[Mlir-commits] [llvm] [mlir] [mlir][tensor] Fold linalg transpose with tensor pack (PR #74206)

[lorenzo chelini]

================
@@ -81,10 +82,91 @@ struct FoldUnpackWithExtractSliceOp : public OpRewritePattern<ExtractSliceOp> {
     return success();
   }
 };
+
+/// Fold 'pack' -> 'transpose' into 'pack' since 'pack' already has transpose
+/// semantics.
+struct FoldProducerPackWithConsumerLinalgTransposeOp
+    : public OpRewritePattern<linalg::TransposeOp> {
+  using OpRewritePattern<linalg::TransposeOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(linalg::TransposeOp transposeOp,
+                                PatternRewriter &rewriter) const override {
+    auto transposeInputTensor = transposeOp.getOperand(0);
+    auto packOp = transposeInputTensor.getDefiningOp<PackOp>();
+
+    if (!packOp)
+      return failure();
+
+    auto packInnerDimsPos = packOp.getInnerDimsPos();
+    auto packInnerTiles = packOp.getStaticInnerTiles();
+    auto packOuterDimsPerm = packOp.getOuterDimsPerm();
+    auto transposePerm = transposeOp.getPermutation();
+    SmallVector<int64_t> newPackOuterDimsPermVec;
+    SmallVector<int64_t> newPackInnerDimsPosVec;
+    SmallVector<int64_t> newPackInnerTilesVec;
+
+    // Variable for storing translated position after considering original
+    // outer_dims_perm and permutation attributes of tensor.pack and
+    // linalg.transpose.
+    int64_t translatedPosition;
+
+    // Process transpose operation for non-tiled outer dimensions of the tensor.
+    for (unsigned int i = 0; i < transposePerm.size() - packInnerTiles.size();
----------------
chelini wrote:

We could make `transposePerm.size() - packInnerTiles.size()` as a variable since it is computed multiple times.

https://github.com/llvm/llvm-project/pull/74206