[Mlir-commits] [mlir] [NVVM] Add FP8 (e4m3/e5m2) support to dense mma.sync (PR #207307)

[Guray Ozen]

================
@@ -210,6 +210,30 @@ func.func @nvvm_mma_m16n8k16_bf16_bf16(%a0 : i32, %a1 : i32, %a2 : i32, %a3 : i3
   llvm.return %0 : !llvm.struct<(f32, f32, f32, f32)>
 }
 
+// CHECK-LABEL: @nvvm_mma_m16n8k16_e4m3_e4m3
+func.func @nvvm_mma_m16n8k16_e4m3_e4m3(%a0 : i32, %a1 : i32,
+                              %b0 : i32,
+                              %c0 : f32, %c1 : f32, %c2 : f32, %c3 : f32) -> !llvm.struct<(f32, f32, f32, f32)> {
+  // CHECK: nvvm.mma.sync A[{{.*}}] B[{{.*}}] C[{{.*}}] {layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<e4m3>, multiplicandBPtxType = #nvvm.mma_type<e4m3>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (i32, i32, f32) -> !llvm.struct<(f32, f32, f32, f32)>
+  %0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
----------------
grypp wrote:

can you add tests for mlir-translate in nvvmir.mlir

https://github.com/llvm/llvm-project/pull/207307