Skip to content

[mlir] Fix bug in pack and unpack op canonicalization for folding dynamic dims #82539

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 4 commits into from
Feb 28, 2024
Merged

[mlir] Fix bug in pack and unpack op canonicalization for folding dynamic dims #82539

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
04877c5
[mlir] Fix bug in pack op canonicalization for folding dynamic dims
Max191 Feb 21, 2024
be565c7
fix tests
Max191 Feb 21, 2024
c21cd97
fix unpack
Max191 Feb 21, 2024
3a1d710
add new test cases
Max191 Feb 21, 2024
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
16 changes: 10 additions & 6 deletions mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4012,15 +4012,17 @@ static bool inferStaticShape(PackOp packOp, SmallVectorImpl<int64_t> &srcShape,
llvm::SmallSetVector<int64_t, 4> innerDims;
innerDims.insert(packOp.getInnerDimsPos().begin(),
packOp.getInnerDimsPos().end());
auto outerDimsPerm = packOp.getOuterDimsPerm();
SmallVector<int64_t> inverseOuterDimsPerm;
if (!packOp.getOuterDimsPerm().empty())
inverseOuterDimsPerm = invertPermutationVector(packOp.getOuterDimsPerm());
int srcRank = packOp.getSourceRank();
for (auto i : llvm::seq<int64_t>(0, srcRank)) {
if (innerDims.contains(i))
continue;
int64_t srcPos = i;
int64_t destPos = i;
if (!outerDimsPerm.empty())
destPos = outerDimsPerm[srcPos];
if (!inverseOuterDimsPerm.empty())
destPos = inverseOuterDimsPerm[srcPos];
if (ShapedType::isDynamic(srcShape[srcPos]) ==
ShapedType::isDynamic(destShape[destPos])) {
continue;
Expand Down Expand Up @@ -4240,15 +4242,17 @@ static bool inferStaticShape(UnPackOp op, SmallVectorImpl<int64_t> &srcShape,
op.getDestType().getShape().end());
llvm::SmallSetVector<int64_t, 4> innerDims;
innerDims.insert(op.getInnerDimsPos().begin(), op.getInnerDimsPos().end());
auto outerDimsPerm = op.getOuterDimsPerm();
SmallVector<int64_t> inverseOuterDimsPerm;
if (!op.getOuterDimsPerm().empty())
inverseOuterDimsPerm = invertPermutationVector(op.getOuterDimsPerm());
int destRank = op.getDestRank();
for (auto i : llvm::seq<int64_t>(0, destRank)) {
if (innerDims.contains(i))
continue;
int64_t srcPos = i;
int64_t destPos = i;
if (!outerDimsPerm.empty())
srcPos = outerDimsPerm[destPos];
if (!inverseOuterDimsPerm.empty())
srcPos = inverseOuterDimsPerm[destPos];
if (ShapedType::isDynamic(srcShape[srcPos]) ==
ShapedType::isDynamic(destShape[destPos])) {
continue;
Expand Down
35 changes: 29 additions & 6 deletions mlir/test/Dialect/Tensor/canonicalize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -822,7 +822,7 @@ func.func @infer_src_shape_pack(%src: tensor<?x?x?x?xf32>, %dest: tensor<10x20x3
// CHECK-LABEL: func.func @infer_src_shape_pack
// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?xf32> to tensor<30x20x?x10xf32>
// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?xf32> to tensor<40x20x?x30xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[CAST_SRC]] {{.+}} into %[[DEST]]
// CHECK: return %[[PACK]]

Expand All @@ -841,13 +841,24 @@ func.func @infer_dest_shape_pack(%src: tensor<30x20x?x10xf32>, %dest: tensor<?x?
// CHECK-LABEL: func.func @infer_dest_shape_pack
// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?x16xf32> to tensor<10x20x30x?x16xf32>
// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?x16xf32> to tensor<?x20x10x30x16xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[SRC]] {{.+}} into %[[CAST_DEST]]
// CHECK: %[[CAST_PACK:.+]] = tensor.cast %[[PACK]] : tensor<10x20x30x?x16xf32> to tensor<?x?x?x?x16xf32>
// CHECK: %[[CAST_PACK:.+]] = tensor.cast %[[PACK]] : tensor<?x20x10x30x16xf32> to tensor<?x?x?x?x16xf32>
// CHECK: return %[[CAST_PACK]]

// -----

func.func @no_infer_pack_shape(%arg0: tensor<?x32x100xf32>, %arg1: index) -> tensor<32x7x?x16x1xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = tensor.empty(%arg1) : tensor<32x7x?x16x1xf32>
%pack = tensor.pack %arg0 padding_value(%cst : f32) outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 0] inner_tiles = [16, 1] into %0 : tensor<?x32x100xf32> -> tensor<32x7x?x16x1xf32>
return %pack : tensor<32x7x?x16x1xf32>
}
// CHECK-LABEL: func.func @no_infer_pack_shape
// CHECK-NOT: tensor.cast

// -----

func.func @fold_padding_value_pack_negative1(%arg0: tensor<1200x499999xf32>) -> tensor<31250x1200x16x1xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = tensor.empty() : tensor<31250x1200x16x1xf32>
Expand Down Expand Up @@ -920,9 +931,9 @@ func.func @infer_dest_shape_unpack(%src: tensor<10x20x30x40x16xf32>, %dest: tens
// CHECK-LABEL: func.func @infer_dest_shape_unpack
// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?xf32> to tensor<30x20x?x10xf32>
// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?xf32> to tensor<40x20x?x30xf32>
// CHECK: %[[UNPACK:.+]] = tensor.unpack %[[SRC]] {{.+}} into %[[CAST_DEST]]
// CHECK: %[[CAST_UNPACK:.+]] = tensor.cast %[[UNPACK]] : tensor<30x20x?x10xf32> to tensor<?x?x?x?xf32>
// CHECK: %[[CAST_UNPACK:.+]] = tensor.cast %[[UNPACK]] : tensor<40x20x?x30xf32> to tensor<?x?x?x?xf32>
// CHECK: return %[[CAST_UNPACK]]

// -----
Expand All @@ -938,12 +949,24 @@ func.func @infer_src_shape_unpack(%src: tensor<?x?x?x?x16xf32>, %dest: tensor<30
// CHECK-LABEL: func.func @infer_src_shape_unpack
// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?x16xf32> to tensor<10x20x30x?x16xf32>
// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?x16xf32> to tensor<?x20x10x30x16xf32>
// CHECK: %[[UNPACK:.+]] = tensor.unpack %[[CAST_SRC]]
// CHECK: return %[[UNPACK]]

// -----

func.func @no_infer_unpack_shape(%arg1: tensor<32x7x?x16x1xf32>, %arg2: index) -> tensor<?x32x100xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = tensor.empty(%arg2) : tensor<?x32x100xf32>
%unpack = tensor.unpack %arg1 outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 0] inner_tiles = [16, 1] into %0 : tensor<32x7x?x16x1xf32> -> tensor<?x32x100xf32>
return %unpack : tensor<?x32x100xf32>
}
// CHECK-LABEL: func.func @no_infer_unpack_shape
// CHECK-NOT: tensor.cast

// -----


// CHECK-LABEL: func @fold_overlapping_insert
// CHECK-SAME: %[[INPUT:.+]]: tensor<?x?x?xf32>, %{{.+}}: tensor<4x?x8xf32>, %[[SLICE2:.+]]: tensor<4x?x8xf32>
func.func @fold_overlapping_insert(%input : tensor<?x?x?xf32>, %slice1: tensor<4x?x8xf32>, %slice2: tensor<4x?x8xf32>, %i: index, %size: index) -> (tensor<?x?x?xf32>) {
Expand Down