[AMD] Fix matmul unit test failure on RDNA4 (#7631)

This PR fixes language/test_matmul.py::test_simple_matmul failures for
fp64 on RDNA4.
There were 4 failing test cases of the form
test_matmul.py::test_simple_matmul[float64 - float64], These failures
were specific to the RDNA4 backend and are now resolved by this PR.

Author: saeid-rostami

third_party/amd/lib/TritonAMDGPUToLLVM/DotOpToLLVM/FMA.cpp

@@ -60,6 +60,8 @@ class AMDFMAVectorMultiplier : public FMAVectorMultiplier {
            dElemTy);
     chosenOp.vectorSize = 1;
     chosenOp.outElemTy = aElemTy;
+    if (aElemTy.isF64())
+      chosenOp.intrinsicName = "llvm.fmuladd.f64";
     if (aElemTy.isF32())
       chosenOp.intrinsicName = "llvm.fmuladd.f32";
     if (aElemTy.isF16())

third_party/amd/lib/TritonAMDGPUTransforms/AccelerateAMDMatmul.cpp

@@ -1194,7 +1194,7 @@ class AccelerateBlocked : public OpRewritePattern<DotOp> {
       auto elTy = opTy.getElementType();
       if (elTy != expectedElTy)
         return false;
-      if (!elTy.isF16() && !elTy.isF32())
+      if (!elTy.isF16() && !elTy.isF32() && !elTy.isF64())
         return false;
     }
     return true;