conversion tritonGPUToLLVM elementwiseOpToLLVMBase decoupling

Author: toudefu

third_party/iluvatar/backend/flagtree_backend_specialization/include/flagtree_spec.h

@@ -1,3 +1,4 @@
 #include "triton/Analysis/iluvatar_AxisInfo.h"
 #include "triton/Analysis/iluvatar_Membar.h"
 #include "triton/Analysis/iluvatar_Utility.h"
+#include "triton/Conversion/TritonGPUToLLVM/iluvatar_ElementwiseOpToLLVMBase.h"

third_party/iluvatar/backend/flagtree_backend_specialization/include/triton/Conversion/TritonGPUToLLVM/iluvatar_ElementwiseOpToLLVMBase.h

@@ -0,0 +1,7 @@
+#ifndef ILUVATAR_TRITON_CONVERSION_TRITONGPU_TO_ELEMENTWISE_OP_H
+#define ILUVATAR_TRITON_CONVERSION_TRITONGPU_TO_ELEMENTWISE_OP_H
+
+#define FLAGTREE_SPEC_ElementwiseOpConversionBase_maybeDeduplicate
+#define FLAGTREE_SPEC_ElementwiseOpConversionBase_matchAndRewrite
+
+#endif // ILUVATAR_TRITON_CONVERSION_TRITONGPU_TO_ELEMENTWISE_OP_H

third_party/iluvatar/include/triton/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVMBase.h

@@ -102,12 +102,14 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
       // test_core::test_fp8_dot_acc
       return resultVals;
     }
+#ifdef FLAGTREE_SPEC_ElementwiseOpConversionBase_maybeDeduplicate
     if (isa<IluvatarMmaEncodingAttr, DotOperandEncodingAttr>(baseEncoding)) {
       // TODO: this logic seems incorrect for mma layout. Skip for now.
       // The following test crashes and some other miscompile:
       // test_core::test_fp8_dot_acc
       return resultVals;
     }
+#endif
 
     SmallVector<unsigned> elemsPerThread = getElemsPerThread(rtType);
     int rank = elemsPerThread.size();
@@ -188,7 +190,7 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
     // element type
     auto resultElementTy = getElementTypeOrSelf(resultTy);
     Type elemTy = this->getTypeConverter()->convertType(resultElementTy);
-#ifdef __ILUVATAR__
+#ifdef FLAGTREE_SPEC_ElementwiseOpConversionBase_matchAndRewrite
     auto srcType = this->getTypeConverter()->convertType(resultTy);
     if (auto structTy = dyn_cast<LLVM::LLVMStructType>(srcType))
       elemTy = structTy.getBody()[0];