will load faster (especially over slow paths, like
modems), memory bandwidth is conserved, and memory footprint is kept low.
Acknowledgments
n
S. Gottschalk, M.C. Lin, D. Manocha OBBTree: A Hierarchical Structure for Rapid
Interference Detection, http://citeseer.nj.nec.com/gottschalk96obbtree.html.
n
C. Gotsman Compression of 3D Mesh Geometry, http://www.cs.technion.ac.il/
~gotsman/primus/Documents/primus-lecture-slides.pdf.
n
Tom Forsyth WGDC and Meltdown 2001 Displaced Subdivision
n
Oscar Cooper General sanity checking
Vertex Decompression in a Shader 187
Team LRN
Shadow Volume
Extrusion Using a Vertex
Shader
Chris Brennan
Introduction
The shadow volume technique of rendering real-time shadows involves drawing geometry that
represents the volume of space that bounds what is in the shadow cast by an object. To calcu-
late if a pixel being drawn is in or out of a shadow, shoot a ray from the eye through the
shadow volume toward the point on the rendered object and count the number of times the
shadow volume entered and exited. If the ray has entered more times than exited, the pixel
being drawn is in shadow. The stencil buffer can be used to emulate this by rendering the back
sides of the shadow volume triangles while incrementing the stencil buffer, followed by the
front sides of the triangles, which decrement it. If the final result adds up to where it started,
then you have entered and exited the shadow an equal number of times and are therefore out-
side the shadow; otherwise, you are inside the shadow. The next step is rendering a light pass
that is masked out by a stencil test.
There are several other very different algorithms to doing shadows [Haines01]. Stencil
shadow volumes have their benefits and drawbacks compared to other shadowing algorithms
like depth buffers. The most important trade-off is that while shadow volumes have infinite
precision and no artifacts, they have a hard edge and an uncertain render complexity, depend-
ing on object shape complexity and the viewer and light positions. Previous major drawbacks
to shadow volumes were the CPU power required to compute the shadow geometry and the
requirement that character animation must be done on the CPU so that a proper shadow geom-
etry could be generated, but a clever vertex shader combined with some preprocessing removes
the need for all CPU computations and therefore allows the GPU to do all the character anima-
tion. A brief comparison of CPU and GPU complexity and their drawbacks can be found in
[Dietrich].
Another historical complexity of shadow volumes that has been solved is what to do if the
viewer is inside the shadow. The problem arises that since the viewer starts in shadow, the
stencil count begins off by one. Many solutions have been proposed [Haines02], and many are
very computationally intensive, but a simple solution exists. Instead of incrementing and
188
Team LRN
decrementing the stencil buffer with the visible portion of the shadow volume, modify the sten-
cil buffer when the volume is hidden by another surface by setting the depth test to fail. This
sounds counterintuitive, but what it does is exactly the same thing, except it counts how many
times the ray from the eye to the pixel exits and enters the shadow volume after the visible
point of the pixel. It still tests to see if the pixel is inside or outside of the shadow volume, but
it eliminates the issues with testing to see if the viewer starts in shadow. It does, however,
emphasize the need to make sure that all shadow volumes are complete and closed as opposed
to previous algorithms, which did not require geometry to cap the front or back of the volume.
Creating Shadow Volumes
The time-consuming part of the algorithm is the detection of all the silhouette edges. These are
normally found by taking a dot product with the light vector and each of the edges two neigh-
boring face normals. If one dot product is positive (toward the light) and one is negative (away
from the light), then it is a silhouette edge. For each silhouette edge, create planes extending
from the edge away from the light creating the minimum geometry needed for the shadow vol-
ume. Unfortunately, not only is it expensive to iterate across all of the edges, but it is also
expensive to upload the new geometry to the video card every frame.
However, hardware vertex shaders can be used to do this work on the GPU. The general
idea is to create geometry that can be modified by a vertex shader to properly create the
shadow volume with any light position so that the geometry can reside on the GPU. At initial-
ization or preprocess time, for each edge of the original object geometry, add a quad that has
two sides consisting of copies of the original edge and two opposite sides of zero length. The
pseudocode for this is as follows:
For each face
Calculate face normal
Create 3 new vertices for this face and face normal
Insert the face into the draw list
For each edge of face
If (edge has been seen before)
Insert degenerate quad into draw list
Remove edge from checklist
Else
Insert edge into a checklist
If (any edges are left in checklist)
flag an error because the geometry is not a closed volume.
Figure 1 shows the geometry with the quads inserted and spread apart slightly so that they can
be seen. These quads will make up the extruded edges of the shadow volume when required.
The original geometry is still present and is used to cap the extruded edges on the front and
back to complete the volume. After the quad insertion, each vertex neighbors only one of the
original faces and should include its faces normal. When rendering the volume, each vertexs
face normal is dotted with the light vector. If the result is negative, the face is facing away
from the light and should therefore be pushed out to the outer extent of the light along the light
vector. Otherwise, it stays exactly where the original geometry lies.
Shadow Volume Extrusion Using a Vertex Shader 189
Team LRN
After this pass is completed, the light pass is rendered using a stencil test to knock out the pix-
els that are in shadow.
The shadowed room application with source code can be found on the companion CD and at
http://www.ati.com/na/pages/resource_centre/dev_rel/sdk/RadeonSDK/Html/Samples/Direct-
3D/RadeonShadowShader.html.
190 Part 2: Vertex Shader Tricks
Figure 1: Illustration of invisible quads inserted into
the original geometry to create a new static geome-
try to be used with a shadow volume vertex shader
Figure 2: Shadow volumes after being extruded away
from the light
Figure 3: The final pass completes the
effect.
Team LRN
Effect File Code
matrix mWVP;
matrix mWV;
matrix mP;
matrix mWVt;
vector cShd;
vector pVL;
vertexshader vShd =
decl
{
stream 0;
float v0[3]; // Position
float v1[3]; // FaceNormal
}
asm
{
; Constants:
; 16..19 - Composite World*View*Proj Matrix
; 20..23 - Composite World*View Matrix
; 24..27 - Projection Matrix
; 28..31 - Inv Trans World*View
; 90 - {light range, debug visualization amount, z near, z far}
; 91 - View Space Light Position
vs.1.0
def c0, 0,0,0,1
; View Space
m4x4 r0, v0, c20 ; World*View Transform of point P (pP)
m3x3 r3, v1, c28 ; World*View Transform of normal (vN)
sub r1, r0, c91 ; Ray from light to the point (vLP)
dp3 r11.x, r1, r1 ; length^2
rsq r11.y, r11.x ; 1/length
mul r1, r1, r11.y ; normalized
rcp r11.y, r11.y ; length
sub r11.z, c90.x, r11.y ; light.Range - len(vLP)
max r11.z, r11.z, c0.x ; extrusion length = clamp0(light.Range -
; len(vLP))
dp3 r10.z, r3, r1 ; vLP dot vN
slt r10.x, r10.z, c0.x ; if (vLP.vN < 0) (is pointing away from light)
mad r2, r1, r11.z, r0 ; extrude along vLP
; Projected Space
m4x4 r3, r2, c24 ; Projected extruded position
m4x4 r0, v0, c16 ; World*View*Proj Transform of original position
; Chose final result
sub r10.y, c0.w, r10.x ; !(vLP.vN >= 0)
mul r1, r3, r10.y
mad oPos, r0, r10.x, r1
};
Shadow Volume Extrusion Using a Vertex Shader 191
Team LRN
technique ShadowVolumes
{
pass P0
{
vertexshader = <vShd>;
VertexShaderConstant[16] = <mWVP>;
VertexShaderConstant[20] = <mWV>;
VertexShaderConstant[24] = <mP>;
VertexShaderConstant[28] = <mWVt>;
VertexShaderConstant[90] = <cShd>;
VertexShaderConstant[91] = <pVL>;
ColorWriteEnable = 0;
ZFunc = Less;
ZWriteEnable = False;
StencilEnable = True;
StencilFunc = Always;
StencilMask = 0xffffffff;
StencilWriteMask = 0xffffffff;
CullMode = CCW;
StencilZFail = IncrSat;
}
pass P1
{
CullMode = CW;
StencilZFail = DecrSat;
}
}
Using Shadow Volumes with Character
Animation
The technique as described is for statically shaped objects and does not include characters that
are skinned or tweened. The biggest advantage of doing shadow volume extrusion in a vertex
shader is that the volumes can exist as static vertex buffers on the GPU and the updated geom-
etry does not have to be uploaded every frame. Therefore, this technique needs to be extended
to work on animated characters as well (see David Gosselins Character Animation with
Direct3D Vertex Shaders article in this book). Otherwise, if the animation of the shadow vol-
ume were to be done on the CPU, it would be possible to do the optimized shadow volume
creation at the same time.
The most straightforward approach is to copy the vertex animation data to the shadow vol-
ume geometry and skin and/or tween the face normal as you would the vertex normal.
Unfortunately, the face normals need to be very accurate and consistent across all vertices of a
face; otherwise, objects will be split and extruded in inappropriate places, resulting in incorrect
shadow volumes.
192 Part 2: Vertex Shader Tricks
Team LRN
These artifacts are the result of the extrusion happening across a face of the original geometry
as opposed to the inserted quads along the face edges. This is caused by the fact that each ver-
tex has different weights for skinning, which yield a different face normal for each of the three
vertices of a face. When that face becomes close to being a silhouette edge, it may have one or
two vertices of the face moved away from the light, while the other one stays behind.
The ideal solution is to animate the positions and regenerate the face normals. However,
generating face normals requires vertex neighbor information that is not normally available in
a vertex shader. One possible solution is to make each vertex contain position information of
its two neighbors, animate the vertex position as well as its two neighbors, and recalculate the
face normal with a cross product. The regular vertex position data would need to be stored and
animated three times. This can sometimes be very expensive, depending on the animation
scheme and the size of the models.
An inexpensive way to fix the variation in face normals across a face is to calculate skin-
ning weights per face in addition to per vertex and use the face weights for the face normal.
This can be done by averaging all of the vertex weights or by extracting them directly from the
original art. Using the same weight for each vertex of a face guarantees that the shadow vol-
ume can only be extruded along the edge quads.
By using face weights for the face normals, the previously seen artifacts are not visible.
This technique can be seen in the ATI Island demos on the companion CD, or at
http://www.ati.com/na/pages/resource_ centre/dev_rel/demos.html.
Shadow Volume Extrusion Using a Vertex Shader 193
Figure 4: Artifacts caused by skinned
face normals. Notice the notch in the
shadow of the shoulder.
Figure 5: Animated shadow volumes
with face normals skinned by face
weights
Team LRN
References
[Dietrich] Sim Dietrich, Practical Priority Buffer Shadows, Game Programming Gems II,
Mark DeLoura, Ed. (Charles River Media, 2001), p. 482.
[Haines01] Eric Haines and Tomas Mller, Real-Time Shadows, GDC 2001 Proceedings,
http://www.gdconf.com/archives/proceedings/2001/haines.pdf.
[Haines02] Eric Haines and Tomas Akenine-Mller, Real-Time Rendering, 2nd edition (A.K.
Peters Ltd., 2002).
194 Part 2: Vertex Shader Tricks
Team LRN
Character Animation
with Direct3D Vertex
Shaders
David Gosselin
Introduction
With the introduction of vertex shaders to modern graphics hardware, we are able to move a
large portion of the character animation processing from the CPU to the graphics hardware.
Two common character animation techniques are tweening (also known as morphing) and skin-
ning (also called skeletal animation). This article describes how both of these techniques and
some related variations can be performed using vertex shaders. We will also cover how to com-
bine animation with the per-pixel lighting techniques described in the articles by Wolfgang
Engel in Part 1. We will conclude with a discussion of some geometry decompression tech-
niques, which allow an application to minimize memory usage and memory bandwidth.
Tweening
Tweening is the technique of linearly interpolating between two (or more) key frames of ani-
mation. This technique is used in many popular games like Quake. The artwork consists of a
character or object in a few poses spaced relatively close together in time. For example, the
figure below shows a few frames of an animation stored as key frames.
195
Figure 1: An object animated with tweening
Team LRN
In order to perform this type of animation, the objects position and normal data are stored for
each frame of the animation. At run time, two frames are selected that represent the current
state of the animation. The vertex buffers containing these two frames are loaded as separate
streams of data. Typically, setting up the vertex streams looks something like:
d3dDevice->SetStreamSource (0, frame0VertexBuffer, frame0Stride);
d3dDevice->SetStreamSource (1, frame1VertexBuffer, frame1Stride);
Additionally, a tween factor is computed at run time. The tween factor represents the animation
time between the two frames, with the value 0.0 corresponding to the first frame and the value
1.0 corresponding to the second frame. This value is used to perform the linear interpolation of
the position data. One way to compute this value is shown below:
float t = time;
if (t < 0)
{
t = (float)fmod ((float)(t), TotalAnimationTime);
t = TotalAnimationTime t;
}
float tween = (float)fmod ((float64)t, TotalAnimationTime);
tween /= TotalAnimationTime;
tween *= (NumberOfAnimationFrames 1);
which = (int32)floor ((float64)tween);
tween = which;
The tween value and any lighting constants you may be using are loaded into the constant
store. The vertex shader code then only needs to implement the following equation in order to
perform the interpolation between the two frames:
A*(1tween) + B*tween
Generally, the vertex shader should also multiply by the concatenated world, view, and projec-
tion matrix. The following vertex shader shows an implementation of tweening:
; position frame 0 in v0
; position frame 1 in v14
; tween in c0.x
; World/View/Projection matrix in c12-c15
; Figure out 1-tween constant
; use the 1.0 from positions w
sub r0, v0.wwww, c0
; Compute the tweened position
mul r1, v0, r0.xxxx
mad r1, v14, c0.xxxx, r1
; Multiply by the view/projection and pass it along
m4x4 oPos, r1, c12
To save some vertex shader instructions, the value of (1 tween factor) can be computed by
the application. Both of these values can be loaded into the constant store, in this case as part
of the same constant vector. The z and w components of this constant register are also good
places to stick handy constants like 1.0, 0.0, 2.0, 0.5, etc., if you happen to need them. The
resulting vertex shader code looks like the following:
; position frame 0 in v0
; normal frame 0 in v3
196 Part 2: Vertex Shader Tricks
Team LRN
; position frame 1 in v14
; normal frame 1 in v15
; tween in c0.x
; 1 tween in c0.y
; View/Projection matrix in c12-c15
; Compute the tweened position
mul r1, v0, c0.yyyy
mad r1, v14, c0.xxxx, r1
; Multiply by the view/projection and pass it along
m4x4 oPos, r1, c12
In this section, we have explored an efficient tweening algorithm that can be easily computed
within a vertex shader. It has the advantages of being very quick to compute and working well
with higher order surface tessellation schemes such as N-Patches. Unfortunately, tweening has
the downside of requiring quite a bit of data to be stored in memory for each animation.
Another popular technique, which addresses the memory usage issue, is skinning. Skinning
requires more vertex shader instructions but less data per frame of animation. We will explore
skinning in more detail in the following section.
Skinning
Skinning is the process of blending the contributions
from several matrices in order to find the final vertex
position. Typically, the character modeling tool
describes the matrices using a series of bones. Fig-
ure 3 shows what this looks like in 3D Studio Max.
The diamond shapes in Figure 2 represent the
various control matrices. The inner triangles represent
the bones. In this tool, the artist can move the
joints, and the models position is updated using
inverse kinematics. During export of this data from
Max, the triangles are grouped according to which
matrices affect their position. Further preprocessing
generates the vertex buffers used by the Direct3D
API. The vertices created from the mesh have bone
weights, which represent the amount of influence
from the corresponding matrices (bones). These
weights are used to blend the contributions of the cor-
responding matrices according to the following
equation:
Final Position = SUM (matrix[n]*position*weight[n])
Before the advent of vertex shading hardware, these computations took place on the CPU (typ-
ically referred to as software skinning). This approach has the benefit of being able to
compute accurate normals after skinning a model, but typically it is fairly costly in terms of the
amount of computation required. Some earlier graphics hardware had fixed-function hardware
dedicated to performing these computations. The user specified a Flexible Vertex Format
Character Animation with Direct3D Vertex Shaders 197
Figure 2: Skeleton of a skinned charac-
ter with overlayed polygon mesh
Team LRN
(FVF), which contained the skinning weights and loaded two to four numbered world matrices
to be blended. With vertex shaders, programmers have the ability to tailor the type of skinning
to their own applications. In this section, we will explore two different kinds of skinning: a
four-matrix skinned approach, which mirrors the functionality available from the fixed func-
tion pipeline, and a paletted approach, which allows for greater batching of drawing calls.
The following vertex shader shows how to perform four-matrix skinning. One common
technique to reduce the amount of per-vertex data is to only store three vertex weights and
compute a fourth one in the vertex shader by assuming that the weights sum to 1.0. At run
time, the matrices for a particular group are loaded into the constant store, and the appropriate
vertex buffers are loaded. The vertex shader code looks like the following:
; position in v0
; matrix weights in v1
; c0.z = 1.0
; View/Projection matrix in c12-c15
; World 0 matrix in c16-c19
; World 1 matrix in c20-c23
; World 2 matrix in c24-c27
; World 3 matrix in c28-31
; Multiply input position by matrix 0
m4x4 r0, v0, c16
mul r1, r0, v1.xxxx
; Multiply input position by matrix 1 and sum
m4x4 r0, v0, c20
mad r1, r0, v1.yyyy, r1
; Multiply input position by matrix 2 and sum
m4x4 r0, v0, c24
mad r1, r0, v1.zzzz, r1
; Multiply input position by matrix 3
m4x4 r0, v0, c28
; Compute fourth weight
dp3 r10, v1, c0.zzzz
sub r11, c0.zzzz, r10
; sum
mad r1, r0, r11.wwww, r1
; Multiply by the view/projection matrix
m4x4 oPos, r1, c20
One variation on this technique is to store a palette of matrices in the constant store. During
preprocessing, four indices are stored per-vertex. These indices determine which matrices from
the palette are used in the blending process. Using this technique, a much larger set of triangles
can be processed without changing constant store state (typically an expensive operation). Note
that it is still worthwhile to sort the triangles by similar bone matrices to take advantage of
pipelined hardware. The vertex shader code takes advantage of the indexing register in order to
reach into the constant store to retrieve the correct matrices. The following vertex shader
shows one way to implement this technique.
198 Part 2: Vertex Shader Tricks
Team LRN
; position in v0
; matrix weights in v1
; matrix indices in v2
; c0.z = 1.0
; c10 = (16.0, 4.0, 1.0, 0.0)
; View/Projection matrix in c12-c15
; World 0 matrix in c16-c19
; World 1 matrix in c20-c23
; . . . Other world matrices follow
; figure out the last weight
dp3 r5, v1, c0.zzzz
sub r5.w, c0.zzzz, r5
; First world matrix constant = index*4 + start index
mad r7, v2.xyzw, c10.y, c10.x
; Skin by Matrix 0
mov a0.x, r7.x
m4x4 r0, v0, c[a0.x]
mul r1, r0, v1.xxxx
; Skin by Matrix 1 and sum
mov a0.x, r7.y
m4x4 r0, v0, c[a0.x]
mad r1, r0, v1.yyyy, r1
; Skin by Matrix 2 and sum
mov a0.x, r7.z
m4x4 r0, v0, c[a0.x]
mad r1, r0, v1.zzzz, r1
; Skin by Matrix 3 and sum
mov a0.x, r7.w
m4x4 r0, v0, c[a0.x]
mad r1, r0, r5.wwww, r1
; Multiply by the view/projection and pass it along
m4x4 oPos, r1, c12
One drawback to using the paletted skinning approach is that it doesnt work well with higher
order surfaces. The issue stems from the fact that vertex shaders are computed post-tessellation
and any vertex data other than positions and normals are linearly interpolated. Interpolating the
index values for blend matrices is nonsensical.
Skinning and Tweening Together
It is also possible to blend skinning and tweening. There are cases where it makes sense to
model a portion of a character using tweening, but you still want some basic movement con-
trolled by bones. For example, you might want the face of a character to be tweened in order to
capture some subtle movements that would require a large number of bones and would be dif-
ficult to manage within an art tool, but you want to use bones for the movement of the head
and body. The following figures show this kind of example, where the mouth and claws of the
character are tweened and skinned.
The following figure shows a few frames of animation using skinning alone.
Character Animation with Direct3D Vertex Shaders 199
Team LRN
The next figure shows just the tweened portion of the animation. Note the complex animation
of the mouth and claws.
The final figure shows the skinning and tweening animations being performed simultaneously
on the character.
When performing this type of animation, the tween portion is computed first and then the
resultant position is skinned. The following vertex shader shows one implementation of this
technique.
; position frame 0 in v0
; position frame 1 in v14
; matrix weights in v1
; matrix indices in v2
200 Part 2: Vertex Shader Tricks
Figure 3
Figure 4
Figure 5
Team LRN
; c0.z = 1.0
; tween in c9.x
; c10 = (16.0, 4.0, 1.0, 0.0)
; View/Projection matrix in c12-c15
; World 0 matrix in c16-c19
; World 1 matrix in c20-c23
; . . . Other world matrices follow
; Figure out tween constant
sub r9, v0.wwww, c9
mov r9.y, c9.x
; Compute the tweened position
mul r10, v0, r9.xxxx
mad r2, v14, r9.yyyy, r10
; figure out the last weight
mov r5, v1
dp3 r3, v1, c0.zzzz
sub r5.w, c0.zzzz, r3
; First world matrix constant = index*4 + start index
mad r7, v2.xyzw, c10.y, c10.x
; Skin by Matrix 0
mov a0.x, r7.x
m4x4 r0, r2, c[a0.x]
mul r1, r0, v1.xxxx
; Skin by Matrix 1 and sum
mov a0.x, r7.y
m4x4 r0, r2, c[a0.x]
mad r1, r0, v1.yyyy, r1
; Skin by Matrix 2 and sum
mov a0.x, r7.z
m4x4 r0, r2, c[a0.x]
mad r1, r0, v1.zzzz, r1
; Skin by Matrix 3 and sum
mov a0.x, r7.w
m4x4 r0, r2, c[a0.x]
mad r1, r0, r5.wwww, r1
; Multiply by the projection and pass it along
m4x4 oPos, r1, c12
In this section, we explored one way to combine two different types of character animation. It
is obviously just one way to mix the two types of animation. One benefit of vertex shaders is
the ability to customize the vertex processing to get different effects that fit the needs of your
particular application. Hopefully, this will provide you with some ideas for how to customize
animation to suit your own needs.
So far, we have only discussed animation of vertex positions and normals. This is suffi-
cient for vertex lighting of animated characters. Modern graphics chips provide very powerful
pixel shading capabilities that allow lighting to be performed per-pixel. To do this, however,
some care must be taken in the vertex shader to properly set up the per-pixel lighting computa-
tions. This will be discussed in the next section.
Character Animation with Direct3D Vertex Shaders 201
Team LRN
Animating Tangent Space for Per-Pixel Lighting
In order to reasonably light a character animated by these techniques, it is important to animate
the normal as well as the position. In the case of tweening, this means storing a normal along
with the position and interpolating the normal in the same way as the position. Getting the
right normal isnt quite as simple in the case of skinning. Since skinning can potentially
move each vertex in very different ways, the only way to get completely correct normals is to
recompute the normal at each vertex once the mesh is reskinned. Typically, this tends to be a
very expensive operation and would require a lot of additional information stored per-vertex.
To top it off, there really isnt a way to index other vertices from within a vertex shader. The
typical compromise, which gives good enough results, is to skin the normal using just the
rotation part of the bone matrices.
Additionally, if you are doing per-pixel bump mapping or other effects requiring tex-
ture/tangent space discussed in the articles by Engel in Part 1, you also need to skin the basis
vectors. The following shader shows the paletted matrix skinned version that also skins the
tangent/texture space basis vectors.
; position in v0
; matrix weights in v1
; matrix indices in v2
; normal in v3
; tangent in v9
; binormal in v10
; c0.z = 1.0
; c10 = (16.0, 4.0, 1.0, 0.0)
; View/Projection matrix in c12-c15
; World 0 matrix in c16-c19
; World 1 matrix in c20-c23
; . . . Other world matrices follow
; figure out the last weight
dp3 r3, v1, c0.zzzz
sub r5.w, c0.zzzz, r3
; First world matrix constant = index*4 + start index
mad r7, v2.xyzw, c10.y, c10.x
; Skin by Matrix 0
mov a0.x, r7.x
m4x4 r0, v0, c[a0.x] ; Position
mul r1, r0, v1.xxxx
m3x3 r0, v9, c[a0.x] ; Tangent
mul r2, r0, v1.xxxx
m3x3 r0, v10, c[a0.x] ; Bi-Normal
mul r3, r0, v1.xxxx
m3x3 r0, v3, c[a0.x] ; Normal
mul r4, r0, v1.xxxx
; Skin by Matrix 1 and sum
mov a0.x, r7.y
m4x4 r0, v0, c[a0.x] ; Position
mad r1, r0, v1.yyyy, r1
m3x3 r0, v9, c[a0.x] ; Tangent
mad r2, r0, v1.yyyy, r2
202 Part 2: Vertex Shader Tricks
Team LRN
m3x3 r0, v10, c[a0.x] ; Bi-Normal
mad r3, r0, v1.yyyy, r3
m3x3 r0, v3, c[a0.x] ; Normal
mad r4, r0, v1.yyyy, r4
; Skin by Matrix 2 and sum
mov a0.x, r7.z
m4x4 r0, v0, c[a0.x] ; Position
mad r1, r0, v1.zzzz, r1
m3x3 r0, v9, c[a0.x] ; Tangent
mad r2, r0, v1.zzzz, r2
m3x3 r0, v10, c[a0.x] ; Bi-Normal
mad r3, r0, v1.zzzz, r3
m3x3 r0, v3, c[a0.x] ; Normal
mad r4, r0, v1.zzzz, r4
; Skin by Matrix 3 and sum
mov a0.x, r7.w
m4x4 r0, v0, c[a0.x] ; Position
mad r1, r0, r5.wwww, r1
m3x3 r0, v9, c[a0.x] ; Tangent
mad r2, r0, r5.wwww, r2
m3x3 r0, v10, c[a0.x] ; Bi-Normal
mad r3, r0, r5.wwww, r3
m3x3 r0, v3, c[a0.x] ; Normal
mad r4, r0, r5.wwww, r4
; Multiply by the projection and pass it along
m4x4 oPos, r1, c12
; >>>> At this point:
; >>>>> r1 contains the skinned vertex position
; >>>>> r2 contains the tangent (v9)
; >>>>> r3 contains the binormal (v10)
; >>>>> r4 contains the normal (v3)
Now, you might ask, what can we do with this basis vector? Well, one common usage is to per-
form per-pixel Dot3 bump mapping. In order to compute this, two textures are needed. The
first texture is the base map, which contains the color on the surface. The second texture is the
bump map that contains the normal at each point in the original texture encoded into the red,
green, and blue channels. Within the vertex shader, the vector from the light to the vertex is
computed, normalized, and converted into tangent/texture space. This vector is then interpo-
lated per-pixel before being sent to the pixel shader.
Per-Pixel Lighting
Additionally, within the vertex shader, you can compute a light falloff. This involves sending
data about the light falloff values in the constant store. The three values sent to the vertex
shader are the distance from the light when the falloff starts, the distance from the light where
the falloff ends, and the difference between those two values. The following shader code builds
upon the previous shader and performs the calculations for three bumped diffuse lights. If
fewer lights are desired, the same shader can be used by setting the color of one or more of the
lights to black (0,0,0,0).
Character Animation with Direct3D Vertex Shaders 203
Team LRN
; c0.z = 1.0
; c1 has the position of light 1
; c2 has the color of light 1
; c3.x has the start of the falloff for light 1
; c3.y has the end of the falloff for light 1
; c3.z has the difference between the end of the falloff and the
; start of the falloff
; c4 has the position of light 2
; c5 has the color of light 2
; c6.x has the start of the falloff for light 2
; c6.y has the end of the falloff for light 2
; c6.z has the difference between the end of the falloff and the
; start of the falloff
; c7 has the position of light 2
; c8 has the color of light 2
; c9.x has the start of the falloff for light 2
; c9.y has the end of the falloff for light 2
; c9.z has the difference between the end of the falloff and the
; start of the falloff
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Compute vector for light 1
sub r0, c1, r1 ; Ray from light to point
dp3 r5.x, r0, r0 ; length^2
rsq r5.y, r5.x ; 1/length
mul r0, r0, r5.y ; normalized
rcp r5.z, r5.y ; length
m3x3 r8, r0, r2 ; Convert to tangent space
dp3 r9.x, r8, r8 ; length^2
rsq r9.y, r9.x ; 1/length
mul oT2.xyz, r8, r9.y ; normalized
sub r7.x, c3.y, r5.z ; fallEnd - length
mul r7.y, c3.z, r7.x ; (fallEnd - length)/
; (fallEnd-fallStart)
min r7.w, r7.y, c0.z ; clamp
mul oD0.xyz, r7.w, c2 ; falloff * light color
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Compute vector for light 2
sub r0, c4, r1 ; Ray from light to the point
dp3 r5.x, r0, r0 ; length^2
rsq r5.y, r5.x ; 1/length
mul r0, r0, r5.y ; normalized
rcp r5.z, r5.y ; length
m3x3 r8, r0, r2 ; Convert to tangent space
dp3 r9.x, r8, r8 ; length^2
rsq r9.y, r9.x ; 1/length
mul oT3.xyz, r8, r9.y ; normalized
sub r7.x, c6.y, r5.z ; fallEnd - length
mul r7.y, c6.z, r7.x ; (fallEnd - length)/
; (fallEnd - fallStart)
min r7.w, r7.y, c0.z ; clamp
mul oD1.xyz, r7.w, c5 ; falloff * light color
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Compute vector for light 3
sub r0, c7, r1 ; Ray from light to the point
204 Part 2: Vertex Shader Tricks
Team LRN
dp3 r5.x, r0, r0 ; length^2
rsq r5.y, r5.x ; 1/length
mul r0, r0, r5.y ; normalized
rcp r5.z, r5.y ; length
m3x3 r8, r0, r2 ; Convert to tangent space
dp3 r9.x, r8, r8 ; length^2
rsq r9.y, r9.x ; 1/length
mul oT4.xyz, r8, r9.y ; normalized
sub r7.x, c9.y, r5.z ; fallEnd - length
mul r7.y, c9.z, r7.x ; (fallEnd - length)/
; (fallEnd- fallStart)
min r7.w, r7.y, c0.z ; clamp
mul oT5.xyz, r7.w, c8 ; falloff * light color
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Pass along the texture coordinates.
mov oT0.xy, v7.xy
mov oT1.xy, v8.xy
In the pixel shader, we compute the dot products and sum up the contributions from all three
lights. In addition, the light colors passed into the above shaders are pre-divided by two to
allow a bigger range of light colors and allow the lights to over-brighten the base textures.
Within the pixel shader, this range is re-expanded by using the _x2 modifier on the colors.
ps.1.4
; c2 contains the ambient lighting color
texld r0, t0 ; DOT3 Bump Map Texture
texld r1, t1 ; Base Texture
texcrd r2.rgb, t2 ; L1 light 1 vector
texcrd r3.rgb, t3 ; L2 light 2 vector
texcrd r4.rgb, t4 ; L3 light 3 vector
; v0 C1 color of light 1 (from above)
; v1 C1 color of light 2 (from above)
texcrd r5.rgb, t5 ; C3 color of light 3
dp3_sat r2, r0_bx2, r2_bx2 ; N.L1
mul r2, r2, v0_x2 ; (N.L1)*C1
dp3_sat r3, r0_bx2, r3_bx2 ; N.L2
mad r2, r3, v1_x2, r2 ; ((N.L1)*C1) + ((N.L2)*C2)
dp3_sat r4, r0_bx2, r4_bx2 ; N.L2
mad r2.rgb, r4, r5_x2, r2 ; ((N.L1)*C1) + ((N.L2)*C2) +
; ((N.L3)*C3)
add r2.rgb, r2, c2 ; add ambient
mul r0.rgb, r2, r1 ; ((N.L1)*C1) + ((N.L2)*C2) +
; ((N.L3)*C3)*base
The following figures show various pieces of this shader in action. The first three characters in
Figure 6 show the contributions from each of the individual lights. The fourth character shows
the summation of all the lighting contributions, and the fifth shows the base texture for the
character.
Character Animation with Direct3D Vertex Shaders 205
Team LRN
The next figure shows a few frames of animation with the full shader in action.
Compression
One downside of doing tweened animation is the large amount of storage space required to
store each individual frame of animation data. Thus, it is desirable to reduce the amount of
memory consumed in storing the character data as well as the amount of data that needs to be
sent to the video card via the buses in the system. One compression technique is done by com-
puting the bounding box of the character and specifying the position data as an integer offset
from one of the corners of the box. The following example packs the position data into 16-bit
shorts for x, y, and z, and pads with one short for w. The preprocessing code to do the packing
looks something like this:
206 Part 2: Vertex Shader Tricks
Figure 6
Figure 7
Team LRN
float half = (maxX minX)/2.0f;
vertex.x = (short)(((x minX half)/half)*32767.5f)
half = (maxY minY)/2.0f;
vertex.y = (short)(((y minY half)/half)*32767.5f)
half = (maxZ minZ)/2.0f;
vertex.z = (short)(((z minZ half)/half)*32767.5f)
The vertex data can then be re-expanded within the vertex shader at a relatively low cost. The
following vertex shader fragment shows how this is accomplished:
; position in v0
; c0.x = ((maxX minX)/2)/32767.5
; c0.y = ((maxY minY)/2)/32767.5
; c0.z = ((maxZ minZ)/2)/32767.5
; c1.x = minX + (maxX minX)/2
; c1.y = minY + (maxY minY)/2
; c1.z = minZ + (maxZ minZ)/2
; scale and bias position
mul r2, v0, c0
add r2, r2, c1
A similar technique can be applied to the weights for the skinning matrices. Again, here is a
small bit of preprocessing code:
vertex.w1 = (unsigned char)((float)(w1)/255.0f);
vertex.w2 = (unsigned char)((float)(w2)/255.0f);
vertex.w3 = (unsigned char)((float)(w3)/255.0f);
A few vertex shader instructions can expand this data back into floating-point data to use for
computing the skinned position.
; v1 contains the weights
; c0.x = 0.003921569 = 1/255
; c0.z = 1.0
; unpack the weights
mul r9, v1, c0.x
dp3 r9.w, r9, c0.zzzz
sub r9.w, c0.zzzz, r9.w
Normal data can also be compressed by sacrificing some quality by quantizing the normals
into bytes or shorts and doing a similar unpacking process within the vertex shader. The fol-
lowing code shows how the data is packed within a preprocessing tool:
vertex.nx = (unsigned char)(nx*127.5 + 127.5);
vertex.ny = (unsigned char)(ny*127.5 + 127.5);
vertex.nz = (unsigned char)(nz*127.5 + 127.5);
A small bit of vertex shader code can be used to decompress the normals:
; v3 contains the normal
; c0.x = 0.007843137 = 1/127.5
; r2.w = 1.0
mad r3, v3, c0.x, r2.w ; scale and bias normal to 1 to 1 range
In this section, we showed a few ways to compress various vertex components. These tech-
niques can be used to significantly reduce the amount of data required per character. They
should, however, be used with some caution since they are lossy methods of compression and
will not work with all data sets. For more on vertex compression, see the Vertex Decompres-
sion Using Vertex Shaders article in this book.
Character Animation with Direct3D Vertex Shaders 207
Team LRN
Summary
This chapter has discussed a few ways to perform character animation within a vertex shader.
This should be considered a starting point for your exploration of character animation through
vertex shaders. Vertex shaders give you the power to customize the graphics pipeline for your
application, allowing you to free up the CPU to make a more engaging game.
208 Part 2: Vertex Shader Tricks
Team LRN
Lighting a Single-Surface
Object
Greg James
In this article, well explore how a programmable vertex shader can be used to light objects
that expose both sides of their triangles to the viewer. Such objects can be problematic with
fixed-function hardware lighting, but a few vertex shader instructions allow us to light such
objects correctly from all points of view without having to modify vertex data, duplicate trian-
gles, or make complex decisions based on the scene. We will also extend the basic technique to
approximate the effect of light transmission through thin scattering objects.
Three-dimensional objects are often represented by single-surface geometry in 3D applica-
tions. Leaves, foliage, blades of grass, particle system objects, hair, and clothing are typically
modeled as sheets of triangles where both sides of the sheet are potentially visible, and as such,
these objects are rendered with back-face culling disabled. This approach is used widely in 3D
games but presents a problem for hardware-accelerated lighting calculations. The problem
stems from the restriction that each vertex is specified with only a single surface normal vector.
This normal is correct when viewing the front faces of triangles but is not correct when the
back faces are in view. When back faces are visible, this single normal vector points opposite
the direction it should, so a lighting calculation based on this normal will give an incorrect
result. Fortunately, we can use five instructions in an OpenGL or DirectX 8 vertex shader pro-
gram to calculate the correct normal and pass it on to the lighting calculation so that a
single-surface object is lit correctly from all orientations. This same correction will also light a
twisted object correctly regardless of how the surfaces front and back faces twist in and out of
view. This approach of detecting the orientation of the normal and correcting it can also be
used to enhance the lighting of thin objects by accounting for transmitted light. A realistic ren-
dering of light shining through leaves or stained glass can be achieved without costly logic
based on the positions of the objects in the scene.
The problem of lighting a single-surface object depends on the viewers orientation rela-
tive to each face. Since surface normals are defined at each vertex, the surface facing is really a
vertex facing defined by each vertexs normal vector. The four possible orientations of the ver-
tex normal, viewer location, and light location are shown in Figure 1. In the first two cases, the
front faces are visible and no correction of the normal is required. In the last two cases, the
back faces are visible and the normal must be reversed to properly represent the face in view.
Note that the position of the light is not relevant to the problem of selecting the proper normal.
Only the eye position relative to the vertex normal matters. Correct illumination will result if
the vertex normal is always chosen to face toward the viewer. The operation of guaranteeing
that the normal always faces the viewer is easily accomplished in a vertex shader program.
209
Team LRN
Calculations for correcting the normal can be done in object space or world space. Since the
vertex input position is in object space, and in a vertex shader this is typically transformed
directly to homogeneous clip space (i.e., the world space position is never calculated), it is
more efficient to perform the calculations in object space. If the calculations were done in
world space, this would require extra vertex shader instructions to transform the vertex posi-
tion and vertex normal to world space. On the other hand, working in object space requires us
to transform the light position and viewer position from world space to object space. Since the
light and viewer position in object space is constant for all vertices of the object being ren-
dered, they should be calculated on the CPU and supplied as vertex shader constants. It would
be wasteful to compute them in the vertex shader itself. To compute them, we will have to
compute the inverse object-to-world transformation matrix, and this is simple if the world
matrix consists only of rotation and translation without scaling [Foley94]. In that case, the
inverse of the 4x4 transform is composed of the transpose of the 3x3 rotation portion along
with the opposite amount of translation.
210 Part 2: Vertex Shader Tricks
Figure 1: The four
possible orienta-
tions of viewer,
vertex normal, and
light. The light
position is irrele-
vant to selecting
the proper vertex
normal for
lighting.
Figure 2: Illustration of vari-
ables calculated in Listing 1
Team LRN
Vertex Shader Code
Figure 2 illustrates the values that well calculate, and Listing 1 shows the vertex shader
instructions. The first step of the correction is to compute a vector from the vertex position to
the viewer position. The viewer position in object space is computed on the CPU and supplied
as a vertex shader constant: c[CV_EYE_POS_OSPACE]. The vector from the vertex to the eye
is a simple subtraction. Next, we test whether the vertex normal is facing toward or away from
the eye. The three component dot-product between the normal and the vector to the eye will
give a value that is negative if the vertex normal points away from the eye or positive if the
vertex normal points toward the eye. The result of the dot-product is written to the ALIGNED
variable, which is positive if no correction is needed or negative if we have to flip the normal.
At this point, we must do one of two things based on the sign of the ALIGNED variable.
The C equivalent of what we want to do is:
if ( ALIGNED < 0 )
normal = V_NORMAL;
else
normal = V_NORMAL;
There are no if or else branch instructions in the vertex shader instruction set, so how can
we perform this decision based on the sign of ALIGNED? The answer is to do one calculation
for both cases which will result in a value as though a branch had been used. We use both the
input normal (V_NORMAL) and the reversed normal (V_NORMAL) in an expression in
which either value can be masked out. Each value is multiplied by a mask and summed. If the
value of mask is 1 or 0, then it can be used in the following equation to select either
V_NORMAL or V_NORMAL.
normal = mask * (V_NORMAL) + ( 1 mask ) * V_NORMAL
In a vertex shader, we could compute mask and 1mask, and this would be expressed as:
sub INV_MASK, c[ONE], MASK
mul COR_NORMAL, MASK, V_NORMAL
mad COR_NORMAL, INV_MASK, V_NORMAL, COR_NORMAL
For our case, the negative normal is computed from the positive normal as in the following
equation, so we can skip the calculation of 1mask and save one instruction in the vertex
shader program.
V_NORMAL = V_NORMAL 2 * V_NORMAL
The vertex shader instructions are then:
mad COR_NORMAL, V_NORMAL, MASK, V_NORMAL
mad COR_NORMAL, V_NORMAL, MASK, COR_NORMAL
The mask value is computed from ALIGNED using the sge (Set if Greater or Equal) instruc-
tion to compare ALIGNED to zero.
sge MASK, ALIGNED, c[ZERO]
MASK will be 0 if ALIGNED is less than c[ZERO], or it will be 1 if ALIGNED is greater or
equal to c[ZERO].
Lighting a Single-Surface Object 211
Team LRN
The five instructions in the middle of Listing 1 (sub, dp3, sge, mad, mad) will always pro-
duce a vertex normal that faces toward the viewer. Lighting based on this normal will be
correct for all light positions.
Performing the correction of the normal in a vertex shader lets us fill a scene with many
single-surface objects with only a small cost for the added calculations. We can avoid more
costly operations such as updating vertex data with the CPU or rendering in multiple passes to
achieve the correct lighting. The single-surface objects may also be animated in the vertex
shader and remain correctly lit. By performing all of the transform and lighting calculations in
a hardware vertex shader, we can populate scenes with a large number of these objects.
Listing 1: Vertex shader for correcting the input vertex normal based on viewer position
// Include definitions of constant indices
#include "TwoSided.h"
// Define names for vertex input data
#define V_POSITION v0
#define V_NORMAL v1
#define V_DIFFUSE v2
#define V_TEXTURE v3
// Define names for temporaries
#define VEC_VERT_TO_EYE r0
#define VEC_VERT_TO_LIGHT r1
#define ALIGNED r2
#define TEMP r6
#define COR_NORMAL r7
// Vertex shader version 1.1
vs.1.1
// Transform position to clip space and output it
dp4 oPos.x, V_POSITION, c[CV_WORLDVIEWPROJ_0]
dp4 oPos.y, V_POSITION, c[CV_WORLDVIEWPROJ_1]
dp4 oPos.z, V_POSITION, c[CV_WORLDVIEWPROJ_2]
dp4 oPos.w, V_POSITION, c[CV_WORLDVIEWPROJ_3]
// Use eye position relative to the vertex to
// determine the correct orientation of the
// vertex normal. Flip the normal if it is
// pointed away from the camera.
// Eye & light positions were transformed to
// object space on the CPU and are supplied
// as constants.
// Make vector from vertex to the eye
sub VEC_VERT_TO_EYE, c[CV_EYE_POS_OSPACE], V_POSITION
// Dot product with the normal to see if they point
// in the same direction or opposite
dp3 ALIGNED, V_NORMAL, VEC_VERT_TO_EYE
// If aligned is positive, no correction is needed
// If aligned is negative, we need to flip the normal
// Do this with SGE to create a mask for a virtual
// branch calculation
// ALIGNED.x = ALIGNED.x >= 0 ? 1 : 0;
212 Part 2: Vertex Shader Tricks
Team LRN
sge ALIGNED.x, ALIGNED, c[CV_ZERO]
mad COR_NORMAL, V_NORMAL, ALIGNED.x, V_NORMAL
// now COR_NORMAL = 0 or V_NORMAL
mad COR_NORMAL, V_NORMAL, ALIGNED.x, COR_NORMAL
// COR_NORMAL = V_NORMAL or V_NORMAL
// Point lighting
// Vector from vertex to the light
add VEC_VERT_TO_LIGHT, c[CV_LIGHT_POS_OSPACE], V_POSITION
// Normalize it using 3 instructions
dp3 TEMP.w, VEC_VERT_TO_LIGHT, VEC_VERT_TO_LIGHT
rsq TEMP.w, TEMP.w
mul VEC_VERT_TO_LIGHT, VEC_VERT_TO_LIGHT, TEMP.w
// dp3 for lighting. Point light is not attenuated
dp3 r4, VEC_VERT_TO_LIGHT, COR_NORMAL
// Use LIT to clamp to zero if r4 < 0
// r5 will have diffuse light value in y component
lit r5, r4
// Light color to output diffuse color
// c[CV_LIGHT_CONST].x is ambient
add oD0, r5.y, c[CV_LIGHT_CONST].x
// Another lit for the light value from opposite normal
lit r5, r4
// Square the value for more attenuation. This
// represents transmitted light, which could fall
// off more sharply with angle of incidence
mul r5.y, r5.y, r5.y
// Attenuate it further and add transmitted ambient factor
mad oD1, r5.y, c[CV_LIGHT_CONST].z, c[CV_LIGHT_CONST].y
// Output texture coordinates
mov oT0, V_TEXTURE
mov oT1, V_TEXTURE
Enhanced Lighting for Thin Objects
We can take this approach a step further by enhancing the lighting model for these thin objects.
Ordinarily, we would most likely compute some function of the dot product of the light vector
and normal vector (NL) at each vertex and be done with it. This would account for light scat-
tering off the surface but not for light shining through the surface. After all, these are thin
objects, and thin objects tend to transmit light. Dramatic lighting effects are possible if we
account for this transmission.
In this case, the transmission is not that of a clear object allowing light to pass directly
through, but it is instead a diffuse scattering of transmitted light. Things behind the object are
not visible, but bright light shining through the object is dispersed and scattered by the mate-
rial. This reveals the inner details or thickness of the object, which have a different appearance
Lighting a Single-Surface Object 213
Team LRN
than the surface illuminated by reflected light. Leaves, stained glass, and lamp shades show a
good contrast between the transmitted and reflected light. Hold a leaf to the ground and it
appears dark and waxy. Hold it up to the sun, and the veins and cells are revealed against a
glowing green background. Stained glass viewed from the outside of a building is entirely
dark, but viewing it from the inside with sunlight streaming through reveals brilliant colors.
We can account for this transmitted light by supplying an additional texture for the object.
This additional texture is simply the colors we would see with bright light shining through
the object. It is a texture representing light transmission. The standard lighting equation is
applied to the ordinary diffuse reflective texture, and we add a new lighting calculation that
will affect the diffuse transmission texture. The new lighting calculation is based on the normal
pointing away from the viewer and should contribute no light when the light is positioned in
front of the vertex in view. We use the vertex shader lit instruction to achieve this clamping to
zero, just as we would for the front-facing lighting. When the front-facing reflective lighting
clamps to zero, the transmissive lighting begins to show through.
The result of scattered reflected light can be placed in one vertex color, and the result of
scattered transmitted light can be placed in the second specular color or any output which will
be interpolated in rasterization. A pixel shader then fetches from each texture, applies the light
contribution for transmission to the transmission texture and for reflection to the ordinary dif-
fuse texture, and adds the results together. As each contribution is clamped to zero based on
whether or not the light is in front or behind, the addition of each term results in correct
lighting.
In this way, dramatic lighting effects are possible without requiring CPU processing or
complex logic based on the placement of objects and lights. As the sun streams through leaves
overhead, they appear different than leaves off in the distance, or as the viewer and lights move
about a scene, fabrics and hair will change appropriately when lit from behind. Other effects
like x-ray illumination or one-way mirrors could also be devised based on the vertex shaders
determination of the correct normal.
About the Demo
Included on the companion CD is the NV Effects Browser demo Two-Sided Lighting, which
demonstrates this technique. To run, simply start the NVEffectsBrowser.exe and click the
effect under the Lighting (per-vertex) section. Keyboard controls can be displayed by hitting
the H or F1 keys. There are controls for twisting, scaling, and moving the object. Complete
DirectX 8 source code is included.
References
[Foley94] James D. Foley, Andries van Dam, et. al., Introduction to Computer Graphics
(Addison-Wesley Co., 1994), p. 182.
214 Part 2: Vertex Shader Tricks
Team LRN
Optimizing Software
Vertex Shaders
Kim Pallister
The introduction of the vertex shader technology in the DirectX APIs marked a significant
milestone in the real-time graphics and gaming industries. It was significant not so much
because it exposed any new capability, but rather because it put innovation back in the hands of
the developer.
By making the pipeline programmable, developers were no longer limited to what the
hardware and API had been explicitly designed to do. Conversely, hardware developers, and
those designing the API, did not need to be clairvoyant in anticipating what game developers
would want the pipeline to do over the coming years.
Thats the good side of things. However, like most things, theres a cloud attached to this
silver lining.
As is often the case with PC game development, developers are challenged to take advan-
tage of new technologies while still supporting older systems. If programmable vertex shaders
were only supported on the highest end of the PC performance spectrum, developers would
have their work cut out for them, having to support multiple code paths for systems without
support.
Thankfully, this isnt the case. Microsoft, with help from CPU vendors like Intel, has seen
to it that software vertex shaders offer impressive performance and can be used in games today.
This is accomplished through a vertex shader compiler within the DirectX run-time that
compiles the vertex shader code into CPU-optimal instruction streams (SSE and SSE2 instruc-
tions in the case of Pentium III and Pentium 4 processors).
Like with any compiler, properly designing and structuring the source code going into the
compiler can improve the code that it generates. In this article, well look at some guidelines
for writing vertex shaders to improve the performance of those shaders when running in
software.
If you are simply interested in knowing a set of canonical rules to apply to your shaders,
you can skip to the optimization guidelines in Table 1. However, if you are interested in some
background about how the software emulation is done, well cover that beforehand. First well
look at some details about the Pentium III and Pentium 4 processor architectures and then look
at how the DirectX run-time compiler works. This is followed by a list of optimization guide-
lines and an explanation of why they help generate optimal code. Finally, well look at a
sample vertex shader and how we can optimize it.
215
Team LRN
Introduction to Pentium 4 Processor Architecture
The Pentium II and Pentium III processors were both based on the same core architecture. This
microarchitecture introduced speculative execution and out-of-order operation to Intel Archi-
tecture. The Pentium III processor made improvements over the Pentium II processor and
added the Streaming SIMD Extensions (SSE).
The Pentium 4 processor introduced a new microarchitecture, the NetBurst
microarchitecture. Like its predecessors, the Pentium 4 processor operates out of order and
does speculative execution.
However, it has a much deeper
pipeline and boasts a number of
other enhancements. For a more
detailed explanation, see Figure
1.
Instructions are specula-
tively fetched by the Branch
Predictor, the part of the diagram
on the far left comprised of the
Branch Target Buffer, or BTB,
and the Instruction Translation
Lookaside Buffer, or I-TLB. The
logic in these units serves to
make predictions about what
instructions will be needed next,
based on a branch history and some prediction rules.
From there, the instructions are fed into the decoder. The Instruction Decoder receives the
IA-32 opcode bytes delivered to it and breaks them into one or more simple micro-ops (Ops).
These Ops are the operations that that core of the Pentium 4 processor executes.
The next stage is the Execution Trace Cache, an innovative microarchitecture redesign
introduced as part of the NetBurst microarchitecture. The Trace Cache caches (i.e., remembers)
the decoded Ops from the instruction decoder. The Trace Cache is the Pentium 4 processors
primary or L1 Instruction cache. It is capable of delivering up to three instructions per clock
down the pipeline to the execution units. It has its own Branch Predictor, labeled here as BTB,
that steers the Trace Cache to its next instruction locations. The micro-code ROM in the block
below contains micro-code sequences for more complex operations like fault handling and
string moves. The addition of the Trace Cache is significant because it means that often, previ-
ously executed instruction streams (e.g., a loop after its first iteration) will not need to go
through the decoder stage, avoiding a possible stall.
The renaming stage maps logical IA-32 registers (see Figure 2) to the Pentium 4 proces-
sors deep physical register file. This abstraction allows for a larger number of physical
registers in the processor without changing the front end of the processor. The allocator stage
assigns all the necessary hardware buffers in the machine for Op execution, at which point
the Ops go into some queues to await scheduling.
The scheduling stage that follows determines when an instruction is ready to execute. An
instruction is ready to execute when both input operands (most instructions have two input
sources) are ready, registers and an execution unit of the required type are free, etc. The
216 Part 2: Vertex Shader Tricks
Figure 1: Pentium 4 processor architectural diagram
Team LRN
scheduler is kind of like a restaurant host, only seating a party when all the party members
have arrived and a table is free, thus ensuring maximum throughput of customers through the
restaurant.
Once an instruction is ready to execute, execution takes place in the integer or float-
ing-point units (depending on the type of instruction, of course). Having these many execution
units operating at once is what makes the processor so highly parallelized and able to execute a
number of integer and floating-point operations per clock cycle.
The Address Generation Units (labeled AGU) are used for doing loads and stores of data
to the L1 data cache (the last block in Figure 1). Since programs have many loads and stores,
having both a load port and a store port keeps this from being a bottleneck. The Pentium 4 pro-
cessor also has an on-chip L2 cache that stores both code and data, along with a fast system
bus to main memory.
In terms of maximizing performance of software running on this architecture, there are
several goals to keep in mind, such as minimizing mis-predicted branches, minimizing stalls,
maximizing cache usage, etc. Well talk about how the vertex shader compiler aims to achieve
some of these goals later on, but in order to understand that discussion, we must first look at
some architectural features aimed at enhancing multimedia performance.
Introduction to the Streaming SIMD Extensions
As mentioned earlier, the Pentium III processor introduced the Streaming SIMD Extensions
(better known as SSE), aimed at offering improved floating-point performance for multimedia
applications. The Pentium 4 processor furthered the enhancements of SSE with the introduc-
tion of the SSE2 instructions. Since the vertex shader compiler uses these, a basic
understanding of how these instructions operate will allow for a better understanding of how
the vertex shader compiler works.
SIMD instructions (Single Instruction, Multiple Data) allow for an operation to be issued
once and simultaneously performed on several pieces of data. This is extremely useful in mul-
timedia applications using 3D graphics, video or audio, where an operation may be performed
thousands of times on data (such as vertices, pixels, audio samples, etc.). Most of the SSE and
SSE2 instructions utilize the eight 128-bit wide XMM registers, pictured on the far right of
Figure 2.
These registers can be used in a number of ways. Depending on the instruction used, the data is
read as four 32-bit single-precision floating-point numbers, two 64-bit double-precision float-
ing-point numbers, eight 16-bit integers, or 16 single bytes. The following assembly example
shows two of these usages:
Optimizing Software Vertex Shaders 217
Figure 2: Pentium 4 processor
register sets
Team LRN
MULPS XMM1, XMM2 // Multiply 4 single precision
// floating-point numbers in
// XMM1 by those in XMM2, place
// results in XMM1.
MULPD XMM1, XMM2 // Multiply 2 double-precision
// floating-point numbers in
// XMM1 by those in XMM2, place
// results in XMM1
In order to understand how the vertex shader compiler works, a full understanding of the SSE
and SSE2 instruction set isnt necessary. However, those interested in learning how to use SSE
instructions in their own code may refer to Intels Developer Services web site
(www.intel.com/IDS).
While the full SSE/SSE2 instruction set isnt covered here, theres one more subject that
will give some insight into how the compiler functions: data arrangement.
Optimal Data Arrangement for SSE
Instruction Usage
Often, execution can be hampered by memory bandwidth. This sometimes occurs when data is
laid out in a format that is intuitive to the author of the code but can result in fragmented
access to memory, unnecessary data being loaded into the cache, etc. This is best illustrated
with an example.
Consider an array of vertices, each of which has of the following structure:
#define numverts 100
struct {
float x, y, z;
float nx, ny, nz ;
float u,v;
} myvertex;
myvertex mVerts[numverts];
In this array-of-structures (AOS) layout, the array of these vertices would sit in memory as
shown in Figure 3. The problem arises that as an operation is performed on all of the vertices,
say doing transformations, only some of the elements are used in that operation, yet an entire
cache line is loaded (anywhere from 32 to 128 bytes, depending on which processor is being
used). This means that some elements of the vertex structure are being loaded into the cache
and then not being used at all.
218 Part 2: Vertex Shader Tricks
Figure 3: Ineffective cache
line load with AOS data
layout
Team LRN
A better approach is to use a structure-of-arrays (SOA) layout, as illustrated below:
#define numverts 100
struct {
float x[numverts]; //x,x,x
float y[numverts]; //y,y,y
float z[numverts];
float nx[numverts];
float ny[numverts];
float nz[numverts];
float u[numverts];
float v[numverts];
} myvertexarray;
myvertexarray mVerts;
While this type of an approach to structuring data may not be as intuitive, it has a couple of
benefits. First, cache line loads consist of data that will be used for the given operation, with no
unnecessary data being loaded into the cache. Secondly, applying SSE to the code becomes
trivial.
Lets look at an example. Consider the following loop to do a simple animation to a vertex:
//assume variables of type float exist, called fTime and fVelocityX/Y/Z
for (n=0;n<numverts;n++)
{
mVerts.x[n] += fTime * fVelocityX;
mVerts.y[n] += fTime * fVelocityY;
mVerts.z[n] += fTime * fVelocityZ;
}
Modifying this to use SSE becomes a matter of iterating through the loop differently. In the
example below, the same loop has been modified slightly. Where we had our variables defined
as floats before, weve now used a C++ class f32Vec4, which is just a container for four floats
with overloaded operators that use the SSE instructions:
//assume variables of type f32Vec4 exist, called fTime and fVelocityX/Y/Z
for (n=0; n<numverts; n += 4)
{
(m128*)(mVerts.x[n]) += fTime * fVelocityX;
(m128*)(mVerts.y[n]) += fTime * fVelocityY;
(m128*)(mVerts.z[n]) += fTime * fVelocityZ;
}
The online training resources available at http://developer.intel.com/software/products/col-
lege/ia32/sse2/ contain more detail about how to use these classes and instructions. Weve just
used them here to illustrate the structure-of-arrays data layout.
Optimizing Software Vertex Shaders 219
Team LRN
How the Vertex Shader Compiler Works
The DirectX 8 API describes a dedicated virtual machine known as the vertex virtual machine,
or VVM, which does processing to a stream of incoming vertex data. This could be any kind of
processing that can be performed autonomously on one vertex (see the articles by Wolfgang
Engel in Part 1). The vertex shader mechanism was designed to run on dedicated hardware, but
as weve been discussing, a software implementation exists. Actually, several software imple-
mentations exist and are invoked based on what processor is detected at run time.
When an Intel Pentium III or Pentium 4 processor is detected and software vertex process-
ing is selected by the application, the vertex shader compiler is used. The compiler takes the
vertex shader program and declaration at shader creation time (done with the IDirect3D-
Device8::CreateVertexShader method) and compiles them to a sequence of Intel Architecture
instructions. The compiler attempts to generate the most optimal sequence of IA instructions to
emulate the functionality required. It does so by making use of the SSE and SSE2 instructions,
minimizing memory traffic, rescheduling the generated assembly instructions, etc.
After this sequence of instructions is generated, it exists until destroyed (done with the
IDirect3DDevice8::DeleteVertexShader method). Whenever any of the IDirect3DDevice8
::Draw methods are used after calling IDirect3DDevice8::SetVertexShader, execution will
use the IA-32 code sequence generated by the compiler.
When following this path, a number of things happen. First, incoming vertex data is
arranged into a structure-of-arrays format for more efficient processing. (As expected, there is
a cost to doing this arrangement.) Vertices are then processed four at a time, taking advantage
of the SIMD processing.
The way that the four-wide VVM registers are mapped to the SSE registers is by turning
them on their side, so to speak. Each VVM register is mapped to four SSE registers. This map-
ping forms a 16-float block that is essentially a four-vertex deep VVM register. The illustration
in Figure 4 shows what this looks like. The vertex shader then only needs to be executed one
quarter as many times, with each iteration simultaneously processing four vertices.
Vertex data is arranged back into the array-of-structures format on the way out to be sent on to
the rendering engine. This also incurs a performance penalty.
220 Part 2: Vertex Shader Tricks
Figure 4: VVM to SSE register
mapping
Team LRN
Performance Expectations
As mentioned above, arranging incoming data from array-of-structures to structure-of-arrays,
and then back again upon exit, incurs some performance penalty. However, this is countered by
the fact that the shader need only be executed one quarter as many times. Additionally, proces-
sor clock speeds tend to be much higher than most graphics hardware.
In the end, the performance benefits depend on the complexity of the shader (as a more
complex shader will amortize the cost of the data arrangement) and on how well the shader is
written. The shader can be written in a way that lets the compiler best do its job of generating
the assembly instructions.
Optimization Guidelines
The optimization guidelines listed below can help improve performance of vertex shaders run-
ning in software. What follows are some details on each one of the guidelines and why it can
affect performance.
Guidelines for optimizing software vertex shaders
Rule
1. Write only the results youll need.
2. Use macros whenever possible.
3. Squeeze dependency chains.
4. Write final arithmetic instructions directly to output registers.
5. Reuse the same temporary register if possible.
6. Dont implicitly saturate color and fog values.
7. When using exp and log functions, use the form with the lowest acceptable accuracy.
8. Avoid using the address register when possible.
9. If the address register must be used, try to order vertices in order of the address field.
10. Profile, profile, profile
Write Only the Results Youll Need
In some cases, only a component of a register is needed. For example, the post-transform
Z-value of a vertex might be used to calculate an alpha or fog value. As we discussed earlier,
each of the components of the VVM registers maps to an individual SSE register. Any instruc-
tion performed on the VVM register will be performed to each of the four SSE registers;
therefore, any unnecessary components of the VVM register that we can avoid using will save
at least a few instructions on one of the SSE registers.
Optimizing Software Vertex Shaders 221
Team LRN
Lets look at an example:
Vertex Shader Code SSE Instruction Cost (for four vertices)
1 mul r0, c8, r0 Four SSE multiplies, 4 to 8 moves
2 mul r0.z, c8.z, r0.z One SSE multiply, 1 to 2 moves
Use Macros Whenever Possible
Macros are traditionally thought of as a convenience more than an optimization technique.
Indeed, in C code, macros are just a shorthand representation of a code sequence. In this case,
however, they give the compiler some extraneous information about what a particular sequence
of instructions is doing. In the m4x4 matrix multiplication macro, for example, the compiler is
aware that the multiplies being executed will be followed by an add (since the matrix multiply
is four dot product operations). Since the compiler is aware of this, it can retain results in regis-
ters, rather than moving them out and back in again.
This is illustrated in the following code sequence:
Before After
dp4 r0.x, v0, c2
dp4 r0.y, v0, c3
dp4 r0.z, v0, c4
dp4 r0.w, v0, c5
add r1, c6,v0
dp3 r2, r1, r1
rsq r2, r2
mov oT0, v2
mul r1,r1,r2
dp3 r3, v1, r1
add r3, r3, c7
mov oD0,r3
mov oPos,r0
m4x4 r0, v0, c2
add r1, c6,v0
dp3 r2, r1, r1
rsq r2, r2
mov oT0, v2
mul r1,r1,r2
dp3 r3, v1, r1
add r3, r3, c7
mov oD0,r3
mov oPos,r0
Squeeze Dependency Chains
Since four SSE registers are used to emulate a VVM register, and there are a total of eight SSE
registers, data often gets moved in and out of registers to make room for the next instructions
operands.
In cases where an operations result is going to be used in another operation, performing
that operation soon afterward can often let the compiler use the registers containing the inter-
mediate result without having to move that intermediate result to memory and back.
This is illustrated in the following example. The 4x4 matrix multiply result is stored in r0,
and then r0 is not used again until the end of the code sequence, where it is moved into oPos.
Therefore, bumping the mov instruction up in the code sequence is possible and can let the
compiler save some instructions in the SSE code sequence.
Before After
m4x4 r0, v0, c2
add r1, c6,v0
dp3 r2, r1, r1
m4x4 r0, v0, c2
mov oPos,r0
add r1, c6,v0
222 Part 2: Vertex Shader Tricks
Team LRN
Before After
rsq r2, r2
mov oT0, v2
mul r1,r1,r2
dp3 r3, v1, r1
add r3, r3, c7
mov oD0,r3
mov oPos,r0
dp3 r2, r1, r1
rsq r2, r2
mul r1,r1,r2
dp3 r3, v1, r1
add r3, r3, c7
mov oD0,r3
mov oT0, v2
Write Final Arithmetic Instructions Directly to Output Registers
A keen eye may have looked at the previous code sequence and wondered why the result of the
matrix multiplication wasnt written directly to the oPos register. Indeed, this is another oppor-
tunity to squeeze a few cycles out of the code sequence, as shown below.
Before After
m4x4 r0, v0, c2
mov oPos,r0
add r1, c6,v0
dp3 r2, r1, r1
rsq r2, r2
mul r1,r1,r2
dp3 r3, v1, r1
add r3, r3, c7
mov oD0,r3
mov oT0, v2
m4x4 oPos, v0, c2
add r1, c6,v0
dp3 r2, r1, r1
rsq r2, r2
mul r1,r1,r2
dp3 r3, v1, r1
add r3, r3, c7
mov oD0,r3
mov oT0, v2
Reuse the Same Temporary Register If Possible
When a temp register is used, there is some work the compiler does in terms of allocating SSE
registers to map to the temp register and the like. Reusing the same temp registers can some-
times save the compiler some work in having to redo this.
Dont Implicitly Saturate Color and Fog Values
This is a simple rule to understand. The compiler saturates color and fog values. Any work
done to saturate them by hand is wasted cycles.
Use the Lowest Acceptable Accuracy with exp and log Functions
The expp and logp instructions offer a variable level of precision. The modifier applied to the
instruction determines the level of precision. As one would imagine, the lower the precision,
the lower the amount of work required by the code sequence the compiler generates.
Avoid Using the Address Register When Possible
The address register allows for the addressing of different constant registers based on the value
in the address register. This is a problem for the compiler generating the SIMD code, which is
Optimizing Software Vertex Shaders 223
Team LRN
trying to process four vertices at a time. Each of the four vertices being processed may end up
referencing a different constant register, in which case the compiler will have to revert to a sca-
lar operation, processing each vertex one at a time.
Try to Order Vertices
Of course, the address register is there for a reason. Some vertex shaders may require its use. If
thats the case, the next best thing is to order vertices in the order of the address field. Ordering
by address field can result in runs of vertices with the same address value. The compiler will
then use SIMD processing on blocks of four vertices with a shared address value.
Profile, Profile, Profile
Less of an optimization guideline than a bit of advice, it should go without saying that
optimizations need to be tested. The guidelines provided are a good start, but theres no substi-
tute for good testing! Test your optimization changes before and after and under a variety of
conditions and input data.
Testing the vertex shader performance can be a bit tricky. Some graphics vendors offer
drivers that output performance statistics, but some do not. The DirectX 8 software vertex
pipeline doesnt output performance statistics. The best place to start is to isolate the shader in
a test harness so that performance differences will be amplified, as the shader will comprise a
larger portion of what gets rendered in a frame.
A Detailed Example
The following vertex shader is
the one used in the sample
code provided on the compan-
ion CD with this book. Its
simply a combination of a cou-
ple of well-known shaders, the
matrix palette skinning shader
and the reflection/refraction
shader often used for soap
bubble type effects. Com-
bining these gives us a glass
character type effects as
shown in Figure 5. Addi-
tionally, a sine wave is
propagated along the charac-
ters skin to provide a rippling
effect.
224 Part 2: Vertex Shader Tricks
Figure 5: Skinned mesh soap bubble shader example. Vertex
shader is performing skinning, sinusoidal peturbation, reflection,
and refraction calculations for texture coordinates.
Team LRN
Heres the shader before and after optimization:
Before After
vs.1.1
; Section 1
; Matrix Palette Skinning
mov a0.x, v3.x
;first bone and normal
dp4 r0.x, v0, c[a0.x + 50]
dp4 r0.y, v0, c[a0.x + 51]
dp4 r0.z, v0, c[a0.x + 52]
dp4 r0.w, v0, c[a0.x + 53]
dp3 r1.x, v4, c[a0.x + 50]
dp3 r1.y, v4, c[a0.x + 51]
dp3 r1.z, v4, c[a0.x + 52]
mov r1.w, c4.y
;second bone and normal
mov a0.x, v3.y
dp4 r2.x, v0, c[a0.x + 50]
dp4 r2.y, v0, c[a0.x + 51]
dp4 r2.z, v0, c[a0.x + 52]
dp4 r2.w, v0, c[a0.x + 53]
dp3 r3.x, v4, c[a0.x + 50]
dp3 r3.y, v4, c[a0.x + 51]
dp3 r3.z, v4, c[a0.x + 52]
mov r3.w, c4.y
;blend between r0, r2
mul r0, r0, v1.x
mad r2, r2, v2.x, r0
mov r2.w, c4.w
;blend between r1, r3
mul r1, r1, v1.x
mad r3, r3, v2.x, r1
; r2 contains final pos
; r3 contains final normal
; Sine-wave calculation
; Transform vert to
; clip-space
dp4 r0.x, r2, c0
dp4 r0.y, r2, c1
dp4 r0.z, r2, c2
dp4 r0.w, r2, c3
; theta from distance & time
mov r1, r0 ; xyz
dp3 r1.x, r1, r1 ; d2
vs.1.1
; Section 1
; Matrix Palette Skinning
mov a0.x, v3.x
;first bone and normal
m4x3 r0.xyz, v0, c[a0.x + 50]
m3x3 r1.xyz, v4, c[a0.x + 50]
;second bone
mov a0.x, v3.y
m4x3 r2.xyz, v0, c[a0.x + 50]
m3x3 r3.xyz, v4, c[a0.x + 50]
;blend between bones r0, r2
mul r0.xyz, r0.xyz, v1.x
mad r2, r2.xyz, v2.x, r0.xyz
mov r2.w, c4.w
;blend between r1, r3
mul r1.xyz, r1.xyz, v1.x mad
r3, r3.xyz, v2.x, r1.xyz
mov r3.w, c4.y
; r2 contains final pos
; r3 contains final normal
; Sine-wave calculation
; Transform vert to
; clip-space
m4x4 r0, r2, c0
; theta from distance & time
mov r1.xyz, r0.xyz ; xyz
dp3 r1.x, r1, r1 ; d2
Optimizing Software Vertex Shaders 225
Team LRN
Before After
rsq r1.x, r1.x
rcp r1.x, r1.x ; d
mul r1, r1, c4.x ; time
; Clamp theta to pi..pi
add r1.x, r1.x, c7.x
mul r1.x, r1.x, c7.y
frc r1.xy, r1.x
mul r1.x, r1.x, c7.z
add r1.x, r1.x, c7.x
; Compute 1
st
4 series values
mul r4.x, r1.x, r1.x ; d^2
mul r1.y, r1.x, r4.x ; d^3
mul r1.z, r4.x, r1.y ; d^5
mul r1.w, r4.x, r1.z ; d^7
mul r1, r1, c10 ; sin
dp4 r1.x, r1, c4.w
; Move vertex sin(x)
mad r0, r1.x, c7.w, r0
mov oPos, r0
; Reflection/Refraction calc
; Trans. vert and normal
; to world-space
dp4 r0.x, r2, c12
dp4 r0.y, r2, c13
dp4 r0.z, r2, c14
dp4 r0.w, r2, c15
dp3 r1.x, r3, c12
dp3 r1.y, r3, c13
dp3 r1.z, r3, c14
; re-normalize normal
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1, r1, r1.w
; Get eye to vertex vector
sub r4, r0, c5
dp3 r3.w, r4, r4
rsq r3.w, r3.w
mul r3, r4, r3.w
; Calculate E 2*(E dot N)*N
dp3 r4.x, r3, r1
add r4.x, r4.x, r4.x
mad oT0, r1, r4.x, r3
; Get refraction normal
mul r1, r1, c6
rsq r1.x, r1.x
rcp r1.x, r1.x ; d
mul r1.xyz, r1, c4.x ; time
; Clamp theta to pi..pi
add r1.x, r1.x, c7.x
mul r1.x, r1.x, c7.y
frc r1.xy, r1.x
mad r1.x, r1.x, c7.z, c7.x
; Compute 1
st
4 series values
mul r4.x, r1.x, r1.x ; d^2
mul r1.y, r1.x, r4.x ; d^3
mul r1.z, r4.x, r1.y ; d^5
mul r1.w, r4.x, r1.z ; d^7
dp4 r1.x, r1, c10 ; sin
; Move vertex sin(x)
mad oPos, r1.x, c7.w, r0
; Reflection/Refraction calc
; Trans. vert and normal
; to world-space
m4x4 r0, r2, c12
m3x3 r1, r3, c12
; re-normalize normal
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1, r1, r1.w
; Get eye to vertex vector
sub r4, r0, c5
dp3 r3.w, r4, r4
rsq r3.w, r3.w
mul r3, r4, r3.w
; Calculate E 2*(E dot N)*N
dp3 r4.x, r3, r1
add r4.x, r4.x, r4.x
mad oT0.xyz, r1.xyz, r4.x, r3.xyz
; Get refraction normal
mul r1.xyz, r1.xyz, c6.xyz
226 Part 2: Vertex Shader Tricks
Team LRN
Before After
; Calculate E 2*(E dot N)*N
dp3 r4.x, r3, r1
add r4.x, r4.x, r4.x
mad oT1, r1, r4.x, r3
; Calculate E 2*(E dot N)*N
dp3 r4.x, r3, r1
add r4.x, r4.x, r4.x
mad oT1.xyz, r1, r4.x, r3
A number of the optimization guidelines are used here, giving an improvement of about 15
percent. The vertex shader before optimization takes approximately 250 cycles per vertex.
After optimization, this is reduced to approximately 220 cycles per vertex. The effect on the
overall frame rate of the application will be less because there will usually be many more
things going on, and the vertex shader comprises only a portion of the total execution time.
Hopefully, the details provided here have given some insight into how the software vertex
shader compiler works and how to optimize code for it. The high performance available
through the software implementation, especially when properly optimized, means that develop-
ers can take advantage of vertex shaders today, even if their games have to run on computers
without vertex shader hardware.
Acknowledgments
Id like to recognize a couple of individuals who contributed to this paper. First off, a large
thanks is due Ronen Zohar at Intel. Ronen worked closely with Microsoft on optimizing the
DirectX run time and provided the optimization guidelines discussed in this article. Thanks as
well to William Damon at Intel, who worked on the sample application referenced in the
examples.
Optimizing Software Vertex Shaders 227
Team LRN
Compendium of Vertex
Shader Tricks
Scott Le Grand
Introduction
There are a surprising number of procedural effects one can express with vertex shaders. As
hardware speed increases, moving such effects from the CPU to the graphics hardware will
free up more CPU time and memory bandwidth for other computations. In addition, a vertex
shader-based approach to procedural rendering allows the use of static vertex buffers within
procedural effects. This allows one to offload both the computation and the data for a proce-
dural effect entirely into the graphics hardware.
However, the conversion of a procedural effect from a CPU-bound calculation to a vertex
shader requires some thought: There is no flow control within a vertex shader, and data cannot
be shared among the vertices. Fortunately, one can generate the equivalent of decision-based
bit masks and use those to emulate flow control. This article will illustrate some of the tricks
one can employ to move procedural effects into 3D hardware and provide a simple example of
doing so.
Periodic Time
Vertex animation requires updating the positions of an effects vertices over the time period of
the effect. In a vertex shader, one wants to use static vertex buffers in order to minimize AGP
bandwidth. So the first trick needed here is to calculate a periodic time value to drive vertex
animation from within the vertex shader itself. The use of this periodic time allows the vertices
to follow a calculated trajectory over the lifetime of an effect and then disappear at the end of
one cycle of the period. If the effect is cyclic, the periodic nature of the calculated time value
will allow each vertex to recycle itself and reappear as a seemingly new component of the
effect during the next cycle of the time period. Mathematically, for this task we need an
approximation to a fractional remainder function to calculate a fractional period. While the
DirectX 8 specification supplies just such a function in the frc macro, there is a simpler
approach. The y component of the destination register in an expp instruction returns the frac-
tional part of the w component from the input register.
To derive a periodic time value for a vertex, first supply each vertex with a phase . Next,
store the current absolute time, t, in a constant register and 1/period, , in another constant
228
Team LRN
register. The fractional period, f, of any vertex can be calculated by passing *(t ) into an
expp instruction as follows:
mov r1, c0.xy ; t is in c0.x, is in c0.y
sub r1.w, c0.x, v0.x ; is in v0.x
mul r1.w, r1.w, c0.y
expp r1.y, r1.w ; f is now in r1.y
Vertex phases can all be identical if you wish the vertices of the effect to move in tandem from
an origin (such as a shockwave from an explosion), random if you wish the effect to always
appear active (such as a particle fountain), or in groups of randomly generated but identical
phases to generate clumps of vertices that are always at the same timepoint in their periodic
trajectories (such as the rings emerging along the path of a railgun shot in Quake).
One-Shot Effect
A variant on the above involves skipping the expp instruction altogether, setting the phases of
all the vertices to the same time as the creation of the effect, and only rendering the vertices up
to the expiration of the time period. This is useful for a one-shot effect such as an explosion.
Random Numbers
While one can supply pregenerated random numbers as a vertex attribute, one will need to gen-
erate a new random number for every cycle of a periodic effects period, or else the effect
could look repetitive without an enormous number of vertices. Fortunately, one can use a sim-
ple equation (a linear congruence) and the aforementioned expp instruction to generate a
pseudorandom number from a supplied seed (such as the phase used above if space is tight and
its random itself) and the integer part of the periodic time from above. To calculate a random
number, store T, the integer part of the periodic time in a constant register, then pass r=T into
an expp instruction:
mov r1.w, c0.z ; T is in c0.z
mul r1.w, r1.w, v0.x ;
expp r2.y, r1.w ; fractional time is in r2
sub r1.w, r1.w, r2.y ; a unit pseudo-random number is in r1.w
Such a random number will remain constant throughout a single period of an effect. This is
useful for initializing a vertexs state. Since no calculated data is retained between invocations
of a vertex shader, this pseudorandom value will be calculated upon every transformation of
each vertex.
The above code fragment also demonstrates how to manipulate the vertex shader input and
output masks. These masks are the register component (x, y, z, and w) sequences following the
period in each input and output register in the above code fragment. While the expp instruction
normally calculates four output values, we can indicate our interest in just the w component
output by leaving out x, y, and z from the output register. The omission of unnecessary compo-
nents of output registers can improve vertex shader performance, especially on machines that
emulate vertex shaders in software. For input registers only, their four components can be
Compendium of Vertex Shader Tricks 229
Team LRN
collectively negated, and each component can be swizzled: replicated or permuted in any
desired combination.
Flow Control
While there are no flow control instructions in the DirectX 8 vertex shader language, there are
two instructions, sgt and sle, which will perform a greater than, equal, or a less than compari-
son between a pair of source registers and then set all of a destination registers components
either to 1.0, if the comparison is true, or 0.0, if it is false. One can use a series of such tests to
conditionally add in multiple components of an effects trajectory.
Cross Products
When either constant or vertex attribute space is tight, one can use two instructions to calculate
the cross product of two vectors in order to generate a vector orthogonal to an existing pair of
normalized vectors. This is also handy for generating the third row of an orthonormal transfor-
mation matrix given only its first two rows. The cross product of two three-dimensional
vectors v
1
and v
2
is defined as the determinant:

i x x
j y y
k z z
x y y x i z x x z j x y y x k
1 2
1 2
1 2
1 2 1 2 1 2 1 2 1 2 1 2
- + - + -
This sequence of two multiplies and one subtraction per vector component can be represented
by:
mul r1.xzx, r2.yxy, r3.xyz
mad r1.yxy, r2,xzx, r3.xyz
where v
1
is stored in r1 and v
2
is stored in r2. In this case, we have used the vertex mask swiz-
zler on registers r1 and r2, along with negation of register r1 to reproduce the input terms of
equation 1.
Examples
The first example for this article on the companion CD shows how to use periodic time to sim-
ulate and render a particle system entirely within a vertex shader. Each particle requires only
four floating-point values. The first three of these values represent a three-dimensional heading
vector, which the particle would follow in the absence of gravity. The fourth floating-point
value is s the phase, which is used to ensure the particle system effect is always in
mid-stream, no matter when it is initiated. This shader only requires five constant registers,
four of which hold the total transformation matrix. The fourth constant register contains, in
order of component, the current time t, normalized by dividing it by T
p
, the desired lifetime of
a particle, followed by the launch speed of all particles, after which comes the total gravita-
tional acceleration experienced by a particle during its lifetime, gT
p
2
, where g is the
230 Part 2: Vertex Shader Tricks
Team LRN
gravitational acceleration, and finally, a constant, 1.0, which is used to convert the
three-dimensional particle position into a four-dimensional homogeneous coordinate.
The second example makes more extensive use of random numbers to create a series of
recurring procedural explosions similar to firework bursts. As in the first example, each parti-
cle has a phase value , which is used here to stagger the times of detonation of each
individual explosion. However, there are two important differences here from the previous
effect. First, each is shared by a group of particles so that the group animates synchronously.
Second, there are three additional shared phase values for each of these particle groups, one for
each spatial dimension. These three phases are used to pseudorandomize the position of each
explosion. They each vary in range from 0.5 to 1.0. With each recurring explosion, the position
of a particle is shifted in each dimension by the respective phase within a periodic box. The
periodicity of the box is enforced, not surprisingly, by the use of the expp instruction.
Summary
Vertex shaders can perform much of the bookkeeping for special effects that is currently han-
dled by the CPU. Converting a CPU-bound effect into a vertex shader effect requires thinking
of the effect as either periodic or turning the effect on and off externally. The advantage of han-
dling the effect inside a vertex shader is that the graphics hardware can render it entirely from
static, pregenerated vertex buffers. The disadvantage is that if geometry bandwidth is tight,
converting an effect to a vertex shader may slow things down. As with most processes, the
only way to find out is to try each way.
Compendium of Vertex Shader Tricks 231
Team LRN
Perlin Noise and
Returning Results from
Shader Programs
Steven Riddle and Oliver C. Zecha
Limitations of Shaders
The hardware-accelerated programmable shader is the most important advance in computer
graphics today. Although early pixel shaders leave a few things to be desired, vertex shaders
are proving to be a robust and versatile solution to many graphics problems. Together, pixel
and vertex shaders allow a degree of realism on screen unprecedented in real-time animation.
From particle systems to geometric deformations, their scope of application seems endless.
The main reason for this is their vector processing capabilities, which allow complex lighting
models and other visually stunning effects.
Currently, if you do not have a graphics card supporting vertex shaders, there are reason-
able implementations that use the CPU to execute the vertex shader program. (The same is true
for pixel shaders, but because the CPU must do all the texturing, it is painfully slow and ren-
ders the graphics processor useless.) However, at the rate with which graphics technology
advances these days, it is difficult to believe that the CPU will be able to keep up with the GPU
in the immediate future. This is due to the nature of the execution units.
A shader program does not allow branching or looping, and consequently, the GPU does
not need to implement the hardware to make these possible. This fact alone allows the GPU to
run much faster. Besides this, a GPU does not allow access to the data of other vertices, so they
may be processed in parallel. Although CPU manufacturers have developed competing tech-
nologies to operate on multiple data with a single instruction, the GPU still has the advantage
when working with larger sets of data. This becomes especially apparent when one considers
all the other work the CPU must do, such as AI, collision detection, or whatever else your
application requires.
Very soon, the number of vertices in average scenes will be so astronomically huge that
performing the math to transform and light them on the CPU will be futile. Couple this with
the memory bandwidth requirements for large meshes, and it becomes clear that the CPUs
cycles are better spent on tasks other than vertex processing.
This leaves us with the vector math capabilities of the GPU. It can compute more vector
operations faster than any PC hardware before it, and the numbers just keep getting bigger and
232
Team LRN
better. Already, the number of vector operations feasible on a GPU exceeds that of a CPU, and
there are no signs of going back. In fact, the trend seems to be that GPUs are doubling in
power every six months, a full three times faster than Moores Law predicts. Therefore, it
seems reasonable that as much graphic computation as possible should be performed using the
GPU. Not only this, but the GPU should be used for as much vector processing as possible.
What this really means, in terms of code, is that as much vector and graphic computation
as achievable should be done using a programmable shader. Here we encounter the first major
limitation of shaders, namely that we may use, at most, 128 instructions in the vertex pipeline
and eight instructions plus four addressing instructions in the pixel pipeline. (Version 1.4 pixel
shaders have increased to six addressing instructions but kept eight regular instructions in each
phase; however, they allow two phases of 14 (8+6) instructions each.) Well, you may
think, even if we are allowed only 140 instructions, at least the instructions themselves are
relatively complicated mathematical operations, and using them we can do all sorts of neat
things. Nevertheless, as graphics programmers, we are the type of people who like to push the
limits, and so we ask, What if we need more? After all, 640K turned out to not be enough for
everybody!
We have always been fascinated by computer graphics and physical simulation, so you can
imagine how excited we were when given the opportunity to harness so much power for
on-screen imaging and vector processing! One of the first applications we saw for vertex
shaders was procedural deformation on a plane. We immediately noticed huge implications for
landscapes and other generated meshes, but something was missing: There was no way of
knowing the positions of the deformed vertices. As far as we were concerned, the objects had
not changed. This presents a whole host of problems. Even doing collision detection with
objects that have been transformed requires knowledge of the transformed vertices. In fact, it
seems that there are many situations where getting results from the shader calculations is use-
ful. Shortly after this, we discovered to our horror that the shader program outputs are write
only. This marks the second major limitation of shaders: the difficulty in returning values.
The first identified limitation, the number of instructions, has been dealt with to some
extent in several documents. The general idea, with which you are certainly familiar by now, is
to utilize the costless modifiers. These consist of swizzle, mask, and negation for the vertex
shader, a whole host of instruction modifiers, and instruction pairing for the pixel shader. This
will allow you to group instructions for speed and possibly allow you to reduce the number of
instructions used. Nevertheless, the gains you can make using these methods are minimal, in
terms of decreasing the number of instructions used. The ideal solution would be to run two
vertex shader programs sequentially, the second processing the output of the first. Unfortu-
nately, this idea was not incorporated into the shader specification, which means that the best
we can do is reduce the problem to getting output from the shader program. Interestingly
enough, this coincides perfectly with the second limitation outlined earlier.
Finally, we have only one solution to discover, and fortunately it exists. Although we can-
not get results directly from the shader program, they do give a sort of result. Generally, shader
programs produce rendered output in memory, usually in video memory for display on screen.
However, it is possible to render to texture. This is the technique used to return values from the
shader program. The trick is to arrange everything so that the texture to which we render is
used as an array of results for our computations. This may seem a bit strange at first, but the
implications are vast. The ability to get values back from the shader pipeline lets us implement
Perlin Noise and Returning Results from Shader Programs 233
Team LRN
a variety of computation-intensive algorithms as fast shader programs which can run on the
GPU.
In order to accomplish this, we need to add an extra piece of information to each vertex.
This extra information represents the texture coordinate where the result will be stored. In
order to justify moving this extra data across the AGP bus, we must be certain that the compu-
tations performed on each vertex are complex enough that doing them on the GPU is
worthwhile. In this article, we use the vertex shader to generate three-dimensional Perlin noise.
This is very intensive, since for each Perlin value, we need to do three cubic evaluations, eight
dot products, and seven linear interpolations, plus a number of moduli, multiplications, and
additions.
Perlin Noise and Fractional Brownian Motion
In general, computer graphics tend to have hard edges and appear too perfect to look real. To
correct for this, noise is often added to give the impression of some blemishes and randomness.
This noise may be added to lines to make them sketchy, textures to give them variation, char-
acter animation to make it more natural, or geometry to add some imperfections. In fact, there
are countless applications of noise in computer imaging. Nevertheless, most noise is too ran-
dom to look natural, unless you need a texture for TV static. For this reason, it is common to
perform some smoothing on the noise, such as a Gaussian blur. There is a problem with this
technique, however, when we consider real-time graphics. A Gaussian blur is quite slow. For
this reason, Perlin noise was developed to approximate this smoothed noise.
Perlin noise, named for its creator Ken Perlin, is a technique for generating smooth
pseudorandom noise. Smooth in this case means that the noise is defined at all real points in
the domain. Pseudorandom means that it appears random, but is in fact a function. The most
important feature of Perlin noise is that it is repeatable. This ensures that for any given point,
the noise will be the same each time it is calculated. A Perlin noise function takes a value in
n
and returns a value in . In simpler terms, this means that it takes an n-dimensional point with
real coordinates as input, and the result is a real number.
In order to produce smooth pseudorandom
noise, we need a random number generator and
a function for interpolation. The random num-
ber generator is used as a noise function, and
the interpolation function is used to produce a
continuous function from the noise.
Figure 1 shows an example of a
one-dimensional random function. Given an
integer, it generates a pseudorandom real num-
ber between zero and one. Each time it is used
to compute a value for a given point, it gives us
the same pseudorandom number, so it is repeat-
able. This is our noise function.
234 Part 2: Vertex Shader Tricks
Figure 1: Random function
Team LRN
In Figure 2, you can see the results of cubic
interpolation on our one-dimensional noise func-
tion. For the moment, dont worry about how
this is done; Ill explain that later. This function
is repeatable like the noise. However, unlike the
noise function, this function is continuous, and
consequently, it can be evaluated at any real
point in the x domain. We call this smooth func-
tion our noise wave.
It is sometimes helpful to think of the noise
wave as a simple wave function, such as sine or
cosine. You can see this analogy illustrated in
Figure 3. The amplitude of a simple wave is the
distance from valley to peak. Similarly, the
amplitude (A) of the noise wave is the maximum
possible distance from valley to peak (this is the
range of the noise function). The wavelength (l)
is taken as the x distance between consecutive
points. Since we know the wavelength (l), we
can calculate the frequency (f) by the following:
f
1

Frequency is one way to describe Perlin noise

waves and compare them with one another.
When one noise wave has a frequency that is
twice the frequency of another, its frequency is
said to be one octave higher. This convention is borrowed from music, where notes exhibit
the same property; if a note is one octave higher than another, it will have a frequency twice
that of the lower note.
By varying the frequency and amplitude, we can create different sorts of Perlin noise. We
scale the x-axis to alter the frequency; fewer points correspond to lower frequency. Addi-
tionally, dividing the output of the random number generator controls amplitude; larger
divisors correspond to smaller amplitude. Below are some smooth noise functions; their range
has been scaled to [0.5,0.5]. In Figure 4, you can see the sum of the first five noise functions
from Table 1.
Perlin Noise and Returning Results from Shader Programs 235
Figure 2: Cubic interpolation
Figure 3: Noise wave
Team LRN
Table 1: Perlin noise
Frequency: 1
Amplitude: 1
Frequency: 2
Amplitude: 1/2
Frequency: 4
Amplitude: 1/4
Frequency: 8
Amplitude: 1/8
Frequency: 16
Amplitude: 1/16
Figure 4: Fractional
Brownian motion
The intricate composite noise function in Figure 4 is called fractional Brownian motion (fBm),
a fractal conceived by Ken Musgrave. The sum is simply a superposition of all the Perlin
noise; this way, each waves features are represented in the result. Summing successive octaves
of Perlin noise is the most common approach to producing fBm noise. At each stage, the
amplitude is divided by two to reduce the effect of the higher frequency. Benoit Mandelbrot
coined the term persistence to describe the high-frequency component of a fractal. Low per-
sistence leads to more high-frequency noise in the final product. In the case of our fractal, the
persistence (p) is equivalent to the amplitudes divisor. Selecting a persistence value other than
two is uncommon but can be used to create interesting effects. Reducing or increasing the per-
sistence will roughen or smooth the noise, respectively. We have a similar value for calculating
frequency; the harmonic interval (h) is the ratio of frequencies for consecutive octaves. Values
between 1.0 and 2.0 really accentuate the fractal nature of the noise.
If n represents the n
th
octave, we can describe its frequency () and amplitude (A) as func-
tions of the harmonic interval (h) and persistence (p):
f h
n
A
p
n

1
So far, all the noise functions we have seen are one-dimensional. The input is a real coordinate
in one dimension, and the output is a real number. One-dimensional noise has its uses, but they
are limited. Two-dimensional noise, on the other hand, has more applications since it can be
used to texture surfaces or generate height maps. Two-dimensional noise takes a two-dimen-
sional point with real coordinates as input and outputs a real number. Table 2 shows some
examples of two-dimensional Perlin noise. Larger values are represented by darker regions and
smaller values by lighter ones. The first row shows five octaves of noise with a persistence of
236 Part 2: Vertex Shader Tricks
Team LRN
one, so the noise functions are clearly visible. The second row shows the same five octaves of
noise with a persistence of two; this is how they are added together.
Table 2: Two-dimensional Perlin noise
Persistence = 1
Persistence = 2
In Figure 5, you can see the fBm noise produced by the sum of the five Perlin noise waves
from the second row of Table 2. Each Perlin noise wave still contributes some features to the
sum just like one-dimensional fBm noise; however, we
now use two dimensions to locate a given point. Since we
are now working with multiple dimensions, it makes more
sense if we think of fBm noise as taking a vector for input
and producing a pseudorandom scalar as output. This
thinking can be applied more easily to three- and
higher-dimensional noise, which we will discuss shortly.
One important issue that has not yet been addressed is
how many octaves of Perlin noise we should use to gener-
ate the fBm noise. Obviously, with a persistence of two,
the amplitude of the Perlin noise quickly approaches zero.
For this special case, Ken Musgrave suggests using no
more than n octaves for noise with a resolution of res pix-
els, where n is given by:

n res - log
2
2
This means that for noise with a 256 x 256 resolution, the recommendation is to use six
octaves. Nevertheless, if you are using persistence values other than two or are concerned with
speed, this rule of thumb may not apply. Ultimately, youre the judge. However, if the noise
fluctuates faster than once about every two pixels, its features become too small to see.
Finally, before getting on to all the fun math stuff, we will take a brief look at three- and
higher-dimensional fBm functions. This is where fBm becomes useful for solids and clouds.
With three-dimensional noise, we can essentially carve out solid objects, which will still have
continuous textures if cut in half. This technique is frequently used to simulate marble and
looks quite good. With higher-dimensional noise, if we use one dimension to represent time,
the object shows a beautiful morphing effect. This last technique is commonly used for cloud
animation and looks incredible.
Perlin Noise and Returning Results from Shader Programs 237
Figure 5: Two-dimensional fBm
Team LRN
Table 3 shows how a marble texture would be generated using three-dimensional fractional
Brownian motion for a cube and how a hole can be cut out of it. (These images were generated
using Justin Legakis Computer Graphics Gallery Marble Applet, which you can play with at
http://graphics.lcs.mit.edu/~legakis/MarbleApplet/marbleapplet.html.) We do not bother to
show how the fBm is composed of Perlin noise because by now you understand this already.
What we do show is a form of procedural texturing.
Table 3: Volumetric textures
This is the base texture,
generated from the function
f(x,y,z)=sin(x) evaluated at
each point on the surface.
This is our fBm noise function
f(x,y,z)=noise also evaluated at
each point on the surface.
This is the final marble texture,
which is generated by adding
the base texture and the fBm
noise: f(x,y,z)=sin(x)+noise.
In our code (based on the Microsoft vertex shader sample), there are numerous examples out-
lining the ideas behind procedural textures using noise. The texture generator project
(texturegenerator\ textureVS.dsw) on the companion CD shows two-dimensional cross sec-
tions of three-dimensional noise. In these particular cases, we do not use a base texture,
although this is also common practice. Instead, for some of the algorithms, we use a function
as a base, similar to the marble in Table 3. Other cases generate textures using only noise, such
as the following:
The manner in which our Perlin noise is returned is an important aspect of how we generate
textures from it. Since the only outputs we are allowed are pixels, we must encode our results
in colors. Our positive results are stored in the blue component, while the negative results are
in the green component. Doing this allows us to return a greater range of values, since we
238 Part 2: Vertex Shader Tricks
Figure 6: Textures using noise. From left: stone, clouds, marble, and wood
Team LRN
effectively double the number of bits. In addition, the vertices are arranged in a grid, so finding
a value corresponding to a vertex is just like looking up the value in a table by using (x, y) to
locate it.
The stone texture is simply lit fractional Brownian motion. Each layer is an octave of
Perlin noise with a persistence of two. Cloud is also a simple form of fBm. Each layer is again
an octave of Perlin noise with a persistence of two, except this time the absolute value of the
noise is taken before summation. This has the effect of adding sharp, defined edges and cusps.
It does not yet look much like clouds because it still needs to drop off to zero faster. This will
make the cloud less dense and more realistic. The cloud was left at its maximum size so that
the divisions would be clearer. Since 1984, this formation has been called turbulence due to
its flow-like appearance.
A good marble texture is much more complicated to generate. Marble is patchy in some
areas and smooth in others. For this reason, we approached it slightly differently than the con-
ventional method outlines in Table 3. For the patchy section, we calculate sin(positive noise
&& negative noise + x). For the smooth regions, we calculate cos(positive noise negative
noise + x) instead. To put these together, we check the value of the noise at each point. If that
value is greater than 10, the patchy equation is used; otherwise, the smooth equation is used.
Wood is very similar to the smooth part of marble, except we add a small modifier in the
sine equation: sin(positive noise negative noise + smallconstant * (positive noise | negative
noise) + x). The bitwise OR operation provides the imperfections in wood, and smallconstant
ensures that their effect is on a small scale. The final thing to do once the texture has the cor-
rect feel is to adjust the color until it looks right. For the wood, we used a gradient from
darker brown to lighter brown to simulate the appearance of grain in natural wood.
Another included project is a planet with a dynamic level of detail (planet\planetVS.dsw).
Here you can see how noise is applied as a height map to a sphere, giving it a realistic terrain
look. We calculate the value of the Perlin noise at each vertex in the mesh using the Perlin ver-
tex shader. Once again, we are doing all the drawing using a separate vertex shader. Our
technique for returning values is important to consider. Although it is not the ideal solution for
all applications, it allows us to use the shader pipeline in very creative and unusual ways.
Additionally, it allows us the versatility to implement both programs using the same shader.
The planet is an excellent example of the detail level possible using fBm. As you zoom in,
the algorithm splits triangles and as a result, adds vertices. The main drawback to this tech-
nique is that we call the vertex shader for any vertex that is added. Ideally, we would put each
vertex in a queue and process the queue every time it is filled. This would minimize the pro-
cessing time required, since the vertex shader is a parallel unit and performs better with larger
datasets.
In both pieces of example code, we use the same vertex shader. This is done to illustrate
how flexible Perlin noise can be. Some performance improvements are possible if a specialized
shader is written specifically for each application. Even so, if collision detection with a
deformed surface is important, you will need to know the positions of the translated vertices.
Perlin Noise and Returning Results from Shader Programs 239
Team LRN
Now lets get into the math of this whole problem. First, it is important to understand that we
are trying to approximate cubic interpolation. Cubic interpolation returns a number between
two known values, given four known values (two before and two after the desired point). We
will label these heights V
1
, V
2
, V
3
, V
4
, in order from left to right, and assume that the x dis-
tance between consecutive V values is always 1. If we let x
1
represent the distance of the
desired point from V
2
(where 0 means we are right on top of V
2
, and 1 means we are right on
top of V
3
), we can calculate the height at A, as follows:
P = ( V
4
V
3
)( V
1
V
2
)
Q = 2( V
1
V
2
)( V
4
V
3
)
R = V
3
V
1
S = V
2
A = Px
1
3
+ Qx
1
2
+ Rx
1
+ S
Although the value at A will be exact if the function f(x) is cubic, this computation is incredi-
bly costly. For this reason, we use linear interpolation instead. In order to linearly interpolate,
we only need to know two values, V
1
and V
2
, on either side of A. Consequently, x
1
now repre-
sents the distance of the desired point from V
1
. A is the value produced by the linear
interpolation instead of the desired f(x), where x represents the x value of the point in question.
A = V
1
*( 1x
1
) + V
2
*x
1
= V
1
x
1
*( V
2
V
1
) f(x)
As you can see, this is much simpler and involves only a fraction of the computations required
to interpolate cubically. The drawback is that linear interpolation does not produce as good of a
result as cubic interpolation. In order to minimize this problem, we use a cubic filter prior to
interpolation. The cubic filter that we use is called the ease curve, shown in Figure 9.
240 Part 2: Vertex Shader Tricks
Figure 7: Cubic interpolation
Figure 8: Linear interpolation
Team LRN
The ease curve brings us from 0 to 1 with a very gradual change in slope and is given by the
function e(x) = 3x
2
2x
3
on the interval 0 x 1. In order to utilize this function as a filter, we
compute its value using x
1
. This results in e(x
1
), which is approximately equal to x
1
for values
of x
1
close to . If x
1
is close to either 0 or 1, e(x
1
) will be even closer to 0 or 1, respectively.
We then proceed to use e(x
1
) in place of x
1
in the linear interpolation. The effects of applying
this filter prior to interpolation are shown in Figure 10.
e(x
1
) = 3x
1
2
2x
1
3
A = V
1
*( 1e(x
1
) ) + V
2
*e(x
1
)
= V
1
e(x
1
)*( V
2
V
1
) f(x)
This added computation is slightly more costly than the simple linear interpolation shown in
Figure 8 but produces much nicer results, as you can see in the comparison that follows. In
addition to this, as we add dimensions to our noise, we only need to calculate one ease filter
per axis. This added overhead is negligible relative to the exponential growth of the number of
interpolations. Figures 11 and 12 show the results of cubic interpolation and linear interpola-
tion, respectively. Figure 13 shows linear interpolation using the ease filter for comparison.
Notice the improvement in smoothness over linear interpolation.
Now that you have a firm grasp on the interpolation that is used, we can continue to the
generation of the actual Perlin noise. In order to compute the value of the noise at a given
point, we need to linearly interpolate on each axis. The exact number of linear interpolations
required is given by (2
d
)1, where d is the number of dimensions over which we want the
noise. You may wonder what it is that we are interpolating between; the answer is that we are
linearly interpolating between scalars, which represent the height at a given point. In
Perlin Noise and Returning Results from Shader Programs 241
Figure 9: Ease curve
Figure 10: Ease interpolation
Team LRN
one-dimensional noise, the scalar represents the height off the x-axis; in two-dimensional
noise, the scalar represents the height off the x-y plane; in three-dimensional noise, the scalar
represents the height out of the x-y-z space (into the fourth dimension). How we arrive at these
scalars is explained later; for the moment, just take it for granted that each point with whole
number coordinates has a scalar associated with it.
Figure 14 shows (x), which is the point at which we would like to calculate the height. n
1
and n
2
represent the scalars associated with the two closest neighbors having whole number
coordinates, and dx represents the distance from n
1
to x.
The scalar value at x is calculated in the following manner:
e(dx) = 3*dx
2
2*dx
3
(x) = n
1
e(dx)*( n
2
n
1
)
Figure 15 shows exactly the same procedure for two dimensions. Now (x,y) represents the
point for which we would like to calculate the height. n
1
, n
2
, n
3
, and n
4
represent the scalars
associated with the four closest neighbors having whole number coordinates, dx represents the
x distance from n
1
to (x,y), and dy represents the y distance from n
1
to (x,y). The two new
variables r
1
and r
2
represent the intermediate interpolated values.
242 Part 2: Vertex Shader Tricks
Figure 11: Cubic interpolation Figure 12: Linear interpolation
Figure 13: Linear interpolation with ease
filter
Figure 14: One-dimensional Perlin noise
Team LRN
The scalar value at (x,y) is calculated in the following manner:
e(dx) = 3*dx
2
2*dx
3
r
1
= n
1
e(dx)*( n
2
n
1
)
r
1
= n
3
e(dx)*( n
4
n
3
)
e(dy) = 3*dy
2
2*dy
3
(x,y) = r
1
e(dy)*( r
2
r
1
)
Finally, Figure 16 shows the same procedure again, for three dimensions. Now (x,y,z) repre-
sents the point for which we would like to calculate the height. n
1
, n
2
, n
3
, n
4
, n
5
, n
6
, n
7
, and n
8
represent the scalars associated with the eight closest neighbors having whole number coordi-
nates; dx, dy, and dz represent the x, y, and z distances respectively from n
1
to (x,y,z). Lastly,
the variables r
1
, r
2
, r
3
, r
4
, r
5
, and r
6
represent the intermediate interpolated values.
Perlin Noise and Returning Results from Shader Programs 243
Figure15: Two-dimensional Perlin noise
Figure 16: Three-dimensional Perlin noise
Team LRN
The scalar value at (x,y,z) is calculated in the following manner:
e(dx) = 3*dx
2
2*dx
3
r
1
= n
1
e(dx)*( n
2
n
1
)
r
2
= n
3
e(dx)*( n
4
n
3
)
r
3
= n
5
e(dx)*( n
6
n
5
)
r
4
= n
7
e(dx)*( n
8
n
7
)
e(dy) = 3*dy
2
2*dy
3
r
5
= r
1
e(dy)*( r
2
r
1
)
r
6
= r
3
e(dy)*( r
4
r
3
)
e(dz) = 3*dz
2
2*dz
3
(x,y,z) = r
5
e(dz)*( r
6
r
5
)
Extending this technique into the fourth dimension and beyond is left as an exercise for the
reader. As you may have guessed, it gets rather long but is not very hard.
Now, you lack only one piece of knowledge to make your own Perlin noise, and that is
how to calculate the scalars at the neighboring points. In the case of Perlin noise, this value is
found by taking the scalar product of the distance vectors and the gradients, as shown below.
Our examples only cover two dimensions since three dimensions make the diagrams overly
cluttered; however, the concepts are the same.
Figure 17 shows the distance vector from
each neighbor point to (x,y) in two dimensions.
We can take advantage of the fact that the neigh-
bor points are regularly spaced to find these
vectors easily. In our case, dx and dy are just the
fractional components of x and y, since each n
m
is a point with integer coordinates.
d
1
= (dx,dy)
d
2
= (dx1,dy)
d
3
= (dx,dy1)
d
4
= (dx1,dy1)
Next, we need to look up the values for the
pseudorandom gradients in a table of gradients,
gradient[t]. In order to avoid symmetry about an
axis, we use a shuffling table, shuffle[t], to fur-
ther randomize the gradients. The construction of
the shuffling table is rather simple; each element
contains a unique integer between zero and t,
where t is the number of elements in the table. The value of t is also equal to the number of
gradients. Building the table of gradients is equally easy. Select a number of random vectors
with magnitudes less than one (|gradient| < 1) and normalize each one. This produces a uni-
form distribution of pseudorandom directions. Figure 18 on the following page shows the four
neighbor points and the gradient associated with each.
Given the point n
m
, we know its coordinates, (i,j,k). We use these coordinates and the
shuffling table to choose a gradient g
t
for the point n
m
, in the following manner:
g
t
= gradient[ (k+shuffle[ (j+shuffle[ i ])%t ])%t ]
244 Part 2: Vertex Shader Tricks
Figure 17: Distance vectors
Team LRN
As you can see here, we do two shuffles, two moduli (represented by %), and three lookups.
The shuffle makes the interval over which we repeat large, and the moduli are used here to
ensure that we never try to read a value from outside the table.
We calculate the height by taking the scalar product of the gradient stored for each neigh-
bor, and the distance vector from the neighbor to the point for which we want the noise. We
could just as well use the magnitude of the cross product, since any continuous function will
work. However this involves more effort and is not beneficial, so we shall stick to the dot prod-
uct. The magnitudes of the vectors shown in Figure 19 represent the value of the scalar product
at each neighbor, hence also the height out of the x-y plane.
n
1
= d
1
g
1
n
2
= d
2
g
2
n
3
= d
3
g
3
n
4
= d
4
g
4
Now that we know the scalar value at each neighbor, we can continue to interpolate, as
described earlier, to find the scalar at any given point. For three-dimensional noise, we first
perform four interpolations in the x-axis between the eight original points, which gives us the
scalars at four new points. We then perform two interpolations in the y-axis, which results in
the scalars for two new points. Finally, we interpolate between these two scalars in the z-axis
to give us our final value, which is the scalar associated with the point in question. This pro-
cess is performed exactly as shown in Figure 16.
To make it explicitly clear, you are interpolating linearly between values that are depend-
ent on your position. As you move around in the domain, the scalars change smoothly and
continuously, due to the smooth and continuous change in the direction vectors.
This concludes our discussion of how Perlin noise works and how to construct fBm. Next,
we will illustrate the implemented algorithm in pseudocode, followed by a clarification of the
actual shader program.
Perlin Noise and Returning Results from Shader Programs 245
Figure 18: Pseudorandom gradients Figure 19: Scalar product of distance and
gradient shown as height
Team LRN
shuffle[2*(number of elements + 1)] is a table of integers
gradient[2*(number of elements + 1)][3] is a table of floats
Initialize Tables
for i=0 to (number of elements - 1)
shuffle[i] = i
do
for j=0 to 2 gradient[i][j] = random from 1.0 to 1.0
while magnitude(gradient[i][j]) > 1.0
normalize(gradient[i])
end for
for i=0 to (number of elements - 1)
temp = shuffle[i]
j = random from 0 and number of elements - 1
shuffle[i] = shuffle[j]
shuffle[j] = temp
end for
for i=0 to (number of elements + 2)
shuffle[number of elements + i] = shuffle[i];
for j=0 to 2 gradient[number of elements + i][j] = gradient[i][j];
end for
End Initialize Tables
The first for loop in this section stores sequential numbers from zero to the number of elements
less one in the shuffle table. Subsequently, another loop ensures that our table of
pseudorandom gradients has a uniform distribution. This is done by continuing to select ran-
dom points until one is found within the unit sphere. The vector from the origin to this point is
then normalized and stored in the gradient list.
The second for loop does nothing other than mix up the shuffling table. It does this by ran-
domly swapping values throughout the array. Randomizing these ensures that the shuffling
table has no pattern, and as a result, the noise will have a large period.
The last for loop in the Initialize Tables section simply duplicates both the shuffle and the
gradient tables. We do this so that we can avoid doing the two moduli when finding the gradi-
ent. This way, performing a lookup beyond the number of elements brings us back into the
appropriate range. In our code, this is all done in C++; it is only the following section that is
implemented as shader assembly.
The next section of pseudocode is implemented entirely in the vertex shader program,
since this part is doing the heavy calculations. The shader assembly program does not follow
the same sequence in order to make use of some optimizations; however, the main ideas are
preserved.
Noise3d(vec[3] is a point in 3 dimensions) returns a float
find 8 neighbor points
look up gradient for each neighbor point, using shuffle
compute distance vector for each neighbor point
compute dot product of distance and gradient for each neighbor
compute ease function for x,y,z (fractional components only)
interpolate 4 times in x using e(x)
interpolate 2 times in y using e(y)
interpolate 1 time in z using e(z)
return the value from the final interpolation
End Noise3d
246 Part 2: Vertex Shader Tricks
Team LRN
To find the neighbor points, we must know the truncated position vector. We copy this vector
and add one to each component. These two vectors provide the bounding box for the neighbor
points, and all the x,y,z data we need to find their gradients. Finding the distance vectors for
each neighbor point involves a similar technique. Beginning with the fractional component of
the position, we copy it and subtract one. This results in the two sets of x,y,z data that we need.
Once we know this, we can perform the dot products at each point. Finally, we get to the inter-
polation, for which we use the ease curve. The ease curve is simply calculated using the
fractional component of the position from earlier. Subsequently, we do seven linear interpola-
tions and arrive at a singular result.
Before we take a detailed look at the vertex shader program itself, we will build some of
the more basic code that we will require later. First, we need a modulus function. The follow-
ing is the only way we have found to do a modulus on the card, which unfortunately eats up
many instructions.

v v frc
v
v
v 0 1
0
1
1 mod =
The frc macroinstruction actually counts as three instructions. Additionally, it can only com-
pute the fractional components of x and y. Consequently, computing the fractional components
of x, y, and z costs us six instructions. Carrying out the operation for all four components is
done as follows:
mov r0.xyzw, v0.xyzw
mov r1.xyzw, v1.xyzw
rcp r2.x, v1.x
rcp r2.y, v1.y
rcp r2.z, v1.z
rcp r2.w, v1.w
mul r3.xyzw, r0.xyzw, r2.xyzw
frc r4.xy, r3,xy
mov r3.xy, r3.zw
frc r5.xy, r3.xy
mov r4.zw, r5.xy
mul r5.xyzw, r4.xyzw, r1.xyzw
Now r5 contains the result of v0 mod v1. We can simplify this slightly if we pass
1
1 v
in as v2,
as shown in the following code:
mov r0, v0
mov r1, v1
mov r2, v2
mul r3, r0, r2
frc r4.xy, r3,xy
mov r3.xy, r3.zw
frc r5.xy, r3.xy
mov r4.zw, r5.xy
mul r5, r4, r1
Perlin Noise and Returning Results from Shader Programs 247
Team LRN
We are required to do this modulus for each component of two vectors, the input vector and the
input vector plus (1,1,1). We are now ready for the main assembly, so lets review our
constants:
c[0].x = 0
c[0].y = 1
c[0].z = 2
c[0].w = 3
c[1].x = 1/42
c[1].y = 42
c[1].z = 1/frequency
c[1].w = frequency
c[2].x = x translation
c[2].y = y translation
c[2].z = z translation
c[2].w = 0.25
c[3-95].xyz gradient table
c[3-95].w shuffling table
v3 is the input position in Perlin space; this is the point for which to find the noise. v0 is the
texture space location; this is where the result of the noise function is written.
First, we calculate our Perlin space position after we have translated and adjusted the fre-
quency of the noise for this octave. Then we make a backup of it in r5.
// r0 = vertex, then translate and scale
mov r0, v3
add r0.xyz, r0.xyz, c[2].xyz
mul r0.xyz, r0.xyz, c[1].www
// r5 = r0
mov r5, r0
Next, we take the translated point in r0 and find its fractional component. The value in r0 is
then the value for calculating the cubic filter.
// r0 = frc(r5) = (dx,dy,dz)
mov r2.xy, r0.zw
frc r2.xy, r2.xy
frc r0.xy, r0.xy
mov r0.z, r2.x
Now we need to find the whole number component of the translated position by subtracting the
fractional component in r0 from the translated position in r5. We have a copy of r5 stored in r4,
to which we add one.
// trunc(r5) and r4 = r5
add r4.xyz, r5.xyz, r0.xyz
mov r5.xyz, r4.xyz
// r4 = r5 + (1,1,1,1)
add r4.xyz, r5.xyz, c[0].yyy
248 Part 2: Vertex Shader Tricks
Team LRN
The following are the moduli. They are found by dividing r4 and r5 by 42, finding the frac-
tional component of the result, and multiplying that fraction by 42 to find the remainder.
// r4 = r4 * 1/42
// r4 = frc(r4)
mul r4.xyz, r4.xyz, c[1].xxx
mov r2.xy, r4.zy
frc r4.xy, r4.xy
frc r2.xy, r2.xy
mov r4.z, r2.x
// r9 = r0+1 % 42
mul r9.xyz, r4.xyz, c[1].yyy
// r4 = r5
// r4 = r4 * 1/42
// r4 = frc(r4)
mov r4.xyz, r5.xyz
mul r4.xyz, r4.xyz, c[1].xxx
mov r2.xy, r4.zy
frc r4.xy, r4.xy
frc r2.xy, r2.xy
mov r4.z, r2.x
// r8 = r0 % 42
mul r8.xyz, r4.xyz, c[1].yyy
Now r8 and r9 contain all the information we need to find the whole number neighbors. r8 and
r9 contain all the possible x,y,z coordinates for the eight neighbors because r8 and r9 define the
bounding box of the neighbors. Next, we use a similar technique to find all the values we need
to construct our distance vectors. We subtract one from the fractional component of our
position.
// r1 = r01 = (dx1,dy1,dz1)
add r1.xyz, r0.xyz, c[0].yyy
Here we begin doing shuffle table lookups. First, we do a lookup for x and put the value in
r10.x, and then we do a lookup for x+1 and put the result in r10.y.
// x lookup
mov a0.x, r8.x
mov r10.x, c[a0.x+3].w
// x+1 lookup
mov a0.x, r9.x
mov r10.y, c[a0.x+3].w
We add the values from the last lookups to our y and y+1 components. We use the resulting
values for the next set of lookups into the table.
add r5.xy, r10.xy, r8.yy
add r6.xy, r10.xy, r9.yy
// x + y
mov a0.x, r5.x
mov r11.x, c[a0.x+3].w
// x+1 + y
mov a0.x, r5.y
mov r11.y, c[a0.x+3].w
// x + y+1
mov a0.x, r6.x
mov r11.z, c[a0.x+3].w
// x+1 + y+1
mov a0.x, r6.y
mov r11.w, c[a0.x+3].w
Perlin Noise and Returning Results from Shader Programs 249
Team LRN
We can now add the z and z+1 values to the results from the preceding lookups. Then, we use
these sums to look up the vectors for each neighbor in our table of gradients. Finally, we com-
pute the dot product of the gradient and distance vectors at each neighbor, carefully arranging
the results into r2 and r3 to ensure that the interpolation is as optimized as possible.
// + z
add r5.xyzw, r11.xyzw, r8.zzzz
// + z+1
add r6.xyzw, r11.xyzw, r9.zzzz
//lookup gradient, and compute dot product
//r2 = (n1,n5,n3,n7)
//r3 = (n2,n6,n4,n8)
mov a0.x, r5.x
dp3 r2.x, r0.xyz, c[a0.x+3].xyz
mov a0.x, r5.y
mov r0.w, r1.x
dp3 r3.x, r0.wyz, c[a0.x+3].xyz
mov a0.x, r5.z
mov r0.w, r1.y
dp3 r2.z, r0.xwz, c[a0.x+3].xyz
mov a0.x, r5.w
mov r1.w, r0.z
dp3 r3.z, r1.xyw, c[a0.x+3].xyz
mov a0.x, r6.x
mov r0.w, r1.z
dp3 r2.y, r0.xyw, c[a0.x+3].xyz
mov a0.x, r6.y
mov r1.w, r0.y
dp3 r3.y, r1.xwz, c[a0.x+3].xyz
mov a0.x, r6.z
mov r1.w, r0.x
dp3 r2.w, r1.wyz, c[a0.x+3].xyz
mov a0.x, r6.w
dp3 r3.w, r1.xyz, c[a0.x+3].xyz
It is almost time to do the linear interpolation, but first we need to calculate the ease function
for each axis. This is done using the fractional component of the position.
// r4 = r0 * r0 * (2*r0 +3) = e(r0)
mad r7.xyz, r0.xyz, c[0].zzz,
c[0].www
mul r7.xyz, r7.xyz, r0.xyz
mul r4.xyz, r7.xyz, r0.xyz
Now we do the linear interpolations using the ease value for each axis. There are four x inter-
polations, two y interpolations, and one z interpolation.
// Linear interpolate along X
add r5, r3, r2
mul r5, r5, r4.xxxx
add r6, r5, r2
// Linear interpolate along Y
add r5.xy, r6.zw, r6.xy
mul r5.xy, r5.xy, r4.yy
add r7.xy, r5.xy, r6.xy
// Linear interpolate along Z
250 Part 2: Vertex Shader Tricks
Team LRN
// r7.z = Noise(x,y,z)
add r5.x, r7.y, r7.x
mul r5.x, r5.x, r4.z
add r7.z, r5.x, r7.x
If Noise is positive, store the value in r7.y; otherwise, store Noise in r7.z. Then move zero
into r7.x and the alpha value into r7.w. The alpha value is used to compute the sum of the vari-
ous Perlin noise layers. As the frequency of the noise increases, we decrease its alpha value to
ensure the amplitude is scaled properly. Then we simply draw the layers we want to the same
texture space with alpha enabled.
// r7.x = 0
// r7.y = Noise if Noise is +ve, otherwise its 0
// r7.z = Noise if Noise is ve, otherwise its 0
// r7.w = 1/frequency (alpha)
mov r7.x, c[0].x
max r7.y, r7.z, c[0].x
max r7.z, r7.z, c[0].x
mov r7.w, c[1].z
The projected position is simply the x,y coordinates to which the results will be written in the
texture. Remember that this is how we get back our results, using the green and blue compo-
nents of the color for the positive and negative noise, respectively, and the alpha as our
amplitude.
// x position = xy position on texture
mov oPos, v0
// Calculate color intensity
mov oD0, r7.xyzw
Final Thoughts
Our vertex consists of two vectors in an FVF, one for Perlin position, and one for texture posi-
tion. This is acceptable, since the amount of data moving across the AGP bus is about 8
megabytes for a 128 x 128 mesh (around 16,384 points). This leaves space for 14 vectors of
additional input. This is good since it allows a fair bit of extra data to be included in each ver-
tex. Not only this, but no textures are used; these could be used for all sorts of data input. On
the downside, 93 instruction spaces were used in the vertex pipeline. This is actually quite
compressed, since we were able to do some major simplifications such as the linear interpola-
tion in x, so there are still 35 instruction spaces remaining. In addition to this, the power of the
pixel pipeline is completely untouched.
The following benchmarks were taken to give you some idea as to the effectiveness of the
GPU vs. CPU. They were taken using slightly modified code without any drawing to screen in
order to strictly test the Perlin throughput. The vertex programs run using an optimized vertex
pipeline when executing in software. This makes vertex shaders a very hardware-independent
method for taking advantage of streamlined CPU assembly instructions such as 3DNOW! and
SSE.
Perlin Noise and Returning Results from Shader Programs 251
Team LRN
Table 4: Shader benchmarks
Vertex shader running on CPU
fps CPU GPU
5 750MHz PIII Coppermine (SSE) GeForce256
11 1.2GHz Athlon Thunderbird (3DNOW!) GeForce2
Vertex shader running on GPU
fps CPU GPU
13 1.2GHz Athlon Thunderbird (3DNOW!) GeForce3
(estimated) 25+ 750MHz PIII Coppermine (SSE) GeForce4
The technique for returning values is slow and may seem questionable, but as GPUs become
faster, getting results back from them will become increasingly desirable. This, coupled with
the simple fact that there is no alternative method to do so, makes this trick not only good to
understand but very unique. In addition, returning values allows us a great deal of versatility in
using the code, as illustrated by the two radically different test programs. Nevertheless, there is
a major problem with the planet program, since it is not queued. Vertex shaders prefer bulk
data; hence it would be much more efficient if we saved a list of added vertices and computed
their noise values as a block.
Keep in mind that this is meant to be an experiment, or possibly a base for your own noise
explorations, and not a production implementation. The constraint of needing to know the
translated position, and thus the Noise value, for collision detection and surface interaction
requires us to use a very inefficient process of writing to texture memory and then reading
back from it. We have shown that this is possible and potentially useful; however, applica-
tion-specific implementations are desirable, and further optimization is required. There is much
room for improvement in the shader, particularly for texturing applications. Nonetheless, it
clearly illustrates the relative power of the GPU vs. the CPU. After all, with the current prod-
uct release cycle, it looks like we will see this single-pass Perlin noise vertex shader running at
100fps in about one year!
252 Part 2: Vertex Shader Tricks
Team LRN
References
Paul Bourke, Perlin Noise and Turbulence, http://astronomy.swin.edu.au/pbourke/
texture/perlin/, January 2000.
David S. Ebert, F. Kenton Musgrave, Darwyn Peachey, Ken Perlin, and Steven Worley,
Texturing and Modeling, A Procedural Approach, 2nd Ed. (AP Professional, Cambridge,
1998).
Hugo Elias, Perlin Noise, http://freespace.virgin.net/hugo.elias/models/m_perlin.htm.
John C. Hart, Perlin Noise Pixel Shaders, http://graphics.cs.uiuc.edu/~jch/papers/
pixelnoise.pdf, University of Illinois, Urbana-Champaign.
Microsoft Corporation, Vertex Shader Reference, http://msdn.microsoft.com/library/
default.asp?url=/library/en-us/dx8_c/directx_cpp/Graphics/Reference/Shader/Vertex/
Vertex.asp, 2002.
Sean ONeil, A Real-Time Procedural Universe, Part One: Generating Planetary Bodies,
http://www.gamasutra.com/features/20010302/oneil_01.htm, March 2001.
Ken Perlin, Noise and Turbulence, http://mrl.nyu.edu/~perlin/doc/oscar.html, MAKING
NOISE, based on a talk presented at GDCHardCore, December 1999.
Matt Zucker, The Perlin Noise Math FAQ, http://students.vassar.edu/mazucker/code/perlin-
noise-math-faq.html, February 2001.
Perlin Noise and Returning Results from Shader Programs 253
Team LRN
254
Blending Textures for Terrain
by Alex Vlachos
Image Processing with 1.4 Pixel Shaders in Direct3D
by Jason L. Mitchell
Hallo World Font Smoothing with Pixel Shaders
by Steffen Bendel
Emulating Geometry with Shaders Imposters
by Steffen Bendel
Smooth Lighting with ps.1.4
by Steffen Bendel
Per-Pixel Fresnel Term
by Chris Brennan
Diffuse Cube Mapping
by Chris Brennan
Accurate Reflections and Refractions by Adjusting for Object Distance
by Chris Brennan
UV Flipping Technique to Avoid Repetition
by Alex Vlachos
Photorealistic Faces with Vertex and Pixel Shaders
by Kenneth L. Hurley
Non-Photorealistic Rendering with Pixel and Vertex Shaders
by Drew Card and Jason L. Mitchell
Team LRN
255
Animated Grass with Pixel and Vertex Shaders
by John Isidoro and Drew Card
Texture Perturbation Effects
by John Isidoro, Guennadi Riguer, and Chris Brennan
Rendering Ocean Water
by John Isidoro, Alex Vlachos, and Chris Brennan
Rippling Reflective and Refractive Water
by Alex Vlachos, John Isidoro, and Chris Oat
Crystal/Candy Shader
by John Isidoro and David Gosselin
Bubble Shader
by John Isidoro and David Gosselin
Per-Pixel Strand-Based Anisotropic Lighting
by John Isidoro and Chris Brennan
A Non-Integer Power Function on the Pixel Shader
by Philippe Beaudoin and Juan Guardado
Bump Mapped BRDF Rendering
by dm Moravnszky
Real-Time Simulation and Rendering of Particle Flows
by Daniel Weiskopf and Matthias Hopf
Team LRN
Blending Textures for
Terrain
Alex Vlachos
Historically, artists have had problems texturing terrain. Hiding texture seams and repetitive
patterns is a difficult problem that most games suffer from. This shader shows how to program-
matically interpolate between two textures based on their normal. The interpolation provides
varying combinations of base textures when rendered as a combination of two textures. In the
following figure, a brown rocky texture is biased toward polygons that are generally facing
toward the sky. A strange blue texture is given to anything that is on an incline. For hills that
are in-between, a blended combination of the two textures is used.
Ideally, using the angle of the normal from the ground plane should be used as the interpolation
value to blend between the two textures. However, to simplify the math, an approximation can
be implemented that simply uses the z-component of each vertex normal. In the example
shown here, z
3
is used to get a sharper cross fade between the two terrain textures. When using
256
Team LRN
just z, the blending is less harsh, which provides a slightly different look. Experimenting with
different powers of z will be necessary to get the desired look given the surrounding scene.
The effect is achieved by computing the blend value in the vertex shader. The pixel shader
will, at every pixel, use the interpolated blend value to blend between the two textures. In
pre-shader hardware, these values can be precomputed and stored in the vertex. However, more
complex schemes can be implemented in a vertex shader if the desired effect would be to blend
four to six textures based on which direction the hill was facing. This technique could be used
to simulate snow falling from a given direction and sticking to the side of a mountain. Over
time, the intensity of the blend value could be increased to show the accumulation of snow.
A further enhancement could be to compute the blend value per-pixel instead of per-ver-
tex. This could provide some very interesting results when used with a bump map. When
computing the per-vertex values, the blend only varies linearly over the face of a polygon. A
bump map could provide more interesting variation within polygons.
Vertex shader:
vs.1.1
m4x4 oPos, v0, c0 //Transform vertex by transformation matrix
mov oT0, v7 //Texture coordinates for texture 0
mov oT1, v8 //Texture coordinates for texture 1
mul r0, v3.z, v3.z //z^2 into temp register
mul oT2, r0, v3.z //z^3 into third set of texture coordinates
Pixel shader:
ps.1.4
texld r0, t0 //Base texture 0
texld r1, t1 //Base texture 1
texcrd r2.rgb, t2 //Interpolation value
lrp r0, r2.r, r1, r0 //Blend between the two textures
//Lighting computations go here
Blending Textures for Terrain 257
Team LRN
Image Processing with
1.4 Pixel Shaders in
Direct3D
Jason L. Mitchell
Introduction
With the ability to perform up to 16 ALU operations and 12 texture accesses in a single pass,
Direct3Ds pixel pipeline has begun to look like a powerful programmable 2D image process-
ing machine. In this article, we will outline several techniques that utilize DirectX 8.1s 1.4
pixel shaders to perform 2D image processing effects such as blurs, edge detection, transfer
functions, and morphology, and will point out applications of these 2D image operations to 3D
game graphics. For an introduction to the structure and capabilities of 1.4 pixel shaders, please
refer back to Wolfgang Engels introductions to vertex and pixel shader programming.
As shown here and in Non-Photorealistic Rendering with the Pixel and Vertex Shaders,
post-processing of 3D frames and a general image-savvy mindset are fundamental to effec-
tively utilizing 3D graphics processors to generate a variety of photorealistic and non-photo-
realistic effects [Saito90]. As shown in Figure 1, filtering an image with a GPU is done by
using the image as a texture and drawing a screen-aligned quadrilateral into the render target
(either the back buffer or another texture).
This results in the pixels of the render target being filled with the results of the pixel shader.
With a very simple pixel shader that samples the source image and passes it directly to the des-
tination without modification, this is equivalent to a Blt. In the simple Blt case, one source
pixel determines the color of one destination pixel. In addition to merely copying the input
image data, this operation can be useful for creating black-and-white or sepia tone images from
the source image. We will illustrate these simple effects and more general transfer functions
(like a heat signature) in the first part of this article.
258
Figure 1: Using a pixel shader for
image processing by rendering from
one image to another
Team LRN
Image processing becomes especially powerful when the color of the destination pixel is
the result of computations done on multiple pixels sampled from the source image. In this case,
we use a filter kernel to define how the different samples from the source image interact. Most
of this chapter will be spent discussing the use of filter kernels to achieve different image pro-
cessing effects, but first we will review basic 1.4 pixel shader syntax and the concept of a
dependent texture read in the following section on simple transfer functions.
Simple Transfer Functions
For the purposes of this chapter, we define a transfer function to be a function (often
one-dimensional) that is used to enhance or segment an input image. Transfer functions are
often used in scientific visualization to enhance the understanding of complex datasets by mak-
ing important details in the image clearly distinguishable. A transfer function might map a
scalar such as heat, pressure, or density to a color to aid in understanding the phenomenon
being visualized. This application of transfer functions is illustrated in 3D Textures and Pixel
Shaders by Evan Hart to apply pseudo-color to a dataset, which is a scalar density acquired
from a medical imaging process.
Here, we show how to use
transfer functions to stylize a
whole rendered 3D frame or
static 2D image. This allows us
to take normal color images and
cause them to become black and
white or sepia toned, as was done
for dramatic effect in recent films
such as Thirteen Days and
Pleasantville. We can also think
of our frame buffer as containing
heat and apply a heat signature
transfer function to give an effect
like that used in the films Preda-
tor and Patriot Games. The
application of the following three
transfer functions are shown in
Color Plate 2 and illustrated at
right for reference.
Image Processing with 1.4 Pixel Shaders in Direct3D 259
Figure 2: Three transfer functions
Team LRN
Black and White Transfer Function
As shown in Figure 2 and in Color Plate 2, we have applied three different transfer functions to
the color input image of an SR-71 Blackbird. The first transfer function is a simple Luminance
calculation which reads in the RGB color pixels from the source image and outputs them as
RGB colors where R = G = B = Luminance. The Luminance operation is performed using a
dot product operation to calculate Luminance = 0.3*red + 0.59*green + 0.11*blue. As shown
in the pixel shader below, we define the constant {0.3f, 0.59f, 0.11f} with the def instruction
and then sample texture 0 with the 0
th
set of texture coordinates using the texld instruction.
Listing 1: Simple Luminance shader for converting to black and white
ps.1.4
def c0, 0.30f, 0.59f, 0.11f, 1.0f
tex1d r0, t0
dp3 r0, r0, c0
The Luminance is calculated with the dp3 instruction and the final pixel color is stored in reg-
ister r0. This is a very simple transfer function that is expressible in the pixel shader assembly
language. Often, it is convenient to use a lookup table to apply a more arbitrary function to the
input pixels. The next transfer function will use this technique by doing a dependent read.
Sepia Tone Transfer Function
An example of a transfer function that is implemented by using a dependent read is a sepia
tone effect. In this example, we apply a 1D sepia map to the luminance of the input image to
make it look like an old photograph. Cut scenes in a 3D game set in the Old West would be a
good candidate for this effect. The bottom right image in Color Plate 2 shows the sepia tone
version of the original image. The colored stripe beneath the sepia image is the 1D texture used
to map from Luminance to Sepia tone. As shown in the shader code below, the color from the
original image is converted to Luminance and then used as a texture coordinate to sample the
1D sepia tone transfer function.
Sampling this 1D texture using texture coordinates that were computed earlier in the
shader is known as a dependent read and is one of the most powerful concepts in pixel shading.
Many of the shaders in following chapters will use dependent reads. Some examples include
application of transfer functions as shown in 3D Textures and Pixel Shaders, perturbation
refraction rays as shown in Accurate Reflections and Refractions by Adjusting for Object Dis-
tance, and indexing reflection and refraction maps as shown in Rippling Refractive and
Reflective Water in this book.
Listing 2: Simple transfer function for sepia or heat signature effects
ps.1.4
def c0, 0.30f, 0.59f, 0.11f, 1.0f
tex1d r0, t0
dp3 r0, r0, c0 // Convert to Luminance
phase
tex1d r5, r0 // Dependent read
mov r0, r5
260 Part 3: Pixel Shader Tricks
Team LRN
Heat Signature
Another example of a transfer function that is implemented by using a dependent read is a heat
signature effect. In this example, we use the same shader as the sepia effect but use the 1D tex-
ture shown in the bottom left of Color Plate 2. This gives the look of an image used to
visualize heat. This further illustrates the power of dependent texture reads: The same shader
can produce radically different results based on simple modification of the texture used in the
dependent read.
In future versions of Direct3D, we look forward to applying tone-mapping transfer func-
tions to real-time high dynamic range scenes with this same technique.
While we have been able to perform some useful operations mapping one pixel in the
input image to one pixel in the output image, we can perform even more useful tasks by sam-
pling multiple pixels from the input image. We will explore filters with this property in the
remainder of this article.
Filter Kernels
Filter kernels are a combination of the locations that the input image is sampled relative to a
reference point and any coefficients which modulate those image samples. A Gaussian blur, for
example, is designed to sample the source image in a radially symmetric way and weight the
samples according to a Gaussian distribution. The coefficients of a Gaussian blur sum to one.
The simple box filter used in this chapter weights all samples evenly, with the coefficients also
summing to one. Edge filters such as the Roberts and Sobel filters shown later in this chapter
have positive and negative lobes. The coefficients of edge filters (including Roberts and Sobel)
typically sum to zero.
Texture Coordinates for Filter Kernels
When rendering the output image, it is necessary to take care when setting the texture coordi-
nates of the quadrilateral that is texture mapped with the input image. To achieve a one-to-one
mapping from input texels to output pixels, the nave 0.0 to 1.0 texture coordinates do not give
the correct mapping. The correct coordinates are 0.0+ to 1.0+ , where is equal to 0.5 /
image_dimension. Obviously, for non-square images (h w), this will result in a separate
h
=
0.5 / h and
w
= 0.5 / w. Filter samples that are offset from the one-to-one mapping must also
take into account. The function ComputeImagePlacement() in the DX81_ImageFilter sample
application on the companion CD illustrates the correct method of setting the texture coordi-
nates on the quadrilateral texture mapped with the input image. Obviously, this texture
coordinate perturbation could be done in a vertex shader [James01], thus saving the cost of
storing and transferring the additional texture coordinates for a multi-tap filter kernel; this is
left out of the sample application for clarity.
Image Processing with 1.4 Pixel Shaders in Direct3D 261
Team LRN
This may seem like a trivial detail and, indeed, getting this wrong may not introduce noticeable
artifacts for some effects, but any ping-pong techniques that repeatedly render to and from a
pair of textures can suffer massively from feedback artifacts if the texture coordinates are set
up incorrectly. A ping-pong blur technique is implemented in the sample application.
Edge Detection
Edge detection is a very common operation performed in machine vision applications. In this
section, we will illustrate two common edge detection filters and their applications to image
processing. The use of image space methods for outlining of real-time non-photorealistic ren-
derings will be explored further in Non-Photorealistic Rendering with Pixel and Vertex
Shaders using the filters introduced here. A related gradient filter, the central-difference filter,
is also discussed in 3D Textures and Pixel Shaders.
Roberts Cross Gradient Filters
One set of inexpensive filters that is commonly used for edge detection is the Roberts cross
gradient. These 22 filters respond to diagonal edges in the input image and amount to doing a
pair of 2-tap first-differences to approximate the image gradients in the diagonal directions.
Since the Roberts filters can be thought of as simply three taps (with one tap reused in both
gradient calculations), it is easy to do the gradient magnitude calculation for both directions in
the same pixel shader. The following pixel shader first computes luminance for all three taps,
262 Part 3: Pixel Shader Tricks
Figure 3: Correctly sam-
pling textures for image
processing
Figure 4: Roberts cross gradient filters
Team LRN
computes the luminance cross gradients, takes their absolute values, adds them together, and
scales and inverts them before compositing them back over the original image.
Listing 3: Roberts cross gradient filter overlayed on original color image
// Roberts Cross Gradient Filter from color image
ps.1.4
Def c0, 0.30f, 0.59f, 0.11f, 1.0f
tex1d r0, t0 // Center Tap
tex1d r1, t1 // Down/Right
tex1d r2, t2 // Down/Left
dp3 r3, r0, c0
dp3 r1, r1, c0
dp3 r2, r1, c0
add r1, r3, r1
add r2, r3, -r2
cmp r1, r1, r1, r1
cmp r2, r2, r2, r2
add_x4 r1, r1, r2
phase
mul r0.rgb, r0, 1r1
+mov r0.a, c0.a
The result of this operation is black outlines layered over the original image in areas that have
sharp luminance discontinuities, as shown in Figure 5.
Sobel Filter
A more robust and popular edge detection filter is the Sobel filter [Sobel90]. The Sobel filters
are 6-tap 33 filters which detect horizontal and vertical edges. Like the Roberts filters, the
Sobel filters are just 90 rotations of each other. The filter is shown below:
Because the Sobel filter for each direction, u and v, is a 6-tap filter and the 1.4 pixel shading
model allows us to sample six textures, it is convenient to handle u and v gradients separately.
This is not strictly necessary, as more taps can be sampled as dependent reads at texture coordi-
nates which are derived from the other input texture coordinates. It should also be mentioned
Image Processing with 1.4 Pixel Shaders in Direct3D 263
Figure 5: Input image and
image with overlayed edges
computed with Roberts filters
Figure 6: The Sobel filter kernels
Team LRN
that it is possible to use the bilinear filtering capability of the hardware to approximate the dou-
ble-weighting of the central samples. In order to keep this example manageable, however, we
will avoid this and take a two-pass approach: one pass for | u| and one for | v|. The two partial
gradient magnitude images can then be composited to form a resulting gradient magnitude
image, which highlights the edges. A Luminance-only version of the original SR-71 Blackbird
image is shown in Figure 7, along with | u|, | v|, and | u|+| v| images.
A given Sobel filter is easily handled in one rendering pass with the following pixel shader that
samples all six taps, accumulates the samples, and takes the absolute value of the results. For a
non-photorealistic object and shadow outlining effect like the one shown in Non-Photo-
realistic Rendering with Pixel and Vertex Shaders, these white edges can be thresholded and
inverted to produce clean, black outlines. The results of one such effect are shown in Color
Plate 6 (see Non-Photorealistic Rendering).
Listing 4: Sobel filter pixel shader
// Sobel Filter (use for U or V, perturbations control which)
ps.1.4
def c0, 0.1666f, 0.1666f, 0.1666f, 1.0f
tex1d r0, t0 // Expect samples in one of two layouts...doesnt matter which
tex1d r1, t1 //
tex1d r2, t2 // t0 t3 t0 2t1 t2
tex1d r3, t3 // 2t1 x 2t4or x
tex1d r4, t4 // t2 t5 t3 2t4 t5
tex1d r5, t5 //
add r0, r0, r2
264 Part 3: Pixel Shader Tricks
Figure 7: Input image and gradient
magnitude images for u and v
directions
Team LRN
add r0, r0, r1_x2
add r0, r0, r3
add r0, r0, r4_x2
add r0, r0, r5
cmp r0, r0, r0, r0
Mathematical Morphology
Morphological operators are used in image processing applications that deal with the extrac-
tion, enhancement, or manipulation of region boundaries, skeletons, or convex hulls. In some
sense, one can think of mathematical morphology as set theory for images. In general, mathe-
matical morphology applies an arbitrary morphological operator to a source image. In the
following discussion, we apply a limited set of morphological operators since we effectively
hard-code them into the pixel shaders themselves. As such, the generality of the operators is
limited by the instruction count of the pixel shaders and the expressiveness of the language.
Nevertheless, morphology is an important area of image processing, and one can see how this
will become even more useful in future generations of hardware, which can access more sam-
ples from the input image and even express the morphological operator as an image in a
texture map. Even for the relatively small number of taps available to us in Direct3Ds 1.4
pixel shaders, we can apply some interesting morphological operators.
Mathematical morphology is typically applied to binary (1-bit) images, but it can just as
easily be applied to scalar (grayscale) images. In many cases, the input image is created from
thresholding another image, often one which has been edge-filtered. This binary image is then
operated on by a morphological operator and is either dilated or eroded by the operator. In this
chapter, we will use morphology to enhance the edges used in the non-photorealistic rendering
application discussed in Non-Photorealistic Rendering. The image processing sample appli-
cation on the companion CD and the non-photorealistic rendering sample application on the
CD can both perform dilation on outline images.
The Definition of Dilation
Figure 8 illustrates a dilation operation. The image on the left is the original binary image.
Each small cell in this figure represents one pixel. The small two-by-two black square in the
center is the operator. The white square specifies which pixel is the origin of the operatorin
this case, the top left pixel of the two-by-two operator.
Image Processing with 1.4 Pixel Shaders in Direct3D 265
Figure 8: Dilation operation
Team LRN
To perform dilation of the original image, the operator is moved over the entire image. For
every position of the operator, if any of the pixels of the operator cover a black pixel in the
input image, then the pixel under the origin of the operator is set to black in the output image.
This procedure results in the image on the right side of the figure where the black rectangle has
been dilated up and to the left. The result of another dilation operation, with a 13 operator
whose origin is centered vertically, is shown in Figure 9. Clearly, different dilation operators
can enhance different features of an input image.
A Dilation Shader
Once we have the binary input image, we apply the morphological operator in exactly the same
way as the filters applied earlier, but the shader performs a different operation on the samples
which lie under the kernel/operator. In the case of dilation, we use the cmp instruction to per-
form a threshold on the summation of the inverted samples. If the summation of the inverted
samples exceeds a threshold, then part of the operator must cover some black pixels and we set
the pixel to black. The source code for a 4-tap dilation operation is shown below:
Listing 5: Dilation shader
// Simple 4-tap dilation operator
ps.1.4
def c0, 0.2f, 1.0f, 0.0f, 0.0f
tex1d r0, t0 // Origin Tap
tex1d r1, t1
tex1d r2, t2
tex1d r3, t3
add r0, 1r0, 1r1 // Sum the inverted samples
add r0, r0, 1r3
add r0, r0, 1r2
add r0, r0, c0.r // Subtract threshold
cmp r0, r0, c0.b, c0.g // Set based on comparison with zero
Executing this shader on an image of object outlines thickens the outlines, as shown in Figure
10. Obviously, we could feed the dilated back through a dilation operation to thicken the lines
further using a technique similar to the ping-pong blur in the image processing sample applica-
tion on the companion CD. Application of dilation to non-photorealistic rendering is discussed
further in Non-Photorealistic Rendering.
266 Part 3: Pixel Shader Tricks
Figure 9: Dilation operation
Team LRN
The Definition of Erosion
Where a dilation operation can thicken images, an erosion operation thins the structures
present in an input binary image. This is often useful to clean up noise or thin down structures
in an image for pattern recognition.
An Erosion Shader
The erosion shader is very similar to the dilation shader. The 4-tap shader below applies a 2x2
box operator like the one shown in Figure 11. If any of the pixels under the box are white, then
the pixel under the origin (upper left) is set to white.
Listing 6: Erosion shader
// Simple 4-tap erosion operator
ps.1.4
def c0, 0.05f, 1.0f, 0.0f, 0.0f
tex1d r0, t0 // Origin Tap
tex1d r1, t1 //
tex1d r2, t2 //
tex1d r3, t3 //
add r0, r0, r1 // Sum the samples
add r0, r0, r2
add r0, r0, r3
add r0, r0, c0.r // Subtract threshold
cmp r0, r0, c0.g, c0.b // If any pixels were white, set to white
Applying this operator to the original edge image, for example, thins down the lines and causes
breaks in some places.
Image Processing with 1.4 Pixel Shaders in Direct3D 267
Figure 10:- An edge
image which has been
dilated to thicken the
edge lines
Figure 11: Erosion operation
Team LRN
Figure 12: Original and eroded outlines
This erosion operator can also get rid of salt-and-pepper noise in a binary image. To illustrate
this, we have added some small black areas (mostly smaller than 2x2 pixels) to the edge image
and applied the erosion shader. The stray black pixels are all eliminated, and the original edges
are thinned.
Figure 13: Noisy outline image, eroded noisy image, and opening of noisy image
Erosion can be followed by a dilation to eliminate salt-and-pepper noise while preserving
larger structures in the input image. This is called the opening of the input image. Dilation
followed by erosion is called the closing of the input image. Naturally, there is much more to
mathematical morphology than just dilation and erosion, but these two operations form the
building blocks of more advanced morphology techniques, including opening and closing.
Conclusion
We have provided an introduction to the concept of image processing in this chapter. This
includes application of analytical and texture-based transfer functions, blurring and edge detec-
tion filters, and mathematical morphology. These filters have applications in data visualization,
depth-of-field, non-photorealistic rendering, and general 3D frame post-processing. This chap-
ter really only scratches the surface of image processing with pixel shaders but will hopefully
serve as a reference for developers interested in post-processing their 3D scenes. For more on
applying these filters to non-photorealistic rendering, see Non-Photorealistic Rendering. For
more on image processing in general, an excellent introduction to image processing concepts
including those discussed here can be found in [Gonzalez92].
268 Part 3: Pixel Shader Tricks
Team LRN
Hardware Support
At publication, only boards based on the ATI RADEON 8500 graphics processor support the
1.4 pixel shading model necessary to perform the image filters illustrated here. Any future
hardware that supports the 1.4 pixel shader model or future pixel shader models such as ps.2.0
will be able to execute these shaders.
Sample Application
The DX81_ImageFilter sample on the companion CD illustrates image processing in Direct3D
and implements all of the filtering examples shown here. Most of the shaders used require sup-
port for the 1.4 pixel shader model. Due to the simplicity of this application, it is reasonable to
run this application on the reference rasterizer if you dont have a board that supports 1.4 pixel
shaders. Rendered output images are also shown on the CD.
References
[Gonzalez92] Rafael C. Gonzalez and Richard E. Woods, Digital Image Processing
(Addison-Wesley, 1992).
[James01] Greg James, Operations for Hardware-Accelerated Procedural Texture
Animation, Game Programming Gems 2, Mark DeLoura, Ed. (Charles River Media,
2001), pp. 497-509.
[Saito90] Takafumi Saito and Tokiichiro Takahashi, Comprehensible Rendering of 3-D
Shapes, SIGGRAPH Proceedings, 1990, pp. 197-206.
[Sobel90] Irvin Sobel, An isotropic 33 image gradient operator, Machine Vision for
Three-Dimensional Scenes, H. Freeman, Ed. (Academic Press, 1990), pp. 376-379.
Image Processing with 1.4 Pixel Shaders in Direct3D 269
Team LRN
Hallo World Font
Smoothing with Pixel
Shaders
Steffen Bendel
Efficient programming is not tricky. A clear design is more important than the knowledge of
special hacks. The shader described next is simple and demonstrates the main vertex shader
principal. As the acronym T&L suggests, vertex shaders are built to transform and light a ver-
tex. And much more is indeed not possible. All vertices are processed independent from each
other and in the same way. Remember, modern CPUs have the same floating-point power as
the vertex shader of current graphic cards. So, if some problems cannot be easily solved with
the vertex shader, use the CPU. Examples of this include complex object construction like pro-
cedural trees or a view-dependent LOD.
An engine needs text output. To be independent from the OS, self-written routines are typ-
ically used. The standard way takes up three steps. First, print the text into a texture. Then
upload the texture into video memory. Finally, render a quad with the text-texture. This allows
full control, but if the texure is big, it will need a lot of texture memory and AGP-bandwidth
for upload.
Because the theoretical triangle throughput is high enough, it is possible to render each
char with an associated quad. If you do so, you save the cost of texture handling at the expense
of vertex buffer upload. This can be an advantage in speed behavior or better handling.
Quads are often used in an engine. Different kinds of sprites are realized by it. In all these
cases, they have the same characteristic. In principle, each quad is its own object, but rendering
each quad with a single DirectX call is slow. Therefore, a vertex buffer is filled with data of
many quads and rendered at once. This can be realized by triangle lists or indexed primitives.
Index buffers are the best solution for most cases because they reduce the number of vertices
used and increase the processing speed.
Since the index layout is the same for different quad buffers (quad 0: vertex 0-3, quad 1:
vertex 4-7 ...), the index buffer can be reused. Some data, like quad corner position or texture
coordinates, are equal for various buffers. To avoid double data in the buffers, it can be split
into several vertex streams. One stream contains the generic data for all buffers, and the other
has special values for each shader. In the samples in this book, the generic stream includes the
corner position (0.0 or 1.0 for x, y coordinates) in register v0, indices for edge (0-3), and quad
in register v1.
270
Team LRN
The char data and layout are in the second stream. It is identical for all vertices of a quad
and must be duplicated in the C code. This cannot be done in the shader. The smallest data for-
mat that can be used is D3DVSDT_UBYTE4: two bytes for char position and two for the
bitmap coordinates that select the char. The bitmap is organized in a 16*16 array. If one more
byte of data is needed (for color selection), perhaps only the ASCII char must be transmitted
by stream and the two bitmap indices can be calculated in the vertex shader. However, that is
not as efficient as the preprocessing by CPU.
; v0 - 2d Position in quad
; v1.x - vertex index in the quad
; v1.y - current number of quad
; v2.xy - char position
; v2.zw - bitmap position for char
; c[0-3] - transformation matrix (not transposed)
; c[4-7] - bitmap : zw scaling, xy position-offset
; c[8-11] - screen : zw scaling, xy position-offset
vs.1.1
; set index
mov a0.x, v1.x
; calculate bitmap position
mad oT0.xy, v2.zw, c[a0.x + 4].zw, c[a0.x + 4].xy
; calculate untransformed screen position
mad r0.xy, v2.xy, c[a0.x + 8].zw, c[a0.x + 8].xy
; transform from object to clip space
; (not complete matrix r0.z=0, r0.w=1)
mul r1, r0.xxxx, c0.xyzw
mad r1, r0.yyyy, c1.xyzw, r1.xyzw
add oPos, c3.xyzw, r1.xyzw
First in the shader, the corner index is set in the constant offset register. Then bitmap position
is calculated. Both x and y indices are scaled by a factor and moved with an offset. The offset
depends on which quad corner is currently used. Scaling factor is always the same, but in the
vertex shader, only one constant can be used per instruction. So scaling and offset are com-
bined in one register. The same is done for the char position. That is, the position is
transformed to clip space. To reduce instruction count, an alternate form of matrix transforma-
tion is used instead of the typical construction with four dp4 commands. Here the matrix is not
transposed. All lines of the matrix are multiplied separately with one coordinate of the vector,
and the products are accumulated. Because the z-coordinate is explicit zero and the w-compo-
nent is one, some calculations can be removed. The result is that one instruction slot is saved.
ps.1.0 ; Declare pixel shader version 1.0
tex t0 ; Read texture
add r0.a, t0, c0 ; set and scale condition parameter
cnd r0, r0.a, c1, c2 ; set color dependend on condition
The pixel shader is not really interesting, but its functionality cannot be emulated by the old
texture combiners in one rendering step. The alpha value of the bitmap is used to decide if font
color or background color is drawn. Both colors are set as constants in the pixel shader. The
alpha of the resulting color can be used for transparency. Before the compare is computed, the
value is shifted. If the bitmap contains continuous alpha values instead of binary 0 or 1, the
shift offset describes the boldness of the font. There is also another advantage, as smoothed
Hallo World Font Smoothing with Pixel Shaders 271
Team LRN
alpha gives better interpolated data. The typical bilinear filtering artifacts are suppressed. Fig-
ure 1 shows a normal and a smoothed object. When it is interpolated and used in the shader
above, you can see the result in Figure 2. Alpha values between 0.4 and 0.6 produce the best
results.
272 Part 3: Pixel Shader Tricks
Figure 1: Normal
(left) and smoothed
object (right)
Figure 2: Contour
lines for different
values when inter-
polating Figure 1
Team LRN
Emulating Geometry
with Shaders
Imposters
Steffen Bendel
The next shader demonstrates the ability to emulate geometry by pixel shader. Of course, real
geometry details require a lot of triangles, but it must not be complex; it is sufficient if it looks
complex. A digital picture consists of pixels. How it is created is not important and cannot be
identified. A painted picture and a rendered image are the same in this manner. The difference
is that a scene can be rendered from any view. The 3D data is easy to transform. To get more
views for a non-3D object, they must be prepared and saved. That means a lot of memory con-
sumption. It is a general rule that memory or CPU power can be reduced at the expense of the
other.
Between the two extremes, there are some good compromises for special cases. Imposters
are quads providing different views of the object. Before the age of real 3D graphics, this was
the only way to draw complex objects in real time, but this allows no continuous turning. Pre-
viously, the quality was good enough relative to other algorithms (like drawing an object with
ten untextured triangles). Now it is relatively bad because it scales bad with memory size. You
need a lot of memory for the precomputed views to get a relatively smooth rotation. That mem-
ory is better used to increase the number of details in a triangle mesh. For general objects,
imposters are only useful for level of detail if an object is far away and only a few pixels big.
A complete viewing description for an object is a five-dimensional array of ray-object
intersection data. For every ray (three values for position and two for direction angles), the
resulting color is saved. This array is too big for use. If the object is symmetric against trans-
formation, the dimension number can be reduced by using alternate coordinates. These
coordinates are computed from position and direction and have lower dimension. The body
with the highest symmetry is the sphere. A point of the sphere rotated by origin is also on the
sphere. The shape is the same, only texture coordinates must be transformed.
The presented algorithm calculates a textured and lighted sphere in the pixel shader. The
sphere normal vector is used as alternate coordinate because it is easy to transform. Position
and normal direction are equivalent and unique. These coordinates are computed by the first
texture lookup. The bitmap contains normals that correlate to a front view. After that, the coor-
dinates are transformed by matrix multiply to get a rotation. A cubemap lookup is done to get
the color for the transformed sphere position.
273
Team LRN
The lighting is simple Gouraud shading with auxiliary ambient color. The dot product
between normal vector and light vector is computed in the coordinates of the original normal
vector.
Because texture values are always positive, the normal data must be compressed into the
range [0.0, 1.0]. The _bx2 modifier expands it to [1.0, 1.0].
The shape of the rendered quad is not a round. The alpha channel of the normal map con-
tains data to mask the sphere. For quality improvement, it is smoothed as in the previous
article.
Early versions of DirectX8 prohibited access to texture color if it was the source for a
dependent texture read. This restriction was abolished in version 8.1.
; t0 - normal map
; t1,t2,t3 - bump matrix vectors
; t3 - cubemap
ps.1.1 ; Declare pixel shader version 1.1
; need DX8.1
tex t0 ; Read normal texture
texm3x3pad t1, t0_bx2 ; transform normal vector and
texm3x3pad t2, t0_bx2 ; look in cubemap
texm3x3tex t3, t0_bx2 ;
dp3_sat r0.rgb, t0_bx2, v0_bx2 ; scalar product for gouraud
add r0.rgb, r0, c0 ; add ambient
mul r0.rgb, r0, t3 ; mul light and color
+mov r0.a, t0.a ; clip border
There are two input streams for the vertex shader, one containing general quad data (see previ-
ous article) and the other the position data for the sphere. This is the simple version; all spheres
have the same radius and orientation. If the spheres differ here, the stream and shader must be
extended.
The first part of the vertex shader performs typical billboard operations. The center posi-
tion is transformed to view space, and then the corner scaled offset is added. Here the mad
operation is used as mov for the z and w components by setting the product to zero. This saves
one instruction, but a suitable constant setting in the registers is needed. The result is trans-
formed to clip space and put into the output register.
Next, the matrix for the normal transformation in the pixel shader is computed. This is an
approximation because perspective correction is not possible. The matrix is built of compo-
nents of the eye vector. So if the sphere is rendered left on screen, you see more of the right
side. The matrix vectors are normalized but not orthographic. At least, the vectors are trans-
formed to world space. Normally, this is done by multiplying with the inverted view matrix,
but in the shaders (pixel and vertex) the matrix must be transposed to realize a matrix multiply
with the dot product operators. For orientation matrixes, inverting and transposing are equiva-
lent operations. So all this can be done by multiplying with the view matrix:
transpose(m*invert(view)) = transpose(m*transpose(view)) = m*view. If every sphere should
have its own orientation, the resulting matrix needs to be transformed by an orientation matrix
(not implemented in the example).
The light position is given in view space. In the shader, the object light vector is computed,
normalized, and scaled by distance-dependent intensity. It is scaled into the range [0.0, 1.0] to
274 Part 3: Pixel Shader Tricks
Team LRN
pass through the vertex color register. Texture coordinates are 0.0 or 1.0 for each component.
They are calculated by resizing the quad coordinates (1.0 or 1.0).
; oPos -- the vertex's clip-space coordinates
; oT0 -- the vertex's uv-coordinates
; oD0 -- the light vector
;
; v0 -- relative quad position
; v1 -- quad indices
; v2 -- position of billboard center
; c0-c3 -- contains matrix to transform to view space
; c4-c7 -- contains matrix to transform pespective
; c8 -- 0.0, 0.5, 1.0, -1.0 constant
; c9 -- radius, -radius, 0.0, 0.0
; c10 -- light position in view space
; Declare vertex shader version 1.1
vs.1.1
; Transform to view coordinates
dp4 r0.x, v2, c0
dp4 r0.y, v2, c1
dp4 r0.z, v2, c2
dp4 r0.w, v2, c3
; perform billboard operation r1.xy = v0.xy + r0.xy and r1.zw = r0.zw
mad r1.xyzw, c9.xyzz, v0.xyzw, r0.xyzw
; Perform perspective transformation
dp4 oPos.x, r1, c4
dp4 oPos.y, r1, c5
dp4 oPos.z, r1, c6
dp4 oPos.w, r1, c7
; create bump matrix vector1 dependent an view position,
; r7.x = r0.z r7.y = 0.0 r7.z = r0.x
mul r7.xyz, r0.zyx, c8.zxx
;normalize vector1
dp3 r1, r7, r7
rsq r1, r1
mul r7.xyz, r7.xyz, r1.xyz
; create bump matrix vector2 dependent an view position
; r8.x = 0.0 r8.y = r0.z r8.z = r0.y
mul r8.xyz, r0.xzy, c8.xzz
;normalize vector2
dp3 r1, r8, r8
rsq r1, r1
mul r8.xyz, r8.xyz, r1.xyz
; create bump matrix vector3 dependent an view position
; r9.x = -r0.x r9.y = -r0.y r9.z = r0.z
mul r9.xyz, r0.xyz, c8.wwz
;normalize vector3
dp3 r1, r9, r9
rsq r1, r1
mul r9.xyz, r9.xyz, r1.xyz
; Put transform the bump mapping matrix with view and put in T1, T2 and T3
dp3 oT1.x, r7, c0
dp3 oT1.y, r7, c1
Emulating Geometry with Shaders Imposters 275
Team LRN
dp3 oT1.z, r7, c2
dp3 oT2.x, r8, c0
dp3 oT2.y, r8, c1
dp3 oT2.z, r8, c2
dp3 oT3.x, r9, c0
dp3 oT3.y, r9, c1
dp3 oT3.z, r9, c2
; calculate light1 vector
sub r1, c10, r0
; resize by distance dependency
dp3 r2, r1, r1
rcp r2, r2
mul r1, r1, r2
; Shift the light vector and put in D0
mad oD0, r1, c8.yyyy, c8.yyyy
; Output texture coordinates
mad oT0.xy, v0.xy, c8.yy, c8.yy
The following figure demonstrates the pixel and vertex shaders discussed in this article.
276 Part 3: Pixel Shader Tricks
Figure 1
Team LRN
Smooth Lighting with
ps.1.4
Steffen Bendel
Vertex lighting uses linear approximation between the vertices and causes visible artifacs.
Exact lighting needs calculation per-pixel. This is the way of the future, but not all necessary
operations can be calculated in the pixel shader, or they are too slow for current hardware.
This article describes how to create better lighting on a per-pixel basis. Only the light
intensity calculation is discussed here. The material color is ignored in this rendering step.
What are the main components determining the brightness? The distance to the light
source is one factor. If the object is farther away, it looks darker. The physical correct equation
for a sphere light is proportional to the reciprocal of the squared distance (Figure 1). This func-
tion falls off too fast. Computer graphics have a smaller color representation range than the
human eye can observe. If using the real equation, most scene parts are too dark or have maxi-
mum light intensity that can be represented. A more linear function, as in Figure 2, is a better
solution for some cases.
Because such operations are not possible in the pixel shader, a lookup table is required. All
functions can be represented in this way. The distance dependency is a one-dimensional func-
tion, but the distance cannot be iterated or calculated in the pixel shader. A volume texture
lookup is the easiest solution. The light vector is scaled and the result is used as coordinates for
the texture that contains the distance function.
Another factor is the angle between normal vector, eye vector, and light vector. The
reflected eye vector is often used instead of the original eye. Take a look at Figure 3 for
notation.
277
Figure 1
Figure 2
Team LRN
How are the vectors that are necessary for the lighting calculated? The normal vector is given
by texture map (bump mapping). The light and eye vector are calculated in the vertex shader
and transformed to the space of the normal map. If no world coordinates are needed in the
pixel shader, a matrix transformation is avoided. The vectors are interated perspectively cor-
rect. They are normalized by cubemap lookups. Sometimes the precision (around 1% for an
8-bit per channel texture) is not enough because specular highlighting is very sensitive. A
rough bump map can hide this problem.
The reflect eye vector is given by:
, are normalized
In general, the intensity equation is separated into two parts, one for and one for .
The diffuse component is view independent. A typical function is I
Diffuse
= k
Diffuse
cos ( ).
That is equal to the dot product between the normalized light and the normal vector. If the
material is a mirror, this component vanishes. The specular highlight is decribed by . This is
an approximation for reflection and micro roughness of the surface. If the object is polished,
the highlight is small and bright. A rough surface has a smooth highlight with low intensity.
This is described by I
Specular
= k
Specular
cos( )
n
. Metallic surfaces have values n 100 for the
exponent. The example uses a lookup table for the exponential function. The distance depend-
ence is also encoded in this texture. Theoretically, this isnt urgent since both components are
linearly independent and the result is a simple product. But texture values have a limited range.
If one calculates with that, it comes to numeric inaccuracies. A combined texture for both val-
ues avoids the problems.
The same is done for the diffuse part. It is calculated by the multiplication of distance
intensity and the dot product between N and L in a simpler variant. If the light source is near
the surfaces, the normal vector changes strongly. As a result, the object looks too dark because
the dot product is small and cannot be compensated for intensity. Here, a correction for small
distances is precalculated in the lookup table.
278 Part 3: Pixel Shader Tricks
-eye vector
-normal vector
-light vector
-reflected eye vector
surface
point light
camera

R
r
N
r
E
r
L
r
Figure 3
Team LRN
There are a lot of improvements that are not included in the example, such as the
lightsource has no color, the alpha channel of the bump map could contain a specular coeffi-
cient, and a spotlight or shadow map can be used instead of one of the lookup tables.
Pixel shader:
ps.1.4
texld r0, t0 ; normal map N
texld r1, t1 ; cubic normalization for light vector L
texld r2, t2 ; cubic normalization for eye vector E
texld r3, t3 ; volume table lookup for distance
dp3 r4, r0_bx2, r2_bx2 ; N*E
mad r4.rgb, r0_bx2, -r4_x2, r2_bx2 ; -N(2*N*E)+E
dp3_sat r5.xyz, r1_bx2, r4 ; L*Reflectted E
mov r5.y, r3.r ; set distant value
dp3_sat r4.xyz, r1_bx2, r0_bx2 ; diffuse shading N*L
mov r4.y, r3.r ; set distant value
phase
texld r4, r4 ; table lookup for diffuse and distance
texld r5, r5 ; table lookup for specular and distance
add r0, r4, r5 ; add diffuse and specular
Vertex shader:
; v0 - 3d Position in quad
; v1-3 - face orientation matrix, v2 is normal
; v4 - bumpmap coordinates
; c[0-3] - object to clip space matrix
; c4 - light position in object space
; c5 - camera position in object space
; c6.x - inverse light range
; c6.y - 0.5
vs.1.1
; calculate position for draw
dp4 oPos.x, v0, c0
dp4 oPos.y, v0, c1
dp4 oPos.z, v0, c2
dp4 oPos.w, v0, c3
; set bumpmap
mov oT0, v4
; calculate light vector in object space
sub r0.xyz, c4, v0
; transform light vector to object(normal) space
dp3 oT1.x, r0, v1
dp3 oT1.y, r0, v2
dp3 oT1.z, r0, v3
; set coordunates for distance lookup
mad oT3.xyz, r0, c6.xxx, c6.yyy
; calculate eye vector in object space
sub r1.xyz, v0, c5
; transform eye vector to object(normal) space
Smooth Lighting with ps.1.4 279
Team LRN
dp3 oT2.x, r1, v1
dp3 oT2.y, r1, v2
dp3 oT2.z, r1, v3
The following figure is an example using the pixel and vertex shaders in this article.
280 Part 3: Pixel Shader Tricks
Figure 4
Team LRN
Per-Pixel Fresnel Term
Chris Brennan
Introduction
Common real-time environment mapping reflection techniques reflect the same amount of
light no matter how the surface is oriented relative to the viewer, much like a mirror or chrome.
Many real materials, such as water, skin, plastic, and ceramics, do not behave in this way. Due
to their unique physical properties, these materials appear to be much more reflective edge-on
than they are when viewed head-on. The change in reflectance as a function of the viewing
angle is called the Fresnel effect, and the term in our lighting equation that we use to approxi-
mate this is called the Fresnel term.
Use of a Fresnel term can be somewhat subtle, but it is fundamental to photorealistic ren-
dering of many materials, especially water. While this has been used as a per-vertex quantity
in [Tso87] and [Vlachos00], we will show how to compute a per-pixel Fresnel term using 1.4
pixel shaders. In addition to photorealistic effects, we will show how this term can be used to
generate some supernatural or just plain fun results as well.
In this chapter, we use a per-pixel Fresnel term in conjunction with cubic environment-
mapped bump mapping. This is not a requirement, but it is a good illustrative example
because a normal and reflection terms are readily available to experiment with.
Fresnel Effects
The Fresnel term is calculated with a simple dot
product between the per-pixel normal N and a nor-
malized eye vector E. The result of the dot product,
NE, is then used as a parameter to a function, F(x),
to get a reflection intensity as seen in Figure 1,
which uses an approximate Fresnel term of F(x) =
(1x)
2
.
281
Figure 1: The Fresnel term, (1NE)
2
, which
will be used as the intensity of the
reflections
Team LRN
The following example shows the look of the ceramic statue with and without the Fresnel
effect modulated into the reflection map:
+ =
+ =
Figure 2: Diffuse + Reflections = final result with (bottom) and without (top) the Fresnel term
Some other effects can be achieved by using a Fresnel term in ways other than reflectance. In
the following example, the Fresnel term is used as a transparency term. NE is used in a
dependent texture read to make F(x) completely arbitrary. This allows transparency to be soft
along the silhouette edges of the skull, opaque just inside, and then transparent again when
viewed straight on. When used with additive alpha blending, this gives a ghostly membrane
effect similar to the alien creatures in the film Final Fantasy: The Spirits Within. Viewing F(x)
= 1 NE directly yields an effect that looks like an electron micrograph.
Figure 3: Ghostly effect and electron micrograph effect
282 Part 3: Pixel Shader Tricks
Team LRN
Effect Code
matrix mW; // World matrix
matrix mWVP; // World*View*Projection matrix
matrix mWrt; // Inverse Transpose World Rotation matrix
matrix mWVt; // Inverse Transpose World*View matrix
vector pOV; // Object space viewer position
vector cnst;
vector cHlf;
texture tBase;
texture tGloss;
texture tCube;
texture tDiffuseCube;
texture tNormal;
texture tNormalizer;
DWord dwCubeMipFilter = 2; // None = 0, Point = 1, Linear = 2
vertexshader vTexTangentmatrixViewLight =
decl
{
stream 0;
float v0[3]; // Position
D3DCOLOR v3; // Normal
float v7[2]; // Tex coord 0
D3DCOLOR v8; // Binormal
D3DCOLOR v9; // Tangent
}
asm
{
; Sample vertex shader which outputs a 3x3 orthogonal transform which
; takes you from world space to tangent space. The result is stored in
; output tex coords 1,2,and 3. Tex coord 4 stores a view vector which
; can be used in a reflect operation.
;
; Tex coordinate 5 stores a scaled version of the centroid to vertex
; vector that is used as a correction factor for environment mapping.
;
; This sample expects the incoming vertex data to be set up as follows:
;
; v0 - Vertex Position
; v3 - Vertex Normal
; v7 - Vertex 2D Bumpmap Texture Coords
; v8 - Vertex Tangent (normalized assumed to be positive v direction for
tangent space)
; v9 - Vertex Tangent (normalized assumed to be positive u direction for
tangent space)
; This sample expects the constant store to be set up in the following manner:
;
; c1 - Color of material
; c12..c15 - World Matrix
; c16..c19 - Composite World-View-Projection Matrix
; c20..c23 - Env Transform Matrix (object to world or view coord transform
; (depends on type of env mapping you want to do) )
; c28 - ModelSpace Camera Position
; c95 - 0, .5, 1, 2
vs.1.0
def c94, 0.1f, 0.33f, 0.20f, 0.18f ; Scalar for cube map adjustment
; = roughly 1/Radius of environment map.
Per-Pixel Fresnel Term 283
Team LRN
; OutPos = ObjSpacePos * Composite World-View-Projection Matrix
m4x4 oPos, v0, c16
mov oT0.xy, v7 ; passthrough texcoord 0 for bump map addressing
;derive per vertex eye vector using camera position
add r8, c28, -v0 ; view direction
m3x3 r9, r8, c20 ; transform view vector from object space to env space
dp3 r11.x, r9, r9 ; magnitude^2
rsq r11.y, r11.x ; 1/mag
mul oT4.xyz, r9, r11.y ; normalize
;Expand from compressed D3DCOLOR to 1 to 1 range.
mad r3, v3, c95.wwwz, -c95.zzzx ; Normal = (v3*2.0)-1.0
mad r7, v8, c95.wwwz, -c95.zzzx ; TangentU = (v8*2.0)-1.0
mad r8, v9, c95.wwwz, -c95.zzzx ; TangentV = (v9*2.0)-1.0
; upper 3x3 of world matrix by upper tangent space transposed
dp3 oT1.x, c20, r7
dp3 oT1.y, c20, r8
dp3 oT1.z, c20, r3
dp3 oT2.x, c21, r7
dp3 oT2.y, c21, r8
dp3 oT2.z, c21, r3
dp3 oT3.x, c22, r7
dp3 oT3.y, c22, r8
dp3 oT3.z, c22, r3
; Calculate adjustment amount to account for cube map center not being at this
object's centroid.
m3x3 r0, v0, c12
mul oT5, r0, c94.w
};
pixelshader ps14_Cube_Bump_Diff_Fresnel_Reflect = asm
{
ps.1.4
texld r0, t0 ; Look up normal map.
texld r1, t4 ; Eye vector through normalizer cube map
texcrd r4.rgb, t1 ; 1st row of environment matrix
texcrd r2.rgb, t2 ; 2st row of environment matrix
texcrd r3.rgb, t3 ; 3rd row of environment matrix
texcrd r5.rgb, t5 ; Cube map position adjustment
dp3 r4.r, r4, r0_bx2 ; 1st row of matrix multiply
dp3 r4.g, r2, r0_bx2 ; 2nd row of matrix multiply
dp3 r4.b, r3, r0_bx2 ; 3rd row of matrix multiply
dp3 r0.rgb, r4, r1_bx2 ; (N.Eye)
mad r3.rgb, r4_x2, r0, r5 ; 2N(N.Eye) + adjustment for pos not being at
center of cube map
dp3 r2.rgb, r4, r4 ; N.N
mad r2.rgb, -r1_bx2, r2, r3 ; 2N(N.Eye) - Eye(N.N)
mov_sat r0.rgb, 1-r0 ; Pass through 1-(N.Eye) for fresnal
calculation
phase
; texcrd r0, r0 ; N.Eye used for fresnel calculation
texld r2, r2 ; Sample cubic reflection map
texld r3, t0 ; Sample base map
texld r4, r4 ; Sample cubic diffuse map
texld r5, t0 ; Sample gloss map
284 Part 3: Pixel Shader Tricks
Team LRN
mul r1.rgb, r5, r2 ; Specular = Gloss*Reflection
+mul r0.a, r0.b, r0.b ; ^2
mul r0.rgb, r1, r0.b ; Specular * Fresnel
mad r0.rgb, r3, r4_x2, r0 ; Base*Diffuse + Specular
};
technique ps14_Cube_Bump_Diff_Fresnel_Reflect
{
pass P0
{
VertexShaderConstant[12] = <mW>; // World
VertexShaderConstant[16] = <mWVP>; // World*View*Proj Matrix
VertexShaderConstant[20] = <mWrt>; // World Rotation Inv Transpose used
// for Eye vector transform
VertexShaderConstant[28] = <pOV>; // Object space Viewer position
VertexShaderConstant[95] = <cnst>;
VertexShader = <vTexTangentmatrixViewLight>;
Texture[0] = <tNormal>;
MagFilter[0] = Linear;
MinFilter[0] = Linear;
MipFilter[0] = Linear;
Texture[1] = <tNormalizer>;
AddressU[1] = Clamp;
AddressV[1] = Clamp;
MagFilter[1] = Linear;
MinFilter[1] = Linear;
MipFilter[1] = Point;
Texture[2] = <tCube>;
AddressU[2] = Clamp;
AddressV[2] = Clamp;
MagFilter[2] = Linear;
MinFilter[2] = Linear;
MipFilter[2] = <dwCubeMipFilter>;
Texture[3] = <tBase>;
AddressU[3] = Wrap;
AddressV[3] = Wrap;
MagFilter[3] = Linear;
MinFilter[3] = Linear;
MipFilter[2] = Linear;
Texture[4] = <tDiffuseCube>;
AddressU[4] = Clamp;
AddressV[4] = Clamp;
MagFilter[4] = Linear;
MinFilter[4] = Linear;
MipFilter[4] = <dwCubeMipFilter>;
Texture[5] = <tGloss>;
AddressU[5] = Wrap;
AddressV[5] = Wrap;
MagFilter[5] = Linear;
MinFilter[5] = Linear;
MipFilter[5] = Linear;
PixelShader = <ps14_Cube_Bump_Diff_Fresnel_Reflect>;
}
}
Per-Pixel Fresnel Term 285
Team LRN
This technique, including shaders and source code, is used in the Treasure Chest demo avail-
able on the companion CD and also at http://www.ati.com/na/pages/resource_centre/dev_rel/
Treasurechest.html. There are menus that can be accessed after going to windowed mode by
pressing Alt+Enter.
References
[Tso87] Pauline Tso and Brian Barsky, Modeling and Rendering Waves: Wave Tracing
Using Beta-Splines and Reflective and Refractive Texture Mapping, ACM Transactions
on Graphics (6) 1987, pp. 191-214.
[Vlachos00] Alex Vlachos and Jason L. Mitchell, Refraction Mapping for Liquids in
Containers, Game Programming Gems, Mark DeLoura, Ed. (Charles River Media, 2000).
286 Part 3: Pixel Shader Tricks
Team LRN
Diffuse Cube Mapping
Chris Brennan
Introduction
Cube mapping has become a common environment mapping technique for reflective objects.
Typically, a reflection vector is calculated either per-vertex or per-pixel and is then used to
index into a cube map that contains a picture of the world surrounding the object being drawn.
To date, cube maps have been used primarily for specular effects that are layered on top of
per-vertex diffuse lighting or static diffuse light maps. In this chapter, we present diffuse cube
maps, which happen to come nearly for free when performing specular or environment map-
ping with a cube map. The main advantage of diffuse cube mapping is that an unlimited
number of lights may be included in a single rendering pass. Diffuse cube maps can also be
generated from light probes [Debevec98] acquired from a real environment to provide a diffuse
or ambient term to really place an object in a scene.
Using Diffuse Cube Maps
Diffuse cube maps are different in usage from reflection maps only in that the surface normal
is used to index into the map with the intention that it returns the total amount of light scattered
from that direction. The reason it is free to use diffuse cube maps when using reflections is that
a normal is required for doing the reflection operation given the normal (N) and eye vector (E):
or if the normal is not of unit length:
Even the unnormalized normal may be used in diffuse cube mapping without any additional
pixel or vertex operations. The following is an example of per-pixel environment mapped
bump mapping with the diffuse component added in bold:
ps.1.4
def c0, 1,1,1,1
texld r0, t0 ; Look up normal map.
texcrd r1.xyz, t4 ; Eye vector
texcrd r4.xyz, t1 ; 1st row of environment matrix
texcrd r2.xyz, t2 ; 2st row of environment matrix
texcrd r3.xyz, t3 ; 3rd row of environment matrix
287
Team LRN
dp3 r4.x, r4, r0_bx2 ; N.x = 1st row of matrix multiply
dp3 r4.y, r2, r0_bx2 ; N.y = 2nd row of matrix multiply
dp3 r4.z, r3, r0_bx2 ; N.z = 3rd row of matrix multiply
dp3_x2 r3.xyz, r4, r1 ; 2(N.Eye)
mul r3.xyz, r4, r3 ; 2N(N.Eye)
dp3 r2.xyz, r4, r4 ; N.N
mad r2.xyz, -r1, r2, r3 ; 2N(N.Eye) - Eye(N.N)
phase
texld r2, r2 ; Sample cubic reflection map
texld r3, t0 ; Sample base map with gloss in alpha
texld r4, r4 ; Sample cubic diffuse map
mul r1.rgb, r3.a, r2 ; Specular = Gloss*Reflection
mad r0.rgb, r3, r4, r1 ; Base*Diffuse + Specular
+mov r0.a, c0.a ; Put 1.0 in alpha
The normal is calculated by looking up a normal map. However, the normal is in tangent
space, which needs to be transformed into world space through a matrix multiply of the concat-
enation of the texture tangent matrix and the world matrix and stored in r4, which is used when
sampling the diffuse map in the next phase. So with the addition of only one texture fetch and
one instruction, an arbitrarily complex diffuse lighting component can be added.
Generating Dynamic Diffuse Cube Maps
The idea is that a cube map is generated dynamically every frame in the same way that a
reflection cube map is generated, but instead of drawing actual world geometry from the point
of view of the object into the six faces, a diffusely lit hemisphere around the object is accumu-
lated into the cube map for each light. The following figure shows the corresponding diffuse
and reflection maps for one light.
Figure 1: Diffuse cube map (right) and reflection cube map (left) for one light
So for four lights, the process would be to clear the cube map to the ambient color, and then for
each light, add in its light component as shown below and in Color Plate 3.
288 Part 3: Pixel Shader Tricks
Team LRN
Figure 2: The accumulation of lights when generating a diffuse cube map
The resulting cube map can then be used to account for all light in the scene with one texture
lookup. Generally, the cube maps do not need to be very large. 32x32 for each face is sufficient
for not having any artifacts. Therefore, the cost of rendering a large number of lights into a
cube map is much less than rendering many light passes per object when rendering the final
image. This property makes diffuse cube mapping a logical choice when an object can be con-
tained in an environment map and the lighting is complex or a reflection environment map is
already being used.
This technique is used in the Treasure Chest demo available on the companion CD
included in this book and also at http://www.ati.com/na/pages/resource_center/dev_rel/
Treasurechest.html. There are menus to view the diffuse cube map that can be accessed after
going to windowed mode by pressing Alt+Enter.
References
[Debevec98] Paul E. Debevec, Rendering Synthetic Objects into Real Scenes: Bridging
Traditional and Image-Based Graphics with Global Illumination and High Dynamic Range
Photography, (SIGGRAPH Proceedings, July 1998).
Diffuse Cube Mapping 289
Team LRN
Accurate Reflections and
Refractions by Adjusting
for Object Distance
Chris Brennan
Introduction
Environment mapping is a technique by which a texture map or maps contain the view of the
environment from a particular point. The texture map is then addressed using a function of a
reflection or refraction vector when rendering an object. There are several forms of environ-
ment mapping: spherical, dual paraboloid, and cubic, to name a few. They are all basically the
same concept, differing only in what function is used to create and address the map. Cubic
environment mapping is a method of sampling a texture consisting of six renderings of the
world from the center of a cube out through each of its faces. The texture is addressed with a
vector that indicates the direction from the center of the cube to the sample point on one of its
faces. Therefore, a calculated reflection vector can be directly used as an address into the envi-
ronment map. Because of these simple properties, cube maps lend themselves to dynamic
real-time creation and addressing within pixel shaders. The following artifacts apply to all
environment mapping, but this example is applied to cubic environment maps.
The Artifacts of Environment Mapping
One of the common problems seen when environment mapping is that flat surfaces do not ren-
der well. There is very little variation in the reflection vector direction itself, and because
environment maps are indexed from the exact center, it does not take into account that a reflec-
tion vector can be originated off-center and therefore end up reflecting a different object. This
deficiency would make a flat mirror appear to reflect all the same point. The result is that only
a few texels are magnified and used in the final image. This artifact is worse the closer that
reflected objects are relative to the viewer. The same artifact appears when using a refraction
vector as opposed to a reflection vector, but it is more obvious where the transparent object
meets the object behind. See Figure 1 for an example of the artifact as seen through a window
using a refraction vector. Notice how the part of the chest seen through the window does not
properly match the chest itself.
290
Team LRN
Figure 1: This example shows a windowpane with environment map artifacts that produce incorrect
looking transparency.
Environment mapping based solely on the reflection vector direction acts as if the objects in
the environment map are infinitely far away. Advanced RenderMan [Apodaca00] describes a
way to alleviate this by adjusting the use of the environment map such that it has a finite radius
that is close to the real proximity of the objects. The most precise way to do this is by intersect-
ing the reflection or refraction ray with the environment map sphere and then using the vector
from the center of the map to the intersection point (CI) to index the map. See Figure 2.
Calculating the intersection of a circle and a vector is difficult in a pixel shader, but another
approximation of the same goal can be achieved by adding a scaled version of the vector from
the center point of the environment map to the point on the object being drawn (CP). Figure
3shows how the approximate corrected version of the vector intersects the same point I on the
environment map. The scaling factor is 1/(environment map radius) and is derived from:
Accurate Reflections and Refractions by Adjusting for Object Distance 291
Figure 2: Corrected vector based on
an intersection with a finite radius
Figure 3: Corrected vector adjusted by
a vector based on the points distance
from center
Team LRN
which when divided by the environment map radius changes to:
The total correction factor of CP/Radius can be calculated in the vertex shader and interpolated
for adding in after the per-pixel reflection calculation. The following figure illustrates the
amount of correction applied.
The corrected version of Figure 1 is shown in Figure 5:
Figure 5: The adjusted version of the refraction vector shows the geometry behind at the correct
scale at all viewing distances.
Notice how the refracted image now matches the
geometry behind it. The correction can be applied
per-pixel as seen in Figure 6 and Color Plate 4.
Figure 6: Stained glass color plate
292 Part 3: Pixel Shader Tricks
Figure 4: The magnitude of the correction vector
displayed as grayscale intensity.
Team LRN
The composite effect was rendered with the following shader:
; Sample vertex shader which outputs a 3x3 orthogonal transform which
; takes you from world space to tangent space. The result is stored in
; output tex coords 1,2,and 3. Tex coord 4 stores a view vector which
; can be used in a reflect operation.
;
; Tex coordinate 5 stores a scaled version of the centroid to vertex
; vector that is used as a correction factor for environment mapping.
;
; This sample expects the incoming vertex data to be set up as follows:
;
; v0 - Vertex Position
; v3 - Vertex Normal
; v8 - Vertex Tangent U
; v9 - Vertex Tangent V
; This sample expects the constant store to be set up in the following manner:
;
; c1 - Color of material
; c12..c15 - World Matrix
; c16..c19 - Composite World-View-Projection Matrix
; c20..c23 - Env Transform Matrix (object to world or view coord transform
; (depends on type of env mapping you want to do) )
; c28 - ModelSpace Camera Position
; c95 - 0, .5, 1, 2
vs.1.0
def c94, 0.18f, 0.18f, 0.18f, 0.18f ; Scalar for cube map adjustment
; = roughly 1/Radius of environment map.
; OutPos = ObjSpacePos * Composite World-View-Projection Matrix
m4x4 oPos, v0, c16
mov oT0.xy, v7 ; passthrough texcoord 0 for bump map addressing
;derive per vertex eye vector using camera position
add r8, c28, -v0 ; view direction
m3x3 r9, r8, c20 ; transform view vector from object space to env space
dp3 r11.x, r9, r9 ; magnitude^2
rsq r11.y, r11.x ; 1/mag
mul oT4.xyz, r9, r11.y ; normalize
; upper 3x3 of world matrix by upper tangent space transposed
dp3 oT1.x, c20, v8
dp3 oT1.y, c20, v9
dp3 oT1.z, c20, v3
dp3 oT2.x, c21, v8
dp3 oT2.y, c21, v9
dp3 oT2.z, c21, v3
dp3 oT3.x, c22, v8
dp3 oT3.y, c22, v9
dp3 oT3.z, c22, v3
; Calculate an adjustment amount to account for cube map center not
; being at this object's center.
m3x3 r0, v0, c12 ; Object->World space position without translation.
mul oT5, r0, c94.w ; Scale by 1/appoximate scene size
ps.1.4
def c0, 0.9f, 0.0f, 0.0f, 0.0f
texld r0, t0 ; Look up normal map.
texcrd r1.rgb, t4 ; Eye vector
texcrd r4.rgb, t1 ; t row of environment matrix
Accurate Reflections and Refractions by Adjusting for Object Distance 293
Team LRN
texcrd r2.rgb, t2 ; 21sst row of environment matrix
texcrd r3.rgb, t3 ; 3rd row of environment matrix
texcrd r5.rgb, t5 ; Cube map position adjustment
dp3 r4.r, r4, r0_bx2 ; 1st row of matrix multiply
dp3 r4.g, r2, r0_bx2 ; 2nd row of matrix multiply
dp3 r4.b, r3, r0_bx2 ; 3rd row of matrix multiply
lrp r2.rgb, c0.r, -r1, -r4 ; Refract by c0 = index of refraction fudge factor
add r2.rgb, r5, r2 ; Adjust for cube map center
phase
texld r2, r2 ; Sample cubic reflection map
texld r3, t0 ; Sample base map
texld r4, r4 ; Sample cubic diffuse map
texld r5, t0 ; Sample gloss map
mul r1.rgb, r5, r2 ; Specular = Gloss*Reflection
mad r0.rgb, r3, r4_x2, r1 ; Base*Diffuse + Specular
+mov r0.a, r3
The bold lines of the shader are the required lines for the environment map adjustment. This
technique makes reflections look more accurate, but it also makes refraction possible because
the eye is much more sensitive to the inaccuracies of refraction. This technique, including
shaders and source code, is used in the Treasure Chest demo available on the companion CD
and also at http://www. ati.com/na/pages/resource_centre/dev_rel/Treasurechest.html. There
are menus to view the reflection cube map that can be accessed after going to windowed mode
by pressing Alt+Enter.
References
[Apodaca00] Anthony A. Apodaca and Larry Gritz, Advanced RenderMan (Morgan Kaufmann
Publishers, 1999), pp. 212-215.
294 Part 3: Pixel Shader Tricks
Team LRN
UV Flipping Technique to
Avoid Repetition
Alex Vlachos
A common problem exists among many shaders that rely on scrolling two copies of the same
texture in slightly different directions. No matter what angle or speed the textures are scrolling,
they will eventually line up exactly and cause a visual hiccup. This scrolling technique is com-
monly used to gain the effect of having random patterns, like water caustics. With caustics, the
pattern appears to stop every once in a while even though it is scrolling at a constant speed.
This was also encountered when implementing the Ocean Water and Reflective and
Refractive Water shaders when we scrolled two copies of the same bump map in opposite
directions to produce high frequency waves (see Rendering Ocean Water and Rippling
Refractive and Reflective Water later in this book). The method we used to solve this problem
is to compute the two sets of texture coordinates like you normally would. Then immediately
before using those texture coordinates to fetch from textures, choose one of the sets of texture
coordinates and swap the U and V. This will effectively flip the texture along the diagonal.
Using this method, the intermittent hiccup that appears due to the two copies being per-
fectly aligned is now impossible since one of the textures is reflected about the diagonal of the
texture.
295
Team LRN
Photorealistic Faces with
Vertex and Pixel Shaders
Kenneth L. Hurley
Introduction
Achieving photorealistic faces in games has always been a goal of game developers. Todays
DX8-compatible hardware allows the game programmer to use lots of polygons for characters.
The developer can now use facial animation and facial shading as an important feature in
todays games. This article introduces the reader to a technique using bump mapping and cap-
tured lighting to get more photorealistic-looking images. It includes instructions on how to use
vertex and pixel shaders to use bump, diffuse, specular, and environment maps to achieve
photorealistic-looking faces. This technique is based on an idea from an artist named Ulf
Lundgren (see http://www.lost.com.tj/Ulf/artwork/3d/behind.html). There are some preprocess-
ing steps that are necessary to use the vertex and pixel shaders for this technique, and they are
explained in detail.
Software
Several commercial software packages were used for development of this technique. 3D Studio
Max 4.2 was used for the modeling portion. 2D image transformations were done in Adobe
Photoshop 5.5 and JASC Paint Shop Pro 7.0. These two image-processing programs are
used since neither one completely provides the functionality needed to accomplish separation
of lighting conditions.
Additional software needed to reproduce this effect is contained on the companion CD.
The software on the CD includes a diffuse cube map generator, a texture mapping cylindrical
wrapper, and the vertex and pixel shaders. Other software that was used includes NVASM, the
macro assembler from NVIDIA, DirectX 8.0, and Visual Studio 6.0. NVIDIAS assembler
(NVASM) was written by the author and is included on the CD for your convenience. NVASM
has built-in #defines and macros and integrates easily into Visual Studio. The project files
included on the CD include integration of the assembler into Visual Studio. You can click on
output lines from the assembler and be taken right to the error.
296
Team LRN
The vertex and pixel shader code uses a left-handed coordinate sys-
tem (Figure 1). The technique described can also be implemented using
a right-handed system such as the one that OpenGL uses.
Resources
The authors face was used in the following figures. The reasoning
behind the decision was one of scheduling. The facial hair makes the
job that much harder. The photographs in Figure 2 are the starting point
for this technique.
Figure 2: Front view (left) and side view (right)
3D Model
There are numerous ways to convert two photographs into a
3D model. The choice was made to use an artist named Ste-
phen Burke, who was kind enough to take a mannequin
model into 3DS Max and move the vertices to fit the photo-
graphs (see Figure 3). There are several other packages that
allow you to take two photographs and change a canonical
model of a face into the likeness of the subject. Most of these
packages were not adequate for what is required for this tech-
nique. Additionally, most software packages cost money and
do not export the data needed.
Most of the original work for converting a canonical
model is based on the original ideas of Takaaki Akimoto, et
al. [Akimoto]. There is also some good work that was done
by Won-Sook Lee, et al [Lee]. One of the best ways to auto-
matically fit a model to a set of photographs is to use DFFDS
(Dirichlet Free Form Deformations). This type of (FFD) Free Form Deformation shows the
best results when transforming a canonical head. All the work on DFFDS is originally based
on hand animation work by Laurent Moccozet, et al. [Moccozet].
Photorealistic Faces with Vertex and Pixel Shaders 297
Figure 1
Figure 3
Team LRN
Setup for Separation of Lighting Elements
Separating lighting equations from the photographs can be accomplished with a technique that
starts out by separation of the diffuse, specular, and bump map components from photographs.
Purists will probably point out that these techniques are not truly the lighting equations, but
then most 3D graphics techniques arent exact. What really matters here is that the technique
looks good.
The photographs in Figure 1 were taken in complete darkness with only the flash used for
lighting. A digital camera with a flash attached was used. The green background was built
from fabric purchased at a fabric store and glued onto some poster board. The color of the fab-
ric isnt the same as the color used in true green screen movie settings, but it is sufficient for
this technique. The digital camera was mounted to a tripod and placed about 4 feet in front of
the subject. Two pictures are taken of the subject, one from the front and one from the side.
The photographs were then cropped around the face and made into power of two textures.
Diffuse Lighting Component
There are a couple of assumptions that were made when trying to extract out the lighting char-
acteristics of the subjects face. First off, the lighting direction is known, as is the flash from
the camera. A complicated algorithm could have been used based on the normals of the 3D
model and fitting a diffuse function to each pixel to extract the diffuse texture map, but a sim-
pler approach was taken.
The diffuse part is a smattering of lights from all directions. For each pixel in the 3D
model, the incoming light sources for that pixel are a result of the environment reflecting light
and the face reflecting light. Skin absorbs most of the light that is being shone on it. Additional
lighting on skin is very subtle. From any given point, the light reflectance can be considered
coming mostly from the first reflection and direction illumination. The light energy falls off
dramatically after first reflection because of the absorption properties of the skin. Some energy
is deflected and the reflected light taken in these conditions is almost all being contributed
from the face itself.
The transformations needed to get a diffuse cube map
can be accomplished using imaging processing software. The
extraction process will use Adobe Photoshop and JASC Paint
Shop Pro. Start Adobe Photoshop and load the front-facing
picture into it. Now choose Filters->Smart blur and set the
radius to 50 and the threshold to 100. Select High in the
Quality drop-down box. The smart blur filter keeps the gen-
eral features of the face when blurring (see Figure 4). One
additional preprocessing step is to paint out the shadows from
the blurred picture. This processed picture will be used as the
diffuse component. In the photographs, there are shadows
below the chin (see Figure 4). It isnt really necessary to
remove these manually, as the smart blur degrades these to
such a degree that they wont be seen in the final rendering.
298 Part 3: Pixel Shader Tricks
Figure 4
Team LRN
Extracting the Bump and Specular Map
Now switch over to JASC Paint Shop Pro. Paint Shop Pro is easier to work with in terms of its
arithmetic operations on images. Load the original image and the blurred image into Paint
Shop Pro. Now use the Image->Arithmetic menu and select the original and the diffuse for the
two pictures to operate on. Use Function Subtract and make sure the operation is set to all
channels by selecting the All channels check box. Unselect the clip color values and make sure
the divisor is set to 1 and the bias is set to 128 (see
Figure 5a).
What this essentially does is take the blurred image
out of the original and adjust the contrast/brightness back
up to midrange values. The result should look like the
picture in Figure 5b. If you zoom in on this image, you
can see where the bumps are. You can also still see spec-
ular parts in the image.
If you look closely at
the tip of the nose, you
will notice that there are
specular highlights
there.
Separating Specular and Bump Maps
The separation of the specular portion of the texture map must be accomplished. Again, inside
of Paint Shop Pro, press Ctrl+C and Ctrl+V to make a copy of the image in a new window.
Now pick the menu item Colors->Histogram Functions->Histogram Adjustments. The trick to
this portion of the extraction is to adjust the histogram to pull out the specular component of
the face.
When the histogram adjustment window is opened, a spike in the center of the histogram
window is visible. This spike is very similar to a specular lobe. This may have to be adjusted
until the specular parts of the photograph show up in the final result. It isnt difficult to see the
specular reflections in the photographs.
Using Figure 6a as a reference, move the black horizontal slider (triangle) to just before
the spike. Then adjust the gray horizontal slider (triangle) to just inside the middle of the spike.
Turn down the midtone compression slider to 50. Now, adjust the maximum output level to
147, which is about half way. The output level was reduced, since the specular component of
the skin isnt very high. Hit the OK button, and you should have something very similar to Fig-
ure 6b.
Photorealistic Faces with Vertex and Pixel Shaders 299
Figure 5a Figure 5b
Team LRN
Since the specular component is only a specular value
and doesnt include color information, it can be put
into the alpha channel of the diffuse map. Load the
diffuse map that was created in Figure 4, and select
the window for this image. Then use the
Masks->New->From Image menu item. Now select
the specular image in the Source Window drop-down
box. Make sure Source Luminance is checked and
Invert Mask Data is unchecked. Dont worry about
what the mask selection looks like, as it isnt really
representative of what the alpha channel will look
like. Now use Masks->Save to Alpha Channel and
click OK. Rename this Alpha instead of the default
Selection #1.
Now the mask must be removed. Otherwise,
when the image is being saved, it will flatten the
mask onto your image layer, creating an unwanted
masked diffuse image. It should now be saved into
one of the formats that supports 32 bits of data. Targa
(tga) files should be used, as the code included on the
companion CD uses this format for 32-bit images.
When the image is saved, Paint Shop Pro will give a
warning that the image is being saved in .tga format
and only one alpha channel will be saved. Click Yes
and let it merge the layers; since there is no longer a
mask, it will do the right thing with the rest of the
channels, creating a 32-bit texture.
The pixel shaders use these 32-bit images to grab
two images at once. One image is encoded in the
RGB values, and the second image is encoded in the
alpha component. The diffuse texture will be stored in the RGB, and the specular component
will be stored in the A (alpha) channel.
Normal Maps
Lastly, a normal map needs to be created for the face. Nor-
mal maps are a close relative to bump maps but are slightly
different. Figure 5b was converted from the diffuse/original
into what really constitutes a height map. This height map is
biased from the value 128, since 128 was added to center the
values in the arithmetic operation that was performed on the
image.
NVIDIAs web site has a plug-in for Photoshop that con-
verts a height map into a normal map. Load in the image
that was created similar to Figure 5b into Photoshop. Now
use the Filter->nvTools->Normal Map Filter menu
300 Part 3: Pixel Shader Tricks
Figure 6a
Figure 6b
Figure 7
Team LRN
selection. The bias RGB radio button should be selected. Also, the scale of the bumps should
be set to 5. The results can be seen in Figure 7. This image might be hard to see, as normals are
encoded in color values in the texture map. Color plates for this chapter are included on the
companion CD.
More on Normal Maps
Normals maps are not bump maps in the Blinn sense of the word. Jim Blinn [Blinn78] talked
about how to perturb a normal. For instance, an interpolated vertex normal can be perturbed by
some amount using a texture map lookup. Normal maps are a slight variation on this technique,
as they dont use the normal in the calculation. Instead they encode the normal in the RGB
color components of the texture map. RGB values are unsigned bytes, and normals are repre-
sented in the 1 to 1 range. We can use a technique to scale and bias a normal into the 0 to
255 range.
Since normals are used in dot product calculations, we want to encode the normals with a
direction facing the light vector in order to get the correct result of the dot product. To encode
these values, the sample vector {0, 0, 1}, which means 0 in the x direction, 0 in the y direc-
tion, and 1 in the z direction, is used. Encoding these values into the 0-255 ranges of each
RGB value involves first scaling the values by multiplying by .5 (divide by 2). Now 1
becomes .5 and 1 become .5. Then a bias is applied to the values so they fall into the 0 to 1
range. To do this, we add .5. Now .5 becomes 0, 0 becomes .5, and .5 becomess 1.0.
This basically remapped the negative and positive ranged floating-point values into posi-
tive-only values. Finally, multiplying these values by 255 puts them into the 8 bits of precision
of an RGB value. Since a normal will be used in the dot product calculation for the diffuse
lighting computation, we can see that the {0, 0, 1} would be remapped to the {128, 128, 255}
values, giving that kind of sky blue color seen in normal maps.
Since all the normals are defined in this flat 2D texture map, how can these normals be
useful? The technique is very similar to texture mapping. Figure 8 shows how a triangle might
be texture mapped with a texture. The flat plane represents a texture map. Only one triangle is
shown on the texture map and one tri-
angle on the sphere for clarity. The
normals on the sphere are at each
vertex.
These normals are usually used in
calculating the diffuse lighting solution.
Since the normals from the object are
not being used, but the normals are
being looked up instead from the tex-
ture map, some calculations will need
to be done. The normals have to either
be projected back onto the sphere for
each pixel, which is computationally
very expensive, or you need to do one
other trick.
Photorealistic Faces with Vertex and Pixel Shaders 301
Figure 8
Team LRN
The following technique is useful to use when optimizing an application. Instead of transform-
ing the texture-mapped normals, the light can be transformed at each of the vertices into the
texture map plane to do the diffuse calculation. Basis vectors are used to transform the light at
each vertex. Normal maps are used on a per-pixel basis, so there is one other thing to consider.
Texture mapping an object involves dividing by the W component to get perspective-cor-
rect texture mapping. This divide happens on a per-pixel basis. The basis vectors are
calculations that take these divides by W into account when texture mapping or looking up the
normal in the texture map.
Basically, a vector is used that encodes three things. The first component is the change in
the U (S for OpenGL) texture coordinate (delta S) divided by the change in the polygon X
value in object space (delta X). The next two components are the change in U with respect to
the delta Y and delta Z.
The second basis vector is the change in the V (T for OpenGL) texture coordinate with
respect to the three deltas in object space: X, Y, and Z. Of course, you need to have a third vec-
tor to have a rotation matrix. This third coordinate can be generated with a cross product of the
first two vectors or the vertex normal can be used, as it is a very good close approximation.
These calculations create the basis vector at each vertex that rotates the light vector into
texture mapping space, also known as texture space. These calculations are done in the vertex
shader. There are a couple of caveats to be aware of when generating the basis vectors. 3D con-
tent creation applications such as 3DS Max allow the artist to create mirrored and duplicate
texture coordinates, and this can be a problem when making the basis vector calculations. The
calculated vector may also have to be flipped if the result of the cross product is pointing in the
opposite direction of the face normal.
A Word About Optimizations
To save substantial amounts of vertex calculation work, a simple calculation on the CPU can
be made. The CPU can be used to do an inverse transform of the light vector to object space.
The vertex calculations rotate all the normals of an object by the objects matrix to put the nor-
mals in world space in order to do the diffuse lighting calculation. Depending on the number of
normals, this can be quite a few rotations. Instead of rotating the normals, you can rotate the
light into the object space and cut out all but one rotation. The first pass of the code presented
in the vertex shader (Listing 1) doesnt use this optimization, but the second code listing (List-
ing 2) shows just how much time can be saved by utilizing this optimization.
Pixel shaders also have this notion of co-issue instructions that is documented in the
DirectX 8.0 documentation. Basically, this means that you can have two instructions execute in
one clock cycle on hardware that supports it. The pixel shader in this chapter uses that tech-
nique and will be described in that section.
Another way to optimize is to use NVIDIAs NVTune tool available on the NVIDIA web
site. It is a free tool that registered developers can use to help optimize applications. It helps
determine where your code is slowing down in terms of how the hardware is dealing with the
data sent to it.
302 Part 3: Pixel Shader Tricks
Team LRN
Half Angle
Half angles are used in the first few vertex shaders. A half angle vector in the context of these
shaders is the vector that is exactly halfway between the light direction vector and the eye vec-
tor. This is really useful in calculating specular without having to use a reflection calculation.
Jim Blinn discovered this information in 1977; consult [Blinn77] for further information.
Vertex Shader Explanation Pass 1 (Listing 1)
The following explanation describes the vertex and pixel shaders that use the maps described
previously for this technique. This vertex shader sets up the pixel shader that will use a diffuse
map, a specular map, and a bump map, all in the same pass. Throughout the code, the term
light direction is used. This really means the opposite of the light direction for the reason
that the arithmetic operations, such as the dot product, require this inverted vector to work
correctly.
Line numbers have been added to the code for correlation between the explanations and
the code snippets.
Lines 20-27 are definitions that allow organization of temporary register usage. Line 29
tells the assembler to expect a vertex shader and which version. In this case, version 1.0 of a
vertex shader was chosen. Lines 32-35 of the code transform the vertex position by the concat-
enated world->view->projection matrix, which puts the vertex position in projection space
for rasterizing.
Lines 38-44 transform the basis vectors to world space. Remember these lines, as they will
play an important role in Listing 2, which is an optimized version of Listing 1. Lines 46-48
transform the normal into world space. Lines 50-51 scale the basis vector to increase or
decrease the effect of the normal with respect to the light. Lines 53-56 transform the light
direction vector by the world transformed basis vectors to put the light direction in texture
space. Lines 59-61 normalize the light vector, which may or may not be necessary. If a bump
scale value was not being used and the basis vectors were all normalized, then the light direc-
tion would only rotate around the unit sphere and the normalization would not be necessary.
Lines 71-73 transform the vertex position to world space so we can make a vertex position
to eye position vector for calculating a half angle. Some of these lines will be optimized out in
Listing 2. Line 82 does the calculation and lines 86-88 normalize the half angle vector. Lines
99-102 transform this vector by the world transformed basis vectors to put the half angle in
texture space. Lines 108-109 output the texture space light direction to the texture coordinates
for use in the pixel shader. Finally, lines 112-113 output the texture coordinates for looking up
texels in the pixel shader for the diffuse and normal texture maps.
Listing 1: Vertex shader unoptimized
1: /******************************************************************************
2: Path: C:\Development\PhotoFace\
3: File: dot3_directional.nvv
4:
5: Copyright (C) 2001 Kenneth L. Hurley - Free for Any use whatsoever
6: This file is provided without support, instruction, or implied warranty of any
7: kind. Kenneth L. Hurley makes no guarantee of its fitness for a particular
purpose
8: and is not liable under any circumstances for any damages or loss whatsoever
9: arising from the use or inability to use this file or items derived from it.
Photorealistic Faces with Vertex and Pixel Shaders 303
Team LRN
10:
11: Comments:
12: Kenneth Hurley - 10/15/2001
13:
14: ******************************************************************************/
15:
16:
17: #include "PF_Constants.h"
18:
19:
20: #define S_WORLD r0
21: #define T_WORLD r1
22: #define SxT_WORLD r2
23: #define LIGHT_LOCAL r3
24: #define VERTEX_WORLD r4
25: #define EYE_VECTOR r5
26:
27: #define HALF_ANGLE r7
28:
29: vs.1.0
30:
31: ; Transform position to clip space and output it
32: dp4 oPos.x, V_POSITION, c[CV_WORLDVIEWPROJ_0]
33: dp4 oPos.y, V_POSITION, c[CV_WORLDVIEWPROJ_1]
34: dp4 oPos.z, V_POSITION, c[CV_WORLDVIEWPROJ_2]
35: dp4 oPos.w, V_POSITION, c[CV_WORLDVIEWPROJ_3]
36:
37: ; Transform basis vectors to world space
38: dp3 S_WORLD.x, V_S, c[CV_WORLD_0]
39: dp3 S_WORLD.y, V_S, c[CV_WORLD_1]
40: dp3 S_WORLD.z, V_S, c[CV_WORLD_2]
41:
42: dp3 T_WORLD.x, V_T, c[CV_WORLD_0]
43: dp3 T_WORLD.y, V_T, c[CV_WORLD_1]
44: dp3 T_WORLD.z, V_T, c[CV_WORLD_2]
45:
46: dp3 SxT_WORLD.x, V_NORMAL, c[CV_WORLD_0]
47: dp3 SxT_WORLD.y, V_NORMAL, c[CV_WORLD_1]
48: dp3 SxT_WORLD.z, V_NORMAL, c[CV_WORLD_2]
49:
50: mul S_WORLD.xyz, S_WORLD.xyz, c[CV_BUMP_SCALE].w
51: mul T_WORLD.xyz, T_WORLD.xyz, c[CV_BUMP_SCALE].w
52:
53: ; transform light by basis vectors to put it into texture space
54: dp3 LIGHT_LOCAL.x, S_WORLD.xyz, c[CV_LIGHT_DIRECTION]
55: dp3 LIGHT_LOCAL.y, T_WORLD.xyz, c[CV_LIGHT_DIRECTION]
56: dp3 LIGHT_LOCAL.z, SxT_WORLD.xyz, c[CV_LIGHT_DIRECTION]
57:
58: ; Normalize the light vector
59: dp3 LIGHT_LOCAL.w, LIGHT_LOCAL, LIGHT_LOCAL
60: rsq LIGHT_LOCAL.w, LIGHT_LOCAL.w
61: mul LIGHT_LOCAL, LIGHT_LOCAL, LIGHT_LOCAL.w
62:
63:
64: /////////////////////////////////////
65: // Calculate half angle vector
66:
67: // transform vertex position to world space
68: // to calculate V, vector to viewer in world
69: // space.
304 Part 3: Pixel Shader Tricks
Team LRN
70:
71: dp4 VERTEX_WORLD.x, V_POSITION, c[CV_WORLD_0]
72: dp4 VERTEX_WORLD.y, V_POSITION, c[CV_WORLD_1]
73: dp4 VERTEX_WORLD.z, V_POSITION, c[CV_WORLD_2]
74:
75: ; Half angle vector is (L+V)/||L+V|| or Normalize( L+V )
76: ; ||a|| is magnitude of a
77: ; L = vec to light from vertex point
78: ; V = vec to viewer from vertex point
79:
80: // vertex position - eye position
81: // eye position - vertex position
82: add EYE_VECTOR, c[ CV_EYE_POS_WORLD ], -VERTEX_WORLD.xyz
83:
84:
85: ; Normalize eye vec
86: dp3 EYE_VECTOR.w, EYE_VECTOR, EYE_VECTOR
87: rsq EYE_VECTOR.w, EYE_VECTOR.w
88: mul EYE_VECTOR, EYE_VECTOR, EYE_VECTOR.w
89:
90: // Add them to average & create half angle vector
91: add HALF_ANGLE, c[CV_LIGHT_DIRECTION], EYE_VECTOR
92:
93:
94: ; Normalize it
95: dp3 HALF_ANGLE.w, HALF_ANGLE, HALF_ANGLE
96: rsq HALF_ANGLE.w, HALF_ANGLE.w
97: mul HALF_ANGLE, HALF_ANGLE, HALF_ANGLE.w
98:
99: dp3 oT3.x, HALF_ANGLE, S_WORLD
100: dp3 oT3.y, HALF_ANGLE, T_WORLD
101: dp3 oT3.z, HALF_ANGLE, SxT_WORLD
102: mov oT3.w, c[ CV_ONE ]
103:
104:
105: /////////////////////////////////////
106: // move light vector to TC 2
107:
108: mov oT2, LIGHT_LOCAL
109: mov oT2.w, c[CV_ONE].w // w to 0
110:
111: ; output tex coords
112: mov oT0, V_TEXTURE
113: mov oT1, V_TEXTURE
Vertex Shader Explanation Pass 1 (Listing 2)
The following vertex shader is an optimized version of the one discussed in Listing 1. It does
the rotation into object space on the CPU on a per-object basis instead of doing it on the GPU
on a per-vertex basis.
Lines 1-35 havent changed from Listing 1, so no further explanation is needed. Lines
38-39 scale the basis vectors that are in object space to increase the amount of bumpiness on
the object. Lines 38-48 of Listing 1 have been completely eliminated. This is due to the fact
that we rotate everything into object space in the C++ code, thereby eliminating the costly
per-vertex calculations.
Photorealistic Faces with Vertex and Pixel Shaders 305
Team LRN
Lines 42-44 transform the light direction, which is in object space, directly into texture
space. Lines 47-49 renormalize the light vector. Again, this might not have been necessary had
there been no scale factor for the light direction to affect the bumps. Line 61 now uses the eye
position in object space to do the calculation of vertex position to eye position, and calcula-
tions to transform the vertex position to world space have been eliminated. Lines 64-66 are the
same and normalize this new eye vector.
Line 69 creates the half angle vector from the object space light vector and the object
space eye vector. Lines 72-74 renormalize the half angle vector. Lines 77-80 transform this
half angle vector by the basis vectors to put the half angle vector into texture space. Lines
86-91 are the same as the end of Listing 1.
With a little math on the CPU, the elimination of quite a few lines of code in the vertex
shader was achieved. These calculations were being performed on a per-vertex basis. The
instruction count was reduced from 40 instructions per vertex to 28 instructions per vertex, a
reduction of 30%.
Listing 2: Vertex shader optimized
1. /******************************************************************************
2. Path: C:\development\PhotoFace\
3. File: dot3_directional.nvv
4.
5. Copyright (C) 2001 Kenneth L. Hurley - Free for Any use whatsoever
5. This file is provided without support, instruction, or implied warranty of any
6. kind. Kenneth L. Hurley makes no guarantee of its fitness for a particular
purpose
7. and is not liable under any circumstances for any damages or loss whatsoever
8. arising from the use or inability to use this file or items derived from it.
9.
10. Comments:
11. Kenneth Hurley - 10/15/2001
12.
13. ******************************************************************************/
14.
15.
16. #include "PF_Constants.h"
17.
18.
19. #define S_OBJECT r0
20. #define T_OBJECT r1
21. #define SxT_OBJECT r2
22. #define LIGHT_OBJECT r3
23. #define VERTEX_WORLD r4
24. #define EYE_VECTOR r5
25.
26. #define HALF_ANGLE r7
27.
28. vs.1.0
29.
30. ; Transform position to clip space and output it
31. dp4 oPos.x, V_POSITION, c[CV_WORLDVIEWPROJ_0]
32. dp4 oPos.y, V_POSITION, c[CV_WORLDVIEWPROJ_1]
33. dp4 oPos.z, V_POSITION, c[CV_WORLDVIEWPROJ_2]
34. dp4 oPos.w, V_POSITION, c[CV_WORLDVIEWPROJ_3]
35.
36. // scale bumps by amount requested
37. mul S_OBJECT.xyz, V_S.xyz, c[CV_BUMP_SCALE].w
38. mul T_OBJECT.xyz, V_T.xyz, c[CV_BUMP_SCALE].w
306 Part 3: Pixel Shader Tricks
Team LRN
39. ; transform parallel light direction by basis vectors to put it into texture
space
40. dp3 LIGHT_OBJECT.x, S_OBJECT, c[CV_LIGHT_DIRECTION]
41. dp3 LIGHT_OBJECT.y, T_OBJECT, c[CV_LIGHT_DIRECTION]
42. dp3 LIGHT_OBJECT.z, V_SxT, c[CV_LIGHT_DIRECTION]
43.
44. ; Normalize the light vector
45. dp3 LIGHT_OBJECT.w, LIGHT_OBJECT, LIGHT_OBJECT
46. rsq LIGHT_OBJECT.w, LIGHT_OBJECT.w
47. mul LIGHT_OBJECT, LIGHT_OBJECT, LIGHT_OBJECT.w
48.
49. /////////////////////////////////////
50. // Calculate half angle vector
51.
52. ; Half angle vector is (L+V)/||L+V|| or Normalize( L+V )
53. ; ||a|| is magnitude of a
54. ; L = vec to light from vertex point
55. ; V = vec to viewer from vertex point
56.
57. // vertex position - eye position
58. // eye position - vertex position
59. add EYE_VECTOR, c[ CV_EYE_POS_OBJECT ], -V_POSITION.xyz
60.
61. ; Normalize eye vec
62. dp3 EYE_VECTOR.w, EYE_VECTOR, EYE_VECTOR
63. rsq EYE_VECTOR.w, EYE_VECTOR.w
64. mul EYE_VECTOR, EYE_VECTOR, EYE_VECTOR.w
65.
66. // Add them to average & create half angle vector
67. add HALF_ANGLE, c[CV_LIGHT_DIRECTION], EYE_VECTOR
68.
69. // Normalize it
70. dp3 HALF_ANGLE.w, HALF_ANGLE, HALF_ANGLE
71. rsq HALF_ANGLE.w, HALF_ANGLE.w
72. mul HALF_ANGLE, HALF_ANGLE, HALF_ANGLE.w
73.
74. // transform half angle into normal map space
75. dp3 oT3.x, HALF_ANGLE, V_S
76. dp3 oT3.y, HALF_ANGLE, V_T
77. dp3 oT3.z, HALF_ANGLE, V_SxT
78. mov oT3.w, c[ CV_ONE ]
79.
80.
81. /////////////////////////////////////
82. // move light vector to TC 2
83.
84. mov oT2.xyz, LIGHT_OBJECT.xyz
85. mov oT2.w, c[CV_ONE].w // w to 0
86.
87. ; output tex coords
88. mov oT0, V_TEXTURE
89. mov oT1, V_TEXTURE
Pixel Shader Explanation (Listing 3)
The following pixel shader modulates the specular map values with the base texture values and
adds the diffuse color. Figure 9 is a texture map that was encoded with a diffuse lighting solu-
tion in the RGB components and a specular lighting solution in the A (alpha) component. This
Photorealistic Faces with Vertex and Pixel Shaders 307
Team LRN
texture is selected into texture stage 3. Line 16 selects a pixel shader and a version of 1.1 for
the assembler.
Line 18 takes the texture coordinates from the vertex shader and does a texture lookup into the
diffuse or base texture of the face (see Figure 4 and Figure 6b). The RGB components contain
the diffuse texture, and the alpha channel contains the specular map. Line 19 takes the texture
coordinates from the vertex shader and does a texture lookup into the normal map texture (see
Figure 7). Lines 20-21 do a 2x2 matrix calculation or two dot product calculations with the
texture coordinates coming in from the vertex shader.
Texture coordinate set 2 is the light vector passed in by the vertex shader. There isnt a tex-
ture lookup for the texm3x2pad instruction. The texm3x2tex instruction first does a dot
product and uses the calculation from the texm3x2 pad to look up a texture. It grabs a texel
from the texture in Figure 9.
The C++ code sets the texture mode to clamp, so the result of the dot product will always
be mapped into the texture map. The postfix _bx2 modifier on the normal is necessary to
bias and scale the value back into the 1 to 1 range before doing the dot product operation.
Line 25 takes the result of N dot L (Normal * Light vector) and multiplies it by the base tex-
ture. The added _x2 modifier makes the diffuse light act a little brighter in the calculation.
Line 26 uses the + prefix modifier to tell the assembler that this instruction is going to
co-issue with the previous instruction. The pixel shader has the capability, when two adjacent
instructions operate on the RGB and the alpha component separately, to execute the instruc-
tions in a one clock cycle. Line 26 also multiplies the specular lobe by the specular map
associated with the face. Line 27 uses what is called a trinary operator. It multiplies the specu-
lar level with the diffuse color to get a specular color and then adds that value into the
computed diffuse solution in one clock cycle.
Listing 3: Pixel shader
1. /******************************************************************************
2. Path: C:\Development\PhotoFace\
3. File: ps_dot3.nvp
4.
5. Copyright (C) 2001 Kenneth L. Hurley - Free for Any use whatsoever
6. This file is provided without support, instruction, or implied warranty of any
7. kind. Kenneth L. Hurley makes no guarantee of its fitness for a particular
purpose
8. and is not liable under any circumstances for any damages or loss whatsoever
9. arising from the use or inability to use this file or items derived from it.
10.
11. Comments:
12. Kenneth Hurley - 10/15/2001
308 Part 3: Pixel Shader Tricks
Diffuse map Specular map
Figure 9
Team LRN
13.
14. ******************************************************************************/
15.
16. ps.1.1
17.
18. tex t0 // fetch base texture + specular in alpha
19. tex t1 // fetch normal map
20. texm3x2pad t2, t1_bx2 // u = ( t1=N ) dot ( t2=L )
21. texm3x2tex t3, t1_bx2 // v = ( t1=N ) dot ( t3=H )
22. // fetch texture 4 at (u,v) rgb = (N dot L)
23. // alpha = (N dot H)
24.
25. mul_x2 r1.rgb, t3, t0 // ( N dot L) * base texture
26. +mul r1.a, t0.a, t3.a // ( N dot H) * specular map
27. mad r0, r1.a, t0, r1 // (specular * base texture) + (diffuse Color)
Full Head Mapping
The real challenge is to get a head mapped all the way around. Cylindrical mapping helps quite
a bit. There are still problems with how to generate the textures. On the companion CD there is
a plug-in for 3D Studio Max that will help generate a single unfolded texture. The source code
for the plug-in is also on the CD, in case you would like to modify/enhance it. This software
aids in merging two orthogonal photographs into one texture map.
Photorealistic Faces with Vertex and Pixel Shaders 309
Diffuse Lighting only Specular Map
Normal Map Diffuse + Specular + Normal Map
Figure 10: Technique 1
Results
Team LRN
Getting Started
For the plug-in to work, a specific setup of the material editor
and mapping channels inside 3D Studio Max needs to be
done. Create a Multi/Sub-Object in the Material Editor of 3D
Studio Max. Set the number of sub-objects to four by clicking
on the Set Number button. Next create four sub-materials that
have textures. The first one should have its material ID set to
1. This one will be used for the cylindrical map that will be
applied to the final stage. The plug-in uses the Material IDS
to figure out which side of the head the 2D texture is being
applied to. The IDS are preset in the plug-in as follows:
1 Final Cylindrical Map
2 Right Side
3 Left Side
4 Top Side
5 Bottom Side
6 Front Side
7 Back Side
For the pictures that were used earlier, we will planar map
these pictures on the left, right (for the side view), and front
(for the front view). Here is how the material editor should
have been set up.
Mapping the Photographs
Now use the UVW Mapping modifiers to map the photographs onto the head. For mapping the
first material ID (1), just use cylindrical mapping with a dummy texture for now. Now map the
two photographs by adding UVW Mapping modifiers for each of the faces. In this example,
the left, right, and front were planar mapped onto the object. The companion CD contains a
sample head named KenHeadTM_front-FINAL.max
Creating a Single Texture Map
Make sure you have installed the unwrap.dlu plug-in in your 3D Studio Max plug-ins direc-
tory. Now select the Utilities panel (the hammer). Next select the More button, and then scroll
down until you find the utility named Unwrap Object Texture and click OK. Most of the
defaults are acceptable, except the Spline Blending radio button should be selected. The differ-
ences between no spline blending and spline blending can be seen in Figure 13. The algorithm
for selecting the triangles to render is based on using the normal of the triangle and doing a dot
product with the direction of the unit cube. This then selects the texture map, which is most
likely to contribute the best pixels to the triangle. There could have been some sort of simple
blending done, but spline blending gives good results.
310 Part 3: Pixel Shader Tricks
Figure 11
Team LRN
Peter Watje originally wrote the unwrap tool. The code has since been modified to include
texture mapping and spline blending. The original texture mapping code was written by Chris
Hecker and was modified to fit into the 3D Studio Max framework. The final processing code
is based on [Burt]. The difficult part of the code was decomposing the images into Gaussian
and Laplacian image pyramids. The code uses memory quite heavily and the plug-in will warn
you that it can take a while to generate the images and then merge them all back together. On a
Pentium III 500 MHz, it takes about 20 seconds. The time to generate the map is acceptable
given the results of the blending. Save the generated image to disk by pressing on the disk
icon.
Photorealistic Faces with Vertex and Pixel Shaders 311
Normal Based Normal Based with Spline Blending
Figure 12
Figure 13
Team LRN
Putting It All Together
Now that a single texture map has been created, the process at the
beginning of the chapter needs to be rerun. From this image, the
diffuse, specular, and bump map needs to be generated. Before you
export the file, you should remove the UVW mapping coordinates
for the front, left, and right views, leaving the cylindrical map
UVW mapping. Now change the map type to a normal bitmap and
load in the image that was saved. Modification of the UVW coor-
dinates is unnecessary because the image was generated from the
cylindrical map and they should line up automatically. The 3D
Studio Max rendering of the model can be seen in Figure 14.
Whats Next?
3D Studio Max does a pretty good job at rendering, but this chapter is aimed at putting this
into a real-time application, like a game. There are several new techniques for getting environ-
mental lighting that have been discussed on the web and in other publications
[Lundgren][Blinn77][Blinn78] [Miller][Debevec]. Most are not practical for todays hardware.
The next few sections will discuss some techniques that can be used with hardware that has
pixel shaders, such as the GeForce 3.
Environmental Lighting
Environmental lighting simulates lighting of objects based on all incoming light in an area.
This area can be outside, or it may be as small as one room. Radiosity has been used quite a
bit, but so far it hasnt been done in real time, except for pre-computed light maps. Another
technique that people have been using recently is called High Dynamic Range (HDR)
image-based rendering (see [Debevec]). The HDR images are stored in floating-point format,
which precludes them from being used in current hardware. There are some techniques that
allow multiple passes with reduced precision to work on GeForce 2 class hardware. After look-
ing at these sites, it was noticed that the lights could be broken up into three components
diffuse, specular, and directional diffuse. These terms are also used for Bidirectional Reflec-
tion Distribution Functions (BRDF). The next few sections will explain the techniques that
can be used to recover these lights and how to possibly use them with todays 3D graphics
hardware.
How to Get Diffusion Cube Maps
There is a program included on the CD with full source code that helps in making cube maps.
Its under the DCM subdirectory which stands for Diffusion Cube Maps. It aids in extraction
of cube maps from digital pictures. The pictures are of a completely reflective ball. The pro-
gram also allows you to draw an exclusion rectangle to remove the picture taker from the cube
map. Since this technique doesnt use anything that is completely reflective in the scene, the
exclusion rectangle can be ignored.
312 Part 3: Pixel Shader Tricks
Figure 14
Team LRN
To extract the reflection maps, first load in the picture and then use the mouse to draw the
ellipse enclosed in a rectangle. This rectangle should be stretched and moved so that the ellipse
falls on the edges of the ball. Then set which direction is associated with the picture in the
menu options. The pictures in the example use the negative X and negative Z direction (see
Figure 15).
Generating the Cube Maps
Now select the Generate menu option. After a few seconds, a window will come up with a
cube map similar to the one in Figure 16a. Now if the window that contains the cube map is
selected, the Generate menu item will generate a diffusion cube map like the one seen in Fig-
ure 16b. The image in 16b was gamma corrected in
Paint Shop Pro since the diffusion routine reduces
the gamma on the image.
For the specular component, the reflection map
was taken into JASC Paint Shop Pro 7.0, and the
Histogram Adjust was used as it was with the pho-
tographs of the faces. It wouldnt be hard to extract
a specular map from the picture in Figure 16a run-
ning a convolution filter on the reflective map with
the user inputting the specular component. These
cube maps can now be used to do environmental
lighting on the face.
Figure 16: Reflection map(a), diffusion map(b), and specular map(c)
Photorealistic Faces with Vertex and Pixel Shaders 313
Figure 15:
Negative X
sphere picture
(left) and nega-
tive Z sphere
picture (right).
Team LRN
The Pixel Shader to Use All These Maps
The pixel shader in Listing 4 is quite interesting for a shader. It uses all the available texture
units on a GeForce 3 and uses three instructions to combine everything. The vertex shader isnt
included in the text, as it is very similar to Listing 5, which is described later. It uses the texture
maps previously described to perform environmental lighting.
Lines 21-24 load up texture RGB values from the base (diffuse) texture from texture stage
0, a darkening map from texture stage 1, the diffusion map (see Figure 16b), and finally the
specular map (see Figure 16c). The darkening map in this technique is similar to the one in
Figure 6b with the specular component removed. This map simply uses the diffusion cube map
and darkens the value, just like a dot product operation would do.
Line 26 multiplies the diffuse map by the environment diffusion cube map. Line 27 multi-
plies the specular map of the face by the environment map specular and adds in the values
calculated in Line 26. Line 29 multiplies the darkening factor by the diffusion environment
map and adds that to the result in line 27. All the operations add environmental lighting on the
face.
Listing 4: Pixel shader
1. /******************************************************************************
2. Path: C:\Development\PhotoFace
3. File: ps_EnvLight.nvp
4.
5. Copyright (C) 2001 Kenneth L. Hurley
6. This file is provided without support, instruction, or implied warranty of any
7. kind. Kenneth L. Hurley makes no guarantee of its fitness for a particular
purpose
8. and is not liable under any circumstances for any damages or loss whatsoever
9. arising from the use or inability to use this file or items derived from it.
10.
11. Comments:
12.
13. ******************************************************************************/
14.
15.
16. #include "PF_Constants.h"
17.
18.
19. ps.1.1
20.
21. tex t0 // fetch base texture
22. tex t1 // fetch bump map
23. tex t2 // fetch diffusion map using normal
24. tex t3 // fetch specular using reflection vector
25.
26. mul r0, t0, t2 // base map * diffusion map
27. mad r1, t0.a, t3, r0 // specular environment cubemap * specular base map +
28. // previous
29. mad r0, r0, t1, r1 // now use diffusion bump map * diffuse + previous
314 Part 3: Pixel Shader Tricks
Team LRN
Environment Mapping for Eyes
Environment mapping for the eyes is very important. Someone once said, The audience
always watches the eyes. This is why the eyes are very important to achieving very realis-
tic-looking faces. The outer portion of an eye is glossy in that it reflects a little bit of the
environment, but it is also clear, showing the inner part of the eye. This visual effect can be
reproduced in one pass on multitexturing hardware.
The vertex shader in Listing 5 sets up a couple of items. Lines 26-29 do the usual trans-
forming of the position. Lines 34-40 calculate the vertex to eye position needed for a reflection
calculation. Lines 45-47 calculate out 2 (NE)NE, which is the reflection vector calculation.
The vertex shader then stores this value into texture coordinate T1 for input to the pixel shader.
Listing 5: Vertex shader
1. /******************************************************************************
2. Path: C:\Development\PhotoFaceFile: vs_glossy.nvv
3.
4. Copyright (C) 2001 Kenneth L. Hurley - Free for Any use whatsoever
5. This file is provided without support, instruction, or implied warranty of any
6. kind. Kenneth L. Hurley makes no guarantee of its fitness for a particular
purpose
7. and is not liable under any circumstances for any damages or loss whatsoever
8. arising from the use or inability to use this file or items derived from it.
9.
10. Comments:
11. Kenneth Hurley - 10/15/2001
12.
13. ******************************************************************************/
14.
15.
16. #include "PF_Constants.h"
17.
18.
19. #define EYE_VECTOR r1
20. #define REFLECTION_VECTOR oT1
21.
22. vs.1.0
23.
24. ; Transform position to clip space and output it
25. dp4 oPos.x, V_POSITION, c[CV_WORLDVIEWPROJ_0]
26. dp4 oPos.y, V_POSITION, c[CV_WORLDVIEWPROJ_1]
27. dp4 oPos.z, V_POSITION, c[CV_WORLDVIEWPROJ_2]
28. dp4 oPos.w, V_POSITION, c[CV_WORLDVIEWPROJ_3]
29.
30. ; output tex coords
31. mov oT0, V_TEXTURE // for base texture lookup
32.
33. // vertex position - eye position
34. // eye position - vertex position
35. add EYE_VECTOR, c[ CV_EYE_POS_OBJECT ], -V_POSITION.xyz
36. // Normalize eye vec
37. dp3 EYE_VECTOR.w, EYE_VECTOR, EYE_VECTOR
38. rsq EYE_VECTOR.w, EYE_VECTOR.w
39. mul EYE_VECTOR, EYE_VECTOR, EYE_VECTOR.w
40.
41. /////////////////////////////////////
42. // Calculate reflection vector
43.
Photorealistic Faces with Vertex and Pixel Shaders 315
Team LRN
44. dp3 r3, V_NORMAL, EYE_VECTOR
45. mul r4, r3, V_NORMAL
46. mad REFLECTION_VECTOR, r4, CV_TWO, -EYE_VECTOR
The pixel shader in Listing 6 is pretty simple. It is passed a reflection vector and the regular
texture coordinates for the diffuse map. It then looks up the texture RGB values in lines 21 and
22. Texture stage 1 was set up with a cube map in the source code. Lines 24-26 simply use a
blend factor to blend between the two maps with values specified in two pixel shader constant
registers. These were also set up in the C++ code.
Listing 6: Pixel shader
1. /******************************************************************************
2. Path: C:\Development\PhotoFace
3. File: ps_Glossy.nvp
4.
5. Copyright (C) 2001 Kenneth L. Hurley
6. This file is provided without support, instruction, or implied warranty of any
7. kind. Kenneth L. Hurley makes no guarantee of its fitness for a particular
purpose
8. and is not liable under any circumstances for any damages or loss whatsoever
9. arising from the use or inability to use this file or items derived from it.
10.
11. Comments:
12.
13. ******************************************************************************/
14.
15.
16. #include "PF_Constants.h"
17.
18.
19. ps.1.1
20.
21. tex t0 // fetch base texture
22. tex t1 // fetch specular using reflection vector
23.
24. mul r1, t0, c[CP_BLEND_DIFFUSE]
25. mul t3, t1, c[CP_BLEND_SPECULAR]
26. add r0, t3, r1
Final Result
Figure 17 and Color Plate 5 show the
final result using the environmental
effects. Another item that should eventu-
ally be tackled is hair. This addition of
hair would visually enhance the render-
ing a great deal.
316 Part 3: Pixel Shader Tricks
Figure 17
Team LRN
Facial Animation
Facial animation is a very important part to creating photorealistic faces. Frederic I. Parke and
Keith Waters wrote an excellent book on facial animation [Parke]. They have released their
public domain software for animating faces. The software requires that you type the data for
the face and its corresponding muscles into a text file. There is another system for facial ani-
mation available at http://expression.sourceforge.net. This system allows you to attach the
muscles to triangles using bones inside 3D Studio Max. To efficiently use this system, some
code would have to be written in order to use vertex shaders. The code from Parke and Waters
can easily be modified to use vertex shaders by simply putting the 18 muscle groups that they
have defined in vertex shader constants. Each muscle requires ten floating-point values that
can be stored in 45 vertex shader constants. The muscle.c file would need modification to be
used as a vertex shader. This would then enable the vertex shader to move the vertices based
on the original Parke and Waters code.
There are more utilities that would need to be written for this system to work well, as it is
very difficult and cumbersome to extract out the exact data needed to tie triangles to the mus-
cle. The jaw, eye, tongue, and teeth animation systems would need to be defined. Finally, in
order to get speech hooked into the system, a speech system would have to be written or one of
the publicly available ones could be used. Microsoft Speech API is a good choice. Another
good choice is the University of Edinburgh free speech API, The Festival Speech Synthesis
System, available at http://www.cstr.ed.ac.uk/projects/festival/. This project is a free and open
source, which might make it more attractive for some.
Another piece of the facial animation puzzle deals with animating the face for realis-
tic-looking speech synchronization. Once the speech API is hooked up to the animation
system, there might be anomalies in the animation. This visually represents itself as a flutter-
ing effect between phoneme or word changes. One way to solve this problem is with the use
of a coarticulation system. Cohen and Massaro developed a fairly robust system [Cohen]. Sim-
ply put, this is a system that blends lip sync positions based on a weighting value. More
information on this system can be obtained from http://mambo.vcsc.edu/psl/ca93.html.
Conclusion
A lot of software was written for this chapter of the book. Some of it is unpolished, and the
author would have liked to add some new features to the code. One good example is the DCM
code. It does not use multi-resolution splines at the moment, and there are some artifacts in the
cube maps, which would have easily been corrected with this code. Also, the DCM code could
have used a specular convolution filter to automatically extract a specular map. The source
code is included in case anyone wants to enhance the code.
Special thanks goes to Greg James at NVIDIA for the original idea of encoding the diffuse
and specular maps into one 32-bit texture and for the basis of some of the code. Matthias
Wloka deserves a special thank you for his help with some of the more difficult math problems
and some ideas for recovering cube maps from photographs.
Photorealistic Faces with Vertex and Pixel Shaders 317
Team LRN
References
[Akimoto] Takaaki Akimoto, Yasuhito Suenaga, and Richard S. Wallace, Automatic Creation
of 3D Facial Models, IEEE Computer Graphics and Applications, 1993.
[Blinn77] Jim Blinn, Models of Light Reflection for Computer Synthesized Pictures.
Computer Graphics (SIGGRAPH Proceedings), July 1977, pp. 192-198.
[Blinn78] Jim Blinn, Simulation of Wrinkled Surfaces, Computer Graphics (SIGGRAPH
Proceedings), Vol. 12, No. 3, August 1978, pp. 286-292.
[Burt] Peter J. Burt and Edward H. Adelson, A Multiresolution Spline with Application to
Image Mosiacs, ACM Transactions on Graphics, Vol. 2. No 4, October 1983.
[Cohen] M. M. Cohen and D. W. Massaro, Modeling coarticulation in synthetic visual
speech. N. M. Thalmann and D. Thalmann (Eds.) Models and Techniques in Computer
Animation. (Springer-Verlag, 1993).
[Debevec] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin,
and Mark Sagar, Acquiring the Reflectance Field of a Human Face (SIGGRAPH
Proceedings), 2000.
[Lee] Won-Sook Lee, Prem Karla and Nadia Magnenat Thalmann, Model Based Face
Reconstruction for Animation, http://miralabwww.unige.ch/ARTICLES/ MMM97t.html,
1998.
[Lundgren] Ulf Lundgren, 3D Artwork Description, http://www.lost.com.tj/Ulf/artwork/3d/
behind.html, May 2001.
[Miller] Gene S. Miller and Robert C. Hoffman, Illumination and Reflection Maps: Simulated
Objects in Simulated and Real Environments, Course Notes for Advanced Computer
Graphics Animation, SIGGRAPH, 1984.
[Moccozet] L. Moccozet and N. Magnenat-Thalmann, Dirichlet Free-Form Deformations and
their Application to Hand Simulation, Computer Animation Proceedings, IEEE Computer
Society, 1997, pp. 93-102.
[Parke] Frederic I. Parke and Keith Waters, Computer Facial Animation (AK Peters LTD,
1996).
318 Part 3: Pixel Shader Tricks
Team LRN
Non-Photorealistic
Rendering with Pixel
and Vertex Shaders
Drew Card and Jason L. Mitchell
Introduction
Development of cutting-edge graphics techniques, like programmable pixel and vertex shaders,
is often motivated by a desire to achieve photorealistic renderings for gaming or simulation. In
this chapter, we apply vertex and pixel shaders to non-photorealistic rendering (NPR). In many
types of images, such as cartoons and technical drawings, photorealism is not desirable. In the
case of technical illustrations, non-photorealistic rendering techniques are used to enhance
understanding of the scene or object being drawn without obscuring important features such as
outlines. In other cases, we simply hope to simulate other media such as cel-shaded cartoons,
wood-block prints, or hatched line drawings for stylistic purposes. In the following sections,
we will apply Direct3D pixel and vertex shaders to implement and extend recent research
efforts in non-photorealistic rendering.
Rendering Outlines
Rendering of object outlines is a common step in non-photorealistic rendering. In this section,
we will present a geometric approach to outline rendering, which uses vertex shaders to deter-
mine silhouette edges. These outlines are used with the NPR shading techniques discussed in
subsequent sections. An image space approach to outlining will be presented at the end of
this article.
We consider the silhouette outline to be the minimum set of lines needed to represent the
contour and shape of the object. Silhouette edges represent not only the outer edges of the
object but also points of surface discontinuity (e.g., a sharp edge or crease in a surface).
319
Team LRN
The basic algorithm involves drawing every edge of an object as a quadrilateral fin and doing
the silhouette determination in the vertex shader. This is very similar to the way that
[Lengyel01] faded the fins used in fur rendering. The goal of this shader is very simple; if
the edge is a silhouette, the vertex shader renders the quad fin; otherwise, the vertex shader
renders a degenerate (unseen) fin.
The vertex shader determines if an edge is a silhouette by comparing the face normals of
the triangles that share the edge (n
face0
and n
face1
). If one normal is front facing with respect to
the viewer and the other is back facing, then the edge is a silhouette. This algorithm works per-
fectly except in the case of edges that are not shared by more than one triangle. These kinds of
edges are considered boundary edges and need to be drawn all of the time. Boundary edges
only have one face normal associated with them, so there is no second normal to be used in the
comparison. In order to ensure that boundary edges are always drawn, the second shared nor-
mal is chosen so that it is the negation of the first normal (n
face1
= n
face0
). This results in
boundary edges always being drawn, since one normal will always be facing the viewer and
the other will always be facing away from the viewer. With this organization of data, one ver-
tex buffer and one rendering call can be used to draw all of the quad fins, and the vertex shader
will handle both regular edges and boundary edges.
The data for drawing the object will typically be stored in memory in an organization similar to
that shown in Figure 2. As shown, every vertex of every edge is composed of the vertex posi-
tion (p), along with three normal values corresponding to the vertexs smooth normal (n
smooth
)
and the face normals of the triangles sharing the edge (n
face0
and n
face1
). The application should
then render each edge quad fin, possibly passing the additional normal information in a
320 Part 3: Pixel Shader Tricks
Figure 1: N L, silhouette outlines,
image filtered outlines
Figure 2: Vertex buffer orga-
nization for silhouette edge
rendering
Team LRN
separate data stream from the rest of the vertex. In order to reduce the memory bandwidth hit
from reading in the extra data (n
face0
and n
face1
), one optimization would be to quantize the
face normals to byte 4-tuples or short 4-tuples, as illustrated in Character Animation with
Direct3D Vertex Shaders and Vertex Decompression Using Vertex Shaders.
The vertex shader computes the view vector in camera space by multiplying the vertex by
the view matrix and normalizing. The shader then transforms the face normals n
face0
and n
face1
into camera space and dots them with the view vector. If the edge is a silhouette edge, one of
these dot products will be negative and the other will be positive. The shader checks for this
condition by multiplying the two dot products together and checking for a value less than zero.
If the value is less than zero, the vertex offset is set to zero (unchanged); otherwise, the vertex
offset is set to one. The vertex offset is then multiplied by the smooth normal and added to the
untransformed vertex position. Note that two of the four vertices for each quad fin have n
smooth
= 0. This acts as a mask of the fin vertex displacement and causes two of the fin vertices to
stick to the model while the other two are displaced to cause the fin to be visible.
Listing 1: Outline vertex shader
//NPR outline shader
// c0-3 view matrix
// c4-7 view projection matrix
// c8
// c9 (0.0, 0.0, 0.0, 1.0f)
// 10 line with scalar
vs 1.1
m4x4 r0, v0, c0 // compute the view vector
dp3 r1, r0, r0 // normalize the view vector
rsq r1, r1
mul r0, r0, r1
m3x3 r1, v7, c0 // multiply normal 1 by the view matrix
m3x3 r2, v8, c0 // multiply normal 2 by the view matrix
dp3 r3, r0, r1 // dot normal 1 with the view vector
dp3 r4, r0, r2 // dot normal 2 with the view vector
mul r3, r3, r4 // multiply the dot products together
slt r3, r3, c9 // check if less than zero
mov oD, c9 // set the output color
dp4 r0, v0, c6 // compute the vertex depth
mul r0, r0, c10 // multiply by a line thickness scalar
mul r3, r3, r0 // multiply the thickness by the smooth normal
mul r3, v3, r3 // multiply by the normal offset
add r0, v0, r3 // add in the offset
mov r0.w, c9.w // swizzle in a one for the w value
m4x4 oPos, r0, c4 // transform the vertex by the model view projection
Hidden line removal is handled via the depth buffer. We assume that a shaded version of the
model is rendered before the outlines to fill the z buffer with values that will cause hidden out-
lines to fail the z test. The following pseudocode outlines this process:
1. Preprocess the geometry into quad fins:
a. For each vertex of each edge, store the edge vertex, the smooth surface normal, and
the two face normals which share said edge; one should have the smooth normal, and
the other should have the smooth normal field set to zero.
b. If edge is unshared (only used in one face), store the negation of the one face normal
as the second normal.
Non-Photorealistic Rendering with Pixel and Vertex Shaders 321
Team LRN
2. Render a shaded version of the geometry to fill the z-buffer.
3. Enable outline vertex shader and initialize the shader constant storage.
4. Set up stream mappings to pass in the additional normal data.
5. Render the edges as triangles.
6. Vertex shader breakdown:
a. Compute the view vector by transforming the vertex into eye space (multiply by the
view matrix) and normalize.
b. Dot each face normal with the view vector.
c. Multiply the resulting dot products together.
d. Check for a value less than zero.
e. Multiply the smooth normal by the result of the less than zero test.
f. Compute the vertex depth (dot the vertex with the third row of the view projection
matrix).
g. Multiply the vertex depth by the line thickness factor to get a normal scale value.
h. Multiply the smooth normal by the normal scale value.
i. Add the smooth normal to the untransformed vertex.
j. Transform the vertex and output.
There are some drawbacks associated with the previous algorithm. Along with the hassle of
preprocessing the geometry and storing extra edge data, boundary edges may potentially be
drawn incorrectly when a quad fin points straight at the viewer. This is because the algorithm
currently only scales the edge along the smooth surface normal, thereby leaving no means to
screen-align the quadrilateral edge. This could be addressed by reworking the algorithm to also
screen-align the quad. Later in this article, we present an image space approach to rendering
outlines, which requires no preprocessing and does not exhibit the same boundary edge issue.
In the next section, we will discuss methods for shading the interior of the object to
achieve different stylized results.
Cartoon Lighting Model
One method of cartoon shading is to create banded regions of color to represent varying levels
of lighting. Recent examples of 3D games using cel-shading techniques are Cel Damage by
Pseudo Interactive and the Jet Set Radio games (called Jet Grind Radio
in some markets) by Sega/Smilebit. A common technique illustrated in
[Lake00] is a technique called hard shading, which shades an object
with two colors that make a hard transition where NL crosses zero.
[Lake00] indexes into a 1D texture map to antialias the transition, while
the method shown here computes the colors analytically. Figure 3
shows an approach that uses three colors to simulate ambient (unlit),
diffuse (lit), and specular (highlight) illumination of the object.
This is accomplished using a vertex shader that computes the light
vector at each vertex and passes it into the pixel shader as the first tex-
ture coordinate. The vertex shader also computes the half angle vector
at each vertex and passes it into the pixel shader as the second texture
coordinate. The pixel shader analytically computes the pixel color. As
322 Part 3: Pixel Shader Tricks
Figure 3: Cartoon
shaded with outlines
Team LRN
shown in Listing 2, the pixel shader first computes NL and NH. If the NL term is above a
specified threshold, the diffuse color is output; otherwise, the ambient color is output. If the
NH term is above a specified threshold, then the specular color replaces the color from the
NL term. This same analytical method could be expanded to use any number of banded
regions.
Listing 2: Cartoon shading vertex shader code
// Cartoon vertex shader
// c9 is the light position
// c10 is the view projection matrix
// c14 is the view matrix
vs.1.1
// output the vertex multipled by the mvp matrix
m4x4 oPos, v0, c10
// compute the normal in eye space
m3x3 r0, v3, c14
mov oT0, r0 // write the normal to tex coord 0
// compute the light vector
sub r0, c9, v0
dp3 r1, r0, r0
rsq r1, r1
mul r0, r0, r1
m3x3 r1, r0, c14 // transform the light vector into eye space
mov oT1, r1 // write the light vector to tex coord 1
// compute half vector
m4x4 r0, v0, c14 // transform the vertex position into eye space
dp3 r3, r0, r0 // normalize to get the view vector
rsq r3, r3
mul r0, r0, r3
add r0, r1, r0 //add the light vector and the view vector = half angle
dp3 r3, r0, r0 // normalize the half angle vector
rsq r3, r3
r0, r0, r3
mov oT2, r0 // write the half angle vector to tex coord 2
Listing 3: Cartoon shading pixel shader code
// Cartoon shading pixel shader
//
ps.1.4
def c0, 0.1f, 0.1f, 0.1f, 0.1f // falloff 1
def c1, 0.8f, 0.8f, 0.8f, 0.8f // falloff 2
def c2, 0.2f, 0.2f, 0.2f, 1.0f // dark
def c3, 0.6f, 0.6f, 0.6f, 1.0f // average
def c4, 0.9f, 0.9f, 1.0f, 1.0f // bright
// get the normal and place it in register 0
texcrd r0.xyz, t0
// get the light vector and put it in register 1
texcrd r1.xyz, t1
// compute n dot 1 and place it in register 3
dp3 r3, r0, r1
// subtract falloff 1 from the n dot 1 computation
sub r4, r3, c0
// check if n dot 1 is greater than zero
Non-Photorealistic Rendering with Pixel and Vertex Shaders 323
Team LRN
if yes use average color otherwise use the darker color
cmp_sat r0, r4, c3, c2
// subtract falloff 2 from the n dot 1 computation
sub r4, r3, c1
// check if n dot 1 is greater than zero
// if yes use bright color otherwise use what's there
cmp_sat r0, r4, c4, r0
The ambient and diffuse bands help to visualize the shape of the object while the specular
highlight gives insight into the properties of the surface of the object. If the goal of the cartoon
shader is only to represent the objects shape, the shader could omit the specular portion and
replace it with any number of additional diffuse regions.
Hatching
Another method of NPR shading is hatching, which is commonly used in pen and ink drawings
to show shape and differentiate between lit and unlit regions of an object. The density of the
hatch pattern signifies how much light the surface is reflecting at that point. The current state
of the art in real-time hatching is illustrated in [Praun01]. This technique uses an array of hatch
patterns ranging from very sparse (well-lit) to very dense (unlit) hatching called tonal art
maps. NL is computed per-vertex and used to determine a weighted-average of the tones in the
tonal art map. Per-pixel, the tonal art maps are blended together according to the weights inter-
polated from the vertices. The result is a hatched image that is well antialiased.
324 Part 3: Pixel Shader Tricks
Figure 4: Cartoon shaded object with
specular and multiple diffuse regions
Figure 5: Hatched object Figure 6: Tonal art map from [Praun01]
Team LRN
Listing 4: Hatching vertex shader
// Hatching vertex shader
//
// c0 0(0.0, 0.0, 0.0, 0.0)
// c1 1(1.0, 1.0, 1.0, 1.0)
// c2 2(2.0, 2.0, 2.0, 2.0)
// c3 3(3.0, 3.0, 3.0, 3.0)
// c4 4(4.0, 4.0, 4.0, 4.0)
// c5 5(5.0, 5.0, 5.0, 5.0)
// c6 6(6.0, 6.0, 6.0, 6.0)
// c7 7(7.0, 7.0, 7.0, 7.0)
// c8 brightness
// c9 light position
// c10 view projection matrix
// c14 view matrix
//
vs.1.1
m4x4 oPos, v0, c10 // output the vertex multiplied by the mvp matrix
mov oT0, v7 // write out the texture coordinate
mov oT1, v7
mov r1, v3 //normalize the normal
mov r1.w, c0
dp3 r2, r1, r1
rsq r2, r2
mul r1, r1, r2
sub r2, c9, v0 // compute the light vector and normalize
dp3 r3, r2, r2
rsq r3, r3
mul r2, r2, r3
dp3 r3, r2, r1 // compute the light factor (n dot 1) times six clamp at zero
mul r3, r3, c8
mov r5.x, c5.x // seed the blend weights
mov r5.y, c4.x
mov r5.z, c3.x
mov r5.w, c0.x
mov r6.x, c2.x
mov r6.y, cl.x
mov r6.z, c0.x
mov r6.w, c0.x
sub r5, r3, r5 // sub each weight's initial value from the light factor
sub r6, r3, r6
max r5, r5, c0 // get rid of everything less than zero
sge r7, c2, r5 // flag all weights that are <= 2
mul r5, r5, r7 // zero out weights > 2
sge r7, r5, c1 // flag all weights that are >= 1
mul r7, r7, c2 // subtract all weights that are greater than or equal to one
from 2
sub r5, r7, r5
slt r7, r5, c0 // flag all weights that are < 0 and negate
sge r8, r5, c0 // flag all spots that are >= 0
add r7, r7, r8 // add the flags
mul r5, r5, r7 // should negate the negatives and leave the positives
max r6, r6, c0 // same as above only on the second set of weights
sge r7, c2, r6
Non-Photorealistic Rendering with Pixel and Vertex Shaders 325
Team LRN
mul r6, r6, r7
sge r7, r6, c1
mul r7, r7, c2
sub r6, r7, r6
slt r7, r6, c0
sge r8, r6, c0
add r7, r7, r8
mul r6, r6, r7
sge r8, c1, r3 // check for total shadow and clamp on the darkest texture
mov r7, c0
mov r7.z, r8.z
add r6, r6, r7
min r6, r6, c1
mov oT2.xyz, r5 // write the 123 weights into tex coord 3
mov oT3.xyz, r6 // write the 456 weights into tex coord 4
Listing 5: Hatching pixel shader
// Hatching pixel shader
ps.1.4
tex1d r0, t0 // sample the first texture map
tex1d r1, t1 // sample the second texture map
texcrd r2.rgb, t2.xyz // get the 123 texture weights and place it in register 2
texcrd r3.rgb, t3.xyz // get the 456 texture weights and place it in register 3
dp3_sat r0, 1r0, r2 // dot the reg0 (texture values) with reg2 (texture
weights)
dp3_sat r1, 1r1, r3 // dot the reg1 (texture values) with reg3 (texture
weights)
add_sat r0, r0, r1 // add reg 0 and reg 1
mov_sat r0, 1r0 // complement and saturate
Gooch Lighting
The Gooch lighting model introduced in [Gooch98] is designed to provide lighting cues with-
out obscuring the shape of the model, the edge lines, or specular highlights. This technique,
designed to model techniques used by technical illustrators, maps the 1 to 1 range of the dif-
fuse NL term into a cool-to-warm color ramp. This results in diffuse lighting cues, which are
shown as hue changes rather than color intensity changes. This diffuse lighting model is
designed to work with the silhouette and feature-edge lines discussed earlier in this article. It
essentially results in a reduction in the dynamic range of the diffuse shading so that the edge
lines and specular highlights are never obscured. A similar technique is used in the game
Half-Life by Valve Software [Birdwell01]. The Half-Life engine first computes a single
approximate aggregate light direction. The 1 to 1 result of the per-vertex NL from this aggre-
gate light direction is then scaled and biased into the 0 to1 range rather than simply clamped at
zero. This eliminates the flat look that would otherwise be apparent on the side of a game char-
acter that faces away from the light.
As shown in [Gooch98], the classic lighting equation illustrated in Rippling Refractive
and Reflective Water can be generalized to the following, which allows us to experiment with
cool-to-warm colors.
326 Part 3: Pixel Shader Tricks
Team LRN
Listing 6: Gooch lighting vertex shader
// Gooch Lighting vertex shader
// c9 is the light position
// c10 is the view projection matrix
// c14 is the view matrix
vs.1.1
m4x4 oPos, v0, c10 // output the vertex multiplied by the vp matrix
sub r0, c9, v0 // compute the light vector and normalize
dp3 r1, r0, r0
rsq r1, r1
mul r0, r0, r1
mov r1, v3 // compute the normal
mov oT0, r1 // write the normal to tex coord 0
mov oT1, r0 // write the light vector to tex coord 1
Listing 7: Gooch lighting pixel shader
// Gooch Lighting pixel shader
// c0 is alpha (eg. {0.4, 0.4, 0.4, 1.0})
// c1 is beta (eg. {0.5, 0.5, 0.5, 1.0})
// c2 is k
yellow
(eg. {0.5, 0.5, 0.0, 1.0})
// c3 is k
blue
(eg. {0.0, 0.0, 0.4, 1.0})
// c4 is k
d
(eg. {0.0, 0.0, 0.4, 1.0})
// c5 is (1.0, 1.0, 1.0, 1.0)
ps.1.4
texcrd r0.xyz, t0 // get the normal and place it in register 0
texcrd r1.xyz, t1 // get the light vector and put it in register 1
dp3 r3, r0, r1 // compute n dot 1 and place it in register 3
add_d2 r3, r3, c5 // normalize the n dot 1 range
mul_sat r0, c4, c0 // compute the cool factor
add_sat r0, r0, c2
mul_sat r0, r0, r3
mul_sat r1, c4, c1 // compute the warm factor
add_sat r1, r1, c3
mad_sat r0, r1, 1r3, r0 // add the warm and cool together and output
In the preceding sections, we have concentrated on shader techniques that render
non-photorealistic images directly into the frame buffer. In the following section, we will look
at image space techniques that require rendering into textures and subsequent processing of
these rendered images to produce non-photorealistic images.
Non-Photorealistic Rendering with Pixel and Vertex Shaders 327
Figure 7: Gooch lighting with outlining on a
teapot and some cartoon characters
Team LRN
Image Space Techniques
As discussed in Image Processing with Pixel Shaders in Direct3D, it is possible to render 3D
scenes into textures for subsequent image processing. One technique developed in [Saito90]
and refined in [Decaudin96] is to render the depth and world space normals of objects in a
scene into a separate buffer. This rendered image is subsequently post-processed to extract
edges which can be composited with a hatched, Gooch shaded, or cartoon shaded scene. We
will show a Direct3D implementation of this technique as well as our own extension, which
thickens the lines using morphological techniques discussed in Image Processing. One
advantage of an image space approach to determining outlines is that it is independent of the
rendering primitives used to render the scene or even whether the models are particularly
well-formed. A scene that contains N-Patch primitives [Vlachos01], for example, will work
perfectly well with an image space approach, as will interpenetrating geometry such as the
classic Utah teapot. This approach even works with user clip planes (or the front clip plane),
correctly outlining areas of normal or depth discontinuity in the final image, without any appli-
cation intervention at the modeling level. Another advantage is that this approach does not
require the creation and storage of the auxiliary outline buffers discussed in the first section of
this article.
The first step of this technique is to use a vertex shader to render the world space normals
and depths of a scene into a texture map. The vertex shader scales and biases the world space
normals from the 1 to 1 range into the 0 to 1 range and writes them to diffuse r, g and b
(oD0.xyz). The eye-space depth is written into diffuse alpha (oD0.w). This interpolator is sim-
ply written out to the RGBA render target (i.e., a texture) by the pixel shader. One important
detail is that the clear color for the scene should be set to world space +z, so the filter will
interact properly with the objects at all orientations. An image of some cartoon characters ren-
dered with this technique is shown in the following figures and in Color Plate 7. The RGBA
texture containing world space normals and eye space depths is shown in Figure 8.
The vertex shader used to render this scene is shown in Listing 8:
Listing 8: World space normals and eye space depth vertex shader
vs.1.1
m4x4 oPos, v0, c0
mov r0, v3
mov r0.w, c12.w
add r0, c8, r0
mul r0, r0, c9
m4x4 r1, v0, c4
328 Part 3: Pixel Shader Tricks
Figure 8: World space normals
and eye space depth as in
[Saito90] and [Decaudin96].
These are rendered to RGB and
A of a renderable texture map.
Team LRN
sub r1.w, r1.z, c10.x
mul r0.w, r1.w, c11.x
mov oD0, r0
The next step is to use image processing techniques like those shown in Image Processing to
extract the desired outlines. The normals and depths are effectively processed independently to
isolate different classes of edges. Discontinuities in the normals occur at internal creases of an
object, while depth discontinuities occur at object silhouettes, as shown in Figure 9. Note that
while the normal discontinuity filter picks up the edge at the top of the leg and the depth dis-
continuity filter does not, the union of the edge pixels from the two filters produces a
reasonable outline for the object.
Figure 9: (a) Edges from world space normal discontinuities, (b) depth discontinuities,
and (c) both together.
Listing 9: Determining edges from an image of a scenes world space normals
// Normal discontinuity filter for Non-Photorealistic Rendering
ps.1.4
def c0, 1.0f, 1.0f, 1.0f, 1.0f
def c1, 0.85f, 0.0f, 1.0f, 1.0f
def c2, 0.0f, 0.0f, 0.0f, 0.0f
// Sample the map five times
tex1d r0, t0 // Center Tap
tex1d r1, t1 // Down/Right
tex1d r2, t2 // Down/Left
tex1d r3, t3 // Up/Left
tex1d r4, t4 // Up/Right
dp3 r1.rgb, r0_bx2, r1_bx2 // Take dot products with center pixel (Signed result
1 to 1)
dp3 r2.rgb, r0_bx2, r2_bx2
dp3 r3.rgb, r0_bx2, r3_bx2
dp3 r4.rgb, r0_bx2, r4_bx2
// Subtract threshold
add r1, r1, c1.r
add r2, r2, c1.r
add r3, r3, c1.r
add r4, r4, c1.r
phase
// Make black/white based on threshold
cmp r1.rgb, r1, c0.r, c2.r
+mov r1.a, c0.a
cmp r2.rgb, r2, c0.r, c2.r
Non-Photorealistic Rendering with Pixel and Vertex Shaders 329
Team LRN
+mov r2.a, c0.a
cmp r3.rgb, r3, c0.r, c2.r
+mov r3.a, c0.a
cmp r4.rgb, r4, c0.r, c2.r
+mov r4.a, c0.a
mul r0.rgb, r1, r2
mul r0.rgb, r0, r3
mul r0.rgb, r0, r4
+mov r0.a, r0.r
Listing 10: Determining edges from an image of a scenes eye-space depth
// 5-tap depth-discontinuity filter
ps.1.4
def c0, 0.02f, 0.0f, 1.0f, 1.0f
tex1d r0, t0 // Center Tap
tex1d r1, t1 // Down/Right
tex1d r2, t2 // Down/Left
tex1d r3, t3 // Up/Left
tex1d r4, t4 // Up/Right
add r1, r0.a, r1.a // Take four deltas
add r2, r0.a, r2.a
add r3, r0.a, r3.a
add r4, r0.a, r4.a
cmp r1, r1, r1, r1 // Take absolute values
cmp r2, r2, r2, r2
cmp r3, r3, r3, r3
cmp r4, r4, r4, r4
phase
add r0.rgb, r1, r2 // Accumulate the absolute values
add r0.rgb, r1, r3
add r0.rgb, r1, r4
add r0.rgb, r0, c0.r // Substract threshold
cmp r0.rgb, r0, c0.g, c0.b
+mov r0.a, r0.r
Compositing the Edges
Once we have the image containing the world-space normals and depth, we composite the
edge-filtered result with the frame buffer, which already contains a hatched, cartoon, or
Gooch-shaded image. The output of the edge detection shader is either black or white, so we
use a multiplicative blend (src*dst) with the image already in the frame buffer:
d3d->SetRenderState (D3DRS_ALPHABLENDENABLE,TRUE);
d3d->SetRenderState (D3DRS_SRCBLEND, D3DBLEND_DESTCOLOR);
d3d->SetRenderState (D3DRS_DESTBLEND, D3DBLEND_ZERO);
This frame buffer operation is nice because we can multi-pass edge filters with the frame
buffer and get the aggregate edges. In the NPR sample on the companion CD, for example, we
do one pass for normal discontinuities and one for depth discontinuities. It is worth noting that
it would be possible to process both normal discontinuities and depth discontinuities using 1.4
pixel shaders and co-issue pixel shader instructions, but we chose to use a larger filter kernel
(and thus more instructions) in the sample shown here.
330 Part 3: Pixel Shader Tricks
Team LRN
Depth Precision
We have found that 8 bits of precision for eye-space depth works well for the simple scenes we
have tested, but we expect this to be a problem for more aggressive NPR applications such as
games with large environments. In scenes of large game environments, using only 8 bits of
precision to represent eye space depth will cause some that are close to each other to fuse
together if their world space normals are also similar. Because of this, it might be necessary to
use techniques that spread the eye space depth across multiple channels or to simply rely upon
future generations of hardware to provide higher precision pixels and texels.
Alpha Test for Efficiency
Since the black edge pixels are a very small subset of the total pixels in the scene, we can alpha
test the edge image to save frame buffer bandwidth during the composite. Note that the last
instruction in the depth and normal discontinuity shaders moves the red channel of the filter
result into the alpha channel of the pixel. This is done so that the alpha test functionality which
follows the pixel shader can be used to kill the pixel rather than composite it with the frame
buffer, speeding up performance. Since we want to kill white pixels, we set an alpha reference
value of something between white and black (0.5f) and use an alpha compare function of
D3DCMP_GREATER:
d3d->SetRenderState(D3DRS_ALPHATESTENABLE, TRUE);
d3d->SetRenderState(D3DRS_ALPHAREF, (DWORD) 0.5f);
d3d->SetRenderState(D3DRS_ALPHAFUNC, D3DCMP_GREATER);
Shadow Outlines
In addition to outlining object silhouettes, it is also desirable to outline shadows in a scene. We
have added this functionality to a stencil shadow application in the RADEON 8500 SDK on
the ATI web site as shown in Figure 10 and Color Plate 8.
Figure 10a shows a scene using the hatching technique from [Praun01] alone. In addition
to this, we use stencil shadows to generate an image in a texture which contains normals and
depths similar to the preceding technique. The application renders the scene into the texture
with world space normals and depths, using the stencil buffer to write to only pixels that are
not in shadow. The application then re-renders the geometry to the pixels in shadow but
negates the world space normals. This results in an image like the one shown in Figure 10b,
where the alpha channel (not shown) contains depths. The same normal and depth discontinu-
ity filters used above are applied to this image to determine both object and shadow edges in
one pass. These edges are composited over a hatched scene that already contains areas in
shadow which have been hatched with the densest hatching pattern to simulate shadow. Figure
10d shows this technique along with coloration of the TAMs and per-pixel TAM weight
determination.
Non-Photorealistic Rendering with Pixel and Vertex Shaders 331
Team LRN
Figure 10: (a) A plain outlined
hatched scene. (b) The
renderable texture containing
world space normals for
non-shadowed pixels and
negated normals for shadowed
pixels. (c) A hatched scene with
shadows, object outlines, and
shadow outlines. (d) Adding base
texture coloring and per-pixel
TAM weight calculation.
Thickening Outlines with Morphology
The outlines generated by the above technique are a few pixels thick and look fine for many
NPR applications, but some applications may want thicker lines. Rather than directly compos-
ite the edges onto a shaded image in the back buffer, we can render the edges into a separate
texture and apply morphological operations as shown in Image Processing. To thicken the
lines, use dilation; to thin them or break them up, use erosion. To give a different style, we
could thicken the lines by ping-pong rendering between renderable textures. After performing
the desired number of dilations, composite the thickened edge image back onto the shaded
frame buffer as discussed above.
Summary of Image Space Technique
Rendering a scene with the image space outlining technique shown here is done in the follow-
ing steps:
1. Render shaded scene to back buffer
2. Render world space normals and depths to a texture map
3. If thickening lines using morphology,
a. Clear renderable texture to white
b. Draw quad into texture using world space normal discontinuity filter. Use alpha test
and src*dst blending
c. Draw quad into texture using depth discontinuity filter. Use alpha test and src*dst
blending
332 Part 3: Pixel Shader Tricks
(a) (b)
(c) (d)
Team LRN
d. Dilate edges
e. Composite with shaded scene in back buffer by drawing full screen quad using alpha
test and src*dst blending
4. Else using edges directly,
a. Draw full-screen quad over whole screen using world space normal discontinuity fil-
ter. Use alpha test and src*dst blending
b. Draw full-screen quad over whole screen using depth discontinuity filter. Use alpha
test and src*dst blending
Conclusion
We have presented Direct3D shader implementations of some recent developments in
non-photorealistic rendering including outlines, cartoon shading, hatching, Gooch lighting, and
image space techniques. For an excellent overview of these and many other NPR techniques,
please refer to the recent book Non-Photorealistic Rendering by Gooch and Gooch. Gooch and
Gooch also have an excellent online reference for NPR research: http://www.cs.utah.edu/npr/.
Acknowledgments
Thanks to Eli Turner for 3D modeling and Chris Brennan for implementing the shadow outlin-
ing technique.
References
[Birdwell01] Ken Birdwell, Valve Software, personal communication, 2001.
[Decaudin96] Philippe Decaudin, Cartoon-looking rendering of 3D scenes, Technical Report
INRIA 2919, Universite de Technologie de Compiegne, France, June 1996.
[Gooch98] Amy Gooch, Bruce Gooch, Peter Shirley, and Elaine Cohen, A non-photorealistic
lighting model for automatic technical illustration, (SIGGRAPH Proceedings, 1998), pp.
447-452.
[Lake00] Adam Lake, Carl Marshall, Mark Harris, and Marc Blackstein, Stylized Rendering
Techniques for Scalable Real-Time 3D Animations, Non-Photorealistic Animation and
Rendering, 2000.
[Lengyel01] Jed Lengyel, Emil Praun, Adam Finkelstein, and Hugues Hoppe, Real-time fur
over arbitrary surfaces, ACM Symposium on Interactive 3D Graphics 2001, pp. 227-232.
[Praun01] Emil Praun, Hugues Hoppe, Matthew Webb, and Adam Finkelstein, Real-time
Hatching, (SIGGRAPH Proceedings, 2001), pp. 581-586. Figure 6 used with permission.
[Saito90] Saito, Takafumi, and Tokiichiro Takahashi, Comprehensible Rendering of 3-D
Shapes. (SIGGRAPH Proceedings, 1990), pp. 197-206.
[Vlachos01] Alex Vlachos, Jrg Peters, Chas Boyd, and Jason L. Mitchell, Curved PN
Triangles, ACM Symposium on Interactive 3D Graphics, 2001.
Non-Photorealistic Rendering with Pixel and Vertex Shaders 333
Team LRN
Animated Grass with
Pixel and Vertex Shaders
John Isidoro and Drew Card
Introduction
The availability of programmable vertex and pixel shaders has allowed programmers to
re-examine how they implement passive background scene geometry. Before the advent of pro-
grammable shaders, deformation animation was only used for characters and a select few scene
objects due to the CPU overhead associated with it. With the ability to transfer this processing
over to the GPU, previously static background objects can now be animated, adding life and
realism to the scene. In this article, an approach is shown for realistically and inexpensively
animating grass with a combination of vertex and pixel shaders.
Waving the Grass
The waving motion of the grass can be accomplished in a vertex shader. Using a traditional
method, the grass is rendered with randomly placed intersecting quads. The quads are texture
mapped and rendered with an alpha test. In a vertex shader, the top two vertices of each quad
are animated using a combination of four sinusoidal waves. The waves are approximated using
a Taylor Series approach laid out in Rendering Ocean Water. This combination of sine waves
using various different frequencies creates a natural waving that does not look like an anima-
tion or overly repetitious.
Lighting the Grass
When grass blades wave in the wind, they also turn and change their orientation with respect to
the sunlight. Because this method involves using a textured quad to represent various blades of
grass, it is impossible to change the lighting on individual grass blades. To simulate the change
in color that would occur as the blades twisted in the wind, the same sine wave that was used
to animate the grass is used to oscillate between two colors. In this case, the green channel of
the color was changed to make the grass color change from intense green to a more
yellowish-brown.
334
Team LRN
Vertex shader code:
//vertex shader for grass..
//sinusoidal vertex motion for waving grass
//pos + sumOverI(wavedirI * texcoordy * sin( xdirI * (xpos+time)) + ydirI *
(ypos+time)))
SetVertexShaderConstant 0 commonConst
SetVertexShaderConstant 1 appConst
SetVertexShaderConstant 2 worldSpaceCamPos
SetVertexShaderConstant 4 wvp
SetVertexShaderConstant 8 sin9
SetVertexShaderConstant 10 frcFixup
SetVertexShaderConstant 11 waveDistortx
SetVertexShaderConstant 12 waveDistorty
SetVertexShaderConstant 13 waveDistortz
SetVertexShaderConstant 14 waveDirx
SetVertexShaderConstant 15 waveDiry
SetVertexShaderConstant 16 waveSpeed
SetVertexShaderConstant 17 piVector
SetVertexShaderConstant 18 lightingWaveScale
SetVertexShaderConstant 19 lightingScaleBias
StartVertexShader
vs.1.1
mul r0, c14, v0.x // use vertex pos x as inputs to sinusoidal warp
mad r0, c15, v0.y, r0 // use vertex pos y as inputs to sinusoidal warp
mov r1, c1.x // get current time
mad r0, r1, c16, r0 // add scaled time to move bumps according to speed
frc r0.xy, r0 // take frac of all 4 components
frc r1.xy, r0.zwzw //
mov r0.zw, r1.xyxy //
Animated Grass with Pixel and Vertex Shaders 335
Figure 1: Grass texture
Figure 2: Results
Team LRN
mul r0, r0, c10.x // multiply by fixup factor (due to inaccuracy of taylor
series)
sub r0, r0, c0.y // subtract 0.5
mul r1, r0, c17.w // *=2pi coords range from(pi to pi)
mul r2, r1, r1 // (wave vec)^2
mul r3, r2, r1 // (wave vec)^3
mul r5, r3, r2 // (wave vec)^5
mul r7, r5, r2 // (wave vec)^7
mul r9, r7, r2 // (wave vec)^9
mad r0, r3, c8.x, r1 // (wave vec) ((wave vec)^3)/3!
mad r0, r5, c8.y, r0 // + ((wave vec)^5)/5!
mad r0, r7, c8.z, r0 // ((wave vec)^7)/7!
mad r0, r9, c8.w, r0 // ((wave vec)^9)/9!
dp4 r3.x, r0, c11
dp4 r3.y, r0, c12
dp4 r3.zw, r0, c13
mul r4, v7.y, v7.y //
mul r3, r3, v7.y // attenuate sinusoidal warping by tex0.y^2
mov r2.w, v0
add r2.xyz, r3, v0 // add sinusoidal warping to grass position
m4x4 oPos, r2, c4
dp4 r1.x, r0, c18 //scale and add sine waves together
mad oD0, c19.xzxz, r1.x, c19.y //scale and bias color values (green is scaled
// more than red and blue)
mov oT0, v7
EndVertexShader
Pixel shader code:
StartPixelShader
ps.1.4
texld r0, t0
mul_x2 r0, r0, v0
EndPixelShader
336 Part 3: Pixel Shader Tricks
Team LRN
Texture Perturbation
Effects
John Isidoro, Guennadi Riguer, and Chris Brennan
Introduction
The recent generation of graphics hardware has brought about a new level of per-pixel pro-
grammability for real-time graphics, especially with regard to dependent texture reads. With
the ability to perform per-pixel math both before and after texture fetches, a variety of new tex-
ture-based effects are now possible. This article will focus on per-pixel texture perturbation
effects, effects that use the results of a texture fetch to change the texture coordinates used to
fetch another texel. Effects such as wispy clouds, billowing smoke, flickering fire, rippling
water, and plasma can be rendered in real time using the techniques shown here.
Wispy Clouds
One of the easiest effects that can be achieved using texture perturbation is wispy clouds.
The idea is to scroll two tileable textures past each other in opposite directions. The first
texture is the perturbation texture, which has u and v perturbations stored in the red and green
channels of the texture.
The second texture is the cloud texture, which has its texture coordinates perturbed by the per-
turbation map. For this section, we use a grayscale cloud texture and pack it into the alpha
channel of the perturbation map. Because of this packing, only one texture is needed for each
337
Figure 1: Red channel of texture
(u perturbation)
Figure 2: Green channel of texture
(v perturbation)
Team LRN
layer of clouds. This texture is fetched from twice in the shader, first to obtain the perturbation
values, then to read the perturbed texture.
Here is the pixel shader code for the basic cloud shader:
ps.1.4
texld r0, t0
texcrd r1.rgb, t1
mad r0, r0, c0, r1 //multiply perturbation by scale factor, and add texcoords
phase
texld r0,r0
mov r0, r0
In order to make the clouds look much more realistic, however, its best to use multiple layers
of clouds. The layers can be combined in quite a few different ways. A straightforward
approach to combining the layers is to linearly interpolate between them within the pixel
shader. This gives the appearance of one layer of clouds on top of another. The cloud layers are
scrolled past each other at different speeds to give the effect of depth through parallax.
The following shader is the result of combining two layers of clouds with a horizon.
In order to give the clouds a color gradient from a grayscale texture map, we use a multi-
ply operation, which is scaled by two and then saturated. In the example, the color to be
modulated with the clouds is orange (RGB values: 1.0, 0.7, 0.5). The gradient that results from
the scale and modulation ranges between brown, orange, yellow, and white.
Cloud texture intensity Color after modulation
0.00 (0.000, 0.000, 0.000) black
0.25 (0.500, 0.375, 0.250) brown
0.50 (1.000, 0.750, 0.500) orange
0.75 (1.000, 1.000, 0.750) yellow
1.00 (1.000, 1.000, 1.000) white
Because the same grayscale texture is used for both intensity and transparency, the black
regions of the clouds are completely transparent. The two cloud layers are then linearly inter-
polated with the background. The background texture is just a gradient from purple to blue.
338 Part 3: Pixel Shader Tricks
Figure 3: Cloud texture Figure 4: Basic cloud shader in action
Team LRN
// SetPixelShaderConstant 0 psConstVec (0.0, 0.5, 1.0, 0.75)
// SetPixelShaderConstant 1 cloud0PerturbationFactor
// SetPixelShaderConstant 2 cloud0Color
// SetPixelShaderConstant 3 cloud1PerturbationFactor
// SetPixelShaderConstant 4 cloud1Color
ps.1.4
texld r0, t1 //perturbation map 0
texld r1, t3 //perturbation map 1
texcrd r2.rgb, t0 //cloud layer 0 tex coord
texcrd r3.rgb, t2 //cloud layer 1 tex coord
mad r0.rgb, c1, r0_bx2, r2 //perturb base map coords by cloud coords
mad r1.rgb, c3, r1_bx2, r3 //perturb base map coords by cloud coords
phase
texld r0, r0
texld r1, r1
texld r2, t4
mul_x2_sat r4.rgb, r0.a, c2 //cloud layer 0, multiply by color
+mul r4.a, r0.a, c2.a //cloud layer 0, alpha scale
mul_x2_sat r3.rgb, r1.a, c4 //cloud layer 1, multiply by color
+mul r3.a, r1.a, c4.a //cloud layer 1, alpha scale
lrp r0.rgb, r1.a, r3, r4 //interpolate between layers 0 and 1
+add_sat r0.a, r3.a, r4.a //add transparency values for layer
lrp_sat r0, r0.a, r0, r2 //interpolate between clouds and horizon
Cloud layers can be combined in ways other than simple linear interpolation. Another tech-
nique is to use the horizon texture to brighten the edges of the clouds around the moon to give
a light halo effect. This shader is quite complex. The images below show each step of the
shader.
Texture Perturbation Effects 339
Figure 5: Scene using two-layer
cloud shader with horizon
Figure 6: Horizon map Figure 7: Alpha of horizon map (glow map)
Team LRN
340 Part 3: Pixel Shader Tricks
Figure 8: Cloud layer 0 Figure 9: Cloud layer 1
Figure 10: Cloud edges of both maps combined
extracted from cloud texture
Figure 11: Combined cloud layers
Figure 12: Combined cloud alphas (transpar-
ency map)
Figure 13: Cloud Edges * Glow
Team LRN
// SetPixelShaderConstant 0 psConstVec (0.0, 0.5, 1.0, 0.75)
// SetPixelShaderConstant 1 cloud0DistortionFactor
// SetPixelShaderConstant 2 cloud0Color
// SetPixelShaderConstant 3 cloud1DistortionFactor
// SetPixelShaderConstant 4 cloud1Color
// SetPixelShaderConstant 5 horizonBoost
ps.1.4
texld r0, t1
texld r1, t3
texcrd r2.rgb, t0
texcrd r3.rgb, t2
mad r0.rgb, c1, r0_bx2, r2 //perturb base map coords by cloud coords
mad r1.rgb, c3, r1_bx2, r3 //perturb base map coords by cloud coords
phase
texld r0, r0
texld r1, r1
texld r2, t4
mov_x2 r3.rgb, r2.a //alpha of horizon (glow map)
+mul_sat r3.a, 1r0.a, 1r1.a //mask edges for cloud glow
mad_sat r0.rgb, 1r0.a, c2, c2.a //scale and bias inverted clouds layer 0
+mov_x2_sat r0.a, r0.a //boost cloud alpha 0
mad_sat r1.rgb, 1r1.a, c4, c4.a //scale and bias inverted clouds layer 1
+mov_x2_sat r1.a, r1.a //boost cloud alpha 1
lrp r0.rgb, r1.a, r1, r0 //combine cloud layers
+mov r4.a, c5.a //
mul_x2 r3, r3.a, r3 //multiply glow by edges of clouds
mad_x2 r3, r3, c5, r4.a //scale and bias glowing sections of clouds
mul r0.rgb, r3, r0 //multiply glow by combination of cloud
layers
+add_d2_sat r0.a, r0.a, r1.a //add alphas of cloud layers
lrp_sat r0, r0.a, r0, r2 //Lerp between clouds and horizon
Texture Perturbation Effects 341
Figure 14: Two cloud layers combined with
Cloud Edges * Glow
Figure 15: Final result: Cloud layers with glow-
ing edges blended with horizon
Team LRN
Perturbation-Based Fire
Another interesting effect that can be created using texture perturbation is flickering fire. Tra-
ditionally, this was accomplished as either an animation, a texture feedback effect, or a particle
system. However, each of these approaches has limitations.
Rendering fire as an animation has the disadvantages of looking unrealistic due to the
cycling animation and requiring a lot of texture memory to store the frames of animation. In
addition to this, if there are many flames in the scene and all use the same animation, the scene
will look even more artificial. To capture the randomness of fire, cycling animations are not a
good choice.
Another approach is to use a texture feedback. The basic idea is for each frame to generate
a row of random pixels at the bottom of the texture, then blur and scroll the image upward in
the texture. Although the visual effect of this is quite random, the repetitive blurring of the tex-
ture feedback process does not have the sharp edges that real fire has, and therefore looks very
unrealistic.
Particle systems can also be used to create fire effects. Even though the fire effects they
can produce are random in nature, calculating the physics for the particle system requires CPU
time and can be prohibitive to perform for the amount of particles needed to create a realistic
fire. In addition, particle fires tend not to have the opaqueness of real fire.
The technique described in this section for generating fire overcomes all of these short-
comings. Because it only requires two textures a perturbation map and a base map the
technique requires much less texture memory than the animated fire. Also, due to the way we
use the perturbation maps, the fire is both non-repetitive and preserves the sharp edges present
in the base map. Also, the shader can be programmed to use vertex data to seed the offsets of
the perturbation maps so that multiple flames in the same scene can have different
appearances.
The idea behind this shader is to scroll three perturbation texture images (all three use the
same texture) past each other vertically and add the results together to get a perturbation value.
The three maps scroll at rates that are not integer multiples of the others, so the fire does not
repeat as time advances. Using three maps ensures that repetition only happens after a very
long time. When two of the perturbation maps line up, the third one is almost always offset
from the other two. Giving the maps slightly different scroll speeds in the horizontal direction
as well prevents the mechanical upward scrolling appearance of the fire.
The perturbation value is attenuated slightly based on the interpolated texture coordinate v
value and then used to perform a perturbed fetch of a pixel in the base map. The attenuation is
used to weaken the perturbation near the base of the fire in order to make the top of the fire
flicker more than the base. After this, the perturbed fetch is performed in the same way as in
the cloud shader, but both RGB components and alpha components are used from the fetched
pixel.
342 Part 3: Pixel Shader Tricks
Team LRN
//SetPixelShaderConstant 0 texLerp
//SetPixelShaderConstant 1 distortAmount0
//SetPixelShaderConstant 2 distortAmount1
//SetPixelShaderConstant 3 distortAmount2
//SetPixelShaderConstant 4 heightAtten
//SetPixelShaderConstant 5 fireBoost
ps.1.4
texcrd r0.rgb, t0
texld r2, t1
texld r3, t2
texld r4, t3
mul r2.rgb, c1, r2_bx2 //noise0 * dis0
mad r2.rgb, c2, r3_bx2, r2 //noise1 * dis1 + noise0 * dis0
mad r2.rgb, c3, r4_bx2, r2 //noise2 * dis2 + noise1 * dis1 + noise0 * dis0
+mad r0.a, r0.g, c4.x, c4.y //scale and bias y coord of base map
mad r2.rgb, r2, r0.a, r0 //mult distortion by scaled biased y coord
Texture Perturbation Effects 343
Figure 16: Fire effect in action
Figure 17: Base map
(rgb)
Figure 18: Opacity map
(alpha)
Figure 19: Distortion
texture
Team LRN
phase
texld r0, r2
mov r0, r0
Plasma Glass
In addition to the simulation of real-world phenomena, there are many other effects possible
using texture perturbation. For instance, many science fiction movie-style special effects such
as warping plasma are possible. The following technique produces something similar to the
glass plasma spheres that emanate streams of plasma toward the surface of the glass when
touched.
The concept behind the plasma effect is to use two scrolling textures again, but use each
one as both a perturbation map and a base map. The base map is stored in the RGB compo-
nents of the texture, and the perturbation map is stored in the alpha component. Each texture
map perturbs the subsequent texel fetch of the other map. The following shader shows the
basic plasma effect.
//SetPixelShaderConstant 1 Perturbation scaling
ps.1.4
texld r1, t1 //Perturbation map 1 / base map 0
texld r2, t2 //Perturbation map 0 / base map 1
texcrd r4.rgb, t1 //Base map coords 0
texcrd r5.rgb, t2 //Base map coords 1
mad r4.rgb, c1, r2.a, r4 //Use alpha for Perturbation
mad r5.rgb, c1, r1.a, r5 //Use alpha for Perturbation
phase
texld r1, r5 //Base map 0
texld r2, r4 //Base map 1
add_d2 r0, r1, r2 //Add two perturbed textures together
Mathematically, this effect is almost identical to the two-layer cloud effect, where the layers
are added together, but the appearance of the effect is drastically different. Most of the differ-
ence in appearance can be attributed to using very different base and perturbation textures.
The plasma effect can be used as a building block to create more complex effects. For
instance, the plasma effect can be combined with a glass effect to create a glass plasma sphere.
344 Part 3: Pixel Shader Tricks
Figure 20: Base texture Figure 21: Perturbation
texture
Figure 22: Simple plasma
effect applied to a sphere
Team LRN
The first step is to use an environment cube map:
To help give this a more glass-like appearance, the enviroment map is combined with a
per-pixel Fresnel term. The per-pixel Fresnel term allows the glass to have a slightly bumpy
appearance, which can be used to create a bumpy or a frosted glass appearance.
The environment map is linearly interpolated with the plasma effect to produce the final result.
For the final alpha value, the per-pixel Fresnel term and the plasma transparency value are
added together.
Here is the pixel shader for the plasma glass effect.
//SetPixelShaderConstant 0 psConstVec (0.0, 0.5, 1.0, 0.75)
//SetPixelShaderConstant 1 psDistortScaleFactor
//SetPixelShaderConstant 2 glassColor
//SetPixelShaderConstant 3 plasmaColor
Texture Perturbation Effects 345
Figure 23: Environment map
Figure 24: Environment map *
Per-pixel Fresnel term * glass color
Figure 25: Final effect (interpolate
between environment and plasma)
Team LRN
ps.1.4
texld r0, t0 //bump map
texld r1, t1 //perturbation map/ base map
texld r2, t2 //perturbation map/ base map
texld r3, t3 //tan space view vector lookup into
normalization cube map
texcrd r4.rgb, t1 //base map coords
texcrd r5.rgb, t2 //base map coords
dp3 r0, r3_bx2, r0_bx2 //V.N
mad r4.rgb, c1, r2.a, r4 //calc perturbed texture coords
+cmp r0.a, r0, r0, r0 //abs(V.N)
mad r5.rgb, c1, r1.a, r5 //calc perturbed texture coords
mov_sat r0, 1r0.a //1abs(V.N) (Fresnel term)
phase
texcrd r0, r0 //pass through fresnel
texld r1, r5 //perturbed base texture 0
texld r2, r4 //perturbed base texture 1
texld r4, t4 //env map reflective lookup.
mul_x2 r4.rgb, r4, c2 //reflection map * glass color
+add_d2_sat r1.a, r1.r, r2.r //textures together
mul r1.a, r1.a, r1.a //square base texture (increases contrast)
mul_x4_sat r1.rgb, r1.a, c3 //(plasma^n) * plasma color
+mov r0.a, r0.r //fresnel term
lrp r0.rgb, r0.a, r4, r1 //lerp plasma and environment map
+mad_sat r0.a, r1.a, c3.a, r0.a //plasma * plasma alpha + fresnel
Summary
A variety of effects have been created using texture perturbation. Several other effects are also
possible using a variation of the above shaders, such as rippling water or billowing smoke. By
merely changing the base and perturbation textures, a vast array of new and interesting effects
can be achieved. It is recommended that you experiment with these shaders yourself. More
information on texture perturbation can be found in the references. The shaders described in
this chapter can be seen in action in the video files included on the companion CD.
References
[Ebert98] David S. Ebert, F. Kenton Musgrave, Darwyn Peachey, Ken Perlin, and Steven
Worley, Texturing & Modeling: A Procedural Approach, 2nd Ed., (AP Professional, July
1998).
346 Part 3: Pixel Shader Tricks
Team LRN
Rendering Ocean Water
John Isidoro, Alex Vlachos, and Chris Brennan
Introduction
In computer graphics, simulating water has always been a topic of much research. In particular,
ocean water is especially difficult due to the shape and combination of multiple waves, in addi-
tion to the sun in the sky and the reflection of the clouds.
The shader in this article is meant to simulate the appearance of ocean water using vertex
and pixel shaders. The interesting part of this shader is that it runs completely in hardware in a
single pass on recent graphics cards (Radeon 8500). This has the advantage of leaving the CPU
free for other calculations, as well as allowing for a courser tessellation of the input geometry
that can be tessellated using N-Patches or other higher order surface schemes. The input geom-
etry is a grid of quads with one set of texture coordinates and tangent space, though in theory
only a position is actually needed if assumptions are made about the orientation of the up vec-
tor and the scale of the ocean waves in the shader.
This shader is best explained by separating the vertex shader and the pixel shader while
keeping in mind the final result, shown in Figure 1.
347
Figure 1
Figure 2
Team LRN
Sinusoidal Perturbation in a Vertex Shader
The vertex shader is responsible for generating the combination of sine waves that perturb the
position and the cosine waves that perturb the tangent space vectors for the vertex. A Taylor
series approximation is used to generate sine and cosine functions within the shader. Due to the
SIMD nature of vertex shaders, four sine and cosine waves are calculated in parallel, and the
results are weighted and combined using a single dp4.
Each sine wave has fully adjustable direction, frequency, speed, and offset that is config-
ured in the constant store.
The first step is to compute each waves starting phase into the sine or cosine function. The
texture coordinates are multiplied by the direction and frequency of the four waves in parallel.
c14 and c15 are the frequencies of the wave relative to S and T, respectively.
mul r0, c14, v7.x //use tex coords as inputs to sinusoidal warp
mad r0, c15, v7.y, r0 //use tex coords as inputs to sinusoidal warp
Next, the time, which is stored in c16.x, is multiplied by the speed of the waves (in c13 and
added to the wave offsets in c12):
mov r1, c16.x //time...
mad r0, r1, c13, r0 //add scaled time to move bumps according to
frequency
add r0, r0, c12 //starting time offset
This computes the input to the cosine function. A Taylor approximation, however, is only
accurate for the range it is created for, and more terms are needed the larger that range is. So
for a repeating function like a cosine wave, the fractional portion of the wave phase can be
extracted and then expanded to the to range before calculating the Taylor series
expansion.
frc r0.xy, r0 //take frac of all 4 components
frc r1.xy, r0.zwzw //
mov r0.zw, r1.xyxy //
mul r0, r0, c10.x //multiply by fixup factor (due to inaccuracy)
sub r0, r0, c0.y //subtract .5
mul r0, r0, c1.w //mult tex coords by 2pi coords range from
(-pi to pi)
Calculate the Taylor series expansion of sine (r4) and cosine (r5):
mul r5, r0, r0 //(wave vec)^2
mul r1, r5, r0 //(wave vec)^3
mul r6, r1, r0 //(wave vec)^4
mul r2, r6, r0 //(wave vec)^5
mul r7, r2, r0 //(wave vec)^6
mul r3, r7, r0 //(wave vec)^7
mul r8, r3, r0 //(wave vec)^8
mad r4, r1, c2.y, r0 //(wave vec) - ((wave vec)^3)/3!
mad r4, r2, c2.z, r4 //+((wave vec)^5)/5!
mad r4, r3, c2.w, r4 //-((wave vec)^7)/7!
mov r0, c0.z //1
mad r5, r5, c3.x ,r0 //-(wave vec)^2/2!
mad r5, r6, c3.y, r5 //+(wave vec)^4/4!
mad r5, r7, c3.z, r5 //-(wave vec)^6/6!
mad r5, r8, c3.w, r5 //+(wave vec)^8/8!
348 Part 3: Pixel Shader Tricks
Team LRN
The results are modulated by relative heights of each of the waves and the scaled sine wave
is used to perturb the position along the normal. The new object space position is then trans-
formed to compute the final position. The vertex input, v5.x, is used to allow artist control of
how high the waves are in different parts of the ocean. This can be useful for shorelines where
the ocean waves will be smaller than those farther out to sea:
sub r0, c0.z, v5.x //... 1-wave scale
mul r4, r4, r0 //scale sin
mul r5, r5, r0 //scale cos
dp4 r0, r4, c11 //multiply wave heights by waves
mul r0.xyz, v3, r0 //multiply wave magnitude at this vertex
by normal
add r0.xyz, r0, v0 //add to position
mov r0.w, c0.z //homogenous component
m4x4 oPos, r0, c4 //OutPos = ObjSpacePos * World-View-Proj
Matrix
The tangent and normal vectors are perturbed in a similar manner using the cosine wave
instead of the sine wave. This is done because the cosine is the first derivative of the sine and
therefore perturbs the tangent and normal vectors by the slope of the wave. The following code
makes the assumption that the source art is a plane along the Z axis.
It is worth mentioning that this vertex perturbation technique can be extended to sinusoi-
dally warp almost any geometry. See Bubble Shader for more details.
mul r1, r5, c11 //cos* waveheight
dp4 r9.x, -r1, c14 //normal x offset
dp4 r9.yzw, -r1, c15 //normal y offset and tangent offset
mov r5, v3 //starting normal
mad r5.xy, r9, c10.y, r5 //warped normal move nx, ny according to
//cos*wavedir*waveeheight
mov r4, v8 //tangent
mad r4.z, -r9.x, c10.y, r4.z //warped tangent vector
dp3 r10.x, r5, r5
rsq r10.y, r10.x
mul r5, r5, r10.y //normalize normal
dp3 r10.x, r4, r4
rsq r10.y, r10.x
mul r4, r4, r10.y //normalize tangent
The binormal is then calculated using a cross product of the warped normal and the warped
tangent vector to create a tangent space basis matrix. This matrix will be used later to trans-
form the bump maps tangent space normal into world space for cube-mapped environment-
mapped bump mapping (CMEMBM).
mul r3, r4.yzxw, r5.zxyw
mad r3, r4.zxyw, -r5.yzxw, r3 //xprod to find binormal
CMEMBM needs the view vector to perform the reflection operation:
sub r2, c8, r0 //view vector
dp3 r10.x, r2, r2
rsq r10.y, r10.x
mul r2, r2, r10.y //normalized view vector
Rendering Ocean Water 349
Team LRN
The height map shown in Figure 3 is used to create a normal map. The incoming texture coor-
dinates are used as a starting point to create two sets of coordinates that are rotated and scroll
across each other based on time. These coordinates are used to scroll two bump maps past each
other to produce the smaller ripples in the ocean. One interesting trick used in this shader is to
swap the u and v coordinates for the second texture before compositing them. This eliminates
the visual artifacts that occur when the scrolling textures align with each other exactly and the
ripples appear to stop for a moment. Swapping the texture coordinates ensure that the maps
never align with each other (unless they are radially symmetric).
mov r0, c16.x
mul r0, r0, c17.xyxy
frc r0.xy, r0 //frc of incoming time
add r0, v7, r0 //add time to tex coords
mov oT0, r0 //distorted tex coord 0
mov r0, c16.x
mul r0, r0, c17.zwzw
frc r0.xy, r0 //frc of incoming time
add r0, v7, r0 //add time to tex coords
mov oT1, r0.yxzw //distorted tex coord 1
The vertex shader is completed by the output of the remaining vectors used by the pixel shader.
The pixel and vertex shader for the ocean water effect can be found in its entirety at the end of
this article.
mov oT2, r2 //pass in view vector (worldspace)
mov oT3, r3 //tangent
mov oT4, r4 //binormal
mov oT5, r5 //normal
CMEMBM Pixel Shader with Fresnel Term
Once the vertex shader has completed, the pixel shader is responsible for producing the
bump-mapped reflective ocean surface.
First, the pixel shader averages the two scrolling RGB normal bump maps to generate a
composite normal. In this particular case, the bumps are softened further by dividing the x and
y components in half. Next, it transforms the tangent space composite normal into world space
and calculates a per-pixel reflection vector. The reflection vector is used to sample a skybox
cubic environment map (Figure 4). The shader also calculates 2*NV and uses it to sample a
Fresnel 1D texture (Figure 5). This Fresnel map gives the water a more greenish appearance
350 Part 3: Pixel Shader Tricks
Figure 3: Height map used to create the
normal map for the ocean shader
Team LRN
when looking straight down into it and a more bluish appearance when looking edge on. The
scale by two is used to expand the range of the Fresnel map.
texld r0, t0 //bump map 0
texld r1, t1 //sample bump map 1
texcrd r2.rgb, t2 //View vector
texcrd r3.rgb, t3 //Tangent
texcrd r4.rgb, t4 //Binormal
texcrd r5.rgb, t5 //Normal
add_d4 r0.xy, r0_bx2, r1_bx2 //Scaled Average of 2 bumpmaps' xy
offsets
mul r1.rgb, r0.x, r3
mad r1.rgb, r0.y, r4, r1
mad r1.rgb, r0.z, r5, r1 //Transform bumpmap normal into world
space
dp3 r0.rgb, r1, r2 //V.N
mad r2.rgb, r1, r0_x2, -r2 //R = 2N(V.N)-V
mov_sat r1, r0_x2 //2 * V.N (sample over range of 1d
map!)
The second phase composites the water color from the Fresnel map, the environment map, and
other specular highlights extracted from the environment map. One trick we use is to square
the environment map color values to make the colors brighter and to enhance the contrast for
compositing. The advantage to doing this in the pixel shader instead of as a preprocessing step
is so the same skybox environment map can be used for other objects in the scene. To get the
specular light sparkles in the water, a specular component is derived from the green channel of
the environment map. For this example, the choice is based on the environment map artwork.
The desired effect is to have the highlights in the water correspond to bright spots in the sky,
and in this case, the green channel seemed to work best. To make sure the specular peaks were
only generated from the brightest areas of the environment map, the specular value extracted
Rendering Ocean Water 351
Figure 4: Cubic environment map used
for ocean water reflections
Figure 5: 1D texture used for the water color addressed by 1NV
Team LRN
from the green channel was raised to the eighth power. This has the effect of darkening all but
the brightest areas of the image.
Another approach for encoding specular highlights in an environment map is to have the
artists specify a glow map as an alpha channel of the environment map. See Bubble Shader
for more details.
texcrd r0.rgb, r0
texld r2, r2 //cubic env map
texld r3, r1 //Index fresnel map using 2*V.N
mul r2.rgb, r2, r2 //Square the environment map
+mul r2.a, r2.g, r2.g //use green channel of env map as
specular
mul r2.rgb, r2, 1-r0.r //Fresnel Term
+mul r2.a, r2.a, r2.a //Specular highlight ^4
add_d4_sat r2.rgb, r2, r3_x2 //+= Water color
+mul r2.a, r2.a, r2.a //Specular highlight ^8
mad_sat r0, r2.a, c1, r2 //+= Specular highlight * highlight
color
Ocean Water Shader Source Code
DefineParam texture rgbNormalBumpMap NULL
SetParamEnum rgbNormalBumpMap EP_TEX0
DefineParam texture waterGradientMap NULL
SetParamEnum waterGradientMap EP_TEX1
DefineParam texture cubeEnvMap NULL
SetParamEnum cubeEnvMap EP_TEX2
//Constant store
DefineParam vector4 commonConst (0.0, 0.5, 1.0, 2.0)
DefineParam vector4 appConst (0.0, 0.0, 0.0, 0.0) //Time, 1.0/lightFalloffDist
SetParamEnum appConst EP_VECTOR3
DefineParam vector4 worldSpaceCamPos (0, 0, 0, 0)
BindParamToState worldSpaceCamPos STATE_VECTOR_CAMERA_POSITION 0 WORLD_SPACE
DefineParam vector4 worldSpaceLightPos (-10000, -25000, 2000, 1)
SetParamEnum worldSpaceLightPos EP_VECTOR0
DefineParam matrix4x4 wvp [(1,0,0,0) (0,1,0,0) (0,0,1,0) (0,0,0,1)]
BindParamToState wvp STATE_MATRIX_PVW
//=============================================================================
//commonly used constants
//heights for waves 4 different fronts
DefineParam vector4 waveHeights (80.0, 100.0, 5.0, 5.0)
//offset in sine wave.. (ranges 0 to 1)
DefineParam vector4 waveOffset (0.0, 0.2, 0.0, 0.0)
//freqency of the waves (e.g., waves per unit time..)
DefineParam vector4 waveSpeed (0.2, 0.15, 0.4, 0.4)
//diection of waves in tangent space (also controls frequency in space)
DefineParam vector4 waveDirx (0.25, 0.0, -0.7, -0.8)
DefineParam vector4 waveDiry (0.0, 0.15, -0.7, 0.1)
//scale factor for distortion of base map coords
352 Part 3: Pixel Shader Tricks
Team LRN
//bump map scroll speed
DefineParam vector4 bumpSpeed (0.031, 0.04, -0.03, 0.02)
DefineParam vector4 piVector (4.0, 1.57079632, 3.14159265, 6.28318530)
//Vectors for taylor's series expansion of sin and cos
DefineParam vector4 sin7 (1, -0.16161616, 0.0083333, -0.00019841)
DefineParam vector4 cos8 (-0.5, 0.041666666, -0.0013888889, 0.000024801587)
//frcFixup.x is a fixup to make the edges of the clamped sin wave match up again
due to // numerical inaccuracy
//frcFixup.y should be equal to the average of du/dx and dv/dy for the base texture
// coords.. this scales the warping of the normal
DefineParam vector4 frcFixup (1.02, 0.003, 0, 0)
DefineParam vector4 psCommonConst (0, 0.5, 1, 0.25)
DefineParam vector4 highlightColor (0.8, 0.76, 0.62, 1)
DefineParam vector4 waterColor (0.50, 0.6, 0.7, 1)
//=============================================================================
// 1 Pass
//=============================================================================
StartShader
Requirement VERTEXSHADERVERSION 1.1
Requirement PIXELSHADERVERSION 1.4
StartPass
SetTexture 0 rgbNormalBumpMap
SetTextureFilter 0 BILINEAR
SetTextureStageState 0 MIPMAPLODBIAS -1.0
SetTexture 1 rgbNormalBumpMap
SetTextureFilter 1 BILINEAR
SetTextureStageState 1 MIPMAPLODBIAS -1.0
SetTexture 2 cubeEnvMap
SetTextureWrap 2 CLAMP CLAMP CLAMP
SetTextureFilter 2 BILINEAR
SetTextureStageState 2 MIPMAPLODBIAS 0.0
SetTexture 3 waterGradientMap
SetTextureWrap 3 CLAMP CLAMP CLAMP
SetTextureFilter 3 LINEAR
SetVertexShaderConstant 0 commonConst
SetVertexShaderConstant 1 piVector
SetVertexShaderConstant 2 sin7
SetVertexShaderConstant 3 cos8
SetVertexShaderConstant 4 wvp
SetVertexShaderConstant 8 worldSpaceCamPos
SetVertexShaderConstant 9 worldSpaceLightPos
SetVertexShaderConstant 10 frcFixup
SetVertexShaderConstant 11 waveHeights
SetVertexShaderConstant 12 waveOffset
SetVertexShaderConstant 13 waveSpeed
SetVertexShaderConstant 14 waveDirx
SetVertexShaderConstant 15 waveDiry
SetVertexShaderConstant 16 appConst
SetVertexShaderConstant 17 bumpSpeed
StartVertexShader
// v0 - Vertex Position
// v3 - Vertex Normal
Rendering Ocean Water 353
Team LRN
// v7 - Vertex Texture Data u,v
// v8 - Vertex Tangent (v direction)
// c0 - { 0.0, 0.5, 1.0, 2.0}
// c1 - { 4.0, .5pi, pi, 2pi}
// c2 - {1, -1/3!, 1/5!, -1/7! } //for sin
// c3 - {1/2!, -1/4!, 1/6!, -1/8! } //for cos
// c4-7 - Composite World-View-Projection Matrix
// c8 - ModelSpace Camera Position
// c9 - ModelSpace Light Position
// c10 - {fixup factor for taylor series imprecision, }
// c11 - {waveHeight0, waveHeight1, waveHeight2, waveHeight3}
// c12 - {waveOffset0, waveOffset1, waveOffset2, waveOffset3}
// c13 - {waveSpeed0, waveSpeed1, waveSpeed2, waveSpeed3}
// c14 - {waveDirX0, waveDirX1, waveDirX2, waveDirX3}
// c15 - {waveDirY0, waveDirY1, waveDirY2, waveDirY3}
// c16 - { time, sin(time)}
// c17 - {basetexcoord distortion x0, y0, x1, y1}
vs.1.1
mul r0, c14, v7.x //use tex coords as inputs to sinusoidal warp
mad r0, c15, v7.y, r0 //use tex coords as inputs to sinusoidal warp
mov r1, c16.x //time...
mad r0, r1, c13, r0 //add scaled time to move bumps according to
frequency
add r0, r0, c12 //starting time offset
frc r0.xy, r0 //take frac of all 4 components
frc r1.xy, r0.zwzw //
mov r0.zw, r1.xyxy //
mul r0, r0, c10.x //multiply by fixup factor (due to inaccuracy)
sub r0, r0, c0.y //subtract .5
mul r0, r0, c1.w //mult tex coords by 2pi coords range from
(-pi to pi)
mul r5, r0, r0 //(wave vec)^2
mul r1, r5, r0 //(wave vec)^3
mul r6, r1, r0 //(wave vec)^4
mul r2, r6, r0 //(wave vec)^5
mul r7, r2, r0 //(wave vec)^6
mul r3, r7, r0 //(wave vec)^7
mul r8, r3, r0 //(wave vec)^8
mad r4, r1, c2.y, r0 //(wave vec) - ((wave vec)^3)/3!
mad r4, r2, c2.z, r4 //+ ((wave vec)^5)/5!
mad r4, r3, c2.w, r4 //- ((wave vec)^7)/7!
mov r0, c0.z //1
mad r5, r5, c3.x ,r0 //-(wave vec)^2/2!
mad r5, r6, c3.y, r5 //+(wave vec)^4/4!
mad r5, r7, c3.z, r5 //-(wave vec)^6/6!
mad r5, r8, c3.w, r5 //+(wave vec)^8/8!
sub r0, c0.z, v5.x //... 1-wave scale
mul r4, r4, r0 //scale sin
mul r5, r5, r0 //scale cos
dp4 r0, r4, c11 //multiply wave heights by waves
mul r0.xyz, v3, r0 //multiply wave magnitude at this vertex by normal
add r0.xyz, r0, v0 //add to position
mov r0.w, c0.z //homogenous component
354 Part 3: Pixel Shader Tricks
Team LRN
m4x4 oPos, r0, c4 //OutPos = ObjSpacePos * World-View-Projection
Matrix
mul r1, r5, c11 //cos* waveheight
dp4 r9.x, -r1, c14 //normal x offset
dp4 r9.yzw, -r1, c15 //normal y offset and tangent offset
mov r5, v3 //starting normal
mad r5.xy, r9, c10.y, r5 //warped normal move nx, ny according to
//cos*wavedir*waveeheight
mov r4, v8 //tangent
mad r4.z, -r9.x, c10.y, r4.z //warped tangent vector
dp3 r10.x, r5, r5
rsq r10.y, r10.x
mul r5, r5, r10.y //normalize normal
dp3 r10.x, r4, r4
rsq r10.y, r10.x
mul r4, r4, r10.y //normalize tangent
mul r3, r4.yzxw, r5.zxyw
mad r3, r4.zxyw, -r5.yzxw, r3 //xprod to find binormal
sub r2, c8, r0 //view vector
dp3 r10.x, r2, r2
rsq r10.y, r10.x
mul r2, r2, r10.y //normalized view vector
mov r0, c16.x
mul r0, r0, c17.xyxy
frc r0.xy, r0 //frc of incoming time
add r0, v7, r0 //add time to tex coords
mov oT0, r0 //distorted tex coord 0
mov r0, c16.x
mul r0, r0, c17.zwzw
frc r0.xy, r0 //frc of incoming time
add r0, v7, r0 //add time to tex coords
mov oT1, r0.yxzw //distorted tex coord 1
mov oT2, r2 //pass in view vector (worldspace)
mov oT3, r3 //tangent
mov oT4, r4 //binormal
mov oT5, r5 //normal
EndVertexShader
SetPixelShaderConstant 0 psCommonConst
SetPixelShaderConstant 1 highlightColor
StartPixelShader
ps.1.4
texld r0, t0 //bump map 0
texld r1, t1 //sample bump map 1
texcrd r2.rgb, t2 //View vector
texcrd r3.rgb, t3 //Tangent
texcrd r4.rgb, t4 //Binormal
texcrd r5.rgb, t5 //Normal
add_d4 r0.xy, r0_bx2, r1_bx2 //Scaled Average of 2 bumpmaps xy
offsets
mul r1.rgb, r0.x, r3
mad r1.rgb, r0.y, r4, r1
mad r1.rgb, r0.z, r5, r1 //Put bumpmap normal into world space
dp3 r0.rgb, r1, r2 //V.N
Rendering Ocean Water 355
Team LRN
mad r2.rgb, r1, r0_x2, -r2 //R = 2N(V.N)-V
mov_sat r1, r0_x2 //2 * V.N (sample over range of 1d
map!)
phase
texcrd r0.rgb, r0
texld r2, r2 //cubic env map
texld r3, r1 //Index fresnel map using 2*V.N
mul r2.rgb, r2, r2 //Square the environment map
+mul r2.a, r2.g, r2.g //use green channel of env map as
specular
mul r2.rgb, r2, 1-r0.r //Fresnel Term
+mul r2.a, r2.a, r2.a //Specular highlight ^4
add_d4_sat r2.rgb, r2, r3_x2 //+= Water color
+mul r2.a, r2.a, r2.a //Specular highlight ^8
mad_sat r0, r2.a, c1, r2 //+= Specular highlight * highlight
color
EndPixelShader
EndPass
EndShader
Sample Applications
This shader can be seen in the Island demos (http://www.ati.com/na/pages/resource_centre/
dev_rel/Demos.html) and the Ocean Screen Saver (http://www.ati.com/na/pages/resource_
centre/dev_rel/screensavers.html), as well as on the companion CD.
356 Part 3: Pixel Shader Tricks
Team LRN
Rippling Reflective and
Refractive Water
Alex Vlachos, John Isidoro, and Chris Oat
Introduction
One of the classic challenges of real-time computer graphics is to generate realistic-looking
water. Because water can look very different depending on the context of the scene, it is impor-
tant to define a few different categories of water effects. For example, techniques used to
render ocean water (see Rendering Ocean Water) may not produce realistic-looking puddle
water. With this in mind, the shaders presented in this chapter will focus on highly realistic
per-pixel lighting effects used to render real-time lake water. Unlike ocean water, lake water
does not have large rolling swells. Instead, the surface geometry consists of high frequency,
non-uniform perturbations that give the surface a subtle choppiness, resulting in slightly dis-
torted reflections and refractions. Because of these surface characteristics, shallow water
reflections and refractions, as shown in Figure 1, can be reasonably approximated with planar
reflections and renderable textures.
Renderable textures allow the programmer to distort the reflection and refraction textures
using a sum of scrolling bump maps to simulate the water surface ripples. One of the nice
things about distorting these textures with a bump map is that the renderable textures can be of
a much lower resolution than the screen resolution, and the end result will still look very com-
pelling. The following example was rendered using two 512x256 renderable textures. Since the
reflected and refracted geometry is drawn once for each renderable texture, rendering to these
maps at a lower resolution provides a significant performance boost.
357
Figure 1: This scene was drawn using the
reflection and refraction techniques
described in this chapter.
Team LRN
Generating Reflection and Refraction Maps
When dealing with fairly flat water surfaces such as lake water, river water, or puddle water,
reflections can be convincingly approximated with planar reflections. A reflection map is cre-
ated by reflecting the scene geometry over the plane of the waters surface and rendering to a
texture. It is important to utilize user clip planes in order to clip any geometry that is already
below the plane of reflection, as this geometry is underwater and should not be included in the
reflection map (see Figure 2).
As illustrated in Figure 3, the refraction map is generated by clipping all scene geometry above
the plane of the waters surface. As with the reflection map, user clip planes are used to cull
the geometry above the waters surface.
Vertex Shader
A vertex shader is used to approximate sine and cosine waves to generate rippling water geom-
etry as well as perturbing the tangent space vectors per vertex. This vertex shader is very
similar to the vertex shader presented in Rendering Ocean Water; please refer to that article
for a more detailed description of the vertex shader. The slightly altered ocean water vertex
shader is presented below; the only difference is the value for wave height stored in c11.
StartVertexShader
vs.1.1
// v0 - Vertex Position
// v7 - Vertex Texture Data u,v
// v8 - Vertex Tangent (v direction)
358 Part 3: Pixel Shader Tricks
Figure 2: Example of a reflection map
created by reflecting scene geometry
about the plane of the waters surface
Figure 3: Example of a refraction map
created by drawing only the geometry
below the waters surface
Team LRN
//
// c0 - { 0.0, 0.5, 1.0, 2.0}
// c1 - { 4.0, .5pi, pi, 2pi}
// c2 - {1, -1/3!, 1/5!, -1/7! } //for sin
// c3 - {1/2!, -1/4!, 1/6!, -1/8! } //for cos
// c4-7 - Composite World-View-Projection Matrix
// c8 - ModelSpace Camera Position
// c9 - ModelSpace Light Position
// c10 - {fixup factor for taylor series imprecision, }
// c11 - {waveHeight0, waveHeight1, waveHeight2, waveHeight3}
// c12 - {waveOffset0, waveOffset1, waveOffset2, waveOffset3}
// c13 - {waveSpeed0, waveSpeed1, waveSpeed2, waveSpeed3}
// c14 - {waveDirX0, waveDirX1, waveDirX2, waveDirX3}
// c15 - {waveDirY0, waveDirY1, waveDirY2, waveDirY3}
// c16 - {time, sin(time)}
// c17 - {basetexcoord distortion x0, y0, x1, y1}
mul r0, c14, v7.x // use tex coords as inputs to sinusoidal warp
mad r0, c15, v7.y, r0 // use tex coords as inputs to sinusoidal warp
mov r1, c16.x // time...
mad r0, r1, c13, r0 // add scaled time to move bumps by frequency
add r0, r0, c12
frc r0.xy, r0 // take frac of all 4 components
frc r1.xy, r0.zwzw
mov r0.zw, r1.xyxy
mul r0, r0, c10.x // multiply by fixup factor (fix taylor series
inaccuracy)
sub r0, r0, c0.y // subtract .5
mul r0, r0, c1.w // mult tex coords by 2pi coords range from
(-pi to pi)
mul r5, r0, r0 // (wave vec)^2
mul r1, r5, r0 // (wave vec)^3
mul r6, r1, r0 // (wave vec)^4
mul r2, r6, r0 // (wave vec)^5
mul r7, r2, r0 // (wave vec)^6
mul r3, r7, r0 // (wave vec)^7
mul r8, r3, r0 // (wave vec)^8
mad r4, r1, c2.y, r0 // (wave vec) - ((wave vec)^3)/3!
mad r4, r2, c2.z, r4 // + ((wave vec)^5)/5!
mad r4, r3, c2.w, r4 // - ((wave vec)^7)/7!
mov r0, c0.z // 1
mad r5, r5, c3.x ,r0 // -(wave vec)^2/2!
mad r5, r6, c3.y, r5 // +(wave vec)^4/4!
mad r5, r7, c3.z, r5 // -(wave vec)^6/6!
mad r5, r8, c3.w, r5 // +(wave vec)^8/8!
dp4 r0, r4, c11 // multiply wave heights by waves
mul r0.xyz, c0.xxzx, r0 // multiply wave magnitude at this vertex by normal
add r0.xyz, r0, v0 // add to position
mov r0.w, c0.z // homogenous component
m4x4 oPos, r0, c4 // OutPos = ObjSpacePos * World-View-Projection
Matrix
mul r1, r5, c11 // cos* waveheight
dp4 r9.x, -r1, c14 // normal x offset
dp4 r9.yzw, -r1, c15 // normal y offset and tangent offset
mov r5, c0.xxzx
Rippling Reflective and Refractive Water 359
Team LRN
mad r5.xy, r9, c10.y, r5 // warped normal move according to
cos*wavedir*waveeheight
mov r4, v9
mad r4.z, -r9.x, c10.y, r4.z // warped tangent vector
dp3 r10.x, r5, r5
rsq r10.y, r10.x
mul r5, r5, r10.y // normalize normal
dp3 r10.x, r4, r4
rsq r10.y, r10.x
mul r4, r4, r10.y // normalize tangent
mul r3, r4.yzxw, r5.zxyw
mad r3, r4.zxyw, -r5.yzxw, r3 // cross product to find binormal
sub r1, c9, r0 // light vector
sub r2, c8, r0 // view vector
dp3 r10.x, r1, r1 // normalize light vector
rsq r10.y, r10.x
mul r1, r1, r10.y
dp3 r6.x, r1, r3
dp3 r6.y, r1, r4
dp3 r6.z, r1, r5 // transform light vector into tangent space
dp3 r10.x, r2, r2
rsq r10.y, r10.x
mul r2, r2, r10.y // normalized view vector
dp3 r7.x, r2, r3
dp3 r7.y, r2, r4
dp3 r7.z, r2, r5 // put view vector in tangent space
mov r0, c16.x
mul r0, r0, c24
frc r0.xy, r0
mul r1, v7, c26
add oT0, r1, r0 // bump map coord1
mov r0, c16.x
mul r0, r0, c25
frc r0.xy, r0
mul r1, v7, c27
add oT1, r1, r0 // bump map coord 2
dp4 r0.x, v0, c20
dp4 r0.y, v0, c21
dp4 r0.zw, v0, c22
mov oT2, r0 // projective tex coords for reflection/
refreaction maps
mov oT3, r7 // tan space view vec
mov oT4, v7 // base map coords
mov oT5, r6 // tan space light vec
EndVertexShader
Pixel Shader
The pixel shader used to render the water has a few novel features. When sampling from the
renderable textures, the texture coordinates must be interpolated linearly in screen space. The
reason for this is that the scene is already rendered from the point of view of the camera, and
thus the contents of the renderable texture are already perspectively correct. To achieve this,
the projection matrix must be altered in the following way:
360 Part 3: Pixel Shader Tricks
Team LRN
Matrix M = { 0.5f, 0.0f, 0.0f, 0.0f,
0.0f, -0.5f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f,
0.5f, 0.5f, 1.0f, 0.0f };
// row major element (1,4) is set to zero
projectionMatrix[11] = 0.0f;
// projection matrix uses scale and bias to get coordinates into
// [0.0, 1.0] range
projectionMatrix = M * projectionMatrix;
Now the projection matrix will move vertices into a normalized texture space for indexing into
the pre-projected reflection and refraction maps. Given the linearly interpolated texture coordi-
nates, the perturbations for the reflection and refraction maps can be performed on a per-pixel
basis by simply adding a scaled and rotated version of the xy offsets from the scrolling bump
maps. It is important to scale and rotate these offsets so that as the bump map scrolls, the per-
turbations move in the direction that the water is flowing. Another interesting trick is to sample
the refraction map using the swapped texture coordinates used to index the reflection map (for
example, index the reflection map using coordinates <u,v>, and then sample the refraction map
using coordinates <v,u>). Swapping the coordinates in this way prevents the two scrolling
bump maps from aligning. After sampling, the refractions are modulated with the water color,
and the perturbed reflections are attenuated with a per-pixel Fresnel term and modulated with a
reflection color. The per-pixel Fresnel term scales the reflections such that the strength of the
reflection is dependent on the angle of the view vector with respect to the water surface nor-
mal. The full shader is shown below along with further comments.
StartPixelShader
ps.1.4
// T0 : Bump map 1 Coordinates
// T1 : Bump map 2 Coordinates
// T2 : Projective Texture Coordinates for Reflection/Refraction Maps
// T3 : Tangent Space View Vector
// T5 : Tangent Space Light Vector
texld r0, t0
texld r1, t1
texcrd r2.xy, t2_dw.xyw // renderable textures
texcrd r4.xyz, t3 // tan space V
texcrd r5.xyz, t5 // tan space L
add_d2 r1.rgb, r0_bx2, r1_bx2
mov r3.xy, r2
dp3 r0.r, r4, r1 // V.N
dp3 r0.g, r1, r5 // L.N
mad_sat r5.rg, r1, c2, r3 // perturb refraction
mad_sat r4.rg, r1, c1, r3 // perturb reflection
phase
texcrd r0.rgb, r0 // V.N, L.N, R.L
texld r2, r4 // Reflection
texld r3, r5 // Refraction
mul_sat r2.rgb, r2, c3 // reflection color
+mul r2.a, r0.g, r0.g
Rippling Reflective and Refractive Water 361
Team LRN
mul_sat r3.rgb, r3, c4 // refraction color
+mul_sat r2.a, r2.a, r2.a
mad_sat r0.rgb, 1-r0.r, r2, r3
+mul_sat r2.a, r2.a, r2.a
mad_sat r0.rgb, r2.a, c6, r0 // add specular highlights to reflection
+mov r0.a, r0.r
EndPixelShader
Note the perturbed refractions where the body of the fish is still underwater and the underwater
vegetation. Perturbed reflections can be seen where the waterfall is reflected in the waters sur-
face in the far right. Also note how the specular highlighting gives the water surface better
definition.
Conclusion
This section focused on using vertex and pixel shaders to render realistic rippling water with
realistic reflections and refractions by building on techniques described in other articles. The
shaders presented here demonstrate the use of Taylor series approximations to compute sine
and cosine waves in a vertex shader as well as the use of renderable textures to create rippling
reflections and refractions.
The shaders described in this section, as well as the accompanying screen grabs, were
taken from ATIs Nature demo. This real-time interactive graphics demo shows rippling reflec-
tive and refractive water in a complex natural environment and is available on the Internet as
an executable program or compressed video in the Demos section of the ATI Developers web
site (www.ati.com/developer/).
362 Part 3: Pixel Shader Tricks
Figure 4: Rippling reflective and refrac-
tive water
Team LRN
Crystal/Candy Shader
John Isidoro and David Gosselin
Introduction
The goal of this article is to create a shader that gives the effect of a very shiny semi-transpar-
ent surface, such as crystal or a piece of hard candy. This is accomplished using a combination
of vertex and pixel shaders. The shader combines a cubic environment map to provide the
reflections, a bump map to provide some detail, a Fresnel term to control opacity, and a base
map to provide the surface color.
Setup
This shader uses a model with one set of texture coordinates for both a base map (color) and
bump map (normals encoded as colors). Additionally, the shader uses a cubic environment
map. In this example, the environment map was prerendered using the ATI Sushi graphics
engine and is shown in Figure 1.
Since the object needs to be semi-transparent, it will be drawn in two passes, both blended into
the frame buffer. The first pass will draw the back faces of the object. This can be accom-
plished by setting the cull mode to either clockwise or counterclockwise depending on how
363
Figure 1: Cubic environment map
Team LRN
your application specifies back/front faces. This is accomplished by setting the Direct3D ren-
der state D3DRS_CULLMODE appropriately. The code for doing this is:
d3dDevice->SetRenderState (D3DRS_CULLMODE, D3DCULL_CCW);
The front faces are drawn by setting the opposite cull mode and drawing the model again:
d3dDevice->SetRenderState (D3DRS_CULLMODE, D3DCULL_CW);
For both passes, alpha blending is enabled in order to give the transparent effect. The code for
setting this up is:
d3dDevice->SetRenderState (D3DRS_ALPHABLENDENABLE, TRUE);
d3dDevice->SetRenderState (D3DRS_SRCBLEND, D3DBLEND_SRCALPHA);
d3dDevice->SetRenderState (D3DRS_DSTBLEND, D3DBLEND_INVSRCALPHA);
Now that we have set these general render states, we will explore the specifics of the crys-
tal/candy vertex and pixel shaders.
Vertex Shader
The vertex shader for this effect is centered on computing the view direction. Additionally, it
will pass the tangent/texture space vectors along so that the bump map normal can be trans-
formed into object space in the pixel shader. The first section of the vertex shader computes the
transformed vertex position. The vertex shader code to perform this transformation follows:
// c0 - (0.125, 0.5, 1.0, 2.0)
// c2 - World Space Camera Position
// c4-c7- World/View/Projection Matrix
// v0 - Vertex Position
// v3 - Vertex Normal
// v7 - Vertex Texture Data u,v
// v8 - Vertex tangent Vector
vs.1.1
m4x4 oPos, v0, c4 // OutPos = ObjSpacePos * WVP Matrix
mov oT0, v7 // output basemap texcoords
The next section of the vertex shader computes the view vector in world space. This is accom-
plished by subtracting the world space camera position from the world space vertex position
and then normalizing the result. Note that in this example, the world transform is identity and
the input vertex position is used directly. If this is not the case in your application, you will
need to transform the input position by the world matrix before performing this calculation.
sub r8, c2, v0 // viewer direction
dp3 r10.x, r8, r8 // magnitude
rsq r10.y, r10.x // normalize
mul r8, r8, r10.y // V normalized view vector
mov oT1, r8 // world space view vector
The next section passes along the tangent/texture space basis matrix. In this example, we com-
pute the binormal by taking the cross product of the tangent and the normal. Note that this is
not a general solution; it relies upon the fact that the texture coordinates for this model are not
reversed on any triangles.
mov r5, v3 // normal (tZ)
mov oT4, r5
mov r3, v8 // tangent (tY)
364 Part 3: Pixel Shader Tricks
Team LRN
mov oT2, r3
mul r4, r5.yzxw, r3.zxyw // compute binormal (tX) as being
mad r4, r5.zxyw, -r3.yzxw, r4 // perpendicular to normal
mov oT3, r4
The last section of this vertex shader transforms the view vector into tangent space. It also
scales and biases the vector so that it can be passed to the pixel shader in one of the color
interpolators.
m3x3 r6, r8, r3 // xform view vec into tangent space
mad oD0, r6, c0.y, c0.y // tan space V vector
Now that the per-vertex computations are described, we can look at what happens at the pixel
level to create the crystal look.
Pixel Shader
The pixel shader for this effect is accomplished in two shader phases. The first phase computes
the reflection vector used for the cubic environment map, while the second phase composites
the color from the dependent read with other maps. The full shader is shown below and a more
detailed description of the various parts follows.
// c0 - (0.0, 0.5, 1.0, 0.75)
// c1 base color (0.5, 0.0, 0.5, 1.0)
// c2 reflection color (1.0, 1.0, 1.0, 1.0)
// c7 - (0.0, 0.0, 1.0, 0.4)
ps.1.4
texld r0, t0 // bump map
texcrd r1.rgb, t1 // view vec
texcrd r2.rgb, t2 // tan x
texcrd r3.rgb, t3 // tan y
texcrd r4.rgb, t4 // tan z
lrp r0.rgb, c7.a, r0, c7 // weaken bump map by lerping with 0,0,1
mul r5.rgb, r0.x, r2 // move bump normal into obj space
mad r5.rgb, r0.y, r3, r5 // move bump normal into obj space
mad r5.rgb, r0.z, r4, r5 // move bump normal into obj space
dp3 r3.rgb, r1, r5 // V.N
mad r2.rgb, r3, r5_x2, -r1 // 2N(V.N)-V
phase
texld r1, t0 // base map
texld r2, r2 // fetch from environment map
dp3 r0.a, v0_bx2, r0 // (N.V)fresnel term
mul r1.rgb, r1, c1 // base map * base color
+mov r0.a, 1-r0.a // 1-(N.V)
mul r2.rgb, r2, c2 // reflection map * reflection color
mad r0.rgb, r2, r0.a, r1 // (1-(N.V)) * reflection map * reflection color
// + base map * base color
+add_sat r0.a, r0.a, c0.y // boost opacity: (1-(N.V)) + 0.5
At the core of the shader is the environment mapping, which is sampled as a dependent read at
the top of the second phase. The following lines from the pixel shader compute the per-pixel
reflection vector and sample the cube map:
Crystal/Candy Shader 365
Team LRN
dp3 r3.rgb, r1, r5 // V.N
mad r2.rgb, r3, r5_x2, -r1 // 2N(V.N)-V
phase
texld r2, r2 // fetch from env map
If the shader was changed to just do this computation, the result would look like Figure 2. One
thing to note is that in this picture, the top portion of the model has intentionally faceted nor-
mals to give the appearance of some kind of cut crystal, while the bottom part of the model has
smooth normals.
To this effect, we now want to add bumpiness based on the normal map. In order for this to
look correct, we take the sampled normal (r0) from the bump map and transform it into object
space using the tangent space matrix that was computed in the vertex shader and interpolated
across the polygon. The following lines show the part of the pixel shader where this computa-
tion takes place.
mul r5.rgb, r0.x, r2 // move bump normal into obj space
mad r5.rgb, r0.y, r3, r5 // move bump normal into obj space
mad r5.rgb, r0.z, r4, r5 // move bump normal into obj space
If this is added to the environment mapping, you get something that looks like Figure 3.
Since this looks a little too bumpy for the effect we want, a small term is linearly interpo-
lated with the bump map prior to doing the bump calculation. The following line of code from
the shader performs this task:
lrp r0.rgb, c7.a, r0, c7 // weaken bump map by lerping with 0,0,1
Figure 4 shows what this looks like on the above model.
This basically wraps up the computation that occurs in the first phase of the pixel shader. In
order to make the inner parts of the model (where you would see through more of the actual
object) seem more opaque and the outer edges seem more transparent, a Fresnel term is used.
(For more uses of the Fresnel term, see Per-Pixel Fresnel Term.) The computation takes a
366 Part 3: Pixel Shader Tricks
Figure 2: Environment
mapping only
Figure 3: Bumpy envi-
ronment mapped
Figure 4: Weaker
bumpy environment
mapped
Team LRN
few additional instructions, with one hidden in a co-issue alpha instruction. The following lines
of code compute the Fresnel term used to give the impression of thicker crystal:
dp3 r0.a, v0_bx2, r0 // (N.V)fresnel term
+mov r0.a, 1-r0.a // 1-(N.V)
Figure 5 shows what the model looks like with this term added to the shader.
Since we are going for a more solid-looking effect, the Fresnel term is tweaked a bit to
make the model look more opaque. This is also done in a co-issued alpha instruction, shown
below:
+add_sat r0.a, r0.a, c0.y // boost opacity: (1-(N.V)) + 0.5
Figure 6 shows what the model looks like with this term applied.
Now we want to add some color to the model using the base map and base color. The fol-
lowing line in the pixel shader computes this contribution:
mul r1.rgb, r1, c1 // base map * base color
Figure 7 shows what this would look like on its own.
The final step is to combine the terms together to create the final image. The following code
performs this computation:
mul r2.rgb, r2, c2 // reflection map * reflection color
mad r0.rgb, r2, r0.a, r1 // (1-(N.V)) * reflection map * reflection color
// + base map * base color
Crystal/Candy Shader 367
Figure 5: Fresnel term
added
Figure 6: Tweaked
Fresnel term
Figure 7: Base map
and base color
Team LRN
This is what the final image looks like.
Figure 9 shows what this looks like within an environment from the ATI Sushi engine.
Summary
This article has shown a method for creating a semi-transparent crystal/candy look on an
object. It used cubic environment mapping for reflections and per-pixel bump mapping. Hope-
fully, this gives you some ideas and techniques that you can use within your applications.
368 Part 3: Pixel Shader Tricks
Figure 8: Putting it all together
Figure 9: The shader within a scene
Team LRN
Bubble Shader
John Isidoro and David Gosselin
Introduction
Some of the most visually interesting objects in the real world are soap bubbles. Their reflec-
tive and partially transparent nature, as well as the way they pulsate, makes them a key
candidate for vertex and pixel shaders. Strangely enough, on a vertex and pixel shader level,
they are implemented in a very similar manner to the ocean water shader described in Render-
ing Ocean Water. There are a variety of different effects that make up the bubble shader.
Setup
As input, this shader uses a model with a position, a normal, the tangent vector (in v direction
of texture coordinates), and one set of texture coordinates for a rainbow-like basemap (shown
in the following figure).
369
Figure 1: Base rainbow map
Team LRN
This shader, as in Crystal/Candy Shader, uses a cubic enviroment map, shown below.
Also, like the candy shader, we want this object to be semi-transparent, which requires drawing
the model in two passes: back faces first then front faces with alpha blending enabled.
Vertex Shader
The vertex shader uses four sine waves traveling in different directions over the surface of the
bubble to perturb the vertex. The vertex is perturbed in the direction of the normal. In order to
perturb the normal, cosine waves traveling in the same four directions are used. To provide a
consistent-looking warping across vertices, the tangent and binormal are used to guide the
direction of the sinusoidal warp. The first portion of the vertex shader looks almost identical to
the ocean shader and is shown below.
// v0 - Vertex Position
// v3 - Vertex Normal
// v7 - Vertex Texture Data u,v
// v8 - Vertex Tangent (v direction of tex coords)
// c0 - { 0.0, 0.5, 1.0, 2.0}
// c1 - { 4.0, .5pi, pi, 2pi}
// c2 - {1, -1/3!, 1/5!, -1/7!} for sin
// c3 - {1/2!, -1/4!, 1/6!, -1/8!} for cos
// c4-7 - Composite World-View-Projection Matrix
// c8 - Model Space Camera Position
// c9 - Model Space Light Position
// c10 - {1.02, 0.04, 0, 0} fixup factor for Taylor series imprecision
// c11 - {0.5, 0.5, 0.25, 0.25} waveHeight0, waveHeight1, waveHeight2, waveHeight3
// c12 - {0.0, 0.0, 0.0, 0.0} waveOffset0, waveOffset1, waveOffset2, waveOffset3
// c13 - {0.6, 0.7, 1.2, 1.4} waveSpeed0, waveSpeed1, waveSpeed2, waveSpeed3
// c14 - {0.0, 2.0, 0.0, 4.0} waveDirX0, waveDirX1, waveDirX2, waveDirX3
// c15 - {2.0, 0.0, 4.0, 0.0} waveDirY0, waveDirY1, waveDirY2, waveDirY3
// c16 - { time, sin(time)}
// c17 - {-0.00015, 1.0, 0.0, 0.0} base texcoord distortion x0, y0, x1, y1
vs.1.1
mul r0, c14, v7.x // use tex coords as inputs to sinusoidal warp
mad r0, c15, v7.y, r0 // use tex coords as inputs to sinusoidal warp
370 Part 3: Pixel Shader Tricks
Figure 2: Cubic environment map color Figure 3: Cubic environment map alpha
Team LRN
mov r1, c16.x // time...
mad r0, r1, c13, r0 // add scaled time to move bumps according to frequency
add r0, r0, c12
frc r0.xy, r0 // take frac of all 4 components
frc r1.xy, r0.zwzw
mov r0.zw, r1.xyxy
mul r0, r0, c10.x // multiply by fixup factor (due to inaccuracy of taylor
series)
sub r0, r0, c0.y // subtract .5
mul r0, r0, c1.w // mult tex coords by 2pi coords range from(-pi to pi)
mul r5, r0, r0 // (wave vec)^2
mul r1, r5, r0 // (wave vec)^3
mul r6, r1, r0 // (wave vec)^4
mul r2, r6, r0 // (wave vec)^5
mul r7, r2, r0 // (wave vec)^6
mul r3, r7, r0 // (wave vec)^7
mul r8, r3, r0 // (wave vec)^8
mad r4, r1, c2.y, r0 //(wave vec) - ((wave vec)^3)/3!
mad r4, r2, c2.z, r4 // + ((wave vec)^5)/5!
mad r4, r3, c2.w, r4 // - ((wave vec)^7)/7!
mov r0, c0.z //1
mad r5, r5, c3.x ,r0 //-(wave vec)^2/2!
mad r5, r6, c3.y, r5 //+(wave vec)^4/4!
mad r5, r7, c3.z, r5 //-(wave vec)^6/6!
mad r5, r8, c3.w, r5 //+(wave vec)^8/8!
dp4 r0, r4, c11 //multiply wave heights by waves
mul r0, r0, v3 //apply deformation in direction of normal
add r10.xyz, r0, v0 //add to position
mov r10.w, c0.z //homogenous component
m4x4 oPos, r10, c4 // OutPos = WorldSpacePos * Composite View-Projection Matrix
mov oT0, v7 // Pass along texture coordinates
This is where the shader starts to diverge a bit from the ocean shader. First the binormal is
computed:
mov r3, v3
mul r4, v8.yzxw, r3.zxyw
mad r4, v8.zxyw, -r3.yzxw, r4 //cross product to find binormal
Then the normal is warped based on the tangent space basis vectors (tangent and binormal):
mul r1, r5, c11 //cos * waveheight
dp4 r9.x, -r1, c14 //amount of normal warping in direction of binormal
dp4 r9.y, -r1, c15 //amount of normal warping in direction of tangent
mul r1, r4, r9.x //normal warping in direction of binormal
mad r1, v8, r9.y, r1 //normal warping in direction of tangent
mad r5, r1, c10.y, v3 //warped normal move nx, ny: cos * wavedir * waveheight
The normal is then renormalized:
dp3 r10.x, r5, r5
rsq r10.y, r10.x
mul r5, r5, r10.y //normalize warped normal
Bubble Shader 371
Team LRN
Next the view vector is computed:
sub r2, c8, r0 //view vector
dp3 r10.x, r2, r2
rsq r10.y, r10.x
mul r2, r2, r10.y //normalized view vector
Then the dot product of the view vector and the warped normal is computed:
dp3 r7, r5, r2 //N.V
mov oT2, r7 //Pass along N.V
This is used to compute the reflection vector:
add r6, r7, r7 //2N.V
mad r6, r6, r5, -r2 //2N(N.V)-V
mov oT1, r6 //reflection vector
Finally, the distance to the camera is computed:
rcp r10.y, r10.y //distance from camera
mad oD0, r10.y, c17.x, c17.y //scale and bias distance
The vertex shader then passes along the texture coordinates of the base map, the reflection vec-
tor, the value of NV, and the distance to the camera along to the pixel shader for further
processing.
Pixel Shader
The pixel shader uses the brightly colored rainbow film map as a base texture. This texture
map is meant to simulate the soap film appearance of a bubble. The environment map is modu-
lated with this texture. The white highlights are stored in the alpha of the environment map as a
glow map and are linearly interpolated into the result. The final alpha value used from blend-
ing with the frame buffer is the (1abs(NV)) + glow map. The full shader is shown below, and
further descriptions follow.
// c0 (0.0, 0.5, 1.0, -0.75)
// c1 - (0.6, 0.1, 0.0, 0.0) Alpha Scale and bias
ps.1.4
texld r0, t0
texld r1, t1
texcrd r2.rgb, t2
cmp r2.r, r2.r, r2.r, -r2.r // abs(V.N)
+mad_x4_sat r1.a, r1.a, r1.a, c0.a
mul_x2_sat r2.rgb, r0, r1 // base * env (may change scale factor later)
+mad r2.a, 1-r2.r, c1.x, c1.y // alphascale * abs(V.N) + alphabias
lrp r0.rgb, r1.a, r1, r2 // Lerp between Env and Base*Env based on glow map
+add r0.a, r2.a, r1.a // Add glow map to Fresnel term for alpha
The first thing that happens in this pixel shader is the sampling of the environment map. This
happens in the following line of the pixel shader:
texld r1, t1
372 Part 3: Pixel Shader Tricks
Team LRN
It produces the following image:
This is multiplied by the base texture color (shown in Figure 5) using the following code:
mul_x2_sat r2.rgb, r0, r1 // base * env (may change scale factor later)
The results are shown in Figure 6.
Now that the basic colors are computed, the opacity needs to be computed. The first step is to
create strong specular lighting term. The alpha of the environment map (shown in Figure 7) is
squared, multiplied by four, and given a bit of a bias. The following pixel shader co-issued
instruction accomplishes this:
+mad_x4_sat r1.a, r1.a, r1.a, c0.a
Bubble Shader 373
Figure 4: Cubic environment map only Figure 5: Base map
Figure 6: Cubic environment map
modulated with base map
Figure 7: Alpha for environment map
Team LRN
The resulting image is shown in Figure 8.
Next, the Fresnel term is computed. This takes place in a few stages. First, take the abso-
lute value of VN, and then invert this value by computing 1abs(VN). This value is then
scaled and biased. The following code shows how this is computed:
cmp r2.r, r2.r, r2.r, -r2.r // abs(V.N)
+mad r2.a, 1-r2.r, c1.x, c1.y // alphascale * abs(V.N) + alphabias
Figure 9 shows the result of this computation.
The next step is to add the Fresnel term with the glow map using the following code:
+add r0.a, r2.a, r1.a // Add glowmap to Fresnel term for alpha
This produces the final alpha result used for blending and is shown in Figure 10.
374 Part 3: Pixel Shader Tricks
Figure 8: Alpha of the environment map
squared and multiplied by four
Figure 9: Scaled and biased Fresnel term
Figure 10: Final alpha values Figure 11: Final color values
Team LRN
Next, the glow map is used to linearly interpolate between just the environment map and the
base multiplied by the environment map based on the alpha value in the environment map:
lrp r0.rgb, r1.a, r1, r2 // Lerp between Env and Base*Env based on glow map
This produces the final color values shown in Figure 11.
Figure 12 shows the blend produced by the alpha
and color to create the final image.
See BubbleShadeTree.jpg on the companion CD
for a graphic representation of all the above steps. Fig-
ure 13 shows what this shader creates within a scene
produced by the ATI Sushi engine.
Figure 12: Final image
Summary
This article has shown a technique to produce an interesting bubble effect built on top of some
techniques described in other articles. It used sine waves computed within a vertex shader to
produce an undulating shape and cubic environment mapping to give interesting reflections.
Bubble Shader 375
Figure 13: Bubble in an environment
Team LRN
Per-Pixel Strand-Based
Anisotropic Lighting
John Isidoro and Chris Brennan
Introduction
This technique uses Direct3D pixel shaders to implement a per-pixel version of the per-vertex
anisotropic lighting technique presented in [Heidrich99]. An anisotropic lighting model is
meant to simulate the effects of lighting on a grooved surface, like a record, a Christmas orna-
ment, human hair, brushed metal, or some fabrics. The per-pixel nature of the shader uses a
direction of an anisotropy map to determine the direction of the grooves. This allows us to be
able to author many different anisotropic surface types by using standard image editing
packages.
Strand-Based Illumination
The basic idea behind the strand-based illumination is the lighting of strands. These strands are
modeled as having a circular cross section. When lighting them, we use the standard Phong
illumination model but with a slight modification. The normal used in the Phong model is
equal to the light vector projected into the normal plane of the strand at the point being lit and
normalized. We call this strand normal vector N'. The modified Phong equation is:
I
o
= k
a
I
a
+ (k
d
(L N') + k
s
(V R')
n
)I
i
The terms k
a,
k
d,
and k
s
are material properties for the ambient, diffuse, and specular lighting,
respectively. I
a
is the ambient light color, and I
i
is the incident light color. L is the light direc-
tion vector, V is the view direction vector, and n is the specular exponent. The reflection vector
R' can be found using:
R' = 2N'(N'V) V
As shown in [Banks94] and [Stalling97], using some trigonometric identities and the knowl-
edge that L, T, and N' are coplanar, LN' and VR' can be expressed in terms of L, V, and T, the
direction of the strand.
LN' = (1 (LT)
2
)
0.5
VR' = (1 (LT)
2
)
0.5
(1 (VT)
2
) )
0.5
(LT) (VT)
376
Team LRN
Luckily, these equations are a function of two dot products, LT and VT. This allows us to use
these dot products to index into a precomputed 2D texture map to compute the diffuse LN' and
specular (VR')
n
terms. Using the fact that L, T, and V are unit length vectors, it is known that
the terms LT and VT range between [1,1]. To use these terms to index into a texture lookup
table, they first must be scaled and biased into the range of [0,1].
Using this knowledge, building the texture lookup table tex(u,v) is straightforward. In the
equations above, substitute 2u1 for LT and 2v1 for VT. The diffuse component can be
packed into the RGB component of the texture, and the specular component can be packed into
the alpha channel of the texture.
This way both diffuse and specular lookups can be done with a single texture fetch per
light source. Also, the specular exponent n can be pre-baked into the specular lookup table by
computing (VR')
n
instead of just computing (VR').
The rest of the math in the Phong model can be computed within the pixel shader.
Rendering Using the Texture Lookup Table
Depending on whether the technique is applied per-pixel or per-vertex, there are a few places
where the dot products, scaling, and biasing can be computed. When rendering on a per-vertex
basis, all the computation can easily be performed in a vertex shader or even in a traditional
texture matrix. The tangent vector T can be stored with the vertex data as a per-vertex
three-element texture coordinate. The light L and view V vectors can be stored as vertex shader
constants. All vectors should be in object space if the texture matrix is used to perform the dot
products. This is explained in more detail in [Heidrich99]. The main focus of this chapter is to
illustrate the implementation of anisotropic lighting on a per-pixel basis and to present a
unique orthonormalization technique that allows anisotropic lighting to be combined with
bump mapping.
Per-Pixel Strand-Based Anisotropic Lighting 377
Figure 1: Diffuse texture lookup table Figure 2: Specular texture lookup
table
Team LRN
Per-Pixel Strand-Based Illumination
Per-pixel strand-based anisotropic lighting has a lot in common with standard per-pixel light-
ing (aka bump mapping). First off, all per-pixel computation is done in tangent space. The
direction of anisotropy is stored in a texture map and, being in tangent space, has the advantage
that the direction map is not coupled to the geometry. The same direction map can be applied
to many different objects and in a way, can be thought of as part of the surface description in
the same way a bump map is. The direction vectors T
t
are encoded in the texture the same way
normals are in a bump map, and the xyz direction is scaled by 0.5, biased by 0.5, and encoded
as an RGB color. The light and view vectors are computed and transformed into tangent space
in the vertex shader. The resulting vectors L
t
and V
t
are interpolated across the triangle.
In the pixel shader, two dot products need to be per-
formed (T
t
L
t
), and (T
t
V
t
), and the resulting vector
should be scaled and biased by 0.5 and used to fetch a
texel in the texture lookup table. However, we can make
a few optimizations here.
The first optimization is to perform the scale by 0.5
inside the vertex shader by multiplying L
t
and V
t
by 0.5
before outputting them. The second optimization is to
offset the texture lookup table in u and v by half the tex-
ture width and height, respectively. By doing this and
turning on texture wrapping in the texture stage contain-
ing the lookup map, this essentially gives us the bias by
0.5 for free. These optimizations save precious pixel
shader instructions and in the case of 1.1 pixel shaders,
make it possible to perform the anisotropic lighting at
all.
378 Part 3: Pixel Shader Tricks
Figure 3: Red channel of direction of
anisotropy map
Figure 4: Green channel of direction
of anisotropy map
Figure 5: Specular component shown
offset by half of the texture width
Team LRN
Per-Pixel Strand-Based Illumination with
Colored Light and Base Map
Here is a pixel shader that implements the anisotropic strand illumination:
ps.1.1
tex t0 ; Contains direction of anisotropy in tangent space
tex t1 ; base map
texm3x2pad t2, t0_bx2 ; Perform fist row of matrix multiply.
texm3x2tex t3, t0_bx2 ; Perform second row of matrix multiply to get a
; 3-vector with which to sample texture 3, which is
; a look-up table for aniso lighting
mad r0, t3, t1, t1.a ; basemap * diffuse + specular
mul r0, r0, c0 ; color * basemap *diffuse
Per-Pixel Strand-Based Illumination with Four
Colored Lights and Base Map in One Pass
Here is a pixel shader that implements
the anisotropic strand shader with four
colored lights:
Figure 9: Anisotropic lighting
with four colored lights. Also
shown in Color Plate 9.
Per-Pixel Strand-Based Anisotropic Lighting 379
Figure 6 Figure 7 Figure 8: Result of per pixel
anisotropic lighting (only two
triangles!)
Team LRN
1.4 pixel shader:
ps.1.4
texld r0, t0 //sample bump map (note alpha channel = 1)
texcrd r1, t1
texcrd r1, t2
texcrd r1, t3
texcrd r1, t4
texcrd r1, t5
dp3_d2 r1.rba, r1, r0_bx2 // .5 L0.T
dp3_d2 r2.rba, r2, r0_bx2 // .5 L1.T
dp3_d2 r3.rba, r3, r0_bx2 // .5 L2.T
dp3_d2 r4.rba, r4, r0_bx2 // .5 L3.T
dp3_d2 r1.g, r5, r0_bx2 // .5 V.T
mov r2.g, r1.g
mov r3.g, r1.g
mov r4.g, r1.g
phase
texld r1, r1 //aniso lookup 0
texld r2, r2 //aniso lookup 1
texld r3, r3 //aniso lookup 2
texld r4, r4 //aniso lookup 3
texld r5, t0 //sample base map
mad r1, r1, r5, r1.a // diffuse0 * base + specular
mad r2, r2, r5, r2.a // diffuse1 * base + specular
mad r3, r3, r5, r3.a // diffuse2 * base + specular
mad r4, r4, r5, r4.a // diffuse3 * base + specular
mul r0, r1, c1 // lightcolor0 * light0 effect
mad r0, r2, c2, r0 // lightcolor1 * light1 effect + previous
mad r0, r3, c3, r0 // lightcolor2 * light2 effect + previous
mad r0, r4, c4, r0 // lightcolor3 * light3 effect + previous
Per-Pixel Bump-Mapped Strand-Based Illumination
Using Gram-Schmidt Orthonormalization
Another technique for bump mapping combines both anisotropic strand-based lighting and
bump mapping at the same time. This has way more surface detail and can be added to objects
using the anisotropic strand shader.
This shader uses the bump map to perturb the direction of anisotropy by using
Gram-Schmidt orthonormalization. The basic idea is to make
the direction of anisotropy orthogonal to the bump map nor-
mal by subtracting off the component of the tangent T vector
projected onto the normal N and renormalizing the resulting
vector:
T
new
= || T (N T) N ||
There are two possible places this math can be performed,
either as a preprocessing step on the direction of anisotropy
texture or inside the pixel shader. Doing the orthonor-
malization as a preprocessing step has the advantage of not
requiring any extra math in the pixel shader. However, it has
380 Part 3: Pixel Shader Tricks
Figure 10
Team LRN
the negative effect of tying the preprocessed direction map to the corresponding bump map. If
the same direction map needs to be used with many different bump maps, additional textures
need to be allocated. In addition, if the texture coordinates for the bump and direction maps are
different, this preprocessing step is usually not possible.
The alternative is to do the orthonormalization in the pixel shader, as is done in the follow-
ing example. All the math is fairly straightforward to perform in the pixel shader except for the
vector normalization step. Luckily for us, this renormalization step has almost no visual impact
due to the way the math works out. In the worst case, the lighting near the severe bumps may
become slightly brighter. In addition, for most surfaces, the direction of anisotropy runs in the
direction of the grooves in the bump map, thus making the bump map vectors perpendicular to
the direction of anisotropy.
; t1 contains tangent-space L vector ; Lx, Ly, Lz, 0
; t2 contains tangent-space reflection vector ; Vx, Vy, Vz, 0
; Map bound at stage 0 is a normal map
; Map bound at stage 3 is a T.L x T.V map as the z positive face of a cube map
ps.1.4
texld r0, t0 ; Contains direction of anisotropy in tangent space
texcrd r2.rgb, t1 ; Light vector
texcrd r3.rgb, t2 ; View vector
texld r4, t3 ; Bump map
dp3 r5, r0_bx2, r4_bx2 ; T.N
mad r0.xyz, r4_bx2, -r5, r0_bx2 ; Tn = T - N(T.N)
; Calculate V.T and L.T for looking up into function map.
dp3_d2 r1.x, r2, r0 ; L.Tn
dp3_d2 r1.y, r3, r0 ; V.Tn
phase
texld r2, r1 ; Normalize Tn, and perform anisotropic lighting
; Function lookup
texld r3, t0 ; Base map
Per-Pixel Strand-Based Anisotropic Lighting 381
Figure 11: Anisotropically lit chalice
without bump mapping
Figure 12: Anisotropically lit chalice
with bump mapping
Team LRN
mul_sat r0.rgb, r3, r2 ; aniso diffuse * basemap
mad_sat r0.rgb, r2.a, r3.a, r0 ; + glossmap * aniso specular
+mov r0.a, c0.b
Summary
A technique for per-pixel strand-based anisotropy was described in this article. This technique
allows a per-texel direction of anisotropy vector as well as an optional per-texel bump map
normal to be specified. Due to the efficiency of the math used, up to four lights per rendering
pass can be achieved using DirectX 8.1 version 1.4 pixel shaders.
References
[Banks94] D. C. Banks, Illumination in diverse codimensions, SIGGRAPH Proceedings,
July 1994, pp. 327-334.
[Heidrich99] Wolfgang Heidrich and Hans-Peter Seidel, Realistic, Hardware-accelerated
Shading and Lighting, SIGGRAPH Proceedings, 1999.
[Stalling97] D. Stalling, M. Zckler, and H. C. Hege, Fast Display of Illuminated Field
Lines, IEEE Transactions on Visualization and Computer Graphics, 3(2), 1997, pp.
118-129.
382 Part 3: Pixel Shader Tricks
Team LRN
A Non-Integer Power
Function on the Pixel
Shader
Philippe Beaudoin and Juan Guardado
Overview
Raising a variable to a non-integer power is an operation often encountered in computer graph-
ics. It is required, for example, to compute the specular term in various shading schemes. It can
also be used to produce a smooth conditional function useful in many pixel shaders. In most
cases, the input variable falls between 0 and 1, while the exponent is greater than 1 and often
quite large.
Like many other techniques, a quality gain can be achieved by computing the function
per-pixel rather than per-vertex. This gain is very noticeable when using large exponents since
the function varies a lot and sampling it at each vertex is bound to miss visually important
details (see Figure 1 and Color Plate 10).
Therefore, we are particularly interested in finding a way to compute such a function on the
pixel shader. Like any pixel shader trick, it is important to minimize the number of textures and
blending stages since these are very limited resources. This article presents a simple shader
trick that performs a good per-pixel approximation of a non-integer power function. The tech-
nique works for input values between 0 and 1 and supports large exponents. The presented
383
Figure 1: Gouraud
shading (left) and
Phong shading (right).
(See also Color Plate
10.)
Team LRN
shader does not require any texture lookup and is scalable, making it possible to spend more
instructions in order to decrease the error or to reach greater exponents.
We first consider and analyze two typical techniques used to compute a power function on
the pixel shader. We then expose some mathematical background used throughout this article.
Finally, we show how the algorithm can be used to perform smooth conditional functions and
complex bump-mapped Phong shading. The actual implementation of the approximation as a
pixel shader program is discussed in detail.
Traditional Techniques
When confronted with the problem of computing a power function on the pixel shader, two
simple techniques come to mind. First, it seems possible to proceed through a 1D texture
lookup, and second, applying successive multiplications looks promising.
Texture Lookup
Linearly interpolated textures can be thought of as piecewise linear functions. In particular, 1D
textures with a linear filter are really a function taking a value between 0 and 1 and mapping it
onto another value in the same range. (Actually, such a function can map the input onto four
values, one for each of the R,G,B, and alpha channels. In the current case, however, were only
interested in one output value.) This looks promising for our problem since an input between 0
and 1 raised to any power greater than 0 yields a result between 0 and 1.
Listing 1 shows a piece of code that builds a 16-bit monochrome 1D texture of resolution
Resi to compute x
n
.
Listing 1: C function to create a 1D power texture
void ComputePowerTexture( int Resi, double n, unsigned short* Texture )
{
int i;
for( i=0; i<Resi; ++i )
Texture[i] = (unsigned short)( pow( (double)i/Resi, n ) * USHRT_MAX );
}
Once this texture has been constructed, a simple texture lookup pixel shader can be used to
perform the computation, provided that the value to raise to power n is placed in an interpo-
lated texture coordinate. The pixel shader in Listing 2 shows how to apply the power function
on the result of a dot product, as is often required. Note that this code only works for pixel
shader versions 1.2 and 1.3. The code for version 1.4 is presented in Listing 3.
Listing 2: Code computing a power function using a 1D texture (pixel shader versions 1.2 and 1.3)
ps.1.2
tex t0 ; Texture #0 lookup a vector
texdp3tex t1, t0 ; Use texture coordinates #1 as a 3D vector,
; performs dot product between this 3D vector and t0,
; using the result, lookup power function in 1D texture #1
mov r0, t1 ; emit the result
384 Part 3: Pixel Shader Tricks
Team LRN
Listing 3: Code computing a power function using a 1D texture (pixel shader version 1.4)
ps.1.4
texld r0, t0 ; Texture #0 look-up a vector
texcrd r1.rgb, t1 ; Load texture coordinates #1 as a 3D vector
dp3 r0, r0, r1 ; Performs dot product between this 3D vector and r0
phase
texld r1, r0.x ; using the result, look-up power function in 1D texture #1
mov r0, r1 ; emit the result
Various problems exist with this particular technique:
n
It uses up one texture stage, which may make it unfit for algorithms requiring many
textures.
n
Changing the value of the exponent n requires regenerating the texture, unless a 2D texture
is used, in which case a limited number of predefined exponents can be used.
n
For pixel shader versions before 1.4, a 1D texture lookup cannot be applied to intermediate
computation results unless multiple passes are used.
This last limitation is often a major drawback since, usually, the power function must be pre-
ceded by a vector renormalization (such as with a cube map) and a dot product. With large
exponents, the vector renormalization is especially important. This is due to the fact that the
maximum value of a dot product is the product of the length of the two vectors. If one of these
is not normalized, the dot product can never reach 1. When raised to a large exponent, a value
smaller than 1 will rapidly move toward 0. This translates to visual details being washed out.
Figure 2 shows a vector interpolated with and without normalization, followed by a dot prod-
uct, and then raised to a high power. It is obvious that the detail (for example, a specular spot)
is washed out in the second version.
A Non-Integer Power Function on the Pixel Shader 385
Figure 2: Result of not
normalizing vectors
before applying a
power function
Team LRN
Successive Multiplications
Since raising a value to an integer power simply requires multiplying a value with itself a num-
ber of times, it seems possible to approximate a non-integer power function through successive
multiplication steps.
For example, the pixel shader in Listing 4 shows how to raise t0 to power 16. Analyzing
this scheme indicates that log
2
n multiplications are required to raise a variable to the power n,
when n is a power of 2.
Listing 4: Power function using successive multiplications
ps.1.0
tex t0
mul r0, t0, t0 ; r0 = t0 *t0 = t0^2
mul r0, r0, r0 ; r0 = t0^2*t0^2 = t0^4
mul r0, r0, r0 ; r0 = t0^4*t0^4 = t0^8
mul r0, r0, r0 ; r0 = t0^8*t0^8 = t0^16
Listing 5 shows a pixel shader that raises t0 to power 31. Analyzing this shows that, in general,
for n in the range [2
a
,2
a+1
], the algorithm can require 2a multiplications and a temporary vari-
ables. (A step can be saved by combining a multiplication and an addition using the mad
instruction.)
Listing 5: Power function with a non-power-of-2 exponent using successive multiplications
ps.1.1
tex t0
mul t1, t0, t0 ; t1 = t0^2
mul t2, t1, t1 ; t2 = t0^4
mul t3, t2, t2 ; t3 = t0^8
mul r0, t3, t3 ; r0 = t0^16
mul r0, r0, t3 ; r0 = t0^16 * t0^8 = t0^24
mul r0, r0, t2 ; r0 = t0^24 * t0^4 = t0^28
mul r0, r0, t1 ; r0 = t0^28 * t0^2 = t0^30
mul r0, r0, t0 ; r0 = t0^30 * t0 = t0^31
This technique has the following limitations:
n
It only supports discrete changes in the exponent, making it impossible to change the value
of n in a continuous fashion.
n
It requires a lot of instructions for a large exponent.
n
It requires a lot of instructions and temporary variables for non-power-of-2 exponents.
These last two problems often limit the usefulness of successive multiplications, since practical
exponents have a tendency to be large and are usually not powers of 2.
The Need for a New Trick
Although the 1D texture lookup and the successive multiplications techniques have no limita-
tions in common, both are too restrictive to be really useful in general cases. In particular,
neither of them is suited to large exponents. In the case of 1D textures, the impossibility to per-
form a per-pixel renormalization before the lookup makes it unsuitable, and for successive
multiplications, the number of instructions required for large exponents is prohibitive.
386 Part 3: Pixel Shader Tricks
Team LRN
The power function approximation technique presented in the rest of this article addresses
all of the preceding issues. Therefore we think it can often be used as an efficient alternative in
pixel shaders that require a power function.
Mathematical Details
If we take x
n
in the range from 0 to
1, and then for a large enough n, the
function is very close to 0 on most
of its domain and abruptly goes to 1
when x approaches 1. This is shown
in Figure 3 for increasing values of
n. This will be the basic concept
used for developing our
approximation.
It is interesting to note that the
result x
n
is greater than 1/256 only
for values of x greater than 256
1/n
.
For example, if we take n = 16, the
function will be greater than 1/256
only if x is over 0.707; with n = 64,
this value becomes 0.917. Since
1/256 is the smallest value display-
able on 8-bits-per-channel
hardware, approximating x
n
with 0
will yield no perceptual error for
values of x between 0 and 256
1/n
.
Now, if we look at x
n
for input
values greater than 256
1/n
, we can
see that it looks pretty much like a
scaled and offset power function of
a lower degree. Figure 4 shows the
function x
16
being approximated by
x
4
scaled and offset horizontally to
reach 0 when x = 256
1/16
.
Therefore, using the null func-
tion from 0 to 256
1/n
and correctly
scaling a lower degree power func-
tion for the rest of the range seems
to constitute an adequate approxi-
mation. If our approximating
function uses an exponent m, a
scaled and offset version of the
function can be written as (Ax +
B)
m
, where A is the scale and B is
A Non-Integer Power Function on the Pixel Shader 387
Figure 3: Shape of a power function with increasing expo-
nents n = 2, 4, 8, 16, 64
Figure 4: Approximating the right part of a power function
with a lower degree exponent m
Team LRN
the offset. Now, naturally, we want this approximating function to be equal to 1 when x is 1.
Also, we want it to reach 0 when x = 256
1/n
. Therefore, we can solve for A and B. We find
that A = 1 / (1 256
1/n
) and B = 1 A. The approximating function can be written as:
By noting that x < 256
1/n
if and only if Ax + B < 0, the above function can be rewritten more
concisely as:
f(x) = max(Ax + B, 0)
m
Note that we will always consider m n in the rest of the text. This is because we want the
approximating function to have a lower degree than the original.
This technique can now be used
with real examples. Figure 5 shows
the result of approximating x
n
with
max(Ax + B, 0)
m
for n = 16 and m =
4. In this case, A and B are com-
puted as described earlier, and their
value is A = 3.4142 and B =
2.4142. The graph displays both
the original and approximated
curves. The normalized error func-
tion is also plotted in order to show
how the error is distributed to the
left and right of the point where
both curves cross.
A first analysis shows that the
approximation gives good results for
these values of A, B, n, and m.
However, if we look more closely at
the graph, we notice that the error is
not distributed equally on each side of the crossing point. This leads us to think that the maxi-
mal approximation error could be lowered by adjusting A and B in order to move the crossing
point.
In fact, for arbitrary values of n and m, the technique weve described to select A and B
doesnt give any guarantee on the maximal approximation error. In practice, however, it may
suit many applications.
In order to optimize the approximation error, one should seek values of A and B for which
the maximal error on the left and right side are equal. We solve this problem using the numeri-
cal approach presented next.
First we need a function that, given approximate values of the optimal A and B, is able to
compute the maximal error on the left and right side of the crossing point. To do so, we simply
find the zero of the derivative of the error function on the left and right sides and evaluate the
error function at these points. This is accomplished with the function EvaluateMaxError.
Then, we perform a binary search, changing our guess values for A and B in order to move
the crossing point. When the left error is greater than the right error, the crossing point is
388 Part 3: Pixel Shader Tricks
Figure 5: Approximation of a power function and normal-
ized error
Team LRN
moved to the left; otherwise, it is moved to the right. This is accomplished by the function
FindAB.
For any value of n and m such that 1 m n, the following algorithm is guaranted to pick
values of A and B that minimize the maximal approximation error.
Listing 6: C code to determine optimal scale and offset for some approximation
const double Epsilon = 0.0001;
const double ThresholdError = 0.0001;
// Find values for A and B minimizing the maximal error given n and m
void FindAB( double n, double m, double &A, double &B )
{
double k0 = 0;
double k1 = 1;
double k;
double ErrorLeft, ErrorRight;
// Binary search for optimal crossing point
do
{
k = (k0 + k1)/2.0;
A = ( pow( k, n/m ) - 1.0 ) / ( k - 1.0 );
B = 1.0 - A;
EvaluateMaxError( n, m, A, B, k, ErrorLeft, ErrorRight );
if( ErrorLeft < ErrorRight )
k0 = k;
else
k1 = k;
} while( fabs( ErrorLeft - ErrorRight ) > ThresholdError );
}
// Evaluate the maximal error to the left and right of the crossing point
// for given values of A and B
// The crossing point k is given to reduce computation
void EvaluateMaxError( double n, double m, double A, double B, double k,
double &ErrorLeft, double &ErrorRight )
{
double x;
double DerErr;
// Find maximal point of the error function on the left of the crossing point
double x0 = 0;
double x1 = k;
do
{
x = (x0 + x1)/2.0;
DerErr = DerError( x, n, m, A, B );
if( DerErr < 0 )
x0 = x;
else
x1 = x;
} while( fabs( x0-x1 ) > Epsilon );
A Non-Integer Power Function on the Pixel Shader 389
Team LRN
// Evaluate the error at that point
ErrorLeft = fabs( Error( x, n, m, A, B ) );
x0 = k;
x1 = 1;
// Find maximal point of the error function on the right of the crossing point
do
{
x = (x0 + x1)/2.0;
DerErr = DerError( x, n, m, A, B );
if( DerErr > 0 )
x0 = x;
else
x1 = x;
} while( fabs( x0-x1 ) > Epsilon );
// Evaluate the error at that point
ErrorRight = fabs( Error( x, n, m, A, B ) );
}
// Evaluate the approximation function
double Approx( double x, double m, double A, double B )
{
if( x < -B/A )
return 0;
return pow( A * x + B, m );
}
// Evaluate the derivative of the approximation function
double DerApprox( double x, double m, double A, double B )
{
if( x < -B/A )
return 0;
return A*m*pow( A * x + B, m-1 );
}
// Evaluate the error function
double Error( double x, double n, double m, double A, double B )
{
return Approx( x, m, A, B ) - pow( x, n );
}
// Evaluate the derivative of the error function
double DerError( double x, double n, double m, double A, double B )
{
return DerApprox( x, m, A, B ) - n*pow( x, n-1 );
}
Using this algorithm with the previous example, where n = 16 and m = 4, yields A = 3.525 and
B = 2.525. This is illustrated in Figure 6. It can be seen that the error is equally distributed to
the left and right of the crossing point, indicating that the maximal error has been minimized.
390 Part 3: Pixel Shader Tricks
Team LRN
It should be noted that by selecting
A and B through the algorithm of
Listing 6, we lose the property that
the max(Ax+B, 0)
m
is equal to 0
only for values of x
n
< 1/256.
However, this doesnt hurt our
approximation, since the maximal
error has been lowered.
Table 1 shows the maximal
approximation error for typical val-
ues of n and m. It also shows the
optimal values of A and B for these
values.
Table 1: Scale, offset, and maximal error for typical approximations
m n A B Max error m n A B Max error
2 3 1.336 0.336 0.025616 8 9 1.115 0.115 0.001862
4 1.683 0.683 0.037531 10 1.229 0.229 0.003314
5 2.033 1.033 0.044382 11 1.344 0.344 0.004514
6 2.385 1.385 0.048856 12 1.459 0.459 0.005527
10 3.802 2.802 0.057565 16 4.918 0.918 0.008200
18 6.645 5.645 0.063146 24 2.838 1.838 0.010886
4 5 1.208 0.208 0.006984 16 17 1.060 0.060 0.000491
6 1.417 0.417 0.011542 18 1.120 0.120 0.000920
7 1.628 0.628 0.014761 19 1.180 0.180 0.001321
8 1.838 0.838 0.017146 20 1.239 0.239 0.001653
12 2.681 1.681 0.022635 24 1.479 0.479 0.002682
20 4.370 3.370 0.026925 32 1.959 0.959 0.004028
Power Function on the Pixel Shader
Approximating a power function on the pixel shader requires us to translate the preceding
mathematical reasoning into the pixel shader assembly language. Doing so requires us to com-
pute the function max(Ax+B, 0)
m
through the set of available microcode instructions. We also
need a way to specify the variables present in this equation, namely A, B, m, and x.
We can rule out a number of these variables easily: The input variable x will simply be
stored in a general-purpose register, and the exponent m will be decided in advance. For vari-
ables A and B, we will consider two scenarios. At first, they will be fixed ahead of time and
A Non-Integer Power Function on the Pixel Shader 391
Figure 6: Approximation of a power function using opti-
mized scale and offset
Team LRN
their content will be stored in constant registers. In the second scenario, we will show how A
and B can be modified dynamically on a per-pixel basis.
Constant Exponent
Lets first study the case where A and B do not change per-pixel. In this scenario, A is placed
in the constant register c0 while B is placed in c1. This means that the exponent n being
approximated is constant as long as c0 and c1 remain unchanged.
Now we need to compute max(Ax+B, 0)
m
. First, the max(, 0) function is taken care of
using the _sat modifier available on the pixel shader. Then, we pick m as a power of 2, selected
to approximate with enough precision the target exponent n. We then perform a bias and scal-
ing with mad, followed by log
2
m self-multiplications with mul. The result is the pixel shader
in Listing 7, where the power function is applied to each element of the vector r0. It should be
noted that log
2
m + 1 is equal to the number of pixel shader stages required. Therefore, in an
actual shader, the number of free instructions could limit m. (Note that if we want to compute
the power function on a scalar rather than on a vector, we can co-issue the instructions, which
could possibly help us gain a number of stages.)
Listing 7: Basic code for approximating a power function
ps.1.0
... ; Place input value x into r0
mad_sat r0, c0, r0, c1 ; r0 = max( Ax + B, 0 )
mul r0, r0, r0 ; r0 = max( Ax + B, 0 )^2
mul r0, r0, r0 ; r0 = max( Ax + B, 0 )^4
.
. ; repeat (log2 m) times the mul instruction
.
mul r0, r0, r0 ; r0 = max( Ax + B, 0 )^m
There is an important problem with the previous pixel shader. In fact, since all constant regis-
ters must be between 1 and 1, we have a very limited range of values for A and B. Table 1
shows that, for typical values of n and m, A is always greater than 1. Therefore, in practice, the
proposed pixel shader is invalid.
To limit ourselves to scale and offset values in the allowed range, we first rewrite the
approximation function as follows:
max(Ax+B, 0)
m
= k max((Ax+B)k
1/m
, 0)
m
= k max(A'x+B', 0)
m
where we introduced variables A' and B' defined as:
A' = Ak
1/m
B' = Bk
1/m
Therefore, A and B can be written:
A = A'k
1/m
B = B'k
1/m
From these last two relations, we can see that with A' and B' between 1 and 1, we can obtain
values of A and B between k
1/m
and k
1/m
. Given that k is greater than 1, this translates to an
increased range for A and B.
392 Part 3: Pixel Shader Tricks
Team LRN
The pixel shader lets us compute k max(A'x+B', 0)
m
with k greater than 1 through its multi-
ply instruction modifiers _x2 and _x4. If we take the program in Listing 7 and apply such
modifiers to some or each of the mad or mul instructions, we will get a k greater than 1.
It is possible to compute k given a sequence of multiply instruction modifiers. This is per-
formed by the function ComputeK in Listing 8. Before calling any function in this listing,
make sure to correctly initialize the global array Multiplier so that it contains the correct multi-
ply instruction modifier for each instruction (either 1, 2, or 4). If we want to know which value
of A' and B' correspond to some values of A and B, we can use the computed k and the equa-
tion presented earlier; this is done with the function ComputeAB.
We go in the opposite direction and find the maximal values for A and B given k and m, as
performed by the function MaxAB. This result can then be converted in a maximal value for n,
as computed by MaxN.
Listing 8: C code for computing corrected scale and offset A and B
// Table containing multiply instruction modifier for each instruction (1, 2 or 4)
int Multiplier[] = { 4, 4, 4, 4, 4, 4, 4, 4 };
// Compute value of k at instruction i given a table of multiply instruction
modifiers
double ComputeK( int i )
{
if( i == 0 )
return Multiplier[i];
double Temp = ComputeK( i-1 );
return Multiplier[i] * Temp * Temp;
}
// Compute values of A and B given A and B and a multiplier table
// LogM: log of m in base 2 (number of instructions 1)
void ComputeApBp( int LogM, double A, double B, double &APrime, double &BPrime )
{
double Temp = 1.0/MaxAB( LogM ); // Note that k 1/m = 1/MaxAB
APrime = A * Temp;
BPrime = B * Temp;
}
// Compute maximum absolute values for A and B given some m and a multiplier table
// LogM: log of m in base 2 (number of instructions 1)
double MaxAB( int LogM )
{
double m = pow( 2.0, LogM ); // Find the value of m
double K = ComputeK( LogM ); // Compute K
return pow( K, 1.0/m );
}
// Compute maximum possible exponent given some m and a multiplier table
// LogM: log of m in base 2 (number of instructions - 1)
double MaxN( int LogM )
{
double m = pow( 2.0, LogM ); // Find the value of m
A Non-Integer Power Function on the Pixel Shader 393
Team LRN
double Max = MaxAB( LogM );
double A;
double B;
double n0 = m; // Lower bound for maximal exponent
double n1 = 5000; // Upper bound for maximal exponent
double n;
do
{
n = (n0 + n1)/2.0;
FindAB( n, m, A, B );
if( fabs(A) > Max || fabs(B) > Max )
n1 = n;
else
n0 = n;
} while( fabs( n0 - n1 ) > Epsilon );
return n;
}
It can be seen that the maximum value n is obtained when the modifier _x4 is used for each
instruction in the code. Given that the value for A' and B' are stored in constant registers c0 and
c1, respectively, the pixel shader in Listing 9 performs this approximation.
Listing 9: Corrected code for approximating a power function
ps.1.0
... ; Place input value x into r0
mad_x4_sat r0, c0, r0, c1 ; r0 = 4 * max( A*x + B, 0 )
mul_x4 r0, r0, r0 ; r0 = 4 * (4*max( A*x + B, 0 ))^2
mul_x4 r0, r0, r0 ; r0 = 4 * (4*(4*max( A*x + B, 0 ))^2)^2
.
. ; repeat (log2 m) times the mul
.
mul_x4 r0, r0, r0 ; r0 = 4^(2m-1) * max( A*x + B, 0 )^m
Table 2 shows the maximal range for A and B and the maximal exponents n that can be
obtained with the previous shader for various values of m.
Table 2: Maximal values of A, B, and n available, depending on approximation exponent m.
m Range for A and B Maximal n
2 [ 8, 8 ] 21.81
4 [ 11.314, 11.314 ] 52.87
8 [ 13.454, 13.454 ] 116.34
16 [ 14.672, 14.672 ] 244.08
32 [ 15.322, 15.322 ] 499.99
Naturally, if for a given m we want to limit ourselves to exponents smaller than the maximal n
listed in this table, we can remove some _x4 modifiers or replace them by _x2. When doing
394 Part 3: Pixel Shader Tricks
Team LRN
this, we figure the new maximal n by using the MaxN function with an updated Multiplier
array.
Not using _x4 modifiers at each instruction is often a good idea since it can help reduce
the numerical imprecision often present in pixel shaders. This lack of precision is mostly
noticeable for small values of n since they translate to values of c0 and c1 close to zero. Such
small numbers may suffer from an internal fixed-point representation and yield visual artifacts.
Per-Pixel Exponent
Until now, we have considered that the exponent to approximate n was fixed per-pixel and
could be stored in constant registers. However, being able to vary the exponent based on the
result of a texture lookup is sometimes required. The proposed pixel shader trick can be
extended to support that. To do so, we must decide a value for m and a sequence of multiplica-
tion modifiers ahead of time. These must be chosen in order to cover all possible values of n,
for n m.
Once we have picked these values, we simply take a texture image of the desired n and
translate each texel to their corresponding value of A. (Depending on its format, this texture
can contain up to four different exponents n per pixel.) The texels are then translated to A'
through a multiplication by k
1/m
. This time, however, instead of storing the result in a con-
stant register, we update the texture. The program in Listing 10 executes the process.
Listing 10: C code to generate an approximation texture based on an exponent texture
// Translates a texture of exponents n into values that
// can be directly used by the pixel shader
// Texture: monochrome texture in custom format
// ResX: X resolution of the texture
// ResY: Y resolution of the texture
// LogM: log of m in base 2 (number of instructions - 1)
void TranslateTexture( double** Texture, int ResX, int ResY, int LogM )
{
double A;
double APrime;
double Dummy;
double m = pow( 2.0, LogM ); // Find the value of m
for( int i=0; i<ResX; ++i )
for( int j=0; j<ResY; ++j )
{
FindAB( Texture[i][j], m, A, Dummy );
APrime = A/MaxAB( LogM ); // Compute A. Note that k 1/m = 1/MaxAB
assert( fabs(APrime) <= 1 ); // If assert fails select another m
// or change the multipliers
Texture[i][j] = APrime; // Update the texture
}
}
Once such a texture has been generated and placed in texture stage 0, we can easily extract A'
inside a pixel shader. We can also extract B' by recalling that B = 1 A. Since B' is the result
of multiplying B by k
1/m
, we can write B' = (1 A)k
1/m
= k
1/m
A'. Since m is fixed, we
store k
1/m
in constant register c0 and perform a simple subtraction to extract B'.
A Non-Integer Power Function on the Pixel Shader 395
Team LRN
The pixel shader of Listing 11 approximates a power function with n varying per pixel.
This shader uses modifiers _x4 for each instruction; the texture and the constant register c0
should therefore be generated accordingly.
Listing 11: Code for approximating a power function with an exponent n varying per pixel
ps.1.0
tex t0 ; Sample the value c0 for approximation
... ; Place input value x into r0
sub r1, c0, t0 ; Compute B = 1/MaxAB(LogM) A
mad_x4_sat r0, t0, r0, r1 ; r0 = 4 * max( A*x + B, 0 )
mul_x4 r0, r0, r0 ; r0 = 4 * 16 * max( A*x + B, 0 )^2
mul_x4 r0, r0, r0 ; r0 = 4 * 16 * 256 * max( A*x + B, 0 )^4
.
. ; repeat (log2 m) times the mul
.
mul_x4 r0, r0, r0 ; r0 = 4^(2m-1) * max( A*x + B, 0 )^m
This shader seems to show that one extra instruction is required to handle an exponent n vary-
ing per pixel. However, this is only true if we cannot spare additional texture space. In the case
where a texture or a texture component is still available, we could precompute the value of B'
and store it there. However, we believe that textures are often a more limited resource than
pixel shader instructions; this is why we suggest you use the approach presented above.
As a last remark, note that the power function often only needs to be computed on a scalar.
Therefore, provided that the input value x is placed in the alpha channel, the pixel shader can
co-issue instructions. In such a case, we can also limit our usage of constant registers by using
only the alpha channel of the constants. This technique is applied in the pixel shaders presented
in the rest of this article.
Applications
The nature of the power function makes it very interesting for a number of applications in
computer graphics. Any effect that should be visible at some point and fade out more or less
rapidly for neighboring points can benefit from such a function. In this section we will present
two techniques that require a per-pixel power function: smooth conditionals and specular
Phong shading.
Smooth Conditional Function
Lets first start by studying the conditional instructions available in the pixel shader. We take
cmp because it behaves similar to cnd while being more general:
cmp dest, src0, src1, scr2
This instruction loads dest with src1 or src2 depending on whether src0 is positive or not. If we
consider this instruction with values of src1 = 1 and src2 = 0, we obtain the function illustrated
in Figure 7.
396 Part 3: Pixel Shader Tricks
Team LRN
Since conditional instructions
can only take one of two values,
they often produce jagged-edged
artifacts. In fact, some neighboring
pixels end up having very different
colors because they sit on the thresh-
old of the condition.
To overcome this problem, we
can use the function illustrated in
Figure 8. We call such a function a
smooth conditional because it
smoothly and rapidly goes from one
value to the other as some threshold
is crossed.
The function illustrated in Fig-
ure 8 corresponds to the following
mathematical formula:
This formula can be computed on
the pixel shader using the code in
Listing 12. This code is similar to
Listing 9. The difference is that it
uses two extra cmp instructions, one
for the absolute value and one for
the condition x > 0. Also, note that
the last mul has a _x2 multiply
instruction modifier, although it
should be considered as a 4 in the
Multiplier table when generating A'
and B'. This is done in order to
account for the multiplicative factor
in the term 0.5 (1 |x|)
n
appearing in
the formula.
Listing 12: Code to perform a smooth conditional function (pixel shader versions 1.2 and 1.3)
ps.1.2
; c2 = -1, -1, -1, -1 ; Used for the sign function
... ; Place input value x into r0
cmp r1, r0, r0, -r0 ; r1 = |x|
mad_x4_sat r1, c0, 1-r1, c1 ; r1 = k0 * max( A*(1-|x|) + B, 0 )
mul_x4 r1, r1, r1 ; r1 = k1 * max( A*(1-|x|) + B, 0 )^2
mul_x4 r1, r1, r1 ; r1 = k2 * max( A*(1-|x|) + B, 0 )^4
.
. ; repeat (log2 m) times the mul
A Non-Integer Power Function on the Pixel Shader 397
Figure 8: Smooth conditional function
Figure 7: Standard conditional function
Team LRN
.
mul_x2 r1, r1, r1, c2 ; r1 = 0.5 * k * max( A*(1-|x|) + B, 0 )^m
cmp r0, r0, 1-r1, r1 ; r0 = 1 0.5*(1-|x|)^n if x > 0
; r0 = 0.5*(1-|x|)^n otherwise
We can use the above code to build a smooth conditional selection between two arbitrary val-
ues src1 and src2. To do so, we simply add a lrp instruction at the end of the previous pixel
shader. This linear interpolation is made to map 0 to src1 and 1 to src2. The result is a function
that mimics the standard cmp in a smooth fashion.
The presented code requires a lot of instructions; however, we can co-issue them if we
want the smooth conditional to act only on the alpha channel. Moreover, the simplified asym-
metric shape in Figure 9 often
constitutes a good enough smooth
conditional for x > 0. Given that the
input x is saturated (greater or equal
to 0), this function can be expressed
as 1 (1 x)
n
. This is computed
with minimal change to Listing 9.
We simply need to use the invert
source modifier at the input and
output.
Other variations on the smooth
conditional pixel shader code allows
for various other arithmetic tests. In
fact, the simple power function x
n
can be considered as the smooth
conditional for x > 1.
Volume Bounded Pixel Shader Effects
A problem that may arise with pixel shaders is that sometimes per-pixel effects need to occur
only within a given volume. For example, say we have a magical spell that affects every pixel
within a given radius of the casting point. If we want that spell to turn on a specific per-pixel
effect, we need to find if each pixel is inside the spell radius using some pixel shader instruc-
tions. This can be done easily with a standard conditional instruction. However, if we do that,
then we will most likely suffer from the jagged-edged artifacts mentioned earlier. Therefore,
this is a good place where the smooth conditional function could be used.
So lets see how we can build a pixel shader that equals 1 for pixels within some radius r
of a point P and smoothly goes to 0 for pixels outside this radius. First, for each vertex, well
require a vector R that joins point P to the vertex. This vector needs to be scaled so that its
length equals 1 for a vertex that is exactly at distance r from P. We place the vector P in
texture coordinates 0. We will directly use this vector; therefore, no texture needs to be bound
to stage 0.
We then select the approximation exponent m for the power function used in the smooth
conditional. In the presented example, we take m = 4. The multiply instruction modifier _x4
was used with each instruction. Since we use a fixed exponent n for the shader, we place A'
and B' in constant registers c0 and c1.
398 Part 3: Pixel Shader Tricks
Figure 9: Simplified smooth conditional function
Team LRN
The pixel shader in Listing 13 uses this technique to apply a simple diffuse lighting equa-
tion to all pixels within some radius of a given point. The diffuse color is placed in interpolated
color v0, texture 1 contains the decal texture to apply, and texture 2 holds an extra diffuse light
map. A constant ambient factor is stored in constant register c2. More complex effects could
use the same pixel shader trick in order to limit themselves to a volume. Also, more complex
volumes could be devised by applying some per-vertex or per-pixel process on P.
Listing 13: Code to perform volume bounded pixel shader lighting
ps.1.0
; c0 ; A
; c1 ; B
; c2 ; Ambient factor
; c7 = 1, 1, 1, 1 ; Uniform white, used in shader
; Texture address ops
texcoord t0 ; Pass on vector P
tex t1 ; Fetch decal texture
; Pixel ops
dp3_sat r0, t0, t0 ; r0 = max(0, P.P) = |P|^2
add r0.rgb, v0, t2 ; Compute diffuse lighting by adding v0 and t2
+mad_x4_sat r0.a, c0, r0, c1 ; r0.a = k0 * max( A*x + B, 0 )
mul r0.rgb, r0, t1 ; Apply diffuse lighting to decal
+mul_x4 r0.a, r0, r0 ; r0.a = k1 * max( A*x + B, 0 )^2
mad r0.rgb, c2, t1, r0 ; Add ambient contribution
+mul_x4 r0.a, r0, r0 ; r0.a = k * max( A*x + B, 0 )^4
mul r0.rgb, 1-r0.a, r0 ; 1-r0.a = SmoothConditional( |P|^2 < 1 )
; Multiply by result of shading
+mov r0.a, c7 ; Clear alpha
Phong Shading
One of the major problems of per-pixel lighting resides in computing the specular component
of the final color. This component is due to the reflection of the light itself on a shiny material.
It is easy to see that the reflection of a point light on a perfectly shiny sphere is a single point.
Unfortunately, very few surfaces are perfectly shiny. To account for non-perfect reflections,
sometimes called dull reflection, Phong (and Warnock before him) introduced a very popular
model. This model relies on the use of a power function in which the exponent depends on the
shininess of the material.
Pixel shader algorithms to perform Phong shading have been described before. However,
they performed the per-pixel power function using the usual techniques described at the begin-
ning and suffered from the problems of such techniques. This often leads to multipass
algorithms or reduces the control over the shininess exponent.
We will now show how, using the power function algorithm exposed earlier, we can per-
form realistic per-pixel Phong shading including a normal map, a color diffuse texture map, a
color specular texture map, and a shininess map.
A Non-Integer Power Function on the Pixel Shader 399
Team LRN
Phong Equation with Blinn Half-Vector
First, we need to express our shading model mathematically. We suggest reading computer
graphics textbooks for more information on the equations presented in this section.
Before we can express this equation, we need to detail the variables that are needed. First,
the shading depends upon the light direction, described by unitary vector L. We also need to
know the surface normal at the shading
point, N. Finally, we need to use some
vector to compute the specular compo-
nent. As is often the case for per-pixel
shading, we do not directly use the
Phong equation. Instead we take the
Blinn half-vector H, which is the unitary
vector falling directly between the view
direction V and the light direction L.
Figure 10 illustrates these vectors.
When using normal mapping, we
distinguish between the perturbed nor-
mal used for lighting, N', and the real
surface normal, N.
The scalar values that are needed for the computation are the ambient, diffuse, and specu-
lar coefficients: MAmb
c
, MDiff
c
, and MSpec
c
for the material and LAmb
c
, LDiff
c
, and LSpec
c
for the light. Here the index c is R, G, or B to indicate one of the color components.
The shading equation can now be written as:
Expressing the Inputs
The Phong equation has a number of inputs that we need to be able to pass down to the pixel
shader. We will first make a number of assumptions that reduce the number of inputs required
for the above function.
First, we suppose that the ambient, diffuse, and specular coefficients do not vary from one
pixel to the next. This means that the algorithm proposed cannot handle slide projectors or
other kinds of textured lights.
We also suppose that the material ambient coefficients do not vary arbitrarily from pixel to
pixel. Instead, these are linked to the material diffuse coefficients using the equation MAmb
c
=
KAmb
c
*MDiff
c
, with KAmb
c
constant within the shader. The equation states that the ambient
color is always the same as the diffuse color up to some constant. This is not a very limiting
assumption, since this relationship between material ambient and diffuse coefficients is often
witnessed in practice.
The three coefficients MDiff
c
and exponent n can vary from pixel to pixel; therefore, they
need to be expressed in a four-component, 2D texture map. Recall that we do not store n
directly; instead we place the corresponding value A' in the textures fourth component. Coeffi-
cients MSpec
c
can also vary at each pixel and can be placed in a separate three-component, 2D
400 Part 3: Pixel Shader Tricks
Figure 10: Vectors required for Phong shading
Team LRN
texture. These coefficients act as a color gloss map, effectively an extension of the traditional
single-component specular gloss map.
The perturbed normal N' is expressed as a tangent-space normal map. We therefore use a
2D texture map containing color-encoded normalized vectors that can be accessed using a sin-
gle 2D texture lookup. We refer the reader to other real-time shading texts to learn more on
tangent space normal maps.
To effectively use a tangent-space normal map in our lighting equation, we need to have a
normal map representation of the light vector L and halfway vector H. As discussed earlier, the
halfway vector needs to be renormalized per pixel;
otherwise, important visual artifacts will occur when
the power function is applied. Therefore, we interpo-
late H through a texture coordinate and use a cube
map lookup to renormalize it. A renormalization cube
map, as illustrated in Figure 11, contains a
color-encoded vector in each texel corresponding to
the unitary vector pointing in the direction of that
texel.
Since the light vector is only used in linear com-
putations, not renormalizing it has almost no impact
on the visual result. Therefore, we skip per-pixel
renormalization in order to make the best usage out of
our texture resources. This means that we can store
the light vector L in an interpolated color.
The Pixel Shader
Lets put all this together into a single pixel shader. First, each texture stage and interpolated
color is assigned as follows:
Table 3: Content of pixel shader interpolators and textures
Register Vector Texture Content
Texture 0 Diffuse u, v (MDiff
R
, MDiff
G
, MDiff
B
, A)
Texture 1 Normal map u, v Normal map N
Texture 2 Specular u, v (MSpec
R
, MSpec
G
, MSpec
B
, 0)
Texture 3 Halfway vector H Renormalization cube map
Color 1 L = Sign-corrected light vector L
Note that since interpolated colors do not support negative values, we place the sign corrected
light vector in color 0. This can be written as L' = 0.5( L + (1,1,1)).
We need to have access to KAmb
c
and LAmb
c
, which are constant within the shader. Due to
the nature of the equation, we can precompute the product of these coefficients. Therefore, we
store the resulting vector (KAmb
R
*LAmb
R
, KAmb
G
*LAmb
G
, KAmb
B
*LAmb
B
) in the first three
components of constant register c0.
We also need the values of constant coefficients LDiff
c
and LSpec
c
. These are stored in the
first three components of registers c1 and c2, respectively.
A Non-Integer Power Function on the Pixel Shader 401
Figure 11: Renormalization cube map
Team LRN
Finally, we need to pick the approximation exponent m used for the power function. In our
example, we use m = 8. We also use _x4 as the multiply instruction modifier for each instruc-
tion of the power function approximation. Since we wish to use a value of n varying per pixel,
we must precompute k
1/m
= 0.074325. We store this value in the alpha channel of constant
register c0. Looking at Table 2, we find that the maximal n that can be achieved is 116.34.
Also, using the function TranslateTexture in Listing 10, we can convert per-texel values of n
into values of A' to be placed in the alpha component of texture 0.
The pixel shader that computes per-pixel Phong shading then becomes:
Listing 14: Code for Phong shading with various features (pixel shader versions 1.1, 1.2, and 1.3)
ps.1.1
; v1 RGB: light vector (L) (sign corrected)
; c0 RGB: ambient light (Kamb*Lamb), A: 1/MaxAB(logM) = 0.074325
; c1 RGB: diff light (Ldiff), A: 0
; c2 RGB: spec light (Lspec), A: 0
tex t0 ; diffuse texture (Mdiff) + A' (exponent')
tex t1 ; normal map (N')
tex t2 ; specular texture (Mspec)
tex t3 ; normalised halfway vector (H)
dp3_sat r0.rgb, v1_bx2, t1_bx2 ; L.N'
+ sub r0.a, c0.a, t0.a ; B'
dp3_sat r1, t3_bx2, t1_bx2 ; H.N'
mad r0.rgb, r0, c1, c0 ; Kamb*Lamb + Ld*L.N'
+ mad_x4_sat r1.a, t0.a, r1.a, r0.a ; p0 = k0*max(Ax+B, 0)
mul r1.rgb, t2, c2 ; Ms*Ls
+ mul_x4 r1.a, r1.a, r1.a ; p1 = p0*p0
mul_x4 r1.a, r1.a, r1.a ; p2 = p1*p1
mul_x4 r1.a, r1.a, r1.a ; p = p2*p2 = H.N'^n
mul r1.rgb, r1, r1.a ; Ms*Ls*H.N'^n
mad r0, r0, t0, r1 ; Ma + Md*Ld*L.N' + Ms*Ls*H.N'^n
Summary
We have presented a method of approximating a non-integer power function on a pixel shader
by using as few texture stages as possible and gracefully degrading in accuracy depending on
the desired exponent and available number of blend stages. Furthermore, the technique can be
used for single or multiple channels, thus adapting nicely to individual shader requirements.
This method has several advantages over traditional exponentiation techniques that use
either a texture lookup or a series of sequential multiplications. Texture lookups are only accu-
rate for pixel shader versions 1.4 and greater, and even then they will require two phases.
Sequential multiplications need a large number of stages to compute high power-of-two expo-
nents and even more for non-power-of-two. Additionally, since the multiplications are
inherently uniform during the entire shader, they do not allow for a smooth variation in power.
A couple of applications were suggested, and Phong shading in particular is covered in the
results below. We believe such a useful technique can be applied to many other algorithms that
402 Part 3: Pixel Shader Tricks
Team LRN
require a power function, especially considering that it can be abstracted to any effect requiring
a sharp yet smooth transition, such as a spotlights cone falloff.
The per-pixel variation of the exponent, which is a handy extension to the basic principle,
can provide important visual cues for surfaces whose specularity varies, such as for a material
including both metallic and organic features. Its main disadvantages are that m constrains the
lower bound of the specular exponent n, as explained in the mathematical details section, and
that one component of a texture must be used to encode the exponent. The latter, however, is
expected of any technique that varies the exponent per-pixel.
The shader code in Listing 14 was applied to red spheres with different properties, such as
faceted per-pixel exponent maps, wrinkled normal maps, and orange specular maps. The
results can be seen in Figure 12 and Color Plate 11. Due to the many iterative multiplications,
there is a large accumulation of error that manifests itself as banding. Generally speaking, the
greater the exponent, the more banding that will be evident; however, this is mostly noticeable
on smooth surfaces, such as those expressed with uniform normal maps. The banding artifacts
are less significant when using normal maps that contain some perturbation because the specu-
lar falloff is quite sharp due to the abrupt changes in the normal. Therefore, visual artifacts are
reduced as detail in the normal map is increased.
Banding can also result from reduced instruction counts. If additional instructions are avail-
able, we recommend using them to maximize precision. Note that since the Phong shading
algorithm presented only requires exponentiation of a single component, the instructions easily
fit into the scalar pipeline of the pixel shader, which reduces the number of dedicated stages.
The shader code in Listing 14 consumes two stages purely for exponentiation purposes, but
instructions in the vector pipeline can be co-issued at these stages if desired.
A Non-Integer Power Function on the Pixel Shader 403
Figure 12: Per-pixel
Phong shading on
spheres. Clockwise from
upper left: uniform nor-
mal and exponent
maps, uniform normal
map and faceted expo-
nent map, noisy normal
and exponent maps
with uniform orange
specular material, and
noisy normal and expo-
nent maps with uniform
white specular material.
(See also Color Plate
11.)
Team LRN
For example, the specular texture can be removed if the diffuse and specular materials are
represented by a single texture. We can then use a cube map at that stage to represent the sur-
rounding environment, even encoding a Fresnel term in the remaining color register. The math
computations can easily be accommodated within the two remaining vector instructions.
Finally, the images in Figure 12 were rendered on hardware with 8 bits of fractional preci-
sion. Other hardware is available that has more than 8 bits of fractional precision and will
suffer less from the banding artifacts.
We hope you found this trick helpful and that you will find many more uses for an approx-
imating power function in your shading endeavors.
404 Part 3: Pixel Shader Tricks
Team LRN
Bump Mapped BRDF
Rendering
dm Moravnszky
Introduction
With this effect, we introduce the concept of the BRDF and show how they can be rendered
efficiently by using vertex shaders. I will combine the BRDF technique with normal mapping
to get a bumpy BRDF effect, taking full advantage of version 1.4 pixel shaders.
405
Figure 1: A teapot drawn with the BRDF effect and illuminated by four lights. The polygonal outline
is caused by the self-shadowing algorithm, not the BRDF shader. (See also Color Plate 12.)
Team LRN
Bidirectional Reflectance Distribution Functions
Bidirectional Reflectance Distribution Functions (BRDFs) are a very important milestone
along the road to photorealistic rendering. The BRDF is a generalization of all shading models,
such as Lambert, Blinn, or Phong. A shading model is an analytic approximation of a surfaces
reflective properties; it describes how incoming light is reflected by a surface. The simplest
shading model, Lambert shading, says that a surface reflects incoming light equally in all
directions. Phong shading, on the other hand, has an additional additive specular term which
permits a surface to reflect different amounts of light in different directions. Thus, each of
these models is a formula which computes, for a given incoming and a given outgoing light
direction, the portion of incoming light that gets reflected in the outgoing direction. Such func-
tions are called BRDFs. Formally, the scalar valued BRDF is a five-dimensional function
because the reflected light may not only depend on the incoming and outgoing direction vec-
tors (which each have two degrees of freedom in three space) but also on the wavelength of the
light. However, in computer graphics, we always decompose wavelengths into a sum of dis-
crete color channels, like red, green, and blue. This way the BRDF becomes a 3D vector
valued function with four parameters. Permitting the BRDF to be color dependent is important
because some materials reflect certain colors differently than others: just look at the rainbow
colored pattern on a CD. This pattern has another important property that neither Lambert nor
Phong shading can capture, but it can be described by more general BRDFs: that of anisotropic
reflection. The CD has tiny grooves, and light coming in orthogonally to them is in for a stron-
ger reflection.
Perhaps the primary advantage of a BRDF is that it can be measured directly from real
world objects; see [Marschner00] for a technique to do this via an inexpensive digital camera.
There are also extensive databases of measured BRDF data available on the Internet. This data
generally comes as a table of color values for a long list of incoming and outgoing directions
(usually in polar coordinates).
With all the advantages, one has to be aware of the limitations: While BRDFs capture
more lighting effects than any of the simple lighting models, even they cannot capture the
effects of transmitted light (glass, water), subsurface light transport (marble, skin), or heteroge-
neous surfaces (wood). There are other, still more general concepts which can do these things,
but they are beyond the scope of this article. On the other hand, BRDFs are great for rendering
ceramics, metals, plastics, leathers, and a whole host of other common materials.
Because of its generality, a BRDF is not only convenient but also relatively expensive. For
quite some time, BRDF rendering was beyond the scope of real-time graphics. In fact, it is still
not sensible to render BRDFs directly, say, from a 4D lookup table; it would simply take too
much memory for any reasonable sampling resolution. Instead, we are forced to go back and
pick an analytic, parametric model that can do a good job at approximating the BRDF. The key
idea is that we want to reduce the BRDF into a sum or product of two- or three-dimensional
functions, so that we can store a lookup table to these functions in a texture map and then use
additive or multiplicative blending to compute the result.
For this effect, I will use the parameterization presented by [McCool01]. The authors
decided to represent BRDFs as a product of three factors, plus an optional additive term for
very bright specular BRDFs: f(u,v) = p(u)q(h)p(v)+s(h), where u is the incoming direction, v
the outgoing direction, and h =
u v
u v

|| ||
the half-angle vector. The factors p, q, and s do not have
406 Part 3: Pixel Shader Tricks
Team LRN
any physical meaning; they are merely a way to factor the function. The full explanation for
this factorization is found in [McCool01], but it suffices to say that this is an expressive for-
mula (it approximates real BRDFs well), while using only three planar or cube map textures
(so it is very economic).
Decomposing the Function
By now, you have either measured or, more likely, downloaded a 4D table of BRDF sample
values for your favorite surface and would like to generate the texture maps we are going to
use as lookup tables for rendering. While a program to do this is not included with this effect,
writing the factorization code (especially in a matrix-oriented language like Matlab) based on
the description in [McCool01] should be straightforward. If you would like to put that off until
you have the real-time rendering portion, you can use the p, q, and s textures computed by the
authors of that paper for a set of interesting BRDFs, which you can get from their web site as
part of their OpenGL-based demo (http://www.cgl.waterloo.ca/Projects/rendering/Papers/
#homomorphic). I use a subset of these textures for my effect, which is implemented as an
Effects Browser plug-in.
Note that the BRDF textures are scaled up by a factor of to take full advantage of the
limited 8-bit range of the images. This factor is saved along with the textures and will be
needed during rendering to rescale the colors to the correct values.
Another trick is to use cube maps instead of simple planar textures. With cube maps, you
can directly use the incident, exitant, and half-angle vectors as texture coordinates, without
having to normalize them. Planar textures would require an additional transformation to project
these 3D vectors into a 2D texture coordinate.
Reconstruction
Reconstruction is the process of taking the BRDF factors tabulated in the textures and, for a
pair of given incident and exitant vectors, computing the value of the BRDF. While this is part
of the rendering process, it is not all of it. You still need the incident light, which you will
reflect. In the real world, various amounts of light are coming from all directions. Unfortu-
nately, you would need to integrate over all these directions per-pixel to simulate this, which is
clearly intractable. What you can do is only have a set of point lights illuminating your object.
Thus, for each light, you have a single incident direction, for which you have to evaluate the
BRDF once. The exitant direction is the only outgoing direction you are interested in: the
direction toward the camera. You can then accumulate the light originating from the different
lights L in the frame buffer via additive multipass blending by evaluating the sum:
The term after the BRDF is the incoming light at the surface. l is the intensity of your light
source, which, in nature, falls off with the square of the distance r from the light to the surface.
Finally, the dot product of the incident direction and the surface normal n models how the
Bump Mapped BRDF Rendering 407
Team LRN
incoming light from the source gets distributed over a larger area of the surface (and thus less
arrives per unit area) if the light is at a low angle relative to the surface.
Now we want to evaluate this lighting equation in Direct3D. We need to write a vertex and
pixel shader, plus the setup code, which evaluates a light term in the sum, plus the routine
which uses them to render a mesh. Then we can call this method repeatedly for different lights
with additive blending turned on to get the final result.
Pseudocode for the single light rendering routine:
d3dDevice->SetTexture(0, P_Texture )
d3dDevice->SetTexture(1, Q_Texture )
d3dDevice->SetTexture(2, P_Texture )
if (BRDF has S texture)
d3dDevice->SetTexture(3, S_Texture )
Set Pixel and Vertex Shader Constants
Set Pixel and Vertex Shader
Call DrawPrimitive() to render the mesh
The vertex shader is more interesting. Its job is to compute the per-vertex incoming (to_light)
and outgoing (to_eye) vectors, which will be our texture coordinates. It is reproduced below,
with each operation commented.
vs.1.0
; incoming vertex is [POS, TANGU, TANGV, NRML]
; Transform vertex position to homogenous clip space position
dp4 oPos.x, c[CMODVIEW0], POS
dp4 oPos.y, c[CMODVIEW1], POS
dp4 oPos.z, c[CMODVIEW2], POS
dp4 oPos.w, c[CMODVIEW3], POS
; compute to_light = normalize(light_position - vertex_position)
sub r1, c[LIGHTPOS], POS
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1.xyz, r1, r1.w
; compute to_eye = normalize(camera_position - vertex_position)
sub r2, c[CAMPOS], POS
dp3 r2.w, r2, r2
rsq r2.w, r2.w
mul r2.xyz, r2, r2.w
408 Part 3: Pixel Shader Tricks
Figure 2: The tangent space vectors
used in shading
Team LRN
; compute (normal . to_light) * LCOLOR term.
; LCOLOR is the combined light color, and the texture scaling factor .
dp3 r0, NRML, r1
mul oD0.xyz, r0, c[LCOLOR]
; texcoord0 = outgoing direction in the tangent space
dp3 r5.x, r2, TANGU
dp3 r5.y, r2, TANGV
dp3 r5.z, r2, NRML
mov oT0.xyz, r5.xyz
; texcoord2 = incoming direction in the tangent space
dp3 r6.x, r1, TANGU
dp3 r6.y, r1, TANGV
dp3 r6.z, r1, NRML
mov oT2.xyz, r6.xyz
; texcoord1 = texcoord3 = half-angle vector = outgoing + incoming,
; normalized implicitly by the cube map.
add r5.xyz, r5.xyz, r6.xyz
mov oT1.xyz, r5.xyz
mov oT3.xyz, r5.xyz
Note the way the vertex shader uses the same normal and tangent vectors to span a per-vertex
tangent space as the dot3 bump mapping techniques do. This means you will not have to
generate any additional mesh attributes for this effect if your engine supports simple bump
mapping. As for the constant registers, this shader needs the camera and light position in
model space and a combined light color* value. I left out the quadratic attenuation (it does
not look good given the limited dynamic range of our lights), but it can easily be added by
squaring the inverse light-vertex distance, which gets computed as part of the normalization of
the to_light vector above.
An additional trick not shown here, but one I use in my production code, is to compute the
normal vector from the two tangent vectors with a cross product, instead of passing it explicitly
per-vertex. If you find that your normal is sometimes the negative of this cross product, you
need to additionally encode its sign as the magnitude of one of your tangent vectors, which
will then need renormalization as well.
The pixel shader is straightforward; it has to sample the textures and compute the product:
ps.1.0 ; out = diffuse*tex0(t0)*tex1(t1)*tex2(t2) + tex3(t3)
tex t0
tex t1
tex t2
tex t3
mul r0, v0, t0
mul r0, r0, t1
mul r0, r0, t2
add r0, r0, t3
Of course, you only need the final add if you used the optional additive S term.
Bump Mapped BRDF Rendering 409
Team LRN
Adding the Bumps
Unfortunately, applying a bump map to the BRDF is not as straightforward as one would hope.
The reason is that the cube maps need their texture coordinates, the to_light, to_eye, and
half-angle vectors, in the surfaces tangent space. When there are no bumps, the basis of this
tangent space is equivalent to the interpolated per-vertex basis from our vertex shader in the
previous section. However, in the bumpy case, the tangent space is spanned by the per-pixel
normal (read from a conventional normal map) and two tangent vectors orthogonal to it. Dot3
bump mapping also uses this exact normal, but in that case, it is only needed for a dot product
operation, and the two tangent vectors are not used. We, on the other hand, have to transform
our per-vertex vectors computed in the vertex shader into this space before doing the texture
lookups. Fortunately, there are some corners that can be cut, but first we will make some
changes to our vertex shader:
vs.1.0
; incoming vertex is [POS, TANGU, TANGV, NRML, TEXCOORD]
; Transform vertex position to homogenous clip space position
dp4 oPos.x, c[CMODVIEW0], POS
dp4 oPos.y, c[CMODVIEW1], POS
dp4 oPos.z, c[CMODVIEW2], POS
dp4 oPos.w, c[CMODVIEW3], POS
; compute to_light = normalize(light_position - vertex_position)
sub r1, c[LIGHTPOS], POS
dp3 r1.w, r1, r1
rsq r1.w, r1.w
mul r1.xyz, r1, r1.w
; compute to_eye = normalize(camera_position - vertex_position)
sub r2, c[CAMPOS], POS
dp3 r2.w, r2, r2
rsq r2.w, r2.w
mul r2.xyz, r2, r2.w
; conventional texture coordinate for the bump map
mov oT0.xy, TEXCOORD.xy
; texcoord1 = light direction
dp3 oT1.x, r1, TANGU
dp3 oT1.y, r1, TANGV
dp3 oT1.z, r1, NRML
; texcoord2 = eye direction
dp3 oT2.x, r2, TANGU
dp3 oT2.y, r2, TANGV
dp3 oT2.z, r2, NRML
We made two changes: First, we took out the code to compute the normal dependent diffuse
color. Second, we no longer compute the half-angle sum here. Since the half-angle direction is
the sum of the other two vectors, regardless of which space we are in, we can do the addition in
the pixel shader. This way, we avoid having to transform it per-pixel along with the other two
vectors. Finally, since we now use a bump map, we also have to pass texture coordinates for it
along with each vertex. Of course, the texture bindings in the rendering routine have to be
updated accordingly.
410 Part 3: Pixel Shader Tricks
Team LRN
Now, on to the pixel shader:
def X_AXIS, 1.0f, 0.0f, 0.0f, 1.0f
def Y_AXIS, 0.0f, 1.0f, 0.0f, 1.0f
; PHASE1-TEXTUREADDRESS 1) fetch vertex space normal
texld r4, t0 ; look up the normal in the normal map
texcrd toLight_vs.xyz, t1 ; just grab the other vectors
texcrd toEye_vs.xyz, t2 ; without using a texture map
; PHASE1 - ARITHMETIC 2) create tangent space basis
; tangU_vs = X_AXIS - normal_vs * DOT(X_AXIS,normal_vs)
mad tangU_vs, -normal_vs_bx2, normal_vs_bx2.x, X_AXIS
; same with y axis
mad tangV_vs, -normal_vs_bx2, normal_vs_bx2.y, Y_AXIS
; [tangU_vs, tangV_vs, normal_vs] now approximates an
; orthogonal basis
; transform toEye from the smooth tangent space to the
; bumpy one.
dp3 toEye_ts.x, tangU_vs , toEye_vs
dp3 toEye_ts.y, tangV_vs , toEye_vs
dp3 toEye_ts.z, normal_vs_bx2, toEye_vs
; not enough instructions available to do another
; full 3x3 multiplication: we just perturb one component
mov toLight_ts.xyz, toLight_vs
dp3 toLight_ts.z, normal_vs_bx2, toLight_vs
; compute the half angle vector as usual
add halfAngle_ts.xyz, toLight_ts, toEye_ts
phase
; PHASE2-TEXTUREADDRESS 4) sample cubemaps
texld r0, toLight_ts
texld r1, halfAngle_ts
texld r2, toEye_ts
texld r3, halfAngle_ts
; PHASE2 - ARITHMETIC 5) do blending math
; compute LCOLOR * (toLight_vs . normal_vs)
dp3 normal_vs, toLight_vs, normal_vs_bx2
mul r4, normal_vs, c[LCOLOR]
mul r0, normal_vs, r0
mul r0, r0, r1
mul r0, r0, r2
add r0, r0, r3
Bump Mapped BRDF Rendering 411
Team LRN
What we are trying to do is totally hopeless in ps.1.3 and below and just about fits into the
ps.1.4 spec. In the texture address part of phase 1, we retrieve the two vectors we computed in
the vertex shader (they have since been interpolated across the triangle and may have been
denormalized). We also look up the normal in the normal map. This normal is relative to the
interpolated vertex basis, as are our other two vectors.
In the arithmetic phase, we need to construct the true surfaces tangent space basis. To
approximate that as compactly as possible, I separately orthogonalize the x and y axis to the
normal. Full Gram-Schmidt orthogonalization would take too many instructions.
We immediately make use of this matrix by transforming the to_eye vector with it. We
would like to do the same to the to_light vector, but due to the limited number of instructions
available, we only succeed in transforming the z component. Fortunately, this approximation
looks quite good regardless.
Finally, the half-angle vector is computed as the sum of the two transformed vectors.
Phase 2 of the shader is the same as the one for the smooth case, with the exception that we
also compute the color term using the perturbed normal.
Conclusion
We have presented a brief introduction to BRDFs and shown how the rendering technique of
[McCool] can be extended to bumpy surfaces. This is only scratching the surface of the
untapped possibilities of BRDF rendering. Other directions worth following are the combina-
tion of BRDF effects with color textures and environment maps.
Acknowledgments
Thanks to Jason Mitchell, Christoph Niederberger, and Tim Weyrich for help with different
aspects of this effect.
412 Part 3: Pixel Shader Tricks
Figure 3: Tori showing the bump mapped BRDF effect
Team LRN
References
[McCool01] Michael D. McCool, Jason Ang, and Anis Ahmad, Homomorphic Factorizations
of BRDFs for High-Performance Rendering, SIGGRAPH Proceedings, 2001, pp.
171-178.
[Marschner00] Stephen R. Marschner, Stephen H. Westin, Eric P. F. Lafortune, and Kenneth E.
Torrance, Image-based BRDF Measurement, Applied Optics, Vol. 39, No. 16, 2000.
Bump Mapped BRDF Rendering 413
Team LRN
Real-Time Simulation
and Rendering of
Particle Flows
Daniel Weiskopf and Matthias Hopf
In this article, we describe a pixel shader technique for simulating and rendering a 2D particle
flow in real time. This physics-based approach allows the environment to affect the particles,
leading to more realistically evolving scenes. Typical applications include water surfaces, lava
streams, and waterfalls, etc.
Motivation
Computer games usually try to immerse the player in a realistic-looking environment. In most
cases, you want the environment to be alive there should be plenty of action in the back-
ground not directly related to the players avatar or his or her friends and foes. For the most
challenging effects, hundreds and thousands of moving elements have to be coordinated; these
effects become time-critical features in the graphics engine.
Depending on the kind of action, either vertex-based or pixel-based approaches seem to be
appropriate. Particle systems are a kind of vertex-based approach that is used for various
effects that have to display many loosely coupled simple features and is already state of the art.
Effects that can only be modeled with tightly coupled data and effects that are not closely
related to distinguishable elements are inappropriate for particle systems. Here and in the case
of 2D effects like TV screens, water surfaces, and clouds, animated textures are the way to go.
Especially for water surfaces, a lot of techniques using bump maps (especially environment
bump maps) already exist. The ideas in this article can be combined with these methods to
achieve an optimal effect.
Even modern computer games do not feature a lot of static animated textures like movie
screens or fire effects due to memory requirements. Dynamic animated textures for mirror
effects and portal rendering are computationally expensive and thus rarely used as well. In con-
trast to particle systems, neither one can display changing surfaces that respond to external
stimuli. For example, a water surface showing the flow of the underlying river should react to a
rock that has been thrown into that river.
414
Team LRN
Procedural textures that are created on-the-fly can be used to incorporate a simulation pro-
cess into the surface model. In the case of water, the resulting flow appears to be much more
realistic; the process can create non-repeating texture animations, and it still does not need a lot
of texture memory. The process can react to environmental changes, creating a kind of interac-
tion with the environment unknown before.
Of course, there is a drawback. There always is.
You really dont want to use procedural textures in the form of textures created by the pro-
cessor at run time, as you dont want to create and download a texture ten times a second or
more. This would only be possible for small textures and very simple procedure models.
Besides, we have to do some regular graphics, sound, AI, and user interaction, dont we?
With respect to 2D flows, we would like to offer a solution: We can use the GPU of mod-
ern graphics systems to solve some simple mathematical differential equations in order to
simulate 2D particle transport. One could think of other equations and solving processes for
different materials and effects, but for now, we will restrict ourselves to fluids. Using the GPU
for the brute force mathematics, the processor just needs to work on the interaction model and
set up the solving process.
Using these particle transport simulations, water surfaces, flowing lava, waterfalls, etc.,
can be rendered easily. We will now take a closer look at the underlying theory and how it can
be implemented by using pixel shaders and texture render targets. Please refer to our technical
description in [Weiskopf01] if you are interested in further details on the theoretical back-
ground of particle transport.
Ingredients
These are the benefits you get from physics-based simulation: Most phenomena look natural
and convincing, they usually allow simple modeling by a few and physically meaningful
parameters, and they provide a quite abstract definition of their dynamic behavior.
If we want to take a closer look at the simulation of moving fluid surfaces, the essential
questions boil down to:
n
What kind of motion is it? How do objects move?
n
What objects are involved? What moves?
The answer to the first question is: The velocity of the flow governs the motion of objects
within this flow. Note that velocity is a vector it has a length (a speed) and a direction.
Moreover, the velocity doesnt need to be constant across the flow. Rather, imagine a distinct
velocity vector attached to each single point of the flow. See Figure 1 for a schematic sketch of
such a velocity distribution.
Real-Time Simulation and Rendering of Particle Flows 415
Team LRN
The answer to the second question is:
We are dealing with particles. These parti-
cles are negligible in size and mass (i.e.,
they are massless point-like objects).
Whats important is that particles carry fur-
ther information such as color, temperature,
and pressure. We can attach these or other
properties to the particles according to our
application. A particle keeps its attached
properties along its way through the flow,
making it easy to follow, or trace, the
particle.
Particles are massless objects and cant
withstand changes in velocity caused by an
outside force. Therefore, a particles veloc-
ity is identical to the velocity of the flow
that surrounds the particle. This provides a
link between the two main ingredients: the
particles and the velocities in the flow.
Now imagine that a large number of particles is spread across the surface of the fluid. The
particles are stitched to the moving surface and are thus a direct means of showing the motion
of the flow.
How Does a Single Particle Move?
Lets first consider the motion of a single particle. If you know how to trace one particle, you
know how to follow all particles.
As a particles velocity is identical to the velocity of the flow, we can directly relate a par-
ticles motion to the behavior of the flow. Lets have a look at the definition of velocity v:
Note that velocity is a vector. In the current situation, we consider a two-dimensional flow.
Therefore, velocity is a vector with two components (i.e., v = (v
x
, v
y
)). Vector r describes a
position and has two components as well (i.e., r = (x, y)). The third quantity involved is time t.
The terms dr and dt are so-called infinitesimal quantities. Generally speaking, dr and dt are
very tiny changes in position and time, respectively. As a good approximation, we can replace
infinitesimal changes by small but finite changes r and t. This leads to the relationship:
Rewriting the above equation in a slightly different form yields:
416 Part 3: Pixel Shader Tricks
Figure 1: Distribution of velocities in a flow. The
arrows show the different speeds and directions of
motion at various points in the flow.
Team LRN
The same expression in explicit component-wise
notation:
See Figure 2 for a graphical representation of this equa-
tion. A particle moves from a starting position r
0
to a
subsequent position r
1
+ r within a time span t. The
direction of the connection between the initial position
and the subsequent position is, of course, identical to the
direction of the particles velocity. Note that the length
of the displacement is the length of the velocity multi-
plied by the time span. How far we move in one time
step depends on the time scale and the velocity.
So far, we know how a particle moves from one position to a subsequent position during a
small time step, but what happens for an extended time span? First, split this longer time span
in small time intervals t, which usually are of equal length. Then apply the above displace-
ment scheme for each single time step. Figure 3 shows this iterative process schematically.
Starting at the initial position r
0
, just add the displacement t v
0
to obtain the subsequent posi-
tion r
1
. In the next step, add the displacement
t v
1
. As the velocity may not be constant
within the flow, you have to consider a differ-
ent velocity at each step. If we perform this
process step after step, we will finally obtain
the complete motion of the particle.
This numerical method is called Euler
integration. Please consult one of the numer-
ous textbooks on numerical mathematics, such
as [Press94], if you are interested in more
background information on this or other
related numerical techniques.
Basic Texture Advection: How Do a Bunch
of Particles Move?
Since we would like to visualize a moving fluid surface by means of many different particles,
we need to consider the motion of a large number of particles. We show how we can exploit
the very fast texture and rasterization units offered by graphics chips to speed up tracing many
particles.
Although the idea of point-like particles is conceptually beneficial, it is not directly appli-
cable to graphics hardware. First, we have to transform point-like particles into a texture-based
representation by means of sampling. Figure 4 shows such a sampling process. In this example,
a grid cell of the texture takes the color of the particle contained in the cell (i.e., nearest neigh-
bor sampling). If more than one particle is attached, the color of the cell is the sum of the
Real-Time Simulation and Rendering of Particle Flows 417
Figure 3: Motion of a particle for an extended
time span
Figure 2: Relationship between veloc-
ity v and spatial displacement ( x,
y)
Team LRN
particles colors. Since this texture holds all information on particles, we call it a particle
texture.
After we spread all particles over the texture, we trace them by mapping the initial particle
texture to the particle texture of the following time step. This approach is generally called for-
ward mapping. Figure 5 illustrates the transformation of the particle texture for one time step.
The arrows in the left part of the figure indicate the displacement of each particle, depending
on the velocity of the underlying flow and on the step size t. The color information in each
single cell of the old texture is transferred to another cell in the new texture, depending on the
local displacement vector.
Unfortunately, this basic idea of texture advection suffers from two important problems. First,
if the velocity or the step size is too small, particles may stay in a cell forever, even though, in
fact, they should move slowly. This is caused by the finite size of a grid cell; by definition, we
assume that particles are located at the center of a cell. If a displacement vector is too short to
push a particle out of the cell, the new position of the particle lies inside the same cell. In the
following step, we repeat the same procedure once again. We could iterate this process over
and over without shifting the particle out of the cell.
The second problem is the possibility of gaps in the particle texture. If, for example, the
flow diverges in some areas, neighboring cells might move into different directions in such a
way that intermediate cells in the new particle texture dont receive any particles from the pre-
vious time step. These blank cells cause unpleasant visual artifacts. In general, gaps are an
issue in forward mapping approaches.
We can overcome these two problems by replacing forward mapping by backward map-
ping. The basic idea remains the same: Just replace the mapping from time t to t + t by a
time-reversed mapping from t back to t t. Figure 6 demonstrates backward mapping.
418 Part 3: Pixel Shader Tricks
Figure 5: Propagating
a particle texture one
step forward in time
Figure 4: Sampling of
point particles on a
regular grid
Team LRN
For the sake of simplicity, let us consider only a single cell in the right part of the figure. By
inverting the direction of the displacement vector, we obtain the position from where this cell
has originated:
Then, we use this position to determine the color of the original particle. As the previous posi-
tion usually is not identical to the center of a cell, we apply a bilinear lookup: The actual color
depends on the four surrounding cells. Finally, the complete texture advection approach con-
sists of backward mappings for all cells in the new particle texture.
With this backward mapping approach, we overcome the two problems of forward map-
ping. First, bilinear interpolation guarantees that even small displacements affect the particle
texture of the following time step. Secondly, all cells of the new particle texture receive new
color values (i.e., no gaps can occur).
Based on the previous considerations, the graphics hardware-based approach to texture
advection is as follows. The particle texture is represented by a standard 2D texture. The
number of components in that texture depends on the type of application; normally, we use
RGB values. The velocity field is stored in another 2D texture which has two components for
(v
x
, v
y
).
The advection algorithm uses a dependent texture lookup in order to shift the particles
along the velocity field. Offset textures are part of hardware-supported pixel operations. (In
DirectX, offset textures are called texbem for bump environment-map transform.) See Figure 7
for a schematic diagram of a dependent texture lookup via an offset texture.
The offset texture functionality transforms signed components (du, dv) of a previous tex-
ture unit by a 2 2 matrix and then uses the result to offset the texture coordinates of the
current stage. This process contains two texture fetch operations. The first step comprises a
lookup in an offset texture based on texture coordinates (u
0
, v
0
). An offset texture is a signed
2D texture with two components describing the displacement of coordinates in the successive
texture fetch operation. In the second step, a dependent texture lookup in a standard RGB tex-
ture is performed. Here, another pair of texture coordinates (u
1
, v
1
) is modified by the
previously fetched offset values. We will discuss the scaling by a factor mag later in the section
titled How Can Particles Drop Out?
Real-Time Simulation and Rendering of Particle Flows 419
Figure 6: Texture
advection backward in
time
Team LRN
Based on offset textures, the pseudocode of the texture advection algorithm is:
// Texture tex1 holds information on particles
initialize tex1 with initial_particle_colors
// Texture tex0 is a signed texture holding
// the velocities
initialize tex0 with velocities
// Set initial time t to zero
t = 0
// Main simulation loop
while(simulation_running) {
// Load velocities as offset texture
load_offset_texture tex0
// Prepare dependent texture lookup
// in particle texture tex1
dependent_texture_look_up in tex1
// Compute new particle information,
// which is written into the intermediate
// texture tex2
draw_textured_rectangle into tex2
// Update particle texture tex1 with
// new data
swap tex1 and tex2
// Draw the actual part of the scene
draw_textured_fluid_surface into framebuffer
// Update time
t = t + delta_t
}
During the advection process, the particles tend to bleed out, and sharp edges get smoother
with every time step. This diffusion effect is intrinsic to our approach using discretized particle
positions and thus cant be removed completely. You can use graphics hardware-based image
sharpening filters or so-called morphological operators [Hopf00] to nearly compensate for this
effect, but that is beyond the scope of this article.
420 Part 3: Pixel Shader Tricks
Figure 7: Dependent
texture lookup via off-
set texture
Team LRN
Inflow: Where are the Particles Born?
The particles we are now able to trace in the texture have to come from somewhere. We call
the incoming particles the inflow. Inflowing particles may be dropped into the texture at any
time and any position if applicable. For instance, a leaf that falls into a river can now be pro-
jected onto the flow and rendered into the particle texture. However, in most cases, you do not
want the particles to suddenly pop up in the middle of the flow.
Another possibility is to define an initial texture and leave it unchanged during run time.
Texture coordinate wrapping prevents particles from flowing out of the texture at one of the
edges. Instead, they are inserted on
the opposing edge as you can see in
Figure 8. You will realize soon that
the texture gets blurry very fast,
due to the diffusion effect, until
you can no longer distinguish any
features.
When we dont use texture
wrap mode, texture elements near
the rendered edges of the flow reg-
ularly address elements outside the
texture boundaries during the
advection step. As no texture is
defined there, these texels get
black, introducing very visible arti-
facts to the scene. However, when
the used texture is a bit larger than
the displayed portion, we can insert
new particle blobs at the edges of
the texture. During the advection,
these new particles are sucked into
the visible portion of the texture
and retain the particle density.
When the underlying flow has a
preferred flowing direction, we can remove or even cancel some of the non-displayed edges
and concentrate on one or two edges as shown in Figure 9. Please keep in mind that the
off-screen part of the texture must be large enough so that no off-boundary texels pop into the
displayed part in a single time step. Additionally, they should be sharp-edged in the usual case,
as they will smear out a bit during the flow due to the diffusion effect.
Real-Time Simulation and Rendering of Particle Flows 421
Figure 8: When using texture coordinate wrapping, an end-
less texture can be simulated and no lookups refer to areas
outside the texture.
Team LRN
The design of the particle blobs, the used color palette, and the insertion pattern are as critical
for achieving good-looking images as the flow pattern itself. Additionally, it largely depends
on the kind of flow you want to render. For flows like lava streams, an interesting approach
could be to completely fill the edge with a partially randomized flow color and add a few crack
blobs of a completely different color afterward. For using the particle texture as a bump map-
ping texture (e.g., for water surfaces), the inserted blobs have to be much smoother. Always
remember that it can be a good idea to add blobs inside the rendered part of the texture when
something in the virtual environment interacts with the flow.
In the pseudocode of the texture advection algorithm, inflow is added just after the new
particle information has been written to the target texture:
...
// Main simulation loop
while(simulation_running) {
...
// Compute new particle information,
// which is written into the intermediate
// texture tex2
draw_textured_rectangle into tex2
+ // Add regular inflow
+ render_offscreen_inflow into tex2
+
+ // Add inflow that occurs due to
+ // interactions with the environment
+ render_interaction_inflow into tex2
// Update particle texture tex1 with
// new data
422 Part 3: Pixel Shader Tricks
Figure 9: New particles are inserted in the off-screen part of the texture that is
addressed in the next time step. The flow on the right side has a preferred
direction; thus, only one edge has to be considered.
Team LRN
swap tex1 and tex2
...
}
How Can Particles Drop Out?
Particles that enter the flow have to exit somewhere. We call the exiting particles the outflow.
Of course, we do not trace particles individually but rather their concentration or features, so
we cant really say that we have particles exiting the flow. However, the contents of texture
cells that are not referred to by the offset texture addressing scheme do not contribute to the
new texture contents. In this sense, the non-referred texels contain the outflow.
For flows with a preferred direction, natural outflow happens at the edge opposed to the
inflow edge. Figure 9 depicts this situation. In some cases, with slow flow or additional blobs
added to the flow, this may not be enough. For circular flows or textures using texture coordi-
nate wrap mode, no natural outflow exists, and we have to apply other means to get the range
of flowing particles into a steady state.
Similar to the inflow situation, you may want to remove particles at dedicated positions in
the flow, for instance, when something is interacting with the flow. You could either subtract or
weight a small part of the texture to remove or reduce the particles in this area. This can be
done either explicitly after the advection process is completed or implicitly by choosing a
luminance weighting at the end of the offset texture process. In the section on basic texture
advection, we indicated that a scaling of fetched texels is possible. Each component of the final
RGB triplet can be multiplied by a factor mag, which is part of the offset texture as an inde-
pendent, third component. By choosing scaling values smaller than 1.0, you can specify
arbitrarily shaped regions of outflow.
Implementation
Our implementation is based on DirectX 8.1 and ps.1.0. Lets start with the actual pixel shader
program:
// Pixel shader 1.0
ps.1.0
// Load offset texture in stage 0
tex t0
// Dependent texture look-up in the particle
// texture t1. Texture coordinates are
// modified by the previous offset texture t0
texbem t1,t0
// Final color
mov r0, t1
First, we load the velocity field into texture stage 0. The velocity field t0 is a 2D texture whose
components may have positive and negative values. A typical texture format is the signed for-
mat D3DFMT_V8U8. Then we perform a dependent texture lookup in the particle texture of
the previous time step. The particle texture t1 is a standard 2D RGB texture. The name texbem
stands for bump environment-map transform, indicating that its standard usage is in per-pixel
bump mapping applications.
Real-Time Simulation and Rendering of Particle Flows 423
Team LRN
Before we can start the pixel shader program, we have to initialize the textures. The veloc-
ity field is stored as a signed 2D texture and can be defined by:
device->CreateTexture( width, height, 1, 0,
D3DFMT_V8U8,
D3DPOOL_MANAGED,
&texture0);
The offset texture, in this case, is limited to 8 bits per component (i.e., the velocity is coded as
signed bytes). Its a good idea to scale the actual velocity data to the maximum value range
(from 128 to +127) in order to achieve the highest numerical accuracy. We can absorb this
scaling in a later step, as described shortly.
The initial particle texture is a standard RGB texture. For example, we can use an RGB
texture with 8 bits per color channel:
device->CreateTexture( width, height, 1, 0,
D3DFMT_X8R8G8B8,
D3DPOOL_MANAGED,
&texture1);
The dependent texture lookup in texture stage 1 transforms the offset values (du, dv) before
adding them to the texture coordinates of stage 1. Set the transformation matrix to a multiple of
the identity matrix in order to uniformly scale the offset values:
float scale = -step_size;
device->SetTextureStageState
(1, D3DTSS_BUMPENVMAT00, F2DW(scale));
device->SetTextureStageState
(1, D3DTSS_BUMPENVMAT01, F2DW(0.0f) );
device->SetTextureStageState
(1, D3DTSS_BUMPENVMAT10, F2DW(0.0f) );
device->SetTextureStageState
(1, D3DTSS_BUMPENVMAT11, F2DW(scale));
The scaling takes care of the factor t in backward particle tracing. In this way, the step size
can be changed without redefining the values of the velocity texture. In addition, scale absorbs
the scaling of the original velocity values to the range of signed bytes, as introduced above.
In order to allow for outflow regions, you simply have to replace the texbem pixel shader
operator with texbeml (which stands for bump environment-map transform with luminance
correction) and use a three-component offset texture. For example, the texture format
D3DFMT_V8U8 could be substituted by the signed format D3DFMT_W8Q8V8U8 or the
mixed formats D3DFMT_X8L8V8U8 or D3DFMT_L6V5U5. The magnification coefficient L
of the texture is multiplied with the texel color directly after lookup and can be set to values
smaller than 1.0 for outflow regions.
The new particle texture of the current time step is the old particle texture of the subse-
quent time step. Therefore, we directly render the new particles into a texture by setting:
device->CreateTexture( width, height, 1,
D3DUSAGE_RENDERTARGET,
D3DFMT_X8R8G8B8,
D3DPOOL_DEFAULT,
&texture2));
424 Part 3: Pixel Shader Tricks
Team LRN
and
renderBuffer->GetSurfaceLevel(0, &surface);
device->SetRenderTarget(surface, NULL);
Since the particle texture for the new time step is also a rectangular grid of the same size, as in
the previous time step, we just have to render such a rectangle into the render target in order to
obtain the new particles.
So far, we have accomplished the simulation of particle flow for one time step. After each
time step, we swap the particle textures texture1 and texture2, so that texture1 holds the current
state of the particle texture. (Note that both textures have to be defined as
D3DUSAGE_RENDERTARGET, slightly changing the original definition of texture1.) In
order to finally draw the surface of the fluid, we render the polygonal representation of the sur-
face to the frame buffer, with the particle texture texture1 used as RGB texture.
Example code for simulating and rendering texture advection is contained on the compan-
ion CD.
Summary
We have described a method to simulate and render texture transport on a moving surface (for
example, the surface of a liquid). This approach allows for dynamic interaction with the envi-
ronment and on-the-fly inclusion of outside effects on the flow. For practical purposes, the
design of both the flow pattern and the inflow procedures is critical. Particle advection can be
used for enhancing liquid surfaces such as water surfaces, lava streams, or waterfalls.
References
[Hopf00] M. Hopf and T. Ertl, Accelerating morphological analysis with graphics hardware,
Workshop on Vision, Modeling, and Visualization VMV 2000 (infix, 2000), pp. 337-345.
[Press94] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical
Recipes in C, 2nd Ed. (Cambridge University Press, 1994).
[Weiskopf01] D. Weiskopf, M. Hopf, and T. Ertl, Hardware-accelerated visualization of
time-varying 2D and 3D vector fields by texture advection via programmable per-pixel
operations, Workshop on Vision, Modeling, and Visualization VMV 2001 (infix, 2001),
pp. 439-446.
Real-Time Simulation and Rendering of Particle Flows 425
Team LRN
426
3D Textures and Pixel Shaders
by Evan Hart
With the introduction of programmable pixel pipelines, graphics fea-
tures that were once somewhat obscure are now finding additional
uses. One example of this is 3D, or volumetric, textures. While 3D tex-
tures have been available for several years, few real applications have
been created to take advantage of them. The reality is that real-time
shading is still largely a discipline of pasting multiple 2D images onto
3D surfaces. In this article, we will define 3D textures and then present
a variety of shaders which use them to achieve effects that are either
awkward or impossible with 1D or 2D textures.
Team LRN
427
Truly Volumetric Effects
by Martin Kraus
This article presents several examples of ps.1.4 programming for volu-
metric effects, ranging from quite basic techniques to recent research
results. These include volume visualization with transfer functions,
animation of volume graphics by blending between different volume
data sets, animation of transfer functions, and a recently developed
rendering technique for real-time interactive, high-quality volume
graphics, called pre-integrated volume rendering.
Team LRN
3D Textures and Pixel
Shaders
Evan Hart
Introduction
With the introduction of programmable pixel pipelines, graphics features that were once some-
what obscure are now finding additional uses. One example of this is 3D or volumetric
textures. While 3D textures have been in high-end graphics hardware for about a decade and
appeared in consumer hardware nearly two years ago, few real applications have been created
to take advantage of them. The reality is that real-time shading is still largely a discipline of
pasting multiple 2D images onto 3D surfaces. Programmable pixel shaders have enabled a
variety of unique algorithms, and 3D textures are a concise way to implement techniques that
might be awkward or impossible with just 1D or 2D textures. In this article, we will define 3D
textures and then present a variety of shaders which use them to achieve effects that are either
awkward or impossible with 1D or 2D textures.
3D Textures
Since 3D textures have appeared relatively recently in consumer-level hardware, many people
are unfamiliar with exactly what they are. The simplest way to think of a 3D texture is as a
stack of 2D textures. There are
multiple conventions for referring
to the coordinates used to index a
3D texture; for this article, we will
use s, t, r, and q. These are the
width, height, depth, and projec-
tive coordinates, respectively. This
is different from the terminology
used in DirectX documentation,
but it avoids confusion as to
whether w is the depth component
or the projective component of a
texture.
428
Figure 1: Relationship between 2D and 3D textures
Team LRN
Filtering
With the addition of an extra dimension, the filtering nomenclature changes a bit.
For a 3D texture, simple linear filtering of a single level of the mip map chain is referred to
as trilinear filtering, as it performs linear filtering along all three axes of the texture, sampling
colors from eight different texels. Most people automatically assume that trilinear filtering
means that multiple mip maps are being blended together, but this is not the case with 3D tex-
tures. With 3D textures, this case is referred to as quadrilinear filtering, as it adds a linear filter
along the fourth dimension of the volumetric mip maps, sampling from 16 different texels. As
this change in nomenclature suggests, the filtering doubles in expense when moving from
two-dimensional textures to three-dimensional textures. Additionally, anisotropic filtering
becomes significantly more complex. With 2D textures, anisotropic filtering is thought of as a
filter along a line across the texture with a width of two and length up to the maximum degree
of anisotropy. This means that an anisotropic filter with a maximum level of anisotropy of 8
requires 16 texels. In three dimensions, this increases to a slab with width two extending in the
other two dimensions up to the maximum degree of anisotropy. Now, our worst case for a max-
imum level of anisotropy becomes 8*8*2 = 128 texels. This is 32 times the cost of the bilinear
fetch of a 2D texture that most programmers consider the basic unit of texture performance.
It is easily seen that anisotropic filtering of a 3D texture is something to avoid, unless the pro-
grammer has a very specific reason. Even if there is good justification, the programmer must
understand that this sort of filtering may be impossible as current accelerators only handle a
single direction of anisotropy at a time.
Storage
The only thing more expensive than filtering a 3D texture is storing it in memory.
Three-dimensional textures grow extremely rapidly with increases in resolution. For example,
a 32-bit 256256256 texture is 64 MB in size. While this is a significant hurdle to using vol-
ume textures, it can often be addressed with a few simple tricks. First, the number of channels
used in a 3D texture should be kept to a minimum. If the data to be stored in the 3D texture
represents a scalar quantity like density or intensity, it should be stored as a luminance or alpha
texture rather than RGBA. Next, if the artifacts of DXTC style compression are not objection-
able, the 3D texture may be compressed to conserve memory. Finally, mirroring and repeating
offer a great way to compress 3D textures. Many of the examples in the following sections use
3D Textures and Pixel Shaders 429
Figure 2: Anisotropic filtering in two
and three dimensions
Team LRN
textures that repeat or are symmetric. By using the wrapping and mirroring modes for the dif-
ferent dimensions, 2x to 8x compression can generally be achieved.
Applications
Despite the high storage and filtering costs, 3D textures do make a lot of tasks easier and can
frequently be more efficient than using 2D textures to implement the same algorithm. Also, the
two primary difficulties previously mentioned will become smaller issues over time. Memory
on graphics accelerators has been doubling almost yearly, and the cost of the filtering operation
is only as expensive as performing two texture fetches of the same quality. Modern cards that
support 1.4 pixel shaders are capable of performing up to 12 2D fetches (which can each be
trilinear, anisotropic, etc.) in a single pass.
Function of Three Independent Variables
The first and most obvious application for 3D textures in shaders is as lookup tables to provide
piecewise linear approximations to arbitrary functions of three variables. Many of the follow-
ing sections show specific examples of functions of three dimensions, but it is important to
remember the general concept rather than particular examples. One simple example would be
an implementation of the per-pixel calculation of Phong-style specular highlights. Note that
this is an alternative to the technique proposed in A Non-Integer Power Function on the Pixel
Shader.
Since the half-angle (H) and reflection (R) vectors dont interpolate linearly across the sur-
face of polygons, they must be recomputed per-pixel to obtain the best results. This leads to the
creation of a vector of non-unit length, as graphics hardware is not yet able to perform a
per-pixel square root. Naturally, the equation for specular highlights requires its inputs to be of
unit length. The problem is that LR or NH, whichever you prefer, is not just cos( ), but cos( )
* |R| or cos( ) * |H|. Clearly, this problem can be corrected by performing a normalizing divide
prior to raising cos( ) to the kth power. Now, the function that needs to be evaluated is one of
three variables: {NH (cos( ) * |H|), |H|, k} or {RL (cos( ) * |R|), |R|, k}. The equation actu-
ally implemented is shown below. The most important item to note about this particular
algorithm is that it can be implemented much more efficiently as a projected 2D texture since
the first operation is just a divide, but this example is good for illustrative purposes.
N H
H H
k
o

where N H o 0 if N H 0
Pixel shader:
ps.1.4
texld r0.rgb, t0.xyz //get normal from a texture lookup
texld r1.rgb, t1.xyz //get light vector in same way
texld r2.rgb, t2.xyz //get view vector in same way
texcrd r3.rgb, t3.xyz //pass in spec power in t3.x
//compute H as V+L/2
add_d2 r1.rgb, r1_bx2, r2_bx2
430 Part 4: Using 3D Textures with Shaders
Team LRN
//compute N.H and store in r0.r
dot3 r0.r, r0_bx2, r1
//compute H.H and store in r0.g
dot3 r0.g, r1, r1
//copy in k
mov r0.b, r3.r
phase
texld r3, r0.rgb //fetch the 3D texture to compute specular
//move it to the output
mov_sat r0, r3
Noise and Procedural Texturing
Noise is a particular type of function texture that is extremely useful in pixel shading. The
noise that is most interesting to us here is not white noise of random numbers, but Perlin noise
that has smoothly varying characteristics [Perlin85]. Noise of this sort has been used for years
in production quality 3D graphics for texture synthesis. It allows the creation of non-repeating
fractal patterns like those found in wood, marble, or granite. This is especially interesting in
that an effectively infinite texture can be created from finite texture memory. Other uses of 2D
noise textures are discussed in Texture Perturbation Effects.
The most classic example for using a 3D noise function is to create turbulence. Turbulence
is an accumulation of multiple levels (octaves) of noise for the purpose of representing a turbu-
lent flow. A great example of this is the appearance of marble. The veins flow through the
material in a turbulent manner. The example we will use is an adaptation for real time of the
Blue Marble shader in [Upstill89]. In this example, the shader uses six textures to create the
effect. The textures are five occurrences of a noise map and a one-dimensional color table.
(Note that this shader will work even better with more noise textures six is just the limit for
DirectX 8.1s 1.4 pixel shaders.) The noise textures are scaled and summed to create the turbu-
lence, and the value of the turbulence is then used to index a 1D texture via a dependent read to
map it to a color. The texture coordinates used to index into the noise maps are all procedurally
generated in the same manner, and they are just different scales of each other.
Pixel shader code:
ps.1.4
//fetch noise maps
texld r0, t0.xyz
texld r1, t1.xyz
texld r2, t2.xyz
texld r3, t3.xyz
texld r4, t4.xyz
//accumulate multiple octaves of noise
// this sets r0 equal to:
// 1/2r0 + 1/4r1 + 1/8r2 + 1/16r3 + 1/32r4
add_d2 r3, r3_x2, r4
add_d2 r3, r2_x2, r3
add_d2 r3, r1_x2, r3
3D Textures and Pixel Shaders 431
Team LRN
add_d4 r0, r0_x2, r3
phase
//fetch 1-D texture to remap
texld r5, r0.xyz
mov r0, r5
Realistically, the interactive 3D graphics field is just now on the edge of getting procedural
noise textures to work. Noise textures tend to require several octaves of noise and dependent
fetches just to implement something as simple as marble. This ignores the application of light-
ing and other effects one would also need to compute in the shader. As a result, it would be
impractical to ship a game that relied upon noise textures of this sort today, but it still certainly
applies as a technique to add that little bit of extra varying detail to help eliminate the repeti-
tive look in many of todays interactive 3D applications. Finally, even if this technique is not
immediately applicable, graphics hardware is evolving at such a pace that we expect this tech-
nique will be very common in a few years.
Attenuation and Distance Measurement
Another procedural trick that 3D textures are useful for is attenuation. The attenuation can cor-
respond to falloff proportional to the distance from a light source, or simply some distance
measurement trick intended to create a cool effect. Typically, a distance measurement like this
is implemented with a combination of pixel and vertex shaders, where the true distance is com-
puted per-vertex and interpolated or a vector representing the distance is interpolated with its
length being computed per-pixel. The deficiency in these methods is that they incorrectly lin-
early interpolate non-linear quantities or that they potentially force some of the computation to
occur in lower precision, leading to poor quality. Additionally, these algorithmic solutions are
only applicable to calculations based on simple primitives such as points, lines, and planes.
By using a 3D texture, the programmer is effectively embedding the object in a distance or
attenuation field. Each element in the field can have arbitrary complexity in its computation
since it is generated off-line. This allows the field to represent a distance from an arbitrary
shape. On the downside, using a 3D texture can be reasonably expensive from a memory
432 Part 4: Using 3D Textures with Shaders
Figure 3: Marble shader using eight
octaves of noise instead of five
Team LRN
standpoint, and it relies on a piecewise linear approximation to the function from the filtering.
The toughest portion of using a 3D texture to perform this sort of operation is defining the
space used for the texture and the transform necessary to get the object into the space. Below is
a general outline of the algorithm:
1. Define a texture space containing the area of interest.
2. For each texel in the space, compute the attenuation.
3. Apply the 3D attenuation texture to the object with appropriate texture coordinate
generation.
The first step involves analyzing the data and understanding its symmetry and extents. If the
data is symmetric along one of its axes, then one of the mirrored texturing modes, such as
D3DTADDRESS_MIRRORONCE, can be used to double the resolution (per axis of symme-
try) for a given amount of data. Additionally, an understanding of the extents of the data is
necessary to optimally utilize the texture memory. If a function falls to a steady state, such as
zero along an axis, then the texture address mode can be set to clamp to represent that portion
of the data. Once the user has fit the data into a space that is efficient at representing the func-
tion, the function must be calculated for each texel. Below are a pair of pseudocode examples
showing how to fill an attenuation field for a point. The first is nave and computes all the
attenuation values in the shape of a sphere; the second takes advantage of the symmetry of the
sphere and only computes one octant. This can be done since the values all mirror against any
plane through the center of the volume.
Complete sphere:
#define VOLUME_SIZE 32
unsigned char volume[VOLUME_SIZE][VOLUME_SIZE][VOLUME_SIZE];
float x, y, z;
float dist;
//walk the volume filling in the attenuation function,
// where the center is considered (0,0,0) and the
// edges of the volume are one unit away in parametric
// space (volume ranges from 1 to 1 in all dimensions
x = -1.0f + 1.0f/32.0f; //sample at texel center
for (int ii=0; ii<VOLUME_SIZE; ii++, x += 1.0f/16.0f)
{
y = -1.0f + 1.0f/32.0f;
for (int jj=0; jj<VOLUME_SIZE; jj++, y += 1.0f/16.0f)
{
z = -1.0f + 1.0f/32.0f;
for (int kk=0; kk<VOLUME_SIZE; kk++, y += 1.0f/16.0f)
{
//compute distance squared
dist = x*x + y*y + z*z;
//compute the falloff and put it into the volume
if (dist > 1.0f)
{
//outside the cutoff
volume[ii][jj][kk] = 0;
}
else
{
3D Textures and Pixel Shaders 433
Team LRN
//inside the cutoff
volume[ii][jj][kk] = (unsigned char)
( 255.0f * (1.0f dist));
}
}
}
}
Sphere octant, used with the D3DTADDRESS_MIRRORONCE texture address mode:
#define VOLUME_SIZE 16
unsigned char volume[VOLUME_SIZE][VOLUME_SIZE][VOLUME_SIZE];
float x, y, z;
float dist;
//walk the volume filling in the attenuation function,
// where the center of the volume is really (0,0,0) in
// texture space.
x = 1.0f/32.0f;
for (int ii=0; ii<VOLUME_SIZE; ii++, x += 1.0f/16.0f)
{
y = 1.0f/32.0f;
for (int jj=0; jj<VOLUME_SIZE; jj++, y += 1.0f/16.0f)
{
z = 1.0f/32.0f;
for (int kk=0; kk<VOLUME_SIZE; kk++, y += 1.0f/16.0f)
{
//compute distance squared
dist = x*x + y*y + z*z;
//compute the falloff and put it into the volume
if (dist > 1.0f)
{
//outside the cutoff
volume[ii][jj][kk] = 0;
}
else
{
//inside the cutoff
volume[ii][jj][kk] = (unsigned char)
( 255.0f * (1.0f dist));
}
}
}
}
With the volume, the next step is to generate the coordinates in the space of the texture. This
can be done with either a vertex shader or the texture transforms available in the fixed-function
vertex processing. The basic algorithm involves two transformations. The first translates the
coordinate space to be centered at the volume, rotates it to be properly aligned, then scales it so
that the area effected by the volume falls into a cube ranging from 1 to 1. The second trans-
formation is dependent on the symmetry being exploited. In the case of the sphere octant, no
additional transformation is needed. With the full sphere, the transformation needs to map the
entire 1 to 1 range to the 0 to 1 range of the texture. This is done by scaling the coordinates by
one half and translating them by one half. All these transformations can be concatenated into a
single matrix for efficiency. The only operation left to perform is the application of the dis-
tance or attenuation to the calculations being performed in the pixel shader.
434 Part 4: Using 3D Textures with Shaders
Team LRN
Representation of Amorphous Volume
One of the original uses for 3D textures was in the visualization of scientific and medical data
[Drebin88]. Direct volume visualization of 3D datasets using 2D textures has been in use for
some time, but it requires three sets of textures, one for each of the axes of the data, and is
unable to address filtering between slices without the use of multitexture. With 3D textures, a
single texture can directly represent the dataset.
Although this article does not focus on scientific visualization, the implementation of this
technique is useful in creating and debugging other effects using volume textures. Using this
technique borrowed from medical volume visualization is invaluable in debugging the code
that a developer might use to generate a 3D distance attenuation texture, for example. It allows
the volume to be inspected for correctness, and the importance of this should not be underesti-
mated, as presently an art tool for creating and manipulating volumes does not exist.
Additionally, this topic is of great interest since a large amount of research is presently ongoing
in the area of using pixel shaders to improve volume rendering. For more information on volu-
metric rendering, see Truly Volumetric Effects.
Application
These general algorithms and techniques for 3D textures are great, but the really interesting
part is the concrete applications. The following sections will discuss a few applications of the
techniques described previously in this article. They range from potentially interesting spot
effects to general improvements on local illumination.
Volumetric Fog and Other Atmospherics
One of the first problems that jumps into many programmers minds as a candidate for volume
texture is the rendering of more complex atmospherics. So far, no one has been able to get
good variable density fog and smoke wafting through a scene, but many have thought of vol-
ume textures as a potential solution. The reality is that many of the rendering techniques that
can produce the images at a good quality level burn enormous amounts of fill-rate. Imagine
having to render an extra 50 to 100 passes over the majority of a 1024 x 768 screen to produce
a nice fog effect. This is not going to happen quite yet. Naturally, hardware is accelerating its
capabilities toward this sort of goal, and with the addition of a few recently discovered
improvements, this sort of effect might be acceptable.
The basic algorithm driving the rendering of the atmospherics is based on the volume visu-
alization shaders discussed earlier. The primary change is an optimization technique presented
at the SIGGRAPH/Eurographics Workshop on Graphics Hardware in 2001 by Klaus Engel
(http://wwwvis.informatik.uni-stuttgart.de/~engel/). The idea is that the accumulation of slices
of the texture map is attempting to perform an integration operation. To speed the process up,
the volume can be rendered in slabs instead. By having each slice sample the volume twice (at
the front and rear of the slab), the algorithm approximates an integral as if the density varied
linearly between those two values. This results in the slice taking the place of several slices
with the same quality. This technique is better described in the following article, Truly Volu-
metric Effects.
3D Textures and Pixel Shaders 435
Team LRN
The first step in the algorithm is to draw the scene as normal, filling in the color values
that the fog will attenuate and the depth values that will clip the slices appropriately. The atmo-
spheric is next added to the scene by rendering screen-aligned slices with the shader described
in the following article. When rendering something like fog that covers most of the view, this
means generating full-screen quads from the present viewpoint. The quads must be rendered
back to front to allow proper compositing using the alpha blender. The downside to the algo-
rithm is in the coarseness of the planes intersecting with the scene. This will lead to the last
section of fog before a surface to be dropped out. In general, this algorithm is best for a fog
volume with relatively little detail, but that tends to be true of most fog volumes.
Light Falloff
A more attainable application of 3D textures and shaders on todays hardware is volumetric
light attenuation. In its simplest form, this becomes the standard light-map algorithm that
everyone knows. In more complex forms, it allows the implementation of per-pixel lighting
derived from something as arbitrary as a fluorescent tube or a light saber.
In simple cases, the idea is to simply use the 3D texture as a ball of light that is modulated
with the base environment to provide a simple lighting effect. Uses for this include the classic
rocket flying down the dark hallway to kill some ugly monster shader. Ideally, that rocket
should emit light from the flame. The light should roughly deteriorate as a sphere centered
around the flame. This obviously oversimplifies the real lighting one might like to do, but it is
likely to be much better than the projected flashes everyone is used to currently. Implementing
this sort of lighting falloff is extremely simple; one only needs to create a coordinate frame
centered around the source of the illumination and map texture coordinates from it. In practice,
this means generating texture coordinates from world or eye-space locations and back, trans-
forming them into this light-centered volume. The texture is then applied via modulation with
the base texture. The wrap modes for the light map texture should be set as described in the
earlier section on the general concept of attenuation.
Extending the simple glow technique to more advanced local lighting models brings in
more complex shader operations. A simple lighting case that might spring to mind is to use this
to implement a point light. This is not really a good use of this technique, as a point light needs
to interpolate its light vector anyway to perform any of the standard lighting models. Once the
light vector is available, the shader is only a dot product and a dependent texture fetch away
from computing an arbitrary falloff function from that point. A more interesting example is to
use a more complex light shape such as a fluorescent tube. A fluorescent tube is a great exam-
ple. The light vector cannot be interpolated between vertices, so a different method needs to be
found. Luckily, a 3D texture provides a great way to encode light vector and falloff data. The
light vector can be encoded in the RGB channels, and the attenuation can simply fit into the
alpha channel. Once the correct texel is fetched, the pixel shader must rotate the light vector
from texture space to tangent space by performing three dot products. Finally, lighting can pro-
ceed as usual.
This algorithm uses the same logic for texture coordinate generation as the attenuation
technique described earlier. The only real difference with respect to the application of the vol-
ume texture is that the light vector mentioned above is encoded in the texture in addition to the
attenuation factor. The application of the light is computed in the following shader using the
light vector encoded in the volume:
436 Part 4: Using 3D Textures with Shaders
Team LRN
ps.1.4
texld r0.rgba, t0.xyz //sample the volume map
texld r1.rgb, t1.xyz //sample the normal map
texcrd r2.rgb, t2.xyz //pass in row 0 of a rotation
texcrd r3.rgb, t3.xyz //pass in row 1 of a rotation
texcrd r4.rgb, t4.xyz //pass in row 2 of a rotation
//rotate Light vector into tangent space
dot3 r5.r, r0_bx2, r2
dot3 r5.g, r0_bx2, r3
dot3 r5.b, r0_bx2, r4
//compute N.L and store in r0
dot3_sat r0.rgb, r5, r1_bx2
//apply the attenuation
mul r0.rgb, r0, r0.a
The only tricky portion of this shader is the rotation into tangent space. The rotation matrix is
stored in texture coordinate sets two through four. This matrix must rotate from texture space
to surface space. This matrix is generated by concatenating the matrix composed of the tan-
gent, binormal, and normal vectors (the one typically used in per-pixel lighting) with a matrix
to rotate the texture-space light vector into object space. This matrix can easily be derived from
the matrix used to transform object coordinates into texture space.
References
[Drebin88] Robert A. Drebin, Loren Carpenter, and Pat Hanrahan, Volume Rendering,
SIGGRAPH Proceedings, Vol. 22, No. 4, August 1988, pp. 65-74.
[Perlin85] Ken Perlin, An Image Synthesizer, SIGGRAPH Proceedings, Vol. 19, No. 3, July
1985, pp. 287-296.
[Upstill89] Steve Upstill, The RenderMan Companion (Addison Wesley, 1989).
3D Textures and Pixel Shaders 437
Team LRN
Truly Volumetric Effects
Martin Kraus
This article presents several examples of ps.1.4 programming for volumetric effects, ranging
from quite basic techniques to recent research results; this includes volume visualization with
transfer functions, animation of volume graphics by blending between different volume data
sets, animation of transfer functions, and a recently developed rendering technique for
real-time interactive, high-quality volume graphics, which is called pre-integrated volume
rendering.
Volumetric effects have been part of computer graphics since transparency became popu-
lar, more than 20 years now. They include the rendering of common effects like explosions,
fire, smoke, fog, clouds, dust, colored liquids, and all other kinds of semi-transparent materials.
However, there are important differences. For example, noninteractive rendering for movie
productions usually tries to achieve the highest possible quality by simulating the physics of
volumetric effects, at least approximately. In contrast to this, graphics in computer games do
not have the privilege of noninteractivity. Therefore, they usually have to fake volumetric
effects, in particular by rendering polygons with precalculated, semi-transparent textures.
Nowadays, this necessity of faking volumetric effects becomes questionable, as modern
graphics boards offer the required resources for real volume graphics. In fact, programmable
pixel shading offers even more than is necessary, allowing us to increase the quality of interac-
tive volume graphics and, maybe more importantly, to animate volumetric effects in ways that
have not been possible in real time before.
The Role of Volume Visualization
One of the most valuable resources for programmers of interactive, truly volumetric effects is
the literature on volume visualization. Since volume visualization is a branch of scientific data
visualization, faking volumetric effects is not an option in this field; however, the demands for
real-time interactivity and high visual quality come close to the demands in the entertainment
industry. Moreover, in recent years, many applications of volume rendering techniques have
emerged that are no longer within the strict limits of scientific visualization. All these technol-
ogies are now often subsumed under the more general term volume graphics, which is a
branch of computer graphics.
Apart from this growth of research topics, there is also a steady transition going on with
respect to the employed hardware. In fact, PCs with OpenGL-capable graphics boards have
been a common benchmark environment for algorithms in volume visualization for several
438
Team LRN
years. Consequently, researchers have also started to make use of specific hardware features of
new PC graphics boards. This transition from scientific volume visualization on high-end
graphics workstations to volume graphics on off-the-shelf PCs in combination with the similar-
ity of challenges is why recent research in volume visualization is so much more relevant for
graphics and game programming than it was a few years ago. Therefore, it is worthwhile to
take a closer look at the basic problem of volume visualization and the way researchers have
solved it with the help of programmable pixel shading.
Basic Volume Graphics
Volume visualization started with the task of generating pictures of three-dimensional scalar
fields (i.e., volumes in which a real number is assigned to each point). This problem arose in
medical imaging and many kinds of computer simulations. While one common way to view
and represent this kind of data was a stack of two-dimensional grayscale images, the best
suited representation on a modern graphics board is a three-dimensional texture with a single
component. Note that this volume texture does not hold any colors or opacities, but only a sin-
gle number per texel (or voxel in the terms of volume graphics). Although it might be
convenient to store this number in the place of the red component of a texel, this in no way
implies that all voxels have to be reddish.
If the volume texture does not specify colors, where do the colors come from? The answer
of volume visualization is the concept of a transfer function. This function maps the numbers
stored in the volume texture to colors and opacities (i.e., to red, green, blue, and alpha compo-
nents). Thus, a transfer function works very much like a one-dimensional dependent texture,
and this is in fact the way transfer functions are implemented with programmable pixel
shading.
Once we can color a single point of a volume texture, we can also render an arbitrary slice
through it by assigning appropriate texture coordinates to the corners of the slicing polygon, as
shown in Figure 1 for a volume texture, which resulted from a volume scan of a teddy bear. If
we are able to render one slice, we can easily render the whole volume by rendering a stack of
Truly Volumetric Effects 439
Note: Of course, there are many different algorithms for volume rendering
(e.g., ray-casting, shear-warp, splatting, cell projection, axis-aligned slicing with
two-dimensional textures, or viewplane-aligned slicing with three-dimensional
textures). While this article only covers the latter, many of the presented ideas
work exactly the same or in a similar way for other rendering algorithms.
Note: If a specific graphics hardware device does not support dependent tex-
tures or similar techniques, then the transfer function has to be applied to the
volume data in a preprocessing step and the resulting colors and opacities are
stored in an RGBA volume texture. (The technical term for this method is
preclassification.) Unfortunately, this alternative has many drawbacks. In partic-
ular, it needs four times more texture memory and does not usually permit
modifications of the transfer function in real timea feature that is crucial for
both volume visualization and the programming of many volumetric effects.
Team LRN
many slices parallel to the viewplane and compositing the slices in the frame buffer with the
blending equation:
destColor = (1 - srcAlpha) * destColor + srcAlpha * srcColor
Figure 2 depicts a coarse rendering with too few slices, while an appropriate number of
slices was chosen for Figure 3. In order to demonstrate the effect of transfer functions, Figure 4
depicts the same object with identical viewing parameters but a different transfer function.
440 Part 4: Using 3D Textures with Shaders
Figure 1: A textured slice through a vol-
ume texture
Figure 2: Several (but too few)
composited slices
Note: In many cases, srcColor is pre-multiplied with srcAlpha, thus the
blending equation simplifies to:
destColor = (1 - srcAlpha) * destColor + srcColor
Figure 3: A volume rendering of a teddy
bear
Figure 4: Same as Figure 3 with a more
colorful transfer function
Team LRN
Here is a simple pixel shader 1.4 program that samples a volume texture, which is stored in
texture stage 0, and applies a transfer function, which is specified in texture stage 1.
ps.1.4
texld r0, t0.xyz // lookup in volume texture
phase
texld r1, r0 // lookup for transfer function
mov r0, r1 // move result to output register
Rendering a stack of viewplane-parallel slices that are textured this way is straightforward. In
fact, it is much more challenging to come up with interesting volume data sets and appropriate
transfer functions for particular volumetric effects.
Animation of Volume Graphics
While animations of faked volumetric effects often require too many resources, texture mem-
ory in particular, or do not offer a satisfying visual quality, there are several ways to animate
true volume graphics without additional costs.
Rotating Volume Graphics
Perhaps the most important advantage of truly volumetric graphics is the ability to rotate them
easily. In order to do this, we may, however, not rotate the geometric positions of the slices but
rather the texture coordinates. This may seem awkward at first glance, but it becomes clear
when you remember that the slices are parallel to the viewplane all the time. (See the source
code for the simple volume renderer on the books companion CD for more details.)
Animated Transfer Functions
In some sense, transfer functions are just the three-dimensional analog of color palettes or
indexed colors for images. Many applications of animated color palettes exist in two-dimen-
sional computer graphics; similarly, the animation of transfer functions has a special
importance for volume graphics and volumetric effects in particular.
While color palettes are often animated by modifying single entries or setting a new pal-
ette, there is a different, very efficient and general way to animate transfer functions that are
implemented by dependent textures. The idea is to store a set of one-dimensional textures as
the rows of a two-dimensional texture. When accessing this two-dimensional dependent tex-
ture, we just need to set a second texture coordinate (with the same value for all fragments),
which selects a specific row of the texture and corresponds to a time parameter. We can easily
implement this idea by extending the previous code example. The volume texture is again set
Truly Volumetric Effects 441
Tip: A common technique to reduce the screen area of the slicing polygons
is to clip them in the space of texture coordinates, such that all fragments have
texture coordinates within the limits of the volume texture and no time is wasted
with the rasterization of superfluous fragments. This technique is demonstrated
in Figures 1 to 4.
Team LRN
in texture stage 0, texture stage 2 specifies the texture containing the transfer functions, and the
green component of the constant vector c0 represents a time parameter.
ps.1.4
texld r0, t0.xyz // lookup in volume texture
mov r0.g, c1.r // copy time parameter
phase
texld r2, r0 // lookup for transfer functions
mov r0, r2 // move result to output register
Blending of Volume Graphics
While blending between images is definitely not a spectacular effect, blending between volume
data can lead to rather surprising and dramatic results. This is partially due to the effect of
transfer functions and can be further enhanced by blending between more than two volume
data sets, especially volume data sets with a complicated structure (e.g. noise or turbulence).
An example is given in Figures 5 to 8. Figure 5 depicts a volume rendering of the radial
distance field of a point, while Figure 6 shows a volume data set consisting of noise. Blending
between these two volumes with different pairs of weights results in Figures 7 and 8.
442 Part 4: Using 3D Textures with Shaders
Figure 5: Volume rendering of a sphere
Tip: One additional advantage of this implementation is the linear interpolation
in the dependent texture lookup (i.e., the possibility to choose the time parameter
continuously). This results in a smooth blending between successive transfer func-
tions, which permits smoother animations.
Tip: As indicated in Figures 7 and 8, we cannot only blend smoothly from one
volume texture to another, but we can also distort, warp, or twist volumes by
blending with an appropriate second volume texture.
Team LRN
The following implementation stores four independent volume data sets in the red, green, blue,
and alpha components of an RGBA volume texture. After the texture lookup, a dot product
with a constant vector of weights is performed in order to blend between the four data sets.
Finally, we apply the transfer function and move the result to the output register. The volume
data is again associated with texture stage 0, time-dependent weights are stored in c0, and tex-
ture stage 1 specifies the transfer function.
ps.1.4
texld r0, t0.xyz // lookup in RGBA volume texture
dp4 r0, r0, c0 // blend components with c0
phase
texld r1, r0 // lookup for transfer function
mov r0, r1 // move result to output register
Truly Volumetric Effects 443
Figure 6: Volume rendering of some noise Figure 7: A blended combination of the
volumes of Figure 5 and 6
Figure 8: Same as Figure 7 with
an increased weight of Figure 6
Team LRN
High-Quality but Fast Volume Rendering
Presumably, the most important drawback of volume rendering is the required rasterization
performance necessary to achieve a satisfying visual quality by rendering a sufficient number
of slices. How many slices are necessary? A first guess is to render enough slices to match the
resolution of the volume texture. This is, in fact, very acceptable for small volume textures.
Unfortunately, the application of transfer functions usually adds further fine structures to the
original volume texture, which leads to visible artifacts if the number of slices is not increased
appropriately. Thus, our dilemma is that we do not want to abandon transfer functions, but we
also do not want to render additional slices.
Fortunately, the literature on volume graphics suggests a solution to this problem, called
pre-integrated volume rendering. In the context of textured slices, this technique actually
requires programmable pixel shading. Its features include full support for transfer functions
and a much higher visual quality than the simple composition of slices, in particular for trans-
fer functions with jumps and sharp peaks. For a detailed description of this algorithm, see
[EKE2001].
Pre-integrated volume rendering reduces the number of slices by rendering slabs instead of
slices. A slab is simply the space between two successive slices. For a single fragment, the
problem reduces to the computation of the color contributed by a ray segment between the two
slices; see Figure 9. If a linear interpolation of the volume data along this ray segment is
assumed, the color will only depend on three numbers: the value on the front slice (at point F
in Figure 9), the value on the back slice (at point B), and the distance between the two slices. If
the slice distance is also assumed to be constant, the color of the ray segment will only depend
on the two values of the volume data, and a two-dimensional lookup table is already enough to
hold this dependency.
In the context of pixel shading, this table is implemented as a two-dimensional dependent
texture, which is generated in a preprocessing step by computing the ray integrals for all possi-
ble pairs of values at the start and end points F and B. (The precomputation of this integral is
the reason for the name pre-integrated volume rendering.)
A corresponding ps. 1.4 program has to fetch the value of a volume texture on the front
slice and the value on the back slice (i.e., on the next slice in the stack of slices). Then these
two values are used as texture coordinates in a dependent texture fetch, which corresponds to
the lookup of the precalculated colors of ray segments.
In the following code example, texture stage 0 specifies the volume texture on the front
slice, while texture stage 3 specifies the volume texture on the back slice (i.e., both refer to the
same volume texture data but use different texture coordinates). The lookup table for the
precomputed ray integrals is specified in texture stage 4.
444 Part 4: Using 3D Textures with Shaders
Tip: Depending on the particular data sets, the vector of weights does not nec-
essarily have to be normalized. In fact, a sum of weights unequal to 1 offers
another degree of freedom for volumetric effects. Another way of enhancing this
example is the animation of transfer functions, as discussed before.
Team LRN
ps.1.4
texld r0, t0.xyz // lookup for front slice
texld r3, t3.xyz // lookup for back slice
mov r0.g, r3.r // merge texture coordinates
phase
texld r4, r0 // lookup for transfer function
mov r0, r4 // move result to output register
Note that an additional mov command is necessary in the first phase in order to merge the two
slice values in register r0, which holds the texture coordinates of the dependent texture lookup
in the second phase.
The following code shows one possibility for the computation of a lookup texture for
pre-integrated volume rendering of slabs of a specified thickness. The RGBA transfer function
with premultiplied colors is stored in an array of floats called tfunc[TABLE_SIZE][4]. Note
that the color and alpha components in this transfer function specify color and opacity densities
(i.e., color and opacity per unit length). The resulting RGBA lookup table is returned in
lutable[TABLE_SIZE][TABLE_SIZE][4].
void createLookupTable(float thickness,
float tfunc[TABLE_SIZE][4],
float lutable[TABLE_SIZE][TABLE_SIZE][4])
{
for (int x = 0; x < TABLE_SIZE; x++)
{
for (int y = 0; y < TABLE_SIZE; y++)
{
int n = 10 + 2 * abs(x-y);
double step = thickness / n;
double dr = 0.0, dg = 0.0, db = 0.0,
dtau = 0.0;
for (int i = 0; i < n; i++)
{
double w = x + (y-x)*(double)i/n;
if ((int)(w + 1) >= TABLE_SIZE)
w = (double)(TABLE_SIZE - 1) - 0.5 / n;
double tau = step *
(tfunc[(int)w][3]*(w-floor(w))+
tfunc[(int)(w+1)][3]*(1.0-w+floor(w)));
double r = exp(-dtau)*step*(
tfunc[(int)w][0]*(s-floor(w))+
Truly Volumetric Effects 445
Figure 9: A slab of the volume between two slices
Team LRN
tfunc[(int)(w+1)][0]*(1.0-w+floor(w)));
double g = exp(-dtau)*step*(
tfunc[(int)w][1]*(w-floor(w))+
tfunc[(int)(w+1)][1]*(1.0-w+floor(w)));
double b = exp(-dtau)*step*(
tfunc[(int)w][2]*(w-floor(w))+
tfunc[(int)(w+1)][2]*(1.0-w+floor(w)));
dr += r;
dg += g;
db += b;
dtau += tau;
}
}
lutable[x][y][0] =
(dr > 1.0 ? 1.0f : (float)dr);
lutable[x][y][1] =
(dg > 1.0 ? 1.0f : (float)dg);
lutable[x][y][2] =
(db > 1.0 ? 1.0f : (float)db);
lutable[x][y][3] =
(1.- exp(-dtau));
}
}
Where to Go from Here
Programmable pixel shading offers a whole world of new possibilities for volumetric effects,
and only a small fraction of them have been actually implemented and described. This article
covers an even smaller fraction of these ideas, focusing on the pixel shader programming.
Thus, there are many paths to take from here, for example:
n
Play with the code! The books companion CD contains a simple volume renderer, which
includes all the code snippets presented and discussed in this chapter. It also displays and
lets you modify the pixel shader programs.
n
Design your own volumetric effects: Try to play with different volume textures and trans-
fer functions.
n
Look for other applications of volume graphics. The research literature about volume visu-
alization and volume graphics is full of interesting applications (e.g., shape modeling,
terrain rendering, simulation of natural phenomena, etc.).
446 Part 4: Using 3D Textures with Shaders
Tip: There are many ways to accelerate the computation of the pre-integrated
lookup table; some of them are discussed in [EKE2001]. However, for static trans-
fer functions, the best solution is probably to save this table to a file and load it
from disk when it is needed.
Team LRN
Acknowledgments
The author would like to thank Klaus Engel, who generated Figures 1 to 8 and programmed
earlier versions of several code samples.
References
[EKE2001] Klaus Engel, Martin Kraus, and Thomas Ertl, High-Quality Pre-Integrated
Volume Rendering Using Hardware-Accelerated Pixel Shading, SIGGRAPH
Proceedings/Eurographics Workshop on Graphics Hardware, August 2001, pp. 9-16
(http://wwwvis.informatik.uni-stuttgart.de/~engel/pre-integrated/).
Truly Volumetric Effects 447
Team LRN
448
First Thoughts on Designing a Shader-Driven
Game Engine
by Steffen Bendel
When you design a game engine, you have to think about the usage of
effects before writing even one line of code. Otherwise, it is possible
that your initial engine design might be trashed by the addition of a
new effect. A single effect is easy to realize, but an engine is more than
the sum of its components. The topics in this article focus on issues
that need to be taken into consideration when designing an engine that
uses shaders.
Team LRN
449
Visualization with the Krass Game Engine
by Ingo Frick
The objective of this article is to present selected visualization tech-
niques that have been applied in a successful game engine. The
structure of these techniques is demonstrated and motivated by their
developmental history, the further requirements in the foreseeable
future, and the main design goal of the entire engine.
Designing a Vertex Shader-Driven 3D Engine
for the Quake III Format
by Bart Sekura
This article explains how vertex programs can be used to implement real
game engine algorithms, in addition to great-looking effects and demos.
Rather than showing a particular trick, my goal is to provide the reader
with the display of applicability which, for the current vertex shader
capable hardware, provides easier coding paths and visible performance
benefits compared to CPU-based implementations of the same effects.
Team LRN
First Thoughts on
Designing a Shader-
Driven Game Engine
Steffen Bendel
When you design a game engine, you have to think about the usage of effects before writing
even one line of code. Otherwise, it is possible that your initial engine design might be trashed
by the addition of a new effect. A single effect is easy to realize, but an engine is more than the
sum of its components. An engine includes a concept and a lot of compromises. It must share
the resources in a meaningful way. Some of the effects described in this book are too expensive
to incorporate into a real-time game engine. In case of an already existing game engine, they
might be too difficult because the overall engine design is not prepared for that. The perfect
engine does not exist, but there is a good engine for a special purpose and for a defined hard-
ware basis. Try to focus on the things that are most important for you and optimize your design
for that.
I would like to present a few thoughts on specific topics that have to be considered while
designing a engine that uses shaders.
Bump Mapping
Bump mapping is futile without real dynamic lighting. If there is no moving light source, you
see no effect and the effort is wasted. So bump mapping is not an isolated effect. It needs a
complete engine infrastructure, not only for light. Objects using bump mapping require addi-
tional tangent space vectors, which must be supported by the rendering pipeline. The enormous
effort is the reason why it is so seldom used.
Real-time Lighting
Real-time lighting is one of the most important things in engine design today. It is not a
post-rendering effect that changes the brightness of the picture; it is a calculation which con-
nects material properties and light intensity. If you decide to use real-time lighting, you have to
implement it into every shader you use. Whereas per-vertex lighting is an easy way to
450
Team LRN
implement lighting, it is very inaccurate. It works only with specifically optimized meshes
because you get artifacts with big triangles. You cant use an LOD system with per-vertex
lights, as lower LOD levels will produce artifacts because of the bigger triangles.
Precalculation is only possible for static lights. Dynamic lightmaps can be calculated only
in real time if there is a simple projection algorithm for the texture coordinates. This is espe-
cially suitable for flat landscapes. One of the big challenges is to give the user a consistent
impression of the lighting in the engine. Using different types of lighting for objects and land-
scape doesnt seem to work. Try to avoid an artifical look here.When using dynamic lights in
the engine, sometimes a very simple lighting model is used. The resulting picture of such a
simple lighting model will be too clear and too easy to analyze for the human brain, which is
confusing. To make it more dusty, a preprocessed part of light might be a good idea. You might
use radiosity, which could be too slow, or precalculated lightmaps to make it more dusty.
Use Detail Textures
In the real world, there are numerous details all the way down to the microscopic level. In
computer graphics, the user should not realize an end to details. Geometry is limited by hard-
ware polygon throughput, but for textures, multilayer detail textures are a good solution.
Approximately one texel of the highest resolution texture should be mapped onto one screen
pixel.
Use Anisotropic Filtering
Blurring is no longer state of the art. If hardware supports it, anisotropic texture filtering can
increase quality in some cases. Without either, the picture is smooth or flickers. The difference
between anisotropic and trilinear filtering is amazing.
Split Up Rendering into Independent Passes
The more complex the rendering model gets, the more important a clear structure of the engine
will get. Two elements determine the color of the object: material and light. The product of
both is the visible color. In the simplest version, this is premodulated in a single texture:

Color r Material r Light r
i
i
i j
j

r r r

10 . ,
,
Fog r r
Fog r r
Camera
Camer
intensity
intensity
r r
r r

a
Fogcolorr Worldposition
r
If you use a different algorithm to render material and light, then there are many combinations.
To reduce the count of vertex and pixel shaders, it makes sense to split the rendering into
independent passes. The speed increase by using two textures instead of one is noticeable.
More than two textures per pass amount is not so noticeable and is not compatible with older
hardware. Thus, it is not necessary if the feature does not request it. A separation of the mate-
rial, light, and fog part allows the combination of every material shader with any light shader.
First Thoughts on Designing a Shader-Driven Game Engine 451
Team LRN
A compact shader is, of course, the better solution if there is only one material map and one or
no lightmap, but these are a limited number of cases.
Two solutions are possible. First, the material color is accumulated in the frame buffer and
the result is blended with the light calculated by the following pass, or the light is rendered and
modulated by material. The best solution depends on your specification. The component that is
rendered second must fit into one pass.
Use _x2
Most rendering values are only in a range between zero and one. This means the light factor is
limited and the product of a color multiply is darker than the original values. To get a brighter
picture, the result of light and material blending should be scaled by a factor of 2. This is pro-
cessed by using the pixel shader _x2 modifier in the single pass solution. If the blending is
done in the frame buffer, SourceBlendingFactor is set to DestColor and DestBlendingFactor is
set to SourceColor to realize the doubled modulation.
452 Part 5: Engine Design with Shaders
Team LRN
Visualization with the
Krass Game Engine
Ingo Frick
Introduction
The objective of this article is to present selected visualization techniques that have been
applied in a successful game engine. The structure of these techniques is demonstrated and
motivated by their developmental history, the further requirements in the foreseeable future,
and the main design goal of the entire engine.
Apart from the theoretical discussion, the practical suitability will also be demonstrated, as
all techniques were used without modification in AquaNox, the recently published game
from Massive Development GmbH in Germany.
We will present techniques which offer a good balance between the current state of tech-
nology, practical usability, and the application in a real game engine. All the techniques are
part of the krass engine, a European game engine that powers some high-end 3D games. It is
important to see the historical development of this engine, as it didnt take place in an ivory
tower with the only focus on technical details. Instead, it originated in the context of gaming
and was first used about six years ago in a game called Archimedean Dynasty. Because of its
origins, its software patterns are mainly designed to solve the problems that arise during the
creation of computer games.
General Structure of the Krass Engine
To explain the location of the rendering systems within the main structure, we will begin with a
short description of the engines architecture. The entire system offers an extensive functional-
ity covering the areas of visualization, networking, simulation, collision, mathematics, physics,
and more. The top structure consists of three separated layers. The topmost layer, the so-called
application layer, holds the game-specific code. It is intended to exchange this layer while
moving to another project. The second layer, the component layer, aggregates functionality,
which had been identified as reusable units and can be shared between different games; for
instance, here you find systems like particle systems, collision systems, animation systems, and
453
Team LRN
more. The third and lowest layer is the system layer. It offers basic system functionality like
I/O, timing, rendering, etc.
Developmental History of the Rendering
Component
The rendering component is implemented in the system and component layers. The basic
functionalities, like textures, vertex buffers, frame buffers, and depth buffers, are located in the
system layer. More specific effects, like render object with fogmap, static- and dynamic
lightmap, are located in the component layer. The best way to understand the structural archi-
tecture is to follow its historical development.
Regarding the development of visualization techniques over the past five years, the devel-
opment of specialized hardware was the biggest change. This specialization led to higher
graphics quality and an unloading of the main CPU. This shift made a lot of reorganization
necessary. Some of them negatively affected the software design and the downstream produc-
tion processes. In earlier (software-driven) rendering systems, it was usually easier to design
the algorithms and data structures regarding flexibility and controllability. Accessing data
structures like triangles, vertices, and materials using the consistent and powerful program-
ming language of the main CPU (C++, Assembler) often led to data structures that were more
object oriented than process oriented. Beyond this, the overall production efficiency was very
high. This meant that the evolution of specialized hardware had to pass through many steps to
regain these initial advantages.
Considered a simple model, the generation of a picture with the rendering approach
requires at least the following data structures:
n
Vertex
n
Triangle (constructed from vertices)
n
Triangles pixel (fragment, generated from color(s), texture(s), etc.)
n
Frames pixel (composed of triangle pixel and frame pixel)
The enumeration above is not complete, but it should be sufficient for the following thoughts.
In the following table, you will find a brief overview of the chronological shift of these data
structures from software to specialized hardware. The data structures that came along with
each shift are labeled with a D and the respective processes are labeled with AP. SW-
means software and HW- means hardware.
Table 1
Period Vertex Triangle Trianglepixel Framepixel
0
Software-
Rendering
D: SW-Vertex
(arbitrary layout)
P: SW-Processing
(programmable)
D: SW-Triangle
(arbitrary layout)
P: SW-Processing
(programmable)
D: SW-Texture
(arbitrary layout)
P: SW-Processing
(programmable)
D: HW-FrameBuffer
(selection)
P: SW-BlendingMode
(programmable)
454 Part 5: Engine Design with Shaders
Team LRN
Period Vertex Triangle Trianglepixel Framepixel
1
Glide
D: SW-Vertex
(predefined)
P: SW-Processing
(programmable)
unchanged D: HW-Texture
(selection)
P: HW-RenderState
(selection)
D: HW-FrameBuffer
(selection)
P: HW-BlendingMode
(selection)
2
Hardware
(T/T&L)
D: HW-VertexBuffer
(selection / (FVF))
P: HW-Processing
(selection (T/T&L))
D: HW-IndexBuffer
(selection)
P: HW-Processing
(selection)
unchanged unchanged
3
Vertex
Shader
D: HW-VertexStream
(selection)
P: HW-VertexShader
(programmable)
unchanged unchanged unchanged
4
Pixel
Shader
unchanged unchanged D: HW-Texture+Data
(selection )
P: HW-PixelShader
(programmable)
unchanged
5
Pixel
Shader++
unchanged unchanged D: HW-Texture+Data
(arbitrary layout)
P: HW-PixelShader
(programmable)
D: HW-FrameBuffer
(arbitrary layout)
P: HW-PixelShader
(programmable)
Comparing period 0 (pure software rendering) with period 5 (maximum hardware support)
shows a nearly equivalent flexibility for the individual organization and interpretation of the
data structures and processes. This closes the evolution circle mentioned above.
Previous Drawbacks of Hardware Development
The main objective of visualization in computer games is the creation of individual effects.
They are the unique selling points over the competitors and show a definite influence on the
success of a product. Restrictions caused by the current hardware technology normally result in
a drastic reduction of the possible amount of individual effects. This consequence could clearly
be noticed right after the introduction of graphics hardware with T&L capabilities. Though the
game developer could configure the T&L pipeline for the standard cases, every aberrant fea-
ture caused serious problems. The Character Skinning problem, for instance, is well suited to
demonstrate this effect. From todays point of view, the approach to solve this problem using
the standard path (e.g., transformation matrix palettes) seems curious. The resulting lack of
individuality caused all 3D games of this period to look nearly the same.
Current Drawbacks
With the return of flexibility (e.g., vertex and pixel shaders), this problem can be avoided. The
variety of visual effects is again determined by the imagination of the developer instead of the
hardware capabilities. Instead, now you have to face another problem, which unfortunately
gets overlooked many times. The programming of individual vertex and pixel shader effects
only leads to a local flexibility (assembly-syntax, unrestricted data layout). In a larger context,
Visualization with the Krass Game Engine 455
Team LRN
you catch a global inflexibility, as we went back to the development of effects. Effects very
often tend to interact with each other, which makes it difficult to recombine simple effects to
more complex ones. Take a simple vertex shader implementing some effects like skinning and
mapping texture coordinates and try to combine it with a second vertex shader doing the illu-
mination. This usually very common and basic combination becomes difficult.
Consequently, you have to implement the skinning shader for each number of possible
light sources. This example seems to be more of a technical shortcoming of the current hard-
ware technology, but it demonstrates the general problem.
Ordering Effects in the Krass Engine
Modeling the rendering process through an accumulation of a sufficient amount of effects
seems to be a reasonable method. It disengages the rendering techniques from using
parameterized standard methods and enters the terrain of individually designed rendering
effects. These thoughts also stood at the beginning of the current krass revision. The next ques-
tion is how to define and associate the effects to the overall structure.
Considering commercial and traditional rendering systems for the non-real-time range of
use (3DSMax, Maya), you will usually not find effects in terms of our definition. In fact, these
applications use standard rendering methods on a per material basis and combine these with a
multipass approach to emulate quasi-effects. Apart from that, some effect-oriented concepts
have been realized, such as the shading concept of the RenderMan systems. Here as well, the
visualization is done by the individual programming of a pixel effect (material) rather than
by the combination of generic texture parameter sets.
In our case (real time, hardware accelerated), it makes no sense to take something like a
material as the only criterion of organization. In most of the cases, we dont have a material in
the common sense (e.g., terrain rendering) or the effect is determined more by a geometric
characteristic (e.g., plankton, skinning, morphing).
In our latest game, AquaNox, it turned out to be very useful to order the effects directly by
the different game entities. Each effect is created by some subeffects combined in a very spe-
cial way to ensure flexibility and recombination. The effects themselves are labeled as render
pipeline (RP) because they can be interpreted as a specialized encapsulation of the global ren-
der pipeline for a single effect.
The following table contains a short description of some render pipelines together with a
qualitative description of the effect and the related main elements:
Table 2
Render Pipeline Elements of Effect
Building
RP for all fixed objects like buildings
n
Caustic texture (animated)
n
Light texture (prelighted)
n
Dynamic lighting
n
Detail texture
n
Diffuse texture
n
Fog texture
456 Part 5: Engine Design with Shaders
Team LRN
Render Pipeline Elements of Effect
Terrain
RP for generating the terrain visualization from
given level-of-detail (LOD) geometry
n
Caustic texture (animated)
n
Light texture (prelighted)
n
Dynamic lighting
n
Detail texture
n
One to six material textures
n
One material selection texture
n
Fog texture
Plankton
RP for translation, rotation, wrapping, and visu-
alization of individual particles
n
Diffuse texture
n
Fog texture
n
Translation with vertex shader
n
Rotation with vertex shader
n
Wrapping in view space with vertex shader
At that point, it is appropriate to return to the above-mentioned problem of recombining effects
with one another (light source difficulty). To solve this problem in an elegant way, we only
have to unify the advantages of both approaches (recombination through multipass rendering
and individual effect programming for each pass). Reconsidering the AquaNox render pipelines
reveals an interesting strategy. Instead of combining the individual render passes in one
inflexible vertex and/or pixel shader, it makes sense to assign the individual passes to larger
function blocks.
We identified the following blocks:
n
Illumination (lightmap, caustic map, vertex-/pixel light, bump map)
n
Material (diffuse map, detail map)
n
Global effects (fog map, environment map)
That settles the matter for the larger group of effects. Creating the three function blocks illu-
mination, material, and global effects seems to be a sensitive concept, which has been dem-
onstrated in the krass engine and AquaNox for the first time. (In the following, this approach
will be abbreviated as IMG = Illumination, Material, Global Effects). This reorganization gives
you the opportunity to implement each individual pass with the highest flexibility and also to
recombine these blocks with each other. In this context, it makes sense to reintroduce the
multipass concept. So far, this concept has mainly been used as an emergency solution in the
case of an insufficient number of texture stages.
If you assume that the frame buffer stores arbitrary values instead of traditional pixels, you
can interpret the IMG technique as calling subroutines for specialized tasks (illumination,
material generation, global effects). If it was possible to call (non-trivial) subroutines within
the vertex and pixel shaders, the multipass IMG system could easily be collapsed to a single
pass concept.
Visualization with the Krass Game Engine 457
Team LRN
Application of the IMG Concept for Terrain
Rendering
The proposed function block organization particularly makes sense, as the first block (illumi-
nation) can be constructed from independent individual passes. This block finally results in the
summation of all lights. As shadow is the absence of light, even shadow passes fall within this
block. In the AquaNox terrain system, the summation of all lights originates in the following
sources:
1. Illumination Stage (I)
n
Light texture (prelighted)
n
Caustic texture (animated)
n
Dynamic lights (rendertarget)
The next stage (material) is composed of a predefined number of materials. Each material is
assembled from a material texture, a fine detail texture, and a coarser color texture.
2. Material Stage (M)
n
Material texture (sand)
n
Color texture (color)
n
Detail texture (sand crystals)
The last stage (global) handles the fog which is realized by using a procedurally generated and
dynamically mapped fog texture.
3. Global Stage (G)
n
Fog texture (procedural, volumetric fog)
Figure 1 shows an overview of the entire process. The three frames show the clearly separated
IMG stages. The interconnections between the different stages are indicated by the arrows. The
combining operations are displayed as well.
We can now go into detail regarding the global stage. AquaNox here uses a volumetric fog
system which allows us to define multiple planes of constant height with real fog density.
Interestingly enough, this feature is not only a visual effect but also offers a gameplay-relevant
element (e.g., hiding). We do not generate the fog itself with the graphics hardware in the stan-
dard manner (vertex and pixel fog), but instead we use the vertex shader to map a real-time
generated texture onto the geometry (in view space). This method doesnt show the annoying
artifacts that normally accompany the use of standard fog in combination with transparent tex-
tures. It also avoids the appearance of the popping effects that always happen when
vertex-related calculations are carried out on the basis of a real-time LOD geometry system.
The 3D texture is generated in a preprocessing step. The generator solves the equation of light
transport for a predefined parameter range through a numerical integration. The integrated val-
ues for fog density (a) and fog color (r,g,b) are stored in the 3D texture array. The texture
mapping itself corresponds more to a parameterized solution lookup of the equation of light
transport.
458 Part 5: Engine Design with Shaders
Team LRN
Visualization with the Krass Game Engine 459
Figure 2
Figure 1
Team LRN
Unfortunately, the major part of the customers presently installed hardwarebase does not sup-
port hardware-accelerated volume textures. To meet this challenge, we had to think of a small
work-around: We put the texture data set into the CPU main memory and extract one fog tex-
ture slice for each rendered frame. This slice is uploaded to the hardware and used as a
standard 2D texture. The restrictive condition for this method is to find one fixed texture coor-
dinate for the entire frame which could be achieved by a well-defined texture parameterization.
In the model we use here, the fog parameters are determined by the following values:
n
Height of the observer (origin of camera in world space)
n
Height difference between vertex and observer
n
Distance between camera and vertex (linear norm is sufficient)
Together with the information about the maximum distance (far plane) and the maximum
height and depth of the covered game world, it becomes possible to map the three values to the
3D texture coordinates (u,v,w). Because the height of the observer remains constant for the
entire frame, this parameter is used as the fixed coordinate (w). The texture slice (u,v) is
extracted from the 3D data array exactly at this coordinate (linear interpolation). Figure 2 dem-
onstrates this process:
Particle Rendering to Exemplify a Specialized
Effect Shader
Beside the shaders operating according to the introduced stage concept, we have developed
other shaders for AquaNox, which show the character of an effect shader. As an example for
these shaders, we will now take a look at the so-called particle shader: Its render pipeline
allows you to translate, rotate, wrap, and display a dense field of individual polygons in the
near range of the observer. At first glance, this sounds similar to the hardware-accelerated
point-primitive. The point-primitive had been invented to render huge amounts of quads with
an identical texture map. Up to now, they still show some serious shortcomings which chal-
lenge their use; for example, they are restricted in their maximum size, and they cannot be
rotated. These problems have been solved with the particle shader. We also additionally satis-
fied the following requirements:
n
Minimum vertex and pixel demand for individual particle
n
Translation of individual particle with graphics hardware
n
Rotation of individual particle with graphics hardware
n
3D position wrapping with graphics hardware (classic star scroller)
In order to minimize the area (pixel demand) and the number of vertices (vertex demand) of an
individual particle, we build up each particle from a single triangle. This avoids a second trian-
gle setup for the graphics hardware, and the amount of rasterized pixels is reduced (depends on
particle shape on texture). The vertex shader also has to process only three vertices per particle
instead of four. This is important because the vertex shader turned out to be the bottleneck
(complex vertex program).
The individual translation and rotation could be achieved by a sophisticated vertex stream
data layout. In contrast to the standard data, each vertex holds not only the position of the
460 Part 5: Engine Design with Shaders
Team LRN
triangles center, but also its relative position
with regard to this center. The individual
rotation is controlled by a rotational velocity
of this relative position around the center.
The wrapping functionality has been realized
with the help of the expp vertex shader
instruction. For this purpose, we first carried
out the translation in a unit particle cube (0 to
1) and later eliminated the integer part of the
position with the instruction mentioned
above. In the sequencing step, the unit cube
particle position is transformed into world
space coordinate system and the rotated rela-
tive vertex position is added. The correct
location (origin, dimension) of the unit cube
in the world space (required to create the
cubeworld transformation) is illustrated in
Figure 3.
The center of the cube (C) is determined
by adding the camera position to the viewing
direction vector which is normalized to a half
cube edge length in world space (D). The
entire vertex shader process is demonstrated in Figure 4.
Operation Space
Translation (P
ABS
+ V
ABS
* t) Unit Space
Wrapping (fmodf(P,1.0f)) Unit Space
Transform (UWV) View Space
Attenuation (~D) View Space
Rotate (P
ABS
+ Va * t * P
REL
) View Space
Transform (VC) Clip Space
Figure 4
Summary
In this article, we presented selected visualization techniques that we have actually used in a
shipped game engine. The structure of these techniques is demonstrated by their developmental
history. We showed that the techniques passed the real-world game test successfully with
AquaNox. The mixture of procedural and classic texture generation (fog), as well as the proce-
dural geometry handling (particle), proved to be adequate procedures. Introducing the IMG
concept offers an interesting alternative to reduce the huge amount of shader permutations
(>100) that game developers need for their games.
Visualization with the Krass Game Engine 461
Figure 3
Team LRN
Concluding Remarks and Outlook
The techniques presented here only demonstrate the current state of development. Regarding
the foreseeable developments in hardware evolution, you can deduce tendencies which will
become important in the near future:
n
The pixel shader will advance to be the most important instrument for rendering effects.
n
The complexity of vertex and pixel shader language will increase and new instructions will
be added (jumps, subroutine calls).
n
The size of programs, as well as the related data, will increase.
n
The hardware internal data formats will develop to a complete 32-bit floating-point
representation.
n
The classic pixel will only be used for input/output purposes.
Eventually, if you compare the graphics hardware processors (GPU) with the normal main pro-
cessors (CPU), the differences will appear rather small. Just like the main processors, the
graphics processors will need continually increasing assembly programs. This development
will finally force people to introduce a higher level programming language. Perhaps one of the
already existing higher level languages can be used, which might also be the one you used to
write the rest of the application.
So were back to square one. The only difference lies within ourselves: We have gained a
much more comprehensive rendering knowledge, and we can look forward to the next hard-
ware evolution in the realm of the remaining global rendering techniques, such as raytracing or
radiosity.
462 Part 5: Engine Design with Shaders
Team LRN
Designing a Vertex
Shader-Driven 3D Engine
for the Quake III Format
Bart Sekura
Some time ago, I developed a Quake III Arena level viewer for the purpose of trying out vari-
ous algorithms and techniques using an industry standard engine format and an abundance of
tools and shading techniques. When DirectX 8.0 came out, I immediately decided to port my
viewer to this API and in particular, take advantage of new programmable pipelines available
for developers. The result was a Quake III Arena level viewer, which runs various effects,
described in a shader script as a vertex program and thus on vertex shader capable cards in
hardware. In this article, I will present the vertex program translation process, which I think is
also very useful outside of the Quake level viewer context, as it introduces some nice tricks
and effects that can be implemented in vertex programs. The reader can find the Quake III
Arena viewer and its full source code on the companion CD. Please consult the viewer docu-
mentation included on the CD for instructions on how to run and compile it.
I have chosen this particular platform to demonstrate how vertex programs can be used to
implement real game engine algorithms, in addition to great looking effects and demos. Rather
than showing this or that particular trick, my goal is to provide the reader with the display of
applicability which, for the current vertex shader capable hardware, provides easier coding
path and visible performance benefits compared to CPU-based implementation of the same
effects.
Quake III Arena Shaders
Quake III Arena allows level designers to create shader scripts for a specific brush/surface dur-
ing mapping process. This shader is then used by the renderer to draw the faces associated with
it. The shader in Quake III Arena is a text script that describes how the surface should be ren-
dered and is preprocessed by the game engine upon startup. In general, the shader scripts can
be complex and can contain information not relevant for the rendering process (such as hints or
commands for the level editor). A description of the full capabilities of the shader system is
beyond the scope of this article. Useful documentation is provided by the game developer, id
software, and is included on the companion CD with this book.
463
Team LRN
Basically, at the beginning of the shader script, general parameters are provided, such as
culling or geometry deformation functions. Consecutive layers are then specified, where each
layer is usually a texture, along with its blending function and various parameters, such as:
n
Texture coordinates function (e.g., deformation function or a specific type of coordinate
generation, such as environment mapping)
n
Diffuse color and alpha source vertex, periodic function, fixed
n
Depth and alpha test parameters
We will focus on some of the most common shader effects that can be easily translated to ver-
tex programs. First, the description of the shader setup in the viewer is presented to give the
reader an idea of how things are organized. Next, we will follow the translation process and
look at some specific shader effects as they are translated into vertex programs. Finally, we
will put it all together and see the whole rendering process.
To minimize confusion, I will refer to Quake III Arena shader scripts as Quake III
shaders and DirectX 8.1 vertex shaders as simply vertex shaders.
Vertex Program Setup in the Viewer
Quake III Arena stores the game data in a package file, which is basically a zipped archive
containing all art content as well as Quake III shaders. The collection of shader scripts resides
in the scripts subdirectory inside the package and each of the text files usually contains many
Quake III shaders.
When the viewer initializes, it scans the package file for any Quake III shader scripts and
parses them into the internal format, represented as C++ classes, which contain shader parame-
ters and a collection of passes, each with specific material properties. Below is a C++
representation of Quake III shaders (reformatted for clarity):
struct shader_t
{
// shader pass definition
struct pass_t
{
pass_t() : flags(0), map_ref(0), vsh_ref(0), tcgen(0) {}
int flags; // misc flags
std::string map; // texture name applied for this pass
int map_ref; // internal renderer reference to texture
int vsh_ref; // internal vertex shader reference
int blend_src; // blend operation (source)
int blend_dst; // blend operation (destination)
int alpha_func; // alpha test function
int alpha_ref; // alpha test reference value
int depth_func; // depth test function
int tcgen; // texture generation type
// texture coordinate modification functions
std::vector<tcmod_t> tcmods;
// concatenated matrix of texture coordination modifications
matrix_t tcmod_mat;
464 Part 5: Engine Design with Shaders
Team LRN
// color (RGB) generation function
struct rgbgen_t {
rgbgen_t() : type(0), color(vec4_t(1,1,1,1)) {}
int type;
wave_func_t func;
vec4_t color;
}
rgbgen;
// animated textures collection
struct anim_t {
// helper to determine current texture from the collection
// based on current time and FPS settings
int frame(double time) const {
int i = (int)(time*fps)%maps.size();
return map_refs[i];
}
float fps; // animation speed
std::vector<std::string> maps;
// textures collection
std::vector<int> map_refs;
// renderer refs
}
anim;
};
// the shader itself
public:
shader_t() : flags(0), type(0), stride(0) {}
std::string name; // name of the shader
std::vector<pass_t> passes; // collection of passes
int stride; // stride in vertex stream
int flags;
int type;
int vertex_base;
int vertex_count;
// vertex deformation function
struct deform_t {
float div;
wave_func_t func;
}
deform;
// parsing function translates text script representation
// into this class internal representation suitable for
// rendering pipeline
static void parse(const byte* data,
int size,
std::vector<shader_t>& shaders);
};
This C++ object is created per the Quake III shader and used during the rendering phase. It
makes it easy to traverse the shader passes and apply various parameters that directly influence
the rendering process. One additional thing that is performed during the setup phase is setting
up all the textures used by Quake III shaders (loading them up and uploading to the card) and
Designing a Vertex Shader-Driven 3D Engine for the Quake III Format 465
Team LRN
translating of various passes into corresponding vertex shaders. I will describe the latter pro-
cess in greater detail now.
Many of the surfaces inside the game level reference Quake III shaders that dont exist in
the form of the script. This is just a handy shortcut for default Quake III shaders that use only a
base texture and lightmap with no additional rendering passes or special parameters. Whenever
the Quake III shader that is referenced from within the level is not found among the parsed
scripts, the viewer will create a default Quake III shader with two passes: default texture,
looked up by the shaders name, and a lightmap.
To facilitate the setup code as well as provide easier and more optimized code paths for
rendering, I divided the shaders into the following categories:
Table 1: Shader categories
Shader Category/Type Description
SHADER_SIMPLE This is the default Quake III shader that uses base texture, lightmap
texture, and constant diffuse. There is no vertex or texture coordinate
deformation, and diffuse color is not coming from vertex stream data.
This is by far the majority of shaders, also known as default shaders,
automatically created when there is no explicit script code. Primarily, it
is used for basic world geometry.
SHADER_DIFFUSE This category is very similar to the above shader type, the difference
being that diffuse color is coming from the vertex stream. It also uses
base texture and lightmap. It is used primarily for mesh models embed-
ded in world geometry.
SHADER_CUSTOM All non-standard shaders fall into this category, which represents arbi-
trarily complex shaders with one or many passes, various vertex and
texture coordinate deformation functions, etc.
The category that the Quake III shader belongs to is stored in a type member variable of
shader_t class. The default vertex shaders, SHADER_SIMPLE and SHADER_DIFFUSE, are
set up in world_t::upload_default_vertex_shaders() method. These never change and are
uploaded when the viewer starts. In addition, I have added one more default vertex shader,
which is not a result of Quake III shader script translation process but is set up to facilitate all
screen space rendering operations (heads-up display, 2D window interface, etc.) and is repre-
sented as SHADER_FONT type. This shader uses one texture and expects the coordinates in
normalized device range (1.0 to +1.0). The primary use of this shader can be found in font_t
class.
The Quake III shaders that fall into the SHADER_SIMPLE, SHADER_DIFFUSE, and
SHADER_FONT categories are hard-coded, since they never change. All SHADER_CUS-
TOM shaders are dynamically constructed depending on the number of passes, their
parameters, etc. This process is constructed from consecutive steps that result in final vertex
shader source code, which is then compiled by DirectX runtime and uploaded to the card. The
basic steps are as follows:
466 Part 5: Engine Design with Shaders
Team LRN
Table 2: Constructing custom shaders
Step Description
Vertex coordinates These are either set up as coming unchanged from the vertex stream or as
a result of a deformation function. In the second case, the vertex deforma-
tion function, sinus, is evaluated using a Taylor series.
Vertex diffuse color The vertex color can be constant, come from the vertex stream, or be evalu-
ated as a result of a periodic function. In any case, the actual evaluation
happens in CPU, and the vertex program code merely loads the appropri-
ate value from constant memory or vertex stream data.
Texture coordinates Texture coordinates either come unchanged from the vertex stream or are
computed using a simplified environment mapping function. The texture
unit used depends on whether were dealing with base texture or lightmap.
Texture coordinates
modification
Texture coordinates computed in the previous step can be optionally
changed, depending on shader pass parameters. If that is the case, they
are multiplied by a concatenated texture coordinate modification matrix in
the vertex shader. The matrix is prepared by CPU.
As you can see, various Quake III shader script effects are translated into the corresponding
vertex shader source. As it turns out, the final vertex shader code is very often identical for
many different Quake III shader scripts, and it would be a waste of resources to upload identi-
cal vertex shaders many times. I have employed a simple string comparison mechanism to
prevent this. Whenever a vertex shader is generated from the Quake III shader definition, after
it is compiled by the DirectX run-time, its source code is put on a map to allow for easy lookup
later on. The source code (a string) becomes a key for the collection and the value is the ID of
the already compiled vertex shader, as it was returned by the DirectX run-time after successful
compilation. Then, whenever the vertex shader source is generated and it is time to compile it,
first the map is consulted to see whether there was previously an identical vertex shader
already compiled. If there were, the same ID is used; otherwise, it is compiled and put on the
source code map.
Vertex Shader Effects
We will now look in more detail at some of the effects that can be created.
Deformable Geometry
One of the simple things that you can do in Quake III shader is deformable geometry.
Basically, you can specify a periodic function such as sin() and provide the parameters, and the
geometry is then deformed during the rendering process and parameterized by time. This is, for
example, how all those waving banners are done in Quake levels. It would be great to do that
in a vertex shader, especially since it involves recomputing the vertex position, something
wed rather avoid. So we need to evaluate the sin function from within vertex shader.
There is no direct way to do this; that is, there is no sin instruction similar to the C library
function that is available from within vertex shader. Therefore, we have to simulate one. This
becomes possible thanks to Taylor series, which allows us to generate approximate sinus val-
ues using available vertex shader instructions.
Designing a Vertex Shader-Driven 3D Engine for the Quake III Format 467
Team LRN
The theory behind Taylor series is beyond the scope of this article. What is interesting to
us is how to generate approximate sinus values using Taylor series. Here are the approximation
formulas:
sin(x) = x (x^3)/3! + (x^5)/5! + (x^7)/7! + ...
sin(x) ~= x*(1-(x^2)*(1/3!-(x^2)(1/5!-(x^2)/7!)))
To compute sinus from within vertex shader using Tayor series, we need to set up some magic
numbers in vertex shader constant memory. These numbers are:
1.0 1/3! 1/5! 1/7!
Additionally, since Taylor series will only work accurately with the input range of {PI, PI},
we need to prepare our input variable accordingly. This is done in the following way:
// r0.x x
// r1.y - temp
expp r1.y, r0.x
// r1.y=x-floor(x)
mad r0.x, r1.y, c[12].y, -c[12].x
// r0.x=x*2PI-PI
As you see, we also need to add 2*PI and PI to vertex shader constant memory. During transla-
tion of shader scripts into vertex shader, we generate vertex shader code with the following:
if(shader.flags&SHF_DEFORM)
{
// vertex deformation code
vsh << "dp3 r1.x, v0, c15.w\n"
<< "add r0.x, r1.x, c14.y\n" // r0.x=off+phase
<< "add r0.x, r0.x, c15.x\n" // r0.x=off+phase+time
<< "mul r0.x, r0.x, c14.w\n" // r0.x=(off+phase+time)*freq
<< "expp r1.y, r0.x\n" // r1.y=x-floor(x)
//r0.x=x*TWO_PI-PI
<< "mad r0.x, r1.y, c12.y, -c12.x\n"
// fast sin
// r0.x=sin(r0.x)
<< "dst r2.xy, r0.x, r0.x\n"
<< "mul r2.z, r2.y, r2.y\n"
<< "mul r2.w, r2.y, r2.z\n"
<< "mul r0, r2, r0.x\n"
<< "dp4 r0.x, r0, c13\n"
// r1=y*amp+base
<< "mad r1.xyz, r0.xxx, c14.zzz, c14.xxx\n"
<< "mov r0,v0\n"
<< "mad r0.xyz, v1.xyz, r1.xyz, r0.xyz\n"
<< "mov r0.w, c10.x\n" // w=1.0f
// transform to clip space
<< "dp4 oPos.x,r0,c0\n"
<< "dp4 oPos.y,r0,c1\n"
<< "dp4 oPos.z,r0,c2\n"
<< "dp4 oPos.w,r0,c3\n";
}
This vertex shader section takes vertex position in world space and deforms in time using the
sinus function. All parameters (particularly current time) are passed in using vertex shader con-
stant memory. The shader computes a new vertex world position and then transforms it to clip
space.
468 Part 5: Engine Design with Shaders
Team LRN
Texture Coordinate Generation
One particular effect that is often used is environment mapping. The exact way this is done in
the Quake III Arena engine is, of course, not public knowledge, so the effect we will show here
is an approximation. It has to be noted that even the Q3A engine does not implement real envi-
ronment mapping (e.g., using cubemaps), but rather a crude approximation that uses skillfully
prepared texture and generats texture coordinates based on the viewing angle. This simple
effect is pretty nice and especially useful for implementing all kinds of chrome effects.
Our approximation works like this: Based on the viewing position, referred to as eye,
and the vertex position in world space, we compute the reflection vector; based on the result,
we generate texture coordinates dynamically. This gives the effect of shiny or chrome-like sur-
faces, which change along with the position of the player and the viewing angle.
The C++ code to compute the texture coordinates for our fake environment mapping is
something like this:
vec3_t dir = cam->eye-vec3_t(vertex_pos);
dir.normalize();
st[0]=dir[0]+normal[0];
st[1]=dir[1]+normal[1];
We basically take the direction vector from the vertex to the eye (camera position) in world
space, normalize it, and generate the texture coordinates by adding the normal of the surface to
the direction vector.
In the vertex shader setup phase, we process the custom shader definition and attempt to
generate the corresponding fragment of vertex shader code. In case of the environment map-
ping texture coordinate generation, we construct the code in the following way:
if((pass.flags & SPF_TCGEN) && pass.tcgen == SP_TCGEN_ENVIRONMENT)
{
vsh << "add r0, r0, -c11\n" // r0=vertex-eye
<< "dp3 r0.w,r0,r0\n" // normalize(r0)
<< "rsq r0.w,r0.w\n"
<< "mul r0,r0,r0.w\n" // r0=normalize(vertex-eye)
<< "add r1.x,r0.x,v1.x\n"
<< "add r1.y,r0.y,v1.y\n"
<< "mov r1.zw, c10.zw\n";
}
Note that this can be easily extended to support full environment mapping using cubemaps on
hardware that supports it. We would need to compute the reflection vector from within the ver-
tex shader and use it to generate the texture coordinates for the cubemap.
Texture Matrix Manipulation
The texture coordinates can be deformed to produce nice-looking effects. With the Quake
shader script, one can specify multiple texture coordinate modification steps, each applied on
top of another. These modification functions are usually time-based periodic functions with
various parameters controlling amplitude, frequency, etc. In our viewer, during the rendering
phase, we evaluate all these functions based on the current time value, producing a single
matrix per step. So if, for example, we had two modification steps specified in a shader script,
one that does the scaling and another that rotates the texture, we would end up with two matri-
ces for each step: a scaling matrix and a rotation matrix. These matrices are then concatenated
Designing a Vertex Shader-Driven 3D Engine for the Quake III Format 469
Team LRN
to produce one texture coordinates deformation matrix, which is then loaded to the vertex
shader. The actual modification of texture coordinates takes place inside the vertex program,
when we simply multiply original texture coordinates by the tcmod matrix, thus producing the
final effect.
Lets look at this in more detail. The following function is called once per frame per shader
pass, which involves texture coordinate modification:
static void
eval_tcmod(shader_t::pass_t& pass, const float time)
{
matrix_t t;
t.translate(+0.5f,+0.5f,0);
for(int i = pass.tcmods.size()-1; i >= 0; i)
{
pass.tcmods[i].eval(time);
t*=pass.tcmods[i].mat;
}
t.translate(-0.5f,-0.5f,0);
matrix_t::transpose(t,pass.tcmod_mat);
}
This function moves through all the texture modification steps for a particular shader pass and
produces a single concatenated matrix. The matrix is transposed in order to be used from
within the vertex shader. Later in the rendering process, when it comes to the actual drawing of
batched geometry data that uses this particular shader, the following piece of code updates the
constant memory of the vertex shader with the final texture modification matrix produced
above, prior to the actual DrawIndexedPrimitive() call:
const int pass_count=s.passes.size();
for(int i=0; i<pass_count; ++i)
{
const shader_t::pass_t& pass = s.passes[i];
[ omitted for clarity ]
// set vertex shader for this pass
d3dev->SetVertexShader(pass.vsh_ref);
// set tcmod matrix if needed
if(pass.flags&SPF_TCMOD)
{
d3dev->SetVertexShaderConstant(4,pass.tcmod_mat,4);
}
[ omitted for clarity ]
}
To put it all together, lets step back to the setup code, which creates a vertex shader for each
shader script. During the translation process, whenever we encounter the pass that involves
texture coordinate modification, we add the following piece of vertex shader code:
if(pass.flags&SPF_TCMOD)
{
vsh << "dp4 " << unit << ".x,r1,c4\n"
<< "dp4 " << unit << ".y,r1,c5\n";
}
470 Part 5: Engine Design with Shaders
Team LRN
This generates the code to transform the texture coordinates of the vertex by our matrix (actu-
ally, the first two vectors from it), which had been set up at constant memory locations 4
through 8. So the final vertex shader code looks like this, assuming were dealing with the first
texture unit:
dp4 oT0.x, r1, c4
dp4 oT0.y, r1, c5
This results in transforming the texture coordinates by our concatenated texture modification
matrix. This is similar to setting the texture matrix when using a fixed programmable pipeline
and letting DirectX 8.0 handle the transformation.
Rendering Process
We will now describe the rendering pipeline of our level viewer, which is greatly simplified
thanks to vertex shaders. Normally, all kinds of effects require the CPU to touch each vertex
and modify the stream accordingly, which makes the code path rather complicated. Since we
are using vertex shaders to do that on GPU instead of CPU, our lower level pipeline boils
down to preparing the vertex stream, managing the rotating dynamic index buffer, and setting
up vertex constant memory.
The rendering process is divided roughly into higher and lower level pipeline code. Table
3 shows a summary of which tasks and functions belong to higher and lower level rendering
code:
Table 3: Rendering code levels
Level Main Functions
Higher level rendering pipeline code
Implementation in: world.cpp
n
Traverses BSP tree
n
Culls BSP nodes and leaves to view frustrum
n
Culls faces to view frustrum
n
Culls faces according to PVS tables
n
Sorts faces by shader and lightmap
n
Arranges faces by solid and translucent
n
Walks sorted, batched shader faces list and pre-
pares vertex constant memory data
Lower level rendering pipeline code
DirectX 8.0 specific driver
Implementation in: d3dev.cpp
n
Device initialization/shutdown
n
Texture upload management
n
Static/dynamic vertex and index buffer
management
n
Batch rendering and device state setup according
to shader pass parameters
The higher level rendering pipeline code is traversing the BSP tree and performing some gross
culling using the view frustrum and bounding boxes for the BSP nodes. At a finer level, each
face is tested using PVS data when it is considered a candidate for drawing. The result of the
higher level rendering pipeline code is a sorted list of faces, batched by shader and lightmap,
and the shader hash map is sorted by shader solidity and translucency. This is performed in
world_t::process_faces() function. The higher level rendering begins in world_t::render_faces()
Designing a Vertex Shader-Driven 3D Engine for the Quake III Format 471
Team LRN
function, where the sorted shader batches are visited and vertex constant memory data is pre-
pared, if needed. This includes texture coordinate modification matrix, RGB generation
functions, etc. When the current batch is ready to be rendered, the control is passed to the
lower level rendering code.
The lower level code is a DirectX 8.0 driver that does the actual rendering of the triangles.
Its main rendering loop consists of managing vertex and index buffers and making low-level
calls for drawing primitives. A single dynamic index buffer is used to prepare batched up verti-
ces for drawing, according to the list passed in from the higher level rendering code. For
hardware vertex processing, all level geometry is uploaded into one static vertex buffer. This
gives significant performance benefits, since the data is effectively cached on GPU and does
not need to travel across the bus. For software vertex processing, this is not possible due to
performance reasons, so I have provided a fallback mechanism, where a fixed-size dynamic
vertex buffer is managed and vertex stream data is uploaded dynamically prior to rendering
calls. The reason behind this stems from the fact that when calling the DrawIndexedPrimitive()
API of DirectX, the range of vertices that needs to be processed is interpreted differently
depending on whether were using hardware or software vertex processing. In hardware mode,
vertices are transformed and lit according to the indices passed to the call, so the range of verti-
ces specified does not influence the process. We can therefore use a single static vertex buffer
that is spatially addressed by indices from our dynamic index buffer. In software mode, how-
ever, the optimized processor-specific pipeline transforms exactly the range of vertices that
were specified during the call. Thus, it is significantly more beneficial for software vertex pro-
cessing pipeline to batch vertices for drawing using a single dynamic vertex buffer so that the
range of vertices is continuous and always corresponds exactly to the number of vertices actu-
ally being drawn.
Assuming hardware vertex processing and a single static vertex buffer with level geometry
uploaded, the actual rendering loop is extremely simple. All the low-level driver does is man-
age the index buffer by filling it up with the current batch triangle indices, set up device state
according to shader pass parameters (culling, blending, z-buffering writes), set up vertex pro-
gram constant memory, initialize texture units with textures, set up current vertex shader, and
make actual calls to DrawIndexedPrimitive. All this is performed in the d3dev_t::ren-
der_faces() function.
Summary
It should be noted that the vertex shaders that result from the translation process do not provide
the full implementation of the Quake III shader scripts in a way that no CPU work is required
for their implementation during rendering process. Because of the great amount of combina-
tions of vertex shaders that are possible as a result of a single Quake III shader script, as well
as size limitations imposed on the current vertex shader implementation, the ideal situation of
having vertex shaders that implement the Quake III shader script in its entirety (sometimes
involving evaluation of a couple of sin functions) is extremely hard to achieve, if not impossi-
ble. For performance reasons, the intention was to eliminate the need for CPU to touch every
vertex in order to perform given shader script functions, such as texture coordinate modifica-
tion or vertex deformation. This was required in the traditional model, before vertex shaders
were available. The fixed rendering pipeline of DirectX does not allow developers to
472 Part 5: Engine Design with Shaders
Team LRN
customize vertex processing according to their needs. Thanks to programmable pipeline and
vertex shaders resulting from translation process, geometry data once uploaded to the card
never needs to be touched by the CPU again, as the GPU makes all the changes to it. The
geometry can therefore be treated as static, even though we are doing changes to individual
components of the vertex stream. Using vertex shaders implemented in hardware, the stream
data does not need to travel across the bus. The balance was achieved in such a way that the
CPU performs more complicated tasks, such as evaluation time-based periodic functions, and
simply updates the constant memory of the vertex shader in each frame, greatly reducing the
complexity of the vertex shader.
Designing a Vertex Shader-Driven 3D Engine for the Quake III Format 473
Team LRN
Glossary
anisotropic Exhibiting different properties when measured from different directions.
Directional dependence. For instance, anisotropic lighting takes into account direction
with respect to the surface of incoming and/or outgoing light when illuminating a surface.
anisotropic filtering With 2D textures, anisotropic filtering is thought of as a filter along a
line across the texture with a width of two and length up to the maximum degree of
anisotropy. In three dimensions, this increases to a slab with width two extending in the
other two dimensions up to the maximum degree of anisotropy. 3D anisotropy is not
supported by current accelerators.
anisotropic lighting See anisotropic.
basis vectors A (complete) set of line or row vectors for a matrix; these might be the three
vectors that represent a rotation matrix.
bilerping/bilinear filtering An image interpolation technique that is based on the averaging
of the four nearest pixels in a digital image. It first computes a texel address, which is
usually not an integer address, and then finds the texel whose integer address is closest to
the computed address. After that, it computes a weighted average of the texels that are
immediately above, below, to the left of, and to the right of the nearest sample point.
binormal To realize per-pixel lighting, a texture space coordinate system is established at
each vertex of a mesh. With the help of this texture space coordinate system, a light vector
or any other vector can be transformed into texture space. The axes of the texture space
coordinate system are called tangent, binormal, and normal.
BRDF Bidirectional Reflectance Distribution Functions. The BRDF is a generalization of all
shading models, such as Lambert, Blinn, or Phong. A shading model is an analytic
approximation of a surfaces reflective properties: It describes how incoming light is
reflected by a surface. The simplest shading model, Lambert shading, says that a surface
reflects incoming light equally in all directions. Phong shading (see Phong shading), on the
other hand, has an additional additive specular term, which permits a surface to reflect
different amounts of light in different directions. Thus, each of these models is a formula
which computes, for a given incoming and a given outgoing light direction, the portion of
incoming light that gets reflected in the outgoing direction. Such functions are called
BRDFs.
bump map Bump mapping adds surface detail (bumpiness) to objects in 3D scenes without
adding more geometry than already exists. It does so by varying the lighting of each pixel
according to values in a bump map texture. As each pixel of a triangle is rendered, a
lookup is done into a texture depicting the surface relief (aka the bump map). The values
in the bump map are then used to perturb the normals for that pixel. The new
bumped-and-wiggled normal is then used during the subsequent color calculation, and the
end result is a flat surface that looks like it has bumps or depressions in it. Usually a
normal map (see normal map) contains the already perturbed normals. In this case, an
additional bump map is not necessary.
474 Glossary
Team LRN
canonical Very common. With respect to faces, it means a face with very common features.
cartoon shading The use of any class of rendering approaches that are designed to mimic
hand-drawn cartoon artwork. Examples include pen-and-inking, pencil sketch, pastel, etc.
convex hull The smallest convex region that can enclose a defined group of points.
cube map Cube mapping has becoming a common environment mapping technique for
reflective objects. Typically, a reflection vector is calculated either per-vertex or per-pixel
and is then used to index into a cube map that contains a picture of the world surrounding
the object being drawn. Cube maps are made up of six square textures of the same size,
representing a cube centered at the origin. Each cube face represents a set of directions
along each major axis (+X, X, +Y, Y, +Z, Z). Think of a unit cube centered about the
origin. Each texel on the cube represents what can be seen from the origin in that
direction.
The cube map is accessed via vectors expressed as 3D texture coordinates (S, T, R or
U, V, W). The greatest magnitude component, S, T, or R, is used to select the cube face.
The other two components are used to select a texel from that face.
Cube mapping is also used for vector normalization with the light vector placed in the
texture coordinates.
dependent read A read from a texture map using a texture coordinate that was calculated
earlier in the pixel shader, often by looking up the coordinate in another texture. This
concept is fundamental to many of the advanced effects that use textures as transfer
functions or complex functions like power functions.
diffuse lighting Diffuse lighting simulates the emission of an object by a particular light
source. Therefore, you are able to see that the light falls onto the surface of an object from
a particular direction by using the diffuse lighting model.
It is based on the assumption that light is reflected equally well in all directions, so the
appearance of the reflection does not depend on the position of the observer. The intensity
of the light reflected in any direction depends only on how much light falls onto the
surface.
If the surface of the object is facing the light source, which means it is perpendicular to
the direction of the light, the density of the incident light is the highest. If the surface is
facing the light source under some angle smaller than 90 degrees, the density is
proportionally smaller.
The diffuse reflection model is based on a law of physics called Lamberts Law, which
states that for ideally diffuse (totally matte) surfaces, the reflected light is determined by
the cosine between the surface normal N and the light vector L.
directional light A directional light is a light source in an infinite distance. This simulates
the long distance the light beams have to travel from the sun. In game programming, these
light beams are treated as being parallel.
displacement map A texture that stores a single value at every texel, representing the
distance to move the surface point along the surface normal.
edge detection A type of filter that has a strong response at edges in the input image and a
weak response in areas of the image with little change. This kind of filter is often used as
the first step in machine vision tasks which must segment an image to identify its contents.
Glossary 475
Team LRN
EMBM Environment Mapped Bump Mapping. Introduced with DirectX6, this is a specific
kind of dependent lookup (see dependent read) allowing values of texels in one texture to
offset the fetching of texels in another texture, often an environment map. There is
spherical, dual paraboloid, and cubic environment mapping.
Fresnel term Common real time environment mapping reflection techniques reflect the
same amount of light on matter depending how the surface is oriented relative to the
viewer, much like a mirror or chrome. Many real materials, such as water, skin, plastic,
and ceramics, do not behave in this way. Due to their unique physical properties, these
materials appear to be much more reflective edge-on than they are when viewed head-on.
The change in reflectance as a function of the viewing angle is called the Fresnel effect.
Gram-Schmidt orthonormalization Given a set of vectors making up a vector space,
Gram-Schmidt orthonormalization creates a new set of vectors which span the same vector
space, but are unit length and orthogonal to each other.
half vector The half-way vector H is useful to calculate specular reflections (see specular
lighting). It was invented by James F. Blinn to prevent the expensive calculation of a
reflection vector (mirror of light incidence around the surface normal). The H vector is
defined as halfway between L and V. H is therefore:
H = (L + V) / 2
When H coincides with N, the direction of the reflection coincides with the viewing
direction V and a specular highlight is observed. This specular reflection model is called
the Blinn-Phong model.
isotropic Exhibiting the same properties when measured from different directions.
Directional independence.
Lambert model See diffuse lighting.
light reflection vector A vector, usually called R, that describes the direction of the
reflection of light.
light vector A vector that describes the light direction.
LUT Look-up table. A precomputed array of values used to replace a function evaluation
often for the sake of performance. Many pixel shader techniques use 1D and 2D textures
as lookup tables to calculate functions that are prohibitive or even impossible to calculate
given the limited size and scope of the pixel shader instruction set and size.
mip mapping Mip maps consist of a series of textures, each containing a progressively
lower resolution of an image that represents the texture. Each level in the mip map
sequence has a height and a width that is half of the height and width of the previous level.
The levels could be either square or rectangular. Using mip maps ensures that textures
retain their realism and quality as you move closer or further away.
normal map A normal map stores vectors that represents the direction in which the normal
vector (see normal vector) points. A normal map is typically constructed by extracting
normal vectors from a height map whose contents represent the height of a flat surface at
each pixel.
normal vector Normals are vectors that define the direction that a face is pointing or the
visible side of a face.
476 Glossary
Team LRN
orthonormal Two vectors are considered to be orthonormal if both vectors are unit length
and have a dot product of 0 between them.
per-pixel lighting For example, diffuse (see diffuse lighting) or specular (see specular
lighting) reflection models calculated on a per-pixel level are called per-pixel lighting.
Per-pixel lighting gets the per-pixel data usually from a normal map, which stores vectors
representing the direction in which the normal vector (see normal vector) points. The
additional needed light (see light vector) and/or half angle vector(s) (see half vector) are
usually transformed via a texture space coordinates system, that is, build per-vertex into
texture space so that the normal vector provided by the normal map and these vectors are
in the same space.
per-vertex lighting Per-vertex lighting means that the actual lighting calculations are done
for each vertex of a triangle, as opposed to each pixel that gets rendered (see per-pixel
lighting). In some cases, per-vertex lighting produces noticeable artifacts. Think of a large
triangle with a light source close to the surface. As long as the light is close to one of the
vertices of the triangle, you can see the lighting effect on the triangle. When the light
moves toward the center of the triangle, then the triangle gradually loses the lighting
effect. In the worst case, the light is directly in the middle of the triangle and there is
nearly no light visible on the triangle, instead of a triangle with a bright spot in the middle.
That means that a smooth-looking object needs a lot of vertices or in other words, a
high level of tessellation; otherwise, the coarseness of the underlying geometry is visible.
Phong shading Phong shading is done by interpolating the vertex normals across the surface
of a polygon or triangle and illuminating the pixel at each point, usually using the phong
lighting model. At each pixel, you need to re-normalize the normal vector and also
calculate the reflection vector.
point light A point light source has color and position within a scene but no single direction.
All light rays originate from one point and illuminate equally in all directions. The
intensity of the rays will remain constant regardless of their distance from the point source,
unless a falloff value is explicitly stated. A point light is useful to simulate light bulb.
post-classification (in volume graphics) The mapping of volume data to RGBA tuples after
interpolating, filtering, and/or sampling of the volume data.
power of 2 texture A texture that has dimensions that fall into the set 1, 2, 4, 8, 16, 32, 64,
128, 256, 512, 1024, 2048, 4096. Some hardware have these limitation, or it may be more
efficient for hardware to deal with these dimensions.
pre-classification (in volume graphics) The mapping of volume data to RGBA tuples
before interpolating, filtering, and/or sampling of the volume data.
pre-integrated classification (in volume graphics) The mapping of (small) line segments
of volume data to RGBA tuples with the help of a pre-computed look-up table.
quadrilinear filtering Adds a linear filter along the fourth dimension of the volumetric mip
maps consisting of 3D textures, sampling from 16 different texels.
reflection The return of light from a surface.
reflection vector The mirror of a vector around a surface normal. In lighting equations, often
substituted by the half-vector (see half-vector).
Glossary 477
Team LRN
refraction Refraction is a phenomenon that simulates the bending of light rays through
semi-transparent objects. There are several properties defined with refraction, Snells law,
and the critical angle. Snells law states that for a light ray going from a less dense medium
to a higher dense medium, the light ray will bend in one direction, and going from a higher
density medium to a lower density medium, it will bend in the other direction. These
properties are known as the refraction index, and it is a ratio of the speed of light through
one medium divided by the speed of light through the other medium.
render target The buffer in memory to which a graphics processor writes pixels. A render
target may be a displayable surface in memory, like the front or back buffer, or the render
target may be a texture.
renderable texture The render target (see render target) is a texture.
scientific data visualization The branch of computer graphics that is concerned with the
(interactive) generation of comprehensible, informative, accurate, and reliable images
from scientific or technical data.
shadow volume The volume of space that forms a shadow for a given object and light.
Objects inside this volume of space are in shadow. Rapidly determining whether
something is in shadow can be done by using the stencil buffer and a vertex shader.
skinning A technique for blending several bone matrices at a vertex level to produce a
smooth animation.
slab (in volume graphics) The space between two consecutive slices through volume data.
The interior of the slab is considered to be filled with the volume data.
slice (in volume graphics) A flat subset of volume data, often implemented as a polygon
that is textured with a three-dimensional texture.
specular lighting Compared to the diffuse reflection (see diffuse lighting) model, the
appearance of the reflection depends on the specular reflection model on the position of
the viewer. When the direction of the viewing coincides, or nearly coincides, with the
direction of specular reflection, a bright highlight is observed. This simulates the reflection
of a light source by a smooth, shiny, and polished surface. To describe reflection from
shiny surfaces, an approximation is commonly used, which is called the Phong
illumination model (not to be confused with Phong shading), named after its creator Phong
Bui Tong. According to this model, a specular highlight is seen when the viewer is close to
the direction of reflection. The intensity of light falls off sharply when the viewer moves
away from the direction of the specular reflection.
spotlight A light source in which all light rays illuminate in the shape of a cone. The falloff
(the attenuation in the intensity of a light source as the distance increases), spread (the
parameter that controls the width of the cone of light produced by the spotlight), and
dropoff (the parameter that controls the way the light intensity fades based on its distance
from the center of the light cone) of a spotlight are adjustable.
SSE Streaming SIMD Extensions. A set of registers and instructions in Pentium III and
Pentium 4 processors for processing multiple floating-point numbers per instruction.
tangent space Tangent space is a 3D coordinate system defined for every point on the
surface of an object. It can be thought of as a transform from object space into texture
space. The x and y axes are tangent to the surface at the point and are equal to the u and v
478 Glossary
Team LRN
directions of the texture coordinates, respectively. The z-axis is equal to the surface
normal. Tangent space is used for a variety of per pixel shaders due to the fact that most
bump maps and other types of maps are defined in tangent space. Tangent space is
sometimes called the Frenet Frame or a Surface Normal Coordinate Frame.
tap Each sample that contributes to a filter result.
Taylor series A Taylor series allows the generation of approximate sinus and cosinus values
without being computationally as expensive as the sinus and cosinus functions.
texels Abbreviation for texture elements, usually meaning 1 RGB or ARGB component of
the texture.
texture perturbation With the ability to perform per-pixel math both before and after
texture fetches, a variety of new texture-based effects are possible. Per-pixel texture
perturbation effects, effects that use the results of a texture fetch to change the texture
coordinates used to fetch another texel (see dependent read).
texture space See tangent space.
topology The arrangement in which the nodes of a 3D object are connected to each other.
transfer function A function (often one-dimensional) which is used to enhance or segment
an input image. Transfer functions are often used in scientific visualization to enhance the
understanding of complex datasets. A transfer function might map a scalar, such as heat,
pressure, or density, to a color to aid in understanding the phenomenon being drawn. In
games, transfer functions can be used on images to stylize them or to simulate effects like
night-vision goggles or heat-sensitive displays. In volume graphics, this is the mapping of
a number to an RGBA tuple; it is used in volume visualization and volume graphics to
assign colors to numeric data.
trilinear filtering In case of 2D textures, this means that multiple linear filtered mip maps
are being blended together. In case of a 3D texture, trilinear filtering performs a linear
filtering along all three axes of the texture, sampling colors from eight different texels. The
filtering method that is called trilinear filtering in conjunction with 2D textures is called
quadrilinear filtering for 3D textures.
trinary operator Operation that can use three components and do two operations
simultaneously.
tweening A technique for interpolating between two or more key frames to produce a
smooth animation.
unit vector A unit vector is a vector with the length 1. To calculate a unit vector, divide the
vector by its magnitude or length. The magnitude of vectors is calculated by using the
Pythagorean theorem:
x
2
+ y
2
+ z
2
= m
2
The length of the vector is retrieved by:
||A|| = sqrt(x
2
+ y
2
+ z
2
)
The magnitude of a vector has a special symbol in mathematics. It is a capital letter
designated with two vertical bars: ||A||. So dividing the vector by its magnitude is:
UnitVector = Vector / sqrt(x
2
+ y
2
+ z
2
)
vector field map A texture that stores a vector at every texel to move the surface point.
Glossary 479
Team LRN
volume data A data set that is indexed by three coordinates, often implemented as
three-dimensional textures (volume textures).
volume graphics The branch of computer graphics that is concerned with the rendering of
volumetric (as opposed to polygonal) primitives.
volume visualization The branch of scientific data visualization that is concerned with the
visualization of volume data.
480 Glossary
Team LRN
About the Authors
Philippe Beaudoin ([email protected])
Philippe recently completed a masters thesis in computer graphics at the University of
Montreal where he developed algorithms for the spreading and rendering of 3D fire
effects. As part of his job as a hardware architect at Matrox Graphics, he studied and
developed various vertex and pixel shader technologies. He is currently employed at Dig-
ital Fiction where he develops games for various next-generation consoles.
Steffen Bendel ([email protected])
Steffen started his career as a game programmer in 1996 at a company called BlueByte.
After that, he studied physics with a specialty in quantum optics at the university in
Rostock. While studying, he worked as a freelancer for Massive Development where he
wrote some parts of the 3D engine of the underwater game AquaNox.
Chris Brennan ([email protected])
Chris graduated with a BS degree in Computer Science and another BS degree in Electri-
cal Engineering from Worcester Polytechnic Institute in 97 and joined Digital
Equipment Corps Workstation Graphics group doing hardware design and verification.
When Digital died, Chris joined ATI as a 3D ASIC designer for the RADEON line of
graphics chips and then moved over to the 3D Application Research Group where he
tries to get those chips to do things that were not originally thought possible.
Dean Calver ([email protected])
Games are fun; Dean figured that out at age 2 and has spent the following years working
out how to make better games. For the last five years, people have even paid him to do it.
Having no real preference for console or PC has meant a mixed career flipping between
them for every project. Professionally, he started on a war game, did three years of a rac-
ing game, followed by an X-COM style game and arcade classic updates. Currently hes
doing a 3D graphic adventure. He still studies various subjects including optics, mathe-
matics, and other geeky things for fun. This preoccupation with learning means that he
has been taking exams every year for over half his life. At least hell be ready to write the
first game for a quantum computer.
Drew Card ([email protected])
Drew is currently a software engineer in the 3D Application Research Group at ATI
Research where he is focusing on the application of shader-based rendering techniques.
He has worked on SDK applications as well as helped out with the demo engine. Drew is
a graduate of the University of South Carolina.
Wolfgang F. Engel ([email protected])
Wolfgang is the author of Beginning Direct3D Game Programming and a co-author of
OS/2 in Team, for which he contributed the introductory chapters on OpenGL and DIVE.
Wolfgang has written several articles in German journals on game programming and
About the Authors 481
Team LRN
many online tutorials that have been published on www.gamedev.net and his own web
site, www.direct3d.net. During his career in the game industry, he built two game devel-
opment units from scratch with four and five people that published, for example, six
online games for the biggest European TV show Wetten das..? As a member of the
board or as a CEO in different companies, he has been responsible for several game
projects.
Ingo Frick ([email protected])
Ingo is co-founder and technical director of Massive Development GmbH. He has played
a leading role in the development of the Krass engine, Archimedean Dynasty
(Schleichfahrt), and AquaNox. In the mid-80s he developed several games (C64, Amiga,
PC) that were mainly distributed by smaller publishers and magazines. His first success-
ful commercial product was the conversion of The Settlers from the Commodore Amiga.
Ingo has a PhD in the area of numerical simulation of the motion of granular media.
David Gosselin ([email protected])
Dave is currently a software engineer in the 3D Application Research Group at ATI
Research. He is involved in various demo and SDK work, focusing mainly on character
animation. Previously, he worked at several companies, including Oracle, Spacetec IMC,
Xyplex, and MIT Lincoln Laboratory on varied projects from low-level networking and
web technologies to image processing and 3D input devices.
Juan Guardado ([email protected])
Juan currently works at Matrox as a graphics architect. He started as the first developer
support engineer in the days before DirectX. Later, his focus shifted to better understand-
ing the requirements of next-generation games and APIs, using this analysis to direct
research projects and argue with his boss.
Evan Hart ([email protected])
Evan is presently a software engineer with ATIs Application Research Group. He works
with new hardware features on all levels from API specification to implementation in
applications. He has been working in real-time 3D for the past four years. He is a gradu-
ate of Ohio State University.
Matthias Hopf
Matthias graduated with a degree in computer science from the FAU Erlangen in Ger-
many. Right now, he is a PhD student in the Visualization and Interactive Systems Group
at the University of Stuttgart. Despite all his research on adaptive and hierarchical algo-
rithms as well as on hardware-based filters, he is still interested in highly tuned low-level
software and hardware. He is mainly known for still being the maintainer of the Aminet
archive mirror at the FAU Erlangen, collecting gigabytes of Amiga software.
Kenneth L. Hurley ([email protected])
Kenneth started his career in the games industry in 1985 with a company called
Dynamix. He has also worked for Activision, Electronic Arts, and Intel and now works in
developer relations at NVIDIA Corp. His current job includes research and development
and instructing developers how to use new technology from NVIDIA. His credits in the
game industry include Sword of Kadash (Atari ST), Rampage (PC, Amiga, Apple II),
482 About the Authors
Team LRN
Copy II ST, Chuck Yeagers Air Combat Simulator (PC), The Immortal (PC), and Wing
Commander III (Playstation). While at NVIDIA, he has contributed the following pack-
ages/demos: NVASM (GeForce3 vertex/pixel shader assembler), NVTune (NVIDIAs
performance analysis tool set), DX7 Refract demo, Minnaert Lighting demo, Particle
Physics demo, and the Brushed Metal effect.
John Isidoro ([email protected])
John is a member of the 3D Application Research Group at ATI Technologies and a grad-
uate student at Boston University. His research interests are in the areas of real-time
graphics, image-based rendering, and machine vision.
Greg James ([email protected])
Greg is a software engineer with NVIDIAs technical developer relations group where he
develops tools and demos for real-time 3D graphics. Prior to this, he worked for a small
game company and as a research assistant in a high-energy physics laboratory. He is very
glad to have avoided graduate school and even happier to be working in computer graph-
ics, which he picked up as a hobby after his father brought home a strange beige Amiga
1000.
Martin Kraus ([email protected])
Since graduating in physics, Martin is a PhD student in the Visualization and Interactive
Systems Group at the University of Stuttgart in Germany. In recent years he has pub-
lished several papers on volume visualization, but he is still best known for his Java
applet LiveGraphics3D. Martin started programming on a C64 in his early teens and
quickly became addicted to computer graphics. Major goals in his life include a long
vacation after receiving his PhD and achieving a basic understanding of quantum
mechanics.
Scott Le Grand ([email protected])
Scott is a senior engineer on the Direct3D driver team at NVIDIA. His previous commer-
cial projects include BattleSphere for the Atari Jaguar and Genesis for the Atari ST. Scott
has been writing video games since 1971 when he played a Star Trek game on a main-
frame and was instantly hooked. In a former life, he picked up a BS in biology from
Siena College and a PhD in biochemistry from Pennsylvania State University. Scotts
current interests are his wife, Stephanie, and developing techniques to render planets in
real time.
Jason L. Mitchell ([email protected])
Jason is the team lead of the 3D Application Research Group at ATI Research, makers of
the RADEON family of graphics processors. Working on the Microsoft campus in
Redmond, Jason has worked with Microsoft to define new Direct3D features, such as the
1.4 pixel shader model in DirectX 8.1. Prior to working at ATI, Jason did work in human
eye tracking for human interface applications at the University of Cincinnati where he
received his masters degree in Electrical Engineering. He received a BS in Computer
Engineering from Case Western Reserve University. In addition to several articles for this
book, Jason has written for the Game Programming Gems books, Game Developer Mag-
azine, Gamasutra.com, and academic publications on graphics and image processing.
About the Authors 483
Team LRN
dm Moravnszky ([email protected] or http://n.ethz.ch/student/adammo/)
Adam is a computer science student at the Swiss Federal Institute of Technology. He is
looking forward to receiving his masters degree in April 2002, after finishing his thesis
on real-time 3D graphics.
Christopher Oat ([email protected])
Christopher is a software engineer in the 3D Application Research Group at ATI
Research where he explores novel rendering techniques for real-time 3D graphics appli-
cations. His current focus is on pixel and vertex shader development in the realm of PC
gaming.
Kim Pallister ([email protected])
Kim is a technical marketing manager and processor evangelist with Intels Software and
Solutions Group. He is currently focused on real-time 3D graphics technologies and
game development.
Steven Riddle ([email protected])
Steven is an independent contractor developing various web and 3D applications. Since
he started programming on the C64, his main interest has been graphics. It has taken him
from writing the first Sega Genesis emulator to writing his own rendering engine in C
and Assembly. Currently, he is focusing on using vertex and pixel shaders to create a vir-
tual universe.
Guennadi Riguer ([email protected])
Guennadi is a software developer at ATI Technologies, where he is helping game engine
developers to adopt new graphics technologies. Guennadi holds a degree in Computer
Science from York University and has previously studied at Belorussian State University
of Computing and Electronics. He began programming in the mid-80s and worked on a
wide variety of software development projects prior to joining ATI.
John Schwab ([email protected], www.shaderstudio.com)
John has been programming for the better part of his life. He is the creator of Shader Stu-
dio, but in the past ten years, he has been developing games professionally and has
worked for Psygnosis and Microsoft. His specialty is computer graphics, and he is
always trying to do what others dont think is possible. Currently, he is software engineer
at Electronic Arts. In his spare time, he builds robots and designs electronic devices.
Bart Sekura ([email protected])
Bart is a software developer with over seven years of professional experience. He loves
computer games and enjoys writing 3D graphics-related code. He spends most of his
time tinkering with DirectX, locking and unlocking vertex buffers, and transposing matri-
ces. Bart is currently a senior developer for People Can Fly working on Painkiller, a
next-generation technology action shooter.
Alex Vlachos ([email protected] or http://alex.vlachos.com)
Alex is currently part of the 3D Application Research Group at ATI Research, where he
has worked since 1998 focusing on 3D engine development. Alex is one of the lead
developers for ATIs graphics demos and screen savers, and he continues to write 3D
484 About the Authors
Team LRN
engines that showcase next-generation hardware features. In addition, he has developed
N-Patches (a curved surface representation which is part of Microsofts DirectX 8). Prior
to working at ATI, he worked at Spacetec IMC as a software engineer for the SpaceOrb
360, a six degrees-of-freedom game controller. He has also been published in Game Pro-
gramming Gems 1, 2, and 3 and ACMs I3DG. Alex is a graduate of Boston University.
Daniel Weiskopf ([email protected])
Daniel is a researcher at the Visualization and Interactive Systems Group at the Univer-
sity of Stuttgart. His scientific interests range from figuring out how the latest graphics
boards can be used to speed up scientific visualization to rather crazy things like visualiz-
ing general relativistic faster-than-light travel in a warp spaceship. Daniel received a PhD
in Theoretical Astrophysics from the University of Tuebingen.
Oliver C. Zecha ([email protected])
Oliver is an independent consultant with five years experience in the real-time 3D graph-
ics field. His primary focus is dynamic physical simulation, for which he has received
several awards and accolades. Recently, he migrated from OpenGL to Direct3D in order
to utilize programmable hardware shaders. At the time of publication, his research
involves the design of new algorithms that utilize consumer grade graphics hardware in
creative and unconventional ways, as well as implementing them for the X-Box console.
About the Authors 485
Team LRN
Team LRN
Index
3D model, converting photographs into, 297
3D noise, 431
3D Studio Max, 16, 296
3D texture lookup table shader, 430-431
3D textures, 428
filtering, 429
storing, 429-430
using, 430-437
A
add instruction, 23, 101-102
address register (vertex shader), using, 29
alpha blending, 78-79
alpha test, 78, 331
amorphous volume, representing with 3D textures, 435
anisotropic
filtering, 429, 451
lighting model, 376
strand illumination shaders, 379, 380, 381-382
arithmetic address instructions (pixel shader), 128
arithmetic instruction set, ps.1.1-ps.1.4, 101-108
array-of-structures vs. structure-of-arrays, 218-219
ATI ShadeLab, 80-81
atmospherics, creating with 3D textures, 435-436
attenuation, 66-67, 432-434 see also light falloff
B
backface culling, 73-74
backward mapping, 418-419
bem instruction, 102-103
Bzier patch, 65
Bzier patch class, using in vertex shader example, 59-66
bias, 114
Bidirectional Reflectance Distribution Functions, see BRDF
black and white transfer function, 260
Blinn half-vector, using with Phong shading, 400
Blinn-Phong reflection, 63-64
BRDF, 406-407
shaders, 408-409
bubble shaders, 370-375
bubbles, rendering, 369-370
bump mapping, 410-411, 450
shaders, 410-411
bump maps, 130
extracting with Paint Shop Pro, 299
separating with Paint Shop Pro, 299-300
C
cartoon shading, 322-323
shaders, 323-324
character animation, using shadow volumes with, 192-193
clip planes, 74
clipping, 74-75
cloud effects, 337-341
shaders, 338, 339, 341
clouds, rendering, 337-341
cmp instruction, 103-104
cnd instruction, 105
color map shader, ps.1.1-1.3, 128
color registers (pixel shader), 83
common files framework, 40-41
complex instructions (vertex shader), 27
compression
of tweened animation, 206-207
shader, 206-207
compression transform, 176-179
ConfirmDevice() function, 42
constant instruction (pixel shader), 128
constant registers (pixel shader), 82
constant registers (vertex shader), 17
setting, 22, 43, 54-55
using, 29
CreateFile() function, 52
CreateFileMapping() function, 52
CreatePixelShader() function, 120, 135
CreatePSFromCompiledFile, 134-135
CreateVertexShader() function, 34, 46
CreateVSFromCompiledFile() function, 51
cross products, 230
crystal shaders, 364-368
crystal texture, rendering, 363-364
cube mapping, 287
cube maps, 287
diffuse, 287
generating, 313
cubic environment mapping, 290
cubic interpolation, 240
custom shaders, constructing, 466-467
D
D3DXAssembleShader() function, 45-46
using for vertex shader example, 38-49
data arrangement, 218-219
deformable geometry, 467-468
DeletePixelShader() function, 36, 121
Index 487
Team LRN
DeleteVertexShader() function, 47
depth precision, 331
depth test, 78
destination register modifiers, 117-118
detail textures, 451
development system recommendations, 5
diffuse and specular point light shader, 68-69
diffuse and specular reflection shader, 61-62
diffuse bump mapping shader, 117-118
diffuse cube maps, 287
generating dynamically, 288-289
shader, 287-288
using, 287-288
diffuse lighting, 298
shader, 55
diffuse reflection, 57-59
shader, 144
diffusion cube map tool, 14-15
dilation, 265-266
shader, 266-267
Direct3D pipeline, 5-6, 73
directional light, 57
displacement maps, 184-186
dithering, 79
DLL Detective, 15
dp3 instruction, 23, 106
dp4 instruction, 23, 106-107
dst instruction, 23
E
ease interpolation, 241
edge detection, 262
filters, 262-265
edge determination shaders, 329-330
edges, compositing, 330
engine design considerations, 450-452
environment mapping, 290-292
artifacts, 290
cubic, 290
for eyes, 315-316
shaders, 293-294, 315-316
with Quake III shader, 469
environmental lighting, 312
shader, 314
erosion, 267
shader, 267-268
expp instruction, 24, 27
eye mapping shaders, 315-316
eye space depth shader, 328-329
F
faces, rendering, 296
facial animation, 317
filter kernels, 261
texture coordinates, 261-262
filtering, 429
fire effect, 342-344
shader, 343-344
fixed-function pipeline, 6
mapping, 20
flow control, 230
fog, 77-78
font smoothing, 270-272
shaders, 271-272
forward mapping, 418
fractional Brownian motion, 236-239
three-dimensional, 238
FrameMove() function, 54-55, 60
frc instruction, 27
frequency, using to compare noise waves, 235
Fresnel effects, 281-282
Fresnel shaders, 283-285
Fresnel term, 281
frustum clipping, 74-75
full head mapping, 309-312
G
geometry emulation, 273-274
shaders, 274-276
gmax, 16
Gooch lighting, 326-327
shaders, 327
Gram-Schmidt orthonormalization, 380-381
grass,
animating, 334
lighting, 334-335
shaders, 335-336
guard band clipping, 74
H
half angles, 303
hard shading, 322
hardware vertex shader, 17
hatching, 324
shaders, 325-326
HCLIP space, 186
heat signature
shader, 260
transfer function, 261
hyperspace, 186
I
illumination,
per-pixel strand-based, 378-382
strand-based, 376-377
image rendering, 328-329
imposters, 273
488 Index
Team LRN
independent rendering passes, 451-452
index buffer, using, 49
inflow, 421-423
input registers (vertex shader), 17
using, 28
instruction modifiers, 116-118
instruction pairing, 119-120
instruction set, complex (vertex shader), 27
instruction set (vertex shader), 23-26
instructions, modifying, 110-111
invert, 114
K
krass engine, 453
development history, 454-455
ordering effects in, 456-457
particle rendering in, 460-461
rendering terrain in, 458-460
structure, 453-454
L
lake water,
rendering, 357
shaders, 358-362
Lambert shading, 406
Lamberts Law, 57-58
light factor scaling, 452
light falloff, see also attenuation
calculating, 203-206
creating with 3D textures, 436
shader, 204-205, 437
light, directional, 57
lighting,
calculating, 57
enhancing, 213-214
improving, 277-278
real-time, 450-451
shaders, 279-280
lighting elements, separating, 298
linear interpolation, 240-241
lit instruction, 24
log instruction, 27
logp instruction, 25
lookup table, using, 140-143, 145-146
lrp instruction, 107
luminance shader, 260
M
m3x2 instruction, 27
m3x3 instruction, 27
m3x4 instruction, 27
m4x3 instruction, 27
m4x4 instruction, 27
mad instruction, 25, 107-108
mapping head, 309-312
MapViewOfFile() function, 53
masking, 32, 117-118
mathematical morphology, 265
max instruction, 25
MFC pixel shader, 80
Microsoft pixel shader assembler, 79-80
Microsoft vertex shader assembler, 11
min instruction, 25
morphing, see tweening
morphology, using to thicken outlines, 332
mov instruction, 25, 108
mul instruction, 25, 108
multiple compression transforms, 180-182
multisampling, 76
N
noise, 234, 431
function, 234-235
one-dimensional, 234-236
Perlin, see Perlin noise
shader, 431-432
two-dimensional, 236-237
using to generate textures, 238-239
noise waves, 235
using frequency to compare, 235
non-integer power function, 383-384
non-photorealistic rendering, 319
nop instruction, 25, 108
normal, transforming, 56
normal map, 130, 301-302
creating for face, 300-301
Normal Map Generator, 13
N-Patches, 6
NVIDIA Effects Browser, 8
NVIDIA NVASM, 10
using for vertex shader example, 49-53
NVIDIA Photoshop plug-ins, 13-14
NVIDIA Shader Debugger, 9
NVLink, 12
O
ocean water,
rendering, 347
rendering using Fresnel term, 350-352
shaders, 348-356
offset textures, 419
one-dimensional noise function, 234-236
one-dimensional Perlin noise, 242
one-time effect, 229
on-the-fly compiled shaders, 120
optimization example (vertex shader), 224-227
optimizations
for creating photorealistic faces, 302
Index 489
Team LRN
vertex compression, 179-180
vertex shader, 221-224
outflow, 423
outline shader, 321
outlines,
rendering, 319-322
thickening, 332
output registers (pixel shader), 81, 82
output registers (vertex shader), 17
using, 30-31
P
Paint Shop Pro, 296
using to extract maps, 299
using to separate maps, 299-300
particle
motion, 416-420
systems, 414
particle flow
rendering, 414
shader, 423
particles, 416
rendering with krass engine, 460-461
Pentium 4 processor architecture, 216-217
periodic effects, randomizing, 229-230
Perlin noise, 234, 431 see also noise
calculating scalars for, 244-245
generating, 241-244
one-dimensional, 242
three-dimensional, 243-244
two-dimensional, 242-243
per-pixel Fresnel term, 281
per-pixel lighting, 130-134, 203-206
per-pixel strand-based illumination, 378-382
persistence, 236
Phong illumination, 62-63
Phong shading, 399-402, 406
shader, 402
using with Blinn half-vector, 400
photorealistic faces, 296
shaders, 303-309
Photoshop Compression plug-in, 13-14
Photoshop, using to extract lighting characteristics, 298
pipelines, 5-6
disadvantages of, 7
Pixar Animation Studios, 4
pixel shader, 72
architecture, 81-84
assembling, 120, 134
compiling, 120
creating, 120-121, 134-135
creating with Shader Studio, 159-160
defining constants for, 86, 127-128, 137
deleting, 121, 136
effects, 73
freeing resources of, 121, 136
programming, 84-121
range, 83-84
registers, 82-84
setting, 121, 136
setting texture flags for, 126-127, 140
setting texture of, 127, 140-141
standards, 77
tools, 79
pixel shader arithmetic instruction set, ps.1.1-ps.1.4, 101-108
pixel shader examples
using lookup table, 140-143
with diffuse reflection and specular reflection, 144-147
with directional light and diffuse reflection, 125-137
with higher precision specular reflection, 140-143
with specular reflection, 137-140
pixel shader instructions, 87-88, 137-139
ps.1.1-1.3, 128
ps.1.4, 128-130
pixel shader support, checking for, 84-85, 126
pixel shader texture addressing instruction set,
ps.1.1-ps.1.3, 89-97
ps.1.4, 98-100
pixel shader texture stages, setting, 85-86
pixel shader texture, setting, 86
pixel shaders
3D texture lookup table, 430-431
anisotropic strand illumination, 379, 380, 381-382
approximating power function, 392, 394, 396
blending textures, 257
bubble effect, 372-375
bump mapping, 411
cartoon shading, 323-324
cloud effects, 338, 339, 341
computing power function, 384-385, 386
converting to black and white, 260
creating photorealistic faces, 307-309
crystal effect, 365-368
determining edges, 329-330
diffuse bump mapping, 117-118
diffuse cube maps, 287-288
diffuse reflection, 144
dilation, 266-267
displaying color map, 128
environment mapping, 293-294
environmental lighting, 314
erosion, 267-268
eye environment mapping, 316
fire effects, 343-344
font smoothing, 271-272
Fresnel effects, 284-285
geometry emulation, 274
Gooch lighting, 327
490 Index
Team LRN
grass, 336
hatching, 326
heat signature effect, 260
lake water, 360-362
light falloff, 437
light rendering, 409
lighting, 279
noise, 431-432
ocean water, 350-352, 352-356
particle flow, 423
Phong shading, 402
plasma effects, 344, 345-346
Roberts cross gradient filter, 263
sepia effect, 260
smooth conditional function, 397-398
Sobel filter, 264-265
specular map, 142
specular power, 145, 146
specular reflection, 139
transfer function, 441, 442
transformation to texture space, 133
volume bounded lighting, 399
volume rendering, 443, 445
pixel shaders,
advantages of, 72
limitations of, 232-234
plasma glass
effect, 344-346
shaders, 344, 345-346
point light source, 66
point lights, attenuation for, 66-67
power function,
approximating in pixel shader, 391-396
computing, 384-387
mathematical background, 387-391
non-integer, 383-384
using, 396-402
precompiled shaders, 120
pre-integrated volume rendering, 438, 444-446
procedural textures, 415
programmable pipeline, 4
ps.1.4, advantages of, 77
Q
quad example shader, 43-45
quadrilinear filtering, 429
Quake III Arena shaders, see Quake III shaders
Quake III shaders, 463-464
categories of, 466
vertex shader effects, 467-471
quantization, 174-175
R
random numbers, generating, 229-230
rcp instruction, 25
read port limit, 109-110
real-time lighting, 450-451
reconstruction, 407-408
reflection,
Blinn-Phong, 63-64
diffuse, 57-59
specular, 62-65
reflection map, generating, 358
refraction map, generating, 358
render pipeline, 456-457
render target, 79, 258
rendering passes, independent, 451-452
rendering process with Quake III shader, 471-472
RenderMan, 4
repetition, avoiding, 295
Roberts cross gradient filter, 262-263
shader, 263
rsq instruction, 26
S
scale, 114
scaled offset, 175-176
scrolling, disadvantage of, 295
sepia effect shader, 260
sepia tone transfer function, 260
SetPixelShader() function, 121, 136
SetPixelShaderConstant() function, 86, 127, 137
SetTexture() function, 86, 127, 140-141
SetVertexShader() function, 34-35, 47
SetVertexShaderConstant() function, 22, 43, 61
sge instruction, 26
shader assembly program, 246-251
Shader City, 10
Shader Studio, 11, 149
browser, 165
coordinate system, 150
creating pixel shader with, 159-160
creating vertex shader with, 152-159
drawbacks of, 151
features, 150-151
files, 161
light source, 152
menu layout, 160
object control, 152
transformations, 165-166
Shader Studio dialogs
Configure, 168
Constant Properties, 163-164
Materials, 164
Settings, 167-168
Shaders, 161-163
Index 491
Team LRN
Statistics, 167
Textures, 165
shaders
in effect files, 120
limitations of, 232-234
shadow volume extrusion, 192-193
shader, 191-192
shadow volumes, 188-189
creating, 189-190
using with character animation, 192-193
shadows, outlining, 331-332
signed scaling, 114
silhouette edges, 319
detecting, 189
single-surface geometry, 209
single-surface object, lighting, 209-210
sinusoidal perturbation, 348-350
skeletal animation, see skinning
skinning, 197-199
blending with tweening, 199-201
shader, 198-199
sliding compression transform, 182-184
slt instruction, 26
smooth conditional function, 396-399
shader, 397-398
Sobel filter, 263-265
shader, 264-265
source register modifiers,
ps.1.1-1.3, 113, 115
ps.1.4, 115
specular map,
extracting with Paint Shop Pro, 299
separating with Paint Shop Pro, 299-300
shader, 142
specular power shader, 145, 146
specular reflection, 62-65
shaders, 138-139
spot light, 66
SSE, 217-218
data arrangement, 218-219
SSE registers, mapping VVM registers to, 220
stencil test, 78
strand-based illumination, 376-377
per-pixel, 378-382
Streaming SIMD Extensions, see SSE
structure-of-arrays vs. array-of-structures, 218-219
sub instruction, 108
swizzling, 31-32, 111-113
T
T&L pipeline, 6
disadvantages of, 7
tangent space, animating for per-pixel lighting, 202-203
temporary registers (pixel shader), 81, 82
temporary registers (vertex shader), 17
using, 30
terrain rendering with krass engine, 458-460
tex instruction, 89
texbem instruction, 89-90
texbeml instruction, 91
texcoord instruction, 91
texcrd instruction, 98
texdepth instruction, 99
texdp3 instruction, 91
texdp3tex instruction, 91-92
texkill instruction, 92, 99
texld instruction, 99-100
texm3x2depth instruction, 92
texm3x2pad instruction, 92-93
texm3x2tex instruction, 93
texm3x3 instruction, 93-94
texm3x3pad instruction, 94
texm3x3spec instruction, 94-95
texm3x3tex instruction, 95-96
texm3x3vspec instruction, 96
texreg2ar instruction, 97
texreg2gb instruction, 97
texreg2rgb instruction, 97
texture
advection, 418
coordinates, 81
matrix manipulation with Quake III shader, 469-471
operations, 81
perturbation effects, 337-346
scrolling, 295
shaders, 257
stages, 81
texture address instructions (pixel shader), 128
texture addressing,
ps.1.1-ps.1.3, 88-89, 97-98
ps.1.4, 98, 100-101
texture addressing instruction set,
ps.1.1-ps.1.3, 89-97
ps.1.4, 98-100
texture registers, 81
pixel shader, 82-83
texture space,
moving light into, 131-134
transforming L into, 132-134
shaders, 132-133
texture space coordinate system, establishing, 131-132
textures,
blending, 256-257
generating using noise, 238-239
three-dimensional Perlin noise, 243-244
time value, calculating, 228-229
tonal art maps, 324
transfer functions, 259, 441-442
492 Index
Team LRN
implementing, 439-441
shader, 441, 442
types of, 260-261
Transform & Lighting pipeline, see T&L pipeline
transformation matrix, using, 56
trilinear filtering, 429
turbulence, 431
tweening, 195-197
blending with skinning, 199-201
shader, 196-197
two-dimensional noise, 236-237
two-dimensional Perlin noise, 242-243
U
unit vector, calculating, 56
user clip planes, 74
UV flipping, 295
V
vector fields, 184-186
version instruction (pixel shader), 128
vertex buffer, using, 48
vertex compression, 172-173
examples, 174-179, 180-186
optimizations, 179-180
techniques, 172-173
vertex lighting, 277
vertex normal, correcting, 211-213
vertex shader,
architecture, 16-18
assembler, 10-11
benchmarks, 252
compiling, 22, 33-34, 45-46
creating, 34, 46, 51-53
creating with Shader Studio, 152-159
data types, 173
declaring, 19-21, 42-43, 54, 59-60
deleting, 35
effects, 7
effects with Quake III shaders, 467-471
freeing resources of, 47
hardware, 17
non-shader specific code, 47-49, 65
optimization example, 224-227
optimizations, 221-224
process of creating, 18, 42
programming, 18-35
setting, 34-35, 47
tools, 8-16
writing, 22, 32-33
vertex shader constant registers, setting, 22, 43, 54-55
vertex shader constants, setting, 60-61, 67
vertex shader examples
using Bzier patch class, 59-66
using combined diffuse and specular reflection, 59-6
using D3DXAssembleShader() function, 38-49
using NVASM, 49-53
with per-vertex diffuse reflection, 53-59
with point light source, 66-70
vertex shader instruction set, 23-26
complex, 27
vertex shader registers, 17-18
using, 28-31
vertex shader resources, freeing, 35
vertex shader support, checking for, 19, 42
vertex shaders
blending textures, 257
blending tweening and skinning, 200-201
bubble effect, 370-372
bump mapping, 410
cartoon shading, 323
compression, 206-207
compression transform, 179
correcting vertex normal, 212-213
creating photorealistic faces, 303-307
crystal effect, 364-365
diffuse and specular point light, 68-69
diffuse and specular reflection, 61-62
diffuse lighting, 55
displacement, 185-186
environment mapping, 293
eye environment mapping, 315-316
eye space depth, 328-329
font smoothing, 271
Fresnel effects, 283-284
geometry emulation, 274-276
Gooch lighting, 327
grass, 335-336
hatching, 325-326
HCLIP displacement map, 186
HCLIP vector field, 186
lake water, 358-360
light falloff, 204-205
light rendering, 408-409
lighting, 279-280
multiple compression transforms, 181-182
ocean water, 348-350, 352-356
optimized compression transform, 180
outlines, 321
quad example, 43-45
quantization, 175
scaled offset, 176
shadow volume extrusion, 191-192
skinning, 198-199
skinning tangent/texture space vectors, 202-203
specular reflection, 138
transformation to texture space, 132
tweening, 196-197
Index 493
Team LRN
vector field, 185-186
world space normals, 328-329
vertex shaders, 6-7
advantages of, 7
limitations of, 232-234
vertex stream declaration example, 174
vertex virtual machine, see VVM
vertices, processing, 6
visualization in computer games, 455
drawbacks of, 455-456
volume bound for pixel shader effects, 398-399
volume graphics
animating, 441-444
blending, 442-444
rotating, 441
volume rendering, 444-446
volume visualization, 438-439
volumetric effects, 438
volumetric fog, creating with 3D textures, 435-436
volumetric textures, see 3D textures
voxel, 439
VVM, 220
VVM registers, mapping to SSE registers, 220
W
water, see ocean water, lake water
world space normals shader, 328-329
494 Index
Team LRN
C
o
l
o
r
P
l
a
t
e
1
Team LRN
Color Plate 2
Team LRN
C
o
l
o
r
P
l
a
t
e
3
Team LRN
C
o
l
o
r
P
l
a
t
e
4
Team LRN
Color Plate 5
Color Plate 6
Color Plate 7
Team LRN
Color Plate 9
Color Plate 8
Team LRN
Color Plate 10
Color Plate 11
Team LRN
C
o
l
o
r
P
l
a
t
e
1
2
Team LRN
About the CD
The companion CD contains vertex and pixel shader code along with a variety of programs
demonstrating the concepts presented throughout the book. When you unzip the CD files,
they will be organized into directories named for the articles.
See the individual readme files for more information.
Additional resources are available at www.shaderx.com
WARNING: By opening the CD package, you accept the terms and conditions
of the CD/Source Code Usage License Agreement.
Additionally, opening the CD package makes this book non-returnable.
Team LRN
CD/Source Code Usage License Agreement
Please read the following CD/Source Code usage license agreement before opening the CD and using the contents
therein:
1. By opening the accompanying software package, you are indicating that you have read and agree to be bound
by all terms and conditions of this CD/Source Code usage license agreement.
2. The compilation of code and utilities contained on the CD and in the book are copyrighted and protected by
both U.S. copyright law and international copyright treaties, and is owned by Wordware Publishing, Inc. Indi-
vidual source code, example programs, help files, freeware, shareware, utilities, and evaluation packages,
including their copyrights, are owned by the respective authors.
3. No part of the enclosed CD or this book, including all source code, help files, shareware, freeware, utilities,
example programs, or evaluation programs, may be made available on a public forum (such as a World Wide
Web page, FTP site, bulletin board, or Internet news group) without the express written permission of
Wordware Publishing, Inc. or the author of the respective source code, help files, shareware, freeware, utili-
ties, example programs, or evaluation programs.
4. You may not decompile, reverse engineer, disassemble, create a derivative work, or otherwise use the
enclosed programs, help files, freeware, shareware, utilities, or evaluation programs except as stated in this
agreement.
5. The software, contained on the CD and/or as source code in this book, is sold without warranty of any kind.
Wordware Publishing, Inc. and the authors specifically disclaim all other warranties, express or implied,
including but not limited to implied warranties of merchantability and fitness for a particular purpose with
respect to defects in the disk, the program, source code, sample files, help files, freeware, shareware, utilities,
and evaluation programs contained therein, and/or the techniques described in the book and implemented in
the example programs. In no event shall Wordware Publishing, Inc., its dealers, its distributors, or the authors
be liable or held responsible for any loss of profit or any other alleged or actual private or commercial dam-
age, including but not limited to special, incidental, consequential, or other damages.
6. One (1) copy of the CD or any source code therein may be created for backup purposes. The CD and all
accompanying source code, sample files, help files, freeware, shareware, utilities, and evaluation programs
may be copied to your hard drive. With the exception of freeware and shareware programs, at no time can any
part of the contents of this CD reside on more than one computer at one time. The contents of the CD can be
copied to another computer, as long as the contents of the CD contained on the original computer are deleted.
7. You may not include any part of the CD contents, including all source code, example programs, shareware,
freeware, help files, utilities, or evaluation programs in any compilation of source code, utilities, help files,
example programs, freeware, shareware, or evaluation programs on any media, including but not limited to
CD, disk, or Internet distribution, without the express written permission of Wordware Publishing, Inc. or the
owner of the individual source code, utilities, help files, example programs, freeware, shareware, or evalua-
tion programs.
8. You may use the source code, techniques, and example programs in your own commercial or private applica-
tions unless otherwise noted by additional usage agreements as found on the CD.
Warning: By opening the CD package, you accept the terms and conditions of
the CD/Source Code Usage License Agreement.
Additionally, opening the CD package makes this book non-returnable.
Team LRN