pytorch
compiling FBGEMM for ARM
I noticed that FBGEMM_GPU now supports ARM. Is there a plan for FBGEMM cpu to support ARM as well? I am trying to compile & run on an M1 mac.
@q10 in #2157 you mentioned "We are currently working on ARM support for FBGEMM, which is incomplete at the moment.". Do you have a very rough ETA for this feature?
I am trying to understand if this will be ready in time to ship something, or we need to go ahead with alternatives like ruy. We would prefer to stick with FBGEMM if possible.
Hi @vrnmthr, I would expect the FBGEMM main library to build on ARM sometime in 2024 H2. For now, I would suggest building the cpu-only version of FBGEMM-GPU and using that.
We have been actively working on optimizing some FBGEMM operators on ARM, such as TBE. However, building the full FBGEMM main library on ARM turned out to be much more complex as we tried this before (needing to completely disable AVX related routines for all operators during compile time or it will complain about "undefined reference" in link time or runtime), and we currently don't have enough time to focus on this.
Hi @excelle08 , glad to know ARM support is coming in 2024 H2. could you please elaborate the below suggestion? Does it support ARM CPU build with minimal kernels? otherwise what exactly does it provide?
For now, I would suggest building the cpu-only version of FBGEMM-GPU and using that.
Can the vcpkg port be looked into, it fails to build on ARM and needs to update to the latest ASMJIT. The source here seems to already support the latest asmjit so just need to PR vcpkg. This would help the asmjit port update go through faster https://github.com/microsoft/vcpkg/pull/39090
Bumping this. Is there an update?
From some of the more recent PRs (i.e. https://github.com/pytorch/FBGEMM/pull/4068, https://github.com/pytorch/FBGEMM/pull/4091, https://github.com/pytorch/FBGEMM/pull/3424, etc) it seems this is being worked on.
Also linking https://github.com/pytorch/FBGEMM/pull/2128 as it seems relevant