Skip to content

DequantizeLinear: infer Y’s shape from X #749

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

Draft
SamNour wants to merge 3 commits into
base: feature/onnx-to-tosa
Choose a base branch
from
Draft

DequantizeLinear: infer Y’s shape from X #749

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d5db607
Store inferred output dims; skip refineDims so Y is not overwritten by
May 8, 2026
9ad8491
chore: formatting
May 8, 2026
6607ae3
[test] update test to preserve dynamic dimensions
May 8, 2026
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
5 changes: 3 additions & 2 deletions src/Dialect/ONNX/ONNXOps/Quantize/DequantizeLinear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,8 +79,9 @@ LogicalResult ONNXDequantizeLinearOpShapeHelper::computeShape() {
}

// Get values.
// Save the final result.
setOutputDims(outputDims);
// Store inferred output dims; skip refineDims so Y is not overwritten by
// stale static types on Y.
setOutputDims(outputDims, /*n=*/0, /*refineShape=*/false);
return success();
}

Expand Down
24 changes: 6 additions & 18 deletions test/mlir/conversion/onnx_to_tosa/NN/DequantizeLinear.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,15 +117,9 @@ func.func @dynamic(%arg0 : tensor<?xi8>, %arg1 : tensor<f32>, %arg2 : tensor<i8>
}

// CHECK-LABEL: func.func @dynamic
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<?xi8>, [[PARAM_1_:%.+]]: tensor<f32>, [[PARAM_2_:%.+]]: tensor<i8>) -> tensor<1xf32> {
// CHECK-DAG: [[VAR_0_:%.+]] = tosa.cast [[PARAM_0_]] : (tensor<?xi8>) -> tensor<?xf32>
// CHECK-DAG: [[VAR_1_:%.+]] = tosa.reshape [[PARAM_2_]] {new_shape = array<i64: 1>} : (tensor<i8>) -> tensor<1xi8>
// CHECK: [[VAR_2_:%.+]] = tosa.cast [[VAR_1_]] : (tensor<1xi8>) -> tensor<1xf32>
// CHECK-DAG: [[VAR_3_:%.+]] = tosa.sub [[VAR_0_]], [[VAR_2_]] : (tensor<?xf32>, tensor<1xf32>) -> tensor<?xf32>
// CHECK-DAG: [[VAR_4_:%.+]] = tosa.reshape [[PARAM_1_]] {new_shape = array<i64: 1>} : (tensor<f32>) -> tensor<1xf32>
// CHECK-DAG: [[VAR_5_:%.+]] = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
// CHECK: [[VAR_6_:%.+]] = tosa.mul [[VAR_3_]], [[VAR_4_]], [[VAR_5_]] : (tensor<?xf32>, tensor<1xf32>, tensor<1xi8>) -> tensor<1xf32>
// CHECK: return [[VAR_6_]] : tensor<1xf32>
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<?xi8>, [[PARAM_1_:%.+]]: tensor<f32>, [[PARAM_2_:%.+]]: tensor<i8>) -> tensor<?xf32> {
// CHECK: [[VAR_0_:%.+]] = "onnx.DequantizeLinear"([[PARAM_0_]], [[PARAM_1_]], [[PARAM_2_]]) {{.*}} : (tensor<?xi8>, tensor<f32>, tensor<i8>) -> tensor<?xf32>
// CHECK: return [[VAR_0_]] : tensor<?xf32>
// CHECK: }
// -----

Expand Down Expand Up @@ -153,15 +147,9 @@ func.func @dynamic3(%arg0 : tensor<2x?xi8>, %arg1 : tensor<f32>, %arg2 : tensor<
}

// CHECK-LABEL: func.func @dynamic3
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<2x?xi8>, [[PARAM_1_:%.+]]: tensor<f32>, [[PARAM_2_:%.+]]: tensor<i8>) -> tensor<2x1xf32> {
// CHECK-DAG: [[VAR_0_:%.+]] = tosa.cast [[PARAM_0_]] : (tensor<2x?xi8>) -> tensor<2x?xf32>
// CHECK-DAG: [[VAR_1_:%.+]] = tosa.reshape [[PARAM_2_]] {new_shape = array<i64: 1, 1>} : (tensor<i8>) -> tensor<1x1xi8>
// CHECK: [[VAR_2_:%.+]] = tosa.cast [[VAR_1_]] : (tensor<1x1xi8>) -> tensor<1x1xf32>
// CHECK-DAG: [[VAR_3_:%.+]] = tosa.sub [[VAR_0_]], [[VAR_2_]] : (tensor<2x?xf32>, tensor<1x1xf32>) -> tensor<2x?xf32>
// CHECK-DAG: [[VAR_4_:%.+]] = tosa.reshape [[PARAM_1_]] {new_shape = array<i64: 1, 1>} : (tensor<f32>) -> tensor<1x1xf32>
// CHECK-DAG: [[VAR_5_:%.+]] = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
// CHECK: [[VAR_6_:%.+]] = tosa.mul [[VAR_3_]], [[VAR_4_]], [[VAR_5_]] : (tensor<2x?xf32>, tensor<1x1xf32>, tensor<1xi8>) -> tensor<2x1xf32>
// CHECK: return [[VAR_6_]] : tensor<2x1xf32>
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<2x?xi8>, [[PARAM_1_:%.+]]: tensor<f32>, [[PARAM_2_:%.+]]: tensor<i8>) -> tensor<2x?xf32> {
// CHECK: [[VAR_0_:%.+]] = "onnx.DequantizeLinear"([[PARAM_0_]], [[PARAM_1_]], [[PARAM_2_]]) {{.*}} : (tensor<2x?xi8>, tensor<f32>, tensor<i8>) -> tensor<2x?xf32>
// CHECK: return [[VAR_0_]] : tensor<2x?xf32>
// CHECK: }
// -----

Expand Down
12 changes: 12 additions & 0 deletions test/mlir/onnx/onnx_shape_inference.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2084,6 +2084,18 @@ func.func @test_dequantize_linear_2(%arg0 : tensor<5x?x3x4xi8>, %arg1 : tensor<*

// -----

// COM: inferShapes derives Y's shape from X.
func.func @test_dequantize_linear_stale_output_batch(%arg0 : tensor<2x4x8xi8>, %arg1 : tensor<f32>, %arg2 : tensor<i8>) -> tensor<1x4x8xf32> {
%1 = "onnx.DequantizeLinear"(%arg0, %arg1, %arg2) {} : (tensor<2x4x8xi8>, tensor<f32>, tensor<i8>) -> tensor<1x4x8xf32>
"onnx.Return"(%1) {} : (tensor<1x4x8xf32>) -> ()
Copy link
Copy Markdown

@jorickert jorickert May 8, 2026

Choose a reason for hiding this comment

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

is'n the model here already broken? Not sure if we should try to support that


// CHECK-LABEL: test_dequantize_linear_stale_output_batch
// CHECK: [[RES:%.+]] = "onnx.DequantizeLinear"(%arg0, %arg1, %arg2) {axis = 1 : si64, block_size = 0 : si64} : (tensor<2x4x8xi8>, tensor<f32>, tensor<i8>) -> tensor<2x4x8xf32>
// CHECK: onnx.Return [[RES]] : tensor<2x4x8xf32>
}

// -----

//===----------------------------------------------------------------------===//
/// Test shape inference for ConvInteger operation and all its attributes.
//===----------------------------------------------------------------------===//
Expand Down
Loading