runtime/onert/backend/cpu/ops/LSTMLayer.h

   1 /*
   2  * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  *      http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 #ifndef __ONERT_BACKEND_CPU_OPS_LSTMLAYER_H__
  18 #define __ONERT_BACKEND_CPU_OPS_LSTMLAYER_H__
  19
  20 #include <backend/IPortableTensor.h>
  21 #include "OperationUtils.h"
  22 #include <ir/InternalType.h>
  23 #include <ir/operation/LSTM.h>
  24 #include <exec/IFunction.h>
  25
  26 namespace nnfw
  27 {
  28 namespace cker
  29 {
  30 class FCTempArena;
  31 }
  32 } // namespace nnfw
  33
  34 namespace onert
  35 {
  36 namespace backend
  37 {
  38 namespace cpu
  39 {
  40 namespace ops
  41 {
  42
  43 // TODO Support LSTM, BiDirectionalSequenceLSTM
  44 class LSTMLayer : public ::onert::exec::IFunction
  45 {
  46 public:
  47   LSTMLayer() = default;
  48
  49 public:
  50   void LSTMFloat();
  51
  52   void configure(const IPortableTensor *input, const IPortableTensor *input_to_input_weights,
  53                  const IPortableTensor *input_to_forget_weights,
  54                  const IPortableTensor *input_to_cell_weights,
  55                  const IPortableTensor *input_to_output_weights,
  56                  const IPortableTensor *recurrent_to_input_weights,
  57                  const IPortableTensor *recurrent_to_forget_weights,
  58                  const IPortableTensor *recurrent_to_cell_weights,
  59                  const IPortableTensor *recurrent_to_output_weights,
  60                  const IPortableTensor *cell_to_input_weights,
  61                  const IPortableTensor *cell_to_forget_weights,
  62                  const IPortableTensor *cell_to_output_weights,
  63                  const IPortableTensor *input_layer_norm_weights,
  64                  const IPortableTensor *forget_layer_norm_weights,
  65                  const IPortableTensor *cell_layer_norm_weights,
  66                  const IPortableTensor *output_layer_norm_weights, const IPortableTensor *aux_input,
  67                  const IPortableTensor *aux_input_to_input_weights,
  68                  const IPortableTensor *aux_input_to_forget_weights,
  69                  const IPortableTensor *aux_input_to_cell_weights,
  70                  const IPortableTensor *aux_input_to_output_weights,
  71                  const IPortableTensor *input_gate_bias, const IPortableTensor *forget_gate_bias,
  72                  const IPortableTensor *cell_gate_bias, const IPortableTensor *output_gate_bias,
  73                  const IPortableTensor *projection_weights, const IPortableTensor *projection_bias,
  74                  const IPortableTensor *output_state_in, const IPortableTensor *cell_state_in,
  75                  const ir::operation::LSTM::Param &params, bool forward_sequence, bool time_major,
  76                  int32_t output_offset, IPortableTensor *scratch_buffer,
  77                  IPortableTensor *output_state, IPortableTensor *cell_state,
  78                  IPortableTensor *output, bool has_output_state_data, bool has_cell_state_data);
  79
  80   void run() override;
  81
  82 private:
  83   const IPortableTensor *_input{nullptr};
  84   const IPortableTensor *_input_to_input_weights{nullptr};
  85   const IPortableTensor *_input_to_forget_weights{nullptr};
  86   const IPortableTensor *_input_to_cell_weights{nullptr};
  87   const IPortableTensor *_input_to_output_weights{nullptr};
  88   const IPortableTensor *_recurrent_to_input_weights{nullptr};
  89   const IPortableTensor *_recurrent_to_forget_weights{nullptr};
  90   const IPortableTensor *_recurrent_to_cell_weights{nullptr};
  91   const IPortableTensor *_recurrent_to_output_weights{nullptr};
  92   const IPortableTensor *_cell_to_input_weights{nullptr};
  93   const IPortableTensor *_cell_to_forget_weights{nullptr};
  94   const IPortableTensor *_cell_to_output_weights{nullptr};
  95   const IPortableTensor *_input_layer_norm_coefficients{nullptr};
  96   const IPortableTensor *_forget_layer_norm_coefficients{nullptr};
  97   const IPortableTensor *_cell_layer_norm_coefficients{nullptr};
  98   const IPortableTensor *_output_layer_norm_coefficients{nullptr};
  99   const IPortableTensor *_aux_input{nullptr};
 100   const IPortableTensor *_aux_input_to_input_weights{nullptr};
 101   const IPortableTensor *_aux_input_to_forget_weights{nullptr};
 102   const IPortableTensor *_aux_input_to_cell_weights{nullptr};
 103   const IPortableTensor *_aux_input_to_output_weights{nullptr};
 104   const IPortableTensor *_input_gate_bias{nullptr};
 105   const IPortableTensor *_forget_gate_bias{nullptr};
 106   const IPortableTensor *_cell_gate_bias{nullptr};
 107   const IPortableTensor *_output_gate_bias{nullptr};
 108   const IPortableTensor *_projection_weights{nullptr};
 109   const IPortableTensor *_projection_bias{nullptr};
 110   const IPortableTensor *_output_state_in{nullptr};
 111   const IPortableTensor *_cell_state_in{nullptr};
 112   IPortableTensor *_scratch_buffer{nullptr};
 113   IPortableTensor *_output_state{nullptr};
 114   IPortableTensor *_cell_state{nullptr};
 115   IPortableTensor *_output{nullptr};
 116   std::vector<uint8_t> _scratch_vec{};
 117   std::vector<uint8_t> _output_state_vec{};
 118   std::vector<uint8_t> _cell_state_vec{};
 119   ir::operation::LSTM::Param _params{};
 120   bool _forward_sequence{true};
 121   bool _time_major{true};
 122   int32_t _output_offset{0};
 123   bool _has_output_state_data{false};
 124   bool _has_cell_state_data{false};
 125 };
 126
 127 } // namespace ops
 128 } // namespace cpu
 129 } // namespace backend
 130 } // namespace onert
 131
 132 #endif // __ONERT_BACKEND_CPU_OPS_LSTMLAYER_H__