[mlir][GPUToNVVM] Add benefit to populate functions

mlir/include/mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h

@@ -10,6 +10,7 @@
 
 #include "mlir/Conversion/LLVMCommon/LoweringOptions.h"
 #include "mlir/Dialect/LLVMIR/LLVMTypes.h"
+#include "mlir/IR/PatternMatch.h"
 #include <memory>
 
 namespace mlir {
@@ -35,18 +36,27 @@ void configureGpuToNVVMConversionLegality(ConversionTarget &target);
 /// GPU dialect to NVVM.
 void configureGpuToNVVMTypeConverter(LLVMTypeConverter &converter);
 
+/// Populate patterns that lower certain arith and math dialect ops to
+/// libdevice calls.
+void populateLibDeviceConversionPatterns(const LLVMTypeConverter &converter,
+                                         RewritePatternSet &patterns,
+                                         PatternBenefit benefit = 1);
+
 /// Collect a set of patterns to convert from the GPU dialect to NVVM.
 void populateGpuToNVVMConversionPatterns(const LLVMTypeConverter &converter,
-                                         RewritePatternSet &patterns);
+                                         RewritePatternSet &patterns,
+                                         PatternBenefit benefit = 1);
 
 /// Populate GpuSubgroupReduce pattern to NVVM. It generates a specific nvvm
 /// op that is not available on every GPU.
 void populateGpuSubgroupReduceOpLoweringPattern(
-    const LLVMTypeConverter &converter, RewritePatternSet &patterns);
+    const LLVMTypeConverter &converter, RewritePatternSet &patterns,
+    PatternBenefit benefit = 1);
 
 /// Collect a set of patterns to convert WMMA ops from GPU dialect to NVVM.
 void populateGpuWMMAToNVVMConversionPatterns(const LLVMTypeConverter &converter,
-                                             RewritePatternSet &patterns);
+                                             RewritePatternSet &patterns,
+                                             PatternBenefit benefit = 1);
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_GPUTONVVM_GPUTONVVMPASS_H_

mlir/include/mlir/Dialect/GPU/TransformOps/GPUTransformOps.td

@@ -26,6 +26,7 @@ def ApplyGPUToNVVMConversionPatternsOp : Op<Transform_Dialect,
     Collects patterns that convert GPU dialect ops to NVVM dialect ops. These
     patterns require an "LLVMTypeConverter".
   }];
+  let arguments = (ins DefaultValuedAttr<I16Attr, "1">:$benefit);
   let assemblyFormat = "attr-dict";
 }
 

mlir/lib/Conversion/GPUCommon/GPUOpsLowering.h

@@ -43,8 +43,9 @@ struct GPUDynamicSharedMemoryOpLowering
   using ConvertOpToLLVMPattern<
       gpu::DynamicSharedMemoryOp>::ConvertOpToLLVMPattern;
   GPUDynamicSharedMemoryOpLowering(const LLVMTypeConverter &converter,
-                                   unsigned alignmentBit = 0)
-      : ConvertOpToLLVMPattern<gpu::DynamicSharedMemoryOp>(converter),
+                                   unsigned alignmentBit = 0,
+                                   PatternBenefit benefit = 1)
+      : ConvertOpToLLVMPattern<gpu::DynamicSharedMemoryOp>(converter, benefit),
         alignmentBit(alignmentBit) {}
 
   LogicalResult
@@ -81,8 +82,9 @@ struct GPUFuncOpLoweringOptions {
 
 struct GPUFuncOpLowering : ConvertOpToLLVMPattern<gpu::GPUFuncOp> {
   GPUFuncOpLowering(const LLVMTypeConverter &converter,
-                    const GPUFuncOpLoweringOptions &options)
-      : ConvertOpToLLVMPattern<gpu::GPUFuncOp>(converter),
+                    const GPUFuncOpLoweringOptions &options,
+                    PatternBenefit benefit = 1)
+      : ConvertOpToLLVMPattern<gpu::GPUFuncOp>(converter, benefit),
         allocaAddrSpace(options.allocaAddrSpace),
         workgroupAddrSpace(options.workgroupAddrSpace),
         kernelAttributeName(options.kernelAttributeName),

mlir/lib/Conversion/GPUCommon/IndexIntrinsicsOpLowering.h

@@ -36,14 +36,16 @@ struct OpLowering : public ConvertOpToLLVMPattern<Op> {
   IntrType intrType;
 
 public:
-  explicit OpLowering(const LLVMTypeConverter &typeConverter)
-      : ConvertOpToLLVMPattern<Op>(typeConverter),
+  explicit OpLowering(const LLVMTypeConverter &typeConverter,
+                      PatternBenefit benefit = 1)
+      : ConvertOpToLLVMPattern<Op>(typeConverter, benefit),
         indexBitwidth(typeConverter.getIndexTypeBitwidth()),
         indexKind(IndexKind::Other), intrType(IntrType::None) {}
 
   explicit OpLowering(const LLVMTypeConverter &typeConverter,
-                      IndexKind indexKind, IntrType intrType)
-      : ConvertOpToLLVMPattern<Op>(typeConverter),
+                      IndexKind indexKind, IntrType intrType,
+                      PatternBenefit benefit = 1)
+      : ConvertOpToLLVMPattern<Op>(typeConverter, benefit),
         indexBitwidth(typeConverter.getIndexTypeBitwidth()),
         indexKind(indexKind), intrType(intrType) {}
 

mlir/lib/Conversion/GPUCommon/OpToFuncCallLowering.h

@@ -57,8 +57,9 @@ struct OpToFuncCallLowering : public ConvertOpToLLVMPattern<SourceOp> {
   explicit OpToFuncCallLowering(const LLVMTypeConverter &lowering,
                                 StringRef f32Func, StringRef f64Func,
                                 StringRef f32ApproxFunc, StringRef f16Func,
-                                StringRef i32Func = "")
-      : ConvertOpToLLVMPattern<SourceOp>(lowering), f32Func(f32Func),
+                                StringRef i32Func = "",
+                                PatternBenefit benefit = 1)
+      : ConvertOpToLLVMPattern<SourceOp>(lowering, benefit), f32Func(f32Func),
         f64Func(f64Func), f32ApproxFunc(f32ApproxFunc), f16Func(f16Func),
         i32Func(i32Func) {}