Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2020 Arm Limited. |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +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_CLQLSTMLAYER_H |
| 25 | #define ARM_COMPUTE_CLQLSTMLAYER_H |
| 26 | |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/CL/kernels/CLCopyKernel.h" |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h" |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h" |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/runtime/CL/functions/CLActivationLayer.h" |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 32 | #include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h" |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 33 | #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" |
| 34 | #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 35 | #include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h" |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 36 | #include "arm_compute/runtime/CL/functions/CLTranspose.h" |
| 37 | |
| 38 | #include "arm_compute/runtime/common/LSTMParams.h" |
| 39 | |
| 40 | namespace arm_compute |
| 41 | { |
| 42 | // Forward declarations |
| 43 | class ICLTensor; |
| 44 | |
| 45 | /** Basic function to run @ref CLQLSTMLayer |
| 46 | * |
| 47 | * This function calls the following CL functions/kernels: |
| 48 | * |
| 49 | * -# @ref CLActivationLayer Activation functions (tanh and logistic) |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 50 | * -# @ref CLCopyKernel Copy kernel for copying output_state_out to output |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 51 | * -# @ref CLArithmeticAddition Elementwise addition and subtraction |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 52 | * -# @ref CLGEMMLowpMatrixMultiplyCore Quantized matrix multiplication core. Accumulators are 32-bit integers |
| 53 | * -# @ref CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint Convert 32-bit integers into QSYMM16 |
| 54 | * -# @ref CLGEMMLowpMatrixAReductionKernel For precomputing effective biases to use |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 55 | * -# @ref CLPixelWiseMultiplication Elementwise multiplication |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 56 | * -# @ref CLTranspose Transpose function for reshaping the weights |
| 57 | * */ |
| 58 | class CLQLSTMLayer : public IFunction |
| 59 | { |
| 60 | public: |
| 61 | /** Default constructor */ |
| 62 | CLQLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
| 63 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 64 | CLQLSTMLayer(const CLQLSTMLayer &) = delete; |
| 65 | /** Default move constructor */ |
| 66 | CLQLSTMLayer(CLQLSTMLayer &&) = default; |
| 67 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 68 | CLQLSTMLayer &operator=(const CLQLSTMLayer &) = delete; |
| 69 | /** Default move assignment operator */ |
| 70 | CLQLSTMLayer &operator=(CLQLSTMLayer &&) = default; |
| 71 | /** Initialize function's tensors. |
| 72 | * |
| 73 | * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED. |
| 74 | * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 75 | * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 76 | * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 77 | * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 78 | * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 79 | * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 80 | * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 81 | * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 82 | * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 83 | * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: QSYMM16. |
| 84 | * @param[in] output_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input. |
| 85 | * @param[out] cell_state_out Destination tensor. Output is a 2D tensor with dimensions [num_units, batch_size]. Data type supported: QSYMM16. |
| 86 | * @param[out] output_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input. |
| 87 | * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input. |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 88 | * @param[in] lstm_params Weights tensors used in peephole, CIFG and layer normalization optimizations: |
| 89 | * input_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate. |
| 90 | * forget_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate. |
| 91 | * cell_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate. |
| 92 | * output_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate. |
| 93 | * hidden_state_zero The zero point of the hidden state. |
| 94 | * hidden_state_scale The scale of the hidden state. |
| 95 | * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 96 | * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 97 | * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16. |
| 98 | * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 99 | * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 100 | * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 101 | * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 102 | * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. S32. |
| 103 | * input_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 104 | * forget_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 105 | * cell_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 106 | * output_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 107 | * cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 108 | * If set to 0.0 then clipping is disabled. |
| 109 | * projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within |
| 110 | * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 111 | */ |
| 112 | void configure(const ICLTensor *input, |
| 113 | const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, |
| 114 | const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, |
| 115 | const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 116 | ICLTensor *cell_state_in, const ICLTensor *output_state_in, |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 117 | ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output, |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 118 | const LSTMParams<ICLTensor> &lstm_params); |
| 119 | |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 120 | /** Initialize function's tensors. |
| 121 | * |
| 122 | * @param[in] compile_context The compile context to be used. |
| 123 | * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED. |
| 124 | * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 125 | * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 126 | * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 127 | * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 128 | * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 129 | * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 130 | * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 131 | * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 132 | * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 133 | * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: QSYMM16. |
| 134 | * @param[in] output_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input. |
| 135 | * @param[out] cell_state_out Destination tensor. Output is a 2D tensor with dimensions [num_units, batch_size]. Data type supported: QSYMM16. |
| 136 | * @param[out] output_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input. |
| 137 | * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input. |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 138 | * @param[in] lstm_params Weights tensors used in peephole, CIFG and layer normalization optimizations: |
| 139 | * input_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate. |
| 140 | * forget_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate. |
| 141 | * cell_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate. |
| 142 | * output_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate. |
| 143 | * hidden_state_zero The zero point of the hidden state. |
| 144 | * hidden_state_scale The scale of the hidden state. |
| 145 | * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 146 | * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 147 | * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16. |
| 148 | * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 149 | * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 150 | * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 151 | * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 152 | * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. S32. |
| 153 | * input_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 154 | * forget_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 155 | * cell_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 156 | * output_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 157 | * cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 158 | * If set to 0.0 then clipping is disabled. |
| 159 | * projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within |
| 160 | * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 161 | */ |
| 162 | void configure(const CLCompileContext &compile_context, const ICLTensor *input, |
| 163 | const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, |
| 164 | const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, |
| 165 | const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 166 | ICLTensor *cell_state_in, const ICLTensor *output_state_in, |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 167 | ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output, |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 168 | const LSTMParams<ICLTensor> &lstm_params); |
| 169 | |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 170 | /** Static function to check if given info will lead to a valid configuration of @ref CLQLSTMLayer |
| 171 | * |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 172 | * @param[in] input Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED. |
| 173 | * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 174 | * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 175 | * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 176 | * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 177 | * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 178 | * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 179 | * @param[in] forget_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: S32. |
| 180 | * @param[in] cell_bias 1D weights tensor info with dimensions [num_units]. Data type supported: S32. |
| 181 | * @param[in] output_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: S32. |
| 182 | * @param[in] cell_state_in 2D tensor info with dimensions [num_units, batch_size]. Data type supported: QSYMM16. |
| 183 | * @param[in] output_state_in 2D tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input. |
| 184 | * @param[in] cell_state_out Destination tensor info. Output is a 2D tensor info with dimensions [num_units, batch_size]. Data type supported: QSYMM16. |
| 185 | * @param[in] output_state_out Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size].Data types supported: Same as @p input. |
| 186 | * @param[in] output Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size].Data types supported: Same as @p input. |
| 187 | * @param[in] lstm_params Weights tensors info used in peephole, CIFG and layer normalization optimizations: |
| 188 | * input_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate. |
| 189 | * forget_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate. |
| 190 | * cell_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate. |
| 191 | * output_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate. |
| 192 | * hidden_state_zero The zero point of the hidden state. |
| 193 | * hidden_state_scale The scale of the hidden state. |
| 194 | * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8. |
| 195 | * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 196 | * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16. |
| 197 | * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 198 | * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 199 | * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32. |
| 200 | * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8. |
| 201 | * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. S32. |
| 202 | * input_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 203 | * forget_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 204 | * cell_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 205 | * output_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16. |
| 206 | * cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 207 | * If set to 0.0 then clipping is disabled. |
| 208 | * projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within |
| 209 | * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 210 | * @return a status |
| 211 | */ |
| 212 | static Status validate(const ITensorInfo *input, |
| 213 | const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights, |
| 214 | const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights, |
| 215 | const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias, |
| 216 | const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in, |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 217 | const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out, const ITensorInfo *output, |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 218 | const LSTMParams<ITensorInfo> &lstm_params); |
| 219 | |
| 220 | // Inherited methods overridden: |
| 221 | void run() override; |
| 222 | void prepare() override; |
| 223 | |
| 224 | private: |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 225 | enum class LayerNormGate : uint8_t |
| 226 | { |
| 227 | Forget, |
| 228 | Cell, |
| 229 | Input, |
| 230 | Output, |
| 231 | Count |
| 232 | }; |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 233 | static constexpr uint8_t _layer_norm_count = static_cast<uint8_t>(LayerNormGate::Count); |
| 234 | static constexpr uint32_t _out_state_output_size_dimension_idx = 0; |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 235 | |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 236 | /** Internal method to configure matrix multiplication plus output stage of each gate. |
| 237 | * |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 238 | * @param[in] compile_context The compile context to be used. |
| 239 | * @param[in] mm Matrix multiplication function to use. |
| 240 | * @param[in] outstage Output stage function to use. |
| 241 | * @param[in] gemmlowp_info GEMMLowp metadata to be used by the output stage. |
| 242 | * @param[in] mm_input Input tensor to matrix multiplication function. |
| 243 | * @param[in] mm_weights Weights tensor to matrix multiplication function. |
| 244 | * @param[in] bias Bias tensor to matrix multiplication function. |
| 245 | * @param[in] outstage_res Tensor to be used for storing the result of the output stage. |
| 246 | * @param[in] gemmlowp_scale Real multiplier to be used computing multiplier and shift for requantization. |
| 247 | * @param[in] mm_res_info Tensor info to be used to initialize matrix multiplication result tensor. |
| 248 | * @param[in] mm_res_info Tensor info to be used to initialize output stage result tensor. |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 249 | * |
| 250 | */ |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 251 | void configure_mm(const CLCompileContext &compile_context, CLGEMMLowpMatrixMultiplyCore &mm, CLGEMMLowpOutputStage &outstage, GEMMLowpOutputStageInfo &gemmlowp_info, |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 252 | const ICLTensor *mm_input, const ICLTensor *mm_weights, const ICLTensor *bias, CLTensor *mm_res, |
| 253 | CLTensor *outstage_res, float gemmlowp_scale, |
| 254 | const TensorInfo &mm_res_info, const TensorInfo &outstage_tensor_info); |
| 255 | |
| 256 | MemoryGroup _memory_group{}; |
| 257 | |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 258 | /** A small internel kernel do the copy between two tensors */ |
| 259 | class TensorCopyKernel |
| 260 | { |
| 261 | static constexpr uint32_t max_dimension_supported = 2; |
| 262 | |
| 263 | ICLTensor *_src{ nullptr }; |
| 264 | ICLTensor *_dst{ nullptr }; |
| 265 | size_t _row_size{}; |
| 266 | Window _window{}; |
| 267 | |
| 268 | public: |
| 269 | /** Static function to check if given info will lead to a valid configuration of @ref CLQLSTMLayer::TensorCopyKernel |
| 270 | * |
| 271 | * @param[in] src Source tensor info. |
| 272 | * @param[in] dst Destination tensor info |
| 273 | * |
| 274 | * @return a status |
| 275 | */ |
| 276 | static Status validate(const ITensorInfo &src, const ITensorInfo &dst); |
| 277 | /** Set the input and output tensors. |
| 278 | * |
| 279 | * @param[in] src Source tensor |
| 280 | * @param[out] dst Destination tensor |
| 281 | */ |
| 282 | void configure(ICLTensor &src, ICLTensor &dst); |
| 283 | /** run the kernel */ |
| 284 | void run(); |
| 285 | }; |
| 286 | |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 287 | // Functions used |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 288 | CLTranspose _transpose_input_to_forget_weights{}; |
| 289 | CLTranspose _transpose_input_to_cell_weights{}; |
| 290 | CLTranspose _transpose_input_to_output_weights{}; |
| 291 | CLTranspose _transpose_input_to_input_weights{}; |
| 292 | CLTranspose _transpose_recurrent_to_forget_weights{}; |
| 293 | CLTranspose _transpose_recurrent_to_cell_weights{}; |
| 294 | CLTranspose _transpose_recurrent_to_output_weights{}; |
| 295 | CLTranspose _transpose_recurrent_to_input_weights{}; |
| 296 | CLTranspose _transpose_projection_weights{}; |
| 297 | CLGEMMLowpMatrixAReductionKernel _input_to_input_reduction{}; |
| 298 | CLGEMMLowpMatrixAReductionKernel _recurrent_to_input_reduction{}; |
| 299 | CLGEMMLowpMatrixAReductionKernel _input_to_forget_reduction{}; |
| 300 | CLGEMMLowpMatrixAReductionKernel _recurrent_to_forget_reduction{}; |
| 301 | CLGEMMLowpMatrixAReductionKernel _input_to_cell_reduction{}; |
| 302 | CLGEMMLowpMatrixAReductionKernel _recurrent_to_cell_reduction{}; |
| 303 | CLGEMMLowpMatrixAReductionKernel _input_to_output_reduction{}; |
| 304 | CLGEMMLowpMatrixAReductionKernel _recurrent_to_output_reduction{}; |
| 305 | CLGEMMLowpMatrixAReductionKernel _projection_reduction{}; |
| 306 | CLArithmeticAddition _projection_bias_add{}; |
| 307 | CLGEMMLowpMatrixMultiplyCore _mm_input_to_forget{}; |
| 308 | CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_forget{}; |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 309 | CLPixelWiseMultiplication _pixelwise_mul_cell_to_forget{}; |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 310 | CLGEMMLowpOutputStage _input_to_forget_outstage{}; |
| 311 | CLGEMMLowpOutputStage _recurrent_to_forget_outstage{}; |
| 312 | CLGEMMLowpOutputStage _cell_to_forget_outstage{}; |
| 313 | CLArithmeticAddition _accumulate_input_recurrent_forget{}; |
| 314 | CLArithmeticAddition _accumulate_cell_forget{}; |
| 315 | CLActivationLayer _forget_gate_sigmoid{}; |
| 316 | CLGEMMLowpMatrixMultiplyCore _mm_input_to_cell{}; |
| 317 | CLGEMMLowpOutputStage _input_to_cell_outstage{}; |
| 318 | CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_cell{}; |
| 319 | CLGEMMLowpOutputStage _recurrent_to_cell_outstage{}; |
| 320 | CLArithmeticAddition _accumulate_input_recurrent_modulation{}; |
| 321 | CLActivationLayer _cell_gate_tanh{}; |
| 322 | CLArithmeticSubtraction _input_gate_sub{}; |
| 323 | CLGEMMLowpMatrixMultiplyCore _mm_input_to_input{}; |
| 324 | CLGEMMLowpOutputStage _input_to_input_outstage{}; |
| 325 | CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_input{}; |
| 326 | CLGEMMLowpOutputStage _recurrent_to_input_outstage{}; |
| 327 | CLArithmeticAddition _accumulate_input_recurrent_input{}; |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 328 | CLPixelWiseMultiplication _pixelwise_mul_cell_to_input{}; |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 329 | CLGEMMLowpOutputStage _cell_to_input_outstage{}; |
| 330 | CLArithmeticAddition _accumulate_cell_input{}; |
| 331 | CLActivationLayer _input_gate_sigmoid{}; |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 332 | CLPixelWiseMultiplication _pixelwise_mul_forget_cell{}; |
| 333 | CLPixelWiseMultiplication _pixelwise_mul_input_cell{}; |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 334 | CLArithmeticAddition _add_forget_cell{}; |
| 335 | CLActivationLayer _cell_clip{}; |
| 336 | CLGEMMLowpMatrixMultiplyCore _mm_input_to_output{}; |
| 337 | CLGEMMLowpOutputStage _input_to_output_outstage{}; |
| 338 | CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_output{}; |
| 339 | CLGEMMLowpOutputStage _recurrent_to_output_outstage{}; |
| 340 | CLArithmeticAddition _accumulate_input_recurrent_output{}; |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 341 | CLPixelWiseMultiplication _pixelwise_mul_cell_to_output{}; |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 342 | CLGEMMLowpOutputStage _cell_to_output_outstage{}; |
| 343 | CLArithmeticAddition _accumulate_cell_to_output{}; |
| 344 | CLActivationLayer _output_gate_sigmoid{}; |
| 345 | CLActivationLayer _hidden_tanh{}; |
Michalis Spyrou | 1009e87 | 2020-07-27 12:48:34 +0100 | [diff] [blame] | 346 | CLPixelWiseMultiplication _pixelwise_mul_hidden{}; |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 347 | CLGEMMLowpOutputStage _hidden_outstage{}; |
| 348 | CLGEMMLowpMatrixMultiplyCore _mm_projection{}; |
| 349 | CLGEMMLowpOutputStage _projection_outstage{}; |
| 350 | CLArithmeticAddition _accumulate_projection{}; |
| 351 | CLActivationLayer _projection_clip{}; |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 352 | std::array<CLQLSTMLayerNormalizationKernel, _layer_norm_count> _layer_norms{ {} }; |
Michele Di Giorgio | beb2d45 | 2020-05-11 16:17:51 +0100 | [diff] [blame] | 353 | CLCopyKernel _copy_output{}; |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 354 | |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 355 | TensorCopyKernel _projection_bias_copy{}; |
| 356 | TensorCopyKernel _projection_output_to_accumulate_copy{}; |
| 357 | TensorCopyKernel _projection_accumulate_to_output_copy{}; |
| 358 | TensorCopyKernel _hidden_to_output_copy{}; |
| 359 | |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 360 | // Tensor pointers |
Michalis Spyrou | ad7515d | 2020-07-24 00:02:23 +0100 | [diff] [blame] | 361 | const ICLTensor *_input_to_input_weights |
| 362 | { |
| 363 | nullptr |
| 364 | }; |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 365 | const ICLTensor *_recurrent_to_input_weights{ nullptr }; |
| 366 | const ICLTensor *_projection_bias{ nullptr }; |
| 367 | const ICLTensor *_input_to_forget_weights{ nullptr }; |
| 368 | const ICLTensor *_input_to_cell_weights{ nullptr }; |
| 369 | const ICLTensor *_input_to_output_weights{ nullptr }; |
| 370 | const ICLTensor *_recurrent_to_forget_weights{ nullptr }; |
| 371 | const ICLTensor *_recurrent_to_cell_weights{ nullptr }; |
| 372 | const ICLTensor *_recurrent_to_output_weights{ nullptr }; |
| 373 | const ICLTensor *_projection_weights{ nullptr }; |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 374 | std::array<const ICLTensor *, _layer_norm_count> _layer_norm_weights{ {} }; |
| 375 | std::array<const ICLTensor *, _layer_norm_count> _layer_norm_bias{ {} }; |
| 376 | |
| 377 | using LayerNormIndexType = typename std::underlying_type<LayerNormGate>::type; |
| 378 | inline LayerNormIndexType getGateIndex(LayerNormGate g) |
| 379 | { |
| 380 | return static_cast<LayerNormIndexType>(g); |
| 381 | } |
| 382 | |
| 383 | inline void set_layer_norm_weight(const ICLTensor *t, LayerNormGate g) |
| 384 | { |
| 385 | _layer_norm_weights[getGateIndex(g)] = t; |
| 386 | } |
| 387 | |
| 388 | inline void set_layer_norm_bias(const ICLTensor *t, LayerNormGate g) |
| 389 | { |
| 390 | _layer_norm_bias[getGateIndex(g)] = t; |
| 391 | } |
| 392 | |
| 393 | inline const ICLTensor *get_layer_norm_weight(LayerNormGate g) |
| 394 | { |
| 395 | return _layer_norm_weights[getGateIndex(g)]; |
| 396 | } |
| 397 | |
| 398 | inline const ICLTensor *get_layer_norm_bias(LayerNormGate g) |
| 399 | { |
| 400 | return _layer_norm_bias[getGateIndex(g)]; |
| 401 | } |
| 402 | |
| 403 | inline CLQLSTMLayerNormalizationKernel &get_layer_norm(LayerNormGate g) |
| 404 | { |
| 405 | return _layer_norms[getGateIndex(g)]; |
| 406 | } |
| 407 | |
| 408 | inline void configure_layer_norm(LayerNormGate g, const ICLTensor *in) |
| 409 | { |
| 410 | ARM_COMPUTE_ERROR_ON(!_has_layer_norm); |
| 411 | |
| 412 | CLTensor *out = &get_layer_norm_output(g); |
| 413 | _memory_group.manage(out); |
| 414 | out->allocator()->init(*(in->info())); |
| 415 | |
| 416 | get_layer_norm(g).configure(in, out, get_layer_norm_weight(g), get_layer_norm_bias(g)); |
| 417 | } |
| 418 | |
| 419 | inline static Status validate_layer_norm(const ITensorInfo &in, const ITensorInfo &weight, const ITensorInfo &bias) |
| 420 | { |
| 421 | // Output quantization scale will be different, but ignored here |
| 422 | // since it will be configured at configure() stage. |
| 423 | const TensorInfo out |
| 424 | { |
| 425 | in |
| 426 | }; |
| 427 | return CLQLSTMLayerNormalizationKernel::validate(&in, &out, &weight, &bias); |
| 428 | } |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 429 | |
| 430 | // Temporary tensors |
| 431 | CLTensor _input_to_forget_weights_transposed{ nullptr }; |
| 432 | CLTensor _input_to_cell_weights_transposed{ nullptr }; |
| 433 | CLTensor _input_to_output_weights_transposed{ nullptr }; |
| 434 | CLTensor _input_to_input_weights_transposed{ nullptr }; |
| 435 | CLTensor _recurrent_to_forget_weights_transposed{ nullptr }; |
| 436 | CLTensor _recurrent_to_cell_weights_transposed{ nullptr }; |
| 437 | CLTensor _recurrent_to_output_weights_transposed{ nullptr }; |
| 438 | CLTensor _recurrent_to_input_weights_transposed{ nullptr }; |
| 439 | CLTensor _projection_weights_transposed{ nullptr }; |
| 440 | CLTensor _input_to_input_eff_bias{ nullptr }; |
| 441 | CLTensor _recurrent_to_input_eff_bias{ nullptr }; |
| 442 | CLTensor _input_to_forget_eff_bias{ nullptr }; |
| 443 | CLTensor _recurrent_to_forget_eff_bias{ nullptr }; |
| 444 | CLTensor _input_to_cell_eff_bias{ nullptr }; |
| 445 | CLTensor _recurrent_to_cell_eff_bias{ nullptr }; |
| 446 | CLTensor _input_to_output_eff_bias{ nullptr }; |
| 447 | CLTensor _recurrent_to_output_eff_bias{ nullptr }; |
| 448 | CLTensor _projection_reduction_res{ nullptr }; |
| 449 | CLTensor _projection_eff_bias{ nullptr }; |
| 450 | CLTensor _mm_input_to_forget_res{ nullptr }; |
| 451 | CLTensor _mm_recurrent_to_forget_res{ nullptr }; |
| 452 | CLTensor _mul_cell_to_forget_res{ nullptr }; |
| 453 | CLTensor _input_to_forget_outstage_res{ nullptr }; |
| 454 | CLTensor _cell_to_forget_outstage_res{ nullptr }; |
| 455 | CLTensor _recurrent_to_forget_outstage_res{ nullptr }; |
| 456 | CLTensor _forget_gate{ nullptr }; |
| 457 | CLTensor _mm_input_to_cell_res{ nullptr }; |
| 458 | CLTensor _input_to_cell_outstage_res{ nullptr }; |
| 459 | CLTensor _mm_recurrent_to_cell_res{ nullptr }; |
| 460 | CLTensor _recurrent_to_cell_outstage_res{ nullptr }; |
| 461 | CLTensor _cell_gate{ nullptr }; |
| 462 | CLTensor _mul_input_cell_res{ nullptr }; |
| 463 | CLTensor _mm_input_to_input_res{ nullptr }; |
| 464 | CLTensor _input_to_input_outstage_res{ nullptr }; |
| 465 | CLTensor _mm_recurrent_to_input_res{ nullptr }; |
| 466 | CLTensor _mul_cell_to_input_res{ nullptr }; |
| 467 | CLTensor _cell_to_input_outstage_res{ nullptr }; |
| 468 | CLTensor _recurrent_to_input_outstage_res{ nullptr }; |
| 469 | CLTensor _input_gate{ nullptr }; |
| 470 | CLTensor _mm_input_to_output_res{ nullptr }; |
| 471 | CLTensor _input_to_output_outstage_res{ nullptr }; |
| 472 | CLTensor _mm_recurrent_to_output_res{ nullptr }; |
| 473 | CLTensor _mul_cell_to_output_res{ nullptr }; |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 474 | CLTensor _cell_to_output_outstage_res{ nullptr }; |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 475 | CLTensor _recurrent_to_output_outstage_res{ nullptr }; |
| 476 | CLTensor _output_gate{ nullptr }; |
| 477 | CLTensor _hidden_mul_res{ nullptr }; |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 478 | CLTensor _hidden_gate{ nullptr }; |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 479 | CLTensor _mm_projection_res{ nullptr }; |
| 480 | CLTensor _projection_outstage_res{ nullptr }; |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 481 | CLTensor _projection_out_res{ nullptr }; |
| 482 | CLTensor _projection_accumulate_res{ nullptr }; |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 483 | CLTensor _ones{ nullptr }; |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 484 | std::array<CLTensor, _layer_norm_count> _layer_norm_output{ {} }; |
| 485 | |
| 486 | inline CLTensor &get_layer_norm_output(LayerNormGate g) |
| 487 | { |
| 488 | return _layer_norm_output[getGateIndex(g)]; |
| 489 | } |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 490 | |
| 491 | bool _is_prepared{ false }; |
| 492 | bool _has_cifg{ false }; |
| 493 | bool _has_cell_clipping{ false }; |
| 494 | bool _has_projection{ false }; |
| 495 | bool _has_projection_clipping{ false }; |
| 496 | bool _has_peephole{ false }; |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 497 | bool _has_layer_norm{ false }; |
Sang-Hoon Park | a7431ae | 2020-05-12 11:13:30 +0100 | [diff] [blame] | 498 | bool _projection_tensor_copy_required{ false }; |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 499 | }; |
| 500 | } // namespace arm_compute |
| 501 | #endif /* ARM_COMPUTE_CLQLSTMLAYER_H */ |