There's already an effort made for AMD gpus to support CUDA and cuDNN codes, that is already in beta stage. Maybe someone can port it into lc0?smatovic wrote: ↑Mon Jan 21, 2019 9:22 pmThanks for the Kudos, but i am really a layman in neural networks.Dann Corbit wrote: ↑Mon Jan 21, 2019 8:38 pmSrdja is an OpenCL expert, so maybe he can fix it.CMCanavessi wrote: ↑Mon Jan 21, 2019 8:33 pmThere are 3 versions of lc0, CUDA, OpenCL and CPU so it's already there. Thing is it's way slower than CUDA.Dann Corbit wrote: ↑Mon Jan 21, 2019 8:18 pmI think the LC0 code is strictly written for Nvidia cards, so a complete rewrite would be needed for OpenCL.
Afaik, GCP wrote Leela Zero initially with an OpenCL backend,
and this was ported to Leela Chess Zero,
then Ankan added an high performant cuDNN backend,
and with the new Nivida RTX series this backend profits from TensorCores.
So i think it is AMDs turn to release some optimized OpenCL NN libraries
and add TensorCores, or a similar technique, to their gpus.
--
Srdja
CUDA code to portable C++ converter (HIP):
https://github.com/ROCm-Developer-Tools/HIP
cuDNN to hipDNN:
https://github.com/ROCmSoftwarePlatform/hipDNN