Michalis Spyrou | bcedf51 | 2018-03-22 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" |
| 30 | #include "arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h" |
| 31 | #include "arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h" |
| 32 | #include "arm_compute/core/CL/kernels/CLCopyKernel.h" |
| 33 | #include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/runtime/CL/CLMemoryGroup.h" |
| 36 | #include "arm_compute/runtime/CL/CLTensor.h" |
| 37 | #include "arm_compute/runtime/CL/functions/CLArithmeticAddition.h" |
| 38 | #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" |
| 39 | #include "arm_compute/runtime/CL/functions/CLGEMM.h" |
| 40 | #include "arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h" |
| 41 | #include "arm_compute/runtime/IMemoryManager.h" |
| 42 | |
| 43 | #include <memory> |
| 44 | |
| 45 | namespace arm_compute |
| 46 | { |
| 47 | class ICLTensor; |
| 48 | |
| 49 | template <typename T> |
| 50 | class LSTMParams |
| 51 | { |
| 52 | public: |
| 53 | /** Constructor */ |
| 54 | LSTMParams() |
| 55 | : _input_to_input_weights(nullptr), _recurrent_to_input_weights(nullptr), _cell_to_input_weights(nullptr), _input_gate_bias(nullptr), _cell_to_forget_weights(nullptr), |
| 56 | _cell_to_output_weights(nullptr), _projection_weights(nullptr), _projection_bias(nullptr), _has_peephole_opt(false), _has_projection(false), _has_cifg_opt(true) |
| 57 | { |
| 58 | } |
| 59 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 60 | LSTMParams(const LSTMParams &) = delete; |
| 61 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 62 | LSTMParams &operator=(const LSTMParams &) = delete; |
| 63 | /** Default destructor */ |
| 64 | ~LSTMParams() = default; |
| 65 | /** Set CIFG tensor parameters. |
| 66 | * |
| 67 | * @param[in] input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data types supported: F16/F32. |
| 68 | * @param[in] recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input_to_input_weights. |
| 69 | * @param[in] cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input_to_input_weights. |
| 70 | * @param[in] input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input_to_input_weights |
| 71 | * |
| 72 | * @return Reference to this LSTMParams object |
| 73 | */ |
| 74 | LSTMParams &set_cifg_params(const T *input_to_input_weights, const T *recurrent_to_input_weights, const T *cell_to_input_weights, const T *input_gate_bias) |
| 75 | { |
| 76 | _input_to_input_weights = input_to_input_weights; |
| 77 | _recurrent_to_input_weights = recurrent_to_input_weights; |
| 78 | _cell_to_input_weights = cell_to_input_weights; |
| 79 | _input_gate_bias = input_gate_bias; |
| 80 | _has_cifg_opt = false; |
| 81 | return *this; |
| 82 | } |
| 83 | /** Set projection tensor parameters. |
| 84 | * |
| 85 | * @param[in] projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Data types supported: F16/F32. |
| 86 | * @param[in] projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p projection_weights. |
| 87 | * |
| 88 | * @return Reference to this LSTMParams object |
| 89 | */ |
| 90 | LSTMParams &set_projection_params(const T *projection_weights, const T *projection_bias) |
| 91 | { |
| 92 | _projection_weights = projection_weights; |
| 93 | _projection_bias = projection_bias; |
| 94 | _has_projection = true; |
| 95 | return *this; |
| 96 | } |
| 97 | /** Set peephole tensor parameters. |
| 98 | * |
| 99 | * @param[in] cell_to_input_weights 1D weights tensor with dimensions [num_units]. Data type supported: Data types supported: F16/F32. |
| 100 | * @param[in] cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p cell_to_input_weights. |
| 101 | * @param[in] cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p cell_to_input_weights. |
| 102 | * |
| 103 | * @return Reference to this LSTMParams object |
| 104 | */ |
| 105 | LSTMParams &set_peephole_params(const T *cell_to_input_weights, const T *cell_to_forget_weights, const T *cell_to_output_weights) |
| 106 | { |
| 107 | _cell_to_input_weights = cell_to_input_weights; |
| 108 | _cell_to_forget_weights = cell_to_forget_weights; |
| 109 | _cell_to_output_weights = cell_to_output_weights; |
| 110 | _has_peephole_opt = true; |
| 111 | return *this; |
| 112 | } |
| 113 | |
| 114 | const T *input_to_input_weights() const |
| 115 | { |
| 116 | return _input_to_input_weights; |
| 117 | } |
| 118 | |
| 119 | const T *recurrent_to_input_weights() const |
| 120 | { |
| 121 | return _recurrent_to_input_weights; |
| 122 | } |
| 123 | |
| 124 | const T *cell_to_input_weights() const |
| 125 | { |
| 126 | return _cell_to_input_weights; |
| 127 | } |
| 128 | |
| 129 | const T *input_gate_bias() const |
| 130 | { |
| 131 | return _input_gate_bias; |
| 132 | } |
| 133 | |
| 134 | const T *cell_to_forget_weights() const |
| 135 | { |
| 136 | return _cell_to_forget_weights; |
| 137 | } |
| 138 | |
| 139 | const T *cell_to_output_weights() const |
| 140 | { |
| 141 | return _cell_to_output_weights; |
| 142 | } |
| 143 | |
| 144 | const T *projection_weights() const |
| 145 | { |
| 146 | return _projection_weights; |
| 147 | } |
| 148 | |
| 149 | const T *projection_bias() const |
| 150 | { |
| 151 | return _projection_bias; |
| 152 | } |
| 153 | |
| 154 | bool has_peephole_opt() const |
| 155 | { |
| 156 | return _has_peephole_opt; |
| 157 | } |
| 158 | |
| 159 | bool has_projection() const |
| 160 | { |
| 161 | return _has_projection; |
| 162 | } |
| 163 | |
| 164 | bool has_cifg_opt() const |
| 165 | { |
| 166 | return _has_cifg_opt; |
| 167 | } |
| 168 | |
| 169 | private: |
| 170 | const T *_input_to_input_weights; |
| 171 | const T *_recurrent_to_input_weights; |
| 172 | const T *_cell_to_input_weights; |
| 173 | const T *_input_gate_bias; |
| 174 | const T *_cell_to_forget_weights; |
| 175 | const T *_cell_to_output_weights; |
| 176 | const T *_projection_weights; |
| 177 | const T *_projection_bias; |
| 178 | bool _has_peephole_opt; |
| 179 | bool _has_projection; |
| 180 | bool _has_cifg_opt; |
| 181 | }; |
| 182 | |
| 183 | /** This function performs a single time step in a Long Short-Term Memory (LSTM) layer. |
| 184 | * |
| 185 | */ |
| 186 | class CLLSTMLayer : public IFunction |
| 187 | { |
| 188 | public: |
| 189 | /** Default constructor */ |
| 190 | CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
| 191 | /** Initialize function's tensors. |
| 192 | * |
| 193 | * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32. |
| 194 | * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 195 | * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 196 | * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 197 | * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 198 | * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 199 | * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 200 | * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 201 | * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 202 | * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 203 | * @param[in, out] output_state 2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input. |
| 204 | * @param[in, out] cell_state 2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input. |
| 205 | * @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. |
| 206 | * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size]. |
| 207 | * Data types supported: Same as @p input. |
| 208 | * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization: |
| 209 | * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 210 | * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 211 | * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input. |
| 212 | * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 213 | * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 214 | * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input |
| 215 | * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 216 | * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input. |
| 217 | * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo. |
| 218 | * @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. |
| 219 | * @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. |
| 220 | */ |
| 221 | void configure(const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, |
| 222 | const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, |
| 223 | const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, ICLTensor *output_state, ICLTensor *cell_state, ICLTensor *scratch_buffer, ICLTensor *output, |
| 224 | const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f); |
| 225 | |
| 226 | /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayer |
| 227 | * |
| 228 | * @param[in] input Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: F16/F32. |
| 229 | * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 230 | * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 231 | * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 232 | * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 233 | * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 234 | * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 235 | * @param[in] forget_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
| 236 | * @param[in] cell_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
| 237 | * @param[in] output_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
| 238 | * @param[in] output_state 2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input. |
| 239 | * @param[in] cell_state 2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input. |
| 240 | * @param[in] scratch_buffer 2D tensor info with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF. Data type supported: Same as @p input. |
| 241 | * @param[in] output Destination tensor info. Output is a 2D tensor with dimensions [output_size, batch_size]. |
| 242 | * Data types supported: Same as @p input. |
| 243 | * @param[in] lstm_params (Optional) Weights tensors used in peephole optimization: |
| 244 | * input_to_input_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 245 | * recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 246 | * cell_to_input_weights 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input. |
| 247 | * cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 248 | * cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 249 | * input_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input |
| 250 | * projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 251 | * projection_bias 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input. |
| 252 | * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo. |
| 253 | * @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. |
| 254 | * @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. |
| 255 | * |
| 256 | * @return a status |
| 257 | */ |
| 258 | static Status validate(const ITensorInfo *input, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights, |
| 259 | const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights, |
| 260 | const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias, |
| 261 | const ITensorInfo *output_state, const ITensorInfo *cell_state, const ITensorInfo *scratch_buffer, const ITensorInfo *output, |
| 262 | const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f); |
| 263 | |
| 264 | // Inherited methods overridden: |
| 265 | void run() override; |
| 266 | |
| 267 | private: |
| 268 | CLMemoryGroup _memory_group; |
| 269 | CLFullyConnectedLayer _fully_connected_input_gate; |
| 270 | CLGEMM _gemm_input_gate1; |
| 271 | CLGEMM _gemm_input_gate2; |
| 272 | CLTransposeKernel _transpose_input_gate1; |
| 273 | CLTransposeKernel _transpose_input_gate2; |
| 274 | CLArithmeticAdditionKernel _accum_input_gate1; |
| 275 | CLArithmeticAddition _accum_input_gate2; |
| 276 | CLArithmeticSubtractionKernel _subtract_input_gate; |
| 277 | CLActivationLayerKernel _activation_input_gate; |
| 278 | CLFullyConnectedLayer _fully_connected_forget_gate; |
| 279 | CLGEMM _gemm_forget_gate1; |
| 280 | CLGEMM _gemm_forget_gate2; |
| 281 | CLTransposeKernel _transpose_forget_gate1; |
| 282 | CLTransposeKernel _transpose_forget_gate2; |
| 283 | CLArithmeticAdditionKernel _accum_forget_gate1; |
| 284 | CLArithmeticAddition _accum_forget_gate2; |
| 285 | CLActivationLayerKernel _activation_forget_gate; |
| 286 | CLFullyConnectedLayer _fully_connected_cell_state; |
| 287 | CLGEMM _gemm_cell_state1; |
| 288 | CLGEMM _gemm_cell_state2; |
| 289 | CLTransposeKernel _transpose_cell_state1; |
| 290 | CLArithmeticAdditionKernel _accum_cell_state1; |
| 291 | CLArithmeticAdditionKernel _accum_cell_state2; |
| 292 | CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state1; |
| 293 | CLActivationLayerKernel _activation_cell_state; |
| 294 | CLActivationLayerKernel _cell_clip; |
| 295 | CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2; |
| 296 | CLFullyConnectedLayer _fully_connected_output; |
| 297 | CLGEMM _gemm_output1; |
| 298 | CLGEMM _gemm_output2; |
| 299 | CLTransposeKernel _transpose_output1; |
| 300 | CLTransposeKernel _transpose_output2; |
| 301 | CLArithmeticAdditionKernel _accum_output1; |
| 302 | CLArithmeticAddition _accum_output2; |
| 303 | CLActivationLayerKernel _activation_output; |
| 304 | CLActivationLayerKernel _activation_output_state; |
| 305 | CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state; |
| 306 | CLFullyConnectedLayer _fully_connected_output_state; |
| 307 | CLGEMM _gemm_output_state; |
| 308 | CLArithmeticAdditionKernel _accum_output_state; |
| 309 | CLActivationLayerKernel _projection_clip; |
| 310 | CLCopyKernel _copy_cell_state; |
| 311 | CLCopyKernel _copy_output; |
| 312 | CLWidthConcatenateLayer _concat_scratch_buffer; |
| 313 | CLTensor _input_gate_out1; |
| 314 | CLTensor _input_gate_out2; |
| 315 | CLTensor _input_gate_out3; |
| 316 | CLTensor _input_gate_out4; |
| 317 | CLTensor _input_gate_out5; |
| 318 | CLTensor _input_gate_out6; |
| 319 | CLTensor _forget_gate_out1; |
| 320 | CLTensor _forget_gate_out2; |
| 321 | CLTensor _forget_gate_out3; |
| 322 | CLTensor _forget_gate_out4; |
| 323 | CLTensor _forget_gate_out5; |
| 324 | CLTensor _forget_gate_out6; |
| 325 | CLTensor _cell_state_out1; |
| 326 | CLTensor _cell_state_out2; |
| 327 | CLTensor _cell_state_out3; |
| 328 | CLTensor _cell_state_out4; |
| 329 | CLTensor _cell_state_out5; |
| 330 | CLTensor _output1; |
| 331 | CLTensor _output2; |
| 332 | CLTensor _output3; |
| 333 | CLTensor _output4; |
| 334 | CLTensor _output5; |
| 335 | CLTensor _output6; |
| 336 | CLTensor _cell_state_activation; |
| 337 | CLTensor _output_projection1; |
| 338 | CLTensor _ones; |
| 339 | bool _run_peephole_opt; |
| 340 | bool _run_cifg_opt; |
| 341 | bool _perform_cell_clipping; |
| 342 | bool _has_projection_weights; |
| 343 | bool _perform_projection_clipping; |
| 344 | }; |
| 345 | } |
| 346 | #endif /* __ARM_COMPUTE_CLLSTMLAYER_H__ */ |