Could we control how many cores (SMs) will be used for a specific kernel?

Hi all,

For my new project, I will try to use multiple kernel at once.
And I got a question, “could I control how many cores (SMs) will be used for a specific kernel?”.

I cannot found any information about that in the programming guide and this forum.
If you have an experience on multiple kernel launch or on controlling the usage of the number of SMs (GPU computing resources), please tell me your knowledge or your experience.

Thanks.
Duksu Kim

The short answer is no. The programming model deliberately hides low level details of the hardware execution. About the only thing that can be done to help “promote” concurrent kernel execution is to keep the kernels block size, shared memory and register usage down. There has to be sufficient resources for blocks from multiple kernels to be run concurrently, if one or more of the kernels uses very large blocks or is very resource hungry, the chances blocks of another kernel can share the hardware are slim.

Thanks for your kind response.

I will try the method you mentioned.