EvNN icon indicating copy to clipboard operation
EvNN copied to clipboard

Published 20 hours ago • verified publisher icon Efficient-Scalable-Machine-Learning

Compilation with CUDA 12

Open pascalkuthe opened this issue 8 months ago • 0 comments

The Makefile in this repo fails if compiling with CUDA12.

The issue is that the GPU_ARCH_FLAGS are hardcoded and include -gencode arch=compute_37,code=compute_37 The compute_37 architecture is not supported anymore in CUDA 12. Simply removing these flags from the makefile fixes the compilation.

That would break compatibility with older Tesla K80 cards. Ideally, instead of hardcoding the architectures, there would be a way to configure the architecture yourself based on the GPU/CUDA version you use.

pascalkuthe avatar Oct 16 '23 15:10 pascalkuthe