Skip to content

[DML EP] Restrict MatMul to 4D #26583

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
wants to merge 7 commits into
base: main
Choose a base branch
from
Draft

[DML EP] Restrict MatMul to 4D #26583

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
cef17ad
Restrict DML EP MatMul to 4D
fdwr Nov 15, 2025
3b73d5a
Update comments
fdwr Nov 15, 2025
6d8eef2
Merge branch 'main' into user/dwayner/DmlEpMatMulRestrictTo4D
fdwr Nov 19, 2025
adb2569
Fix EinSum failure. Support more MatMul broadcasting via AreStridesCo…
fdwr Nov 19, 2025
63f1311
Address MatMul 1D broadcasting to enable more test cases
fdwr Nov 20, 2025
26e432b
Fold output rank of MatMulIntegerToFloat and QLinearMatMul
fdwr Nov 20, 2025
6a4634a
Merge branch 'main' into user/dwayner/DmlEpMatMulRestrictTo4D
fdwr Nov 21, 2025
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
2 changes: 1 addition & 1 deletion onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorEinSum.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -368,7 +368,7 @@ class DmlOperatorEinSum : public DmlOperator, public EinSumHelper
}
}
tensorDesc.SetDimensionsAndStrides(newSizes, newStrides);
tensorDesc.EnsureDimensionCount(1, TensorAxis::RightAligned);
tensorDesc.EnsureMinimumDimensionCount(1, TensorAxis::RightAligned);
}

// Reproject a tensor to the given axis arrangement.
Expand Down
20 changes: 16 additions & 4 deletions onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorMatMul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,16 +22,28 @@ class DmlOperatorMatMul : public DmlOperator
std::vector<DimensionType> inputShape0 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(0);
std::vector<DimensionType> inputShape1 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(1);
std::vector<DimensionType> outputShape = kernelInfo.GetTensorShapeDescription().GetOutputTensorShape(0);
std::vector<DimensionType> inputShape0Broadcasted;
std::vector<DimensionType> inputShape1Broadcasted;

OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape);
OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape, inputShape0Broadcasted, inputShape1Broadcasted);

// Initialize the input descriptions with broadcasting
m_inputTensorDescs[0] = CreateTensorDescFromInput(kernelInfo, 0, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape0);
m_inputTensorDescs[1] = CreateTensorDescFromInput(kernelInfo, 1, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape1);
// Initialize the input descriptions without broadcasting yet, since MatMul has special rules where broadcasting the
// original shape (notably when 1D) to the output shape would mess up because the dimensions are shifted.
m_inputTensorDescs[0] = CreateTensorDescFromInput(kernelInfo, 0);
m_inputTensorDescs[1] = CreateTensorDescFromInput(kernelInfo, 1);

// Initialize the output description while overriding the shape
m_outputTensorDescs[0] = CreateTensorDescFromOutput(kernelInfo, 0, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, outputShape);

// Broadcast the inputs to their broadcasted shapes.
m_inputTensorDescs[0].SetBroadcastedShape(inputShape0Broadcasted, inputShape0, outputShape.size());
m_inputTensorDescs[1].SetBroadcastedShape(inputShape1Broadcasted, inputShape1, outputShape.size());

// DirectML only supports ranks up to 4D for GEMM, and so any leading dimensions must be folded.
m_inputTensorDescs[0].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);
m_inputTensorDescs[1].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);
m_outputTensorDescs[0].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

std::vector<DML_TENSOR_DESC> inputDescs = GetDmlInputDescs();
std::vector<DML_TENSOR_DESC> outputDescs = GetDmlOutputDescs();

Expand Down
17 changes: 13 additions & 4 deletions ...untime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorMatMulInteger.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,12 +25,20 @@ class DmlOperatorMatMulInteger : public DmlOperator
std::vector<DimensionType> inputShape0 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(0);
std::vector<DimensionType> inputShape1 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(1);
std::vector<DimensionType> outputShape = kernelInfo.GetTensorShapeDescription().GetOutputTensorShape(0);
std::vector<DimensionType> inputShape0Broadcasted;
std::vector<DimensionType> inputShape1Broadcasted;

OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape);
OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape, inputShape0Broadcasted, inputShape1Broadcasted);

// Initialize the input descriptions with broadcasting
m_inputTensorDescs[IN_A] = CreateTensorDescFromInput(kernelInfo, 0/*OnnxIndex*/, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape0);
m_inputTensorDescs[IN_B] = CreateTensorDescFromInput(kernelInfo, 1/*OnnxIndex*/, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape1);
// Initialize the input descriptions without broadcasting
m_inputTensorDescs[IN_A] = CreateTensorDescFromInput(kernelInfo, 0/*OnnxIndex*/);
m_inputTensorDescs[IN_B] = CreateTensorDescFromInput(kernelInfo, 1/*OnnxIndex*/);

// Broadcast the inputs to their broadcasted shapes.
m_inputTensorDescs[IN_A].SetBroadcastedShape(inputShape0Broadcasted, inputShape0, outputShape.size());
m_inputTensorDescs[IN_B].SetBroadcastedShape(inputShape1Broadcasted, inputShape1, outputShape.size());
m_inputTensorDescs[IN_A].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);
m_inputTensorDescs[IN_B].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

uint32_t dmlDimSize = m_inputTensorDescs[0].GetDimensionCount();
// Resize the A ZeroPoint to be the same dimension as the input tensor.
Expand All @@ -49,6 +57,7 @@ class DmlOperatorMatMulInteger : public DmlOperator

// Initialize the output description while overriding the shape
m_outputTensorDescs[0] = CreateTensorDescFromOutput(kernelInfo, 0, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, outputShape);
m_outputTensorDescs[0].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

std::vector<DML_TENSOR_DESC> inputDescs = GetDmlInputDescs();
std::vector<DML_TENSOR_DESC> outputDescs = GetDmlOutputDescs();
Expand Down
34 changes: 26 additions & 8 deletions ...core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorMatMulIntegerToFloat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,20 +42,37 @@ class DmlOperatorMatMulIntegerToFloat : public DmlOperator
std::vector<DimensionType> inputShape0 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(OrtInputTensors::ortA);
std::vector<DimensionType> inputShape1 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(OrtInputTensors::ortB);
std::vector<DimensionType> outputShape = kernelInfo.GetTensorShapeDescription().GetOutputTensorShape(0);
std::vector<DimensionType> inputShape0Broadcasted;
std::vector<DimensionType> inputShape1Broadcasted;

OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape);
OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape, inputShape0Broadcasted, inputShape1Broadcasted);

// Initialize the input descriptions with broadcasting
m_inputTensorDescs[DmlInputIndex::dmlA] = CreateTensorDescFromInput(kernelInfo, OrtInputTensors::ortA, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape0);
m_inputTensorDescs[DmlInputIndex::dmlB] = CreateTensorDescFromInput(kernelInfo, OrtInputTensors::ortB, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape1);
// Initialize the input descriptions without broadcasting
m_inputTensorDescs[DmlInputIndex::dmlA] = CreateTensorDescFromInput(kernelInfo, OrtInputTensors::ortA);
m_inputTensorDescs[DmlInputIndex::dmlB] = CreateTensorDescFromInput(kernelInfo, OrtInputTensors::ortB);

// Broadcast the inputs to their broadcasted shapes.
m_inputTensorDescs[DmlInputIndex::dmlA].SetBroadcastedShape(inputShape0Broadcasted, inputShape0, outputShape.size());
m_inputTensorDescs[DmlInputIndex::dmlB].SetBroadcastedShape(inputShape1Broadcasted, inputShape1, outputShape.size());
m_inputTensorDescs[DmlInputIndex::dmlA].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);
m_inputTensorDescs[DmlInputIndex::dmlB].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

uint32_t dmlDimSize = m_inputTensorDescs[DmlInputIndex::dmlA].GetDimensionCount();

// Broadcast Bias tensor to the shape of the output tensor.
if(kernelInfo.IsInputValid(OrtInputTensors::ortBias)) {
m_inputTensorDescs[DmlInputIndex::dmlBias] = CreateTensorDescFromInput(kernelInfo, OrtInputTensors::ortBias, TensorAxis::DoNotCoerce,
TensorAxis::W, TensorAxis::RightAligned, outputShape);
if (kernelInfo.IsInputValid(OrtInputTensors::ortBias))
{
m_inputTensorDescs[DmlInputIndex::dmlBias] = CreateTensorDescFromInput(
kernelInfo,
OrtInputTensors::ortBias,
TensorAxis::DoNotCoerce,
TensorAxis::W,
TensorAxis::RightAligned,
outputShape,
dmlDimSize
);
}

uint32_t dmlDimSize = m_inputTensorDescs[DmlInputIndex::dmlA].GetDimensionCount();
// Resize the A Scale to be the same dimension as the input tensor.
// The 1D tensor needs to be moved to the H channel.
m_inputTensorDescs[DmlInputIndex::dmlAScale] = CreateTensorDescFromInput(
Expand Down Expand Up @@ -87,6 +104,7 @@ class DmlOperatorMatMulIntegerToFloat : public DmlOperator

// Initialize the output description while overriding the shape
m_outputTensorDescs[0] = CreateTensorDescFromOutput(kernelInfo, 0, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, outputShape);
m_outputTensorDescs[0].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

std::vector<DML_TENSOR_DESC> inputDescs = GetDmlInputDescs();
std::vector<DML_TENSOR_DESC> outputDescs = GetDmlOutputDescs();
Expand Down
17 changes: 13 additions & 4 deletions ...untime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorQLinearMatMul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,12 +29,20 @@ class DmlOperatorQLinearMatMul : public DmlOperator
std::vector<DimensionType> inputShape0 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(0/*A OnnxIndex*/);
std::vector<DimensionType> inputShape1 = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(3/*B OnnxIndex*/);
std::vector<DimensionType> outputShape = kernelInfo.GetTensorShapeDescription().GetOutputTensorShape(0);
std::vector<DimensionType> inputShape0Broadcasted;
std::vector<DimensionType> inputShape1Broadcasted;

OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape);
OperatorHelper::MatMulShapeMapping(inputShape0, inputShape1, outputShape, inputShape0Broadcasted, inputShape1Broadcasted);

// Initialize the input descriptions with broadcasting
m_inputTensorDescs[IN_A] = CreateTensorDescFromInput(kernelInfo, 0/*A OnnxIndex*/, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape0);
m_inputTensorDescs[IN_B] = CreateTensorDescFromInput(kernelInfo, 3/*B OnnxIndex*/, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, inputShape1);
// Initialize the input descriptions without broadcasting
m_inputTensorDescs[IN_A] = CreateTensorDescFromInput(kernelInfo, 0/*A OnnxIndex*/);
m_inputTensorDescs[IN_B] = CreateTensorDescFromInput(kernelInfo, 3/*B OnnxIndex*/);

// Broadcast the inputs to their broadcasted shapes.
m_inputTensorDescs[IN_A].SetBroadcastedShape(inputShape0Broadcasted, inputShape0, outputShape.size());
m_inputTensorDescs[IN_B].SetBroadcastedShape(inputShape1Broadcasted, inputShape1, outputShape.size());
m_inputTensorDescs[IN_A].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);
m_inputTensorDescs[IN_B].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

uint32_t dmlDimSize = m_inputTensorDescs[0].GetDimensionCount();
// Resize the A Scale to be the same dimension as the input tensor.
Expand Down Expand Up @@ -89,6 +97,7 @@ class DmlOperatorQLinearMatMul : public DmlOperator

// Initialize the output description while overriding the shape
m_outputTensorDescs[0] = CreateTensorDescFromOutput(kernelInfo, 0, TensorAxis::DoNotCoerce, TensorAxis::W, TensorAxis::RightAligned, outputShape);
m_outputTensorDescs[0].SetDimensionCount(4, TensorAxis::RightAligned, /*foldEndDimensions*/ true);

std::vector<DML_TENSOR_DESC> inputDescs = GetDmlInputDescs();
std::vector<DML_TENSOR_DESC> outputDescs = GetDmlOutputDescs();
Expand Down
Loading
Loading