Getting Explicit

How hard is Vulkan Really?

“Vulkan is not a good first-step path to
learning graphics. It’s maintainers will tell you
to start out by learning OpenGL or DX11.”
– Internet Person
“To be fair to Vulkan, it was designed to be
used by somebody who knew what they
were doing.” – Another Internet Person
“Eh I would say Vulkan decreases
reliability. Precisely because Vulkan
drivers are much thinner and do no error
checking, … “ – Also an Internet Person
“I thought about the Vulkan programming guide,
but ended up deciding modern OpenGL would be
an easier starting point...

… In a twist of fate, the SuperBible example code

only supports OpenGL 4.0+ and my primary
computer is stuck at 3.3.” – Same Internet Person
“So at the end of the day, Vulkan is about giving
graphics developers foot-guns in order to reduce CPU
overhead. That was explicitly its mission statement
from the beginning!” – Proud Internet Person
Background

• 2005: Started in web services.

• 2008-2011: Contract to hire Quake Live
• 2011-2013: Dev Services
• 2013-Present: Generalist idTech
• 8yrs QuakeCon Support
!= Graphics Programmer

• Occasional day-to-day fixes.

• Wrote a handful of shaders.
• Played with ShaderToy, OpenCV, Processing, etc
• Read Peter Shirley’s “Ray Tracing in One Weekend”

• More capable hands at id.

• Tiago, Billy, Axel, Jean, etc
• Great talks given at previous events regarding state of graphics in idTech.
• Great work on Wolfenstein II. Hopefully talk about that soon.
Dawn of Vulkan at id

• 2015 – Breaking the ice.

• Q1 2016 – Began work in earnest
close to DOOM2016 launch.
• Q3 2016 – DOOM2016 Vulkan binary
released to the wild.

• Clear benefits
• Solidifying marriage with idTech
• Eager for others to Adopt
Early Public Perceptions

• HARD
• A LOT of Code
• Graphics++ for developers at the bleeding edge.
Hobbyists and beginners need not apply.
• Expensive investment to refactor. ( 4 – 6mo, $$$$ )
• “Unpacking” the driver revealed more
responsibilities.
The Early Days – Body of Information

Top 10 %
Expert

Middle
???

Bottom 10-20 %
Intro
Evangelizing the id way – Open Source

• VkQuake:
• Author: Axel Gneiting
• Developed: Q2 2016
• Released: Q3 2016
• Link: https://github.com/Novum/vkQuake

• VkDOOM3:
• Author: Dustin Land
• Developed: Q3 2016
• Released: Q3 2017
• Link: https://github.com/DustinHLand/vkDOOM3
Blog - I Am Graphics And So Can You

• 7 Part series articulating my

experience with the port.
• Thought I’d get 3x dozen views TOTAL.
• ~30,000 views first month. ( broke
squarespace’s analytics report )
• Average ~1,000 view every month
after.
• Handful of conversations every
month; mostly with beginners.
• Considerable interest in Vulkan

https://www.fasterthan.life/
Impetus for Porting DOOM 3 BFG

• Expanding Skillset
• Explicit APIs are a nice inflection point.
• Everyone still learning.
• Good Fit
• Manageable by a single person.
• Familiar with idTech.
• Open Source
• Fixed Assets. ( actually a problem later on )
• Evangelizing
• OpenGL -> Vulkan a nice example of old to new.
• Partners asking how to make Vulkan more approachable.
Video
Port Effort

• Code:
• Commits: 718
• LoC: ~5,000
• GLSL LoC: ~3,000
• Time:
• Total: 4x full time months.
• Regularity: Waxed and Waned
• When: 10pm – 3am
Renderer Lines of Code

12000
10000
8000
6000
4000
2000
0
Frontend OpenGL Vulkan GLSL
Vulkan Lines of Code

2500

2000

1500

1000

500

0
Allocator Staging Renderer Pipelines Buffers Images
Vulkan Summoning Ritual – Step 1: Stopping For Intuition
 Commercial Break

RenderDoc
Intuition Begins And Ends In Understanding How Data Is
Transformed

Author: Baldur Karlsson

Link: https://renderdoc.org/
GitHub: https://github.com/baldurk/renderdoc
RenderDoc – Dynamic Shadows
RenderDoc – Color Channels
RenderDoc – Post Processing
Vulkan Summoning Ritual – Step 2: The First 1,000 LoC
The First 1,000 LoC – “An” Entry Trajectory Into Creating Vulkan Applications

Create
Create Create Render
Swapchain &
Instance Pass(es)
Get Images

For Each Swapchain Image

Create Surface Create Pools
Create Command
Buffer

Create
Allocate Your
Create Logical Frame Objects
Synchronization
Select Suitable Primitives
Device & Get
Physical Device
Queues
Create Frame
Buffer
The First 1,000 LoC - RenderPass

• Area of interest for Vulkan adopters.

• Doesn’t have to be complicated. ( Can be if you’re into that. )
• vkDOOM3 has 1x RenderPass with 1x subpass and 3x attachments.
• Manually transition layout to PRESENT_SRC_KHR.

Depth Color Resolve - MSAA

Vulkan Summoning Ritual – Step 3: Resources Take Center Stage
Resources Take Center Stage

• Buffers in vkDOOM3
• Static Geo (DEVICE_LOCAL)
• Dynamic Geo double buffered.
(HOST_VISIBLE)
• UBOs for skinning and render parms.
• idBuffer treats range as discrete
• Images in vkDOOM3
• 12x Formats
• All DEVICE_LOCAL.
• Track VkImageLayout for descriptor
updates.
• TRANSFER_DST ->
SHADER_READ_ONLY
Resources Take Center Stage - Allocator

• There are options.

• Simple Apps – Just use the API directly.
Watch your numbers.
• Small to Medium Apps – Simple block
allocator. Be cognizant of the rules.
• Medium to Large – You really need to
invest in something serious.

https://vulkan.gpuinfo.org/
Resources Take Center Stage - What I did for allocations

VkResult ret = vkCreateBuffer( struct vulkanAllocation_t {

vkcontext.device, idVulkanBlock * block;
&bufferCreateInfo, uint32 id;
NULL, VkDeviceMemory deviceMemory;
&m_apiObject ); VkDeviceSize offset;
VkDeviceSize size;
VkMemoryRequirements memoryRequirements;
byte * data;
vkGetBufferMemoryRequirements(
};
vkcontext.device,
m_apiObject,
&memoryRequirements );

// determine mem usage

m_allocation = vulkanAllocator.Allocate(
memoryRequirements.size,
memoryRequirements.alignment,
memoryRequirements.memoryTypeBits,
memUsage,
VULKAN_ALLOCATION_TYPE_BUFFER );
Resources Take Center Stage - Allocator Breakdown

Get Heap Index

Allocate Block Chain Per Heap

Don’t shoot yourself in the face. Subdivide On Allocation

Only 120/600 LoC header/source.

Merge On Free
VMA – Vulkan Memory Awards
 Commercial Break

Vulkan Memory
Allocator
Author: Adam Sawicki
Link: https://gpuopen.com/
GitHub: https://github.com/GPUOpen-
LibrariesAndSDKs/VulkanMemoryAllocator
Vulkan Summoning Ritual – Step 4: Coupling to idTech 4
Coupling to idTech 4

• Long history with OpenGL pedigree. ( 2000 –

Present )
• GL code had “grown roots”. Present in 18x
files. Not bad, but still leaky
• PS3 and 360 code had gone down a different
path.
• Wanted to keep OpenGL working side-by-side
for comparison.
• Large part of port effort was “putting the
beast back in its cage”.
• In the end, GL code sequestered to 8x files.
Coupling to idTech 4 – State Tracking

• Designed to change state “immediately”. ( // example snippet

take that driver ) static const uint64
static const uint64
GLS_SRCBLEND_ONE
GLS_SRCBLEND_ZERO
=
=
0
1
<<
<<
0;
0;
• GL_State static const uint64 GLS_SRCBLEND_DST_COLOR = 2 << 0;
• GL_SelectTexture / GL_BindTexture static const uint64
static const uint64
GLS_SRCBLEND_ONE_MINUS_DST_COLOR
GLS_SRCBLEND_SRC_ALPHA
=
=
3
4
<<
<<
0;
0;
• GL_Clear static const uint64
static const uint64
GLS_SRCBLEND_ONE_MINUS_SRC_ALPHA
GLS_SRCBLEND_DST_ALPHA
=
=
5
6
<<
<<
0;
0;

• Required state tracking till data could be static const uint64

static const uint64
GLS_SRCBLEND_ONE_MINUS_DST_ALPHA
GLS_SRCBLEND_BITS
=
=
7
7
<<
<<
0;
0;
used at appropriate times.
• Descriptor Set Updates
• Pipeline Creation / Selection // In StencilShadowPass
• Current state saved to GLS bits, but not GL_State(
glState |
the data. ( uint64_t ) GLS_STENCIL_OP_FAIL_KEEP |
• Blend Ops/Functions GLS_STENCIL_OP_ZFAIL_KEEP |
• Depth State GLS_STENCIL_OP_PASS_INCR |
GLS_STENCIL_MAKE_REF( STENCIL_SHADOW_TEST_VALUE ) |
• Cull Mode GLS_STENCIL_MAKE_MASK( STENCIL_SHADOW_MASK_VALUE ) |
• Stencil State GLS_POLYGON_OFFSET |
GLS_CULL_TWOSIDED );
• etc
Coupling to idTech 4 – State Tracking Example

OpenGL::GL_State() Vulkan::GL_State()
• 210 LoC • 5 LoC
• glEnable/glDisable
• Save off GLS bits
• glCullFace
• glDepthFunc • Submitted at pipeline creation
• glDepthMask
• glBlendFunc
• glColorMask
• glPolygonMode
• glPolygonOffset
• glStencilOp
Vulkan Summoning Ritual – Step 5: Pixels & Pipelines
Pixels & Pipelines

• Surface based drawing - two routines

• Main lighting
• Stencil shadows
• Steps
• Find offsets into index / vertex
buffers.
• Set joint cache handle if we’re
skinning.
• renderProgManager.CommitCurrent(
GLS bits )
• vkCmdBindIndexBuffer
• vkCmdBindVertexBuffer
• vkCmdDrawIndexed
• Profit
Pixels & Pipelines - CommitCurrent

• Backend sets a current “render prog” index. ::BindProgram( int index )

• CommitCurrent uses index to retrieve render prog which has a list of viable
pipelines.
• Lookup pipeline using GLS bits.
• If one doesn’t exist, create it.
• No noticeable perf hit, so didn’t bother with pipeline cache.
• Desc Set – Brute force approach, but still quite fast.
• Allocate a DescSet
• Update it using preallocated UBO and bound textures.
• Bind
• Bind Pipeline
Pixels & Pipelines – Graphics Is a Series of Tubes

Create Built-In Shaders

Create Descriptor Set Layouts

Create Vertex Descriptions

28 built in shaders + change

Create Descriptor Pools
Under 100 pipelines total at runtime.

Preallocate UBOs For Render Parms

Summoning Complete
In Closing - Views On Process Work Load

Legacy Explicit Lite Explicit

Init Setup Frame 1 Frame 2 Frame 3 Frame 4 Frame 5

 Special Thanks
Blog Readers
Axel Gneiting – vkQuake, Advice
Tiago Sousa – Talk Review, Advice
Stephanie Hurlburt – Encouragement
John Carmack – Open Source
OG & BFG DOOM 3 Teams

Jan Paul van Waveren Steven Serafin

 Getting Explicit:
How hard is Vulkan Really?
Panel Discussion
Khronos Dev Day – GDC 2018
Copyright ©2018 The Khronos™ Group Inc. - Page 47
Getting Explicit: How hard is Vulkan Really?
Panelists:
• Matthäus Chajdas, AMD
@NIV_Anteru
• Arseny Kapoulkine, Roblox
@zeuxcg
• Dustin Land, id Software
@DustinHLand
Chair:
• Alon Or-bach, Samsung Electronics
@alonorbach

Copyright ©2018 The Khronos™ Group Inc. - Page 48