cublas and streams

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

Is it possible to tie a cublasSgemm call to a stream?

2

Not yet. There was a rumour that a streams interface for CUBLAS would appear during the 3.1 beta release cycle. I don’t know if that is still the case.

For GEMM, there is code available for very high performance implementations here, which you can wrap in streams yourself if you really need it now.

3

Thank you avidday. I have considered the possibility of a custom implementation but wasnt aware of the code you linked.

I’ve looked through the code. Unfortunately, I’m not able to control that matrix dimensions are multiples of 16 or 64 in my particular application, are you aware of any custom implementations that accept arbitrary number of rows/cols and have recieved the same amount of love as Volkovs’ or Lung-Sheng Chiens’ implementations? (Even though this might hurt performance)

4

I believe there’s stream support in the beta–cublasSetStream, maybe?

5

Thanks, I’ll check it out.