[MLIR] [Vector] Added canonicalizer for folding from_elements + transpose

[keshavvinayak01]

Description

Adds a new canonicalizer that folds vector.from_elements(vector.transpose)) => vector.from_elements. This canonicalization reorders the input elements for vector.from_elements, adjusts the output shape to match the effect of the transpose op and eliminating its need.

Testing

Added a 2D vector lit test that verifies the working of the rewrite.

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-vector

Author: Keshav Vinayak Jha (keshavvinayak01)

Changes

Description

Adds a new canonicalizer that folds vector.from_elements(vector.broadcast)) => vector.from_elements. This canonicalization reorders the input elements for vector.from_elements, adjusts the output shape to match the effect of the broadcast op and eliminating its need.

Testing

Added a 2D vector lit test that verifies the working of the rewrite.

---

Full diff: https://github.com/llvm/llvm-project/pull/161841.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/Vector/IR/VectorOps.cpp (+57-1)
• (modified) mlir/test/Dialect/Vector/canonicalize.mlir (+12)

diff --git a/mlir/lib/Dialect/Vector/IR/VectorOps.cpp b/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
index b0132e889302f..7f6313c11ea18 100644
--- a/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
+++ b/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
@@ -6723,6 +6723,61 @@ class FoldTransposeShapeCast final : public OpRewritePattern<TransposeOp> {
   }
 };
 
+/// Folds transpose(from_elements(...)) into a new from_elements with permuted
+/// operands matching the transposed shape.
+class FoldTransposeFromElements final : public OpRewritePattern<TransposeOp> {
+public:
+  using Base::Base;
+  LogicalResult matchAndRewrite(vector::TransposeOp transposeOp,
+                                PatternRewriter &rewriter) const override {
+    auto fromElementsOp =
+        transposeOp.getVector().getDefiningOp<vector::FromElementsOp>();
+    if (!fromElementsOp)
+      return failure();
+
+    VectorType srcTy = fromElementsOp.getDest().getType();
+    VectorType dstTy = transposeOp.getType();
+
+    ArrayRef<int64_t> permutation = transposeOp.getPermutation();
+    int64_t rank = srcTy.getRank();
+
+    // Build inverse permutation to map destination indices back to source.
+    SmallVector<int64_t, 4> inversePerm(rank, 0);
+    for (int64_t i = 0; i < rank; ++i)
+      inversePerm[permutation[i]] = i;
+
+    ArrayRef<int64_t> srcShape = srcTy.getShape();
+    ArrayRef<int64_t> dstShape = dstTy.getShape();
+    SmallVector<int64_t, 4> srcIdx(rank, 0);
+    SmallVector<int64_t, 4> dstIdx(rank, 0);
+    SmallVector<int64_t, 4> srcStrides = computeStrides(srcShape);
+    SmallVector<int64_t, 4> dstStrides = computeStrides(dstShape);
+
+    auto elements = fromElementsOp.getElements();
+    SmallVector<Value> newElements;
+    int64_t dstNumElements = dstTy.getNumElements();
+    newElements.reserve(dstNumElements);
+
+    // For each element in destination row-major order, pick the corresponding
+    // source element.
+    for (int64_t lin = 0; lin < dstNumElements; ++lin) {
+      // Pick the destination element index.
+      dstIdx = delinearize(lin, dstStrides);
+      // Map the destination element index to the source element index.
+      for (int64_t j = 0; j < rank; ++j)
+        srcIdx[j] = dstIdx[inversePerm[j]];
+      // Linearize the source element index.
+      int64_t srcLin = linearize(srcIdx, srcStrides);
+      // Add the source element to the new elements.
+      newElements.push_back(elements[srcLin]);
+    }
+
+    rewriter.replaceOpWithNewOp<FromElementsOp>(transposeOp, dstTy,
+                                                newElements);
+    return success();
+  }
+};
+
 /// 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.
@@ -6823,7 +6878,8 @@ class FoldTransposeBroadcast : public OpRewritePattern<vector::TransposeOp> {
 void vector::TransposeOp::getCanonicalizationPatterns(
     RewritePatternSet &results, MLIRContext *context) {
   results.add<FoldTransposeCreateMask, FoldTransposeShapeCast, TransposeFolder,
-              FoldTransposeSplat, FoldTransposeBroadcast>(context);
+              FoldTransposeSplat, FoldTransposeFromElements,
+              FoldTransposeBroadcast>(context);
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/Vector/canonicalize.mlir b/mlir/test/Dialect/Vector/canonicalize.mlir
index 5448976f84760..5f34d144cd472 100644
--- a/mlir/test/Dialect/Vector/canonicalize.mlir
+++ b/mlir/test/Dialect/Vector/canonicalize.mlir
@@ -308,6 +308,18 @@ func.func @constant_mask_transpose_to_transposed_constant_mask() -> (vector<2x3x
 
 // -----
 
+// CHECK-LABEL: transpose_from_elements_2d
+func.func @transpose_from_elements_2d(%a0: i32, %a1: i32, %a2: i32,
+                                      %a3: i32, %a4: i32, %a5: i32) -> vector<3x2xi32> {
+  %v = vector.from_elements %a0, %a1, %a2, %a3, %a4, %a5 : vector<2x3xi32>
+  %t = vector.transpose %v, [1, 0] : vector<2x3xi32> to vector<3x2xi32>
+  return %t : vector<3x2xi32>
+  // CHECK: %[[R:.*]] = vector.from_elements %arg0, %arg3, %arg1, %arg4, %arg2, %arg5 : vector<3x2xi32>
+  // CHECK-NOT: vector.transpose
+}
+
+// -----
+
 func.func @extract_strided_slice_of_constant_mask() -> (vector<2x2xi1>) {
   %0 = vector.constant_mask [2, 2] : vector<4x3xi1>
   %1 = vector.extract_strided_slice %0

[banach-space]

Thanks!

[banach-space]

What does lin represent?

[keshavvinayak01]

So "lin" is short for "linear index" - it's a 1D index that represents the position of an element when the multi-dimensional vector is laid out in row-major order in memory. I felt it was a good iter name.

[banach-space]

I find lin a bit too enigmatic. Why not linearIdx?

[banach-space]

I find lin a bit too enigmatic. Why not linearIdx?

[dcaballe]

LGTM, thanks! (modulo pending comments)