ROCm / composable_kernel

Composable Kernel: Performance Portable Programming Model for Machine Learning Tensor Operators

https://rocm.docs.amd.com/projects/composable_kernel/en/latest/

Other

297 stars 113 forks source link

fix build issues on gfx1100 alone #1358

Closed junliume closed 3 months ago

junliume commented 3 months ago

1357

illsilin commented 3 months ago

The changes are made to 2 files with xdl instances. Those files should never get built for gfx11 targets at all.

junliume commented 3 months ago

The changes are made to 2 files with xdl instances. Those files should never get built for gfx11 targets at all.

with -DGPU_TARGETS="gfx1100" -DINSTANCES_ONLY=ON I think the second option overrides the former :(

illsilin commented 3 months ago

I see. Yes, setting GPU_TARGETS and INSTANCES_ONLY flags would break the cmake logic. If you need to build CK just for gfx1100, I recommend just removing the INSTANCES_ONLY. I may need to eventually come up with another build flag, so we should be able to build just instances for a select target and introduce some new flag to build the libraries for all supported targets.

junliume commented 3 months ago

@illsilin this problem also manifests when we set

CXX=/opt/rocm/bin/amdclang++ cmake -DCMAKE_PREFIX_PATH=/opt/rocm -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS="gfx1100;gfx90a;gfx942" ..

Hence GPU_TARGETS="gfx1100;gfx90a;gfx942"

junliume commented 3 months ago

replaced by #1372