Skip to content

[WebGPU] DequantizeLinear op fixes #27706

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
edgchen1 merged 10 commits into from
Mar 20, 2026
Merged

[WebGPU] DequantizeLinear op fixes #27706

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
e18f42c
fix existing dequantizelinear.json test failures
edgchen1 Mar 16, 2026
fa1e534
fix zp handling for non packed block quantization
edgchen1 Mar 16, 2026
a0df1c3
add test, rename other test
edgchen1 Mar 16, 2026
9d2a6ce
update test names
edgchen1 Mar 16, 2026
bd61484
move test around
edgchen1 Mar 17, 2026
8d3e7e9
remove unneeded u32 cast
edgchen1 Mar 17, 2026
b312d1a
add comments
edgchen1 Mar 17, 2026
10462e0
Merge remote-tracking branch 'origin/main' into edgchen1/webgpu_fix_t…
edgchen1 Mar 18, 2026
23a130f
try removing check for no zero point if int32 input from jsep webgpu …
edgchen1 Mar 19, 2026
f9fcee8
add more int32 with zp tests
edgchen1 Mar 19, 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
3 changes: 0 additions & 3 deletions js/web/lib/wasm/jsep/webgpu/ops/quantize-linear.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,9 +31,6 @@ const validateInputs = (inputs: readonly TensorView[], attributes: DequantizeLin
if (inputs.length === 3 && inputs[0].dataType !== inputs[2].dataType) {
throw new Error('x and x-zero-point must have the same data type.');
}
if (inputs[0].dataType === DataType.int32 && inputs.length > 2) {
throw new Error('In the case of dequantizing int32 there is no zero point.');
}
if (inputs[1].dims.length !== 0 && inputs[1].dims.length !== 1 && inputs[1].dims.length !== inputs[0].dims.length) {
throw new Error('scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.');
}
Expand Down
183 changes: 173 additions & 10 deletions js/web/test/data/ops/dequantizelinear.jsonc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
"attributes": [],
"cases": [
{
"name": "T[1]",
"name": "uint8 per-tensor with zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
Expand Down Expand Up @@ -41,7 +41,7 @@
"attributes": [],
"cases": [
{
"name": "T[2]",
"name": "int32 per-tensor no zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
Expand All @@ -64,6 +64,41 @@
}
]
},
{
"name": "dequantizelinear",
"operator": "DequantizeLinear",
"opset": { "domain": "", "version": 10 },
"attributes": [],
"cases": [
{
"name": "int32 per-tensor with zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
"dims": [4],
"type": "int32"
},
{
"data": [0.1],
"dims": [1],
"type": "float32"
},
{
"data": [1],
"dims": [1],
"type": "int32"
}
],
"outputs": [
{
"data": [0.0, 0.1, 0.2, 0.3],
"dims": [4],
"type": "float32"
}
]
}
]
},
{
"name": "dequantizelinear",
"operator": "DequantizeLinear",
Expand All @@ -77,7 +112,7 @@
],
"cases": [
{
"name": "T[3]",
"name": "uint8 2D per-axis scalar scale with zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
Expand Down Expand Up @@ -118,7 +153,7 @@
],
"cases": [
{
"name": "T[4]",
"name": "int32 2D per-axis scalar scale no zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
Expand Down Expand Up @@ -154,7 +189,48 @@
],
"cases": [
{
"name": "T[5]",
"name": "int32 2D per-axis scalar scale with zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
"dims": [2, 2],
"type": "int32"
},
{
"data": [0.1],
"dims": [1],
"type": "float32"
},
{
"data": [1],
"dims": [1],
"type": "int32"
}
],
"outputs": [
{
"data": [0.0, 0.1, 0.2, 0.3],
"dims": [2, 2],
"type": "float32"
}
]
}
]
},
{
"name": "dequantizelinear",
"operator": "DequantizeLinear",
"opset": { "domain": "", "version": 13 },
"attributes": [
{
"name": "axis",
"data": 1,
"type": "int"
}
],
"cases": [
{
"name": "uint8 3D per-axis uniform scale with zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
Expand Down Expand Up @@ -195,7 +271,7 @@
],
"cases": [
{
"name": "T[6]",
"name": "uint8 3D per-axis varying scale with zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
Expand Down Expand Up @@ -236,7 +312,7 @@
],
"cases": [
{
"name": "T[7]",
"name": "int32 3D per-axis scalar scale no zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
Expand All @@ -259,6 +335,47 @@
}
]
},
{
"name": "dequantizelinear",
"operator": "DequantizeLinear",
"opset": { "domain": "", "version": 13 },
"attributes": [
{
"name": "axis",
"data": 1,
"type": "int"
}
],
"cases": [
{
"name": "int32 3D per-axis scalar scale with zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
"dims": [2, 2, 2],
"type": "int32"
},
{
"data": [0.1],
"dims": [1],
"type": "float32"
},
{
"data": [1],
"dims": [1],
"type": "int32"
}
],
"outputs": [
{
"data": [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7],
"dims": [2, 2, 2],
"type": "float32"
}
]
}
]
},
{
"name": "dequantizelinear",
"operator": "DequantizeLinear",
Expand All @@ -277,7 +394,7 @@
],
"cases": [
{
"name": "T[8]",
"name": "uint8 3D blocked with zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
Expand Down Expand Up @@ -323,7 +440,7 @@
],
"cases": [
{
"name": "T[9]",
"name": "int32 3D blocked no zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
Expand All @@ -346,6 +463,52 @@
}
]
},
{
"name": "dequantizelinear block dequantization",
"operator": "DequantizeLinear",
"opset": { "domain": "", "version": 21 },
"attributes": [
{
"name": "axis",
"data": 1,
"type": "int"
},
{
"name": "block_size",
"data": 2,
"type": "int"
}
],
"cases": [
{
"name": "int32 3D blocked with zero point",
"inputs": [
{
"data": [1, 2, 3, 4, 5, 6, 7, 8],
"dims": [2, 2, 2],
"type": "int32"
},
{
"data": [0.1, 0.2, 0.3, 0.4],
"dims": [2, 1, 2],
"type": "float32"
},
{
"data": [0, 1, 0, 1],
"dims": [2, 1, 2],
"type": "int32"
}
],
"outputs": [
{
"data": [0.1, 0.2, 0.3, 0.6, 1.5, 2.0, 2.1, 2.8],
"dims": [2, 2, 2],
"type": "float32"
}
]
}
]
},
{
"name": "dequantizelinear",
"operator": "DequantizeLinear",
Expand All @@ -359,7 +522,7 @@
],
"cases": [
{
"name": "T[3]",
"name": "uint8 2D per-axis scalar scale no zero point",
"inputs": [
{
"data": [1, 2, 3, 4],
Expand Down
15 changes: 9 additions & 6 deletions onnxruntime/core/providers/webgpu/quantization/quantize_linear.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ Status DequantizeLinearProgram::GenerateShaderCode(ShaderHelper& shader) const {
if (packed_) {
shader.MainFunctionBody()
<< "let zero_point_index = " << output.IndicesGet("output_indices", "uniforms.axis") << ";\n"
<< "let zero_point_input = " << zero_point.GetByOffset("u32(zero_point_index / 4)") << ";\n"
<< "let zero_point_input = " << zero_point.GetByOffset("zero_point_index / 4") << ";\n"
<< "let zero_point_vec = " << unpack << ";\n"
<< "let zero_point_value = zero_point_vec[zero_point_index % 4];\n";
} else {
Expand All @@ -88,16 +88,17 @@ Status DequantizeLinearProgram::GenerateShaderCode(ShaderHelper& shader) const {
<< "let zero_point_value = " << zero_point.GetByOffset("zero_point_index") << ";\n";
}
} else {
// BlockedQuantization. The zero-point input shape is same as the input shape except along axis.
// BlockedQuantization. The zero-point input shape is the same as the scale input shape.
if (packed_) {
shader.MainFunctionBody()
<< "let zero_point_offset = " << scale.GetByIndices("scale_indices") << ";\n"
<< "let zero_point_input = " << zero_point.GetByOffset("u32(zero_point_offset / 4)") << ";\n"
<< "let zero_point_offset = " << scale.IndicesToOffset("scale_indices") << ";\n"
<< "let zero_point_input = " << zero_point.GetByOffset("zero_point_offset / 4") << ";\n"
<< "let zero_point_vec = " << unpack << ";\n"
<< "let zero_point_value = zero_point_vec[zero_point_offset % 4];\n";
Comment thread
edgchen1 marked this conversation as resolved.
} else {
shader.MainFunctionBody()
<< "let zero_point_value = " << zero_point.GetByIndices("scale_indices") << ";\n";
<< "let zero_point_offset = " << scale.IndicesToOffset("scale_indices") << ";\n"
<< "let zero_point_value = " << zero_point.GetByOffset("zero_point_offset") << ";\n";
}
}
} else {
Expand Down Expand Up @@ -145,7 +146,9 @@ Status DequantizeLinear::ComputeInternal(ComputeContext& context) const {
program
.AddInputs({{x, ProgramTensorMetadataDependency::TypeAndRank, ProgramInput::Flatten, packed ? 4 : input_component}})
.AddInputs({{x_scale, ProgramTensorMetadataDependency::TypeAndRank}})
.AddOutput({output_tensor, ProgramTensorMetadataDependency::Rank, components})
.AddOutput(use_components
? ProgramOutput{output_tensor, ProgramTensorMetadataDependency::Rank, ProgramOutput::Flatten, components}
: ProgramOutput{output_tensor, ProgramTensorMetadataDependency::Rank, components})
.SetDispatchGroupSize((x_size / components + WORKGROUP_SIZE - 1) / WORKGROUP_SIZE)
Comment thread
edgchen1 marked this conversation as resolved.
.AddUniformVariables({{static_cast<uint32_t>(axis)}})
.AddUniformVariables({{static_cast<uint32_t>(block_size_)}})
Expand Down
Loading