[Mlir-commits] [mlir] Extend GPU and NVVM mma ops to support fp64 (PR #165380)

[Guray Ozen]

================
@@ -463,6 +463,43 @@ llvm.func @nvvm_wmma_mma(%0 : i32, %1 : i32, %2 : i32, %3 : i32, %4 : i32, %5 :
   llvm.return
 }
 
+// CHECK-LABEL: @nvvm_wmma_load_a_f64
+llvm.func @nvvm_wmma_load_a_f64(%arg0: !llvm.ptr, %arg1 : i32) {
+  // CHECK: call double @llvm.nvvm.wmma.m8n8k4.load.a.row.stride.f64.p0(ptr %{{.*}}, i32 %{{.*}})
+  %0 = nvvm.wmma.load %arg0, %arg1
----------------
grypp wrote:

I recommend splitting the PR into 2 pieces. First PR is about `f64` in nvvm dialect, and second PR could using them in gpu dialect.

https://github.com/llvm/llvm-project/pull/165380