Skip to content

[MLIR] Add shape propagation through tensor.pad #136681

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 11 commits into
base: main
Choose a base branch
from
Open

[MLIR] Add shape propagation through tensor.pad #136681

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
0d8c636
Add FoldReshapeWithProducerPadOpByExpansion
ita9naiwa Apr 22, 2025
cf1e560
Merge branch 'main' into ita9naiwa/fold
ita9naiwa Jul 12, 2025
57ec657
add collapse_shape
ita9naiwa Jul 14, 2025
737d4a4
fix upon review
ita9naiwa Jul 18, 2025
d8ca036
fix upon review
ita9naiwa Jul 18, 2025
17a2447
fix upon review
ita9naiwa Jul 22, 2025
9ee8e08
fix upon review
ita9naiwa Jul 22, 2025
3b91645
fix upon review
ita9naiwa Jul 26, 2025
9c38ad5
fix upon review
ita9naiwa Jul 26, 2025
0faf084
fix upon review
ita9naiwa Jul 29, 2025
9cbd032
fix comments
ita9naiwa Aug 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
199 changes: 199 additions & 0 deletions mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1100,6 +1100,198 @@ class FoldPadWithProducerReshapeOpByExpansion
ControlFusionFn controlFoldingReshapes;
};

/// Pattern to fold a tensor.expand_shape op with its producer tensor.pad op
/// by bubbling the expand_shape before the pad.
Copy link
Contributor

@banach-space banach-space Aug 1, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

To me, "folding" would mean that tensor.expand_shape disappears (i.e. is folded away), but that's not what is happening here, is it? This is merely "bubbling up".

Please update the description accordingly and add some example IR before and after. As an example: https://github.com/banach-space/llvm-project/blob/7d35eb58959c0ab398a9739f38bfb9754c5ba5e5/mlir/lib/Dialect/Tensor/Transforms/ReshapePatterns.cpp#L305-L317

struct FoldReshapeWithProducerPadOpByExpansion
: public OpRewritePattern<tensor::ExpandShapeOp> {

FoldReshapeWithProducerPadOpByExpansion(MLIRContext *context,
ControlFusionFn foldReshapes,
PatternBenefit benefit = 1)
: OpRewritePattern<tensor::ExpandShapeOp>(context, benefit),
controlFoldingReshapes(std::move(foldReshapes)) {}

LogicalResult matchAndRewrite(tensor::ExpandShapeOp expandOp,
PatternRewriter &rewriter) const override {
tensor::PadOp padOp = expandOp.getSrc().getDefiningOp<tensor::PadOp>();
if (!padOp)
return failure();

if (!padOp->hasOneUse())
return failure();

if (!controlFoldingReshapes(&expandOp.getSrcMutable())) {
return rewriter.notifyMatchFailure(expandOp,
"fusion blocked by control function");
}

Value constantPaddingValue = padOp.getConstantPaddingValue();
if (!constantPaddingValue) {
return rewriter.notifyMatchFailure(
expandOp, "cannot fold with non-constant padding value");
}

SmallVector<ReassociationIndices> reassociations =
expandOp.getReassociationIndices();
SmallVector<OpFoldResult> low = padOp.getMixedLowPad();
SmallVector<OpFoldResult> high = padOp.getMixedHighPad();

for (auto [reInd, l, h] : llvm::zip_equal(reassociations, low, high)) {
if (reInd.size() > 1 &&
(!isConstantIntValue(l, 0) || !isConstantIntValue(h, 0)))
return rewriter.notifyMatchFailure(
expandOp, "fusion blocked by non-zero padding");
}

SmallVector<OpFoldResult> newLow, newHigh;
for (auto [idx, reInd] : llvm::enumerate(reassociations)) {
newLow.append(reInd.size(), low[idx]);
newHigh.append(reInd.size(), high[idx]);
}
Copy link
Contributor

@Max191 Max191 Jul 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: combine these 2 loops.


Location loc = expandOp.getLoc();
ArrayRef<int64_t> finalShape = expandOp.getResultType().getShape();
SmallVector<OpFoldResult> expandedShape = expandOp.getMixedOutputShape();

for (auto [inDimIdx, reInd] : llvm::enumerate(reassociations)) {
OpFoldResult l = low[inDimIdx];
OpFoldResult h = high[inDimIdx];

if (!isConstantIntValue(l, 0) || !isConstantIntValue(h, 0)) {
assert(reInd.size() == 1 && "expected single dimension");
expandedShape[reInd[0]] =
tensor::getMixedSize(rewriter, loc, padOp.getSource(), inDimIdx);
;
}
}

for (auto [outDimIdx, dimSize] : llvm::enumerate(finalShape)) {
if (dimSize == ShapedType::kDynamic &&
!isa<Value>(expandedShape[outDimIdx]) &&
!isa<Attribute>(expandedShape[outDimIdx])) {
expandedShape[outDimIdx] =
tensor::getMixedSize(rewriter, loc, expandOp.getSrc(), outDimIdx);
}
}
Copy link
Contributor

@Max191 Max191 Jul 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You can delete this loop because expandOp.getMixedOutputShape() will already populate the expandedShape with the right dynamic values.


SmallVector<int64_t> staticExpandedShape;
std::tie(staticExpandedShape, std::ignore) =
decomposeMixedValues(expandedShape);

auto newExpandOp = rewriter.create<tensor::ExpandShapeOp>(
loc,
RankedTensorType::get(staticExpandedShape,
padOp.getSource().getType().getElementType()),
padOp.getSource(), reassociations, expandedShape);

rewriter.replaceOpWithNewOp<tensor::PadOp>(
expandOp, expandOp.getType(), newExpandOp.getResult(), newLow, newHigh,
padOp.getConstantPaddingValue(), padOp.getNofold());
return success();
}

private:
ControlFusionFn controlFoldingReshapes;
};

/// Pattern to fold a tensor.collapse_shape op with its producer tensor.pad op
/// by bubbling the collapse_shape before the pad.
struct FoldReshapeWithProducerPadOpByCollapsing
: public OpRewritePattern<tensor::CollapseShapeOp> {

FoldReshapeWithProducerPadOpByCollapsing(MLIRContext *context,
ControlFusionFn foldReshapes,
PatternBenefit benefit = 1)
: OpRewritePattern<tensor::CollapseShapeOp>(context, benefit),
controlFoldingReshapes(std::move(foldReshapes)) {}

LogicalResult matchAndRewrite(tensor::CollapseShapeOp collapseOp,
PatternRewriter &rewriter) const override {
tensor::PadOp padOp = collapseOp.getSrc().getDefiningOp<tensor::PadOp>();

if (!padOp)
return failure();

if (!padOp->hasOneUse())
return failure();

if (!controlFoldingReshapes(&collapseOp.getSrcMutable())) {
return rewriter.notifyMatchFailure(collapseOp,
"fusion blocked by control function");
}

Value constantPaddingValue = padOp.getConstantPaddingValue();
if (!constantPaddingValue) {
return rewriter.notifyMatchFailure(
collapseOp, "cannot fold with non-constant padding value");
}

SmallVector<ReassociationIndices> reassociations =
collapseOp.getReassociationIndices();
SmallVector<OpFoldResult> low = padOp.getMixedLowPad();
SmallVector<OpFoldResult> high = padOp.getMixedHighPad();

for (auto [idx, reInd] : llvm::enumerate(reassociations)) {
if (reInd.size() > 1) {
for (auto dimIdx : reInd) {
if (!isConstantIntValue(low[dimIdx], 0) ||
!isConstantIntValue(high[dimIdx], 0)) {
return failure();
}
}
}
}

SmallVector<OpFoldResult> newLow, newHigh;
for (auto [idx, reInd] : llvm::enumerate(reassociations)) {
newLow.push_back(low[reInd[0]]);
newHigh.push_back(high[reInd[0]]);
}

Copy link
Contributor

@Max191 Max191 Jul 28, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: combine this loop with the loop above.

Location loc = collapseOp.getLoc();
auto resultType = collapseOp.getResultType();

ArrayRef<int64_t> finalShape = collapseOp.getResultType().getShape();

SmallVector<OpFoldResult> collapsedShape;
for (int64_t dimSize : finalShape) {
if (dimSize == ShapedType::kDynamic) {
collapsedShape.push_back(OpFoldResult{});
} else {
collapsedShape.push_back(rewriter.getI64IntegerAttr(dimSize));
}
}
Copy link
Contributor

@Max191 Max191 Jul 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You shouldn't need the mixed sizes for the collapsed shape. It is only used for getting the new type, so you can just collect static sizes instead (SmallVector<int64_t>).


for (auto [inDimIdx, reInd] : llvm::enumerate(reassociations)) {
OpFoldResult l = low[reInd[0]];
OpFoldResult h = high[reInd[0]];

if (!isConstantIntValue(l, 0) || !isConstantIntValue(h, 0)) {
collapsedShape[inDimIdx] =
tensor::getMixedSize(rewriter, loc, padOp.getSource(), reInd[0]);
}
}

SmallVector<int64_t> staticCollapsedShape;
std::tie(staticCollapsedShape, std::ignore) =
decomposeMixedValues(collapsedShape);

auto newCollapseType = RankedTensorType::get(
staticCollapsedShape, padOp.getSource().getType().getElementType());
auto newCollapseOp = rewriter.create<tensor::CollapseShapeOp>(
loc, newCollapseType, padOp.getSource(), reassociations);

rewriter.replaceOpWithNewOp<tensor::PadOp>(
collapseOp, resultType, newCollapseOp.getResult(), newLow, newHigh,
padOp.getConstantPaddingValue(), padOp.getNofold());

return success();
}

private:
ControlFusionFn controlFoldingReshapes;
};

/// Pattern to fold a tensor.expand_shape op with its producer generic op
/// by expanding the dimensionality of the loop in the producer op.
struct FoldReshapeWithGenericOpByExpansion
Expand Down Expand Up @@ -2235,6 +2427,8 @@ void mlir::linalg::populateFoldReshapeOpsByExpansionPatterns(
controlFoldingReshapes);
patterns.add<FoldPadWithProducerReshapeOpByExpansion>(patterns.getContext(),
controlFoldingReshapes);
patterns.add<FoldReshapeWithProducerPadOpByExpansion>(patterns.getContext(),
controlFoldingReshapes);
patterns.add<FoldWithProducerReshapeOpByExpansion>(patterns.getContext(),
controlFoldingReshapes);
}
Expand All @@ -2246,6 +2440,11 @@ void mlir::linalg::populateFoldReshapeOpsByCollapsingPatterns(
controlFoldingReshapes);
patterns.add<FoldPadWithProducerReshapeOpByCollapsing>(
patterns.getContext(), controlFoldingReshapes);
patterns.add<FoldReshapeWithProducerPadOpByCollapsing>(
patterns.getContext(), controlFoldingReshapes);

patterns.add<FoldReshapeWithProducerPadOpByCollapsing>(
patterns.getContext(), controlFoldingReshapes);
patterns.add<FoldReshapeWithGenericOpByCollapsing>(patterns.getContext(),
controlFoldingReshapes);
}
Expand Down
53 changes: 52 additions & 1 deletion mlir/test/Dialect/Linalg/fuse-with-reshape-by-collapsing.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,7 +232,7 @@ func.func @fuse_by_collapsing_dynamic_2(%arg0 : tensor<?xf32>, %sz0: index, %sz1
%1 = linalg.generic {
indexing_maps = [#map0, #map0],
iterator_types = ["parallel", "parallel"]}
ins(%0 : tensor<?x?xf32>)
ins(%0 : tensor<?x?xf32>)
outs(%init : tensor<?x?xf32>) {
^bb0(%b0 : f32, %b1 : f32):
%out = arith.negf %b0 : f32
Expand Down Expand Up @@ -858,3 +858,54 @@ func.func @partial_fuse_by_collapsing(%arg0: tensor<4x?x32x128x192xf16>, %arg1:
// CHECK: %[[COLLAPSED:.+]] = tensor.collapse_shape %[[EXPANDED]]
// CHECK-SAME: tensor<4x128x192x?x32xf32> into tensor<512x192x?xf32>
// CHECK: return %[[COLLAPSED]] : tensor<512x192x?xf32>

// -----

func.func @fold_tensor_pad_with_collapse(%arg0: tensor<32x16x256x256xf32>) -> tensor<512x258x258xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = linalg.fill ins(%cst : f32) outs(%arg0 : tensor<32x16x256x256xf32>) -> tensor<32x16x256x256xf32>
%padded = tensor.pad %0 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%i0: index, %i1: index, %i2: index, %i3: index):
tensor.yield %cst : f32
} : tensor<32x16x256x256xf32> to tensor<32x16x258x258xf32>
%collapsed = tensor.collapse_shape %padded [[0, 1], [2], [3]]
: tensor<32x16x258x258xf32> into tensor<512x258x258xf32>
return %collapsed : tensor<512x258x258xf32>
}
// CHECK: func @fold_tensor_pad_with_collapse(
// CHECK-SAME: %[[ARG0:[^:]+]]: tensor<32x16x256x256xf32>
// CHECK-DAG: %[[CST:.+]] = arith.constant 0.000000e+00 : f32
// CHECK: %[[FILLED:.*]] = linalg.fill ins(%[[CST]] : f32) outs(%[[ARG0]] : tensor<32x16x256x256xf32>)
// CHECK: %[[COLLAPSED:.+]] = tensor.collapse_shape %[[FILLED]] {{\[}}[0, 1], [2], [3]{{\]}}
// CHECK-SAME: : tensor<32x16x256x256xf32> into tensor<512x256x256xf32>
// CHECK: %[[PADDED:.*]] = tensor.pad %[[COLLAPSED]] low[0, 1, 1] high[0, 1, 1]
// CHECK: ^bb0(%[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index):
// CHECK: tensor.yield %[[CST]] : f32
// CHECK: } : tensor<512x256x256xf32> to tensor<512x258x258xf32>
// CHECK: return %[[PADDED]] : tensor<512x258x258xf32>

// -----

func.func @fold_tensor_pad_with_collapse_dynamic_pad_zero(%arg0: tensor<32x16x256x256xf32>) -> tensor<512x258x258xf32> {
%cst = arith.constant 0.000000e+00 : f32
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = linalg.fill ins(%cst : f32) outs(%arg0 : tensor<32x16x256x256xf32>) -> tensor<32x16x256x256xf32>
%padded = tensor.pad %0 low[%c0, %c0, %c1, %c1] high[%c0, %c0, %c1, %c1] {
^bb0(%i0: index, %i1: index, %i2: index, %i3: index):
tensor.yield %cst : f32
} : tensor<32x16x256x256xf32> to tensor<32x16x258x258xf32>
%collapsed = tensor.collapse_shape %padded [[0, 1], [2], [3]]
: tensor<32x16x258x258xf32> into tensor<512x258x258xf32>
return %collapsed : tensor<512x258x258xf32>
}
// CHECK: func @fold_tensor_pad_with_collapse_dynamic_pad_zero(
// CHECK-SAME: %[[ARG0:[^:]+]]: tensor<32x16x256x256xf32>
// CHECK: %[[CST:.+]] = arith.constant 0.000000e+00 : f32
// CHECK: %[[FILLED:.*]] = linalg.fill ins(%[[CST]] : f32) outs(%[[ARG0]] : tensor<32x16x256x256xf32>)
// CHECK: %[[COLLAPSED:.+]] = tensor.collapse_shape %[[FILLED]] {{\[}}[0, 1], [2], [3]{{\]}}
// CHECK-SAME: : tensor<32x16x256x256xf32> into tensor<512x256x256xf32>
// CHECK: %[[PADDED:.*]] = tensor.pad %[[COLLAPSED]] low[0, 1, 1] high[0, 1, 1]
// CHECK: ^bb0(
// CHECK: tensor.yield %[[CST]] : f32
// CHECK: return %[[PADDED]]
51 changes: 49 additions & 2 deletions mlir/test/Dialect/Linalg/reshape_fusion.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -247,7 +247,7 @@ func.func @indexed_consumer_reshape_producer_fusion(%arg0 : tensor<?x?x4x?xi32>,

#map0 = affine_map<(d0, d1) -> (d0, d1)>
func.func @indexed_producer_reshape_consumer_fusion(%arg0 : tensor<?x?xi32>,
%arg1 : tensor<?x?xi32>,
%arg1 : tensor<?x?xi32>,
%sz0: index, %sz1: index) ->
tensor<?x?x4x5xi32>
{
Expand Down Expand Up @@ -515,7 +515,7 @@ func.func @fuse_dynamic_dims(%arg0: tensor<?x?xf32>) -> tensor<?xf32> {
// -----

func.func @reshape_as_consumer_permutation_with_multiple_results
(%a : tensor<?x?x?xf32>, %b : tensor<?x?xf32>, %sz0: index,
(%a : tensor<?x?x?xf32>, %b : tensor<?x?xf32>, %sz0: index,
%sz1: index, %sz2: index, %sz3: index, %sz4: index)
-> (tensor<?x2x?x3x4x?xf32>, tensor<?x?x2x3x4x?xf32>) {
%c:2 = linalg.generic {
Expand Down Expand Up @@ -893,3 +893,50 @@ func.func @move_operand_deps(%arg0 : tensor<?x128xf16>,
// CHECK: %[[GENERIC:.+]] = linalg.generic
// CHECK-SAME: ins(%[[EXPANDED]] :
// CHECK: return %[[GENERIC]]

// -----

func.func @fold_tensor_pad_with_expand(%arg0: tensor<512x256x256xf32>) -> tensor<32x16x258x258xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = linalg.fill ins(%cst : f32) outs(%arg0 : tensor<512x256x256xf32>) -> tensor<512x256x256xf32>
%padded = tensor.pad %0 low[0, 1, 1] high[0, 1, 1] {
^bb0(%i: index, %j: index, %k: index):
tensor.yield %cst : f32
} : tensor<512x256x256xf32> to tensor<512x258x258xf32>
%expanded = tensor.expand_shape %padded [[0, 1], [2], [3]] output_shape [32, 16, 258, 258] : tensor<512x258x258xf32> into tensor<32x16x258x258xf32>
return %expanded : tensor<32x16x258x258xf32>
}
// CHECK: func @fold_tensor_pad_with_expand(
// CHECK-SAME: %[[ARG0:[^:]+]]: tensor<512x256x256xf32>
// CHECK-DAG: %[[CST:.+]] = arith.constant 0.000000e+00 : f32
// CHECK-DAG: %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]]
// CHECK: %[[FILLED:.*]] = linalg.fill ins(%[[CST]] : f32) outs(%[[EXPANDED]] : tensor<32x16x256x256xf32>)
// CHECK: %[[PADDED:.*]] = tensor.pad %[[FILLED]] low[0, 0, 1, 1] high[0, 0, 1, 1]
// CHECK: ^bb0(%[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index, %[[ARG4:.*]]: index):
// CHECK: tensor.yield %[[CST]] : f32
// CHECK: } : tensor<32x16x256x256xf32> to tensor<32x16x258x258xf32>
// CHECK: return %[[PADDED]] : tensor<32x16x258x258xf32>

// -----

func.func @fold_tensor_pad_with_expand_dynamic_pad_zero(%arg0: tensor<512x256x256xf32>) -> tensor<32x16x258x258xf32> {
%cst = arith.constant 0.000000e+00 : f32
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = linalg.fill ins(%cst : f32) outs(%arg0 : tensor<512x256x256xf32>) -> tensor<512x256x256xf32>
%padded = tensor.pad %0 low[%c0, %c1, %c1] high[%c0, %c1, %c1] {
^bb0(%i: index, %j: index, %k: index):
tensor.yield %cst : f32
} : tensor<512x256x256xf32> to tensor<512x258x258xf32>
%expanded = tensor.expand_shape %padded [[0, 1], [2], [3]] output_shape [32, 16, 258, 258] : tensor<512x258x258xf32> into tensor<32x16x258x258xf32>
return %expanded : tensor<32x16x258x258xf32>
}
// CHECK: func @fold_tensor_pad_with_expand_dynamic_pad_zero(
// CHECK-SAME: %[[ARG0:[^:]+]]: tensor<512x256x256xf32>
// CHECK: %[[CST:.+]] = arith.constant 0.000000e+00 : f32
// CHECK: %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]]
// CHECK: %[[FILLED:.*]] = linalg.fill ins(%[[CST]] : f32) outs(%[[EXPANDED]]
// CHECK: %[[PADDED:.*]] = tensor.pad %[[FILLED]] low[0, 0, 1, 1] high[0, 0, 1, 1]
// CHECK: ^bb0(
// CHECK: tensor.yield %[[CST]] : f32
// CHECK: return %[[PADDED]]