Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / df4cf57c7394265b27d051cb1cf0152c53659126 / . / arm_compute / runtime / NEON / functions / NELSTMLayer.h
blob: 183745c185ec296637b4803b536a367ef8cc909f [file] [log] [blame]
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]1/*
John Kesapides917959c2019-02-04 12:37:29 +0000[diff] [blame]2 * Copyright (c) 2018-2019 ARM Limited.
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[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 */
24#ifndef __ARM_COMPUTE_NELSTMLAYER_H__
25#define __ARM_COMPUTE_NELSTMLAYER_H__
26
27#include "arm_compute/core/NEON/kernels/NEActivationLayerKernel.h"
28#include "arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h"
29#include "arm_compute/core/NEON/kernels/NEArithmeticSubtractionKernel.h"
30#include "arm_compute/core/NEON/kernels/NECopyKernel.h"
31#include "arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h"
32
33#include "arm_compute/core/Types.h"
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]34#include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h"
Georgios Pinitas09f24972019-05-17 18:14:40 +0100[diff] [blame]35#include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]36#include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
37#include "arm_compute/runtime/NEON/functions/NEGEMM.h"
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]38#include "arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h"
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]39#include "arm_compute/runtime/common/LSTMParams.h"
40
41namespace arm_compute
42{
43// Forward declarations
44class ITensor;
45
46/** Basic function to run @ref NELSTMLayer */
47class NELSTMLayer : public IFunction
48{
49public:
50 /** Default constructor */
51 NELSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
52 /** Initialize function's tensors.
53 *
54 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
55 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
56 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
57 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
58 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
59 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
60 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
61 * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
62 * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
63 * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
64 * @param[in] output_state_in 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
65 * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
66 * @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.
67 * @param[out] output_state_out 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
68 * @param[out] cell_state_out 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
69 * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
70 * Data types supported: Same as @p input.
71 * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization:
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]72 * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
73 * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
74 * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
75 * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
76 * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
77 * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
78 * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
79 * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
80 * input_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
81 * forget_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
82 * cell_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
83 * output_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]84 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
85 * @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.
86 * @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.
87 */
88 void configure(const ITensor *input,
89 const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights,
90 const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights,
91 const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
92 const ITensor *output_state_in, const ITensor *cell_state_in,
93 ITensor *scratch_buffer, ITensor *output_state_out, ITensor *cell_state_out, ITensor *output,
94 const LSTMParams<ITensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
95
96 /** Static function to check if given info will lead to a valid configuration of @ref NELSTMLayer
97 *
98 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
99 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
100 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
101 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
102 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
103 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
104 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
105 * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
106 * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
107 * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
108 * @param[in] output_state_in 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
109 * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
110 * @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.
111 * @param[in] output_state_out 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
112 * @param[in] cell_state_out 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
113 * @param[in] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
114 * Data types supported: Same as @p input.
115 * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization:
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]116 * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
117 * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
118 * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
119 * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
120 * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
121 * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
122 * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
123 * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
124 * input_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
125 * forget_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
126 * cell_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
127 * output_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]128 * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo.
129 * @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.
130 * @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.
131 *
132 * @return a status
133 */
134 static Status validate(const ITensorInfo *input,
135 const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
136 const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
137 const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
138 const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
139 const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output,
140 const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f);
141
142 // Inherited methods overridden:
143 void run() override;
John Kesapides917959c2019-02-04 12:37:29 +0000[diff] [blame]144 void prepare() override;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]145
146private:
147 MemoryGroup _memory_group;
148 NEFullyConnectedLayer _fully_connected_input_gate;
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]149 NEArithmeticAddition _accum_input_gate1;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]150 NEArithmeticSubtractionKernel _subtract_input_gate;
151 NEPixelWiseMultiplicationKernel _pixelwise_mul_input_gate;
152 NEActivationLayerKernel _activation_input_gate;
153 NEFullyConnectedLayer _fully_connected_forget_gate;
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]154 NEArithmeticAddition _accum_forget_gate1;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]155 NEPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate;
156 NEActivationLayerKernel _activation_forget_gate;
157 NEFullyConnectedLayer _fully_connected_cell_state;
158 NEGEMM _gemm_cell_state1;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]159 NETransposeKernel _transpose_cell_state;
160 NEArithmeticAdditionKernel _accum_cell_state1;
161 NEArithmeticAdditionKernel _accum_cell_state2;
162 NEPixelWiseMultiplicationKernel _pixelwise_mul_cell_state1;
163 NEActivationLayerKernel _activation_cell_state;
164 NEActivationLayerKernel _cell_clip;
165 NEPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2;
166 NEFullyConnectedLayer _fully_connected_output;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]167 NEPixelWiseMultiplicationKernel _pixelwise_mul_output_state1;
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]168 NEArithmeticAddition _accum_output1;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]169 NEActivationLayerKernel _activation_output;
170 NEActivationLayerKernel _activation_output_state;
171 NEPixelWiseMultiplicationKernel _pixelwise_mul_output_state2;
172 NEFullyConnectedLayer _fully_connected_output_state;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]173 NEActivationLayerKernel _projection_clip;
174 NECopyKernel _copy_cell_state;
175 NECopyKernel _copy_output;
Georgios Pinitas09f24972019-05-17 18:14:40 +0100[diff] [blame]176 NEConcatenateLayer _concat_scratch_buffer;
177 NEConcatenateLayer _concat_inputs_forget_gate;
178 NEConcatenateLayer _concat_weights_forget_gate;
179 NEConcatenateLayer _concat_weights_input_gate;
180 NEConcatenateLayer _concat_weights_output;
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]181 NEMeanStdDevNormalizationLayer _mean_std_norm_input_gate;
182 NEPixelWiseMultiplicationKernel _pixelwise_mul_input_gate_coeff;
183 NEArithmeticAdditionKernel _accum_input_gate_bias;
184 NEMeanStdDevNormalizationLayer _mean_std_norm_forget_gate;
185 NEPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate_coeff;
186 NEArithmeticAdditionKernel _accum_forget_gate_bias;
187 NEMeanStdDevNormalizationLayer _mean_std_norm_cell_gate;
188 NEPixelWiseMultiplicationKernel _pixelwise_mul_cell_gate_coeff;
189 NEArithmeticAdditionKernel _accum_cell_gate_bias;
190 NEMeanStdDevNormalizationLayer _mean_std_norm_output_gate;
191 NEPixelWiseMultiplicationKernel _pixelwise_mul_output_gate_coeff;
192 NEArithmeticAdditionKernel _accum_output_gate_bias;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]193 Tensor _input_gate_out1;
194 Tensor _input_gate_out2;
195 Tensor _input_gate_out3;
196 Tensor _input_gate_out4;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]197 Tensor _forget_gate_out1;
198 Tensor _forget_gate_out2;
199 Tensor _forget_gate_out3;
200 Tensor _forget_gate_out4;
201 Tensor _forget_gate_out5;
John Kesapides917959c2019-02-04 12:37:29 +0000[diff] [blame]202 Tensor _forget_gate_out6;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]203 Tensor _cell_state_out1;
204 Tensor _cell_state_out2;
205 Tensor _cell_state_out3;
206 Tensor _cell_state_out4;
207 Tensor _cell_state_out5;
208 Tensor _output1;
209 Tensor _output2;
210 Tensor _output3;
211 Tensor _output4;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]212 Tensor _cell_state_activation;
213 Tensor _output_state1;
214 Tensor _ones;
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]215 Tensor _input_layer_norm_out1;
216 Tensor _input_layer_norm_out2;
217 Tensor _forget_layer_norm_out1;
218 Tensor _forget_layer_norm_out2;
219 Tensor _cell_layer_norm_out1;
220 Tensor _cell_layer_norm_out2;
221 Tensor _output_layer_norm_out1;
222 Tensor _output_layer_norm_out2;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]223 bool _run_peephole_opt;
224 bool _run_cifg_opt;
225 bool _perform_cell_clipping;
226 bool _has_projection_weights;
227 bool _perform_projection_clipping;
John Kesapides917959c2019-02-04 12:37:29 +0000[diff] [blame]228 bool _is_prepared;
Michele Di Giorgio0cbfda62019-06-13 17:01:29 +0100[diff] [blame]229 bool _is_layer_norm_lstm;
Michalis Spyrou25f45a42018-08-08 12:53:05 +0100[diff] [blame]230};
231} // namespace arm_compute
232#endif /* __ARM_COMPUTE_NELSTMLAYER_H__ */