Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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/graph/nodes/ConvolutionLayer.h" |
| 25 | |
| 26 | #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h" |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 27 | #include "arm_compute/runtime/CL/functions/CLDirectConvolutionLayer.h" |
| 28 | #include "arm_compute/runtime/IFunction.h" |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 29 | #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h" |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 30 | #include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h" |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 31 | #include "support/ToolchainSupport.h" |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 32 | #include "utils/GraphTypePrinter.h" |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 33 | #include "utils/TypePrinter.h" |
| 34 | |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 35 | #include <tuple> |
| 36 | #include <vector> |
| 37 | |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 38 | using namespace arm_compute::graph; |
| 39 | |
| 40 | namespace |
| 41 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 42 | /** Calculates the output shaped of the convolution layer |
| 43 | * |
| 44 | * @param[in] input_shape Input tensor shape |
| 45 | * @param[in] weights_shape Weights shape |
| 46 | * @param[in] conv_info Convolution information (padding, stride, etc.) |
| 47 | * |
| 48 | * @return The expected output tensor shape |
| 49 | */ |
| 50 | TensorShape calculate_convolution_layer_output_shape(const TensorShape &input_shape, const TensorShape &weights_shape, const PadStrideInfo &conv_info) |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 51 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 52 | unsigned int output_width = 0; |
| 53 | unsigned int output_height = 0; |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 54 | |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 55 | // Get output width and height |
| 56 | std::tie(output_width, output_height) = arm_compute::scaled_dimensions(input_shape.x(), input_shape.y(), weights_shape.x(), weights_shape.y(), conv_info); |
| 57 | |
| 58 | // Create output shape |
| 59 | TensorShape output_shape = input_shape; |
| 60 | output_shape.set(0, output_width); |
| 61 | output_shape.set(1, output_height); |
| 62 | output_shape.set(2, weights_shape[3]); |
| 63 | |
| 64 | return output_shape; |
| 65 | } |
| 66 | |
| 67 | // Instantiate GEMM based convolution layer |
| 68 | template <typename ConvolutionType, typename TensorType, Hint hint> |
| 69 | std::unique_ptr<arm_compute::IFunction> instantiate_function(ITensor *input, ITensor *weights, ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) |
| 70 | { |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 71 | auto conv = arm_compute::support::cpp14::make_unique<ConvolutionType>(); |
| 72 | conv->configure( |
| 73 | dynamic_cast<TensorType *>(input), |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 74 | dynamic_cast<TensorType *>(weights), |
| 75 | dynamic_cast<TensorType *>(biases), |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 76 | dynamic_cast<TensorType *>(output), |
| 77 | conv_info, weights_info); |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 78 | return std::move(conv); |
| 79 | } |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 80 | |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 81 | // Instantiate direct convolution layer |
| 82 | template <typename ConvolutionType, typename TensorType, Hint hint> |
| 83 | std::unique_ptr<arm_compute::IFunction> instantiate_direct_function(ITensor *input, ITensor *weights, ITensor *biases, ITensor *output, const PadStrideInfo &conv_info) |
| 84 | { |
| 85 | auto conv = arm_compute::support::cpp14::make_unique<ConvolutionType>(); |
| 86 | conv->configure( |
| 87 | dynamic_cast<TensorType *>(input), |
| 88 | dynamic_cast<TensorType *>(weights), |
| 89 | dynamic_cast<TensorType *>(biases), |
| 90 | dynamic_cast<TensorType *>(output), |
| 91 | conv_info); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 92 | return std::move(conv); |
| 93 | } |
| 94 | |
| 95 | template <Hint hint> |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 96 | std::unique_ptr<arm_compute::IFunction> instantiate(ITensor *input, ITensor *weights, ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, |
| 97 | ConvolutionMethodHint conv_method); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 98 | |
| 99 | template <> |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 100 | std::unique_ptr<arm_compute::IFunction> instantiate<Hint::OPENCL>(ITensor *input, ITensor *weights, ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, |
| 101 | ConvolutionMethodHint conv_method) |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 102 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 103 | if(conv_method == ConvolutionMethodHint::GEMM) |
| 104 | { |
| 105 | return instantiate_function<arm_compute::CLConvolutionLayer, arm_compute::ICLTensor, Hint::OPENCL>(input, weights, biases, output, conv_info, weights_info); |
| 106 | } |
| 107 | else |
| 108 | { |
| 109 | return instantiate_direct_function<arm_compute::CLDirectConvolutionLayer, arm_compute::ICLTensor, Hint::OPENCL>(input, weights, biases, output, conv_info); |
| 110 | } |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 111 | } |
| 112 | |
| 113 | template <> |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 114 | std::unique_ptr<arm_compute::IFunction> instantiate<Hint::NEON>(ITensor *input, ITensor *weights, ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, |
| 115 | ConvolutionMethodHint conv_method) |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 116 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 117 | if(conv_method == ConvolutionMethodHint::GEMM) |
| 118 | { |
| 119 | return instantiate_function<arm_compute::NEConvolutionLayer, arm_compute::ITensor, Hint::NEON>(input, weights, biases, output, conv_info, weights_info); |
| 120 | } |
| 121 | else |
| 122 | { |
| 123 | return instantiate_direct_function<arm_compute::NEDirectConvolutionLayer, arm_compute::ITensor, Hint::NEON>(input, weights, biases, output, conv_info); |
| 124 | } |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 125 | } |
| 126 | } // namespace |
| 127 | |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 128 | /** Grouped Convolution function */ |
| 129 | class GroupedConvolutionFunction final : public arm_compute::IFunction |
| 130 | { |
| 131 | public: |
| 132 | /** Default Constructor */ |
| 133 | GroupedConvolutionFunction() |
| 134 | : _convolutions() |
| 135 | { |
| 136 | } |
| 137 | /** Default Destructor */ |
| 138 | ~GroupedConvolutionFunction() final = default; |
| 139 | /** Prevent instances from being copy constructed */ |
| 140 | GroupedConvolutionFunction(const GroupedConvolutionFunction &) = delete; |
| 141 | /** Prevent instances from being copy assigned */ |
| 142 | GroupedConvolutionFunction &operator=(const GroupedConvolutionFunction &) = delete; |
| 143 | /** Allow instances to be move constructed */ |
| 144 | GroupedConvolutionFunction(GroupedConvolutionFunction &&) noexcept = default; |
| 145 | /** Allow instances to be move assigned */ |
| 146 | GroupedConvolutionFunction &operator=(GroupedConvolutionFunction &&) noexcept = default; |
| 147 | /** Adds a convolution |
| 148 | * |
| 149 | * @param convolution Convolution function to add |
| 150 | */ |
| 151 | void add_convolution_function(std::unique_ptr<IFunction> convolution) |
| 152 | { |
| 153 | _convolutions.emplace_back(std::move(convolution)); |
| 154 | } |
| 155 | |
| 156 | // Inherited methods overriden: |
| 157 | void run() override |
| 158 | { |
| 159 | for(auto &c : _convolutions) |
| 160 | { |
| 161 | c->run(); |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | private: |
| 166 | std::vector<std::unique_ptr<IFunction>> _convolutions; |
| 167 | }; |
| 168 | |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 169 | std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_node(Hint hint, ITensor *input, ITensor *output) |
| 170 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 171 | // Set weights and biases info |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 172 | if(_weights.tensor() == nullptr) |
| 173 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 174 | _weights.set_info(TensorInfo(TensorShape(_conv_width, _conv_height, input->info()->dimension(2) / _num_groups, _ofm), |
| 175 | input->info()->num_channels(), input->info()->data_type(), |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 176 | input->info()->fixed_point_position())); |
| 177 | } |
| 178 | if(_biases.tensor() == nullptr) |
| 179 | { |
| 180 | _biases.set_info(TensorInfo(TensorShape(_ofm), input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position())); |
| 181 | } |
| 182 | |
| 183 | std::unique_ptr<arm_compute::IFunction> func; |
| 184 | _hint = hint; |
| 185 | _input = input; |
| 186 | _output = output; |
| 187 | |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 188 | // Check if the weights and biases are loaded |
| 189 | bool weights_are_loaded = _weights.tensor() != nullptr; |
| 190 | bool biases_are_loaded = _weights.tensor() != nullptr; |
| 191 | |
| 192 | // Set bias and weights target |
| 193 | _weights.set_target(_hint); |
| 194 | _biases.set_target(_hint); |
| 195 | |
| 196 | // Calculate output shape |
| 197 | TensorShape output_shape = calculate_convolution_layer_output_shape(_input->info()->tensor_shape(), _weights.info().tensor_shape(), _conv_info); |
| 198 | |
| 199 | // Output auto inizialitation if not yet initialized |
| 200 | arm_compute::auto_init_if_empty(*_output->info(), output_shape, 1, _input->info()->data_type(), _input->info()->fixed_point_position()); |
| 201 | |
| 202 | // Create appropriate convolution function |
| 203 | // TODO(geopin01): Fix convolution layer hints once the GraphContext has been added |
| 204 | if(_num_groups == 1) |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 205 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 206 | func = instantiate_convolution(ConvolutionMethodHint::GEMM); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 207 | } |
| 208 | else |
| 209 | { |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 210 | func = instantiate_grouped_convolution(ConvolutionMethodHint::GEMM); |
| 211 | } |
| 212 | |
| 213 | // Fill weights |
| 214 | if(!weights_are_loaded) |
| 215 | { |
| 216 | _weights.allocate_and_fill_if_needed(); |
| 217 | } |
| 218 | // Fill biases |
| 219 | if(!biases_are_loaded) |
| 220 | { |
| 221 | _biases.allocate_and_fill_if_needed(); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 222 | } |
| 223 | |
| 224 | return func; |
| 225 | } |
| 226 | |
| 227 | void ConvolutionLayer::print_info() |
| 228 | { |
| 229 | if(_hint == Hint::OPENCL) |
| 230 | { |
| 231 | std::cout << "Instantiating CLConvolutionLayer"; |
| 232 | } |
| 233 | else |
| 234 | { |
| 235 | std::cout << "Instantiating NEConvolutionLayer"; |
| 236 | } |
Georgios Pinitas | 6f669f0 | 2017-09-26 12:32:57 +0100 | [diff] [blame^] | 237 | std::cout << " Data Type: " << _input->info()->data_type() |
| 238 | << " Input Shape: " << _input->info()->tensor_shape() |
| 239 | << " Weights shape: " << _weights.info().tensor_shape() |
| 240 | << " Biases Shape: " << _biases.info().tensor_shape() |
| 241 | << " Output Shape: " << _output->info()->tensor_shape() |
| 242 | << " PadStrideInfo: " << _conv_info |
| 243 | << " Groups: " << _num_groups |
| 244 | << " WeightsInfo: " << _weights_info |
| 245 | << std::endl; |
| 246 | } |
| 247 | |
| 248 | std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_convolution(ConvolutionMethodHint conv_method_hint) |
| 249 | { |
| 250 | std::unique_ptr<arm_compute::IFunction> func; |
| 251 | if(_hint == Hint::OPENCL) |
| 252 | { |
| 253 | func = instantiate<Hint::OPENCL>(_input, _weights.tensor(), _biases.tensor(), _output, _conv_info, _weights_info, conv_method_hint); |
| 254 | } |
| 255 | else |
| 256 | { |
| 257 | func = instantiate<Hint::NEON>(_input, _weights.tensor(), _biases.tensor(), _output, _conv_info, _weights_info, conv_method_hint); |
| 258 | } |
| 259 | return func; |
| 260 | } |
| 261 | |
| 262 | std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_grouped_convolution(ConvolutionMethodHint conv_method_hint) |
| 263 | { |
| 264 | // Get tensor shapes |
| 265 | TensorShape input_shape = _input->info()->tensor_shape(); |
| 266 | TensorShape output_shape = _output->info()->tensor_shape(); |
| 267 | TensorShape weights_shape = _weights.info().tensor_shape(); |
| 268 | TensorShape biases_shape = _biases.info().tensor_shape(); |
| 269 | |
| 270 | ARM_COMPUTE_ERROR_ON_MSG((input_shape.z() % _num_groups) != 0, "Input depth not multiple of the number of groups!"); |
| 271 | ARM_COMPUTE_ERROR_ON_MSG((output_shape.z() % _num_groups) != 0, "Output depth not multiple of the number of groups!"); |
| 272 | ARM_COMPUTE_ERROR_ON_MSG((weights_shape[3] % _num_groups) != 0, "Number of kernels not multiple of the number of groups!"); |
| 273 | ARM_COMPUTE_ERROR_ON_MSG((biases_shape.x() % _num_groups) != 0, "Biases not multiple of the number of groups!"); |
| 274 | |
| 275 | // Create a grouped convolution function |
| 276 | auto grouped_conv = arm_compute::support::cpp14::make_unique<GroupedConvolutionFunction>(); |
| 277 | |
| 278 | // Create sub-tensors vectors |
| 279 | _is = arm_compute::support::cpp14::make_unique<SubTensor[]>(_num_groups); |
| 280 | _os = arm_compute::support::cpp14::make_unique<SubTensor[]>(_num_groups); |
| 281 | _ws = arm_compute::support::cpp14::make_unique<SubTensor[]>(_num_groups); |
| 282 | _bs = arm_compute::support::cpp14::make_unique<SubTensor[]>(_num_groups); |
| 283 | |
| 284 | // Calculate sub-tensor splits |
| 285 | const int input_split = input_shape.z() / _num_groups; |
| 286 | const int output_split = output_shape.z() / _num_groups; |
| 287 | const int weights_split = weights_shape[3] / _num_groups; |
| 288 | const int biases_split = biases_shape.x() / _num_groups; |
| 289 | |
| 290 | // Calculate sub-tensor shapes |
| 291 | input_shape.set(2, input_split); |
| 292 | output_shape.set(2, output_split); |
| 293 | weights_shape.set(3, weights_split); |
| 294 | biases_shape.set(0, biases_split); |
| 295 | |
| 296 | // Configure sub-tensors |
| 297 | for(int i = 0; i < static_cast<int>(_num_groups); ++i) |
| 298 | { |
| 299 | // Create convolution function |
| 300 | std::unique_ptr<arm_compute::IFunction> func; |
| 301 | |
| 302 | // Calculate sub-tensors starting coordinates |
| 303 | Coordinates input_coord(0, 0, input_split * i); |
| 304 | Coordinates output_coord(0, 0, output_split * i); |
| 305 | Coordinates weights_coord(0, 0, 0, weights_split * i); |
| 306 | Coordinates biases_coord(biases_split * i); |
| 307 | |
| 308 | // Create sub-tensors for input, output, weights and bias |
| 309 | auto hint_to_use = (_hint == Hint::OPENCL) ? Hint::OPENCL : Hint::NEON; |
| 310 | _is[i] = SubTensor(_input, input_shape, input_coord, hint_to_use); |
| 311 | _os[i] = SubTensor(_output, output_shape, output_coord, hint_to_use); |
| 312 | _ws[i] = SubTensor(_weights.tensor(), weights_shape, weights_coord, hint_to_use); |
| 313 | _bs[i] = SubTensor(_biases.tensor(), biases_shape, biases_coord, hint_to_use); |
| 314 | |
| 315 | // Instantiate convolution function |
| 316 | if(_hint == Hint::OPENCL) |
| 317 | { |
| 318 | func = instantiate<Hint::OPENCL>(_is[i].tensor(), _ws[i].tensor(), _bs[i].tensor(), _os[i].tensor(), _conv_info, _weights_info, conv_method_hint); |
| 319 | } |
| 320 | else |
| 321 | { |
| 322 | func = instantiate<Hint::NEON>(_is[i].tensor(), _ws[i].tensor(), _bs[i].tensor(), _os[i].tensor(), _conv_info, _weights_info, conv_method_hint); |
| 323 | } |
| 324 | |
| 325 | // Add convolution function to the list of convolutions for the grouped convolution |
| 326 | grouped_conv->add_convolution_function(std::move(func)); |
| 327 | } |
| 328 | |
| 329 | return std::move(grouped_conv); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 330 | } |