clearnature wrote: @krzysz00 @searlmc1 The virtual FP4 types were originally submitted as part of ROCm/rocm-libraries#8061. That PR was closed with the instruction:
> "I would advise you to reopen such a PR against upstream LLVM if you believe > it is relevant there." — Torrezuk (rocBLAS maintainer) So we are now caught in a circular dependency: ``` rocWMMA says → "submit to LLVM" LLVM says → "submit to rocWMMA" ``` The FP4 types (`VirtualFp4Hw.h`, etc.) define numerical format representations (E2M1/E3M0/Q16) that are used by the WMMA/SWMMAC instruction lowering in the AMDGPU backend. They are infrastructure headers, not application code. **Request for direction:** Please clarify where these FP4 type definitions should live: - `lib/Support/` (LLVM infrastructure, alongside APFloat) - `lib/Target/AMDGPU/` (backend-only) - Some other location you recommend I will move them wherever you direct. The rest of this PR (WMMA256bInsts, SISchedule, TargetParser) is backend-only and uncontroversial. https://github.com/llvm/llvm-project/pull/201810 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
