CUDA + OpenGL CUDA + OpenGL

• CUDA used for calculation, • CUDA C uses familiar C memory management

data generation, image techniques (malloc, pointers)
manipulation
• OpenGL stores data in abstract generic buffers
• OpenGL used to draw pixels or called buffer objects
vertices on the screen • CUDA/OpenGL interop uses one simple concept:
• Interop is very fast! They share – Map/Unmap an OpenGL buffer into CUDA’s
data through common memory memory space
in the framebuffer

Setup Steps to OpenGL with CUDA 1)

1 Creating the window
1 Create a window (OS specific)
• Each OS does this differently.
2 Create a GL context (also OS specific) We’ll use Win32 for examples here:
– CreateWindowEx() is the Win32 function to
3 Set up the GL viewport and coordinate system create a window. Returns an HWND.

Create the CUDA Context

– Also need the windows HDC:
44 HDC hDC;
hDC=GetDC(hWnd);
5 Generate one or more GL buffers to be shared with CUDA

6 Register these buffers with CUDA

1a Set the Pixel Format for the Window 2)
2 Create the OpenGL Context
static PIXELFORMATDESCRIPTOR pfd=
{ // Create a wGL rendering context
sizeof(PIXELFORMATDESCRIPTOR), // Size Of This Pixel Format Descriptor
1, // Version Number HGLRC hGLRC;
PFD_DRAW_TO_WINDOW | // Format Must Support Window hGLRC=wglCreateContext(hDC);
PFD_SUPPORT_OPENGL | // Format Must Support OpenGL
PFD_DOUBLEBUFFER, // Must Support Double Buffering
PFD_TYPE_RGBA, // Request An RGBA Format // Activate the rendering context
8, // Select Our Color Depth, 8 bits / channel
0, 0, 0, 0, 0, 0, // Color Bits Ignored wglMakeCurrent(hDC,hGLRC);
0, // No Alpha Buffer
0, // Shift Bit Ignored
0, // No Accumulation Buffer // loads OpenGL extensions to support buffers
0, 0, 0, 0, // Accumulation Bits Ignored glewInit();
32, // 32 bit Z-Buffer (Depth Buffer)
0, // No Stencil Buffer
0, // No Auxiliary Buffer
PFD_MAIN_PLANE, // Main Drawing Layer
0, // Reserved
0, 0, 0 // Layer Masks Ignored
};
GLuint PixelFormat;
// create the pixel pixel format descriptor
Note: Use the PFD_STEREO flag on NVIDIA Interested in off-screen rendering? Use GPU Affinity on NVIDIA
PixelFormat=ChoosePixelFormat(hDC,&pfd; Quadro cards for OpenGLStereo Support! Quadro cards to create an OpenGL context without a window
// set the pixel format descriptor
SetPixelFormat(hDC,PixelFormat,&pfd);

3)
3 Set Up Our Viewport 3)
3 Set Up (Cont.)
(0,1.0,-1.0) (0,1.0,1.0)
// Set up which portion of the • Enable depth sorting
// window is being used
glViewport(0, 0, width, height); glEnable(GL_DEPTH_TEST);
(1.0,1.0,-1.0) (1.0,1.0,1.0)
// Just set up an orthogonal system
glMatrixMode(GL_PROJECTION); • Set the clear color and clear the viewport
(0,0,-1.0) (0,0,1.0)
glLoadIdentity(); glClearColor(1.0f, 1.0f, 1.0f, 1.5f);
glOrtho(0,1.0f,0,1.0f,-1.0f,1.0f);
glMatrixMode(GL_MODELVIEW); glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glLoadIdentity(); (1.0,0,-1.0) (1.0,0,1.0)

More on OpenGL coordinates later; for now,

we’ll just set up a simple orthogonal view
4 Create the CUDA Context 4)
5 Create a OpenGL Buffer(s)
• OpenGL context must be created first GLuint bufferID;

// Generate a buffer ID
• To create the CUDA context: glGenBuffers(1,&bufferID);
– Driver API: Use cuGLCtxCreate() instead of // Make this the current UNPACK buffer (OpenGL is state-based)
cuCtxCreate() glBindBuffer(GL_PIXEL_UNPACK_BUFFER, bufferID);
– Runtime API: Call cudaGLSetGLDevice() before any other // Allocate data for the buffer
API calls glBufferData(GL_PIXEL_UNPACK_BUFFER, width * height * 4,
NULL, GL_DYNAMIC_COPY);
• CUDA/OpenGL interop functions defined in:
– cudagl.h for driver API
– cuda_gl_interop.h in C Runtime for CUDA

5)
6 Register Buffers for CUDA Now, Let’s Actually Draw Something
• Driver API: • Common use case: drawing images
– cuGLRegisterBufferObject( GLuint bufferobj );
• Use Textures
– Unregister before freeing buffer:
cuGLUnregisterBufferObject( GLuint bufferobj );
• Textures are a ubiquitous feature of 3D graphics
• Runtime API: • Simple case: Just draw a texture on a Quad
– cudaGLRegisterBufferObject( GLuint bufObj );
– Unregister before freeing buffer:
cudaGLUnregisterBufferObject( GLuint bufObj );

These commands simply inform the OpenGL and CUDA

drivers that this buffer will be used by both
Textures Steps to Draw an Image From CUDA
1 Allocate a GL buffer the size of the image
2 Allocate a GL texture the size of the image
3 Map the GL buffer to CUDA memory
4 Write the image from CUDA to the mapped memory
5 Unmap the GL buffer
6 Create the texture from the GL buffer
7 Draw a Quad, specify the texture coordinates for each corner
8 Swap front and back buffers to draw to the display

1)
1 Allocate the GL Buffer 2)
2 Create a GL Texture
• Same as before, compute the number of bytes based upon the // Enable Texturing
image data type (avoid 3 byte pixels) glEnable(GL_TEXTURE_2D);
// Generate a texture ID
• Do once at startup, don’t reallocate unless buffer needs to grow glGenTextures(1,&textureID);
─ this is expensive // Make this the current texture (remember that GL is state-based)
GLuint bufferID; glBindTexture( GL_TEXTURE_2D, textureID);
// Generate a buffer ID
// Allocate the texture memory. The last parameter is NULL since we only
glGenBuffers(1,&bufferID); // want to allocate memory, not initialize it
// Make this the current UNPACK buffer (OpenGL is state-based) glTexImage2D( GL_TEXTURE_2D, 0, GL_RGBA8, Width, Height, 0, GL_BGRA,
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, bufferID); GL_UNSIGNED_BYTE, NULL);
// Must set the filter mode, GL_LINEAR enables interpolation when scaling
// Allocate data for the buffer. 4-channel 8-bit image glTexParameteri(GL_TEXTURE_2D,GL_TEXTURE_MIN_FILTER,GL_LINEAR);
glBufferData(GL_PIXEL_UNPACK_BUFFER, Width * Height * 4, glTexParameteri(GL_TEXTURE_2D,GL_TEXTURE_MAG_FILTER,GL_LINEAR);
NULL, GL_DYNAMIC_COPY);
Note: GL_TEXTURE_RECTANGLE_ARB may be used instead of GL_TEXTURE_2D
cudaGLRegisterBufferObject( bufferID );
for improved performance if linear interpolation is not desired. Replace
An OpenGL buffer used for pixels and bound as GL_PIXEL_UNPACK_BUFFER GL_LINEAR with GL_NEAREST in the glTexParameteri() call.
is commonly called a PBO (Pixel Buffer Object)
3)
3 Map the GL Buffer to CUDA 4)
4 Write to the Image
• Provides a CUDA pointer to the GL buffer─on • CUDA C kernels may now use the mapped memory
a single GPU no data is moved (Win & Linux) just like regular GMEM
• When mapped to CUDA, OpenGL should not use
this buffer • CUDA copy functions can use the mapped memory
as a source or destination
• Driver API:
– cuGLMapBufferObject( CUdeviceptr *dptr,
unsigned int *size, GLuint bufferobj );

• C Runtime for CUDA:

– cudaGLMapBufferObject(void **devPtr, GLuint bufObj);

5)
5 Unmap the GL Buffer 6)
6 Create a Texture From the Buffer
// Select the appropriate buffer
• Driver API: glBindBuffer( GL_PIXEL_UNPACK_BUFFER, bufferID);

– cuGLUnmapBufferObject( GLuint bufferobj ); // Select the appropriate texture

glBindTexture( GL_TEXTURE_2D, textureID);
• Runtime API:
// Make a texture from the buffer
– cudaGLUnmapBufferObject( GLuint bufObj); glTexSubImage2D( GL_TEXTURE_2D, 0, 0, 0, Width, Height,
GL_BGRA, GL_UNSIGNED_BYTE, NULL);

Source parameter is NULL, Data is

These functions wait for all previous GPU activity to coming from a PBO, not host memory
complete (asynchronous versions also available). Note: glTexSubImage2D will perform a format conversion if the buffer is a
different format from the texture. We created the texture with format
GL_RGBA8. In glTexSubImage2D we specified GL_BGRA and GL_UNSIGNED_INT.
This is a fast-path combination.
7)
7 Draw the Image! 8)
8 Swap
S Buffers
Just draw a single Quad with texture coordinates Eariler we specified a double buffered pixel format
for each vertex: (PFD_DOUBLEBUFFER).
glBegin(GL_QUADS); All drawing is done to a off-screen framebuffer. When
glTexCoord2f( 0, 1.0f);
glVertex3f(0,0,0); finished just swap the front & back buffers.
glTexCoord2f(0,0);
glVertex3f(0,1.0f,0);
SwapBuffers(hDC);
glTexCoord2f(1.0f,0);
glVertex3f(1.0f,1.0f,0);
glTexCoord2f(1.0f,1.0f); Note: Buffer swapping normally occurs at the vertical refresh interval to avoid
glVertex3f(1.0f,0,0); tearing (commonly 60 hz). You can turn off v-sync in the control panel to make
glEnd(); the swap instant (e.g., when benchmarking).

3D Geometry Coordinate Matrices

The Camera Analogy OpenGL’s coordinate systems
1. Position & Point the Camera at • Model-View Matrix: defines the camera
the Scene (View transform)
position and direction (alternatively the
2. Arrange the scene composition model’s position and orientation)
(Model transform)
3. Adjust the camera zoom • Projection Matrix: Defines the cameras
(Projection Transform) field-of-view and perspective
View Frustum
4. Choose the final size
(Viewport Transform)

See the OpenGL Red Book, page 106 Matrices are states. Manipulating a matrix applies to subsequent calls.
Model-View Transform Projection Transform
• Select the Model-View Matrix • Select the projection Matrix Just choose your lens!
– glMatrixMode(GL_MODELVIEW) – glMatrixMode(GL_PROJECTION)

• Common Operations • Useful Functions:

– glLoadIdentity() Resets the matrix – glLoadIdentity()

– glRotatef() – glOrtho()
– glFrustum()
– glTranslatef()
– gluLookAt()
– glScalef()
– gluPerspective()

Drawing Simple Geometry Vertex Arrays

• Primitives are stored in an OpenGL buffer
• glBegin() / glEnd() ─ Lots of options:
– Can be GL_POINTS, GL_LINES, GL_TRIANGLES, etc.
– GL_POINTS, GL_LINES, GL_LINE_STRIP, GL_LINE_LOOP,
GL_TRIANGLES, GL_TRIANGEL_STRIP, GL_TRIANGLE_FAN,
GL_QUAD_STRIP, GL_POLYGON
• Properties including Color, Texture Coordinates, Surface
Normals can also be stored in the array
• Use a glVertex*() function with glColor*(),
glTexCoord*() • glDrawArrays() is a very powerful mega-function;
Draws whatever is in the array to the screen
• Not very efficient, use only for simple geometry
• Mapping the Vertex Buffer to CUDA allows arbitrary
Note: Many OpenGL functions, such as glVertex*() actually refer to a group
data creation or manipulation!
of functions with different parameter options, e.g., glVertex3f(), glVertex2f(), An OpenGL buffer used for vertices and bound as GL_ARRAY_BUFFER is
glVertex3i()… commonly called a VBO (Vertex Buffer Object)
Using a Vertex Array With CUDA 1)
1 Allocate & Register the Buffer
Allocate the GL buffer for the Vertex array,
1 Register it for CUDA E.g., Each vertex contains 3 floating point coordinates (x,y,z)
and 4 color bytes (RGBA): total 16 bytes per vertex
2 Use CUDA to create/manipulate the data
GLuint vertexArray;
• Map the GL Buffer to CUDA glGenBuffers( 1,&vertexArray);
• Set the values for all vertices in the array glBindBuffer( GL_ARRAY_BUFFER, vertexArray);
• Unmap the GL Buffer glBufferData( GL_ARRAY_BUFFER, numVertices * 16, NULL,
GL_DYNAMIC_COPY );
3 Use OpenGL to Draw the Vertex Data cudaGLRegisterBufferObject( vertexArray );

• Bind the buffer as the GL_ARRAY_BUFFER

• Set the type and array pointers for the type of data in the array
• Draw the array (glDrawArrays() )

4 Swap Buffers

2)
2 Use CUDA to Create the Data 3)
3 Use GL to Draw the Array
void * vertexPointer; // Bind the Buffer
glBindBuffer( GL_ARRAY_BUFFER, vertexBuffer );
// Map the buffer to CUDA
cudaGLMapBufferObject(&ptr, vertexBuffer); // Enable Vertex and Color arrays
// Run a kernel to create/manipulate the data
glEnableClientState( GL_VERTEX_ARRAY );
MakeVerticiesKernel<<<gridSz,blockSz>>>(ptr,numVerticies); glEnableClientState( GL_COLOR_ARRAY );

// Unmap the buffer // Set the pointers to the vertices and colors
cudaGLUnmapbufferObject(vertexBuffer); glVertexPointer(3,GL_FLOAT,16,0);
glColorPointer(4,GL_UNSIGNED_BYTE,16,12);
This is how we
tell OpenGL what type
of data is in the buffer.
More on the Pointers Final Step to Draw
Each Vertex contains 3 coordinates + color: glDrawArrays(GL_POINTS,0, numVerticies);
X Y Z R G B A – Can also use: GL_LINES,
GL_LINE_STRIP,GL_LINE_LOOP,
0 4 8 12 16 GL_TRIANGLES, GL_TRIANGLE_STRIP,
GL_TRIANGLE_FAN, GL_QUADS,
GL_QUAD_STRIP, GL_POLYGON

glVertexPointer(3,GL_FLOAT,16,0)
16 bytes between
elements SwapBuffer();
3 coordinates/ float type
vertex 8-bit type
4 Channels NVIDIA APEX Turbulence Demo. CUDA
Fluid simulation creates particles
which are rendered in OpenGL.
glColorPointer(4,GL_UNSIGNED_BYTE,16,12)