James Conroy | cc34093 | 2020-05-12 18:08:52 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <armnn/Descriptors.hpp> |
| 9 | #include <armnn/LstmParams.hpp> |
| 10 | #include <backendsCommon/Workload.hpp> |
| 11 | #include <backendsCommon/WorkloadData.hpp> |
| 12 | |
| 13 | #include "arm_compute/graph/Tensor.h" |
| 14 | #include "arm_compute/runtime/NEON/functions/NEQLSTMLayer.h" |
| 15 | |
| 16 | namespace armnn |
| 17 | { |
| 18 | |
| 19 | class NeonQLstmWorkload : public BaseWorkload<QLstmQueueDescriptor> |
| 20 | { |
| 21 | public: |
| 22 | NeonQLstmWorkload(const QLstmQueueDescriptor& descriptor, const WorkloadInfo& info); |
| 23 | virtual void Execute() const override; |
| 24 | |
| 25 | private: |
| 26 | mutable arm_compute::NEQLSTMLayer m_QLstmLayer; |
| 27 | |
| 28 | std::unique_ptr<arm_compute::Tensor> m_InputToInputWeightsTensor; |
| 29 | std::unique_ptr<arm_compute::Tensor> m_InputToForgetWeightsTensor; |
| 30 | std::unique_ptr<arm_compute::Tensor> m_InputToCellWeightsTensor; |
| 31 | std::unique_ptr<arm_compute::Tensor> m_InputToOutputWeightsTensor; |
| 32 | |
| 33 | std::unique_ptr<arm_compute::Tensor> m_RecurrentToInputWeightsTensor; |
| 34 | std::unique_ptr<arm_compute::Tensor> m_RecurrentToForgetWeightsTensor; |
| 35 | std::unique_ptr<arm_compute::Tensor> m_RecurrentToCellWeightsTensor; |
| 36 | std::unique_ptr<arm_compute::Tensor> m_RecurrentToOutputWeightsTensor; |
| 37 | |
| 38 | std::unique_ptr<arm_compute::Tensor> m_CellToInputWeightsTensor; |
| 39 | std::unique_ptr<arm_compute::Tensor> m_CellToForgetWeightsTensor; |
| 40 | std::unique_ptr<arm_compute::Tensor> m_CellToOutputWeightsTensor; |
| 41 | |
| 42 | std::unique_ptr<arm_compute::Tensor> m_InputGateBiasTensor; |
| 43 | std::unique_ptr<arm_compute::Tensor> m_ForgetGateBiasTensor; |
| 44 | std::unique_ptr<arm_compute::Tensor> m_CellBiasTensor; |
| 45 | std::unique_ptr<arm_compute::Tensor> m_OutputGateBiasTensor; |
| 46 | |
| 47 | std::unique_ptr<arm_compute::Tensor> m_ProjectionWeightsTensor; |
| 48 | std::unique_ptr<arm_compute::Tensor> m_ProjectionBiasTensor; |
| 49 | |
| 50 | std::unique_ptr<arm_compute::Tensor> m_InputLayerNormWeightsTensor; |
| 51 | std::unique_ptr<arm_compute::Tensor> m_ForgetLayerNormWeightsTensor; |
| 52 | std::unique_ptr<arm_compute::Tensor> m_CellLayerNormWeightsTensor; |
| 53 | std::unique_ptr<arm_compute::Tensor> m_OutputLayerNormWeightsTensor; |
| 54 | |
| 55 | void FreeUnusedTensors(); |
| 56 | }; |
| 57 | |
| 58 | arm_compute::Status NeonQLstmWorkloadValidate(const TensorInfo& input, |
| 59 | const TensorInfo& cellStateIn, |
| 60 | const TensorInfo& outputStateIn, |
| 61 | const TensorInfo& cellStateOut, |
| 62 | const TensorInfo& outputStateOut, |
| 63 | const TensorInfo& output, |
| 64 | const QLstmDescriptor& descriptor, |
| 65 | const LstmInputParamsInfo& paramsInfo); |
| 66 | |
| 67 | } //namespace armnn |