Tensile
Tensile copied to clipboard
Add Navi GFX1010 support
I was able to build rocBLAS for GFX1010 by adding "gfx1010" to this list: https://github.com/ROCmSoftwarePlatform/Tensile/blob/79f2eaec38754e4c0acf3399710f68d2b8586450/Tensile/TensileCreateLibrary.py#L1262
https://github.com/ROCmSoftwarePlatform/pytorch/issues/718 https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/issues/1106
Works and tests running happily (using the GPU) for rocBLAS (after adding archs to it too) :D, using Tensile & rocBLAS branch gfx10
Any update on officially enabling gfx1010? @wbgilmartin @benjaminulmer
I would also be interested in any updates for GFX10xx :3
gfx1010 has been enabled by default in rocBLAS builds since ROCm 4.3.0. https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/rocm-4.3.0/CMakeLists.txt#L169
Any updates wether someone build pytorch with gfx1010?