cudaBindTexture

Accelerated Computing CUDA CUDA Programming and Performance
1

Host calls the compute function in every frame to execute ‘bind-call kernel-unbind’.

Inside the kernel function, I modify the content of the ‘mybuffer’.

I’m wondering which case is correct.

case1)

void compute(…)

{

cudaBindTexture(0, texref, mybuffer, desc, size );

//call kernel

K_compute<<numBlocks, numThreads>>( mybuffer );

cudaUnbindTexture( 0, texref );

}

case2)

cudaBindTexture(0, texref, mybuffer, desc, size ); // bind once at initialization

//call compute for several times

void compute(…)

{

 //call kernel

 K_compute<<numBlocks, numThreads>>( mybuffer );

}

cudaUnbindTexture( 0, texref ); // unbind at program termination

Does ‘bind’ mean sharing pointer(or reference) of certain memory?

2

It’s fine (and probably faster) to only bind once if the texture to buffer association isn’t changing.

3

It’s fine (and probably faster) to only bind once if the texture to buffer association isn’t changing.

4

Thanks a lot!

5

Thanks a lot!