Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +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/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h" |
| 29 | #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h" |
| 30 | #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/IAccessWindow.h" |
| 33 | #include "arm_compute/core/ITensor.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/core/Validate.h" |
| 36 | #include "support/ToolchainSupport.h" |
| 37 | |
| 38 | using namespace arm_compute; |
| 39 | |
| 40 | template <unsigned int kernel_size> |
| 41 | GCDirectConvolutionLayerKernel<kernel_size>::GCDirectConvolutionLayerKernel() |
| 42 | : _input(nullptr), _bias(nullptr), _weights(nullptr), _output(nullptr), _border_size(0), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_x(0), _conv_pad_y(0), _lws(gles::NDRange(1U, 1U, 1U)) |
| 43 | { |
| 44 | } |
| 45 | |
| 46 | template <unsigned int kernel_size> |
| 47 | BorderSize GCDirectConvolutionLayerKernel<kernel_size>::border_size() const |
| 48 | { |
| 49 | return _border_size; |
| 50 | } |
| 51 | |
| 52 | template <unsigned int kernel_size> |
| 53 | void GCDirectConvolutionLayerKernel<kernel_size>::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *bias, IGCTensor *output, const PadStrideInfo &conv_info) |
| 54 | { |
| 55 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); |
| 56 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 57 | ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2)); |
| 58 | ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != weights->info()->dimension(1)); |
| 59 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 60 | ARM_COMPUTE_ERROR_ON_MSG((kernel_size == 3 && std::get<0>(conv_info.stride()) > 2), "Strides larger than 2 not supported in 3x3 direct convolution!"); |
| 61 | ARM_COMPUTE_ERROR_ON(kernel_size != weights->info()->dimension(0)); |
| 62 | |
| 63 | if(bias != nullptr) |
| 64 | { |
| 65 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, bias); |
| 66 | // FIXME: Bug in framework, workaround it in tests currently. |
| 67 | //ARM_COMPUTE_ERROR_ON(bias->info()->dimension(0) != weights->info()->dimension(3)); |
| 68 | ARM_COMPUTE_ERROR_ON(bias->info()->num_dimensions() > 1); |
| 69 | } |
| 70 | |
| 71 | _conv_stride_x = std::get<0>(conv_info.stride()); |
| 72 | _conv_stride_y = std::get<1>(conv_info.stride()); |
| 73 | _conv_pad_x = std::get<0>(conv_info.pad()); |
| 74 | _conv_pad_y = std::get<1>(conv_info.pad()); |
| 75 | |
| 76 | _input = input; |
| 77 | _weights = weights; |
| 78 | _output = output; |
| 79 | _bias = bias; |
| 80 | _border_size = BorderSize(_conv_pad_y, _conv_pad_x); |
| 81 | |
| 82 | std::set<std::string> options; |
| 83 | |
| 84 | options.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(_lws[0])); |
| 85 | options.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(_lws[1])); |
| 86 | options.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(_lws[2])); |
| 87 | options.emplace("#define STRIDE_X " + support::cpp11::to_string(_conv_stride_x)); |
| 88 | |
| 89 | std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16"; |
| 90 | options.emplace(("#define " + dt_name)); |
| 91 | |
| 92 | unsigned int num_elems_read_per_iteration_x = kernel_size * _conv_stride_x; |
| 93 | unsigned int num_elems_read_per_iteration_y = 1; |
| 94 | unsigned int num_elems_written_per_iteration_x = 1; |
| 95 | unsigned int num_elems_written_per_iteration_y = 1; |
| 96 | unsigned int num_elems_written_per_iteration_z = 1; |
| 97 | |
| 98 | if(kernel_size == 3) |
| 99 | { |
| 100 | if((_conv_stride_x == 1) && (_conv_stride_y == 1)) |
| 101 | { |
| 102 | switch(input->info()->data_type()) |
| 103 | { |
| 104 | // TODO(APPBROWSER-299): Choose the most optimal path and remove others. |
| 105 | #define PROCESS_X_4ELEMENTS_Y_3ELEMENTS_FP16 |
| 106 | |
| 107 | case DataType::F16: |
| 108 | #if defined(PROCESS_X_8ELEMENTS_Y_3ELEMENTS_FP16) |
| 109 | options.emplace("#define PROCESS_X_8ELEMENTS_Y_3ELEMENTS_FP16"); |
| 110 | num_elems_read_per_iteration_x = 16; |
| 111 | num_elems_read_per_iteration_y = 5; |
| 112 | num_elems_written_per_iteration_x = 8; |
| 113 | num_elems_written_per_iteration_y = 3; |
| 114 | #elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS_FP16) |
| 115 | options.emplace("#define PROCESS_X_4ELEMENTS_Y_3ELEMENTS_FP16"); |
| 116 | num_elems_read_per_iteration_x = 8; |
| 117 | num_elems_read_per_iteration_y = 5; |
| 118 | num_elems_written_per_iteration_x = 4; |
| 119 | num_elems_written_per_iteration_y = 3; |
| 120 | #elif defined(PROCESS_X_4ELEMENTS_Y_4ELEMENTS_FP16) |
| 121 | options.emplace("#define PROCESS_X_4ELEMENTS_Y_4ELEMENTS_FP16"); |
| 122 | num_elems_read_per_iteration_x = 8; |
| 123 | num_elems_read_per_iteration_y = 6; |
| 124 | num_elems_written_per_iteration_x = 4; |
| 125 | num_elems_written_per_iteration_y = 4; |
| 126 | #elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS_Z_2ELEMENTS_FP16) |
| 127 | options.emplace("#define PROCESS_X_4ELEMENTS_Y_3ELEMENTS_Z_2ELEMENTS_FP16"); |
| 128 | num_elems_read_per_iteration_x = 8; |
| 129 | num_elems_read_per_iteration_y = 5; |
| 130 | num_elems_written_per_iteration_x = 4; |
| 131 | num_elems_written_per_iteration_y = 3; |
| 132 | num_elems_written_per_iteration_z = 2; |
| 133 | #endif /* PROCESS_X_8ELEMENTS_Y_3ELEMENTS_FP16 */ |
| 134 | break; |
| 135 | |
| 136 | case DataType::F32: |
| 137 | options.emplace("#define PROCESS_X_4ELEMENTS_Y_3ELEMENTS"); |
| 138 | num_elems_read_per_iteration_x = 8; |
| 139 | num_elems_read_per_iteration_y = 5; |
| 140 | num_elems_written_per_iteration_x = 4; |
| 141 | num_elems_written_per_iteration_y = 3; |
| 142 | break; |
| 143 | |
| 144 | default: |
| 145 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 146 | break; |
| 147 | } |
| 148 | } |
| 149 | // FIXME: Just keep one in release |
| 150 | else |
| 151 | { |
| 152 | switch(input->info()->data_type()) |
| 153 | { |
| 154 | case DataType::F16: |
| 155 | options.emplace("#define PROCESS_X_4ELEMENTS_FP16"); |
| 156 | num_elems_read_per_iteration_x = 8; |
| 157 | num_elems_written_per_iteration_x = 4; |
| 158 | break; |
| 159 | |
| 160 | case DataType::F32: |
| 161 | // TODO(APPBROWSER-299): Choose the most optimal path and remove others. |
| 162 | #define PROCESS_4_ELEMENT |
| 163 | |
| 164 | #if defined(PROCESS_1_ELEMENT) |
| 165 | options.emplace("#define PROCESS_1_ELEMENT"); |
| 166 | num_elems_read_per_iteration_x = 3; |
| 167 | num_elems_written_per_iteration_x = 1; |
| 168 | #elif defined(PROCESS_4_ELEMENT) |
| 169 | options.emplace("#define PROCESS_4_ELEMENT"); |
| 170 | num_elems_read_per_iteration_x = 8; |
| 171 | num_elems_written_per_iteration_x = 4; |
| 172 | #elif defined(PROCESS_8_ELEMENT) |
| 173 | options.emplace("#define PROCESS_8_ELEMENT"); |
| 174 | num_elems_read_per_iteration_x = 12; |
| 175 | num_elems_written_per_iteration_x = 8; |
| 176 | #else /* PROCESS_1_ELEMENT */ |
| 177 | #error Have to declare how many elements to process in one thread. |
| 178 | #endif /* PROCESS_1_ELEMENT */ |
| 179 | break; |
| 180 | |
| 181 | default: |
| 182 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 183 | break; |
| 184 | } |
| 185 | } |
| 186 | } |
| 187 | else if(kernel_size == 1) |
| 188 | { |
| 189 | switch(input->info()->data_type()) |
| 190 | { |
| 191 | case DataType::F16: |
| 192 | num_elems_read_per_iteration_x = 8; |
| 193 | num_elems_written_per_iteration_x = 8; |
zhenglin | 666635c | 2017-12-04 14:38:09 +0800 | [diff] [blame] | 194 | if(weights->info()->dimension(2) % 2 == 0) |
| 195 | { |
| 196 | options.emplace("#define WEIGHTS_OPTIMIZATION"); |
| 197 | } |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 198 | break; |
| 199 | |
| 200 | case DataType::F32: |
| 201 | num_elems_read_per_iteration_x = 1; |
| 202 | num_elems_written_per_iteration_x = 1; |
| 203 | break; |
| 204 | |
| 205 | default: |
| 206 | break; |
| 207 | } |
| 208 | } |
| 209 | else if(kernel_size == 5) |
| 210 | { |
| 211 | switch(input->info()->data_type()) |
| 212 | { |
| 213 | case DataType::F16: |
ASIAPAC\steli01 | 23ac91b | 2017-11-07 16:14:44 +0800 | [diff] [blame^] | 214 | options.emplace("#define PROCESS_4X_1Y_1Z"); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 215 | num_elems_read_per_iteration_x = 8; |
| 216 | num_elems_written_per_iteration_x = 4; |
| 217 | |
| 218 | default: |
| 219 | break; |
| 220 | } |
| 221 | } |
| 222 | else |
| 223 | { |
| 224 | } |
| 225 | |
| 226 | if(_bias != nullptr) |
| 227 | { |
| 228 | options.emplace("#define BIAS"); |
| 229 | } |
| 230 | |
| 231 | std::stringstream kernel_name; |
| 232 | kernel_name << "direct_convolution" << kernel_size << "x" << kernel_size; |
| 233 | |
| 234 | _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel(kernel_name.str(), options)); |
| 235 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 236 | unsigned int idx = (_bias == nullptr) ? 3 * num_arguments_per_3D_tensor() : (num_arguments_per_1D_tensor() + 3 * num_arguments_per_3D_tensor()); |
| 237 | |
| 238 | // Calculate output right and bottom border |
| 239 | const int output_width = output->info()->dimension(0); |
| 240 | const int output_height = output->info()->dimension(1); |
| 241 | const int output_padding_right = ceil_to_multiple(output_width, num_elems_written_per_iteration_x * _lws[0]) - output_width; |
| 242 | const int output_padding_bottom = ceil_to_multiple(output_height, num_elems_written_per_iteration_y * _lws[1]) - output_height; |
| 243 | |
| 244 | // Calculate input right and bottom border |
| 245 | const int input_width = input->info()->dimension(0); |
| 246 | const int input_height = input->info()->dimension(1); |
| 247 | const int upper_bound_w = ceil_to_multiple(((output_width + output_padding_right) * _conv_stride_x + (kernel_size - 1)), num_elems_read_per_iteration_x * _lws[0]) - _conv_pad_x - input_width; |
| 248 | const int upper_bound_h = ceil_to_multiple(((output_height + output_padding_bottom) * _conv_stride_y + (kernel_size - 1)), num_elems_read_per_iteration_y * _lws[1]) - _conv_pad_y - input_height; |
| 249 | const int padding_right = std::max(upper_bound_w, _conv_pad_x); |
| 250 | const int padding_bottom = std::max(upper_bound_h, _conv_pad_y); |
| 251 | |
| 252 | BorderSize border = BorderSize(0, output_padding_right, output_padding_bottom, 0); |
| 253 | |
| 254 | Window win = calculate_max_enlarged_window(*output->info(), Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y, num_elems_written_per_iteration_z), border); |
| 255 | |
| 256 | AccessWindowStatic input_access(input->info(), -_conv_pad_x, -_conv_pad_y, input_width + padding_right, input_height + padding_bottom); |
| 257 | AccessWindowStatic weights_access = AccessWindowStatic(nullptr, 0, 0, 0, 0); |
| 258 | AccessWindowStatic bias_access = AccessWindowStatic(nullptr, 0, 0, 0, 1); |
| 259 | |
| 260 | switch(weights->info()->data_type()) |
| 261 | { |
| 262 | case DataType::F16: |
zhenglin | 666635c | 2017-12-04 14:38:09 +0800 | [diff] [blame] | 263 | if((weights->info()->dimension(2) % 2 != 0) || (kernel_size != 1)) |
| 264 | { |
| 265 | weights_access = AccessWindowStatic(weights->info(), 0, 0, kernel_size + 1, kernel_size); |
| 266 | } |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 267 | if(_bias != nullptr) |
| 268 | { |
| 269 | bias_access = AccessWindowStatic(_bias->info(), 0, 0, _bias->info()->dimension(0) + 1, 1); |
| 270 | } |
| 271 | break; |
| 272 | |
| 273 | case DataType::F32: |
| 274 | weights_access = AccessWindowStatic(weights->info(), 0, 0, kernel_size, kernel_size); |
| 275 | if(_bias != nullptr) |
| 276 | { |
| 277 | bias_access = AccessWindowStatic(_bias->info(), 0, 0, _bias->info()->dimension(0), 1); |
| 278 | } |
| 279 | break; |
| 280 | |
| 281 | default: |
| 282 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 283 | break; |
| 284 | } |
| 285 | |
| 286 | AccessWindowStatic output_access(output->info(), 0, 0, output_width + output_padding_right, output_height + output_padding_bottom); |
| 287 | |
| 288 | if(_bias != nullptr) |
| 289 | { |
| 290 | update_window_and_padding(win, input_access, weights_access, bias_access, output_access); |
| 291 | } |
| 292 | else |
| 293 | { |
| 294 | update_window_and_padding(win, input_access, weights_access, output_access); |
| 295 | } |
| 296 | |
| 297 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); |
| 298 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 299 | _kernel.set_argument(idx++, _weights->info()->strides_in_bytes()[3]); // weights_stride_w |
| 300 | _kernel.set_argument(idx++, _weights->info()->dimension(2)); // weights_depth |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 301 | |
| 302 | IGCKernel::configure(win); |
| 303 | } |
| 304 | |
| 305 | template <unsigned int kernel_size> |
| 306 | void GCDirectConvolutionLayerKernel<kernel_size>::run(const Window &window) |
| 307 | { |
| 308 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 309 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 310 | |
| 311 | _kernel.use(); |
| 312 | |
| 313 | // Get initial windows |
| 314 | Window slice = window.first_slice_window_3D(); |
| 315 | Window win_in = window; |
| 316 | |
| 317 | win_in.adjust(Window::DimX, -_conv_pad_x, true); |
| 318 | win_in.adjust(Window::DimY, -_conv_pad_y, true); |
| 319 | win_in.set_dimension_step(Window::DimX, window.x().step() * _conv_stride_x); |
| 320 | win_in.set_dimension_step(Window::DimY, window.y().step() * _conv_stride_y); |
| 321 | |
| 322 | Window slice_in = win_in.first_slice_window_3D(); |
| 323 | |
| 324 | unsigned int idx1 = 2 * num_arguments_per_3D_tensor(); |
| 325 | add_3D_tensor_argument(idx1, _weights, BufferParam(3, 2), slice); |
| 326 | |
| 327 | if(_bias != nullptr) |
| 328 | { |
| 329 | Window slice_bias; |
| 330 | slice_bias.use_tensor_dimensions(_bias->info()->tensor_shape()); |
| 331 | add_1D_tensor_argument(idx1, _bias, BufferParam(4, 2), slice_bias); |
| 332 | } |
| 333 | |
| 334 | do |
| 335 | { |
| 336 | unsigned int idx = 0; |
| 337 | |
| 338 | switch(_input->info()->data_type()) |
| 339 | { |
| 340 | case DataType::F16: |
| 341 | switch(kernel_size) |
| 342 | { |
| 343 | case 1: |
| 344 | add_3D_tensor_argument(idx, _input, BufferParam(1, 4), slice_in); |
| 345 | add_3D_tensor_argument(idx, _output, BufferParam(2, 4), slice); |
| 346 | break; |
| 347 | |
| 348 | case 3: |
| 349 | add_3D_tensor_argument(idx, _input, BufferParam(1, 3), slice_in); |
| 350 | add_3D_tensor_argument(idx, _output, BufferParam(2, 3), slice); |
| 351 | break; |
| 352 | |
| 353 | case 5: |
| 354 | add_3D_tensor_argument(idx, _input, BufferParam(1, 3), slice_in); |
| 355 | add_3D_tensor_argument(idx, _output, BufferParam(2, 3), slice); |
| 356 | break; |
| 357 | |
| 358 | default: |
| 359 | ARM_COMPUTE_ERROR("Current kernel size %d is not supported", kernel_size); |
| 360 | break; |
| 361 | } |
| 362 | break; |
| 363 | |
| 364 | case DataType::F32: |
| 365 | switch(kernel_size) |
| 366 | { |
| 367 | case 1: |
| 368 | case 5: |
| 369 | add_3D_tensor_argument(idx, _input, BufferParam(1, 2), slice_in); |
| 370 | add_3D_tensor_argument(idx, _output, BufferParam(2, 2), slice); |
| 371 | break; |
| 372 | |
| 373 | case 3: |
| 374 | add_3D_tensor_argument(idx, _input, BufferParam(1, 4), slice_in); |
| 375 | add_3D_tensor_argument(idx, _output, BufferParam(2, 4), slice); |
| 376 | break; |
| 377 | |
| 378 | default: |
| 379 | ARM_COMPUTE_ERROR("Current kernel size %d is not supported", kernel_size); |
| 380 | break; |
| 381 | } |
| 382 | break; |
| 383 | |
| 384 | default: |
| 385 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 386 | break; |
| 387 | } |
| 388 | |
| 389 | _kernel.update_shader_params(); |
| 390 | enqueue(*this, slice, _lws); |
| 391 | } |
| 392 | while(window.slide_window_slice_3D(slice) && win_in.slide_window_slice_3D(slice_in)); |
| 393 | } |
| 394 | |
| 395 | template class arm_compute::GCDirectConvolutionLayerKernel<1>; |
| 396 | template class arm_compute::GCDirectConvolutionLayerKernel<3>; |
| 397 | template class arm_compute::GCDirectConvolutionLayerKernel<5>; |