saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-29 20:14:01 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[DML EP] Force layer norm inputs to be 4D to better target metacommands (#14022)

Browse source 
### Description
Force layer norm inputs to be 4D to better target metacommands

### Motivation and Context
This may improve performance on some hardware by allowing the driver to
return valid layouts to DML when querying for metacommand support.
This commit is contained in:
Patrice Vignola 2023-01-03 12:46:33 -08:00 committed by GitHub
parent 587e891cae
commit 589612106a
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 25 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorLayerNormalization.cpp
View file

@ -25,12 +25,33 @@ public:
std::nullopt,
kernelCreationContext.GetTensorShapeDescription().GetInputTensorDimensionCount(0));
const float epsilon = kernelCreationContext.GetOptionalAttribute<float>(AttrName::Epsilon, DefaultEpsilon);
constexpr static uint32_t minimumDimensionCount = 4;
// Pad the input and the output with trailing 1's until they are at least 4D
for (uint32_t i = 0; i < kernelCreationContext.GetInputCount(); ++i)
{
auto sizes = m_inputTensorDescs[i].GetSizes();
std::vector<uint32_t> tensorShape(sizes.begin(), sizes.end());
tensorShape.resize(std::max<size_t>(tensorShape.size(), minimumDimensionCount), 1);
if (m_inputTensorDescs[i].GetDmlDataType() != DML_TENSOR_TYPE_INVALID)
{
m_inputTensorDescs[i] = TensorDesc(
m_inputTensorDescs[i].GetDmlDataType(),
tensorShape);
}
}
m_outputTensorDescs[0] = TensorDesc(
m_outputTensorDescs[0].GetDmlDataType(),
m_inputTensorDescs[0].GetSizes());
const float epsilon = kernelCreationContext.GetOptionalAttribute<float>(AttrName::Epsilon, DefaultEpsilon);
int32_t onnxAxis = kernelCreationContext.GetOptionalAttribute<int32_t>(AttrName::Axis, -1);
uint32_t inputDimCount = kernelCreationContext.GetTensorShapeDescription().GetInputTensorDimensionCount(0);
onnxAxis = OperatorHelper::HandleNegativeAxis(onnxAxis, inputDimCount);
std::vector<uint32_t> onnxAxes(static_cast<size_t>(inputDimCount) - static_cast<size_t>(onnxAxis));
uint32_t onnxDimCount = kernelCreationContext.GetTensorShapeDescription().GetInputTensorDimensionCount(0);
uint32_t dmlDimCount = m_inputTensorDescs[0].GetDimensionCount();
onnxAxis = OperatorHelper::HandleNegativeAxis(onnxAxis, onnxDimCount);
std::vector<uint32_t> onnxAxes(static_cast<size_t>(dmlDimCount) - static_cast<size_t>(onnxAxis));
std::iota(onnxAxes.begin(), onnxAxes.end(), onnxAxis);
assert(m_inputTensorDescs.size() == 3);