[clang] [llvm] AMDGPU: Support v_wmma_f32_16x16x128_f8f6f4 on gfx1250 (PR #149684)

[Changpeng Fang via cfe-commits]

================
@@ -6627,6 +6627,54 @@ void Verifier::visitIntrinsicCall(Intrinsic::ID ID, CallBase &Call) {
           "invalid vector type for format", &Call, Src1, Call.getArgOperand(5));
     break;
   }
+  case Intrinsic::amdgcn_wmma_f32_16x16x128_f8f6f4: {
+    Value *Src0 = Call.getArgOperand(1);
+    Value *Src1 = Call.getArgOperand(3);
+
+    unsigned FmtA = cast<ConstantInt>(Call.getArgOperand(0))->getZExtValue();
+    unsigned FmtB = cast<ConstantInt>(Call.getArgOperand(2))->getZExtValue();
+    Check(FmtA <= 4, "invalid value for matrix format", Call,
+          Call.getArgOperand(0));
+    Check(FmtB <= 4, "invalid value for matrix format", Call,
+          Call.getArgOperand(2));
+
+    // AMDGPU::MatrixFMT values
----------------
changpeng wrote:

> Most of this looks identical to the gfx950 mfma one, is it possible to merge them

The operand layout is different from mfma, and merging them needs operand selection in multiple places.
Maybe keeping the current (separate) implementation is better for reading and understanding.

https://github.com/llvm/llvm-project/pull/149684