[Mlir-commits] [mlir] [mlir][rocdl] Add cvt scale intrinsics (PR #159649)

[Erick Ochoa Lopez]

================
@@ -1298,6 +1298,54 @@ llvm.func @rocdl_last_use(%ptr: !llvm.ptr<1>) -> i32 {
   llvm.return %ret : i32
 }
 
+// CHECK-LABEL: rocdl.cvt.scale.pk8
+// CHECK-SAME:(i32 %[[I32:.+]], <2 x i32> %[[V2I32:.+]], i32 %[[SCALE:.+]])
+llvm.func @rocdl.cvt.scale.pk8(%i32: i32, %v2xi32: vector<2xi32>, %scale: i32) {
+
+  // CHECK:   call <8 x half> @llvm.amdgcn.cvt.scale.pk8.f16.fp4(i32 %[[I32]], i32 %[[SCALE]], i32 0)
+  %0 =                               rocdl.cvt.scale.pk8.f16.fp4 %i32, %scale[0] : vector<8xf16>
+  // CHECK: call <8 x bfloat> @llvm.amdgcn.cvt.scale.pk8.bf16.fp4(i32 %[[I32]], i32 %[[SCALE]], i32 0)
+  %1 =                               rocdl.cvt.scale.pk8.bf16.fp4 %i32, %scale[0] : vector<8xbf16>
+  // CHECK:  call <8 x float> @llvm.amdgcn.cvt.scale.pk8.f32.fp4(i32 %[[I32]], i32 %[[SCALE]], i32 0)
+  %2 =                               rocdl.cvt.scale.pk8.f32.fp4 %i32, %scale[0] : vector<8xf32>
+
+  // CHECK:   call <8 x half> @llvm.amdgcn.cvt.scale.pk8.f16.fp8(<2 x i32> %[[V2I32]], i32 %[[SCALE]], i32 0)
+  %3 =                               rocdl.cvt.scale.pk8.f16.fp8 %v2xi32, %scale[0] : vector<8xf16>
+  // CHECK: call <8 x bfloat> @llvm.amdgcn.cvt.scale.pk8.bf16.fp8(<2 x i32> %[[V2I32]], i32 %[[SCALE]], i32 0)
+  %4 =                               rocdl.cvt.scale.pk8.bf16.fp8 %v2xi32, %scale[0] : vector<8xbf16>
+  // CHECK:   call <8 x float> @llvm.amdgcn.cvt.scale.pk8.f32.fp8(<2 x i32> %[[V2I32]], i32 %[[SCALE]], i32 0)
+  %5 =                                rocdl.cvt.scale.pk8.f32.fp8 %v2xi32, %scale[0] : vector<8xf32>
----------------
amd-eochoalo wrote:

I fixed the indentation here https://github.com/llvm/llvm-project/pull/159649/commits/5d52e0337919212a69d9e20d5b4c52d6f8d7b0c7 . I wanted to align them to easily verify the check with the operation below it, but I am happy to change it back.

https://github.com/llvm/llvm-project/pull/159649