[mlir][AMDGPU] Add canonicalizer for folding casts into gather_to_lds

Author: qedawkins

mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td

@@ -921,6 +921,7 @@ def AMDGPU_GatherToLDSOp :
     $src `[` $srcIndices `]` `,` $dst `[` $dstIndices `]` attr-dict `:` $transferType `,` type($src) `,` type($dst)
   }];
   let hasVerifier = 1;
+  let hasCanonicalizer = 1;
 }
 
 def AMDGPU_TransposeLoadOp :

mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp

@@ -510,6 +510,10 @@ LogicalResult DPPOp::verify() {
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// GatherToLDSOp
+//===----------------------------------------------------------------------===//
+
 LogicalResult GatherToLDSOp::verify() {
   MemRefType srcType = cast<MemRefType>(getSrc().getType());
   MemRefType dstType = cast<MemRefType>(getDst().getType());
@@ -546,6 +550,59 @@ LogicalResult GatherToLDSOp::verify() {
   return success();
 }
 
+namespace {
+/// If the source/target of a CopyOp is a CastOp that does not modify the shape
+/// and element type, the cast can be skipped. Such CastOps only cast the layout
+/// of the type.
+struct FoldGatherToLDSOfCast : public OpRewritePattern<GatherToLDSOp> {
+  using OpRewritePattern<GatherToLDSOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(GatherToLDSOp gatherOp,
+                                PatternRewriter &rewriter) const override {
+    bool modified = false;
+
+    // Check source.
+    if (auto castOp = gatherOp.getSrc().getDefiningOp<memref::CastOp>()) {
+      auto fromType = llvm::dyn_cast<MemRefType>(castOp.getSource().getType());
+      auto toType = llvm::dyn_cast<MemRefType>(castOp.getSource().getType());
+
+      if (fromType && toType &&
+          fromType.getElementType() == toType.getElementType()) {
+        rewriter.modifyOpInPlace(gatherOp, [&] {
+          gatherOp.getSrcMutable().assign(castOp.getSource());
+        });
+        modified = true;
+      }
+    }
+
+    // Check target.
+    if (auto castOp = gatherOp.getDst().getDefiningOp<memref::CastOp>()) {
+      auto fromType = llvm::dyn_cast<MemRefType>(castOp.getSource().getType());
+      auto toType = llvm::dyn_cast<MemRefType>(castOp.getSource().getType());
+
+      if (fromType && toType &&
+          fromType.getElementType() == toType.getElementType()) {
+        rewriter.modifyOpInPlace(gatherOp, [&] {
+          gatherOp.getDstMutable().assign(castOp.getSource());
+        });
+        modified = true;
+      }
+    }
+
+    return success(modified);
+  }
+};
+} // namespace
+
+void GatherToLDSOp::getCanonicalizationPatterns(RewritePatternSet &results,
+                                                MLIRContext *context) {
+  results.add<FoldGatherToLDSOfCast>(context);
+}
+
+//===----------------------------------------------------------------------===//
+// TransposeLoadOp
+//===----------------------------------------------------------------------===//
+
 LogicalResult TransposeLoadOp::verify() {
   MemRefType srcType = cast<MemRefType>(getSrc().getType());
 

mlir/test/Dialect/AMDGPU/canonicalize.mlir

@@ -130,3 +130,17 @@ func.func @dead_atomic_add(%arg0: memref<4xf32>, %arg1: f32) {
   amdgpu.raw_buffer_atomic_fadd {boundsCheck = true} %arg1 -> %arg0[%c4_i32] : f32 -> memref<4xf32>, i32
   func.return
 }
+
+// -----
+
+// CHECK-LABEL: func @fold_gather_to_lds_of_cast
+func.func @fold_gather_to_lds_of_cast(%global: memref<128x72xf32, 1>, %lds: memref<64x64xf32, 3>) {
+// CHECK-SAME: %[[GLOBAL:[A-Za-z0-9]+]]: memref<128x72xf32, 1>
+  %c0 = arith.constant 0 : index
+  %0 = memref.cast %global : memref<128x72xf32, 1> to memref<?x?xf32, 1>
+  // CHECK: amdgpu.gather_to_lds %[[GLOBAL]]
+  // CHECK-SAME: : f32, memref<128x72xf32, 1>
+  amdgpu.gather_to_lds %0[%c0, %c0], %lds[%c0, %c0]
+    : f32, memref<?x?xf32, 1>, memref<64x64xf32, 3>
+  func.return
+}