Skip to content

[mlir][vector] Use notifyMatchFailure instead of assert in VectorLinearize #93590

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
Jun 21, 2024
Merged

[mlir][vector] Use notifyMatchFailure instead of assert in VectorLinearize #93590

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
35 changes: 22 additions & 13 deletions mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,10 +151,12 @@ struct LinearizeVectorExtractStridedSlice final
LogicalResult
matchAndRewrite(vector::ExtractStridedSliceOp extractOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Type dstType = getTypeConverter()->convertType(extractOp.getType());
assert(!(extractOp.getVector().getType().isScalable() ||
cast<VectorType>(dstType).isScalable()) &&
"scalable vectors are not supported.");
VectorType dstType =
getTypeConverter()->convertType<VectorType>(extractOp.getType());
assert(dstType && "vector type destination expected.");
if (extractOp.getVector().getType().isScalable() || dstType.isScalable())
return rewriter.notifyMatchFailure(extractOp,
"scalable vectors are not supported.");
if (!isLessThanTargetBitWidth(extractOp, targetVectorBitWidth))
return rewriter.notifyMatchFailure(
extractOp, "Can't flatten since targetBitWidth <= OpSize");
Expand Down Expand Up @@ -264,10 +266,14 @@ struct LinearizeVectorShuffle final
LogicalResult
matchAndRewrite(vector::ShuffleOp shuffleOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Type dstType = getTypeConverter()->convertType(shuffleOp.getType());
VectorType dstType =
getTypeConverter()->convertType<VectorType>(shuffleOp.getType());
assert(dstType && "vector type destination expected.");
// The assert is used because vector.shuffle does not support scalable
// vectors.
assert(!(shuffleOp.getV1VectorType().isScalable() ||
shuffleOp.getV2VectorType().isScalable() ||
cast<VectorType>(dstType).isScalable()) &&
dstType.isScalable()) &&
"scalable vectors are not supported.");
if (!isLessThanTargetBitWidth(shuffleOp, targetVectorBitWidth))
return rewriter.notifyMatchFailure(
Expand Down Expand Up @@ -336,9 +342,10 @@ struct LinearizeVectorExtract final
matchAndRewrite(vector::ExtractOp extractOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Type dstTy = getTypeConverter()->convertType(extractOp.getType());
assert(!(extractOp.getVector().getType().isScalable() ||
cast<VectorType>(dstTy).isScalable()) &&
"scalable vectors are not supported.");
if (extractOp.getVector().getType().isScalable() ||
cast<VectorType>(dstTy).isScalable())
Copy link
Contributor

Choose a reason for hiding this comment

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

Actually the dstTy here is not always a vector type. It could be a scalar type too.
e.g., vector.extract %1 [0, 0]: f32 from vector<1024x1024xf32>. So, cast(dstTy) may cause the crash.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Would you say the following is better?

if ( (auto vecDstTy = cast<VectorType>(dstTy) && vecDstTy.isScalable()) || extractOp.getVector().getType().isScalable() )

Copy link
Contributor

Choose a reason for hiding this comment

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

So, cast(dstTy) may cause the crash.

Good point, but let's stick to one change per PR 😅 My recommendation:

  1. identify a test case that would indeed crash,
  2. fix the crash and use the test from 1. for a follow-up PR.

WDYT?

Copy link
Contributor

Choose a reason for hiding this comment

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

What's the update here?

Copy link
Contributor Author

@akroviakov akroviakov Jun 13, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

The test case is there:

%0 = vector.extract %arg0[1,1,1]: f32 from vector<2x8x2xf32>

but fixing it can get a bit tricky because right now there are a lot of vector result assumptions (e.g., isLessThanTargetBitWidth(), populateVectorLinearizeShuffleLikeOpsPatterns()), so yes, it should be another PR.
Any suggestions to make the fix least invasive are welcomed.

Copy link
Contributor

Choose a reason for hiding this comment

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

Any suggestions to make the fix least invasive are welcomed.

You should disable this pattern when the rank of the output is <= 1.

return rewriter.notifyMatchFailure(extractOp,
"scalable vectors are not supported.");
if (!isLessThanTargetBitWidth(extractOp, targetVectorBitWidth))
return rewriter.notifyMatchFailure(
extractOp, "Can't flatten since targetBitWidth <= OpSize");
Expand Down Expand Up @@ -394,10 +401,12 @@ struct LinearizeVectorInsert final
LogicalResult
matchAndRewrite(vector::InsertOp insertOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Type dstTy = getTypeConverter()->convertType(insertOp.getDestVectorType());
assert(!(insertOp.getDestVectorType().isScalable() ||
cast<VectorType>(dstTy).isScalable()) &&
"scalable vectors are not supported.");
VectorType dstTy = getTypeConverter()->convertType<VectorType>(
insertOp.getDestVectorType());
assert(dstTy && "vector type destination expected.");
Copy link
Contributor

Choose a reason for hiding this comment

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

Same comments as above regarding to assert.

if (insertOp.getDestVectorType().isScalable() || dstTy.isScalable())
return rewriter.notifyMatchFailure(insertOp,
"scalable vectors are not supported.");

if (!isLessThanOrEqualTargetBitWidth(insertOp.getSourceType(),
targetVectorBitWidth))
Expand Down
32 changes: 32 additions & 0 deletions mlir/test/Dialect/Vector/linearize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,6 +177,17 @@ func.func @test_extract_strided_slice_1(%arg0 : vector<4x8xf32>) -> vector<2x2xf
return %0 : vector<2x2xf32>
}

// ALL-LABEL: func.func @test_extract_strided_slice_1_scalable(
// ALL-SAME: %[[VAL_0:.*]]: vector<4x[8]xf32>) -> vector<2x[8]xf32> {
func.func @test_extract_strided_slice_1_scalable(%arg0: vector<4x[8]xf32>) -> vector<2x[8]xf32> {
// ALL-NOT: vector.shuffle
// ALL-NOT: vector.shape_cast
// ALL: %[[RES:.*]] = vector.extract_strided_slice %[[VAL_0]] {offsets = [1, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x[8]xf32> to vector<2x[8]xf32>
%0 = vector.extract_strided_slice %arg0 { sizes = [2, 8], strides = [1, 1], offsets = [1, 0] } : vector<4x[8]xf32> to vector<2x[8]xf32>
// ALL: return %[[RES]] : vector<2x[8]xf32>
return %0 : vector<2x[8]xf32>
}

// -----
// ALL-LABEL: test_extract_strided_slice_2
// ALL-SAME: (%[[ORIG_ARG:.*]]: vector<2x8x2xf32>) -> vector<1x4x2xf32> {
Expand Down Expand Up @@ -246,6 +257,16 @@ func.func @test_vector_extract(%arg0: vector<2x8x2xf32>) -> vector<8x2xf32> {
return %0 : vector<8x2xf32>
}

// ALL-LABEL: func.func @test_vector_extract_scalable(
// ALL-SAME: %[[VAL_0:.*]]: vector<2x8x[2]xf32>) -> vector<8x[2]xf32> {
func.func @test_vector_extract_scalable(%arg0: vector<2x8x[2]xf32>) -> vector<8x[2]xf32> {
// ALL-NOT: vector.shuffle
// ALL-NOT: vector.shape_cast
// ALL: %[[RES:.*]] = vector.extract %[[VAL_0]][1] : vector<8x[2]xf32> from vector<2x8x[2]xf32>
%0 = vector.extract %arg0[1]: vector<8x[2]xf32> from vector<2x8x[2]xf32>
// ALL: return %[[RES]] : vector<8x[2]xf32>
return %0 : vector<8x[2]xf32>
}
// -----
// ALL-LABEL: test_vector_insert
// ALL-SAME: (%[[DEST:.*]]: vector<2x8x4xf32>, %[[SRC:.*]]: vector<8x4xf32>) -> vector<2x8x4xf32> {
Expand Down Expand Up @@ -274,3 +295,14 @@ func.func @test_vector_insert(%arg0: vector<2x8x4xf32>, %arg1: vector<8x4xf32>)
%0 = vector.insert %arg1, %arg0[0]: vector<8x4xf32> into vector<2x8x4xf32>
return %0 : vector<2x8x4xf32>
}

// ALL-LABEL: func.func @test_vector_insert_scalable(
// ALL-SAME: %[[VAL_0:.*]]: vector<2x8x[4]xf32>, %[[VAL_1:.*]]: vector<8x[4]xf32>) -> vector<2x8x[4]xf32> {
func.func @test_vector_insert_scalable(%arg0: vector<2x8x[4]xf32>, %arg1: vector<8x[4]xf32>) -> vector<2x8x[4]xf32> {
// ALL-NOT: vector.shuffle
// ALL-NOT: vector.shape_cast
// ALL: %[[RES:.*]] = vector.insert %[[VAL_1]], %[[VAL_0]] [0] : vector<8x[4]xf32> into vector<2x8x[4]xf32>
%0 = vector.insert %arg1, %arg0[0]: vector<8x[4]xf32> into vector<2x8x[4]xf32>
// ALL: return %[[RES]] : vector<2x8x[4]xf32>
return %0 : vector<2x8x[4]xf32>
}
Loading