[mlir][vector] Add ElementwiseToOuterproduct

Author: nujaa

mlir/include/mlir/Dialect/Vector/IR/VectorOps.h

@@ -80,6 +80,10 @@ void populateVectorToVectorCanonicalizationPatterns(RewritePatternSet &patterns,
 /// into vector contract for the backends with native support.
 void populateFoldArithExtensionPatterns(RewritePatternSet &patterns);
 
+/// Collect a set of patterns that fold elementwise op on vectors to the vector
+/// dialect.
+void populateElementwiseToVectorOpsPatterns(RewritePatternSet &patterns);
+
 /// Returns the integer type required for subscripts in the vector dialect.
 IntegerType getVectorSubscriptType(Builder &builder);
 

mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td

@@ -392,6 +392,17 @@ def ApplyFoldArithExtensionPatternsOp : Op<Transform_Dialect,
   let assemblyFormat = "attr-dict";
 }
 
+def ApplyFoldElementwiseToVectorPatternsOp : Op<Transform_Dialect,
+    "apply_patterns.vector.elementwise_to_vector",
+    [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {
+  let description = [{
+    Collect a set of patterns that fold elementwise op on vectors to the vector 
+    dialect.
+  }];
+
+  let assemblyFormat = "attr-dict";
+}
+
 def ApplyVectorReductionToContractPatternsOp : Op<Transform_Dialect,
     "apply_patterns.vector.reduction_to_contract",
     [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {

mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp

@@ -59,6 +59,11 @@ void transform::ApplyFoldArithExtensionPatternsOp::populatePatterns(
   vector::populateFoldArithExtensionPatterns(patterns);
 }
 
+void transform::ApplyFoldElementwiseToVectorPatternsOp::populatePatterns(
+    RewritePatternSet &patterns) {
+  vector::populateElementwiseToVectorOpsPatterns(patterns);
+}
+
 void transform::ApplyVectorReductionToContractPatternsOp::populatePatterns(
     RewritePatternSet &patterns) {
   vector::populateVectorReductionToContractPatterns(patterns);

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

@@ -1795,6 +1795,75 @@ struct BreakDownVectorReduction final : OpRewritePattern<vector::ReductionOp> {
   unsigned maxNumElementsToExtract = 0;
 };
 
+/// Pattern aiming to fold a series of ops mulf(tr(broadcast(A)), broadcast(B))
+/// into vector.outerproduct(A, B) such as :
+/// ```mlir
+///  %lhsBcast = vector.broadcast %lhs : vector<4xi32> to vector<4x4xi32>
+///  %lhsT = vector.transpose %lhsBcast, [1, 0] : vector<4x4xi32> to
+///  vector<4x4xi32> %rhsBcast = vector.broadcast %rhs : vector<4xi32> to
+///  vector<4x4xi32> %mul = arith.muli %lhsT, %rhsBcast : vector<4x4xi32>
+///```
+/// Becomes :
+///```mlir
+///  %res = vector.outerproduct %lhs, %rhs : vector<4xi32>, vector<4xi32>
+///```
+/// Edge Cases where broadcast ops are not 1D to 2D as follow are not handled.
+/// %ex1 = vector.broadcast %lhsCast : vector<1x4xf32> to vector<4x4xf32>
+/// %ex2 = vector.broadcast %lhsCast : f32 to vector<4x4xf32>
+/// %ex3 = vector.broadcast %lhsCast : vector<1x1xf32> to vector<4x4xf32>
+
+template <typename MulOpType>
+struct ElementwiseToOuterproduct : public OpRewritePattern<MulOpType> {
+  using OpRewritePattern<MulOpType>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(MulOpType mulOp,
+                                PatternRewriter &rewriter) const override {
+    auto VT = llvm::cast<VectorType>(mulOp.getResult().getType());
+    if (!VT)
+      return failure();
+    if (VT.getRank() != 2)
+      return failure();
+
+    auto canonicalize = [&](Value OperandA,
+                            Value OperandB) -> vector::OuterProductOp {
+      vector::TransposeOp transposedLhs =
+          dyn_cast_or_null<vector::TransposeOp>(OperandA.getDefiningOp());
+      if (!transposedLhs)
+        return vector::OuterProductOp();
+      // Fail unless this is a true 2-D matrix transpose.
+      ArrayRef<int64_t> permutation = transposedLhs.getPermutation();
+      if (permutation[0] != 1 || permutation[1] != 0)
+        return vector::OuterProductOp();
+
+      // Fail in case it is not a 1-to-2 dimension to broadcast to avoid
+      // generating shape_casts/broadcasts which do not belong in this pattern.
+      vector::BroadcastOp broadcastedLhs = dyn_cast<vector::BroadcastOp>(
+          transposedLhs.getVector().getDefiningOp());
+      if (!broadcastedLhs ||
+          !broadcastedLhs.computeBroadcastedUnitDims().empty())
+        return vector::OuterProductOp();
+      // Avoid broadcast f32 or vector<f32> -> ResType
+      auto srcVT = dyn_cast<VectorType>(broadcastedLhs.getSourceType());
+      if (!srcVT || srcVT.getRank() != 1)
+        return vector::OuterProductOp();
+
+      vector::BroadcastOp broadcastedRhs =
+          dyn_cast<vector::BroadcastOp>(OperandB.getDefiningOp());
+      if (!broadcastedRhs || broadcastedRhs.getSourceType() != srcVT)
+        return vector::OuterProductOp();
+
+      return rewriter.replaceOpWithNewOp<vector::OuterProductOp>(
+          mulOp, VT, broadcastedLhs.getSource(), broadcastedRhs.getSource(),
+          Value(), vector::CombiningKind::ADD);
+    };
+    Value a = mulOp->getOperand(0), b = mulOp->getOperand(1);
+    vector::OuterProductOp outerP = canonicalize(a, b);
+    // Handle commutativity, the transposed op is the outerproduct LHS.
+    outerP = outerP ? outerP : canonicalize(b, a);
+    return outerP ? success() : failure();
+  }
+};
+
 } // namespace
 
 void mlir::vector::populateFoldArithExtensionPatterns(
@@ -1882,6 +1951,12 @@ void mlir::vector::populateBreakDownVectorReductionPatterns(
                                          maxNumElementsToExtract, benefit);
 }
 
+void mlir::vector::populateElementwiseToVectorOpsPatterns(
+    RewritePatternSet &patterns) {
+  patterns.add<ElementwiseToOuterproduct<arith::MulFOp>,
+               ElementwiseToOuterproduct<arith::MulIOp>>(patterns.getContext());
+}
+
 //===----------------------------------------------------------------------===//
 // TableGen'd enum attribute definitions
 //===----------------------------------------------------------------------===//

mlir/test/Dialect/Vector/transform-vector.mlir

@@ -92,3 +92,41 @@ module attributes {transform.with_named_sequence} {
     transform.yield
   }
 }
+
+// -----
+
+// CHECK-LABEL: func.func @ewise_outerproduct
+//  CHECK-SAME:   %[[LHS:.*]]: vector<[4]xi32>,
+//  CHECK-SAME:   %[[RHS:.*]]: vector<[4]xi32>) -> vector<[4]x[4]xi32> {
+//       CHECK:     %[[RES:.*]] = vector.outerproduct %[[LHS]], %[[RHS]] : vector<[4]xi32>, vector<[4]xi32>
+//       CHECK:     return %[[RES]] : vector<[4]x[4]xi32>
+func.func @ewise_outerproduct(%lhs: vector<[4]xi32>, %rhs: vector<[4]xi32>) -> vector<[4]x[4]xi32> {
+  %lhsBcast = vector.broadcast %lhs : vector<[4]xi32> to vector<[4]x[4]xi32>
+  %lhsT = vector.transpose %lhsBcast, [1, 0] : vector<[4]x[4]xi32> to vector<[4]x[4]xi32>
+  %rhsBcast = vector.broadcast %rhs : vector<[4]xi32> to vector<[4]x[4]xi32>
+  %mul = arith.muli %lhsT, %rhsBcast : vector<[4]x[4]xi32>
+  return %mul: vector<[4]x[4]xi32>
+}
+
+// CHECK-LABEL: func.func @ewise_outerproduct_transposed_rhs
+//  CHECK-SAME:   %[[LHS:.*]]: vector<16xf32>,
+//  CHECK-SAME:   %[[RHS:.*]]: vector<16xf32>) -> vector<16x16xf32> {
+//       CHECK:     %[[RES:.*]] = vector.outerproduct %[[RHS]], %[[LHS]] : vector<16xf32>, vector<16xf32>
+//       CHECK:     return %[[RES]] : vector<16x16xf32>
+func.func @ewise_outerproduct_transposed_rhs(%lhs: vector<16xf32>, %rhs: vector<16xf32>) -> vector<16x16xf32> {
+  %rhsBcast = vector.broadcast %rhs : vector<16xf32> to vector<16x16xf32>
+  %rhsT = vector.transpose %rhsBcast, [1, 0] : vector<16x16xf32> to vector<16x16xf32>
+  %lhsBcast = vector.broadcast %lhs : vector<16xf32> to vector<16x16xf32>
+  %mul = arith.mulf %lhsBcast, %rhsT : vector<16x16xf32>
+  return %mul: vector<16x16xf32>
+}
+
+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.vector.elementwise_to_vector
+    } : !transform.any_op
+    transform.yield
+  }
+}