Is it possible to specify when launching a cuDNN convolution operation that it should be performed using high priority streams?
Hi @friedj
Please refer to below links:
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html
Thanks
Thanks! I’m wondering about operations launched using the cuDNN interface. According to https://forums.developer.nvidia.com/t/does-cudnncreate-call-create-multiple-streams-internally/157422, “cuDNN can use internal streams for work.” How do I ensure that those internal streams are running with high priority?
Hi @friedj ,
Would you like to check the below link to understand the priorities in cudnn
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM
Thanks!
cuDNN creates its own streams that it uses internally. Is there a way to specify that those streams run with high or low priority?
Hi @friedj ,
Internal stream priority cannot be controlled by you. However i believe you can create a stream with priority and associate it with the handle.
Thanks!