Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | 807ce59 | 2020-01-03 14:39:37 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 ARM Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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 | #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h" |
| 25 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 26 | #include "arm_compute/core/Helpers.h" |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Size2D.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/Validate.h" |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 31 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 32 | |
| 33 | #include <algorithm> |
| 34 | #include <cmath> |
| 35 | |
Michele Di Giorgio | f29d1b7 | 2019-10-29 10:58:13 +0000 | [diff] [blame] | 36 | namespace arm_compute |
| 37 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 38 | using namespace arm_compute::misc::shape_calculator; |
| 39 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 40 | namespace |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | { |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 42 | // Get min, max bound of a quantized assymetric output tensor, with the effect of fused activation |
| 43 | std::pair<PixelValue, PixelValue> get_quantized_asymmetric_output_min_max(const QuantizationInfo &q_info, const ActivationLayerInfo &act_info, DataType data_type) |
| 44 | { |
| 45 | PixelValue type_min{}; |
| 46 | PixelValue type_max{}; |
| 47 | std::tie(type_min, type_max) = get_min_max(data_type); |
| 48 | const UniformQuantizationInfo q_unif = q_info.uniform(); |
| 49 | |
| 50 | if(act_info.enabled()) |
| 51 | { |
| 52 | switch(act_info.activation()) |
| 53 | { |
| 54 | case ActivationLayerInfo::ActivationFunction::RELU: |
| 55 | type_min = PixelValue(q_unif.offset); |
| 56 | break; |
| 57 | case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: |
| 58 | type_min = PixelValue(q_unif.offset); |
| 59 | type_max = PixelValue(act_info.a(), data_type, q_info); |
| 60 | break; |
| 61 | case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: |
| 62 | type_min = PixelValue(act_info.b(), data_type, q_info); |
| 63 | type_max = PixelValue(act_info.a(), data_type, q_info); |
| 64 | break; |
| 65 | default: |
| 66 | ARM_COMPUTE_ERROR("Activation function not supported."); |
| 67 | break; |
| 68 | } |
| 69 | } |
| 70 | |
| 71 | return std::make_pair(type_min, type_max); |
| 72 | } |
| 73 | |
| 74 | Status get_gemmlowp_output_stage_info(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const ActivationLayerInfo &act, |
| 75 | GEMMLowpOutputStageInfo &gemmlowp_output_stage_info) |
| 76 | { |
| 77 | const auto data_type = input->data_type(); |
| 78 | const QuantizationInfo oq_info = output->quantization_info(); |
| 79 | const UniformQuantizationInfo iq_unif = input->quantization_info().uniform(); |
| 80 | const UniformQuantizationInfo wq_unif = weights->quantization_info().uniform(); |
| 81 | const UniformQuantizationInfo oq_unif = oq_info.uniform(); |
| 82 | |
| 83 | float multiplier = (iq_unif.scale * wq_unif.scale) / oq_unif.scale; |
| 84 | int32_t output_multiplier; |
| 85 | int32_t output_shift; |
| 86 | |
| 87 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift)); |
| 88 | |
| 89 | PixelValue type_min{}; |
| 90 | PixelValue type_max{}; |
| 91 | std::tie(type_min, type_max) = get_quantized_asymmetric_output_min_max(oq_info, act, data_type); |
| 92 | |
| 93 | gemmlowp_output_stage_info.gemmlowp_multiplier = output_multiplier; |
| 94 | gemmlowp_output_stage_info.gemmlowp_shift = output_shift; |
| 95 | gemmlowp_output_stage_info.gemmlowp_offset = oq_unif.offset; |
| 96 | gemmlowp_output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
| 97 | gemmlowp_output_stage_info.gemmlowp_min_bound = type_min.get<int32_t>(); |
| 98 | gemmlowp_output_stage_info.gemmlowp_max_bound = type_max.get<int32_t>(); |
| 99 | |
| 100 | return Status{}; |
| 101 | } |
| 102 | |
| 103 | Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const ActivationLayerInfo &act) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 104 | { |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 105 | if(is_data_type_quantized_asymmetric(input->data_type())) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 106 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 107 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 108 | // Extract and negate input and weights offset |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 109 | const QuantizationInfo input_quantization_info(input->quantization_info().uniform().scale, -input->quantization_info().uniform().offset); |
| 110 | const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset); |
| 111 | |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 112 | GEMMLowpOutputStageInfo gemmlowp_output_stage_info; |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 113 | ARM_COMPUTE_RETURN_ON_ERROR(get_gemmlowp_output_stage_info(input, weights, output, act, gemmlowp_output_stage_info)); |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 114 | |
| 115 | GEMMInfo gemm_info; |
| 116 | gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 117 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 118 | // Validate gemmlowp function |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 119 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input->clone()->set_quantization_info(input_quantization_info), |
| 120 | &weights->clone()->set_quantization_info(weights_quantization_info), |
| 121 | biases, |
| 122 | output, |
| 123 | gemm_info)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 124 | } |
| 125 | else |
| 126 | { |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 127 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMM::validate(input, weights, biases, output, 1.f, 1.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */))); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 128 | } |
| 129 | |
| 130 | return Status{}; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 131 | } |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 132 | } // namespace |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 133 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 134 | void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITensor *output) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 135 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 136 | auto k = arm_compute::support::cpp14::make_unique<NETransposeKernel>(); |
| 137 | k->configure(input, output); |
| 138 | _kernel = std::move(k); |
| 139 | } |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 140 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 141 | Status NEFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output) |
| 142 | { |
| 143 | return NETransposeKernel::validate(input, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 144 | } |
| 145 | |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 146 | NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager) |
| 147 | : _memory_group(std::move(memory_manager)), _weights_manager(weights_manager), _flatten_kernel(), _convert_weights(), _convert_weights_managed(), _reshape_weights_function(), |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 148 | _reshape_weights_managed_function(), _mm_gemm(nullptr, weights_manager), _mm_gemmlowp(nullptr, weights_manager), _flatten_output(), _converted_weights_output(), _reshape_weights_output(), |
| 149 | _original_weights(nullptr), _are_weights_converted(true), _are_weights_reshaped(false), _is_fc_after_conv(false), _is_quantized_asymmetric(false), _is_prepared(false) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 150 | { |
| 151 | } |
| 152 | |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 153 | void NEFullyConnectedLayer::configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act) |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 154 | { |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 155 | if(_is_quantized_asymmetric) |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 156 | { |
| 157 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 158 | // Extract and negate input and weights offset |
| 159 | const QuantizationInfo input_quantization_info = input->info()->quantization_info(); |
| 160 | const QuantizationInfo weights_quantization_info = weights->info()->quantization_info(); |
| 161 | |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 162 | input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset)); |
| 163 | weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset)); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 164 | |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 165 | // Configure gemmlowp function and output stage for asymmetric quantized types |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 166 | GEMMLowpOutputStageInfo gemmlowp_output_stage_info; |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 167 | const Status status = get_gemmlowp_output_stage_info(input->info(), weights->info(), output->info(), act, gemmlowp_output_stage_info); |
| 168 | ARM_COMPUTE_ERROR_ON(status.error_code() != ErrorCode::OK); |
| 169 | |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 170 | GEMMInfo gemm_info; |
| 171 | gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info); |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 172 | gemm_info.set_activation_info(act); |
SiCong Li | adb3291 | 2020-02-17 16:39:27 +0000 | [diff] [blame] | 173 | _mm_gemmlowp.configure(input, weights, biases, output, gemm_info); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 174 | |
| 175 | // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers |
| 176 | input->info()->set_quantization_info(input_quantization_info); |
| 177 | weights->info()->set_quantization_info(weights_quantization_info); |
| 178 | } |
| 179 | else |
| 180 | { |
| 181 | // Configure matrix multiply kernel |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 182 | GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */); |
| 183 | gemm_info.set_activation_info(act); |
| 184 | _mm_gemm.configure(input, weights, biases, output, 1.f, 1.0f, gemm_info); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 185 | } |
| 186 | } |
| 187 | |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 188 | void NEFullyConnectedLayer::configure_conv_fc(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act) |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 189 | { |
| 190 | ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)))); |
| 191 | |
| 192 | // If the fully connected layer is called after a convolution layer, the input tensor must be linearized |
| 193 | |
Giorgio Arena | 368e635 | 2018-08-20 15:06:07 +0100 | [diff] [blame] | 194 | // Initialize output tensor for flatten |
| 195 | TensorShape shape_flatten = compute_flatten_shape(input->info()); |
| 196 | _flatten_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_flatten)); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 197 | |
Giorgio Arena | 368e635 | 2018-08-20 15:06:07 +0100 | [diff] [blame] | 198 | // Configure flatten kernel |
| 199 | _memory_group.manage(&_flatten_output); |
| 200 | _flatten_kernel.configure(input, &_flatten_output); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 201 | |
| 202 | // Configure matrix multiply kernel |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 203 | configure_mm(&_flatten_output, weights, biases, output, act); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 204 | |
Giorgio Arena | 368e635 | 2018-08-20 15:06:07 +0100 | [diff] [blame] | 205 | // Allocate the output tensor for flatten once all the configure methods have been called |
| 206 | _flatten_output.allocator()->allocate(); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 207 | } |
| 208 | |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 209 | void NEFullyConnectedLayer::configure_fc_fc(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act) |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 210 | { |
| 211 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1)); |
| 212 | |
| 213 | // Configure matrix multiply kernel |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 214 | configure_mm(input, weights, biases, output, act); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 215 | } |
| 216 | |
Georgios Pinitas | 7d66a8e | 2018-07-17 12:28:42 +0100 | [diff] [blame] | 217 | void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, |
| 218 | FullyConnectedLayerInfo fc_info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 219 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 220 | // Perform validate step |
Michele Di Giorgio | 9c70037 | 2020-01-08 11:33:44 +0000 | [diff] [blame] | 221 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 222 | ARM_COMPUTE_ERROR_THROW_ON(NEFullyConnectedLayer::validate(input->info(), |
| 223 | weights->info(), |
| 224 | biases != nullptr ? biases->info() : nullptr, |
| 225 | output->info(), |
Georgios Pinitas | 7d66a8e | 2018-07-17 12:28:42 +0100 | [diff] [blame] | 226 | fc_info)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 227 | |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 228 | _are_weights_converted = true; |
| 229 | _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true; |
| 230 | _is_fc_after_conv = true; |
| 231 | _is_quantized_asymmetric = is_data_type_quantized_asymmetric(input->info()->data_type()); |
| 232 | _original_weights = weights; |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 233 | |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 234 | if(_weights_manager) |
| 235 | { |
| 236 | _weights_manager->manage(weights); |
| 237 | } |
| 238 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 239 | // With the Fully Connected layer we can have 4 different cases: |
| 240 | // 1) Convolution layer -> Fully Connected layer without batches |
| 241 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 242 | // 3) Convolution layer -> Fully Connected layer with batches |
| 243 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 244 | |
| 245 | const ITensor *weights_to_use = weights; |
| 246 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 247 | // Check if we have a fully connected layer with batches |
| 248 | const bool is_batched_fc_layer = output->info()->dimension(1) > 1; |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 249 | if(is_batched_fc_layer) |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 250 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 251 | _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3, |
| 252 | input->info()->tensor_shape().cend(), |
| 253 | output->info()->tensor_shape().cbegin() + 1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 254 | } |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 255 | else |
| 256 | { |
| 257 | _is_fc_after_conv = input->info()->num_dimensions() > 1; |
| 258 | } |
| 259 | |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 260 | // Reshape weights if needed |
| 261 | if(!_are_weights_reshaped) |
| 262 | { |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 263 | if(_weights_manager && _weights_manager->are_weights_managed(weights)) |
| 264 | { |
| 265 | _reshape_weights_managed_function.configure(weights); |
| 266 | weights_to_use = _weights_manager->acquire(weights, &_reshape_weights_managed_function); |
| 267 | } |
| 268 | else |
| 269 | { |
| 270 | // Reshape the weights |
| 271 | _reshape_weights_function.configure(weights, &_reshape_weights_output); |
| 272 | weights_to_use = &_reshape_weights_output; |
| 273 | } |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 274 | } |
| 275 | |
| 276 | // Convert weights if needed |
| 277 | if(_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout)) |
| 278 | { |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 279 | if(_weights_manager && _weights_manager->are_weights_managed(weights_to_use)) |
| 280 | { |
| 281 | _convert_weights_managed.configure(weights_to_use, |
| 282 | input->info()->tensor_shape(), |
| 283 | fc_info.weights_trained_layout); |
| 284 | weights_to_use = _weights_manager->acquire(weights, &_convert_weights_managed); |
| 285 | } |
| 286 | else |
| 287 | { |
| 288 | // Convert weights |
| 289 | _convert_weights.configure(weights_to_use, |
| 290 | &_converted_weights_output, |
| 291 | input->info()->tensor_shape(), |
| 292 | fc_info.weights_trained_layout); |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 293 | |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 294 | weights_to_use = &_converted_weights_output; |
| 295 | } |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 296 | _are_weights_converted = false; |
| 297 | } |
| 298 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 299 | if(_is_fc_after_conv) |
| 300 | { |
| 301 | // Fully Connected layer after a Convolution Layer without batches |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 302 | configure_conv_fc(input, weights_to_use, biases, output, fc_info.activation_info); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 303 | } |
| 304 | else |
| 305 | { |
| 306 | // Fully Connected layer after a Fully Connected Layer without batches |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 307 | configure_fc_fc(input, weights_to_use, biases, output, fc_info.activation_info); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 308 | } |
| 309 | |
| 310 | _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 311 | } |
| 312 | |
Georgios Pinitas | 7d66a8e | 2018-07-17 12:28:42 +0100 | [diff] [blame] | 313 | Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, |
| 314 | FullyConnectedLayerInfo fc_info) |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 315 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 316 | ARM_COMPUTE_UNUSED(fc_info.retain_internal_weights); |
| 317 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); |
Michele Di Giorgio | 9c70037 | 2020-01-08 11:33:44 +0000 | [diff] [blame] | 318 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 319 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 320 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); |
Georgios Pinitas | c6aef87 | 2020-04-29 13:37:09 +0100 | [diff] [blame^] | 321 | ARM_COMPUTE_RETURN_ERROR_ON(biases != nullptr && biases->num_dimensions() > 1); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 322 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 323 | bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true; |
| 324 | bool is_fc_after_conv = true; |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 325 | |
Giorgio Arena | 368e635 | 2018-08-20 15:06:07 +0100 | [diff] [blame] | 326 | const ITensorInfo &flatten_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(input))); |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 327 | const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights))); |
Georgios Pinitas | 195b0ba | 2018-08-02 17:18:51 +0100 | [diff] [blame] | 328 | const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone()); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 329 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 330 | // With the Fully Connected layer we can have 4 different cases: |
| 331 | // 1) Convolution layer -> Fully Connected layer without batches |
| 332 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 333 | // 3) Convolution layer -> Fully Connected layer with batches |
| 334 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 335 | |
| 336 | const ITensorInfo *input_to_use = input; |
| 337 | const ITensorInfo *weights_to_use = weights; |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 338 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 339 | // Check if we have a fully connected layer with batches |
| 340 | const bool is_batched_fc_layer = output->dimension(1) > 1; |
| 341 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 342 | if(is_batched_fc_layer) |
| 343 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 344 | is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->tensor_shape().cbegin() + 3, |
| 345 | input->tensor_shape().cend(), |
| 346 | output->tensor_shape().cbegin() + 1)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 347 | } |
| 348 | else |
| 349 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 350 | is_fc_after_conv = input->num_dimensions() > 1; |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 351 | } |
| 352 | |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 353 | if(!weights_reshaped) |
| 354 | { |
| 355 | // Validate reshape weights kernel |
| 356 | ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights)); |
| 357 | weights_to_use = &reshaped_weights; |
| 358 | } |
| 359 | |
| 360 | if(is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout)) |
| 361 | { |
| 362 | // Validate convert weights kernel |
| 363 | ARM_COMPUTE_RETURN_ON_ERROR(NEConvertFullyConnectedWeights::validate(weights_to_use, |
| 364 | &converted_weights, |
| 365 | input->tensor_shape(), |
| 366 | fc_info.weights_trained_layout)); |
| 367 | weights_to_use = &converted_weights; |
| 368 | } |
| 369 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 370 | if(is_fc_after_conv) |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 371 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 372 | // Fully Connected layer after a Convolution Layer without batches |
| 373 | ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (input->dimension(0) * input->dimension(1) * input->dimension(2)))); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 374 | |
Giorgio Arena | 368e635 | 2018-08-20 15:06:07 +0100 | [diff] [blame] | 375 | // Validate flatten kernel |
| 376 | ARM_COMPUTE_RETURN_ON_ERROR(NEFlattenLayerKernel::validate(input, &flatten_input)); |
| 377 | input_to_use = &flatten_input; |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 378 | } |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 379 | else |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 380 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 381 | // Fully Connected layer after a Fully Connected Layer without batches |
| 382 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 383 | } |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 384 | // Validate matrix multiply kernel |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 385 | ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(input_to_use, weights_to_use, biases, output, fc_info.activation_info)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 386 | |
| 387 | return Status{}; |
| 388 | } |
| 389 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 390 | void NEFullyConnectedLayer::run() |
| 391 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 392 | prepare(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 393 | |
Georgios Pinitas | da953f2 | 2019-04-02 17:27:03 +0100 | [diff] [blame] | 394 | MemoryGroupResourceScope scope_mg(_memory_group); |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 395 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 396 | // Linearize input if it comes from a convolutional layer |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 397 | if(_is_fc_after_conv) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 398 | { |
Giorgio Arena | 368e635 | 2018-08-20 15:06:07 +0100 | [diff] [blame] | 399 | NEScheduler::get().schedule(&_flatten_kernel, Window::DimY); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 400 | } |
| 401 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 402 | // Run matrix multiply |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 403 | if(_is_quantized_asymmetric) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 404 | { |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 405 | _mm_gemmlowp.run(); |
| 406 | } |
| 407 | else |
| 408 | { |
| 409 | _mm_gemm.run(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 410 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 411 | } |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 412 | |
| 413 | void NEFullyConnectedLayer::prepare() |
| 414 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 415 | if(!_is_prepared) |
| 416 | { |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 417 | if(!_weights_manager) |
| 418 | { |
| 419 | ARM_COMPUTE_ERROR_ON(!_original_weights->is_used()); |
| 420 | } |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 421 | |
| 422 | auto release_unused = [](Tensor * w) |
| 423 | { |
| 424 | if(!w->is_used()) |
| 425 | { |
| 426 | w->allocator()->free(); |
| 427 | } |
| 428 | }; |
| 429 | |
| 430 | // Pointer to current weights |
| 431 | const ITensor *cur_weights = _original_weights; |
| 432 | |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 433 | // Reshape of the weights (happens only once) |
| 434 | if(!_are_weights_reshaped) |
| 435 | { |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 436 | if(_weights_manager && _weights_manager->are_weights_managed(_original_weights)) |
| 437 | { |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 438 | cur_weights = _weights_manager->run(cur_weights, &_reshape_weights_managed_function); |
| 439 | } |
| 440 | else |
| 441 | { |
| 442 | // Reshape of the weights (happens only once) |
| 443 | if(!_are_weights_reshaped) |
| 444 | { |
| 445 | // Run reshape weights kernel and mark weights as unused |
| 446 | _reshape_weights_output.allocator()->allocate(); |
| 447 | _reshape_weights_function.run(); |
| 448 | } |
| 449 | cur_weights->mark_as_unused(); |
| 450 | cur_weights = &_reshape_weights_output; |
| 451 | } |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 452 | _are_weights_reshaped = true; |
| 453 | } |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 454 | |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 455 | // Convert weights if needed (happens only once) |
| 456 | if(!_are_weights_converted) |
| 457 | { |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 458 | if(_weights_manager && _weights_manager->are_weights_managed(cur_weights)) |
| 459 | { |
| 460 | _weights_manager->run(cur_weights, &_convert_weights_managed); |
| 461 | } |
| 462 | else |
| 463 | { |
| 464 | _converted_weights_output.allocator()->allocate(); |
| 465 | _convert_weights.run(); |
Michalis Spyrou | 20c2b50 | 2019-10-01 15:39:42 +0100 | [diff] [blame] | 466 | cur_weights->mark_as_unused(); |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 467 | } |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 468 | |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 469 | _are_weights_converted = true; |
| 470 | } |
| 471 | |
| 472 | // Release reshaped weights if unused |
| 473 | release_unused(&_reshape_weights_output); |
| 474 | |
| 475 | // Prepare GEMM prepare and release unused weights |
SiCongLi | 2e5fd63 | 2020-03-02 15:39:15 +0000 | [diff] [blame] | 476 | if(!_is_quantized_asymmetric) |
Georgios Pinitas | ef776a8 | 2018-07-25 17:57:49 +0100 | [diff] [blame] | 477 | { |
| 478 | _mm_gemm.prepare(); |
| 479 | } |
| 480 | |
| 481 | // Release converted weights if unused |
| 482 | release_unused(&_reshape_weights_output); |
| 483 | release_unused(&_converted_weights_output); |
| 484 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 485 | _is_prepared = true; |
| 486 | } |
Michele Di Giorgio | f29d1b7 | 2019-10-29 10:58:13 +0000 | [diff] [blame] | 487 | } |
Georgios Pinitas | c6aef87 | 2020-04-29 13:37:09 +0100 | [diff] [blame^] | 488 | } // namespace arm_compute |