Hello,
Absolute CUDA novice and lapsed (20yrs) PGI user here.
I’ve been given code to compile that previously was compiled using CUDA9. We now have a new GPU PC with an RTX2080Ti. Reading the release notes PGI 19.4 has CUDA Fortran support for the Turing architecture and so does Cuda10. So I suppose -Mcuda=10.1 is needed. But is -ta=tesla still the correct switch?
Cheers!
Hi domuhe,
So I suppose -Mcuda=10.1 is needed.
The compiler should auto-detect what CUDA driver version you’re using and match the CUDA version used to that of the driver. So using -Mcuda=cuda10.1 may be unnecessary, but doesn’t hurt.
But is -ta=tesla still the correct switch?
-ta=tesla enables OpenACC so isn’t needed for use with CUDA Fortran. For Turing, the architecture flag would be “-Mcuda=cc75”.
Hope this helps,
Mat
Hi,
Thank you for responding. I am only still confused as to whether I can use OpenACC with a Turing architecture GPU. IF so, what switch would be correct?
thanks!
For CUDA Fortran, use “-Mcuda=cc75”,
For OpenACC use “-ta=tesla:cc75”.
“cc75” stands for Compute Capability (CC) 7.5, also known as the Turning architecture.
Note, for a complete list of sub-options to either flag, run “pgfortran -help -ta -Mcuda”, from the command line.
Hope this helps,
Mat
Thank you!