Skip to content

[mlir] Reuse pack dest in tensor.pack decomposition #108025

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

Merged
merged 1 commit into from
Sep 10, 2024
Merged

[mlir] Reuse pack dest in tensor.pack decomposition #108025

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 2 additions & 5 deletions mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -305,8 +305,6 @@ FailureOr<LowerPackResult> linalg::lowerPack(RewriterBase &rewriter,
if (rankReduces == SliceVerificationResult::Success) {
// This pack is just a plain pad.
// Just insert the pad in the higher ranked tensor.
auto emptyOp =
rewriter.create<tensor::EmptyOp>(loc, packedTensorType, ValueRange{});
// Offsets.
SmallVector<OpFoldResult> zeros(packOp.getDestRank(),
rewriter.getIndexAttr(0));
Expand All @@ -317,9 +315,8 @@ FailureOr<LowerPackResult> linalg::lowerPack(RewriterBase &rewriter,
tensor::getMixedSizes(rewriter, loc, packOp.getDest());

auto insertSliceOp = rewriter.create<tensor::InsertSliceOp>(
loc, /*source=*/padOp, /*dest=*/emptyOp,
/*offsets=*/zeros, sizes,
/*strides=*/ones);
loc, /*source=*/padOp, /*dest=*/packOp.getDest(),
/*offsets=*/zeros, sizes, /*strides=*/ones);

LLVM_DEBUG(DBGS() << "insert_slice op: " << insertSliceOp; DBGSNL(););

Expand Down
14 changes: 8 additions & 6 deletions mlir/test/Dialect/Linalg/transform-lower-pack.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,14 +62,15 @@ module attributes {transform.with_named_sequence} {
// -----

// CHECK-LABEL: func.func @pack_as_pad(
// CHECK: %[[SRC:.+]]: tensor<129x47x16x16xf32>,
// CHECK: %[[OUT:.+]]: tensor<1x1x1x1x136x64x16x16xf32>)
Comment on lines +65 to +66
Copy link
Contributor

Choose a reason for hiding this comment

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

Tip: you can write %[[SRC:[a-zA-Z0-9]+]] to avoid capturing the types, which is not important in the lit tests.

func.func @pack_as_pad(%arg0: tensor<129x47x16x16xf32>, %arg1: tensor<1x1x1x1x136x64x16x16xf32>) -> tensor<1x1x1x1x136x64x16x16xf32> {
%cst_0 = arith.constant 0.0 : f32

// tensor.pack is lowered to tensor.pad + tensor.insert_slice
// CHECK: %[[PAD:.*]] = tensor.pad {{.*}} low[0, 0, 0, 0]
// CHECK: %[[PAD:.*]] = tensor.pad %[[SRC]] low[0, 0, 0, 0] high[7, 17, 0, 0]
// CHECK: : tensor<129x47x16x16xf32> to tensor<136x64x16x16xf32>
// CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<1x1x1x1x136x64x16x16xf32>
// CHECK: %[[RES:.*]] = tensor.insert_slice %[[PAD]] into %[[EMPTY]]
// CHECK: %[[RES:.*]] = tensor.insert_slice %[[PAD]] into %[[OUT]]
// offsets.
// CHECK-SAME: [0, 0, 0, 0, 0, 0, 0, 0]
// sizes.
Expand Down Expand Up @@ -387,14 +388,15 @@ module attributes {transform.with_named_sequence} {
// -----

// CHECK-LABEL: func.func @pack_as_pad_with_outer_dims_perm(
// CHECK: %[[SRC:.+]]: tensor<129x47x16x16xf32>,
// CHECK: %[[OUT:.+]]: tensor<1x1x1x1x136x64x16x16xf32>)
func.func @pack_as_pad_with_outer_dims_perm(%arg0: tensor<129x47x16x16xf32>, %arg1: tensor<1x1x1x1x136x64x16x16xf32>) -> tensor<1x1x1x1x136x64x16x16xf32> {
%cst_0 = arith.constant 0.0 : f32

// tensor.pack is lowered to tensor.pad + tensor.insert_slice
// CHECK: %[[PAD:.*]] = tensor.pad {{.*}} low[0, 0, 0, 0]
// CHECK: %[[PAD:.*]] = tensor.pad %[[SRC]] low[0, 0, 0, 0] high[7, 17, 0, 0]
// CHECK: : tensor<129x47x16x16xf32> to tensor<136x64x16x16xf32>
// CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<1x1x1x1x136x64x16x16xf32>
// CHECK: %[[RES:.*]] = tensor.insert_slice %[[PAD]] into %[[EMPTY]]
// CHECK: %[[RES:.*]] = tensor.insert_slice %[[PAD]] into %[[OUT]]
// offsets.
// CHECK-SAME: [0, 0, 0, 0, 0, 0, 0, 0]
// sizes.
Expand Down
Loading