IVGCVSW-7109: Add Batch MatMul front end support - Reference
* Descriptors added for BatchMatMul
* Layer definition added
* Input validation added (will likely change when opt. param support comes in)
* Ref workload implementation for BatchMatMul added (will also change with opt. param support)
* Ref layer tests made for BatchMatMul
* CMake and other build files updated
Signed-off-by: Samuel Yap <samuel.yap@arm.com>
Change-Id: Ic885301da543ee0fbe7922b85e7f9658c4efc617
diff --git a/src/armnn/BackendHelper.cpp b/src/armnn/BackendHelper.cpp
index 5b5bece..6638709 100644
--- a/src/armnn/BackendHelper.cpp
+++ b/src/armnn/BackendHelper.cpp
@@ -179,6 +179,22 @@
reasonIfUnsupported);
}
+bool LayerSupportHandle::IsBatchMatMulSupported(const TensorInfo& input0,
+ const TensorInfo& input1,
+ const TensorInfo& output,
+ const BatchMatMulDescriptor& descriptor,
+ Optional<std::string&> reasonIfUnsupported)
+{
+ TensorInfos infos{input0, input1, output};
+
+ return m_LayerSupport->IsLayerSupported(LayerType::BatchMatMul,
+ infos,
+ descriptor,
+ EmptyOptional(),
+ EmptyOptional(),
+ reasonIfUnsupported);
+}
+
bool LayerSupportHandle::IsBatchNormalizationSupported(const TensorInfo& input,
const TensorInfo& output,
const TensorInfo& mean,
diff --git a/src/armnn/Descriptors.cpp b/src/armnn/Descriptors.cpp
index c740fd0..f957627 100644
--- a/src/armnn/Descriptors.cpp
+++ b/src/armnn/Descriptors.cpp
@@ -455,4 +455,86 @@
return armnn::GetNumInputs(m_BiasEnabled);
}
+std::pair<std::pair<unsigned int, unsigned int>, std::pair<unsigned int, unsigned int>>
+BatchMatMulDescriptor::GetAxesToMul(
+ const BatchMatMulDescriptor& desc,
+ const TensorShape& tensorXShape,
+ const TensorShape& tensorYShape)
+{
+ // May refactor to just work on one input per call - makes it less confusing and also
+ // allows more flexibility (i.e. in Layer output shape inference)
+
+ auto xNumDims = tensorXShape.GetNumDimensions();
+ auto yNumDims = tensorYShape.GetNumDimensions();
+
+ std::pair<unsigned int, unsigned int> xAxes = { xNumDims-2, xNumDims-1 };
+ std::pair<unsigned int, unsigned int> yAxes = { yNumDims-2, yNumDims-1 };
+
+ if(desc.m_DataLayoutX.has_value())
+ {
+ switch(desc.m_DataLayoutX.value())
+ {
+ case DataLayout::NDHWC:
+ case DataLayout::NHWC:
+ xAxes.first -= 1;
+ xAxes.second -= 1;
+ break;
+ case DataLayout::NCDHW:
+ case DataLayout::NCHW:
+ default:
+ break;
+ }
+ }
+
+ if(desc.m_DataLayoutY.has_value())
+ {
+ switch(desc.m_DataLayoutY.value())
+ {
+ case DataLayout::NDHWC:
+ case DataLayout::NHWC:
+ yAxes.first -= 1;
+ yAxes.second -= 1;
+ break;
+ case DataLayout::NCDHW:
+ case DataLayout::NCHW:
+ default:
+ break;
+ }
+ }
+
+ return { xAxes, yAxes};
+}
+
+std::pair<std::vector<unsigned int>, std::vector<unsigned int>> BatchMatMulDescriptor::GetAxesNotMul(
+ const BatchMatMulDescriptor& desc,
+ const TensorShape& inputXShape,
+ const TensorShape& inputYShape)
+{
+ // May refactor to just work on one input per call - makes it less confusing and also
+ // allows more flexibility (i.e. in Layer output shape inference)
+ auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(desc, inputXShape, inputYShape);
+
+ std::vector<unsigned int> axesXNotMul;
+ std::vector<unsigned int> axesYNotMul;
+
+ for(unsigned int i = 0; i < inputXShape.GetNumDimensions(); i++)
+ {
+ if(i == axesToMul.first.first || i == axesToMul.first.second)
+ {
+ continue;
+ }
+ axesXNotMul.push_back(i);
+ }
+ for(unsigned int i = 0; i < inputYShape.GetNumDimensions(); i++)
+ {
+ if(i == axesToMul.second.first || i == axesToMul.second.second)
+ {
+ continue;
+ }
+ axesYNotMul.push_back(i);
+ }
+
+ return { axesXNotMul, axesYNotMul };
+}
+
}
diff --git a/src/armnn/ILayerSupport.cpp b/src/armnn/ILayerSupport.cpp
index 5366b13..8099782 100644
--- a/src/armnn/ILayerSupport.cpp
+++ b/src/armnn/ILayerSupport.cpp
@@ -13,7 +13,7 @@
{
ARMNN_NO_DEPRECATE_WARN_BEGIN
-// IsLayerSupport() forwards to the deprecated virtual methods depending on input LayerType.
+// IsLayerSupported() forwards to the deprecated virtual methods depending on input LayerType.
// Allows backends continue to behave as before maintaining backward compatibility.
bool ILayerSupport::IsLayerSupported(const LayerType& type,
const std::vector<TensorInfo>& infos,
diff --git a/src/armnn/LayersFwd.hpp b/src/armnn/LayersFwd.hpp
index dcfb91b..acac1f9 100644
--- a/src/armnn/LayersFwd.hpp
+++ b/src/armnn/LayersFwd.hpp
@@ -1,5 +1,5 @@
//
-// Copyright © 2017 Arm Ltd. All rights reserved.
+// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
@@ -9,6 +9,7 @@
#include "layers/ActivationLayer.hpp"
#include "layers/AdditionLayer.hpp"
#include "layers/ArgMinMaxLayer.hpp"
+#include "layers/BatchMatMulLayer.hpp"
#include "layers/BatchNormalizationLayer.hpp"
#include "layers/BatchToSpaceNdLayer.hpp"
#include "layers/CastLayer.hpp"
@@ -110,6 +111,7 @@
DECLARE_LAYER(Activation)
DECLARE_LAYER(Addition)
DECLARE_LAYER(ArgMinMax)
+DECLARE_LAYER(BatchMatMul)
DECLARE_LAYER(BatchNormalization)
DECLARE_LAYER(BatchToSpaceNd)
DECLARE_LAYER(Cast)
diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp
index 5d44306..ef9f4e7 100644
--- a/src/armnn/Network.cpp
+++ b/src/armnn/Network.cpp
@@ -456,6 +456,12 @@
return pNetworkImpl->AddChannelShuffleLayer(descriptor, name);
}
+IConnectableLayer* INetwork::AddBatchMatMulLayer(const BatchMatMulDescriptor &descriptor,
+ const char* name)
+{
+ return pNetworkImpl->AddBatchMatMulLayer(descriptor, name);
+}
+
void INetwork::ExecuteStrategy(IStrategy& strategy) const
{
return pNetworkImpl->ExecuteStrategy(strategy);
@@ -2876,6 +2882,11 @@
return layer;
}
+IConnectableLayer* NetworkImpl::AddBatchMatMulLayer(const BatchMatMulDescriptor& desc, const char* name)
+{
+ return m_Graph->AddLayer<BatchMatMulLayer>(desc, name);
+}
+
IConnectableLayer* NetworkImpl::AddPrecompiledLayer(const PreCompiledDescriptor& preCompiledDescriptor,
CompiledBlobPtr compiledBlobPtr,
const Optional<BackendId>& backend,
diff --git a/src/armnn/Network.hpp b/src/armnn/Network.hpp
index a4387e6..19a0286 100644
--- a/src/armnn/Network.hpp
+++ b/src/armnn/Network.hpp
@@ -49,6 +49,9 @@
IConnectableLayer* AddArgMinMaxLayer(const ArgMinMaxDescriptor& desc,
const char* name = nullptr);
+ IConnectableLayer* AddBatchMatMulLayer(const BatchMatMulDescriptor& desc,
+ const char* name = nullptr);
+
IConnectableLayer* AddBatchNormalizationLayer(const BatchNormalizationDescriptor& desc,
const ConstTensor& mean,
const ConstTensor& variance,
diff --git a/src/armnn/layers/BatchMatMulLayer.cpp b/src/armnn/layers/BatchMatMulLayer.cpp
new file mode 100644
index 0000000..501de2d
--- /dev/null
+++ b/src/armnn/layers/BatchMatMulLayer.cpp
@@ -0,0 +1,97 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#include "BatchMatMulLayer.hpp"
+
+#include <armnn/backends/WorkloadFactory.hpp>
+#include "layers/LayerCloneBase.hpp"
+
+namespace armnn
+{
+
+BatchMatMulLayer::BatchMatMulLayer(const BatchMatMulDescriptor& param, const char* name)
+ : LayerWithParameters(2, 1, LayerType::BatchMatMul, param, name)
+{}
+
+std::unique_ptr<IWorkload> BatchMatMulLayer::CreateWorkload(const IWorkloadFactory& factory) const
+{
+ BatchMatMulQueueDescriptor descriptor;
+ SetAdditionalInfo(descriptor);
+
+ return factory.CreateWorkload(LayerType::BatchMatMul, descriptor, PrepInfoAndDesc(descriptor));
+}
+
+BatchMatMulLayer* BatchMatMulLayer::Clone(Graph& graph) const
+{
+ auto layer = CloneBase<BatchMatMulLayer>(graph, m_Param, GetName());
+
+ return std::move(layer);
+}
+
+std::vector<TensorShape> BatchMatMulLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const
+{
+ ARMNN_ASSERT(inputShapes.size() == 2);
+
+ TensorShape inputXShape = inputShapes[0];
+ TensorShape inputYShape = inputShapes[1];
+
+ // Note: Take into account what pre-adjoint or pre-transposing will do to the inferred output shape
+
+ TensorShape& longerInput = inputXShape.GetNumDimensions() >= inputYShape.GetNumDimensions()?
+ inputXShape:inputYShape;
+ TensorShape& shorterInput = inputXShape.GetNumDimensions() >= inputYShape.GetNumDimensions()?
+ inputYShape:inputXShape;
+
+ unsigned int inputNumDimsOffset = longerInput.GetNumDimensions() - shorterInput.GetNumDimensions();
+
+ unsigned int outputNumDimensions = longerInput.GetNumDimensions();
+
+ std::vector<unsigned int> tensorDimensions(outputNumDimensions, 0);
+
+ auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(m_Param, inputXShape, inputYShape);
+ const auto& longerAxesToMul = (axesToMul.first.first >= axesToMul.second.first &&
+ axesToMul.first.second >= axesToMul.second.second) ?
+ axesToMul.first : axesToMul.second;
+
+ for (unsigned int i = 0; i < outputNumDimensions; ++i)
+ {
+ if (i == longerAxesToMul.first)
+ {
+ tensorDimensions[i] = &shorterInput == &inputXShape ? inputXShape[i - inputNumDimsOffset] : inputXShape[i];
+ }
+ else if(i == longerAxesToMul.second)
+ {
+ tensorDimensions[i] = &shorterInput == &inputYShape ? inputYShape[i - inputNumDimsOffset] : inputYShape[i];
+ }
+ else // The other dimensions not to be multiplied (but may be broadcasted)
+ {
+ // Does NOT validate whether it's a valid broadcast - that's done in the validate func in WorkloadData.cpp
+ tensorDimensions[i] = static_cast<int>(i) - static_cast<int>(inputNumDimsOffset) < 0 ?
+ longerInput[i] :
+ std::max(longerInput[i], shorterInput[i - inputNumDimsOffset]);
+ }
+ }
+
+ auto outputShape = TensorShape(outputNumDimensions, tensorDimensions.data());
+ return std::vector<TensorShape>({ outputShape });
+}
+
+void BatchMatMulLayer::ValidateTensorShapesFromInputs()
+{
+ VerifyLayerConnections(2, CHECK_LOCATION());
+
+ const TensorShape& outputShape = GetOutputSlot(0).GetTensorInfo().GetShape();
+
+ VerifyShapeInferenceType(outputShape, m_ShapeInferenceMethod);
+
+ auto inferredShapes = InferOutputShapes({
+ GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(),
+ GetInputSlot(1).GetConnection()->GetTensorInfo().GetShape() });
+
+ ARMNN_ASSERT(inferredShapes.size() == 1);
+
+ ValidateAndCopyShape(outputShape, inferredShapes[0], m_ShapeInferenceMethod, "BatchMatMulLayer");
+}
+
+} // namespace armnn
\ No newline at end of file
diff --git a/src/armnn/layers/BatchMatMulLayer.hpp b/src/armnn/layers/BatchMatMulLayer.hpp
new file mode 100644
index 0000000..8dc79d3
--- /dev/null
+++ b/src/armnn/layers/BatchMatMulLayer.hpp
@@ -0,0 +1,46 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "LayerWithParameters.hpp"
+
+namespace armnn
+{
+
+class BatchMatMulLayer : public LayerWithParameters<BatchMatMulDescriptor>
+{
+public:
+ /// Makes a workload for the BatchMatMul type.
+ /// @param [in] graph The graph where this layer can be found.
+ /// @param [in] factory The workload factory which will create the workload.
+ /// @return A pointer to the created workload, or nullptr if not created.
+ virtual std::unique_ptr<IWorkload> CreateWorkload(const IWorkloadFactory &factory) const override;
+
+ /// Creates a dynamically-allocated copy of this layer.
+ /// @param [in] graph The graph into which this layer is being cloned.
+ BatchMatMulLayer* Clone(Graph &graph) const override;
+
+ /// Infers the output shape from the given input shapes.
+ /// @param [in] inputShapes The vector of input shapes for BatchMatMul.
+ /// @return A vector of inferred output shape.
+ std::vector<TensorShape> InferOutputShapes(const std::vector<TensorShape>& inputShapes) const override;
+
+ /// Check if the input tensor shapes
+ /// will lead to a valid configuration of @ref BatchMatMulLayer.
+ /// @param [in] shapeInferenceMethod Indicates if output shape shall be overwritten or just validated.
+ void ValidateTensorShapesFromInputs() override;
+
+protected:
+ /// Constructor to create a BatchMatMulLayer.
+ /// @param [in] param BatchMatMulDescriptor to configure optional parameters for batch matrix multiplication
+ /// @param [in] name Optional name for the layer
+ BatchMatMulLayer(const BatchMatMulDescriptor& param, const char* name);
+
+ /// Default destructor
+ ~BatchMatMulLayer() = default;
+};
+
+}
\ No newline at end of file