CUDA-OpenGL Interoperability Accessing all layers of OpenGL array texture in CUDA

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I’m just wondering if anyone has any experience using CUDA-OpenGL interoperability, specifically with accessing all layers of a GL_TEXTURE_2D_ARRAY texture in CUDA.

What I’ve done so far is, I created the layered texture in OpenGL, register it using cudaGraphicsGLRegisterImage() with GL_TEXTURE_2D_ARRAY target. When I want to use it I map it using cudaGraphicsMapResources() and then get access to the cudaArray* using cudaGraphicsSubResourceGetMappedArray, specifying 0 for both the array index and the mipmap level.

This set up seems to be working. The issue I’m having is that I only seem to be able to access the base level of the array texture (I guess from specifying 0 for the array index?). Is it possible to be able to access all layers from within the kernel using a bound texture reference?

I assume this is possible because the sample simpleLayeredTexture creates 3D cudaArray which is bound to a texture reference of type cudaTextureType2DLayered and it can access all layers in the kernel.

Since both the simpleLayeredTexture sample and my code are binding the cudaArray* to the texture reference of the same type, I can only think of the cudaGraphicsSubResourceGetMappedArray() call as the one causing the issue. Is there any other way of getting a cudaArray* to the array texture?

texture<float4, cudaTextureType2DLayered, cudaReadModeElementType> texture_reference;

cudaGraphicsResource_t hemicube_resource;

Gluint render_texture_array;

glGenTextures(1, &render_texture_array);

glBindTexture(GL_TEXTURE_2D_ARRAY, render_texture_array);

glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);

glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);

glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);

glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);

glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);

glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_COMPARE_FUNC, GL_LEQUAL);

glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, internal_format, texture_width, texture_height, texture_depth, 0, format, type, NULL);

cudaError_t status = cudaSuccess;

status = cudaGraphicsGLRegisterImage(&texture_resource, render_texture_array, GL_TEXTURE_2D_ARRAY, cudaGraphicsMapFlagsReadOnly);

// Render to the render_texture_array

cudaError_t status = cudaGraphicsMapResources(1, &texture_resource, 0);

assert(status == cudaSuccess);

cudaArray* texture_array = NULL;

status = cudaGraphicsSubResourceGetMappedArray(&texture_array, texture_resource, 0, 0);

assert(status == NULL);

cudaChannelFormatDesc channel_description = cudaCreateChannelDescHalf4();

cudaError_t status = cudaBindTextureToArray(texture_reference, texture_array, channel_description);

assert(status == cudaSuccess);

// Launch kernel

cudaError_t status = cudaUnbindTexture(texture_reference);

assert(status == cudaSuccess);

status = cudaGraphicsUnmapResources(1, &hemicube_resource, 0);

assert(status == cudaSuccess);

Thank you

Surya

2

Did you solve this problem?
Did anyone know if it is possible to map the whole texture array to be used by cuda, instead of just one texture layer?

Thanks

3

Hi Bob,

I found a work around to be able to access all layers of the texture array. What I’ve done is still use the cudaGraphicsSubResourceGetMappedArray() and bind a texture reference for each layer. Then in the kernel I have a function to sample the right texture reference based on the group index, in this case I used a switch case statement.

I’ve read in another post, it’s possible to create an array of function pointer, one function to access each texture reference. I think the main advantage of this is that there’s no conditional statement in the kernel, which should make it faster, but I haven’t tried it. In my case, the texture array is limited to 8 layers, so it’s not too bad. If there’s more layers I think the function pointer array would be better.

Hope it helps.

Surya