Michele Di Giorgio | 47a8990 | 2020-03-09 19:32:33 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2020 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 | #include "arm_compute/runtime/NEON/functions/NEQLSTMLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/KernelDescriptors.h" |
| 27 | #include "arm_compute/core/QuantizationInfo.h" |
| 28 | #include "arm_compute/core/Utils.h" |
| 29 | #include "arm_compute/core/Validate.h" |
| 30 | #include "arm_compute/core/utils/misc/InfoHelpers.h" |
| 31 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 32 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 33 | |
| 34 | namespace arm_compute |
| 35 | { |
| 36 | using namespace arm_compute::utils::info_helpers; |
| 37 | namespace |
| 38 | { |
| 39 | Status validate_mm(GEMMLowpOutputStageInfo &gemmlowp_info, const ITensorInfo *mm_input, const ITensorInfo *mm_weights, const ITensorInfo *bias, |
| 40 | float gemmlowp_scale, const TensorInfo *mm_res_info, const TensorInfo *outstage_tensor_info) |
| 41 | { |
| 42 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(mm_input, mm_weights, nullptr, mm_res_info)); |
| 43 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(gemmlowp_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); |
| 44 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOutputStage::validate(mm_res_info, bias, outstage_tensor_info, gemmlowp_info)); |
| 45 | return Status{}; |
| 46 | } |
| 47 | } // namespace |
| 48 | |
| 49 | NEQLSTMLayer::NEQLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager) |
| 50 | { |
| 51 | _memory_group = MemoryGroup(std::move(memory_manager)); |
| 52 | } |
| 53 | |
| 54 | void NEQLSTMLayer::configure_mm(NEGEMMLowpMatrixMultiplyCore &mm, NEGEMMLowpOutputStage &outstage, GEMMLowpOutputStageInfo &gemmlowp_info, |
| 55 | const ITensor *mm_input, const ITensor *mm_weights, const ITensor *bias, |
| 56 | Tensor *mm_res, Tensor *outstage_res, float gemmlowp_scale, |
| 57 | const TensorInfo &mm_res_info, const TensorInfo &outstage_tensor_info) |
| 58 | { |
| 59 | _memory_group.manage(mm_res); |
| 60 | _memory_group.manage(outstage_res); |
| 61 | |
| 62 | mm_res->allocator()->init(mm_res_info); |
| 63 | outstage_res->allocator()->init(outstage_tensor_info); |
| 64 | |
| 65 | // Configure matrix-multiplication |
| 66 | mm.configure(mm_input, mm_weights, nullptr, mm_res); |
| 67 | |
| 68 | // Configure output stage |
| 69 | quantization::calculate_quantized_multiplier(gemmlowp_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift); |
| 70 | outstage.configure(mm_res, bias, outstage_res, gemmlowp_info); |
| 71 | mm_res->allocator()->allocate(); |
| 72 | } |
| 73 | |
| 74 | void NEQLSTMLayer::configure(const ITensor *input, |
| 75 | const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights, |
| 76 | const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights, |
| 77 | const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias, |
| 78 | const ITensor *cell_state_in, const ITensor *output_state_in, |
| 79 | ITensor *cell_state_out, ITensor *output_state_out, |
| 80 | const LSTMParams<ITensor> &lstm_params) |
| 81 | { |
| 82 | ARM_COMPUTE_UNUSED(forget_gate_bias); |
| 83 | ARM_COMPUTE_UNUSED(cell_bias); |
| 84 | ARM_COMPUTE_UNUSED(output_gate_bias); |
| 85 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, |
| 86 | recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, |
| 87 | forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out); |
| 88 | |
| 89 | // Set lstm parameters |
| 90 | LSTMParams<ITensorInfo> lstm_params_info{}; |
| 91 | build_lstm_params_tensor_info(lstm_params, &lstm_params_info); |
| 92 | |
| 93 | // Validate |
| 94 | ARM_COMPUTE_ERROR_THROW_ON(NEQLSTMLayer::validate(input->info(), input_to_forget_weights->info(), input_to_cell_weights->info(), input_to_output_weights->info(), |
| 95 | recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(), |
| 96 | forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(), |
| 97 | cell_state_in->info(), output_state_in->info(), cell_state_out->info(), output_state_out->info(), lstm_params_info)); |
| 98 | |
| 99 | const int batch_size = input->info()->dimension(1); |
| 100 | const int num_units = input_to_output_weights->info()->dimension(1); |
| 101 | |
| 102 | const UniformQuantizationInfo qinput = input->info()->quantization_info().uniform(); |
| 103 | const UniformQuantizationInfo qcell_state_in = cell_state_in->info()->quantization_info().uniform(); |
| 104 | const UniformQuantizationInfo qoutput_state_in = output_state_in->info()->quantization_info().uniform(); |
| 105 | |
| 106 | _projection_bias = lstm_params.projection_bias(); |
| 107 | _input_to_forget_weights = input_to_forget_weights; |
| 108 | _input_to_cell_weights = input_to_cell_weights; |
| 109 | _input_to_output_weights = input_to_output_weights; |
| 110 | _recurrent_to_forget_weights = recurrent_to_forget_weights; |
| 111 | _recurrent_to_cell_weights = recurrent_to_cell_weights; |
| 112 | _recurrent_to_output_weights = recurrent_to_output_weights; |
| 113 | _projection_weights = lstm_params.projection_weights(); |
| 114 | |
| 115 | _has_cifg = lstm_params.has_cifg_opt(); |
| 116 | _has_projection = lstm_params.has_projection(); |
| 117 | _has_peephole = lstm_params.has_peephole_opt(); |
| 118 | |
| 119 | // Calculate and decompose effective scales for optimizing matmul calculation |
| 120 | const int32_t cell_shift = log2(qcell_state_in.scale); |
| 121 | |
| 122 | // Calculate quantized parameters for clipping. |
| 123 | int16_t quantized_cell_clip = 0; |
| 124 | if(lstm_params.cell_clip() > 0.0f) |
| 125 | { |
| 126 | quantized_cell_clip = quantize_qsymm16(lstm_params.cell_clip(), qcell_state_in); |
| 127 | } |
| 128 | _has_cell_clipping = quantized_cell_clip > 0; |
| 129 | |
| 130 | // Precompute effective bias for optimizing the matmul computations. |
| 131 | if(!_has_cifg) |
| 132 | { |
| 133 | _input_to_input_weights = lstm_params.input_to_input_weights(); |
| 134 | _recurrent_to_input_weights = lstm_params.recurrent_to_input_weights(); |
| 135 | |
| 136 | _input_to_input_reduction.configure(_input_to_input_weights, &_input_to_input_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)); |
| 137 | _recurrent_to_input_reduction.configure(_recurrent_to_input_weights, &_recurrent_to_input_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)); |
| 138 | } |
| 139 | _input_to_forget_reduction.configure(input_to_forget_weights, &_input_to_forget_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)); |
| 140 | _recurrent_to_forget_reduction.configure(recurrent_to_forget_weights, &_recurrent_to_forget_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)); |
| 141 | _input_to_cell_reduction.configure(input_to_cell_weights, &_input_to_cell_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)); |
| 142 | _recurrent_to_cell_reduction.configure(recurrent_to_cell_weights, &_recurrent_to_cell_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)); |
| 143 | _input_to_output_reduction.configure(input_to_output_weights, &_input_to_output_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)); |
| 144 | _recurrent_to_output_reduction.configure(recurrent_to_output_weights, &_recurrent_to_output_eff_bias, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)); |
| 145 | if(_projection_bias != nullptr) |
| 146 | { |
| 147 | _projection_reduction.configure(_projection_weights, &_projection_reduction_res, GEMMLowpReductionKernelInfo(num_units, false, lstm_params.hidden_state_zero(), true)); |
| 148 | _projection_bias_add.configure(_projection_bias, &_projection_reduction_res, &_projection_eff_bias, ConvertPolicy::SATURATE); |
| 149 | } |
| 150 | |
| 151 | // Pre-transpose weights to be used in GEMM. |
| 152 | _transpose_input_to_forget_weights.configure(input_to_forget_weights, &_input_to_forget_weights_transposed); |
| 153 | _transpose_input_to_cell_weights.configure(input_to_cell_weights, &_input_to_cell_weights_transposed); |
| 154 | _transpose_input_to_output_weights.configure(input_to_output_weights, &_input_to_output_weights_transposed); |
| 155 | _transpose_recurrent_to_forget_weights.configure(recurrent_to_forget_weights, &_recurrent_to_forget_weights_transposed); |
| 156 | _transpose_recurrent_to_cell_weights.configure(recurrent_to_cell_weights, &_recurrent_to_cell_weights_transposed); |
| 157 | _transpose_recurrent_to_output_weights.configure(recurrent_to_output_weights, &_recurrent_to_output_weights_transposed); |
| 158 | if(!_has_cifg) |
| 159 | { |
| 160 | _transpose_input_to_input_weights.configure(lstm_params.input_to_input_weights(), &_input_to_input_weights_transposed); |
| 161 | _transpose_recurrent_to_input_weights.configure(lstm_params.recurrent_to_input_weights(), &_recurrent_to_input_weights_transposed); |
| 162 | } |
| 163 | if(_has_projection) |
| 164 | { |
| 165 | _transpose_projection_weights.configure(_projection_weights, &_projection_weights_transposed); |
| 166 | } |
| 167 | |
| 168 | GEMMLowpOutputStageInfo gemmlowp_info; |
| 169 | gemmlowp_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
| 170 | gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int16_t>::lowest(); |
| 171 | gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int16_t>::max(); |
| 172 | gemmlowp_info.output_data_type = DataType::QSYMM16; |
| 173 | |
| 174 | const TensorInfo mm_out_info(TensorShape(num_units, batch_size), 1, DataType::S32); |
| 175 | // Forget gate. |
| 176 | const TensorInfo forget_gate_outstage_info(mm_out_info.tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0)); |
| 177 | const float input_to_forget_scale = input_to_forget_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.forget_intermediate_scale(); |
| 178 | configure_mm(_mm_input_to_forget, _input_to_forget_outstage, gemmlowp_info, |
| 179 | input, &_input_to_forget_weights_transposed, &_input_to_forget_eff_bias, |
| 180 | &_mm_input_to_forget_res, &_input_to_forget_outstage_res, input_to_forget_scale, |
| 181 | mm_out_info, forget_gate_outstage_info); |
| 182 | |
| 183 | const float recurrent_to_forget_scale = recurrent_to_forget_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.forget_intermediate_scale(); |
| 184 | configure_mm(_mm_recurrent_to_forget, _recurrent_to_forget_outstage, gemmlowp_info, |
| 185 | output_state_in, &_recurrent_to_forget_weights_transposed, &_recurrent_to_forget_eff_bias, |
| 186 | &_mm_recurrent_to_forget_res, &_recurrent_to_forget_outstage_res, recurrent_to_forget_scale, |
| 187 | mm_out_info, forget_gate_outstage_info); |
| 188 | |
| 189 | _accumulate_input_recurrent_forget.configure(&_input_to_forget_outstage_res, &_recurrent_to_forget_outstage_res, &_recurrent_to_forget_outstage_res, ConvertPolicy::SATURATE); |
| 190 | _input_to_forget_outstage_res.allocator()->allocate(); |
| 191 | |
| 192 | if(_has_peephole) |
| 193 | { |
| 194 | _memory_group.manage(&_mul_cell_to_forget_res); |
| 195 | _pixelwise_mul_cell_to_forget.configure(cell_state_in, lstm_params.cell_to_forget_weights(), &_mul_cell_to_forget_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); |
| 196 | _cell_to_forget_outstage_res.allocator()->init(TensorInfo(_mul_cell_to_forget_res.info()->tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0))); |
| 197 | _memory_group.manage(&_cell_to_forget_outstage_res); |
| 198 | const float cell_to_forget_scale = std::pow(2, cell_shift) * lstm_params.cell_to_forget_weights()->info()->quantization_info().uniform().scale / lstm_params.forget_intermediate_scale(); |
| 199 | quantization::calculate_quantized_multiplier(cell_to_forget_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift); |
| 200 | _cell_to_forget_outstage.configure(&_mul_cell_to_forget_res, nullptr, &_cell_to_forget_outstage_res, gemmlowp_info); |
| 201 | _mul_cell_to_forget_res.allocator()->allocate(); |
| 202 | _accumulate_cell_forget.configure(&_recurrent_to_forget_outstage_res, &_cell_to_forget_outstage_res, &_recurrent_to_forget_outstage_res, ConvertPolicy::SATURATE); |
| 203 | _cell_to_forget_outstage_res.allocator()->allocate(); |
| 204 | } |
| 205 | |
| 206 | // Output quantization info of Sigmoid and Tanh activations |
| 207 | const QuantizationInfo sigmoid_tanh_outqinfo(1.f / 32768.f, 0); |
| 208 | |
| 209 | const TensorInfo forget_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 210 | _memory_group.manage(&_forget_gate); |
| 211 | _forget_gate.allocator()->init(forget_gate_info); |
| 212 | _forget_gate_sigmoid.configure(&_recurrent_to_forget_outstage_res, &_forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); |
| 213 | _recurrent_to_forget_outstage_res.allocator()->allocate(); |
| 214 | |
| 215 | // Modulation gate. |
| 216 | const TensorInfo cell_outstage_info(mm_out_info.tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.cell_intermediate_scale(), 0)); |
| 217 | const float input_to_cell_scale = input_to_cell_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.cell_intermediate_scale(); |
| 218 | configure_mm(_mm_input_to_cell, _input_to_cell_outstage, gemmlowp_info, |
| 219 | input, &_input_to_cell_weights_transposed, &_input_to_cell_eff_bias, |
| 220 | &_mm_input_to_cell_res, &_input_to_cell_outstage_res, input_to_cell_scale, |
| 221 | mm_out_info, cell_outstage_info); |
| 222 | |
| 223 | const float recurrent_to_cell_scale = recurrent_to_cell_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.cell_intermediate_scale(); |
| 224 | configure_mm(_mm_recurrent_to_cell, _recurrent_to_cell_outstage, gemmlowp_info, |
| 225 | output_state_in, &_recurrent_to_cell_weights_transposed, &_recurrent_to_cell_eff_bias, |
| 226 | &_mm_recurrent_to_cell_res, &_recurrent_to_cell_outstage_res, recurrent_to_cell_scale, |
| 227 | mm_out_info, cell_outstage_info); |
| 228 | |
| 229 | _accumulate_input_recurrent_modulation.configure(&_input_to_cell_outstage_res, &_recurrent_to_cell_outstage_res, &_recurrent_to_cell_outstage_res, ConvertPolicy::SATURATE); |
| 230 | _input_to_cell_outstage_res.allocator()->allocate(); |
| 231 | |
| 232 | const TensorInfo cell_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 233 | _memory_group.manage(&_cell_gate); |
| 234 | _cell_gate.allocator()->init(cell_gate_info); |
| 235 | _cell_gate_tanh.configure(&_recurrent_to_cell_outstage_res, &_cell_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); |
| 236 | _recurrent_to_cell_outstage_res.allocator()->allocate(); |
| 237 | |
| 238 | // Input gate. |
| 239 | const TensorInfo input_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 240 | _input_gate.allocator()->init(input_gate_info); |
| 241 | _memory_group.manage(&_input_gate); |
| 242 | if(_has_cifg) |
| 243 | { |
| 244 | _ones.allocator()->init(*_forget_gate.info()); |
| 245 | _input_gate_sub.configure(&_ones, &_forget_gate, &_input_gate, ConvertPolicy::SATURATE); |
| 246 | _ones.allocator()->allocate(); |
| 247 | } |
| 248 | else |
| 249 | { |
| 250 | const TensorInfo input_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0)); |
| 251 | const float input_to_input_scale = _input_to_input_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.input_intermediate_scale(); |
| 252 | configure_mm(_mm_input_to_input, _input_to_input_outstage, gemmlowp_info, |
| 253 | input, &_input_to_input_weights_transposed, &_input_to_input_eff_bias, |
| 254 | &_mm_input_to_input_res, &_input_to_input_outstage_res, input_to_input_scale, |
| 255 | mm_out_info, input_outstage_info); |
| 256 | |
| 257 | const float recurrent_to_input_scale = _recurrent_to_input_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.input_intermediate_scale(); |
| 258 | configure_mm(_mm_recurrent_to_input, _recurrent_to_input_outstage, gemmlowp_info, |
| 259 | input, &_recurrent_to_input_weights_transposed, &_recurrent_to_input_eff_bias, |
| 260 | &_mm_recurrent_to_input_res, &_recurrent_to_input_outstage_res, recurrent_to_input_scale, |
| 261 | mm_out_info, input_outstage_info); |
| 262 | _accumulate_input_recurrent_input.configure(&_input_to_input_outstage_res, &_recurrent_to_input_outstage_res, &_recurrent_to_input_outstage_res, ConvertPolicy::SATURATE); |
| 263 | _input_to_input_outstage_res.allocator()->allocate(); |
| 264 | |
| 265 | if(_has_peephole) |
| 266 | { |
| 267 | _memory_group.manage(&_mul_cell_to_input_res); |
| 268 | _pixelwise_mul_cell_to_input.configure(cell_state_in, lstm_params.cell_to_input_weights(), &_mul_cell_to_input_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); |
| 269 | const float cell_to_input_scale = std::pow(2, cell_shift) * lstm_params.cell_to_input_weights()->info()->quantization_info().uniform().scale / lstm_params.input_intermediate_scale(); |
| 270 | quantization::calculate_quantized_multiplier(cell_to_input_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift); |
| 271 | _cell_to_input_outstage_res.allocator()->init(TensorInfo(_mul_cell_to_input_res.info()->tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0))); |
| 272 | _memory_group.manage(&_cell_to_input_outstage_res); |
| 273 | _cell_to_input_outstage.configure(&_mul_cell_to_input_res, nullptr, &_cell_to_input_outstage_res, gemmlowp_info); |
| 274 | _mul_cell_to_input_res.allocator()->allocate(); |
| 275 | _accumulate_cell_input.configure(&_recurrent_to_input_outstage_res, &_cell_to_input_outstage_res, &_recurrent_to_input_outstage_res, ConvertPolicy::SATURATE); |
| 276 | _cell_to_input_outstage_res.allocator()->allocate(); |
| 277 | } |
| 278 | |
| 279 | _input_gate_tanh.configure(&_recurrent_to_input_outstage_res, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); |
| 280 | _recurrent_to_input_outstage_res.allocator()->allocate(); |
| 281 | } |
| 282 | // Cell. |
| 283 | // TODO(COMPMID-3395): Perform multiplication in the quantized domain in NEPixelWiseMultiplicationKernel |
| 284 | _pixelwise_mul_forget_cell.configure(&_forget_gate, cell_state_in, &_forget_gate, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); |
| 285 | const float cell_gate_scale = _cell_gate.info()->quantization_info().uniform().scale; |
| 286 | const float mul_input_cell_scale = cell_gate_scale * std::pow(2, 15 + cell_shift); |
| 287 | const TensorInfo mul_input_cell_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(mul_input_cell_scale, 0)); |
| 288 | _memory_group.manage(&_mul_input_cell_res); |
| 289 | _mul_input_cell_res.allocator()->init(mul_input_cell_info); |
| 290 | _pixelwise_mul_input_cell.configure(&_input_gate, &_cell_gate, &_mul_input_cell_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); |
| 291 | _cell_gate.allocator()->allocate(); |
| 292 | _add_forget_cell.configure(&_forget_gate, &_mul_input_cell_res, cell_state_out, ConvertPolicy::SATURATE); |
| 293 | _mul_input_cell_res.allocator()->allocate(); |
| 294 | _forget_gate.allocator()->allocate(); |
| 295 | if(_has_cell_clipping) |
| 296 | { |
| 297 | _cell_clip.configure(cell_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_cell_clip, quantized_cell_clip)); |
| 298 | } |
| 299 | // Output gate. |
| 300 | const TensorInfo output_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.output_intermediate_scale(), 0)); |
| 301 | const float input_to_output_scale = input_to_output_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.output_intermediate_scale(); |
| 302 | configure_mm(_mm_input_to_output, _input_to_output_outstage, gemmlowp_info, |
| 303 | input, &_input_to_output_weights_transposed, &_input_to_output_eff_bias, |
| 304 | &_mm_input_to_output_res, &_input_to_output_outstage_res, input_to_output_scale, |
| 305 | mm_out_info, output_outstage_info); |
| 306 | |
| 307 | const float recurrent_to_output_scale = recurrent_to_output_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.output_intermediate_scale(); |
| 308 | configure_mm(_mm_recurrent_to_output, _recurrent_to_output_outstage, gemmlowp_info, |
| 309 | output_state_in, &_recurrent_to_output_weights_transposed, &_recurrent_to_output_eff_bias, |
| 310 | &_mm_recurrent_to_output_res, &_recurrent_to_output_outstage_res, recurrent_to_output_scale, |
| 311 | mm_out_info, output_outstage_info); |
| 312 | |
| 313 | _accumulate_input_recurrent_output.configure(&_recurrent_to_output_outstage_res, &_input_to_output_outstage_res, &_recurrent_to_output_outstage_res, ConvertPolicy::SATURATE); |
| 314 | _input_to_output_outstage_res.allocator()->allocate(); |
| 315 | |
| 316 | if(_has_peephole) |
| 317 | { |
| 318 | // TODO(COMPMID-3395): Perform multiplication in the quantized domain in NEPixelWiseMultiplicationKernel |
| 319 | // Here we are not using the output stage because all operations are done in float |
| 320 | // const float cell_to_output_scale = std::pow(2, cell_shift) * lstm_params.cell_to_output_weights()->info()->quantization_info().uniform().scale / lstm_params.output_intermediate_scale(); |
| 321 | // quantization::calculate_quantized_multiplier(cell_to_output_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift); |
| 322 | _memory_group.manage(&_mul_cell_to_output_res); |
| 323 | _pixelwise_mul_cell_to_output.configure(cell_state_out, lstm_params.cell_to_output_weights(), &_mul_cell_to_output_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); |
| 324 | _accumulate_cell_to_output.configure(&_recurrent_to_output_outstage_res, &_mul_cell_to_output_res, &_recurrent_to_output_outstage_res, ConvertPolicy::SATURATE); |
| 325 | _mul_cell_to_output_res.allocator()->allocate(); |
| 326 | } |
| 327 | |
| 328 | const TensorInfo output_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 329 | _memory_group.manage(&_output_gate); |
| 330 | _output_gate.allocator()->init(output_gate_info); |
| 331 | _output_gate_sigmoid.configure(&_recurrent_to_output_outstage_res, &_output_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); |
| 332 | _recurrent_to_output_outstage_res.allocator()->allocate(); |
| 333 | |
| 334 | // Hidden. |
| 335 | _hidden_tanh.configure(cell_state_out, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); |
| 336 | // TODO(COMPMID-3395): Perform multiplication in the quantized domain in NEPixelWiseMultiplicationKernel |
| 337 | _memory_group.manage(&_hidden_mul_res); |
| 338 | const TensorInfo hidden_mul_res(_input_gate.info()->tensor_shape(), 1, DataType::S32); |
| 339 | _hidden_mul_res.allocator()->init(hidden_mul_res); |
| 340 | _pixelwise_mul_hidden.configure(&_output_gate, &_input_gate, &_hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); |
| 341 | _output_gate.allocator()->allocate(); |
| 342 | _input_gate.allocator()->allocate(); |
| 343 | const float hidden_state_scale = std::pow(2, -15) / lstm_params.hidden_state_scale() * std::pow(2, -15); |
| 344 | quantization::calculate_quantized_multiplier(hidden_state_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift); |
| 345 | gemmlowp_info.gemmlowp_offset = lstm_params.hidden_state_zero(); |
| 346 | gemmlowp_info.output_data_type = output_state_in->info()->data_type(); |
| 347 | _hidden_outstage.configure(&_hidden_mul_res, nullptr, output_state_out, gemmlowp_info); |
| 348 | _hidden_mul_res.allocator()->allocate(); |
| 349 | |
| 350 | // Projection. |
| 351 | if(_has_projection) |
| 352 | { |
| 353 | const TensorInfo projection_outstage_info(*output_state_out->info()); |
| 354 | const UniformQuantizationInfo qprojection = _projection_weights->info()->quantization_info().uniform(); |
| 355 | const float projection_scale = qprojection.scale * lstm_params.hidden_state_scale() / qoutput_state_in.scale; |
| 356 | gemmlowp_info.gemmlowp_offset = qoutput_state_in.offset; |
| 357 | gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int8_t>::lowest(); |
| 358 | gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int8_t>::max(); |
| 359 | gemmlowp_info.output_data_type = DataType::QASYMM8_SIGNED; |
| 360 | |
| 361 | configure_mm(_mm_projection, _projection_outstage, gemmlowp_info, |
| 362 | output_state_out, &_projection_weights_transposed, &_projection_eff_bias, |
| 363 | &_mm_projection_res, &_projection_outstage_res, projection_scale, |
| 364 | mm_out_info, projection_outstage_info); |
| 365 | |
| 366 | _accumulate_projection.configure(&_projection_outstage_res, output_state_out, output_state_out, ConvertPolicy::SATURATE); |
| 367 | _projection_outstage_res.allocator()->allocate(); |
| 368 | |
| 369 | int8_t quantized_projection_clip{ 0 }; |
| 370 | if(lstm_params.projection_clip() > 0.0f) |
| 371 | { |
| 372 | quantized_projection_clip = utility::clamp<int8_t>(lstm_params.projection_clip() / qprojection.scale, -128, 127); |
| 373 | } |
| 374 | |
| 375 | if(quantized_projection_clip > 0) |
| 376 | { |
| 377 | _projection_clip.configure(output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_projection_clip, quantized_projection_clip)); |
| 378 | _has_projection_clipping = true; |
| 379 | } |
| 380 | } |
| 381 | } |
| 382 | |
| 383 | Status NEQLSTMLayer::validate(const ITensorInfo *input, |
| 384 | const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights, |
| 385 | const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights, |
| 386 | const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias, |
| 387 | const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in, |
| 388 | const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out, |
| 389 | const LSTMParams<ITensorInfo> &lstm_params) |
| 390 | { |
| 391 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, |
| 392 | recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out); |
| 393 | |
| 394 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED); |
| 395 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() != 2, "Input must have exactly 2 dimensions"); |
| 396 | |
| 397 | const unsigned int input_size = input->dimension(0); |
| 398 | const unsigned int batch_size = input->dimension(1); |
| 399 | const unsigned int num_units = input_to_output_weights->dimension(1); |
| 400 | const unsigned int output_size = recurrent_to_output_weights->dimension(0); |
| 401 | |
| 402 | ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->num_dimensions() != 2); |
| 403 | ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->dimension(0) != input_size); |
| 404 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_to_output_weights, input_to_forget_weights, input_to_cell_weights); |
| 405 | ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->num_dimensions() != 2); |
| 406 | ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->dimension(1) != num_units); |
| 407 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(recurrent_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights); |
| 408 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_to_forget_weights, 1, DataType::QSYMM8); |
| 409 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_to_forget_weights, input_to_cell_weights, input_to_output_weights, |
| 410 | recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights); |
| 411 | |
| 412 | ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() != 1); |
| 413 | ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->dimension(0) != num_units); |
| 414 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(forget_gate_bias, cell_bias, output_gate_bias); |
| 415 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(forget_gate_bias, 1, DataType::S32); |
| 416 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, cell_bias, output_gate_bias); |
| 417 | |
| 418 | ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() != 2); |
| 419 | ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->dimension(0) != num_units); |
| 420 | ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->dimension(1) != batch_size); |
| 421 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(cell_state_in, 1, DataType::QSYMM16); |
| 422 | |
| 423 | ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() != 2); |
| 424 | ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->dimension(0) != output_size); |
| 425 | ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->dimension(1) != batch_size); |
| 426 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output_state_in); |
| 427 | |
| 428 | // Check whether peephole weights are all there or none |
| 429 | if(lstm_params.has_peephole_opt()) |
| 430 | { |
| 431 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights()); |
| 432 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_forget_weights(), 1, DataType::QSYMM16); |
| 433 | ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->num_dimensions() != 1); |
| 434 | ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->dimension(0) != num_units); |
| 435 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights()); |
| 436 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights()); |
| 437 | |
| 438 | if(!lstm_params.has_cifg_opt()) |
| 439 | { |
| 440 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_input_weights()); |
| 441 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_input_weights()); |
| 442 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_input_weights()); |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | const UniformQuantizationInfo qinput = input->quantization_info().uniform(); |
| 447 | const UniformQuantizationInfo qcell_state_in = cell_state_in->quantization_info().uniform(); |
| 448 | const UniformQuantizationInfo qoutput_state_in = output_state_in->quantization_info().uniform(); |
| 449 | |
| 450 | // Calculate and decompose effective scales for optimizing matmul calculation |
| 451 | const int32_t cell_shift = log2(qcell_state_in.scale); |
| 452 | ARM_COMPUTE_RETURN_ERROR_ON(cell_shift > -9); |
| 453 | |
| 454 | // Calculate quantized parameters for clipping. |
| 455 | int16_t quantized_cell_clip = 0; |
| 456 | if(lstm_params.cell_clip() > 0.0f) |
| 457 | { |
| 458 | quantized_cell_clip = quantize_qsymm16(lstm_params.cell_clip(), qcell_state_in); |
| 459 | } |
| 460 | |
| 461 | // Precompute effective bias for optimizing the matmul computations. |
| 462 | const TensorInfo eff_bias_info(TensorShape(num_units), 1, DataType::S32); |
| 463 | if(!lstm_params.has_cifg_opt()) |
| 464 | { |
| 465 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(lstm_params.input_to_input_weights(), &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true))); |
| 466 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(lstm_params.recurrent_to_input_weights(), &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, |
| 467 | true))); |
| 468 | } |
| 469 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(input_to_forget_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true))); |
| 470 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(recurrent_to_forget_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true))); |
| 471 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(input_to_cell_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true))); |
| 472 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(recurrent_to_cell_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true))); |
| 473 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(input_to_output_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true))); |
| 474 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(recurrent_to_output_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true))); |
| 475 | if(lstm_params.projection_bias() != nullptr) |
| 476 | { |
| 477 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(lstm_params.projection_weights(), &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, lstm_params.hidden_state_zero(), |
| 478 | true))); |
| 479 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(lstm_params.projection_bias(), &eff_bias_info, &eff_bias_info, ConvertPolicy::SATURATE)); |
| 480 | } |
| 481 | |
| 482 | const TensorInfo input_weights_transposed(TensorShape(num_units, input_size), 1, input_to_forget_weights->data_type(), input_to_forget_weights->quantization_info()); |
| 483 | const TensorInfo recurrent_weights_transposed(TensorShape(num_units, output_size), 1, recurrent_to_forget_weights->data_type(), recurrent_to_forget_weights->quantization_info()); |
| 484 | |
| 485 | // Validate weights transpose |
| 486 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(input_to_forget_weights, &input_weights_transposed)); |
| 487 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(input_to_cell_weights, &input_weights_transposed)); |
| 488 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(input_to_output_weights, &input_weights_transposed)); |
| 489 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(recurrent_to_forget_weights, &recurrent_weights_transposed)); |
| 490 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(recurrent_to_cell_weights, &recurrent_weights_transposed)); |
| 491 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(recurrent_to_output_weights, &recurrent_weights_transposed)); |
| 492 | if(!lstm_params.has_cifg_opt()) |
| 493 | { |
| 494 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(lstm_params.input_to_input_weights(), &input_weights_transposed)); |
| 495 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(lstm_params.recurrent_to_input_weights(), &recurrent_weights_transposed)); |
| 496 | } |
| 497 | if(lstm_params.has_projection()) |
| 498 | { |
| 499 | ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(lstm_params.projection_weights(), &recurrent_weights_transposed)); |
| 500 | } |
| 501 | |
| 502 | GEMMLowpOutputStageInfo gemmlowp_info; |
| 503 | gemmlowp_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
| 504 | gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int16_t>::lowest(); |
| 505 | gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int16_t>::max(); |
| 506 | gemmlowp_info.output_data_type = DataType::QSYMM16; |
| 507 | |
| 508 | // Forget gate. |
| 509 | const TensorInfo forget_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0)); |
| 510 | const TensorInfo mm_out_info(TensorShape(num_units, batch_size), 1, DataType::S32); |
| 511 | const float input_to_forget_scale = input_to_forget_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.forget_intermediate_scale(); |
| 512 | validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_forget_scale, &mm_out_info, &forget_outstage_info); |
| 513 | |
| 514 | const float recurrent_to_forget_scale = recurrent_to_forget_weights->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.forget_intermediate_scale(); |
| 515 | validate_mm(gemmlowp_info, input, &recurrent_weights_transposed, &eff_bias_info, recurrent_to_forget_scale, &mm_out_info, &forget_outstage_info); |
| 516 | |
| 517 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&forget_outstage_info, &forget_outstage_info, &forget_outstage_info, ConvertPolicy::SATURATE)); |
| 518 | |
| 519 | if(lstm_params.has_peephole_opt()) |
| 520 | { |
| 521 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_forget_weights(), 1, DataType::QSYMM16); |
| 522 | ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE, |
| 523 | RoundingPolicy::TO_ZERO)); |
| 524 | const float cell_to_forget_scale = std::pow(2, cell_shift) * lstm_params.cell_to_forget_weights()->quantization_info().uniform().scale / lstm_params.forget_intermediate_scale(); |
| 525 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_forget_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); |
| 526 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOutputStage::validate(&mm_out_info, nullptr, &forget_outstage_info, gemmlowp_info)); |
| 527 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&forget_outstage_info, &forget_outstage_info, &forget_outstage_info, ConvertPolicy::SATURATE)); |
| 528 | } |
| 529 | |
| 530 | // Output quantization info of Sigmoid and Tanh activations |
| 531 | const QuantizationInfo sigmoid_tanh_outqinfo(1.f / 32768.f, 0); |
| 532 | |
| 533 | const TensorInfo forget_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 534 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&forget_outstage_info, &forget_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); |
| 535 | |
| 536 | // Modulation gate. |
| 537 | const TensorInfo cell_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.cell_intermediate_scale(), 0)); |
| 538 | const float input_to_cell_scale = input_to_cell_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.cell_intermediate_scale(); |
| 539 | validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_cell_scale, &mm_out_info, &cell_outstage_info); |
| 540 | |
| 541 | const float recurrent_to_cell_scale = recurrent_to_cell_weights->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.cell_intermediate_scale(); |
| 542 | validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, recurrent_to_cell_scale, &mm_out_info, &cell_outstage_info); |
| 543 | |
| 544 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&cell_outstage_info, &cell_outstage_info, &cell_outstage_info, ConvertPolicy::SATURATE)); |
| 545 | |
| 546 | const TensorInfo cell_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 547 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&cell_outstage_info, &cell_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f))); |
| 548 | |
| 549 | // Input gate. |
| 550 | const TensorInfo input_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 551 | if(lstm_params.has_cifg_opt()) |
| 552 | { |
| 553 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lstm_params.input_gate_bias() != nullptr, "Input gate bias must not be present when CIFG is used"); |
| 554 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticSubtractionKernel::validate(&input_gate_info, &forget_gate_info, &forget_gate_info, ConvertPolicy::SATURATE)); |
| 555 | } |
| 556 | else |
| 557 | { |
| 558 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), lstm_params.input_gate_bias()); |
| 559 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_to_forget_weights, lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights()); |
| 560 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_to_forget_weights, lstm_params.input_to_input_weights()); |
| 561 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(recurrent_to_forget_weights, lstm_params.recurrent_to_input_weights()); |
| 562 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.input_gate_bias()); |
| 563 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(forget_gate_bias, lstm_params.input_gate_bias()); |
| 564 | |
| 565 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(input, lstm_params.input_to_input_weights(), nullptr, &mm_out_info)); |
| 566 | const TensorInfo input_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0)); |
| 567 | const float input_to_input_scale = lstm_params.input_to_input_weights()->quantization_info().uniform().scale * qinput.scale / lstm_params.input_intermediate_scale(); |
| 568 | validate_mm(gemmlowp_info, input, lstm_params.input_to_input_weights(), &eff_bias_info, input_to_input_scale, &mm_out_info, &input_outstage_info); |
| 569 | |
| 570 | const float recurrent_to_input_scale = lstm_params.recurrent_to_input_weights()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.input_intermediate_scale(); |
| 571 | validate_mm(gemmlowp_info, input, lstm_params.recurrent_to_input_weights(), &eff_bias_info, recurrent_to_input_scale, &mm_out_info, &input_outstage_info); |
| 572 | |
| 573 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&input_outstage_info, &input_outstage_info, &input_outstage_info, ConvertPolicy::SATURATE)); |
| 574 | |
| 575 | if(lstm_params.has_peephole_opt()) |
| 576 | { |
| 577 | ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_outstage_info, 1.f, ConvertPolicy::SATURATE, |
| 578 | RoundingPolicy::TO_ZERO)); |
| 579 | const float cell_to_input_scale = std::pow(2, cell_shift) * lstm_params.cell_to_input_weights()->quantization_info().uniform().scale / lstm_params.input_intermediate_scale(); |
| 580 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_input_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); |
| 581 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOutputStage::validate(&input_outstage_info, &eff_bias_info, &input_outstage_info, gemmlowp_info)); |
| 582 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&input_outstage_info, &input_outstage_info, &input_outstage_info, ConvertPolicy::SATURATE)); |
| 583 | } |
| 584 | |
| 585 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&input_outstage_info, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f))); |
| 586 | } |
| 587 | // Cell. |
| 588 | ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&forget_gate_info, cell_state_in, &forget_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); |
| 589 | ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&input_gate_info, cell_state_in, &cell_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); |
| 590 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&forget_gate_info, &cell_gate_info, cell_state_out, ConvertPolicy::SATURATE)); |
| 591 | if(quantized_cell_clip > 0) |
| 592 | { |
| 593 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(cell_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_cell_clip, |
| 594 | quantized_cell_clip))); |
| 595 | } |
| 596 | // Output gate. |
| 597 | const TensorInfo output_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.output_intermediate_scale(), 0)); |
| 598 | const float input_to_output_scale = input_to_output_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.output_intermediate_scale(); |
| 599 | validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_output_scale, &mm_out_info, &output_outstage_info); |
| 600 | |
| 601 | const float recurrent_to_output_scale = recurrent_to_output_weights->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.output_intermediate_scale(); |
| 602 | validate_mm(gemmlowp_info, output_state_in, &recurrent_weights_transposed, &eff_bias_info, recurrent_to_output_scale, &mm_out_info, &output_outstage_info); |
| 603 | |
| 604 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&output_outstage_info, &output_outstage_info, &output_outstage_info, ConvertPolicy::SATURATE)); |
| 605 | if(lstm_params.has_peephole_opt()) |
| 606 | { |
| 607 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_output_weights(), 1, DataType::QSYMM16); |
| 608 | // TODO(COMPMID-3395): Perform multiplication in the quantized domain in NEPixelWiseMultiplicationKernel |
| 609 | // Here we are not using the output stage because all operations are done in float |
| 610 | // const float cell_to_output_scale = std::pow(2, cell_shift) * lstm_params.cell_to_output_weights()->quantization_info().uniform().scale / lstm_params.output_intermediate_scale(); |
| 611 | // ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_output_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); |
| 612 | ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(cell_state_out, lstm_params.cell_to_output_weights(), &output_outstage_info, 1.f, ConvertPolicy::SATURATE, |
| 613 | RoundingPolicy::TO_ZERO)); |
| 614 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&output_outstage_info, &output_outstage_info, &output_outstage_info, ConvertPolicy::SATURATE)); |
| 615 | } |
| 616 | |
| 617 | const TensorInfo output_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo); |
| 618 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&output_outstage_info, &output_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); |
| 619 | |
| 620 | // Hidden. |
| 621 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(cell_state_out, &input_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f))); |
| 622 | const TensorInfo hidden_mul_res(TensorShape(num_units, batch_size), 1, DataType::S32); |
| 623 | ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&output_gate_info, &input_gate_info, &hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); |
| 624 | const float hidden_state_scale = std::pow(2, -15) / lstm_params.hidden_state_scale() * std::pow(2, -15); |
| 625 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(hidden_state_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); |
| 626 | gemmlowp_info.gemmlowp_offset = lstm_params.hidden_state_zero(); |
| 627 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOutputStage::validate(&hidden_mul_res, nullptr, output_state_out, gemmlowp_info)); |
| 628 | |
| 629 | // Projection. |
| 630 | if(lstm_params.has_projection()) |
| 631 | { |
| 632 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_forget_weights, lstm_params.projection_weights()); |
| 633 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.projection_bias()); |
| 634 | |
| 635 | const UniformQuantizationInfo qprojection = lstm_params.projection_weights()->quantization_info().uniform(); |
| 636 | const float projection_scale = qprojection.scale * lstm_params.hidden_state_scale() / qoutput_state_in.scale; |
| 637 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(projection_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); |
| 638 | gemmlowp_info.gemmlowp_offset = qoutput_state_in.offset; |
| 639 | gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int8_t>::lowest(); |
| 640 | gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int8_t>::max(); |
| 641 | gemmlowp_info.output_data_type = DataType::QASYMM8_SIGNED; |
| 642 | |
| 643 | const TensorInfo projection_outstage_info(*output_state_out); |
| 644 | validate_mm(gemmlowp_info, output_state_out, &recurrent_weights_transposed, &eff_bias_info, input_to_output_scale, &mm_out_info, &projection_outstage_info); |
| 645 | |
| 646 | ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(output_state_out, output_state_out, output_state_out, ConvertPolicy::SATURATE)); |
| 647 | |
| 648 | int8_t quantized_projection_clip{ 0 }; |
| 649 | if(lstm_params.projection_clip() > 0.0f) |
| 650 | { |
| 651 | quantized_projection_clip = quantize_qasymm8_signed(lstm_params.projection_clip(), qprojection); |
| 652 | } |
| 653 | |
| 654 | if(quantized_projection_clip > 0) |
| 655 | { |
| 656 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_projection_clip, |
| 657 | quantized_projection_clip))); |
| 658 | } |
| 659 | } |
| 660 | |
| 661 | if(cell_state_out->total_size() > 0) |
| 662 | { |
| 663 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(cell_state_in, cell_state_out); |
| 664 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(cell_state_in, cell_state_out); |
| 665 | } |
| 666 | |
| 667 | if(output_state_out->total_size() > 0) |
| 668 | { |
| 669 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output_state_out); |
| 670 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output_state_in, output_state_out); |
| 671 | } |
| 672 | |
| 673 | return Status{}; |
| 674 | } |
| 675 | |
| 676 | void NEQLSTMLayer::run() |
| 677 | { |
| 678 | prepare(); |
| 679 | |
| 680 | // Acquire all the temporaries |
| 681 | MemoryGroupResourceScope scope_mg(_memory_group); |
| 682 | |
| 683 | // Forget gate. |
| 684 | _mm_input_to_forget.run(); |
| 685 | _input_to_forget_outstage.run(); |
| 686 | |
| 687 | _mm_recurrent_to_forget.run(); |
| 688 | _recurrent_to_forget_outstage.run(); |
| 689 | NEScheduler::get().schedule(&_accumulate_input_recurrent_forget, Window::DimY); |
| 690 | |
| 691 | if(_has_peephole) |
| 692 | { |
| 693 | NEScheduler::get().schedule(&_pixelwise_mul_cell_to_forget, Window::DimY); |
| 694 | _cell_to_forget_outstage.run(); |
| 695 | NEScheduler::get().schedule(&_accumulate_cell_forget, Window::DimY); |
| 696 | } |
| 697 | |
| 698 | _forget_gate_sigmoid.run(); |
| 699 | |
| 700 | // Modulation gate. |
| 701 | _mm_input_to_cell.run(); |
| 702 | _input_to_cell_outstage.run(); |
| 703 | |
| 704 | _mm_recurrent_to_cell.run(); |
| 705 | _recurrent_to_cell_outstage.run(); |
| 706 | NEScheduler::get().schedule(&_accumulate_input_recurrent_modulation, Window::DimY); |
| 707 | |
| 708 | _cell_gate_tanh.run(); |
| 709 | |
| 710 | // Input gate |
| 711 | if(_has_cifg) |
| 712 | { |
| 713 | NEScheduler::get().schedule(&_input_gate_sub, Window::DimY); |
| 714 | } |
| 715 | else |
| 716 | { |
| 717 | _mm_input_to_input.run(); |
| 718 | _input_to_input_outstage.run(); |
| 719 | _mm_recurrent_to_input.run(); |
| 720 | _recurrent_to_input_outstage.run(); |
| 721 | NEScheduler::get().schedule(&_accumulate_input_recurrent_input, Window::DimY); |
| 722 | |
| 723 | if(_has_peephole) |
| 724 | { |
| 725 | NEScheduler::get().schedule(&_pixelwise_mul_cell_to_input, Window::DimY); |
| 726 | _cell_to_input_outstage.run(); |
| 727 | NEScheduler::get().schedule(&_accumulate_cell_input, Window::DimY); |
| 728 | } |
| 729 | |
| 730 | _input_gate_tanh.run(); |
| 731 | } |
| 732 | |
| 733 | // Cell. |
| 734 | NEScheduler::get().schedule(&_pixelwise_mul_forget_cell, Window::DimY); |
| 735 | NEScheduler::get().schedule(&_pixelwise_mul_input_cell, Window::DimY); |
| 736 | NEScheduler::get().schedule(&_add_forget_cell, Window::DimY); |
| 737 | if(_has_cell_clipping) |
| 738 | { |
| 739 | _cell_clip.run(); |
| 740 | } |
| 741 | |
| 742 | // Output gate. |
| 743 | _mm_input_to_output.run(); |
| 744 | _input_to_output_outstage.run(); |
| 745 | _mm_recurrent_to_output.run(); |
| 746 | _recurrent_to_output_outstage.run(); |
| 747 | NEScheduler::get().schedule(&_accumulate_input_recurrent_output, Window::DimY); |
| 748 | if(_has_peephole) |
| 749 | { |
| 750 | NEScheduler::get().schedule(&_pixelwise_mul_cell_to_output, Window::DimY); |
| 751 | NEScheduler::get().schedule(&_accumulate_cell_to_output, Window::DimY); |
| 752 | } |
| 753 | |
| 754 | _output_gate_sigmoid.run(); |
| 755 | |
| 756 | // Hidden. |
| 757 | _hidden_tanh.run(); |
| 758 | NEScheduler::get().schedule(&_pixelwise_mul_hidden, Window::DimY); |
| 759 | _hidden_outstage.run(); |
| 760 | |
| 761 | // Projection. |
| 762 | if(_has_projection) |
| 763 | { |
| 764 | _mm_projection.run(); |
| 765 | _projection_outstage.run(); |
| 766 | NEScheduler::get().schedule(&_accumulate_projection, Window::DimY); |
| 767 | if(_has_projection_clipping) |
| 768 | { |
| 769 | _projection_clip.run(); |
| 770 | } |
| 771 | } |
| 772 | } |
| 773 | |
| 774 | void NEQLSTMLayer::prepare() |
| 775 | { |
| 776 | if(!_is_prepared) |
| 777 | { |
| 778 | // Pre-transpose weights to be used in GEMM. |
| 779 | _input_to_forget_weights_transposed.allocator()->allocate(); |
| 780 | _input_to_cell_weights_transposed.allocator()->allocate(); |
| 781 | _input_to_output_weights_transposed.allocator()->allocate(); |
| 782 | _recurrent_to_forget_weights_transposed.allocator()->allocate(); |
| 783 | _recurrent_to_cell_weights_transposed.allocator()->allocate(); |
| 784 | _recurrent_to_output_weights_transposed.allocator()->allocate(); |
| 785 | _transpose_input_to_forget_weights.run(); |
| 786 | _transpose_input_to_cell_weights.run(); |
| 787 | _transpose_input_to_output_weights.run(); |
| 788 | _transpose_recurrent_to_forget_weights.run(); |
| 789 | _transpose_recurrent_to_cell_weights.run(); |
| 790 | _transpose_recurrent_to_output_weights.run(); |
| 791 | |
| 792 | // Precompute effective biases |
| 793 | if(_has_cifg) |
| 794 | { |
| 795 | std::fill_n(reinterpret_cast<int16_t *>(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 32767); |
| 796 | } |
| 797 | else |
| 798 | { |
| 799 | _input_to_input_eff_bias.allocator()->allocate(); |
| 800 | _recurrent_to_input_eff_bias.allocator()->allocate(); |
| 801 | NEScheduler::get().schedule(&_input_to_input_reduction, Window::DimY); |
| 802 | NEScheduler::get().schedule(&_recurrent_to_input_reduction, Window::DimY); |
| 803 | |
| 804 | _input_to_input_weights_transposed.allocator()->allocate(); |
| 805 | _recurrent_to_input_weights_transposed.allocator()->allocate(); |
| 806 | _transpose_input_to_input_weights.run(); |
| 807 | _transpose_recurrent_to_input_weights.run(); |
| 808 | _input_to_input_weights->mark_as_unused(); |
| 809 | _recurrent_to_input_weights->mark_as_unused(); |
| 810 | } |
| 811 | _input_to_forget_eff_bias.allocator()->allocate(); |
| 812 | _recurrent_to_forget_eff_bias.allocator()->allocate(); |
| 813 | _input_to_cell_eff_bias.allocator()->allocate(); |
| 814 | _recurrent_to_cell_eff_bias.allocator()->allocate(); |
| 815 | _input_to_output_eff_bias.allocator()->allocate(); |
| 816 | _recurrent_to_output_eff_bias.allocator()->allocate(); |
| 817 | NEScheduler::get().schedule(&_input_to_forget_reduction, Window::DimY); |
| 818 | NEScheduler::get().schedule(&_recurrent_to_forget_reduction, Window::DimY); |
| 819 | NEScheduler::get().schedule(&_input_to_cell_reduction, Window::DimY); |
| 820 | NEScheduler::get().schedule(&_recurrent_to_cell_reduction, Window::DimY); |
| 821 | NEScheduler::get().schedule(&_input_to_output_reduction, Window::DimY); |
| 822 | NEScheduler::get().schedule(&_recurrent_to_output_reduction, Window::DimY); |
| 823 | |
| 824 | if(_has_projection) |
| 825 | { |
| 826 | if(_projection_bias != nullptr) |
| 827 | { |
| 828 | _projection_eff_bias.allocator()->allocate(); |
| 829 | NEScheduler::get().schedule(&_projection_reduction, Window::DimY); |
| 830 | _projection_bias->mark_as_unused(); |
| 831 | } |
| 832 | |
| 833 | _projection_weights_transposed.allocator()->allocate(); |
| 834 | _transpose_projection_weights.run(); |
| 835 | _projection_weights->mark_as_unused(); |
| 836 | } |
| 837 | |
| 838 | // Mark weights as unused |
| 839 | _input_to_forget_weights->mark_as_unused(); |
| 840 | _input_to_cell_weights->mark_as_unused(); |
| 841 | _input_to_output_weights->mark_as_unused(); |
| 842 | _recurrent_to_forget_weights->mark_as_unused(); |
| 843 | _recurrent_to_cell_weights->mark_as_unused(); |
| 844 | _recurrent_to_output_weights->mark_as_unused(); |
| 845 | |
| 846 | _is_prepared = true; |
| 847 | } |
| 848 | } |
| 849 | |
| 850 | } // namespace arm_compute |