Added new AMDGPU gfx950 features in Rocdl dialect.

New MFMA variants.
ds.read.r*.* and global.load.lds

Author: makslevental

third_party/amd/lib/TritonAMDGPUToLLVM/BufferOpsEmitter.cpp

@@ -1,10 +1,8 @@
 #include "PatternTritonGPUOpToLLVM.h"
 #include "TargetInfo.h"
-// clang-format off
-#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "Utility.h"
-// clang-format on
 #include "mlir/Dialect/LLVMIR/LLVMTypes.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "mlir/IR/PatternMatch.h"
 #include "triton/Conversion/TritonGPUToLLVM/TypeConverter.h"
 #include "triton/Dialect/Triton/IR/Dialect.h"

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/ConvertLayoutOpToLLVM.cpp

@@ -1,10 +1,8 @@
 #include "PatternTritonGPUOpToLLVM.h"
 #include "TargetInfo.h"
-// clang-format off
-#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "Utility.h"
-// clang-format on
 #include "mlir/Conversion/LLVMCommon/Pattern.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "mlir/IR/PatternMatch.h"
 #include "triton/Conversion/TritonGPUToLLVM/PatternTritonGPUOpToLLVM.h"
 #include "triton/Conversion/TritonGPUToLLVM/Utility.h"

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM/MMAv5.cpp

@@ -61,7 +61,8 @@ enum class mxfpKind { mxf8f6f4 = 0, mxf4 = 1, mxf4nvf4 = 2 };
 inline mxfpKind getMXFPKind(ScaleDotElemType typeA, ScaleDotElemType typeB,
                             Type scaleAType, Type scaleBType) {
   if (typeA == ScaleDotElemType::E2M1 && typeB == ScaleDotElemType::E2M1) {
-    if (scaleAType.isFloat8E4M3FN() && scaleBType.isFloat8E4M3FN()) {
+    if (llvm::isa<Float8E4M3FNType>(scaleAType) &&
+        llvm::isa<Float8E4M3FNType>(scaleBType)) {
       return mxfpKind::mxf4nvf4;
     }
     return mxfpKind::mxf4;
@@ -102,9 +103,9 @@ static Value createInstDescriptor(ConversionPatternRewriter &rewriter,
       return 1;
     if (type.isF32())
       return 2;
-    if (type.isFloat8E4M3FN())
+    if (llvm::isa<Float8E4M3FNType>(type))
       return 0;
-    if (type.isFloat8E5M2())
+    if (llvm::isa<Float8E5M2Type>(type))
       return 1;
     llvm_unreachable("Unsupported type.");
   };
@@ -227,7 +228,7 @@ static void createGen5MMA(ConversionPatternRewriter &rewriter, Location loc,
     opcode += "f16";
   else if (srcElementTy.isF32())
     opcode += "tf32";
-  else if (srcElementTy.isFloat8E4M3FN() || srcElementTy.isFloat8E5M2())
+  else if (llvm::isa<Float8E4M3FNType, Float8E5M2Type>(srcElementTy))
     opcode += "f8f6f4";
   else
     assert(0 && "Unsupported type.");

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/TargetInfo.cpp

@@ -1,12 +1,10 @@
 #include "TargetInfo.h"
 #include "Dialect/NVGPU/IR/Dialect.h"
 #include "TritonNVIDIAGPUToLLVM/PTXAsmFormat.h"
-// clang-format off
-#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "Utility.h"
-// clang-format on
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
 #include "mlir/Dialect/LLVMIR/LLVMTypes.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "llvm/Support/MathExtras.h"
 
 using namespace mlir;

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/Utility.cpp

@@ -1,8 +1,6 @@
-// clang-format off
-#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "Utility.h"
-// clang-format on
 #include "Dialect/NVGPU/IR/Dialect.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "triton/Conversion/TritonGPUToLLVM/TypeConverter.h"
 
 namespace mlir {