From b0baff73b1574a198e57d46fcd704cedc43cea16 Mon Sep 17 00:00:00 2001 From: Cathal Corbett Date: Fri, 28 Jan 2022 12:17:19 +0000 Subject: IVGCVSW-6268 Add support of Unidirectional Sequence Lstm fp32/fp16 to Neon !ComputeLibrary:7150 Signed-off-by: Cathal Corbett Change-Id: I3de48ffc8d08c95a22705e2b68d069791bddae73 --- ...NeonUnidirectionalSequenceLstmFloatWorkload.hpp | 92 ++++++++++++++++++++++ 1 file changed, 92 insertions(+) create mode 100644 src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp (limited to 'src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp') diff --git a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp new file mode 100644 index 0000000000..10c2ecbd19 --- /dev/null +++ b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp @@ -0,0 +1,92 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include +#include +#include +#include + +#include "arm_compute/graph/Tensor.h" +#include "arm_compute/runtime/NEON/functions/NELSTMLayer.h" +#include "arm_compute/runtime/NEON/functions/NEPermute.h" +#include "arm_compute/runtime/NEON/functions/NESplit.h" +#include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h" + +namespace armnn +{ + +class NeonUnidirectionalSequenceLstmFloatWorkload : public FloatWorkload +{ +public: + NeonUnidirectionalSequenceLstmFloatWorkload(const UnidirectionalSequenceLstmQueueDescriptor& descriptor, + const WorkloadInfo& info); + virtual void Execute() const override; + +private: + + // + // ACL layers required to fully form a Unidirectional Sequence LSTM layer. + // + mutable std::unique_ptr m_Permute1; + mutable std::unique_ptr m_Splitter; + mutable std::vector> m_Layers; + mutable std::unique_ptr m_Concat; + mutable std::unique_ptr m_Permute2; + + // + // ACL LSTM arm_compute::Tensors. + // + std::unique_ptr m_InputToInputWeightsTensor; + std::unique_ptr m_InputToForgetWeightsTensor; + std::unique_ptr m_InputToCellWeightsTensor; + std::unique_ptr m_InputToOutputWeightsTensor; + std::unique_ptr m_RecurrentToInputWeightsTensor; + std::unique_ptr m_RecurrentToForgetWeightsTensor; + std::unique_ptr m_RecurrentToCellWeightsTensor; + std::unique_ptr m_RecurrentToOutputWeightsTensor; + std::unique_ptr m_CellToInputWeightsTensor; + std::unique_ptr m_CellToForgetWeightsTensor; + std::unique_ptr m_CellToOutputWeightsTensor; + std::unique_ptr m_InputGateBiasTensor; + std::unique_ptr m_ForgetGateBiasTensor; + std::unique_ptr m_CellBiasTensor; + std::unique_ptr m_OutputGateBiasTensor; + std::unique_ptr m_ProjectionWeightsTensor; + std::unique_ptr m_ProjectionBiasTensor; + + std::unique_ptr m_ScratchBuffer; + + std::unique_ptr m_InputLayerNormWeightsTensor; + std::unique_ptr m_ForgetLayerNormWeightsTensor; + std::unique_ptr m_CellLayerNormWeightsTensor; + std::unique_ptr m_OutputLayerNormWeightsTensor; + + // + // Additional ACL arm_compute::Tensors and std::vector. + // Required to perform splitting, concatenation and permutations. + // + arm_compute::Tensor m_PermuteFirstOut; + std::vector m_SplitterOutputsTensors; + std::vector m_ConcatInputsTensors; + std::vector m_SplitterOutputs; + std::vector m_ConcatInputs; + arm_compute::Tensor concat_out; + + void FreeUnusedTensors(); +}; + +arm_compute::Status +NeonUnidirectionalSequenceLstmFloatWorkloadValidate(const TensorInfo& input, + const TensorInfo& outputStateIn, + const TensorInfo& cellStateIn, + const TensorInfo& output, + const Optional& hiddenStateOutput, + const Optional& cellStateOutput, + const UnidirectionalSequenceLstmDescriptor& descriptor, + const LstmInputParamsInfo& paramsInfo); + +} //namespace armnn -- cgit v1.2.1