Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 164a2727d3bbce0e575d24b7db787c85e2e2c203 / . / arm_compute / runtime / CL / functions / CLLSTMLayer.h
blob: 87fb1190b731fd092d896646373a96f98cfbda60 [file] [log] [blame]
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]1/*
2 * Copyright (c) 2018 ARM Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#ifndef __ARM_COMPUTE_CLLSTMLAYER_H__
25#define __ARM_COMPUTE_CLLSTMLAYER_H__
26
27#include "arm_compute/runtime/IFunction.h"
28
29#include "arm_compute/core/CL/kernels/CLActivationLayerKernel.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]30#include "arm_compute/core/CL/kernels/CLCopyKernel.h"
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame^]31#include "arm_compute/core/CL/kernels/CLElementwiseOperationKernel.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]32#include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h"
33#include "arm_compute/core/Types.h"
34#include "arm_compute/runtime/CL/CLMemoryGroup.h"
35#include "arm_compute/runtime/CL/CLTensor.h"
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame^]36#include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]37#include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h"
38#include "arm_compute/runtime/CL/functions/CLGEMM.h"
39#include "arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h"
40#include "arm_compute/runtime/IMemoryManager.h"
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]41#include "arm_compute/runtime/common/LSTMParams.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]42
43#include <memory>
44
45namespace arm_compute
46{
47class ICLTensor;
48
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]49/** This function performs a single time step in a Long Short-Term Memory (LSTM) layer.
50 *
51 */
52class CLLSTMLayer : public IFunction
53{
54public:
55 /** Default constructor */
56 CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
57 /** Initialize function's tensors.
58 *
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]59 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
60 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
61 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
62 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
63 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
64 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
65 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
66 * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
67 * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
68 * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
69 * @param[in] output_state_in 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
70 * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
71 * @param[out] scratch_buffer 2D tensor with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF. Data type supported: Same as @p input.
72 * @param[out] output_state_out 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
73 * @param[out] cell_state_out 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
74 * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
75 * Data types supported: Same as @p input.
76 * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization:
77 * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
78 * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
79 * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
80 * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
81 * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
82 * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
83 * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
84 * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
85 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
86 * @param[in] cell_threshold The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0 then clipping is disabled.
87 * @param[in] projection_threshold The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]88 */
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]89 void configure(const ICLTensor *input,
90 const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]91 const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]92 const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
93 const ICLTensor *output_state_in, const ICLTensor *cell_state_in,
94 ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]95 const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
96
97 /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayer
98 *
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]99 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
100 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
101 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
102 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
103 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
104 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
105 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
106 * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
107 * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
108 * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
109 * @param[in] output_state_in 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
110 * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
111 * @param[in] scratch_buffer 2D tensor with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF. Data type supported: Same as @p input.
112 * @param[in] output_state_out 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
113 * @param[in] cell_state_out 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
114 * @param[in] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]115 * Data types supported: Same as @p input.
116 * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization:
117 * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
118 * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
119 * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
120 * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
121 * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
122 * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
123 * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
124 * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
125 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
126 * @param[in] cell_threshold The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0 then clipping is disabled.
127 * @param[in] projection_threshold The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
128 *
129 * @return a status
130 */
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]131 static Status validate(const ITensorInfo *input,
132 const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]133 const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
134 const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]135 const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
136 const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output,
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]137 const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
138
139 // Inherited methods overridden:
140 void run() override;
141
142private:
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame^]143 CLMemoryGroup _memory_group;
144 CLFullyConnectedLayer _fully_connected_input_gate;
145 CLGEMM _gemm_input_gate;
146 CLTransposeKernel _transpose_input_gate;
147 CLSaturatedArithmeticOperationKernel _accum_input_gate1;
148 CLArithmeticAddition _accum_input_gate2;
149 CLSaturatedArithmeticOperationKernel _subtract_input_gate;
150 CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate;
151 CLActivationLayerKernel _activation_input_gate;
152 CLFullyConnectedLayer _fully_connected_forget_gate;
153 CLGEMM _gemm_forget_gate;
154 CLTransposeKernel _transpose_forget_gate;
155 CLSaturatedArithmeticOperationKernel _accum_forget_gate1;
156 CLArithmeticAddition _accum_forget_gate2;
157 CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate;
158 CLActivationLayerKernel _activation_forget_gate;
159 CLFullyConnectedLayer _fully_connected_cell_state;
160 CLGEMM _gemm_cell_state1;
161 CLGEMM _gemm_cell_state2;
162 CLTransposeKernel _transpose_cell_state;
163 CLSaturatedArithmeticOperationKernel _accum_cell_state1;
164 CLSaturatedArithmeticOperationKernel _accum_cell_state2;
165 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state1;
166 CLActivationLayerKernel _activation_cell_state;
167 CLActivationLayerKernel _cell_clip;
168 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2;
169 CLFullyConnectedLayer _fully_connected_output;
170 CLGEMM _gemm_output;
171 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state1;
172 CLTransposeKernel _transpose_output;
173 CLSaturatedArithmeticOperationKernel _accum_output1;
174 CLArithmeticAddition _accum_output2;
175 CLActivationLayerKernel _activation_output;
176 CLActivationLayerKernel _activation_output_state;
177 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state2;
178 CLFullyConnectedLayer _fully_connected_output_state;
179 CLGEMM _gemm_output_state;
180 CLSaturatedArithmeticOperationKernel _accum_output_state;
181 CLActivationLayerKernel _projection_clip;
182 CLCopyKernel _copy_cell_state;
183 CLCopyKernel _copy_output;
184 CLWidthConcatenateLayer _concat_scratch_buffer;
185 CLTensor _input_gate_out1;
186 CLTensor _input_gate_out2;
187 CLTensor _input_gate_out3;
188 CLTensor _input_gate_out4;
189 CLTensor _input_gate_out5;
190 CLTensor _forget_gate_out1;
191 CLTensor _forget_gate_out2;
192 CLTensor _forget_gate_out3;
193 CLTensor _forget_gate_out4;
194 CLTensor _forget_gate_out5;
195 CLTensor _cell_state_out1;
196 CLTensor _cell_state_out2;
197 CLTensor _cell_state_out3;
198 CLTensor _cell_state_out4;
199 CLTensor _cell_state_out5;
200 CLTensor _output1;
201 CLTensor _output2;
202 CLTensor _output3;
203 CLTensor _output4;
204 CLTensor _output5;
205 CLTensor _cell_state_activation;
206 CLTensor _output_state1;
207 CLTensor _ones;
208 bool _run_peephole_opt;
209 bool _run_cifg_opt;
210 bool _perform_cell_clipping;
211 bool _has_projection_weights;
212 bool _perform_projection_clipping;
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]213};
214}
215#endif /* __ARM_COMPUTE_CLLSTMLAYER_H__ */