[mlir][linalg][elementwise] Fold broadcast into new elementwise

Author: someoneinjd

mlir/lib/Dialect/Linalg/Transforms/FoldIntoElementwise.cpp

@@ -64,6 +64,41 @@ struct FoldTransposePattern : public OpRewritePattern<ElementwiseOp> {
   }
 };
 
+struct FoldBroadcastPattern : public OpRewritePattern<ElementwiseOp> {
+  using OpRewritePattern<ElementwiseOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(ElementwiseOp op,
+                                PatternRewriter &rewriter) const override {
+    bool changed = false;
+    SmallVector<Value> newIns;
+    SmallVector<AffineMap> newMaps;
+    for (OpOperand *operand : op.getDpsInputOperands()) {
+      AffineMap map = op.getMatchingIndexingMap(operand);
+      auto broadcastOp = operand->get().getDefiningOp<BroadcastOp>();
+
+      if (!map.isIdentity() || !broadcastOp) {
+        // push in original operand and its map.
+        newIns.push_back(operand->get());
+        newMaps.push_back(map);
+        continue;
+      }
+      newIns.push_back(broadcastOp.getInput());
+      // push in broadcastOp's broadcast map.
+      newMaps.push_back(broadcastOp.getMatchingIndexingMap(
+          broadcastOp.getDpsInputOperand(0)));
+      changed = true;
+    }
+    if (!changed)
+      return failure();
+    newMaps.push_back(op.getIndexingMapsArray().back());
+
+    rewriter.replaceOpWithNewOp<ElementwiseOp>(
+        op, newIns, op.getDpsInits()[0], op.getKindAttr(),
+        rewriter.getAffineMapArrayAttr(newMaps));
+    return success();
+  }
+};
+
 struct LinalgFoldIntoElementwisePass
     : public impl::LinalgFoldIntoElementwisePassBase<
           LinalgFoldIntoElementwisePass> {
@@ -84,4 +119,5 @@ struct LinalgFoldIntoElementwisePass
 void mlir::linalg::populateLinalgFoldIntoElementwisePatterns(
     RewritePatternSet &patterns) {
   patterns.add<FoldTransposePattern>(patterns.getContext());
+  patterns.add<FoldBroadcastPattern>(patterns.getContext());
 }

mlir/test/Dialect/Linalg/elementwise/fold.mlir

@@ -41,3 +41,47 @@ func.func @binary_transposed(%A : tensor<?x?xf32>, %B: tensor<?x?xf32>, %C: tens
                           outs(%C: tensor<?x?xf32>) -> tensor<?x?xf32>
   return %result : tensor<?x?xf32>
 }
+
+// -----
+
+// CHECK-DAG: #[[IDENTITY:.+]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
+// CHECK-DAG: #[[BROADCASTED:.+]] = affine_map<(d0, d1, d2) -> (d0, d2)>
+//
+// CHECK:  func.func @unary_broadcasted(%[[A:.+]]: tensor<8x32xf32>, %[[B:.+]]: tensor<8x16x32xf32>) -> tensor<8x16x32xf32> {
+// CHECK-NEXT:  %[[RES:.+]] = linalg.elementwise kind=#linalg.elementwise_kind<exp>
+// CHECK-SAME:       indexing_maps = [#[[BROADCASTED]], #[[IDENTITY]]]
+// CHECK-SAME:       ins(%[[A]] : tensor<8x32xf32>) outs(%[[B]] : tensor<8x16x32xf32>) -> tensor<8x16x32xf32>
+// CHECK-NEXT:    return %[[RES]] : tensor<8x16x32xf32>
+//
+func.func @unary_broadcasted(%A : tensor<8x32xf32>, %B: tensor<8x16x32xf32>) ->  tensor<8x16x32xf32> {
+  %empty = tensor.empty() : tensor<8x16x32xf32>
+  %broadcasted_A = linalg.broadcast ins(%A : tensor<8x32xf32>) outs(%empty :  tensor<8x16x32xf32>) dimensions = [1]
+  %result = linalg.elementwise kind=#linalg.elementwise_kind<exp>
+                          ins(%broadcasted_A : tensor<8x16x32xf32>) outs(%B: tensor<8x16x32xf32>) -> tensor<8x16x32xf32>
+  return %result : tensor<8x16x32xf32>
+}
+
+// -----
+
+// CHECK-DAG: #[[IDENTITY:.+]] = affine_map<(d0, d1) -> (d0, d1)>
+// CHECK-DAG: #[[BROADCASTED:.+]] = affine_map<(d0, d1) -> (d0)>
+//
+// CHECK:  func.func @binary_broadcasted(%[[A:.+]]: tensor<?x?xf32>, %[[B:.+]]: tensor<?xf32>, %[[C:.+]]: tensor<?x?xf32>) -> tensor<?x?xf32> {
+// CHECK-NEXT:  %[[RES:.+]] = linalg.elementwise kind=#linalg.elementwise_kind<add>
+// CHECK-SAME:              indexing_maps = [#[[IDENTITY]], #[[BROADCASTED]], #[[IDENTITY]]]
+// CHECK-SAME:              ins(%[[A]], %[[B]] : tensor<?x?xf32>, tensor<?xf32>) outs(%[[C]] : tensor<?x?xf32>) -> tensor<?x?xf32>
+// CHECK-NEXT:  return %[[RES]] : tensor<?x?xf32>
+//
+func.func @binary_broadcasted(%A : tensor<?x?xf32>, %B: tensor<?xf32>, %C: tensor<?x?xf32>) ->  tensor<?x?xf32> {
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %dim0 = tensor.dim %A, %c0 : tensor<?x?xf32>
+  %dim1 = tensor.dim %A, %c1 : tensor<?x?xf32>
+
+  %empty = tensor.empty(%dim1, %dim0) : tensor<?x?xf32>
+  %broadcasted_B = linalg.broadcast ins(%B : tensor<?xf32>) outs(%empty :  tensor<?x?xf32>) dimensions = [1] 
+  %result = linalg.elementwise kind=#linalg.elementwise_kind<add>
+                          ins(%A, %broadcasted_B : tensor<?x?xf32>,  tensor<?x?xf32>)
+                          outs(%C: tensor<?x?xf32>) -> tensor<?x?xf32>
+  return %result : tensor<?x?xf32>
+}