how to create streams in CUDA?

hello, so I'm using CUDA for a project. my code is simple- many blocks that have the same amount of threads each and calculate the same equation. I want to create a few streams of blocks, each stream runs a few blocks. how can I create streams in Matlab? in all the examples I see people use C. from what I read I concluded that the streams should be created outside my kernel. I'll note that my kernel is written in C, and the rest of my code in Matlab. thank you!

Answer 1

[Francisco Ramírez Gil]

Hi, I have the same question.

I know that running a mex-function it is possible to setup the kernel launch using a code similar to the following:

cudaStream_t stream;       // CUDA streams are of type `cudaStream_t`.
cudaStreamCreate(&stream); // Note that a pointer must be passed to `cudaCreateStream`.

someKernel<<<number_of_blocks, threads_per_block, 0, stream>>>(); // `stream` is passed as 4th EC argument.

cudaStreamDestroy(stream); // Note that a value, not a pointer, is passed to `cudaDestroyStream`.

However, is there any way to create a non-default CUDA stream while constructing the CUDAKernel Objec? For instance, is there a command in which the "stream" can be specified? e.g.

k = parallel.gpu.CUDAKernel('myfun.ptx','myfun.cu','stream');

Or maybe during CUDAKernel Object Properties? e.g.

k.ThreadBlockSize = [1024,1,1];
k.GridSize = [500,1,1];
k.Stream = 'stream1'

If someone knows something... Please, help me!