saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-07 04:39:07 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merged PR 4852260: DML EP remove redundant rank checks for higher dimension support

Browse source 
Remove redundant checks in the DML EP which should instead rely on DML's validation. At least one of these checks wrongly prevents legitimate execution (5D Concat is supported in DML, but the DML EP blocks it 🤦‍♀️). Note this is a small aspect of the larger work (not sufficient to make the models fully work) that I thought I'd flush now while I had the change ready anyway due to investigation.

Related work items: #23232293, #25707941
This commit is contained in:
Dwayne Robinson 2020-06-30 21:26:16 +00:00
parent ee48b89350
commit 7cb2c3f025
7 changed files with 11 additions and 16 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionProvider.cpp
View file

@ -376,8 +376,9 @@ namespace Dml
{
assert(!m_closed);
const size_t sourceSizeInBytes = ComputeByteSizeFromTensor(*src);
const size_t dataSizeInBytes = ComputeByteSizeFromTensor(*dst);
THROW_HR_IF(E_INVALIDARG, dataSizeInBytes != ComputeByteSizeFromTensor(*src)); // Tensors must be the same size
THROW_HR_IF(E_INVALIDARG, dataSizeInBytes != sourceSizeInBytes); // Tensors must be the same size
if (dataSizeInBytes == 0)
{

2
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorConcat.cpp
View file

@ -42,7 +42,7 @@ public:
for (size_t i = 0; i < m_inputTensorDescs.size(); i++)
{
// DML doesn't support empty tensors for concat, so we ignore them
if (!OperatorHelper::ContainsEmptyDimensions(m_inputTensorDescs[i].GetDmlSizes()))
if (!OperatorHelper::ContainsEmptyDimensions(m_inputTensorDescs[i].GetSizes()))
{
inputDescs.push_back(m_inputTensorDescs[i].GetDmlDesc());
}

8
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorExpand.cpp
View file

@ -25,8 +25,8 @@ public:
TensorDesc inputTensorDesc =
TensorDesc(
kernelCreationContext.GetInputEdgeDescription(0).tensorDataType,
m_outputTensorDescs[0].GetDmlSizes(),
m_inputTensorDescs[0].GetDmlSizes(),
m_outputTensorDescs[0].GetSizes(),
m_inputTensorDescs[0].GetSizes(),
TensorAxis::DoNotCoerce,
TensorAxis::W,
TensorAxis::RightAligned,
@ -36,8 +36,8 @@ public:
TensorDesc outputTensorDesc =
TensorDesc(
kernelCreationContext.GetOutputEdgeDescription(0).tensorDataType,
m_outputTensorDescs[0].GetDmlSizes(),
m_outputTensorDescs[0].GetDmlSizes(),
m_outputTensorDescs[0].GetSizes(),
m_outputTensorDescs[0].GetSizes(),
TensorAxis::DoNotCoerce,
TensorAxis::W,
TensorAxis::RightAligned,

1
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorGather.cpp
View file

@ -27,7 +27,6 @@ public:
auto outputTensorShapeDescription = kernelCreationContext.GetTensorShapeDescription();
std::vector<DimensionType> dataDimensions = outputTensorShapeDescription.GetInputTensorShape(0);
std::vector<DimensionType> indicesDimensions = outputTensorShapeDescription.GetInputTensorShape(1);
ML_CHECK_VALID_ARGUMENT(dataDimensions.size() <= OperatorHelper::NchwDimensionCount);
uint32_t dmlAxis = GetDmlAdjustedAxis(m_axis, kernelCreationContext, m_inputTensorDescs.front().GetDimensionCount());
DML_GATHER_OPERATOR_DESC operatorDesc = {};

3
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorReduce.cpp
View file

@ -24,9 +24,10 @@ public:
std::vector<uint32_t> dmlAxes;
std::vector<DimensionType> reducedDims = kernelInfo.GetTensorShapeDescription().GetInputTensorShape(0);
int dimOffset = gsl::narrow_cast<int>(OperatorHelper::NchwDimensionCount - reducedDims.size());
int dimOffset = gsl::narrow_cast<int>(m_inputTensorDescs[0].GetDimensionCount() - reducedDims.size());
for (auto& dim : m_axes)
{
assert(dim < reducedDims.size()); // ReduceHelperBase already validated this.
reducedDims[dim] = 1;
dmlAxes.push_back(static_cast<uint32_t>(dim + dimOffset));
}

1
onnxruntime/core/providers/dml/DmlExecutionProvider/src/TensorDesc.h
View file

@ -36,7 +36,6 @@ namespace Dml
inline DML_TENSOR_DATA_TYPE GetDmlDataType() const { return m_bufferTensorDesc.DataType; }
inline MLOperatorTensorDataType GetMlOperatorDataType() const { return m_mlOperatorTensorDataType; }
inline gsl::span<const uint32_t> GetDmlSizes() const { return m_sizes; }
void ForceUnsignedDataType();
void Remap64bitDmlDataTypeTo32bit();
bool WasRemapped64bitTo32bit() const;

9
onnxruntime/core/providers/dml/OperatorAuthorHelper/OperatorHelper.cpp
View file

@ -703,7 +703,7 @@ namespace OperatorHelper
ML_CHECK_VALID_ARGUMENT(inputDimensions.size() >= 1);
ML_CHECK_VALID_ARGUMENT(indicesDimensions.size() >= 0);
int outDimCount = gsl::narrow_cast<int>(inputDimensions.size() + indicesDimensions.size() - 1);
ML_CHECK_VALID_ARGUMENT(outDimCount >= 0 && outDimCount <= NchwDimensionCount);
ML_CHECK_VALID_ARGUMENT(outDimCount >= 0);
std::vector<DimensionType> outputDimensions(outDimCount, 1);
@ -746,7 +746,7 @@ namespace OperatorHelper
const uint32_t numberOfOutputDimensionsFromInput = static_cast<uint32_t>(inputDimensions.size()) - numberOfCoordinatesPerIndex;
const uint32_t numberOfOutputDimensionsFromIndices = static_cast<uint32_t>(indicesDimensions.size()) - 1; // Strip off last dimension.
uint32_t outputDimensionCount = gsl::narrow_cast<uint32_t>(numberOfOutputDimensionsFromIndices + numberOfOutputDimensionsFromInput);
ML_CHECK_VALID_ARGUMENT(outputDimensionCount > 0 && outputDimensionCount <= NchwDimensionCount);
ML_CHECK_VALID_ARGUMENT(outputDimensionCount > 0);
// Form the full expected size by concatenating the prefix part of the indices tensor shape
// with the suffix of the input tensor shape.
@ -812,8 +812,6 @@ namespace OperatorHelper
// Dim Offset : 1
std::vector<DimensionType> reducedDims = shapeInfo.GetInputTensorShape(0);
ML_CHECK_VALID_ARGUMENT(reducedDims.size() <= NchwDimensionCount);
std::vector<bool> reduced(reducedDims.size(), false);
for (auto& dim : m_axes)
@ -847,8 +845,6 @@ namespace OperatorHelper
void ReduceHelperBase::AdjustAxesAndOutputShape(const std::vector<uint32_t>& inputShape)
{
ML_CHECK_VALID_ARGUMENT(inputShape.size() <= NchwDimensionCount);
// If axes is not specified, reduce over all the dimensions
if (m_axes.empty())
{
@ -1001,7 +997,6 @@ namespace OperatorHelper
std::vector<EdgeShapes> ConcatHelper::GetOutputShapes(const MLShapeInferenceContext& shapeInfo) const
{
auto outputShape = shapeInfo.GetInputTensorShape(0);
ML_CHECK_VALID_ARGUMENT(outputShape.size() <= NcdhwDimensionCount);
uint32_t inputCount = shapeInfo.GetInputCount();