Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 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/CL/functions/CLFullyConnectedLayer.h" |
| 25 | |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/Size2D.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Validate.h" |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 28 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 30 | #include "arm_compute/runtime/CL/CLScheduler.h" |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 31 | #include "support/ToolchainSupport.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 32 | |
| 33 | #include <algorithm> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 34 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 35 | using namespace arm_compute; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 36 | using namespace arm_compute::misc::shape_calculator; |
| 37 | |
| 38 | namespace |
| 39 | { |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 40 | Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output) |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 41 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 42 | if(is_data_type_quantized_asymmetric(input.data_type())) |
| 43 | { |
| 44 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 45 | // Extract and negate input and weights offset |
| 46 | const QuantizationInfo input_quantization_info(input.quantization_info().scale, -input.quantization_info().offset); |
| 47 | const QuantizationInfo weights_quantization_info(weights.quantization_info().scale, -weights.quantization_info().offset); |
| 48 | |
| 49 | // Validate gemmlowp function |
| 50 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info), |
| 51 | &weights.clone()->set_quantization_info(weights_quantization_info), |
| 52 | &output)); |
| 53 | } |
| 54 | else |
| 55 | { |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 56 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input, &weights, nullptr, &output, 1.f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */))); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 57 | } |
| 58 | |
| 59 | return Status{}; |
| 60 | } |
| 61 | } // namespace |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 62 | |
| 63 | void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output) |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 64 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 65 | auto k = arm_compute::support::cpp14::make_unique<CLTransposeKernel>(); |
| 66 | k->configure(input, output); |
| 67 | _kernel = std::move(k); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 68 | } |
| 69 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 70 | Status CLFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output) |
| 71 | { |
| 72 | return CLTransposeKernel::validate(input, output); |
| 73 | } |
| 74 | |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 75 | CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager) |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 76 | : _memory_group(memory_manager), _im2col_kernel(), _reshape_weights_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _accumulate_biases_kernel(), |
| 77 | _im2col_output(), _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false), _original_weights(nullptr) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 78 | { |
| 79 | } |
| 80 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 81 | void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 82 | { |
| 83 | if(_is_quantized) |
| 84 | { |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 85 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 86 | // Extract and negate input and weights offset |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 87 | const QuantizationInfo input_quantization_info = input->info()->quantization_info(); |
| 88 | const QuantizationInfo weights_quantization_info = weights->info()->quantization_info(); |
| 89 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 90 | input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); |
| 91 | weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 92 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 93 | // Configure gemmlowp function |
| 94 | _mm_gemmlowp.configure(input, weights, output); |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 95 | |
| 96 | // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers |
| 97 | input->info()->set_quantization_info(input_quantization_info); |
| 98 | weights->info()->set_quantization_info(weights_quantization_info); |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 99 | } |
| 100 | else |
| 101 | { |
| 102 | // Configure matrix multiply kernel |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 103 | _mm_gemm.configure(input, weights, nullptr, output, 1.f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */)); |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 104 | } |
| 105 | } |
| 106 | |
| 107 | void CLFullyConnectedLayer::configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 108 | { |
| 109 | ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)))); |
| 110 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 111 | // If the fully connected layer is called after a convolution layer, the input tensor must be linearized |
| 112 | |
| 113 | // Initialize output tensor for im2col |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 114 | TensorShape shape_im2col = compute_im2col_fc_shape(input->info()); |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 115 | _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 116 | |
| 117 | // Configure im2col kernel |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 118 | _memory_group.manage(&_im2col_output); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 119 | _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); |
| 120 | |
| 121 | // Configure matrix multiply kernel |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 122 | configure_mm(&_im2col_output, weights, output); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 123 | |
| 124 | // Allocate the output tensor for im2col once all the configure methods have been called |
| 125 | _im2col_output.allocator()->allocate(); |
| 126 | } |
| 127 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 128 | void CLFullyConnectedLayer::configure_fc_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 129 | { |
| 130 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1)); |
| 131 | |
| 132 | // Configure matrix multiply kernel |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 133 | configure_mm(input, weights, output); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 134 | } |
| 135 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 136 | void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped) |
| 137 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 138 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 139 | |
| 140 | // Perform validate step |
| 141 | ARM_COMPUTE_ERROR_THROW_ON(CLFullyConnectedLayer::validate(input->info(), |
| 142 | weights->info(), |
| 143 | biases != nullptr ? biases->info() : nullptr, |
| 144 | output->info(), |
| 145 | transpose_weights, |
| 146 | are_weights_reshaped)); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 147 | |
| 148 | _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true; |
| 149 | _is_fc_after_conv = true; |
| 150 | _accumulate_biases = false; |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 151 | _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 152 | _original_weights = weights; |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 153 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 154 | // Configure gemmlowp output |
| 155 | if(_is_quantized) |
| 156 | { |
| 157 | _gemmlowp_output.allocator()->init(output->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32)); |
| 158 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 159 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 160 | // Configure accumulate biases kernel for non quantized asymmetric types |
| 161 | if(biases != nullptr && !_is_quantized) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 162 | { |
| 163 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 164 | |
| 165 | _accumulate_biases = true; |
| 166 | |
| 167 | // Configure accumulate biases kernel |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 168 | _accumulate_biases_kernel.set_target(CLScheduler::get().target()); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 169 | _accumulate_biases_kernel.configure(output, biases); |
| 170 | } |
| 171 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 172 | // With the Fully Connected layer we can have 4 different cases: |
| 173 | // 1) Convolution layer -> Fully Connected layer without batches |
| 174 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 175 | // 3) Convolution layer -> Fully Connected layer with batches |
| 176 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 177 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 178 | const ICLTensor *weights_to_use = weights; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 179 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 180 | if(!_are_weights_reshaped) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 181 | { |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 182 | weights_to_use = &_reshape_weights_output; |
| 183 | |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 184 | // Reshape the weights |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 185 | _reshape_weights_kernel.configure(weights, &_reshape_weights_output); |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 186 | } |
| 187 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 188 | // Check if we have a fully connected layer with batches |
| 189 | const bool is_batched_fc_layer = output->info()->dimension(1) > 1; |
| 190 | |
| 191 | if(is_batched_fc_layer) |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 192 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 193 | _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3, |
| 194 | input->info()->tensor_shape().cend(), |
| 195 | output->info()->tensor_shape().cbegin() + 1)); |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 196 | } |
| 197 | else |
| 198 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 199 | _is_fc_after_conv = input->info()->num_dimensions() > 1; |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 200 | } |
| 201 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 202 | ICLTensor *tmp_output = (_is_quantized) ? &_gemmlowp_output : output; |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 203 | if(_is_fc_after_conv) |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 204 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 205 | // Fully Connected layer after a Convolution Layer without batches |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 206 | configure_conv_fc(input, weights_to_use, tmp_output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 207 | } |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 208 | else |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 209 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 210 | // Fully Connected layer after a Fully Connected Layer without batches |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 211 | configure_fc_fc(input, weights_to_use, tmp_output); |
| 212 | } |
| 213 | |
| 214 | // Configure output stage for asymmetric quantized types |
| 215 | if(_is_quantized) |
| 216 | { |
| 217 | float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; |
| 218 | int output_multiplier, output_shift; |
| 219 | quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 220 | _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset); |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 221 | _gemmlowp_output.allocator()->allocate(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 222 | } |
| 223 | |
| 224 | // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 225 | if(!_are_weights_reshaped) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 226 | { |
Moritz Pflanzer | 768e9f1 | 2017-08-11 15:33:30 +0100 | [diff] [blame] | 227 | // Allocate the tensor for the weights reshaped |
| 228 | _reshape_weights_output.allocator()->allocate(); |
| 229 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 230 | } |
| 231 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 232 | Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped) |
| 233 | { |
| 234 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); |
| 235 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); |
| 236 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 237 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); |
| 238 | |
| 239 | bool weights_reshaped = transpose_weights ? are_weights_reshaped : true; |
| 240 | bool is_fc_after_conv = true; |
| 241 | bool is_quantized = is_data_type_quantized_asymmetric(input->data_type()); |
| 242 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 243 | |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 244 | const ITensorInfo &im2col_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_fc_shape(input))); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 245 | const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights))); |
| 246 | const ITensorInfo &gemmlowp_output = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32)); |
| 247 | |
| 248 | // Configure accumulate biases kernel for non quantized asymmetric types |
| 249 | if(biases != nullptr && !is_quantized) |
| 250 | { |
| 251 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 252 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixAccumulateBiasesKernel::validate(output, biases, gpu_target)); |
| 253 | } |
| 254 | |
| 255 | // With the Fully Connected layer we can have 4 different cases: |
| 256 | // 1) Convolution layer -> Fully Connected layer without batches |
| 257 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 258 | // 3) Convolution layer -> Fully Connected layer with batches |
| 259 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 260 | |
| 261 | const ITensorInfo *input_to_use = input; |
| 262 | const ITensorInfo *weights_to_use = weights; |
| 263 | const ITensorInfo *tmp_output = (is_quantized) ? &gemmlowp_output : output; |
| 264 | |
| 265 | if(!weights_reshaped) |
| 266 | { |
| 267 | // Validate reshape weights kernel |
| 268 | ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights)); |
| 269 | weights_to_use = &reshaped_weights; |
| 270 | } |
| 271 | |
| 272 | // Check if we have a fully connected layer with batches |
| 273 | const bool is_batched_fc_layer = output->dimension(1) > 1; |
| 274 | |
| 275 | if(is_batched_fc_layer) |
| 276 | { |
| 277 | is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->tensor_shape().cbegin() + 3, |
| 278 | input->tensor_shape().cend(), |
| 279 | output->tensor_shape().cbegin() + 1)); |
| 280 | } |
| 281 | else |
| 282 | { |
| 283 | is_fc_after_conv = input->num_dimensions() > 1; |
| 284 | } |
| 285 | |
| 286 | if(is_fc_after_conv) |
| 287 | { |
| 288 | // Fully Connected layer after a Convolution Layer without batches |
| 289 | ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (input->dimension(0) * input->dimension(1) * input->dimension(2)))); |
| 290 | |
| 291 | // Validate im2col kernel |
| 292 | ARM_COMPUTE_RETURN_ON_ERROR(CLIm2ColKernel::validate(input, &im2col_input, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false)); |
| 293 | input_to_use = &im2col_input; |
| 294 | } |
| 295 | else |
| 296 | { |
| 297 | // Fully Connected layer after a Fully Connected Layer without batches |
| 298 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1)); |
| 299 | } |
| 300 | // Validate matrix multiply kernel |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 301 | ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*input_to_use, *weights_to_use, *tmp_output)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 302 | |
| 303 | // Validate output stage for asymmetric quantized types |
| 304 | if(is_quantized) |
| 305 | { |
| 306 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(&gemmlowp_output, biases, output)); |
| 307 | } |
| 308 | |
| 309 | return Status{}; |
| 310 | } |
| 311 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 312 | void CLFullyConnectedLayer::run() |
| 313 | { |
| 314 | // Reshape of the weights (happens only once) |
| 315 | if(!_are_weights_reshaped) |
| 316 | { |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 317 | ARM_COMPUTE_ERROR_ON(!_original_weights->is_used()); |
| 318 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 319 | _are_weights_reshaped = true; |
| 320 | _reshape_weights_kernel.run(); |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 321 | |
| 322 | // Mark original weights tensor as unused |
| 323 | _original_weights->mark_as_unused(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 324 | } |
| 325 | |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 326 | _memory_group.acquire(); |
| 327 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 328 | // Linearize input if it comes from a convolutional layer |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 329 | if(_is_fc_after_conv) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 330 | { |
| 331 | CLScheduler::get().enqueue(_im2col_kernel, false); |
| 332 | } |
| 333 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 334 | // Run matrix multiply |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 335 | if(_is_quantized) |
| 336 | { |
| 337 | _mm_gemmlowp.run(); |
| 338 | } |
| 339 | else |
| 340 | { |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame^] | 341 | _mm_gemm.run(); |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 342 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 343 | |
| 344 | // Accumulate biases if provided |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 345 | if(_is_quantized) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 346 | { |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 347 | _gemmlowp_output_stage.run(); |
| 348 | } |
| 349 | else |
| 350 | { |
| 351 | if(_accumulate_biases) |
| 352 | { |
| 353 | CLScheduler::get().enqueue(_accumulate_biases_kernel); |
| 354 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 355 | } |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 356 | |
| 357 | _memory_group.release(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 358 | } |