[MLIR][Vector] Implement TransferOpReduceRank as MaskableOpRewritePattern

Author: nujaa

mlir/lib/Dialect/Vector/Transforms/LowerVectorTransfer.cpp

@@ -322,14 +322,19 @@ struct TransferWriteNonPermutationLowering
 ///     %v = vector.transfer_read ...
 ///         permutation_map: (d0, d1, d2, d3) -> (d1, 0, d3)
 ///     vector.broadcast %v
-struct TransferOpReduceRank : public OpRewritePattern<vector::TransferReadOp> {
-  using OpRewritePattern::OpRewritePattern;
+struct TransferOpReduceRank
+    : public MaskableOpRewritePattern<vector::TransferReadOp> {
+  using MaskableOpRewritePattern::MaskableOpRewritePattern;
 
-  LogicalResult matchAndRewrite(vector::TransferReadOp op,
-                                PatternRewriter &rewriter) const override {
+  FailureOr<mlir::Value>
+  matchAndRewriteMaskableOp(vector::TransferReadOp op,
+                            MaskingOpInterface maskOp,
+                            PatternRewriter &rewriter) const override {
     // TODO: support 0-d corner case.
     if (op.getTransferRank() == 0)
       return rewriter.notifyMatchFailure(op, "0-d corner case not supported");
+    if (maskOp)
+      return rewriter.notifyMatchFailure(op, "Masked case not supported");
 
     AffineMap map = op.getPermutationMap();
     unsigned numLeadingBroadcast = 0;
@@ -369,9 +374,9 @@ struct TransferOpReduceRank : public OpRewritePattern<vector::TransferReadOp> {
             op.getLoc(), originalVecType.getElementType(), op.getSource(),
             op.getIndices());
       }
-      rewriter.replaceOpWithNewOp<vector::BroadcastOp>(op, originalVecType,
-                                                       newRead);
-      return success();
+      return rewriter
+          .create<vector::BroadcastOp>(op.getLoc(), originalVecType, newRead)
+          .getVector();
     }
 
     SmallVector<int64_t> newShape(
@@ -393,9 +398,9 @@ struct TransferOpReduceRank : public OpRewritePattern<vector::TransferReadOp> {
         op.getLoc(), newReadType, op.getSource(), op.getIndices(),
         AffineMapAttr::get(newMap), op.getPadding(), op.getMask(),
         newInBoundsAttr);
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(op, originalVecType,
-                                                     newRead);
-    return success();
+    return rewriter
+        .create<vector::BroadcastOp>(op.getLoc(), originalVecType, newRead)
+        .getVector();
   }
 };
 

mlir/test/Dialect/Vector/vector-transfer-to-vector-load-store.mlir

@@ -219,6 +219,21 @@ func.func @transfer_broadcasting_2D(%mem : memref<8x8xf32>, %i : index) -> vecto
   return %res : vector<4x4xf32>
 }
 
+// CHECK-LABEL:   func @masked_transfer_read_reduce_rank_with_broadcast(
+// CHECK-SAME:                                %[[MEM:.*]]: memref<8x8x8x8xf32>,
+// CHECK-SAME:                                %[[MASK:.*]]: vector<4x4xi1>,
+// CHECK-SAME:                                %[[IDX:.*]]: index) -> vector<4x4x4x4xf32> {
+//      CHECK:      %[[RES:.*]] = vector.mask %[[MASK]] { vector.transfer_read %[[MEM]][%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]], %cst {in_bounds = [true, true, true, true], permutation_map = #map2} : memref<8x8x8x8xf32>, vector<4x4x4x4xf32> } : vector<4x4xi1> -> vector<4x4x4x4xf32>
+// CHECK-NEXT:      return %[[RES]] : vector<4x4x4x4xf32>
+#rank_reducing = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, d3)>
+func.func @masked_transfer_read_reduce_rank_with_broadcast(%mem : memref<8x8x8x8xf32>, %mask : vector<4x4xi1>, %i : index) -> vector<4x4x4x4xf32> {
+  %cf0 = arith.constant 0.0 : f32
+  %res = vector.mask %mask {vector.transfer_read %mem[%i, %i, %i, %i], %cf0
+    {in_bounds = [true, true, true, true], permutation_map = #rank_reducing}
+      : memref<8x8x8x8xf32>, vector<4x4x4x4xf32>} : vector<4x4xi1> -> vector<4x4x4x4xf32>
+  return %res : vector<4x4x4x4xf32>
+}
+
 // More complex broadcasting case (here a `vector.load` is generated).
 // CHECK-LABEL:   func @transfer_broadcasting_complex(
 // CHECK-SAME:                                %[[MEM:.*]]: memref<10x20x30x8x8xf32>,