Closed ggruszczynski closed 1 year ago
I decided to change this mechanism, for TCLB to ask the GPU for the CC, and if that fail, just regress to nvcc
default. Of course one will be still able to specify the CC by hand with --with-cuda-arch=
, but now it will accept whatever you give it. the changes are here:
https://github.com/llaniewski/TCLB/compare/feature/hip...llaniewski:TCLB:feature/sm_auto?expand=1
I will test it and then modify the readme and pull request (this combines with #376 )
Merged in #376
I found that the landing instruction https://github.com/CFD-GO/TCLB is not working
My configuration:
I get this error:
Reason
Support for compute_30 has been removed for versions after CUDA 10.2.
https://stackoverflow.com/questions/64774548/unsupported-gpu-architecture-compute-30-on-a-cuda-5-capable-gpu
https://docs.nvidia.com/cuda/archive/10.2/pdf/CUDA_Toolkit_Release_Notes.pdf
I suggest pushing up the instruction to
sm_60
(Pascal, next aftersm_30
in TCLB config), as Maxwell (sm_50
) is also deprecated from CUDA 11.6 onwards: https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/