[MLIR] Add patterns to bubble-up pack and push-down unpack through collapse/expand shape ops

[pzread]

Add DataLayoutPropagation patterns to bubble-up pack and push-down unpack through collapse/expand shape ops.

This is possible when the inner tile sizes of pack/unpack can divide the projected dims after being swapped with collapse/expand shape ops. For example:

%collapsed = tensor.collapse_shape %1 [[0], [1, 2]] : tensor<3072x64x4xf32> into tensor<3072x256xf32>
%pack = tensor.pack %collapsed outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [8, 1] into %2 : tensor<3072x256xf32> -> tensor<384x256x8x1xf32>

can be transformed into:

%pack = tensor.pack %1 outer_dims_perm = [0, 1, 2] inner_dims_pos = [0, 2] inner_tiles = [8, 1] into %2 : tensor<3072x64x4xf32> -> tensor<384x64x4x8x1xf32>
%collapsed = tensor.collapse_shape %pack [[0], [1, 2], [3], [4]] : tensor<384x64x4x8x1xf32> into tensor<384x256x8x1xf32>

The current support cases are:

• Bubble-up pack op through tensor.collapse_shape
• Push-down unpack op through tensor.expand_shape

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-linalg

Author: Jerry Wu (pzread)

Changes

Add DataLayoutPropagation patterns to bubble-up pack and push-down unpack through collapse/expand shape ops.

The current support cases are:

• Bubble-up pack op through tensor.collapse_shape
• Push-down unpack op through tensor.expand_shape

---

Patch is 22.13 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/85297.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/Linalg/Transforms/DataLayoutPropagation.cpp (+286-1)
• (modified) mlir/test/Dialect/Linalg/data-layout-propagation.mlir (+124)

diff --git a/mlir/lib/Dialect/Linalg/Transforms/DataLayoutPropagation.cpp b/mlir/lib/Dialect/Linalg/Transforms/DataLayoutPropagation.cpp
index 5ceb85e7d9903b..992a8916d90937 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/DataLayoutPropagation.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/DataLayoutPropagation.cpp
@@ -17,6 +17,7 @@
 #include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/IR/Dominance.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/ADT/TypeSwitch.h"
 #include "llvm/Support/Debug.h"
 #include <optional>
 
@@ -552,6 +553,289 @@ class BubbleUpPackThroughPadOp final : public OpRewritePattern<tensor::PackOp> {
   ControlPropagationFn controlFn;
 };
 
+/// Project dimsPos to the inner-most non unit dim pos with reassocIndices.
+/// For example: Given dimsPos: [0, 2], reassocIndices: [[0, 1], [2, 3]], and
+/// targetShape: [3, 4, 5, 1], it returns [1, 2]. Because for pos 0, the
+/// inner-most projected dim in [0, 1] is 1. And for pos 2, the inner-most
+/// non-unit projected dims in [2, 3] is 2.
+///
+/// If all projected dims are unit dims, it chooses the inner-most dim pos.
+static SmallVector<int64_t>
+projectToInnerMostNonUnitDimsPos(ArrayRef<int64_t> dimsPos,
+                                 ArrayRef<ReassociationIndices> reassocIndices,
+                                 ArrayRef<int64_t> targetShape) {
+  SmallVector<int64_t> projectedDimsPos;
+  for (auto pos : dimsPos) {
+    // In the case all dims are unit, this will return the inner-most one.
+    int64_t projectedPos = reassocIndices[pos].back();
+    for (auto it = reassocIndices[pos].rbegin();
+         it != reassocIndices[pos].rend(); ++it) {
+      int64_t dim = targetShape[*it];
+      if (dim > 1 || ShapedType::isDynamic(dim)) {
+        projectedPos = *it;
+        break;
+      }
+    }
+    projectedDimsPos.push_back(projectedPos);
+  }
+  return projectedDimsPos;
+}
+
+/// Check if all dims in dimsPos are divisible by the corresponding tile sizes.
+static bool isDimsDivisibleByTileSizes(ArrayRef<int64_t> dimsPos,
+                                       ArrayRef<int64_t> shape,
+                                       ArrayRef<int64_t> tileSizes) {
+  for (auto [pos, tileSize] : llvm::zip_equal(dimsPos, tileSizes)) {
+    int64_t dim = shape[pos];
+    if (ShapedType::isDynamic(dim) || (dim % tileSize) != 0)
+      return false;
+  }
+  return true;
+}
+
+/// Permutate the reassociation indices and reindex them in the sequence order.
+/// For example: given reassociationIndices: [[0, 1], [2]] and permutation: [1,
+/// 0], it applies the permutation to get [[2], [0, 1]] and reindexes the
+/// indices into [[0], [1, 2]].
+static int64_t applyPermutationAndReindexReassoc(
+    SmallVector<ReassociationIndices> &reassociationIndices,
+    ArrayRef<int64_t> permutation) {
+  applyPermutationToVector<ReassociationIndices>(reassociationIndices,
+                                                 permutation);
+  int64_t lastPos = 0;
+  for (ReassociationIndices &indices : reassociationIndices) {
+    for (auto &index : indices) {
+      index = lastPos;
+      lastPos += 1;
+    }
+  }
+  return lastPos;
+}
+
+/// Bubble up pack op through collapse shape op when the packed dims can be
+/// projected to the dims before collapsing. This is possible when the inner
+/// tile sizes can divide the projected dims.
+///
+/// For example:
+///
+/// %collapsed = tensor.collapse_shape %in [[0, 1], 2]
+///     : tensor<?x16x4xf32> into tensor<?x4xf32>
+/// %pack = tensor.pack %collapsed outer_dims_perm = [0, 1]
+///     inner_dims_pos = [0, 1] inner_tiles = [8, 1] into %empty
+///     : tensor<?x4xf32> -> tensor<?x4x8x1xf32>
+///
+/// Can be transformed into:
+///
+/// %pack = tensor.pack %in outer_dims_perm = [1, 2]
+///     inner_dims_pos = [1, 2] inner_tiles = [8, 1] into %empty
+///     : tensor<?x16x4xf32> -> tensor<?x2x4x8x1xf32>
+/// %collapsed = tensor.collapse_shape %pack [[0, 1], 2, 3, 4]
+///     : tensor<?x2x4x8x1xf32> into tensor<?x4x8x1>
+static LogicalResult
+bubbleUpPackOpThroughCollapseShape(tensor::CollapseShapeOp collapseOp,
+                                   tensor::PackOp packOp,
+                                   PatternRewriter &rewriter) {
+  SmallVector<int64_t> innerTileSizes = packOp.getStaticTiles();
+  ArrayRef<int64_t> innerDimsPos = packOp.getInnerDimsPos();
+  ArrayRef<int64_t> outerDimsPerm = packOp.getOuterDimsPerm();
+
+  ArrayRef<int64_t> srcShape = collapseOp.getSrcType().getShape();
+  SmallVector<ReassociationIndices> reassocIndices =
+      collapseOp.getReassociationIndices();
+  SmallVector<int64_t> projectedInnerDimsPos =
+      projectToInnerMostNonUnitDimsPos(innerDimsPos, reassocIndices, srcShape);
+
+  if (!isDimsDivisibleByTileSizes(projectedInnerDimsPos, srcShape,
+                                  innerTileSizes)) {
+    return failure();
+  }
+  // Expand the outer dims permutation with the associated source dims for the
+  // new permutation after bubbling. This is because moving a collapsed dim is
+  // equivalent to moving the associated source dims together.
+  SmallVector<int64_t> newOuterDimsPerm;
+  for (auto outerPos : outerDimsPerm) {
+    newOuterDimsPerm.insert(newOuterDimsPerm.end(),
+                            reassocIndices[outerPos].begin(),
+                            reassocIndices[outerPos].end());
+  }
+
+  auto emptyOp = tensor::PackOp::createDestinationTensor(
+      rewriter, packOp.getLoc(), collapseOp.getSrc(), packOp.getMixedTiles(),
+      projectedInnerDimsPos, newOuterDimsPerm);
+  auto newPackOp = rewriter.create<tensor::PackOp>(
+      packOp.getLoc(), collapseOp.getSrc(), emptyOp, projectedInnerDimsPos,
+      packOp.getMixedTiles(), packOp.getPaddingValue(), newOuterDimsPerm);
+
+  SmallVector<ReassociationIndices> newReassocIndices = reassocIndices;
+  // First apply the permutation on the reassociations of the outer dims.
+  // For example given the permutation [1, 0], the reassociations: [[0, 1], [2]]
+  // -> [[0], [1, 2]]
+  int64_t lastPos =
+      applyPermutationAndReindexReassoc(newReassocIndices, outerDimsPerm);
+  // Then add direct mapping for the inner tile dims.
+  for (size_t i = 0; i < innerDimsPos.size(); ++i) {
+    newReassocIndices.push_back({lastPos});
+    lastPos += 1;
+  }
+
+  auto newCollapseOp = rewriter.create<tensor::CollapseShapeOp>(
+      collapseOp.getLoc(), packOp.getType(), newPackOp, newReassocIndices);
+  rewriter.replaceOp(packOp, newCollapseOp);
+
+  return success();
+}
+
+class BubbleUpPackOpThroughReshapeOp final
+    : public OpRewritePattern<tensor::PackOp> {
+public:
+  BubbleUpPackOpThroughReshapeOp(MLIRContext *context, ControlPropagationFn fun)
+      : OpRewritePattern<tensor::PackOp>(context), controlFn(std::move(fun)) {}
+
+  LogicalResult matchAndRewrite(tensor::PackOp packOp,
+                                PatternRewriter &rewriter) const override {
+    // User controlled propagation function.
+    if (!controlFn(packOp))
+      return failure();
+
+    Operation *srcOp = packOp.getSource().getDefiningOp();
+    // Currently only support when the pack op is the only user.
+    if (!srcOp || !(srcOp->getNumResults() == 1) ||
+        !srcOp->getResult(0).hasOneUse()) {
+      return failure();
+    }
+    // Currently only support static inner tile sizes.
+    if (llvm::any_of(packOp.getStaticTiles(), [](int64_t size) {
+          return ShapedType::isDynamic(size);
+        })) {
+      return failure();
+    }
+
+    return TypeSwitch<Operation *, LogicalResult>(srcOp)
+        .Case([&](tensor::CollapseShapeOp op) {
+          return bubbleUpPackOpThroughCollapseShape(op, packOp, rewriter);
+        })
+        .Default([](Operation *) { return failure(); });
+  }
+
+private:
+  ControlPropagationFn controlFn;
+};
+
+/// Push down unpack op through expand shape op when the packed dims can be
+/// projected to the dims after expanding. This is possible when the inner tile
+/// sizes can divide the projected dims.
+///
+/// For example:
+///
+/// %unpack = tensor.unpack %in outer_dims_perm = [0, 1]
+///     inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %empty
+///     : tensor<?x32x8x8xf32> -> tensor<?x256xf32>
+/// %expanded = tensor.expand_shape %unpack [[0, 1], [2]]
+///     : tensor<?x256xf32> into tensor<?x256x256xf32>
+///
+/// Can be transformed into:
+///
+/// %expanded = tensor.expand_shape %ain [[0, 1], [2], [3], [4]]
+///     : tensor<?x32x8x8xf32> into tensor<?x32x32x8x8xf32>
+/// %unpack = tensor.unpack %expanded outer_dims_perm = [0, 1, 2]
+///     inner_dims_pos = [1, 2] inner_tiles = [8, 8] into %empty
+///     : tensor<?x32x32x8x8xf32> -> tensor<?x256x256xf32>
+static LogicalResult
+pushDownUnPackOpThroughExpandShape(tensor::UnPackOp unPackOp,
+                                   tensor::ExpandShapeOp expandOp,
+                                   PatternRewriter &rewriter) {
+  SmallVector<int64_t> innerTileSizes = unPackOp.getStaticTiles();
+  ArrayRef<int64_t> innerDimsPos = unPackOp.getInnerDimsPos();
+  ArrayRef<int64_t> outerDimsPerm = unPackOp.getOuterDimsPerm();
+
+  ArrayRef<int64_t> dstShape = expandOp.getType().getShape();
+  SmallVector<ReassociationIndices> reassocIndices =
+      expandOp.getReassociationIndices();
+  SmallVector<int64_t> projectedInnerDimsPos =
+      projectToInnerMostNonUnitDimsPos(innerDimsPos, reassocIndices, dstShape);
+
+  if (!isDimsDivisibleByTileSizes(projectedInnerDimsPos, dstShape,
+                                  innerTileSizes)) {
+    return failure();
+  }
+  // Expand the outer dims permutation with the associated expanded dims for the
+  // new permutation after pushing. This is because moving a source dim is
+  // equivalent to moving the associated expanded dims together.
+  SmallVector<int64_t> newOuterDimsPerm;
+  for (auto outerPos : outerDimsPerm) {
+    newOuterDimsPerm.insert(newOuterDimsPerm.end(),
+                            reassocIndices[outerPos].begin(),
+                            reassocIndices[outerPos].end());
+  }
+
+  SmallVector<ReassociationIndices> newReassocIndices = reassocIndices;
+  // First apply the permutation on the reassociations of the outer dims.
+  // For example given the permutation [1, 0], the reassociations: [[0, 1], [2]]
+  // -> [[0], [1, 2]]
+  int64_t lastPos =
+      applyPermutationAndReindexReassoc(newReassocIndices, outerDimsPerm);
+  // Then add direct mapping for the inner tile dims.
+  for (size_t i = 0; i < innerDimsPos.size(); ++i) {
+    newReassocIndices.push_back({lastPos});
+    lastPos += 1;
+  }
+
+  RankedTensorType newExpandType =
+      tensor::PackOp::inferPackedType(expandOp.getType(), innerTileSizes,
+                                      projectedInnerDimsPos, newOuterDimsPerm);
+  auto newExpandOp = rewriter.create<tensor::ExpandShapeOp>(
+      expandOp.getLoc(), newExpandType, unPackOp.getSource(),
+      newReassocIndices);
+
+  auto emptyOp = tensor::UnPackOp::createDestinationTensor(
+      rewriter, unPackOp.getLoc(), newExpandOp, unPackOp.getMixedTiles(),
+      projectedInnerDimsPos, newOuterDimsPerm);
+  auto newUnPackOp = rewriter.create<tensor::UnPackOp>(
+      unPackOp.getLoc(), newExpandOp.getResult(), emptyOp,
+      projectedInnerDimsPos, unPackOp.getMixedTiles(), newOuterDimsPerm);
+  rewriter.replaceOp(expandOp, newUnPackOp);
+
+  return success();
+}
+
+class PushDownUnPackOpThroughReshapeOp final
+    : public OpRewritePattern<tensor::UnPackOp> {
+public:
+  PushDownUnPackOpThroughReshapeOp(MLIRContext *context,
+                                   ControlPropagationFn fun)
+      : OpRewritePattern<tensor::UnPackOp>(context), controlFn(std::move(fun)) {
+  }
+
+  LogicalResult matchAndRewrite(tensor::UnPackOp unPackOp,
+                                PatternRewriter &rewriter) const override {
+    // User controlled propagation function.
+    if (!controlFn(unPackOp))
+      return failure();
+
+    Value result = unPackOp.getResult();
+    // Currently only support unpack op with the single user.
+    if (!result.hasOneUse()) {
+      return failure();
+    }
+    // Currently only support static inner tile sizes.
+    if (llvm::any_of(unPackOp.getStaticTiles(), [](int64_t size) {
+          return ShapedType::isDynamic(size);
+        })) {
+      return failure();
+    }
+
+    Operation *userOp = *result.user_begin();
+    return TypeSwitch<Operation *, LogicalResult>(userOp)
+        .Case([&](tensor::ExpandShapeOp op) {
+          return pushDownUnPackOpThroughExpandShape(unPackOp, op, rewriter);
+        })
+        .Default([](Operation *) { return failure(); });
+  }
+
+private:
+  ControlPropagationFn controlFn;
+};
+
 // TODO: Relax this restriction. We should unpack a generic op also
 // in the presence of multiple unpack ops as producers.
 /// Return the unpacked operand, if present, for the current generic op.
@@ -774,6 +1058,7 @@ void mlir::linalg::populateDataLayoutPropagationPatterns(
     const ControlPropagationFn &controlPackUnPackPropagation) {
   patterns
       .insert<BubbleUpPackOpThroughGenericOpPattern, BubbleUpPackThroughPadOp,
-              PushDownUnPackOpThroughGenericOp, PushDownUnPackThroughPadOp>(
+              BubbleUpPackOpThroughReshapeOp, PushDownUnPackOpThroughGenericOp,
+              PushDownUnPackThroughPadOp, PushDownUnPackOpThroughReshapeOp>(
           patterns.getContext(), controlPackUnPackPropagation);
 }
diff --git a/mlir/test/Dialect/Linalg/data-layout-propagation.mlir b/mlir/test/Dialect/Linalg/data-layout-propagation.mlir
index e036695a2ac9fd..10c9f5bafb5c03 100644
--- a/mlir/test/Dialect/Linalg/data-layout-propagation.mlir
+++ b/mlir/test/Dialect/Linalg/data-layout-propagation.mlir
@@ -905,3 +905,127 @@ func.func @unpack_different_destination_shape(%arg0: tensor<1x1x1080x1920x16xi32
 // CHECK-SAME:      inner_dims_pos = [0] inner_tiles = [16]
 // CHECK-SAME:      into %[[UNPACK_NEW_DEST]]
 // CHECK:         return %[[UNPACK]] : tensor<16x540x960xi32>
+
+// -----
+
+func.func @bubble_up_pack_through_collapse(%1: tensor<?x16x4xf32>, %dim : index) -> tensor<?x4x8x1xf32> {
+  %collapsed = tensor.collapse_shape %1 [[0, 1], [2]] : tensor<?x16x4xf32> into tensor<?x4xf32>
+  %2 = tensor.empty(%dim) : tensor<?x4x8x1xf32>
+  %pack = tensor.pack %collapsed outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [8, 1] into %2 : tensor<?x4xf32> -> tensor<?x4x8x1xf32>
+  func.return %pack : tensor<?x4x8x1xf32>
+}
+// CHECK-LABEL: func.func @bubble_up_pack_through_collapse
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK-SAME:      %[[ARG1:[a-zA-Z0-9]+]]
+// CHECK:         %[[C0:.+]] = arith.constant 0 : index
+// CHECK:         %[[DIM:.+]] = tensor.dim %[[ARG0]], %[[C0]] : tensor<?x16x4xf32>
+// CHECK:         %[[EMPTY:.+]] = tensor.empty(%[[DIM]]) : tensor<?x2x4x8x1xf32>
+// CHECK:         %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [0, 1, 2] inner_dims_pos = [1, 2] inner_tiles = [8, 1] into %[[EMPTY]] : tensor<?x16x4xf32> -> tensor<?x2x4x8x1xf32>
+// CHECK:         %[[COLLAPSED:.+]] = tensor.collapse_shape %[[PACK]] {{\[}}[0, 1], [2], [3], [4]] : tensor<?x2x4x8x1xf32> into tensor<?x4x8x1xf32>
+// CHECK:         return %[[COLLAPSED]] : tensor<?x4x8x1xf32>
+
+// -----
+
+func.func @bubble_up_permuted_pack_through_collapse(%1: tensor<4x192x16x256xf32>) -> tensor<4x32x3072x8x1xf32> {
+  %collapsed = tensor.collapse_shape %1 [[0], [1, 2], [3]] : tensor<4x192x16x256xf32> into tensor<4x3072x256xf32>
+  %2 = tensor.empty() : tensor<4x32x3072x8x1xf32>
+  %pack = tensor.pack %collapsed outer_dims_perm = [0, 2, 1] inner_dims_pos = [2, 1] inner_tiles = [8, 1] into %2 : tensor<4x3072x256xf32> -> tensor<4x32x3072x8x1xf32>
+  func.return %pack : tensor<4x32x3072x8x1xf32>
+}
+// CHECK-LABEL: func.func @bubble_up_permuted_pack_through_collapse
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK:         %[[EMPTY:.+]] = tensor.empty() : tensor<4x32x192x16x8x1xf32>
+// CHECK:         %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [0, 3, 1, 2] inner_dims_pos = [3, 2] inner_tiles = [8, 1] into %[[EMPTY]] : tensor<4x192x16x256xf32> -> tensor<4x32x192x16x8x1xf32>
+// CHECK:         %[[COLLAPSED:.+]] = tensor.collapse_shape %pack {{\[}}[0], [1], [2, 3], [4], [5]] : tensor<4x32x192x16x8x1xf32> into tensor<4x32x3072x8x1xf32>
+// CHECK:         return %[[COLLAPSED]] : tensor<4x32x3072x8x1xf32>
+
+// -----
+
+func.func @bubble_up_pack_through_unit_collapse(%1: tensor<1x64x1x4xf32>) -> tensor<8x4x8x1xf32> {
+  %collapsed = tensor.collapse_shape %1 [[0, 1, 2], [3]] : tensor<1x64x1x4xf32> into tensor<64x4xf32>
+  %2 = tensor.empty() : tensor<8x4x8x1xf32>
+  %pack = tensor.pack %collapsed outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [8, 1] into %2 : tensor<64x4xf32> -> tensor<8x4x8x1xf32>
+  func.return %pack : tensor<8x4x8x1xf32>
+}
+// CHECK-LABEL: func.func @bubble_up_pack_through_unit_collapse
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK:         %[[EMPTY:.+]] = tensor.empty() : tensor<1x8x1x4x8x1xf32>
+// CHECK:         %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [0, 1, 2, 3] inner_dims_pos = [1, 3] inner_tiles = [8, 1] into %[[EMPTY]] : tensor<1x64x1x4xf32> -> tensor<1x8x1x4x8x1xf32>
+// CHECK:         %[[COLLAPSED:.+]] = tensor.collapse_shape %[[PACK]] {{\[}}[0, 1, 2], [3], [4], [5]] : tensor<1x8x1x4x8x1xf32> into tensor<8x4x8x1xf32>
+// CHECK:         return %[[COLLAPSED]] : tensor<8x4x8x1xf32>
+
+// -----
+
+func.func @no_bubble_up_pack_through_non_divisible_collapse(%1: tensor<3072x64x4xf32>) -> tensor<384x32x8x8xf32> {
+  %collapsed = tensor.collapse_shape %1 [[0], [1, 2]] : tensor<3072x64x4xf32> into tensor<3072x256xf32>
+  %2 = tensor.empty() : tensor<384x32x8x8xf32>
+  %pack = tensor.pack %collapsed outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %2 : tensor<3072x256xf32> -> tensor<384x32x8x8xf32>
+  func.return %pack : tensor<384x32x8x8xf32>
+}
+// CHECK-LABEL: func.func @no_bubble_up_pack_through_non_divisible_collapse
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK:         %[[COLLAPSED:.+]] = tensor.collapse_shape %[[ARG0]] {{\[}}[0], [1, 2]] : tensor<3072x64x4xf32> into tensor<3072x256xf32>
+// CHECK:         %[[PACK:.+]] = tensor.pack %[[COLLAPSED]]
+// CHECK:         return %[[PACK]] : tensor<384x32x8x8xf32>
+
+// -----
+
+func.func @push_down_unpack_through_expand(%5: tensor<?x32x8x8xf32>, %dim: index) -> tensor<?x256x256xf32> {
+  %6 = tensor.empty(%dim) : tensor<?x256xf32>
+  %unpack = tensor.unpack %5 outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %6 : tensor<?x32x8x8xf32> -> tensor<?x256xf32>
+  %expanded = tensor.expand_shape %unpack [[0, 1], [2]] : tensor<?x256xf32> into tensor<?x256x256xf32>
+  func.return %expanded : tensor<?x256x256xf32>
+}
+// CHECK-LABEL: func.func @push_down_unpack_through_expand
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK-SAME:      %[[ARG1:[a-zA-Z0-9]+]]
+// CHECK:         %[[C0:.+]] = arith.constant 0 : index
+// CHECK:         %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0, 1], [2], [3], [4]] : tensor<?x32x8x8xf32> into tensor<?x32x32x8x8xf32>
+// CHECK:         %[[DIM:.+]] = tensor.dim %[[EXPANDED]], %[[C0]] : tensor<?x32x32x8x8xf32>
+// CHECK:         %[[EMPTY:.+]] = tensor.empty(%[[DIM]]) : tensor<?x256x256xf32>
+// CHECK:         %[[UNPACK:.+]] = tensor.unpack %[[EXPANDED:.+]] outer_dims_perm = [0, 1, 2] inner_dims_pos = [1, 2] inner_tiles = [8, 8] into %[[EMPTY]] : tensor<?x32x32x8x8xf32> -> tensor<?x256x256xf32>
+// CHECK:         return %[[UNPACK]] : tensor<?x256x256xf32>
+
+// -----
+
+func.func @push_down_permuted_unpack_through_expand(%5: tensor<4x32x384x8x8xf32>) -> tensor<4x12x256x256xf32> {
+  %6 = tensor.empty() : tensor<4x3072x256xf32>
+  %unpack = tensor.unpack %5 outer_dims_perm = [0, 2, 1] inner_dims_pos = [2, 1] inner_tiles = [8, 8] into %6 : tensor<4x32x384x8x8xf32> -> tensor<4x3072x256xf32>
+  %expanded = tensor.expand_shape %unpack [[0], [1, 2], [3]] : tensor<4x3072x256xf32> into tensor<4x12x256x256xf32>
+  func.return %expanded : tensor<4x12x256x256xf32>
+}
+// CHECK-LABEL: @push_down_permuted_unpack_through_expand
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK:         %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0], [1], [2, 3], [4], [5]] : tensor<4x32x384x8x8xf32> into tensor<4x32x12x32x8x8xf32>
+// CHECK:         %[[EMPTY:.+]] = tensor.empty() : tensor<4x12x256x256xf32>
+// CHECK:         %[[UNPACK:.+]] = tensor.unpack %[[EXPANDED]] ...
[truncated]

[pzread]

Kindly ping : )

[qedawkins]

Sorry for the late review, LGTM, thanks!