[mlir][gpu] Patterns to promote gpu.shuffle to specialized AMDGPU ops

Only swizzle promotion for now, may add DPP ops support later.

Author: Hardcode84

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

@@ -132,24 +132,24 @@ def MapNestedForallToThreads :
      TransformEachOpTrait,
      TransformOpInterface]> {
   let description = [{
-      Target the `gpu.launch op` and rewrite all `scf.forall` nested in it to 
+      Target the `gpu.launch op` and rewrite all `scf.forall` nested in it to
       distributed `gpu.thread_id` attribute.
 
       The operation searches for `scf.forall` ops nested under `target` and maps
-      each such op to GPU threads. 
-      
+      each such op to GPU threads.
+
       `scf.forall` induction variables are rewritten to `gpu.thread_id` according
       to the `mapping` attribute.
 
       Different types of mappings attributes are supported:
         - the block_dims is a list of integers that specifies the number of
           threads in each dimension. This is a mandatory attribute that is used
-          to constrain the number of threads in each dimension. If an 
+          to constrain the number of threads in each dimension. If an
           `scf.forall` op is mapped to fewer threads, predication occurs.
         - the warp_dims is a list of integers that specifies the number of
           warps in each dimension. This is an optional attribute that is used
           to constrain the number of warps in each dimension. When present, this
-          attribute must be specified in a way that is compatible with the 
+          attribute must be specified in a way that is compatible with the
           block_dims attribute. If an `scf.forall` op is mapped to fewer warps,
           predication occurs.
 
@@ -164,7 +164,7 @@ def MapNestedForallToThreads :
       inserted after each scf.forall op. At this time, this is an all or nothing
       choice. This will need to be tightened in the future.
 
-      The operation alters the block size of the given gpu_launch using the 
+      The operation alters the block size of the given gpu_launch using the
       mandatory block_dims argument.
 
       #### Return modes:
@@ -268,7 +268,7 @@ def MapForallToBlocks :
     Only scf.forall distributed to **at most 3 dimensions** are
     currently supported.
 
-    The operation alters the block size of the given gpu_launch using the 
+    The operation alters the block size of the given gpu_launch using the
     grid_dims argument.
 
     #### Return modes:
@@ -300,7 +300,7 @@ def MapForallToBlocks :
     `:` functional-type($target, $result)
   }];
   let hasVerifier = 1;
-  
+
   let extraClassDeclaration = [{
     ::mlir::DiagnosedSilenceableFailure applyToOne(
         ::mlir::transform::TransformRewriter &rewriter,
@@ -310,4 +310,15 @@ def MapForallToBlocks :
   }];
 }
 
+def ApplyGPUPromoteShuffleToAMDGPUPatternsOp : Op<Transform_Dialect,
+    "apply_patterns.gpu.gpu_shuffle_to_amdgpu",
+    [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {
+  let description = [{
+    Collects patterns that are tryin to promote `gpu.shuffle`s to specialized
+    AMDGPU intrinsics.
+  }];
+  let assemblyFormat = "attr-dict";
+}
+
+
 #endif // GPU_TRANSFORM_OPS

mlir/include/mlir/Dialect/GPU/Transforms/Passes.h

@@ -94,6 +94,9 @@ void populateGpuDecomposeMemrefsPatterns(RewritePatternSet &patterns);
 /// Erase barriers that do not enforce conflicting memory side effects.
 void populateGpuEliminateBarriersPatterns(RewritePatternSet &patterns);
 
+/// Tries to promote `gpu.shuffle`s to specialized AMDGPU intrinsics.
+void populateGpuPromoteShuffleToAMDGPUPatterns(RewritePatternSet &patterns);
+
 /// Generate the code for registering passes.
 #define GEN_PASS_REGISTRATION
 #include "mlir/Dialect/GPU/Transforms/Passes.h.inc"

mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp

@@ -150,8 +150,6 @@ struct GPUShuffleOpLowering : public ConvertOpToLLVMPattern<gpu::ShuffleOp> {
         rewriter.create<LLVM::AndOp>(loc, int32Type, add, negwidth);
     Value dstLane;
     // TODO: Add support for gpu::ShuffleMode::UP and gpu::ShuffleMode::DOWN.
-    // TODO: Use ds_swizzle for XOR when step/offsets are constants for better
-    // perf.
     switch (op.getMode()) {
     case gpu::ShuffleMode::DOWN:
       dstLane = rewriter.create<LLVM::AddOp>(loc, int32Type, srcLaneId,

mlir/lib/Dialect/GPU/CMakeLists.txt

@@ -37,9 +37,10 @@ add_mlir_dialect_library(MLIRGPUTransforms
   Transforms/ModuleToBinary.cpp
   Transforms/NVVMAttachTarget.cpp
   Transforms/ParallelLoopMapper.cpp
+  Transforms/PromoteShuffleToAMDGPU.cpp
   Transforms/ROCDLAttachTarget.cpp
-  Transforms/ShuffleRewriter.cpp
   Transforms/SPIRVAttachTarget.cpp
+  Transforms/ShuffleRewriter.cpp
   Transforms/SubgroupReduceLowering.cpp
 
   OBJECT
@@ -52,6 +53,7 @@ add_mlir_dialect_library(MLIRGPUTransforms
   MLIRParallelLoopMapperEnumsGen
 
   LINK_LIBS PUBLIC
+  MLIRAMDGPUDialect
   MLIRAffineUtils
   MLIRArithDialect
   MLIRAsyncDialect
@@ -66,11 +68,11 @@ add_mlir_dialect_library(MLIRGPUTransforms
   MLIRMemRefDialect
   MLIRNVVMTarget
   MLIRPass
+  MLIRROCDLTarget
   MLIRSCFDialect
-  MLIRSideEffectInterfaces
   MLIRSPIRVTarget
+  MLIRSideEffectInterfaces
   MLIRSupport
-  MLIRROCDLTarget
   MLIRTransformUtils
   MLIRVectorDialect
   )

mlir/lib/Dialect/GPU/TransformOps/GPUTransformOps.cpp

@@ -11,6 +11,7 @@
 #include "mlir/Conversion/GPUCommon/GPUCommonPass.h"
 #include "mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h"
 #include "mlir/Conversion/LLVMCommon/TypeConverter.h"
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
 #include "mlir/Dialect/Affine/IR/AffineOps.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
@@ -136,6 +137,11 @@ void ApplyGPURewritePatternsOp::populatePatterns(RewritePatternSet &patterns) {
   populateGpuRewritePatterns(patterns);
 }
 
+void transform::ApplyGPUPromoteShuffleToAMDGPUPatternsOp::populatePatterns(
+    RewritePatternSet &patterns) {
+  populateGpuPromoteShuffleToAMDGPUPatterns(patterns);
+}
+
 //===----------------------------------------------------------------------===//
 // ApplyUnrollVectorsSubgroupMmaOp
 //===----------------------------------------------------------------------===//
@@ -914,9 +920,10 @@ class GPUTransformDialectExtension
   MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(GPUTransformDialectExtension)
 
   GPUTransformDialectExtension() {
-    declareGeneratedDialect<scf::SCFDialect>();
-    declareGeneratedDialect<arith::ArithDialect>();
     declareGeneratedDialect<GPUDialect>();
+    declareGeneratedDialect<amdgpu::AMDGPUDialect>();
+    declareGeneratedDialect<arith::ArithDialect>();
+    declareGeneratedDialect<scf::SCFDialect>();
     registerTransformOps<
 #define GET_OP_LIST
 #include "mlir/Dialect/GPU/TransformOps/GPUTransformOps.cpp.inc"

mlir/lib/Dialect/GPU/Transforms/PromoteShuffleToAMDGPU.cpp

@@ -0,0 +1,64 @@
+//===- PromoteShuffleToAMDGPU.cpp - Promote shuffle to AMDGPU -------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains patterns to try to promote `gpu.shuffle`s to specialized
+// AMDGPU intrinsics.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/GPU/Transforms/Passes.h"
+
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/GPU/IR/GPUDialect.h"
+#include "mlir/IR/PatternMatch.h"
+
+using namespace mlir;
+
+namespace {
+/// Try to promote `gpu.shuffle` to `amdgpu.swizzle_bitmode`, width must be 64
+/// and offset must be a constant integer in the range [0, 31].
+struct PromoteShuffleToSwizzlePattern
+    : public OpRewritePattern<gpu::ShuffleOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(gpu::ShuffleOp op,
+                                PatternRewriter &rewriter) const override {
+    if (op.getMode() != gpu::ShuffleMode::XOR)
+      return rewriter.notifyMatchFailure(op,
+                                         "only xor shuffle mode is supported");
+
+    if (!isConstantIntValue(op.getWidth(), 64))
+      return rewriter.notifyMatchFailure(op,
+                                         "only 64 width shuffle is supported");
+
+    std::optional<int64_t> offset = getConstantIntValue(op.getOffset());
+    if (!offset)
+      return rewriter.notifyMatchFailure(op,
+                                         "offset must be a constant integer");
+
+    int64_t offsetValue = *offset;
+    if (offsetValue < 0 || offsetValue >= 32)
+      return rewriter.notifyMatchFailure(op,
+                                         "offset must be in the range [0, 31]");
+
+    Location loc = op.getLoc();
+    Value res = rewriter.create<amdgpu::SwizzleBitModeOp>(
+        loc, op.getResult(0).getType(), op.getValue(), /*andMask=*/31,
+        /*orMask=*/0, /*xorMask=*/offsetValue);
+    Value valid = rewriter.create<arith::ConstantIntOp>(loc, 1, /*width*/ 1);
+    rewriter.replaceOp(op, {res, valid});
+    return success();
+  }
+};
+} // namespace
+
+void mlir::populateGpuPromoteShuffleToAMDGPUPatterns(
+    RewritePatternSet &patterns) {
+  patterns.add<PromoteShuffleToSwizzlePattern>(patterns.getContext());
+}

mlir/test/Dialect/GPU/promote-shuffle-amdgpu.mlir

@@ -0,0 +1,23 @@
+// RUN: mlir-opt --transform-interpreter --split-input-file %s | FileCheck %s
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%module_op: !transform.any_op {transform.readonly}) {
+    %func = transform.structured.match ops{["func.func"]} in %module_op : (!transform.any_op) -> !transform.any_op
+    transform.apply_patterns to %func {
+      transform.apply_patterns.gpu.gpu_shuffle_to_amdgpu
+    } : !transform.any_op
+    transform.yield
+  }
+}
+
+  // CHECK-LABEL: func @gpu_shuffle_swizzle
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+func.func @gpu_shuffle_swizzle(%arg0: i32) -> (i32, i1) {
+  // CHECK:  %[[TRUE:.*]] = arith.constant true
+  // CHECK:  %[[RES:.*]] = amdgpu.swizzle_bitmode %[[ARG]] 31 0 23 : i32
+  // CHECK:  return %[[RES]], %[[TRUE]] : i32, i1
+  %width = arith.constant 64 : i32
+  %offset = arith.constant 23 : i32
+  %shfl, %pred = gpu.shuffle xor %arg0, %offset, %width : i32
+  func.return %shfl, %pred : i32, i1
+}