Skip to content
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

Fix ASAN/UBSAN issues in DimAnalysis #3052

Merged
merged 1 commit into from
Jan 29, 2025
Merged

Fix ASAN/UBSAN issues in DimAnalysis #3052

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
16 changes: 9 additions & 7 deletions src/Dialect/ONNX/ONNXDimAnalysis.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -361,11 +361,11 @@ static bool exploreSameDimsUsingShapeHelper(const DimAnalysis::DimT &dim,
ONNXOpShapeHelper *shapeHelper =
shape_op.getShapeHelper(op, {}, nullptr, nullptr);
// If no shape helper, or unimplemented, just abort.
if (!shapeHelper || !shapeHelper->isImplemented())
if (!shapeHelper)
Copy link
Collaborator

Choose a reason for hiding this comment

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

Could you explain what's wrong with calling shapeHelper->isImplemented()?

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

There is nothing wrong with it, but if the shape helper is not implemented we still need to delete it or we will leak it's memory. This happens a few lines down

Copy link
Collaborator

Choose a reason for hiding this comment

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

I got it! That's why we have to distinguish the two conditions. Thanks for the explanation!

return false;

// Compute shape.
if (failed(shapeHelper->computeShape())) {
if (!shapeHelper->isImplemented() || failed(shapeHelper->computeShape())) {
delete shapeHelper;
return false;
}
Expand Down Expand Up @@ -961,12 +961,14 @@ void DimAnalysis::visitDim(
bool outputHasOneDynamicDim =
(llvm::count(outputType.getShape(), ShapedType::kDynamic) == 1);
// Check if the products of static sizes in the data and output are equal.
// It's ok to count ShapedType::kDynamic (dynamic dimension) in the size.
int64_t dataStaticSize = 1, outputStaticSize = 1;
for (int64_t i = 0; i < dataType.getRank(); ++i)
dataStaticSize *= dataType.getShape()[i];
for (int64_t i = 0; i < outputType.getRank(); ++i)
outputStaticSize *= outputType.getShape()[i];
for (int64_t i = 0; i < dataType.getRank(); ++i) {
dataStaticSize *= dataType.isDynamicDim(i) ? -1 : dataType.getShape()[i];
}
for (int64_t i = 0; i < outputType.getRank(); ++i) {
outputStaticSize *=
outputType.isDynamicDim(i) ? -1 : outputType.getShape()[i];
Copy link
Collaborator

Choose a reason for hiding this comment

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

Thanks for fixing this!

}
// Conditions hold, the dynamic dimension can be from the data.
if (dataHasOneDynamicDim && outputHasOneDynamicDim &&
(dataStaticSize == outputStaticSize)) {
Expand Down
34 changes: 17 additions & 17 deletions test/mlir/onnx/onnx_dim_analysis.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,38 +184,38 @@ func.func @test_matmul_batchsize(%arg0: tensor<?x8x16x16xf32>) -> tensor<?x8x16x

// -----

func.func @test_matmul_batchsize_diff_rank(%arg0: tensor<8x?x16x4xf32>) -> tensor<8x?x16x32xf32> {
func.func @test_matmul_batchsize_diff_rank(%arg0: tensor<8x?x16x4xf32>) -> tensor<8x?x16x128xf32> {
%shape = onnx.Constant dense<[-1, 4, 128]> : tensor<3xi64>
%0 = "onnx.Reshape"(%arg0, %shape) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x32xf32>
%1 = "onnx.MatMul"(%arg0, %0) : (tensor<8x?x16x4xf32>, tensor<?x4x32xf32>) -> tensor<8x?x16x32xf32>
"onnx.Return"(%1) : (tensor<8x?x16x32xf32>) -> ()
%0 = "onnx.Reshape"(%arg0, %shape) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x128xf32>
Copy link
Collaborator

Choose a reason for hiding this comment

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

Thanks for fixing this!

%1 = "onnx.MatMul"(%arg0, %0) : (tensor<8x?x16x4xf32>, tensor<?x4x128xf32>) -> tensor<8x?x16x128xf32>
"onnx.Return"(%1) : (tensor<8x?x16x128xf32>) -> ()

// CHECK-LABEL: func.func @test_matmul_batchsize_diff_rank
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<8x?x16x4xf32>) -> tensor<8x?x16x32xf32> {
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<8x?x16x4xf32>) -> tensor<8x?x16x128xf32> {
// CHECK-DAG: "onnx.DimGroup"([[PARAM_0_]]) {axis = 1 : si64, group_id = 0 : si64} : (tensor<8x?x16x4xf32>) -> ()
// CHECK-DAG: [[VAR_0_:%.+]] = onnx.Constant dense<[-1, 4, 128]> : tensor<3xi64>
// CHECK: [[VAR_1_:%.+]] = "onnx.Reshape"([[PARAM_0_]], [[VAR_0_]]) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x32xf32>
// CHECK: "onnx.DimGroup"([[VAR_1_]]) {axis = 0 : si64, group_id = 0 : si64} : (tensor<?x4x32xf32>) -> ()
// CHECK: [[VAR_2_:%.+]] = "onnx.MatMul"([[PARAM_0_]], [[VAR_1_]]) : (tensor<8x?x16x4xf32>, tensor<?x4x32xf32>) -> tensor<8x?x16x32xf32>
// CHECK: "onnx.DimGroup"([[VAR_2_]]) {axis = 1 : si64, group_id = 0 : si64} : (tensor<8x?x16x32xf32>) -> ()
// CHECK: onnx.Return [[VAR_2_]] : tensor<8x?x16x32xf32>
// CHECK: [[VAR_1_:%.+]] = "onnx.Reshape"([[PARAM_0_]], [[VAR_0_]]) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x128xf32>
// CHECK: "onnx.DimGroup"([[VAR_1_]]) {axis = 0 : si64, group_id = 0 : si64} : (tensor<?x4x128xf32>) -> ()
// CHECK: [[VAR_2_:%.+]] = "onnx.MatMul"([[PARAM_0_]], [[VAR_1_]]) : (tensor<8x?x16x4xf32>, tensor<?x4x128xf32>) -> tensor<8x?x16x128xf32>
// CHECK: "onnx.DimGroup"([[VAR_2_]]) {axis = 1 : si64, group_id = 0 : si64} : (tensor<8x?x16x128xf32>) -> ()
// CHECK: onnx.Return [[VAR_2_]] : tensor<8x?x16x128xf32>
// CHECK: }
}

// -----

func.func @test_reshape_single_dyn_dim(%arg0: tensor<8x?x16x4xf32>) -> tensor<?x4x32xf32> {
func.func @test_reshape_single_dyn_dim(%arg0: tensor<8x?x16x4xf32>) -> tensor<?x4x128xf32> {
%shape = onnx.Constant dense<[-1, 4, 128]> : tensor<3xi64>
%0 = "onnx.Reshape"(%arg0, %shape) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x32xf32>
"onnx.Return"(%0) : (tensor<?x4x32xf32>) -> ()
%0 = "onnx.Reshape"(%arg0, %shape) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x128xf32>
"onnx.Return"(%0) : (tensor<?x4x128xf32>) -> ()

// CHECK-LABEL: func.func @test_reshape_single_dyn_dim
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<8x?x16x4xf32>) -> tensor<?x4x32xf32> {
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<8x?x16x4xf32>) -> tensor<?x4x128xf32> {
// CHECK-DAG: "onnx.DimGroup"([[PARAM_0_]]) {axis = 1 : si64, group_id = 0 : si64} : (tensor<8x?x16x4xf32>) -> ()
// CHECK-DAG: [[VAR_0_:%.+]] = onnx.Constant dense<[-1, 4, 128]> : tensor<3xi64>
// CHECK: [[VAR_1_:%.+]] = "onnx.Reshape"([[PARAM_0_]], [[VAR_0_]]) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x32xf32>
// CHECK: "onnx.DimGroup"([[VAR_1_]]) {axis = 0 : si64, group_id = 0 : si64} : (tensor<?x4x32xf32>) -> ()
// CHECK: onnx.Return [[VAR_1_]] : tensor<?x4x32xf32>
// CHECK: [[VAR_1_:%.+]] = "onnx.Reshape"([[PARAM_0_]], [[VAR_0_]]) {allowzero = 0 : si64} : (tensor<8x?x16x4xf32>, tensor<3xi64>) -> tensor<?x4x128xf32>
// CHECK: "onnx.DimGroup"([[VAR_1_]]) {axis = 0 : si64, group_id = 0 : si64} : (tensor<?x4x128xf32>) -> ()
// CHECK: onnx.Return [[VAR_1_]] : tensor<?x4x128xf32>
// CHECK: }
}

Expand Down