Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 70d33bdfd36e1b44b0573189dca67ed7c63dd59e / . / arm_compute / runtime / CL / functions / CLLSTMLayer.h
blob: 9509016075e28c688b813ea4ccdf7c5a58863cd1 [file] [log] [blame]
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]1/*
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]2 * Copyright (c) 2018-2019 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.
79 * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization:
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]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_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
89 * forget_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
90 * cell_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
91 * output_layer_norm_coefficients 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 Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]93 * @param[in] cell_threshold The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0f then clipping is disabled.
94 * @param[in] projection_threshold The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip].
95 * If set to 0.0f then clipping is disabled.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]96 */
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]97 void configure(const ICLTensor *input,
98 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]99 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]100 const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
101 const ICLTensor *output_state_in, const ICLTensor *cell_state_in,
102 ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]103 const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
104
105 /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayer
106 *
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]107 * @param[in] input Source tensor info. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
108 * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
109 * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
110 * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
111 * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
112 * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
113 * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
114 * @param[in] forget_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
115 * @param[in] cell_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
116 * @param[in] output_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
117 * @param[in] output_state_in 2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
118 * @param[in] cell_state_in 2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
119 * @param[in] scratch_buffer 2D tensor info with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF.
120 * Data type supported: Same as @p input.
121 * @param[in] output_state_out 2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
122 * @param[in] cell_state_out 2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
123 * @param[in] output Destination tensor info. Output is a 2D tensor with dimensions [output_size, batch_size]. Data types supported: Same as @p input.
124 * @param[in] lstm_params (Optional) Weights tensors info used in peephole optimization:
125 * input_to_input_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
126 * recurrent_to_input_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
127 * cell_to_input_weights 1D weights tensor info with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
128 * cell_to_forget_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
129 * cell_to_output_weights 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
130 * input_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input
131 * projection_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
132 * projection_bias 1D weights tensor info with dimensions [output_size]. Data type supported: Same as @p input.
133 * input_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
134 * forget_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
135 * cell_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
136 * output_layer_norm_coefficients 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]137 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]138 * @param[in] cell_threshold The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0f then clipping is disabled.
139 * @param[in] projection_threshold The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip].
140 * If set to 0.0f then clipping is disabled.
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]141 *
142 * @return a status
143 */
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]144 static Status validate(const ITensorInfo *input,
145 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]146 const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
147 const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
Georgios Pinitas8bc745d2018-07-18 19:51:24 +0100[diff] [blame]148 const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
149 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]150 const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
151
152 // Inherited methods overridden:
153 void run() override;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]154 void prepare() override;
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]155
156private:
Georgios Pinitas26014cf2019-09-09 19:00:57 +0100[diff] [blame]157 MemoryGroup _memory_group;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]158 CLFullyConnectedLayer _fully_connected_input_gate;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]159 CLArithmeticAddition _accum_input_gate1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]160 CLSaturatedArithmeticOperationKernel _subtract_input_gate;
161 CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate;
162 CLActivationLayerKernel _activation_input_gate;
163 CLFullyConnectedLayer _fully_connected_forget_gate;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]164 CLArithmeticAddition _accum_forget_gate1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]165 CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate;
166 CLActivationLayerKernel _activation_forget_gate;
167 CLFullyConnectedLayer _fully_connected_cell_state;
168 CLGEMM _gemm_cell_state1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]169 CLTransposeKernel _transpose_cell_state;
170 CLSaturatedArithmeticOperationKernel _accum_cell_state1;
171 CLSaturatedArithmeticOperationKernel _accum_cell_state2;
172 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state1;
173 CLActivationLayerKernel _activation_cell_state;
174 CLActivationLayerKernel _cell_clip;
175 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2;
176 CLFullyConnectedLayer _fully_connected_output;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]177 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state1;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]178 CLArithmeticAddition _accum_output1;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]179 CLActivationLayerKernel _activation_output;
180 CLActivationLayerKernel _activation_output_state;
181 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state2;
182 CLFullyConnectedLayer _fully_connected_output_state;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]183 CLActivationLayerKernel _projection_clip;
184 CLCopyKernel _copy_cell_state;
185 CLCopyKernel _copy_output;
Georgios Pinitas09f24972019-05-17 18:14:40 +0100[diff] [blame]186 CLConcatenateLayer _concat_scratch_buffer;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]187 CLWidthConcatenate2TensorsKernel _concat_inputs_forget_gate;
188 CLWidthConcatenate2TensorsKernel _concat_weights_forget_gate;
189 CLWidthConcatenate2TensorsKernel _concat_weights_input_gate;
190 CLWidthConcatenate2TensorsKernel _concat_weights_output;
Georgios Pinitasdbfc2dc2019-04-02 12:51:21 +0100[diff] [blame]191 CLMemsetKernel _ones_memset_kernel;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]192 CLMeanStdDevNormalizationLayer _mean_std_norm_input_gate;
193 CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate_coeff;
194 CLSaturatedArithmeticOperationKernel _accum_input_gate_bias;
195 CLMeanStdDevNormalizationLayer _mean_std_norm_forget_gate;
196 CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate_coeff;
197 CLSaturatedArithmeticOperationKernel _accum_forget_gate_bias;
198 CLMeanStdDevNormalizationLayer _mean_std_norm_cell_gate;
199 CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_gate_coeff;
200 CLSaturatedArithmeticOperationKernel _accum_cell_gate_bias;
201 CLMeanStdDevNormalizationLayer _mean_std_norm_output_gate;
202 CLPixelWiseMultiplicationKernel _pixelwise_mul_output_gate_coeff;
203 CLSaturatedArithmeticOperationKernel _accum_output_gate_bias;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]204 CLTensor _input_gate_out1;
205 CLTensor _input_gate_out2;
206 CLTensor _input_gate_out3;
207 CLTensor _input_gate_out4;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]208 CLTensor _forget_gate_out1;
209 CLTensor _forget_gate_out2;
210 CLTensor _forget_gate_out3;
211 CLTensor _forget_gate_out4;
212 CLTensor _forget_gate_out5;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]213 CLTensor _forget_gate_out6;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]214 CLTensor _cell_state_out1;
215 CLTensor _cell_state_out2;
216 CLTensor _cell_state_out3;
217 CLTensor _cell_state_out4;
218 CLTensor _cell_state_out5;
219 CLTensor _output1;
220 CLTensor _output2;
221 CLTensor _output3;
222 CLTensor _output4;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]223 CLTensor _cell_state_activation;
224 CLTensor _output_state1;
225 CLTensor _ones;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]226 CLTensor _input_layer_norm_out1;
227 CLTensor _input_layer_norm_out2;
228 CLTensor _forget_layer_norm_out1;
229 CLTensor _forget_layer_norm_out2;
230 CLTensor _cell_layer_norm_out1;
231 CLTensor _cell_layer_norm_out2;
232 CLTensor _output_layer_norm_out1;
233 CLTensor _output_layer_norm_out2;
giuros01164a2722018-11-20 18:34:46 +0000[diff] [blame]234 bool _run_peephole_opt;
235 bool _run_cifg_opt;
236 bool _perform_cell_clipping;
237 bool _has_projection_weights;
238 bool _perform_projection_clipping;
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]239 bool _is_prepared;
Michele Di Giorgio39438b42019-06-04 12:41:45 +0100[diff] [blame]240 bool _is_layer_norm_lstm;
Michalis Spyroubcedf512018-03-22 14:55:08 +0000[diff] [blame]241};
John Kesapidescafec8f2019-02-19 15:53:59 +0000[diff] [blame]242} // namespace arm_compute
Michalis Spyrouf4643372019-11-29 16:17:13 +0000[diff] [blame]243#endif /* ARM_COMPUTE_CLLSTMLAYER_H */