[Mlir-commits] [mlir] [mlir][vector] Add ElementwiseToOuterproduct (PR #93664)

Andrzej WarzyƄski llvmlistbot at llvm.org
Mon Jun 3 01:52:27 PDT 2024


================
@@ -1795,6 +1795,81 @@ 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;
+  // Helper function returning the source of the input broadcast if it matches
+  // requirements for an outerproduct pattern.
+  Value getValidBroadcastSource(vector::BroadcastOp broadcastOp) const {
+    // Fail if it is not a 1-to-2 dimension to broadcast to avoid generating
+    // shape_casts/broadcasts which does not belong in this pattern.
+    if (!broadcastOp.computeBroadcastedUnitDims().empty())
+      return Value();
+    // Avoid broadcast like f32 or vector<f32> -> ResType
+    auto srcVT = dyn_cast<VectorType>(broadcastOp.getSourceType());
+    if (!srcVT || srcVT.getRank() != 1)
+      return Value();
+    return broadcastOp.getSource();
+  }
+
+  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,
----------------
banach-space wrote:

1. I think that long lambdas (more than 2-3 lines) obscure readability. Why not create a helper hook outside this method? Others might disagree (looking at you @MacDue  :) ).
2. In any case, please document what this lambda does and use less generic name. "canonicalize" is a very powerful term and based on past experience I feel that it can mean different things depending on context. Also, we don't really want people hitting this when searching for "canonicalize" in MLIR.

https://github.com/llvm/llvm-project/pull/93664


More information about the Mlir-commits mailing list