diff options
author | James Conroy <james.conroy@arm.com> | 2020-03-20 08:49:33 +0000 |
---|---|---|
committer | James Conroy <james.conroy@arm.com> | 2020-03-20 14:53:44 +0000 |
commit | 586a9aac99312eb9cb304cbbd18cec46b9158e23 (patch) | |
tree | 6d620eae6dcfb920ac04eae43424548dc602a1eb /src/armnn/layers/QLstmLayer.hpp | |
parent | c94d3f7107b84b586791aa096f8641e6efa18c90 (diff) | |
download | armnn-586a9aac99312eb9cb304cbbd18cec46b9158e23.tar.gz |
IVGCVSW-4549 Add front end for new QLSTM layer
* Added new layer QLstm (Android R HAL 1.3)
* Made necessary updates to APIs
* Added unit tests
* This layer is functionally equivalent to the
original unquantized LSTM layer with some
additonal quantization features added. Due
to this, original LstmParams are used for
this layer.
Signed-off-by: James Conroy <james.conroy@arm.com>
Change-Id: I5b7f2d2fb6e17e81573b41a31bc55f49ae79608f
Diffstat (limited to 'src/armnn/layers/QLstmLayer.hpp')
-rw-r--r-- | src/armnn/layers/QLstmLayer.hpp | 124 |
1 files changed, 124 insertions, 0 deletions
diff --git a/src/armnn/layers/QLstmLayer.hpp b/src/armnn/layers/QLstmLayer.hpp new file mode 100644 index 0000000000..2d40b7e29e --- /dev/null +++ b/src/armnn/layers/QLstmLayer.hpp @@ -0,0 +1,124 @@ +// +// Copyright © 2020 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "LayerWithParameters.hpp" + +namespace armnn +{ + +class ScopedCpuTensorHandle; + +struct QLstmBasicParameters +{ + /// A unique pointer to represent 2D weights tensor with dimensions [num_units, inputSize] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_InputToForgetWeights; + /// A unique pointer to represent 2D weights tensor with dimensions [num_units, inputSize] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_InputToCellWeights; + /// A unique pointer to represent 2D weights tensor with dimensions [num_units, inputSize] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_InputToOutputWeights; + + /// A unique pointer to represent 2D weights tensor with dimensions [num_units, outputSize] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_RecurrentToForgetWeights; + /// A unique pointer to represent 2D weights tensor with dimensions [num_units, outputSize] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_RecurrentToCellWeights; + /// A unique pointer to represent 2D weights tensor with dimensions [num_units, outputSize] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_RecurrentToOutputWeights; + + /// A unique pointer to represent 1D bias tensor with dimensions [num_units] (int32). + std::unique_ptr<ScopedCpuTensorHandle> m_ForgetGateBias; + /// A unique pointer to represent 1D bias tensor with dimensions [num_units] (int32). + std::unique_ptr<ScopedCpuTensorHandle> m_CellBias; + /// A unique pointer to represent 1D bias tensor with dimensions [num_units] (int32). + std::unique_ptr<ScopedCpuTensorHandle> m_OutputGateBias; +}; + +struct QLstmOptProjectionParameters +{ + /// A unique pointer to represent 2D weights tensor with dimensions [output_size, num_units] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_ProjectionWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [output_size] (int32). + std::unique_ptr<ScopedCpuTensorHandle> m_ProjectionBias; +}; + +struct QLstmOptPeepholeParameters +{ + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_CellToInputWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_CellToForgetWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_CellToOutputWeights; +}; + +struct QLstmOptCifgParameters +{ + /// A unique pointer to represent 2D weights tensor with dimensions [input_size, num_units] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_InputToInputWeights; + /// A unique pointer to represent 2D weights tensor with dimensions [input_size, num_units] (QSymmS8). + std::unique_ptr<ScopedCpuTensorHandle> m_RecurrentToInputWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (int32). + std::unique_ptr<ScopedCpuTensorHandle> m_InputGateBias; +}; + +struct QLstmOptLayerNormParameters +{ + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_InputLayerNormWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_ForgetLayerNormWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_CellLayerNormWeights; + /// A unique pointer to represent 1D weights tensor with dimensions [num_units] (QSymmS16). + std::unique_ptr<ScopedCpuTensorHandle> m_OutputLayerNormWeights; +}; + +/// This layer represents a QLstm operation. +class QLstmLayer : public LayerWithParameters<QLstmDescriptor> +{ +public: + + QLstmBasicParameters m_BasicParameters; + QLstmOptCifgParameters m_CifgParameters; + QLstmOptProjectionParameters m_ProjectionParameters; + QLstmOptPeepholeParameters m_PeepholeParameters; + QLstmOptLayerNormParameters m_LayerNormParameters; + + /// Makes a workload for the QLstm 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. + QLstmLayer* Clone(Graph& graph) const override; + + /// Check if the input tensor shape(s) + /// will lead to a valid configuration of @ref QLstmLayer. + void ValidateTensorShapesFromInputs() override; + + /// By default returns inputShapes if the number of inputs are equal to number of outputs, + /// otherwise infers the output shapes from given input shapes and layer properties. + /// @param [in] inputShapes The input shapes layer has. + /// @return A vector to the inferred output shape. + std::vector<TensorShape> InferOutputShapes(const std::vector<TensorShape>& inputShapes) const override; + + void Accept(ILayerVisitor& visitor) const override; + +protected: + /// Constructor to create a QLstmLayer. + /// @param [in] name Optional name for the layer. + QLstmLayer(const QLstmDescriptor& param, const char* name); + + /// Default destructor + ~QLstmLayer() = default; + + /// Retrieve the handles to the constant values stored by the layer. + /// @return A vector of the constant tensors stored by this layer. + Layer::ConstantTensors GetConstantTensorsByRef() override; +}; + +} // namespace armnn |