Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 25d9775b08c59be8d17700f75026a8457aab9838 / . / arm_compute / runtime / CL / functions / CLLSTMLayer.h
blob: a94f239472e5d4ff19814c8f9be66051b0c84b43 [file] [log] [blame]
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]1/*
Michele Di Giorgio25d97752020-03-04 18:08:47 +0000[diff] [blame^]2 * Copyright (c) 2018-2020 ARM Limited.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]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 */
Michalis Spyrouf4643372019-11-29 16:17:13 +0000[diff] [blame]24#ifndef ARM_COMPUTE_CLLSTMLAYER_H
25#define ARM_COMPUTE_CLLSTMLAYER_H
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]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"
Georgios Pinitasdbfc2dc2019-04-02 12:51:21 +0100[diff] [blame]32#include "arm_compute/core/CL/kernels/CLMemsetKernel.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]33#include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h"
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]34#include "arm_compute/core/CL/kernels/CLWidthConcatenate2TensorsKernel.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]35#include "arm_compute/core/Types.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]36#include "arm_compute/runtime/CL/CLTensor.h"
Georgios Pinitas09f24972019-05-17 18:14:40 +0100[diff] [blame]37#include "arm_compute/runtime/CL/functions/CLConcatenateLayer.h"
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]38#include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]39#include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h"
40#include "arm_compute/runtime/CL/functions/CLGEMM.h"
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]41#include "arm_compute/runtime/CL/functions/CLMeanStdDevNormalizationLayer.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]42#include "arm_compute/runtime/IMemoryManager.h"
Georgios Pinitas26014cf2019-09-09 19:00:57 +0100[diff] [blame]43#include "arm_compute/runtime/MemoryGroup.h"
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]44#include "arm_compute/runtime/common/LSTMParams.h"
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]45
46#include <memory>
47
48namespace arm_compute
49{
50class ICLTensor;
51
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]52/** This function performs a single time step in a Long Short-Term Memory (LSTM) layer.
53 *
54 */
55class CLLSTMLayer : public IFunction
56{
57public:
58 /** Default constructor */
59 CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
60 /** Initialize function's tensors.
61 *
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]62 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
63 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
64 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
65 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
66 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
67 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
68 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
69 * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
70 * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
71 * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
72 * @param[in] output_state_in 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
73 * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
74 * @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.
75 * @param[out] output_state_out 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
76 * @param[out] cell_state_out 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
77 * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
78 * Data types supported: Same as @p input.
Michele Di Giorgio25d97752020-03-04 18:08:47 +0000[diff] [blame^]79 * @param[in] lstm_params Weights tensors used in peephole optimization:
80 * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
81 * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
82 * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
83 * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
84 * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
85 * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
86 * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
87 * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
88 * input_layer_norm_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
89 * forget_layer_norm_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
90 * cell_layer_norm_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
91 * output_layer_norm_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]92 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
Michele Di Giorgio25d97752020-03-04 18:08:47 +0000[diff] [blame^]93 * @param[in] cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip].
94 * If set to 0.0f then clipping is disabled.
95 * @param[in] projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip].
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]96 * If set to 0.0f then clipping is disabled.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]97 */
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]98 void configure(const ICLTensor *input,
99 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]100 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]101 const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
102 const ICLTensor *output_state_in, const ICLTensor *cell_state_in,
103 ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]104 const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
105
106 /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayer
107 *
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]108 * @param[in] input Source tensor info. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
109 * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
110 * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
111 * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
112 * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
113 * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
114 * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
115 * @param[in] forget_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
116 * @param[in] cell_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
117 * @param[in] output_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
118 * @param[in] output_state_in 2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
119 * @param[in] cell_state_in 2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
120 * @param[in] scratch_buffer 2D tensor info with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF.
121 * Data type supported: Same as @p input.
122 * @param[in] output_state_out 2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
123 * @param[in] cell_state_out 2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
124 * @param[in] output Destination tensor info. Output is a 2D tensor with dimensions [output_size, batch_size]. Data types supported: Same as @p input.
Michele Di Giorgio25d97752020-03-04 18:08:47 +0000[diff] [blame^]125 * @param[in] lstm_params Weights tensors info used in peephole optimization:
126 * input_to_input_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
127 * recurrent_to_input_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
128 * cell_to_input_weights 1D weights tensor info with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
129 * cell_to_forget_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
130 * cell_to_output_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
131 * input_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input
132 * projection_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
133 * projection_bias 1D weights tensor info with dimensions [output_size]. Data type supported: Same as @p input.
134 * input_layer_norm_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
135 * forget_layer_norm_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
136 * cell_layer_norm_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
137 * output_layer_norm_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]138 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
Michele Di Giorgio25d97752020-03-04 18:08:47 +0000[diff] [blame^]139 * @param[in] cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip].
140 * If set to 0.0f then clipping is disabled.
141 * @param[in] projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip].
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]142 * If set to 0.0f then clipping is disabled.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]143 *
144 * @return a status
145 */
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]146 static Status validate(const ITensorInfo *input,
147 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]148 const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
149 const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]150 const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
151 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]152 const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
153
154 // Inherited methods overridden:
155 void run() override;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]156 void prepare() override;
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]157
158private:
Georgios Pinitas26014cf2019-09-09 19:00:57 +0100[diff] [blame]159 MemoryGroup _memory_group;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]160 CLFullyConnectedLayer _fully_connected_input_gate;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]161 CLArithmeticAddition _accum_input_gate1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]162 CLSaturatedArithmeticOperationKernel _subtract_input_gate;
163 CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate;
164 CLActivationLayerKernel _activation_input_gate;
165 CLFullyConnectedLayer _fully_connected_forget_gate;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]166 CLArithmeticAddition _accum_forget_gate1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]167 CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate;
168 CLActivationLayerKernel _activation_forget_gate;
169 CLFullyConnectedLayer _fully_connected_cell_state;
170 CLGEMM _gemm_cell_state1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]171 CLTransposeKernel _transpose_cell_state;
172 CLSaturatedArithmeticOperationKernel _accum_cell_state1;
173 CLSaturatedArithmeticOperationKernel _accum_cell_state2;
174 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state1;
175 CLActivationLayerKernel _activation_cell_state;
176 CLActivationLayerKernel _cell_clip;
177 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2;
178 CLFullyConnectedLayer _fully_connected_output;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]179 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state1;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]180 CLArithmeticAddition _accum_output1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]181 CLActivationLayerKernel _activation_output;
182 CLActivationLayerKernel _activation_output_state;
183 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state2;
184 CLFullyConnectedLayer _fully_connected_output_state;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]185 CLActivationLayerKernel _projection_clip;
186 CLCopyKernel _copy_cell_state;
187 CLCopyKernel _copy_output;
Georgios Pinitas09f24972019-05-17 18:14:40 +0100[diff] [blame]188 CLConcatenateLayer _concat_scratch_buffer;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]189 CLWidthConcatenate2TensorsKernel _concat_inputs_forget_gate;
190 CLWidthConcatenate2TensorsKernel _concat_weights_forget_gate;
191 CLWidthConcatenate2TensorsKernel _concat_weights_input_gate;
192 CLWidthConcatenate2TensorsKernel _concat_weights_output;
Georgios Pinitasdbfc2dc2019-04-02 12:51:21 +0100[diff] [blame]193 CLMemsetKernel _ones_memset_kernel;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]194 CLMeanStdDevNormalizationLayer _mean_std_norm_input_gate;
195 CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate_coeff;
196 CLSaturatedArithmeticOperationKernel _accum_input_gate_bias;
197 CLMeanStdDevNormalizationLayer _mean_std_norm_forget_gate;
198 CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate_coeff;
199 CLSaturatedArithmeticOperationKernel _accum_forget_gate_bias;
200 CLMeanStdDevNormalizationLayer _mean_std_norm_cell_gate;
201 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_gate_coeff;
202 CLSaturatedArithmeticOperationKernel _accum_cell_gate_bias;
203 CLMeanStdDevNormalizationLayer _mean_std_norm_output_gate;
204 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_gate_coeff;
205 CLSaturatedArithmeticOperationKernel _accum_output_gate_bias;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]206 CLTensor _input_gate_out1;
207 CLTensor _input_gate_out2;
208 CLTensor _input_gate_out3;
209 CLTensor _input_gate_out4;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]210 CLTensor _forget_gate_out1;
211 CLTensor _forget_gate_out2;
212 CLTensor _forget_gate_out3;
213 CLTensor _forget_gate_out4;
214 CLTensor _forget_gate_out5;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]215 CLTensor _forget_gate_out6;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]216 CLTensor _cell_state_out1;
217 CLTensor _cell_state_out2;
218 CLTensor _cell_state_out3;
219 CLTensor _cell_state_out4;
220 CLTensor _cell_state_out5;
221 CLTensor _output1;
222 CLTensor _output2;
223 CLTensor _output3;
224 CLTensor _output4;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]225 CLTensor _cell_state_activation;
226 CLTensor _output_state1;
227 CLTensor _ones;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]228 CLTensor _input_layer_norm_out1;
229 CLTensor _input_layer_norm_out2;
230 CLTensor _forget_layer_norm_out1;
231 CLTensor _forget_layer_norm_out2;
232 CLTensor _cell_layer_norm_out1;
233 CLTensor _cell_layer_norm_out2;
234 CLTensor _output_layer_norm_out1;
235 CLTensor _output_layer_norm_out2;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]236 bool _run_peephole_opt;
237 bool _run_cifg_opt;
238 bool _perform_cell_clipping;
239 bool _has_projection_weights;
240 bool _perform_projection_clipping;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]241 bool _is_prepared;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]242 bool _is_layer_norm_lstm;
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]243};
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]244} // namespace arm_compute
Michalis Spyrouf4643372019-11-29 16:17:13 +0000[diff] [blame]245#endif /* ARM_COMPUTE_CLLSTMLAYER_H */