[mlir][linalg] Add transpose support for reshape as consumer fusion #130344
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
nirvedhmeshram
requested review from
MaheshRavishankar and
nicolasvasilache
as code owners
Member
|
@llvm/pr-subscribers-mlir @llvm/pr-subscribers-mlir-linalg Author: Nirvedh Meshram (nirvedhmeshram) ChangesDuring #129128 adding reshape as consumer fusion handling of linalg.transpose was missed. This PR adds that. Full diff: https://github.com/llvm/llvm-project/pull/130344.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp b/mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp
index a45b5c43f5d33..a35afc60d6cb0 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp
@@ -816,17 +816,51 @@ validateDynamicDimExpansion(LinalgOp linalgOp,
}
// Create an expanded transpose op.
+// For bubbling a collapse : transpose(collapse_shape),
+// all expanded groups are permuted together. We just permute the reassocation
+// map of the collapse and flatten it. For example,
+//
+// reassociation_map = [[0], [1, 2, 3], [4, 5]]
+// permutation = [2, 0, 1]
+//
+// Becomes
+//
+// permutation = [4, 5, 0 , 1, 2, 3]
+//
+// For sinking expand : expand_shape(transpose),
+// the reassociation map is already permuted hence we inverse permutate and then
+// flatten it. Then we inverse permute it again to get the final expanded
+// transpose permutation. For example,
+//
+// permutation = [2, 0, 1]
+// reassociation_map = [[0, 1], [2], [3, 4, 5]]
+//
+// inverse permutation = [1, 2, 0]
+// applied to reassocation_map and then flattened becomes
+// flatened permutation = [2, 3, 4, 5, 0, 1]
+// final permuation is the inverse of the flattened permutation.
+//
+// Becomes
+//
+// permutation=[4, 5, 0, 1, 2, 3]
+
static Operation *
createExpandedTransposeOp(PatternRewriter &rewriter, TransposeOp transposeOp,
SmallVector<ReassociationIndices> reassociation,
- Value expandedInput, Value output) {
- applyPermutationToVector(reassociation, transposeOp.getPermutation());
+ Value expandedInput, Value output, bool isExpanding) {
+ ArrayRef<int64_t> permutation =
+ isExpanding ? invertPermutationVector(transposeOp.getPermutation())
+ : transposeOp.getPermutation();
+ applyPermutationToVector(reassociation, permutation);
SmallVector<int64_t> newPerm;
for (auto reassoc : reassociation) {
for (auto dim : reassoc) {
newPerm.push_back(dim);
}
}
+ if (isExpanding) {
+ newPerm = invertPermutationVector(newPerm);
+ }
return rewriter.create<TransposeOp>(transposeOp.getLoc(), expandedInput,
output, newPerm);
}
@@ -866,12 +900,13 @@ static Operation *createExpandedOp(
PatternRewriter &rewriter, LinalgOp linalgOp, TypeRange resultTypes,
ArrayRef<Value> expandedOpOperands, ArrayRef<Value> outputs,
ArrayRef<AffineMap> expandedOpIndexingMaps, ExpansionInfo &expansionInfo,
- SmallVector<ReassociationIndices> reassociation) {
+ SmallVector<ReassociationIndices> reassociation, bool isExpanding) {
return TypeSwitch<Operation *, Operation *>(linalgOp.getOperation())
.Case<TransposeOp>([&](TransposeOp transposeOp) {
return createExpandedTransposeOp(rewriter, transposeOp, reassociation,
- expandedOpOperands[0], outputs[0]);
+ expandedOpOperands[0], outputs[0],
+ isExpanding);
})
.Case<FillOp, CopyOp>([&](Operation *op) {
return clone(rewriter, linalgOp, resultTypes,
@@ -994,9 +1029,10 @@ fuseWithReshapeByExpansion(LinalgOp linalgOp, Operation *reshapeOp,
SmallVector<ReassociationIndices> reassociationBeforeExpansion =
isExpanding ? expandingReshapeOp.getReassociationIndices()
: collapsingReshapeOp.getReassociationIndices();
- Operation *fusedOp = createExpandedOp(
- rewriter, linalgOp, resultTypes, expandedOpOperands, outputs,
- expandedOpIndexingMaps, expansionInfo, reassociationBeforeExpansion);
+ Operation *fusedOp =
+ createExpandedOp(rewriter, linalgOp, resultTypes, expandedOpOperands,
+ outputs, expandedOpIndexingMaps, expansionInfo,
+ reassociationBeforeExpansion, isExpanding);
// Reshape the result values to their original shape if this is a collapsing
// reshape folded into its consumer.
SmallVector<Value> resultVals;
diff --git a/mlir/test/Dialect/Linalg/reshape_fusion.mlir b/mlir/test/Dialect/Linalg/reshape_fusion.mlir
index 4da9c0851ac70..7c2b55ca745ff 100644
--- a/mlir/test/Dialect/Linalg/reshape_fusion.mlir
+++ b/mlir/test/Dialect/Linalg/reshape_fusion.mlir
@@ -195,7 +195,7 @@ func.func @generic_op_reshape_consumer_static(%arg0: tensor<264x4xf32>)
// CHECK-SAME: : tensor<8x33x4xf32>
// CHECK-DAG: %[[INIT:.+]] = tensor.empty()
// CHECK: %[[T0:.+]] = tensor.expand_shape %[[ARG0]] {{\[\[}}0, 1], [2]] output_shape [8, 33, 4] : tensor<264x4xf32> into tensor<8x33x4xf32>
-// CHECK: %[[T1:.+]] = tensor.expand_shape %[[VAL_0]] {{\[\[}}0, 1], [2]] output_shape [8, 33, 4] : tensor<264x4xf32> into tensor<8x33x4xf32>
+// CHECK: %[[T1:.+]] = tensor.expand_shape %[[INIT]] {{\[\[}}0, 1], [2]] output_shape [8, 33, 4] : tensor<264x4xf32> into tensor<8x33x4xf32>
// CHECK: %[[T2:.+]] = linalg.generic
// CHECK-SAME: indexing_maps = [#[[MAP2]], #[[MAP2]], #[[MAP2]]]
// CHECK-SAME: ["parallel", "parallel", "parallel"]
@@ -203,6 +203,29 @@ func.func @generic_op_reshape_consumer_static(%arg0: tensor<264x4xf32>)
// CHECK-SAME: outs(%[[T1]] : tensor<8x33x4xf32>)
// CHECK: return %[[T2]] : tensor<8x33x4xf32>
+// -----
+
+func.func @reshape_as_consumer_transpose
+ (%a : tensor<4x210x6xf32>)
+ -> tensor<2x3x4x5x6x7xf32> {
+ %b = tensor.empty() : tensor<6x4x210xf32>
+ %c = linalg.transpose
+ ins(%a : tensor<4x210x6xf32>)
+ outs(%b : tensor<6x4x210xf32>) permutation = [2, 0, 1]
+ %d = tensor.expand_shape %c [[0, 1], [2], [3, 4, 5]] output_shape [2, 3, 4, 5, 6, 7] : tensor<6x4x210xf32> into tensor<2x3x4x5x6x7xf32>
+ return %d : tensor<2x3x4x5x6x7xf32>
+}
+// CHECK: func @reshape_as_consumer_transpose
+// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: tensor<4x210x6xf32>
+// CHECK-DAG: %[[INIT:.+]] = tensor.empty()
+// CHECK-DAG: %[[T0:.+]] = tensor.expand_shape %[[ARG0]] {{\[\[}}0], [1, 2, 3], [4, 5]] output_shape [4, 5, 6, 7, 2, 3] : tensor<4x210x6xf32> into tensor<4x5x6x7x2x3xf32>
+// CHECK-DAG: %[[T1:.+]] = tensor.expand_shape %[[INIT]] {{\[\[}}0, 1], [2], [3, 4, 5]] output_shape [2, 3, 4, 5, 6, 7] : tensor<6x4x210xf32> into tensor<2x3x4x5x6x7xf32
+// CHECK: %[[T2:.+]] = linalg.transpose ins(%[[T0]] : tensor<4x5x6x7x2x3xf32>)
+// CHECK-SAME: outs(%[[T1]] : tensor<2x3x4x5x6x7xf32>)
+// CHECK-SAME: permutation = [4, 5, 0, 1, 2, 3]
+// CHECK: return %[[T2]] : tensor<2x3x4x5x6x7xf32>
+
+
// -----
#map0 = affine_map<(d0, d1, d2) -> (d2, d0, d1)>
@@ -859,37 +882,29 @@ func.func @linalg_copy_reshape_producer_fusion(%arg0 : tensor<?x7x?x8xf32>,
// -----
-func.func @linalg_transpose_reshape_producer_fusion(%arg0 : tensor<?x7x?x8xf32>,
- %arg1 : tensor<?x?xf32>) ->
- tensor<?x?xf32>
-{
- %0 = tensor.collapse_shape %arg0 [[0, 1], [2, 3]] :
- tensor<?x7x?x8xf32> into tensor<?x?xf32>
- %1 = linalg.transpose ins(%0 : tensor<?x?xf32>)
- outs(%arg1 : tensor<?x?xf32>) permutation = [1, 0]
- return %1 : tensor<?x?xf32>
+
+func.func @reshape_as_producer_transpose
+ (%a : tensor<4x5x6x7x2x3xf32>)
+ -> tensor<6x4x210xf32> {
+ %b = tensor.empty() : tensor<6x4x210xf32>
+ %c = tensor.collapse_shape %a [[0], [1, 2, 3], [4, 5]] :
+ tensor<4x5x6x7x2x3xf32> into tensor<4x210x6xf32>
+ %d = linalg.transpose
+ ins(%c : tensor<4x210x6xf32>)
+ outs(%b : tensor<6x4x210xf32>) permutation = [2, 0, 1]
+ return %d : tensor<6x4x210xf32>
}
-// CHECK: func @linalg_transpose_reshape_producer_fusion
-// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: tensor<?x7x?x8xf32>
-// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
-// CHECK-DAG: %[[C8:.+]] = arith.constant 8 : index
-// CHECK-DAG: %[[C7:.+]] = arith.constant 7 : index
-// CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
-// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
-// CHECK-DAG: %[[DIM:.+]] = tensor.dim %[[ARG1]], %[[C0]] : tensor<?x?xf32>
-// CHECK-DAG: %[[DIM_0:.+]] = tensor.dim %[[ARG1]], %[[C1]] : tensor<?x?xf32>
-// CHECK-DAG: %[[VAL_0:.+]] = arith.divsi %[[DIM_0]], %[[C7]] : index
-// CHECK-DAG: %[[VAL_1:.+]] = arith.divsi %[[DIM]], %[[C8]] : index
-// CHECK: %[[T1:.+]] = tensor.expand_shape %[[ARG1]] {{\[\[}}0, 1], [2, 3]] output_shape [%[[VAL_1]], 8, %[[VAL_0]], 7] : tensor<?x?xf32> into tensor<?x8x?x7xf32>
-// CHECK: %[[T2:.+]] = linalg.transpose
-// CHECK-SAME: ins(%[[ARG0]] : tensor<?x7x?x8xf32>)
-// CHECK-SAME: outs(%[[T1]] : tensor<?x8x?x7xf32>)
-// CHECK-SAME: permutation = [2, 3, 0, 1]
-// CHECK: %[[T3:.+]] = tensor.collapse_shape %[[T2]]
-// CHECK-SAME: [0, 1], [2, 3]
-// CHECK-SAME: tensor<?x8x?x7xf32> into tensor<?x?xf32>
-// CHECK: return %[[T3]]
+// CHECK: func @reshape_as_producer_transpose
+// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: tensor<4x5x6x7x2x3xf32>
+// CHECK-DAG: %[[INIT:.+]] = tensor.empty()
+// CHECK-DAG: %[[T0:.+]] = tensor.expand_shape %[[INIT]] {{\[\[}}0, 1], [2], [3, 4, 5]] output_shape [2, 3, 4, 5, 6, 7] : tensor<6x4x210xf32> into tensor<2x3x4x5x6x7xf32>
+// CHECK: %[[T1:.+]] = linalg.transpose ins(%[[ARG0]] : tensor<4x5x6x7x2x3xf32>)
+// CHECK-SAME: outs(%[[T0]] : tensor<2x3x4x5x6x7xf32>)
+// CHECK-SAME: permutation = [4, 5, 0, 1, 2, 3]
+// CHECK: %[[T2:.+]] = tensor.collapse_shape %[[T1]] {{\[\[}}0, 1], [2], [3, 4, 5]] : tensor<2x3x4x5x6x7xf32> into tensor<6x4x210xf32>
+// CHECK: return %[[T2]] : tensor<6x4x210xf32>
+
// -----
|
nirvedhmeshram
force-pushed
the
reshape_as_consumer
branch
from
6ee32fc to
14cb232
Compare
Signed-off-by: Nirvedh Meshram <[email protected]>
Signed-off-by: Nirvedh Meshram <[email protected]>
Signed-off-by: Nirvedh Meshram <[email protected]>
nirvedhmeshram
force-pushed
the
reshape_as_consumer
branch
from
9adaa31 to
75d238a
Compare
MaheshRavishankar
approved these changes
Mar 11, 2025
Contributor
|
That makes sense. Thanks. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
During #129128 adding reshape as consumer fusion handling of linalg.transpose was missed. This PR adds that.
Also transpose reshape as producer fusion test is updated to static sizes as that is more likely to catch any issues with the permutation vector in the verifier if the shapes dont match up.