Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | 47a8990 | 2020-03-09 19:32:33 +0000 | [diff] [blame] | 2 | * Copyright (c) 2018-2020 ARM Limited. |
Michalis Spyrou | 25f45a4 | 2018-08-08 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 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_NELSTMLAYER_H |
| 25 | #define ARM_COMPUTE_NELSTMLAYER_H |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/core/NEON/kernels/NEActivationLayerKernel.h" |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/NEON/kernels/NECopyKernel.h" |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 29 | |
| 30 | #include "arm_compute/core/Types.h" |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 31 | #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h" |
Michalis Spyrou | 173ba9b | 2020-06-23 17:25:43 +0100 | [diff] [blame] | 32 | #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h" |
| 33 | #include "arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h" |
Georgios Pinitas | 09f2497 | 2019-05-17 18:14:40 +0100 | [diff] [blame] | 34 | #include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h" |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 35 | #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h" |
| 36 | #include "arm_compute/runtime/NEON/functions/NEGEMM.h" |
Michele Di Giorgio | 0cbfda6 | 2019-06-13 17:01:29 +0100 | [diff] [blame] | 37 | #include "arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h" |
Michalis Spyrou | 6eb7345 | 2020-07-02 17:39:25 +0100 | [diff] [blame] | 38 | #include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h" |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 39 | #include "arm_compute/runtime/common/LSTMParams.h" |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | // Forward declarations |
| 44 | class ITensor; |
| 45 | |
| 46 | /** Basic function to run @ref NELSTMLayer */ |
| 47 | class NELSTMLayer : public IFunction |
| 48 | { |
| 49 | public: |
| 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. |
Michele Di Giorgio | 47a8990 | 2020-03-09 19:32:33 +0000 | [diff] [blame] | 71 | * @param[in] lstm_params Weights tensors used in peephole optimization: |
| 72 | * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 73 | * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 74 | * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input. |
| 75 | * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 76 | * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 77 | * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input |
| 78 | * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 79 | * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input. |
| 80 | * input_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 81 | * forget_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 82 | * cell_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 83 | * output_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
Michalis Spyrou | 25f45a4 | 2018-08-08 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. |
Michele Di Giorgio | 47a8990 | 2020-03-09 19:32:33 +0000 | [diff] [blame] | 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]. |
| 87 | * If set to 0.0 then clipping is disabled. |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 88 | */ |
| 89 | void configure(const ITensor *input, |
| 90 | const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights, |
| 91 | const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights, |
| 92 | const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias, |
| 93 | const ITensor *output_state_in, const ITensor *cell_state_in, |
| 94 | ITensor *scratch_buffer, ITensor *output_state_out, ITensor *cell_state_out, ITensor *output, |
| 95 | const LSTMParams<ITensor> &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 NELSTMLayer |
| 98 | * |
| 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]. |
| 115 | * Data types supported: Same as @p input. |
Michele Di Giorgio | 47a8990 | 2020-03-09 19:32:33 +0000 | [diff] [blame] | 116 | * @param[in] lstm_params Weights tensors used in peephole optimization: |
| 117 | * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input. |
| 118 | * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 119 | * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input. |
| 120 | * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 121 | * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input. |
| 122 | * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input |
| 123 | * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input. |
| 124 | * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input. |
| 125 | * input_layer_norm_weights (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
| 126 | * forget_layer_norm_weights (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
| 127 | * cell_layer_norm_weights (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
| 128 | * output_layer_norm_weights (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input. |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 129 | * @param[in] activation_info Contains activation information described in @ref ActivationLayerInfo. |
| 130 | * @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. |
Michele Di Giorgio | 47a8990 | 2020-03-09 19:32:33 +0000 | [diff] [blame] | 131 | * @param[in] projection_threshold The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip]. |
| 132 | * If set to 0.0 then clipping is disabled. |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 133 | * |
| 134 | * @return a status |
| 135 | */ |
| 136 | static Status validate(const ITensorInfo *input, |
| 137 | const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights, |
| 138 | const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights, |
| 139 | const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias, |
| 140 | const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in, |
| 141 | const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output, |
| 142 | const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f); |
| 143 | |
| 144 | // Inherited methods overridden: |
| 145 | void run() override; |
John Kesapides | 917959c | 2019-02-04 12:37:29 +0000 | [diff] [blame] | 146 | void prepare() override; |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 147 | |
| 148 | private: |
Michalis Spyrou | 6eb7345 | 2020-07-02 17:39:25 +0100 | [diff] [blame] | 149 | MemoryGroup _memory_group; |
| 150 | NEFullyConnectedLayer _fully_connected_input_gate; |
| 151 | NEArithmeticAddition _accum_input_gate1; |
| 152 | NEArithmeticSubtraction _subtract_input_gate; |
| 153 | NEPixelWiseMultiplication _pixelwise_mul_input_gate; |
| 154 | NEActivationLayer _activation_input_gate; |
| 155 | NEFullyConnectedLayer _fully_connected_forget_gate; |
| 156 | NEArithmeticAddition _accum_forget_gate1; |
| 157 | NEPixelWiseMultiplication _pixelwise_mul_forget_gate; |
| 158 | NEActivationLayer _activation_forget_gate; |
| 159 | NEFullyConnectedLayer _fully_connected_cell_state; |
| 160 | NEGEMM _gemm_cell_state1; |
| 161 | NETransposeKernel _transpose_cell_state; |
| 162 | NEArithmeticAddition _accum_cell_state1; |
| 163 | NEArithmeticAddition _accum_cell_state2; |
| 164 | NEPixelWiseMultiplication _pixelwise_mul_cell_state1; |
| 165 | NEActivationLayer _activation_cell_state; |
| 166 | NEActivationLayer _cell_clip; |
| 167 | NEPixelWiseMultiplication _pixelwise_mul_cell_state2; |
| 168 | NEFullyConnectedLayer _fully_connected_output; |
| 169 | NEPixelWiseMultiplication _pixelwise_mul_output_state1; |
| 170 | NEArithmeticAddition _accum_output1; |
| 171 | NEActivationLayer _activation_output; |
| 172 | NEActivationLayer _activation_output_state; |
| 173 | NEPixelWiseMultiplication _pixelwise_mul_output_state2; |
| 174 | NEFullyConnectedLayer _fully_connected_output_state; |
| 175 | NEActivationLayer _projection_clip; |
| 176 | NECopyKernel _copy_cell_state; |
| 177 | NECopyKernel _copy_output; |
| 178 | NEConcatenateLayer _concat_scratch_buffer; |
| 179 | NEConcatenateLayer _concat_inputs_forget_gate; |
| 180 | NEConcatenateLayer _concat_weights_forget_gate; |
| 181 | NEConcatenateLayer _concat_weights_input_gate; |
| 182 | NEConcatenateLayer _concat_weights_output; |
| 183 | NEMeanStdDevNormalizationLayer _mean_std_norm_input_gate; |
| 184 | NEPixelWiseMultiplication _pixelwise_mul_input_gate_coeff; |
| 185 | NEArithmeticAddition _accum_input_gate_bias; |
| 186 | NEMeanStdDevNormalizationLayer _mean_std_norm_forget_gate; |
| 187 | NEPixelWiseMultiplication _pixelwise_mul_forget_gate_coeff; |
| 188 | NEArithmeticAddition _accum_forget_gate_bias; |
| 189 | NEMeanStdDevNormalizationLayer _mean_std_norm_cell_gate; |
| 190 | NEPixelWiseMultiplication _pixelwise_mul_cell_gate_coeff; |
| 191 | NEArithmeticAddition _accum_cell_gate_bias; |
| 192 | NEMeanStdDevNormalizationLayer _mean_std_norm_output_gate; |
| 193 | NEPixelWiseMultiplication _pixelwise_mul_output_gate_coeff; |
| 194 | NEArithmeticAddition _accum_output_gate_bias; |
| 195 | Tensor _input_gate_out1; |
| 196 | Tensor _input_gate_out2; |
| 197 | Tensor _input_gate_out3; |
| 198 | Tensor _input_gate_out4; |
| 199 | Tensor _forget_gate_out1; |
| 200 | Tensor _forget_gate_out2; |
| 201 | Tensor _forget_gate_out3; |
| 202 | Tensor _forget_gate_out4; |
| 203 | Tensor _forget_gate_out5; |
| 204 | Tensor _forget_gate_out6; |
| 205 | Tensor _cell_state_out1; |
| 206 | Tensor _cell_state_out2; |
| 207 | Tensor _cell_state_out3; |
| 208 | Tensor _cell_state_out4; |
| 209 | Tensor _cell_state_out5; |
| 210 | Tensor _output1; |
| 211 | Tensor _output2; |
| 212 | Tensor _output3; |
| 213 | Tensor _output4; |
| 214 | Tensor _cell_state_activation; |
| 215 | Tensor _output_state1; |
| 216 | Tensor _ones; |
| 217 | Tensor _input_layer_norm_out1; |
| 218 | Tensor _input_layer_norm_out2; |
| 219 | Tensor _forget_layer_norm_out1; |
| 220 | Tensor _forget_layer_norm_out2; |
| 221 | Tensor _cell_layer_norm_out1; |
| 222 | Tensor _cell_layer_norm_out2; |
| 223 | Tensor _output_layer_norm_out1; |
| 224 | Tensor _output_layer_norm_out2; |
| 225 | bool _run_peephole_opt; |
| 226 | bool _run_cifg_opt; |
| 227 | bool _perform_cell_clipping; |
| 228 | bool _has_projection_weights; |
| 229 | bool _perform_projection_clipping; |
| 230 | bool _is_prepared; |
| 231 | bool _is_layer_norm_lstm; |
Michalis Spyrou | 25f45a4 | 2018-08-08 12:53:05 +0100 | [diff] [blame] | 232 | }; |
| 233 | } // namespace arm_compute |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 234 | #endif /* ARM_COMPUTE_NELSTMLAYER_H */ |