How to set the priority of a stream in pycuda?

user_11829634 0
Viewed 67

The title says it all, but here is my problem in more detail: I'm implementing a finite elements solver in python + pycuda that should run on distributed systems.

To hide the communication latency, I'm trying to overlap computation and communication (with 2 separate streams). My problem is that the kernels used for the communication (on one stream) are executed at the end of the main computation kernel (see pic below).

Overlapp of the kernels

My question is: how can I tell my GPU to first execute the communication kernels? I'm using a RTX2060M, so stream priority is supported, and the presence of the attribute STREAM_PRIORITIES_SUPPORTED in pycuda makes me think that it's possible to set stream priorities from pycuda.

pycuda cuda python
1 Answers

It appears that at the date of writing (February 2022), PyCUDA has not implemented stream creation with priorities. So while what you want to do can be done with the CUDA driver API (which PyCUDA uses), that feature is not presently exposed in PyCUDA.

user_6818650
Related