[Mlir-commits] [mlir] [mlir][vector] transpose(broadcast) -> broadcast canonicalization (PR #135096)

[James Newling]

================
@@ -6155,12 +6155,115 @@ class FoldTransposeCreateMask final : public OpRewritePattern<TransposeOp> {
   }
 };
 
+/// Folds transpose(broadcast(x)) to broadcast(x) if the transpose is
+/// 'order preserving', where 'order preserving' means the flattened
+/// inputs and outputs of the transpose have identical (numerical) values.
+///
+/// Example:
+/// ```
+///  %0 = vector.broadcast %input : vector<1x1xi32> to vector<1x8xi32>
+///  %1 = vector.transpose %0, [1, 0] : vector<1x8xi32>
+///                                                 to vector<8x1xi32>
+/// ```
+/// can be rewritten as the equivalent
+/// ```
+///  %0 = vector.broadcast %input : vector<1x1xi32> to vector<8x1xi32>.
+/// ```
+/// The algorithm works by partitioning dimensions into groups that can be
+/// locally permuted while preserving order, and checks that the transpose
+/// only permutes within these groups.
+class FoldTransposeBroadcast : public OpRewritePattern<vector::TransposeOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  FoldTransposeBroadcast(MLIRContext *context, PatternBenefit benefit = 1)
+      : OpRewritePattern<vector::TransposeOp>(context, benefit) {}
+
+  static bool canFoldIntoPrecedingBroadcast(vector::TransposeOp transpose) {
+
+    vector::BroadcastOp broadcast =
+        transpose.getVector().getDefiningOp<vector::BroadcastOp>();
+    if (!broadcast)
+      return false;
+
+    auto inputType = dyn_cast<VectorType>(broadcast.getSourceType());
+    bool inputIsScalar = !inputType;
+    ArrayRef<int64_t> inputShape = inputType.getShape();
+    int64_t inputRank = inputType.getRank();
+    int64_t outputRank = transpose.getType().getRank();
+    int64_t deltaRank = outputRank - inputRank;
+
+    // transpose(broadcast(scalar)) -> broadcast(scalar) is always valid
+    if (inputIsScalar)
+      return true;
+
+    // Return true if all permutation destinations for indices in [low, high)
+    // are in [low, high), so the permutation is local to the group.
+    auto isGroupBound = [&](int low, int high) {
+      ArrayRef<int64_t> permutation = transpose.getPermutation();
+      for (int j = low; j < high; ++j) {
+        if (permutation[j] < low || permutation[j] >= high) {
+          return false;
+        }
+      }
+      return true;
+    };
+
+    // Groups are either contiguous sequences  of 1s and non-1s (1-element
+    // groups). Consider broadcasting 4x1x1x7 to 2x3x4x5x6x7. This is equivalent
+    // to broadcasting from 1x1x4x1x1x7.
+    //                      ^^^ ^ ^^^ ^
+    //               groups: 0  1  2  3
+    // Order preserving permutations for this example are ones that only permute
+    // within the groups [0,1] and [3,4], like (1 0 2 4 3 5 6).
+    int low = 0;
+    for (int inputIndex = 0; inputIndex < inputRank; ++inputIndex) {
+      bool notOne = inputShape[inputIndex] != 1;
+      bool prevNotOne = (inputIndex != 0 && inputShape[inputIndex - 1] != 1);
+      bool groupEndFound = notOne || prevNotOne;
+      if (groupEndFound) {
+        int high = inputIndex + deltaRank;
+        if (!isGroupBound(low, high)) {
+          return false;
+        }
+        low = high;
+      }
+    }
+    if (!isGroupBound(low, outputRank)) {
+      return false;
+    }
+
+    // The preceding logic ensures that by this point, the ouutput of the
+    // transpose is definitely broadcastable from the input shape. So we don't
+    // need to call 'vector::isBroadcastableTo', but asserting here just as a
+    // sanity check:
+    bool isBroadcastable =
+        vector::isBroadcastableTo(inputType, transpose.getResultVectorType()) ==
+        vector::BroadcastableToResult::Success;
+    assert(isBroadcastable &&
+           "(I think) it must be broadcastable at this point.");
+
+    return true;
+  }
+
+  LogicalResult matchAndRewrite(vector::TransposeOp transpose,
+                                PatternRewriter &rewriter) const override {
+    if (!canFoldIntoPrecedingBroadcast(transpose))
+      return failure();
+
+    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(
+        transpose, transpose.getResultVectorType(), transpose.getVector());
+
+    return success();
+  }
+};
+
 } // namespace
 
 void vector::TransposeOp::getCanonicalizationPatterns(
     RewritePatternSet &results, MLIRContext *context) {
   results.add<FoldTransposeCreateMask, FoldTransposedScalarBroadcast,
-              TransposeFolder, FoldTransposeSplat>(context);
+              TransposeFolder, FoldTransposeSplat, FoldTransposeBroadcast>(
+      context);
----------------
newling wrote:

I think so. The new broadcast is the same rank and 'volume' as the old one, so i can't think of sense in which it'll be more complex. So the removal of the transpose clinches it in my mind!

As an aside: It would be nice if MLIR/dialects defined an 'energy function' defining what classifies as a canonicalization (i.e. something to guarantee every rewrite takes us closer to a fixed point = energy minimum). 

https://github.com/llvm/llvm-project/pull/135096