1/*
2 * Copyright (C) 2017 The Android Open Source Project
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 FRAMEWORKS_ML_NN_LSTMCELL_H
18#define FRAMEWORKS_ML_NN_LSTMCELL_H
19
20#include "ActivationFunctor.h"
21
22#include <algorithm>
23#include <cmath>
24
25namespace android {
26namespace hardware {
27namespace neuralnetworks {
28namespace V1_0 {
29struct Operation;
30}
31}  // namespace neuralnetworks
32}  // namespace hardware
33}  // namespace android
34
35namespace android {
36namespace nn {
37
38struct LSTMParams {
39  ActivationFn activation_;
40  float cell_clip_;
41  float proj_clip_;
42};
43
44struct RunTimeOperandInfo;
45struct Shape;
46
47class LSTMCell {
48 public:
49  LSTMCell(const android::hardware::neuralnetworks::V1_0::Operation &operation,
50           std::vector<RunTimeOperandInfo> &operands);
51
52  static bool Prepare(const android::hardware::neuralnetworks::V1_0::Operation &operation,
53                      std::vector<RunTimeOperandInfo> &operands,
54                      Shape *scratchShape,
55                      Shape *outputStateShape,
56                      Shape *cellStateShape,
57                      Shape *outputShape);
58  bool Eval();
59
60  // Input Tensors of size {n_batch, n_input}
61  static constexpr int kInputTensor = 0;
62
63  // Input weight tensors of size: {n_cell, n_input}
64  static constexpr int kInputToInputWeightsTensor = 1;  // Optional
65  static constexpr int kInputToForgetWeightsTensor = 2;
66  static constexpr int kInputToCellWeightsTensor = 3;
67  static constexpr int kInputToOutputWeightsTensor = 4;
68
69  // Recurrent weight tensors of size {n_cell, n_output}
70  static constexpr int kRecurrentToInputWeightsTensor = 5;  // Optional
71  static constexpr int kRecurrentToForgetWeightsTensor = 6;
72  static constexpr int kRecurrentToCellWeightsTensor = 7;
73  static constexpr int kRecurrentToOutputWeightsTensor = 8;
74
75  // Peephole weights tensors of size {n_cell}, representing a diagonal matrix.
76  static constexpr int kCellToInputWeightsTensor = 9;    // Optional
77  static constexpr int kCellToForgetWeightsTensor = 10;  // Optional
78  static constexpr int kCellToOutputWeightsTensor = 11;  // Optional
79
80  // Gates bias tensors of size {n_cell}
81  static constexpr int kInputGateBiasTensor = 12;  // Optional
82  static constexpr int kForgetGateBiasTensor = 13;
83  static constexpr int kCellGateBiasTensor = 14;
84  static constexpr int kOutputGateBiasTensor = 15;
85
86  // Projection weight tensor of size {n_output, n_cell}
87  static constexpr int kProjectionWeightsTensor = 16;  // Optional
88  // Projection bias tensor of size {n_output}
89  static constexpr int kProjectionBiasTensor = 17;  // Optional
90
91  static constexpr int kOutputStateInTensor = 18;
92  static constexpr int kCellStateInTensor = 19;
93
94  static constexpr int kActivationParam = 20;
95  static constexpr int kCellClipParam = 21;
96  static constexpr int kProjClipParam = 22;
97
98  // Output tensors.
99  static constexpr int kScratchBufferTensor = 0;
100  static constexpr int kOutputStateOutTensor = 1;
101  static constexpr int kCellStateOutTensor = 2;
102  static constexpr int kOutputTensor = 3;
103
104 private:
105  static bool CheckInputTensorDimensions(
106      const android::hardware::neuralnetworks::V1_0::Operation &operation,
107      std::vector<RunTimeOperandInfo> &operands, uint32_t n_input,
108      uint32_t n_output, uint32_t n_cell);
109  LSTMParams params_;
110
111  const RunTimeOperandInfo *input_;
112
113  const RunTimeOperandInfo *input_to_input_weights_;
114  const RunTimeOperandInfo *input_to_forget_weights_;
115  const RunTimeOperandInfo *input_to_cell_weights_;
116  const RunTimeOperandInfo *input_to_output_weights_;
117
118  const RunTimeOperandInfo *recurrent_to_input_weights_;
119  const RunTimeOperandInfo *recurrent_to_forget_weights_;
120  const RunTimeOperandInfo *recurrent_to_cell_weights_;
121  const RunTimeOperandInfo *recurrent_to_output_weights_;
122
123  const RunTimeOperandInfo *cell_to_input_weights_;
124  const RunTimeOperandInfo *cell_to_forget_weights_;
125  const RunTimeOperandInfo *cell_to_output_weights_;
126
127  const RunTimeOperandInfo *input_gate_bias_;
128  const RunTimeOperandInfo *forget_gate_bias_;
129  const RunTimeOperandInfo *cell_bias_;
130  const RunTimeOperandInfo *output_gate_bias_;
131
132  const RunTimeOperandInfo *projection_weights_;
133  const RunTimeOperandInfo *projection_bias_;
134
135  const RunTimeOperandInfo *output_state_in_;
136  const RunTimeOperandInfo *cell_state_in_;
137
138  RunTimeOperandInfo *output_state_out_;
139  RunTimeOperandInfo *cell_state_out_;
140  RunTimeOperandInfo *output_;
141
142  RunTimeOperandInfo *scratch_buffer_;
143};
144
145}  // namespace nn
146}  // namespace android
147
148#endif  // FRAMEWORKS_ML_NN_LSTMCELL_H
149