[llvm] [AMDGPU] Allow dpp in v_pk_fmac_f16 for GFX9 and GFX10 (PR #144782)
Frederik Harwath via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 10 03:27:02 PDT 2025
================
@@ -2172,6 +2172,7 @@ defm V_LDEXP_F16 : VOP2_Real_gfx10<0x03b>;
let IsSingle = 1 in {
defm V_PK_FMAC_F16 : VOP2_Real_e32_gfx10<0x03c>;
}
+defm V_PK_FMAC_F16 : VOP2_Real_dpp_gfx10<0x03c>, VOP2_Real_dpp8_gfx10<0x03c>;
----------------
frederik-h wrote:
I am working on fixing a problem with the representation of the source modifiers in the compiler - `neg` and `sext` using the same enum values although it would technically be possible to support both at the same time. The test cases using the `sext` modifier on `v_pk_fmac_f16` introduced by this PR are the only ones which cause problems to my changes.
> Actually gfx9 does support SDWA for this instruction.
Are you sure? The ISA docs list the `OPF_NOSDWA` flag for `v_pk_fmac_f16` and I don't understand what a SDWA variant would do. Is it a bug in the compiler that this gets accepted, as the test cases from this PR show, or is the SDWA variant of this instruction in fact supported by the hardware somehow, contrary to what the documentation says?
https://github.com/llvm/llvm-project/pull/144782
More information about the llvm-commits
mailing list