[clang] [llvm] AMDGPU: Add back half and bfloat support for global_load_tr16 pats (PR #99540)

Changpeng Fang via llvm-commits llvm-commits at lists.llvm.org
Thu Jul 18 11:33:22 PDT 2024


================
@@ -1590,10 +1590,14 @@ let OtherPredicates = [isGFX12Plus] in {
   let WaveSizePredicate = isWave32 in {
     defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B64_w32, int_amdgcn_global_load_tr_b64, v2i32>;
     defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B128_w32, int_amdgcn_global_load_tr_b128, v8i16>;
+    defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B128_w32, int_amdgcn_global_load_tr_b128, v8f16>;
+    defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B128_w32, int_amdgcn_global_load_tr_b128, v8bf16>;
   }
   let WaveSizePredicate = isWave64 in {
     defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B64_w64, int_amdgcn_global_load_tr_b64, i32>;
     defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B128_w64, int_amdgcn_global_load_tr_b128, v4i16>;
+    defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B128_w64, int_amdgcn_global_load_tr_b128, v4f16>;
+    defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR_B128_w64, int_amdgcn_global_load_tr_b128, v4bf16>;
----------------
changpeng wrote:

Do you mean we should loop over [v8i16, v8f16, v8bf16]? Other types may not be relevant to load transpose of 16-bit elements. 

https://github.com/llvm/llvm-project/pull/99540


More information about the llvm-commits mailing list