Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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 | #ifndef __UTILS_UTILS_H__ |
| 25 | #define __UTILS_UTILS_H__ |
| 26 | |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
| 29 | #include "arm_compute/core/Types.h" |
| 30 | #include "arm_compute/core/Validate.h" |
steniu01 | bee466b | 2017-06-21 16:45:41 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/Window.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 32 | #include "arm_compute/runtime/Tensor.h" |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 33 | #include "libnpy/npy.hpp" |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 34 | #include "support/ToolchainSupport.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 35 | |
| 36 | #ifdef ARM_COMPUTE_CL |
| 37 | #include "arm_compute/core/CL/OpenCL.h" |
| 38 | #include "arm_compute/runtime/CL/CLTensor.h" |
| 39 | #endif /* ARM_COMPUTE_CL */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 40 | #ifdef ARM_COMPUTE_GC |
| 41 | #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" |
| 42 | #endif /* ARM_COMPUTE_GC */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 43 | |
| 44 | #include <cstdlib> |
| 45 | #include <cstring> |
| 46 | #include <fstream> |
| 47 | #include <iostream> |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 48 | #include <random> |
| 49 | #include <string> |
| 50 | #include <tuple> |
| 51 | #include <vector> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 52 | |
| 53 | namespace arm_compute |
| 54 | { |
| 55 | namespace utils |
| 56 | { |
| 57 | /** Signature of an example to run |
| 58 | * |
| 59 | * @param[in] argc Number of command line arguments |
| 60 | * @param[in] argv Command line arguments |
| 61 | */ |
| 62 | using example = void(int argc, const char **argv); |
| 63 | |
| 64 | /** Run an example and handle the potential exceptions it throws |
| 65 | * |
| 66 | * @param[in] argc Number of command line arguments |
| 67 | * @param[in] argv Command line arguments |
| 68 | * @param[in] func Pointer to the function containing the code to run |
| 69 | */ |
| 70 | int run_example(int argc, const char **argv, example &func); |
| 71 | |
| 72 | /** Draw a RGB rectangular window for the detected object |
| 73 | * |
| 74 | * @param[in, out] tensor Input tensor where the rectangle will be drawn on. Format supported: RGB888 |
| 75 | * @param[in] rect Geometry of the rectangular window |
| 76 | * @param[in] r Red colour to use |
| 77 | * @param[in] g Green colour to use |
| 78 | * @param[in] b Blue colour to use |
| 79 | */ |
| 80 | void draw_detection_rectangle(arm_compute::ITensor *tensor, const arm_compute::DetectionWindow &rect, uint8_t r, uint8_t g, uint8_t b); |
| 81 | |
| 82 | /** Parse the ppm header from an input file stream. At the end of the execution, |
| 83 | * the file position pointer will be located at the first pixel stored in the ppm file |
| 84 | * |
| 85 | * @param[in] fs Input file stream to parse |
| 86 | * |
| 87 | * @return The width, height and max value stored in the header of the PPM file |
| 88 | */ |
| 89 | std::tuple<unsigned int, unsigned int, int> parse_ppm_header(std::ifstream &fs); |
| 90 | |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 91 | /** Parse the npy header from an input file stream. At the end of the execution, |
| 92 | * the file position pointer will be located at the first pixel stored in the npy file //TODO |
| 93 | * |
| 94 | * @param[in] fs Input file stream to parse |
| 95 | * |
| 96 | * @return The width and height stored in the header of the NPY file |
| 97 | */ |
| 98 | std::tuple<std::vector<unsigned long>, bool, std::string> parse_npy_header(std::ifstream &fs); |
| 99 | |
| 100 | /** Obtain numpy type string from DataType. |
| 101 | * |
| 102 | * @param[in] data_type Data type. |
| 103 | * |
| 104 | * @return numpy type string. |
| 105 | */ |
| 106 | inline std::string get_typestring(DataType data_type) |
| 107 | { |
| 108 | // Check endianness |
| 109 | const unsigned int i = 1; |
| 110 | const char *c = reinterpret_cast<const char *>(&i); |
| 111 | std::string endianness; |
| 112 | if(*c == 1) |
| 113 | { |
| 114 | endianness = std::string("<"); |
| 115 | } |
| 116 | else |
| 117 | { |
| 118 | endianness = std::string(">"); |
| 119 | } |
| 120 | const std::string no_endianness("|"); |
| 121 | |
| 122 | switch(data_type) |
| 123 | { |
| 124 | case DataType::U8: |
| 125 | return no_endianness + "u" + support::cpp11::to_string(sizeof(uint8_t)); |
| 126 | case DataType::S8: |
| 127 | return no_endianness + "i" + support::cpp11::to_string(sizeof(int8_t)); |
| 128 | case DataType::U16: |
| 129 | return endianness + "u" + support::cpp11::to_string(sizeof(uint16_t)); |
| 130 | case DataType::S16: |
| 131 | return endianness + "i" + support::cpp11::to_string(sizeof(int16_t)); |
| 132 | case DataType::U32: |
| 133 | return endianness + "u" + support::cpp11::to_string(sizeof(uint32_t)); |
| 134 | case DataType::S32: |
| 135 | return endianness + "i" + support::cpp11::to_string(sizeof(int32_t)); |
| 136 | case DataType::U64: |
| 137 | return endianness + "u" + support::cpp11::to_string(sizeof(uint64_t)); |
| 138 | case DataType::S64: |
| 139 | return endianness + "i" + support::cpp11::to_string(sizeof(int64_t)); |
| 140 | case DataType::F32: |
| 141 | return endianness + "f" + support::cpp11::to_string(sizeof(float)); |
| 142 | case DataType::F64: |
| 143 | return endianness + "f" + support::cpp11::to_string(sizeof(double)); |
| 144 | case DataType::SIZET: |
| 145 | return endianness + "u" + support::cpp11::to_string(sizeof(size_t)); |
| 146 | default: |
| 147 | ARM_COMPUTE_ERROR("NOT SUPPORTED!"); |
| 148 | } |
| 149 | } |
| 150 | |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 151 | /** Maps a tensor if needed |
| 152 | * |
| 153 | * @param[in] tensor Tensor to be mapped |
| 154 | * @param[in] blocking Specified if map is blocking or not |
| 155 | */ |
| 156 | template <typename T> |
Gian Marco Iodice | ae27e94 | 2017-09-28 18:31:26 +0100 | [diff] [blame] | 157 | inline void map(T &tensor, bool blocking) |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 158 | { |
| 159 | ARM_COMPUTE_UNUSED(tensor); |
| 160 | ARM_COMPUTE_UNUSED(blocking); |
| 161 | } |
| 162 | |
| 163 | /** Unmaps a tensor if needed |
| 164 | * |
| 165 | * @param tensor Tensor to be unmapped |
| 166 | */ |
| 167 | template <typename T> |
Gian Marco Iodice | ae27e94 | 2017-09-28 18:31:26 +0100 | [diff] [blame] | 168 | inline void unmap(T &tensor) |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 169 | { |
| 170 | ARM_COMPUTE_UNUSED(tensor); |
| 171 | } |
| 172 | |
| 173 | #ifdef ARM_COMPUTE_CL |
| 174 | /** Maps a tensor if needed |
| 175 | * |
| 176 | * @param[in] tensor Tensor to be mapped |
| 177 | * @param[in] blocking Specified if map is blocking or not |
| 178 | */ |
Gian Marco Iodice | ae27e94 | 2017-09-28 18:31:26 +0100 | [diff] [blame] | 179 | inline void map(CLTensor &tensor, bool blocking) |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 180 | { |
| 181 | tensor.map(blocking); |
| 182 | } |
| 183 | |
| 184 | /** Unmaps a tensor if needed |
| 185 | * |
| 186 | * @param tensor Tensor to be unmapped |
| 187 | */ |
Gian Marco Iodice | ae27e94 | 2017-09-28 18:31:26 +0100 | [diff] [blame] | 188 | inline void unmap(CLTensor &tensor) |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 189 | { |
| 190 | tensor.unmap(); |
| 191 | } |
| 192 | #endif /* ARM_COMPUTE_CL */ |
| 193 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 194 | #ifdef ARM_COMPUTE_GC |
| 195 | /** Maps a tensor if needed |
| 196 | * |
| 197 | * @param[in] tensor Tensor to be mapped |
| 198 | * @param[in] blocking Specified if map is blocking or not |
| 199 | */ |
| 200 | inline void map(GCTensor &tensor, bool blocking) |
| 201 | { |
| 202 | tensor.map(blocking); |
| 203 | } |
| 204 | |
| 205 | /** Unmaps a tensor if needed |
| 206 | * |
| 207 | * @param tensor Tensor to be unmapped |
| 208 | */ |
| 209 | inline void unmap(GCTensor &tensor) |
| 210 | { |
| 211 | tensor.unmap(); |
| 212 | } |
| 213 | #endif /* ARM_COMPUTE_GC */ |
| 214 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 215 | /** Class to load the content of a PPM file into an Image |
| 216 | */ |
| 217 | class PPMLoader |
| 218 | { |
| 219 | public: |
| 220 | PPMLoader() |
| 221 | : _fs(), _width(0), _height(0) |
| 222 | { |
| 223 | } |
| 224 | /** Open a PPM file and reads its metadata (Width, height) |
| 225 | * |
| 226 | * @param[in] ppm_filename File to open |
| 227 | */ |
| 228 | void open(const std::string &ppm_filename) |
| 229 | { |
| 230 | ARM_COMPUTE_ERROR_ON(is_open()); |
| 231 | try |
| 232 | { |
| 233 | _fs.exceptions(std::ifstream::failbit | std::ifstream::badbit); |
| 234 | _fs.open(ppm_filename, std::ios::in | std::ios::binary); |
| 235 | |
| 236 | unsigned int max_val = 0; |
| 237 | std::tie(_width, _height, max_val) = parse_ppm_header(_fs); |
| 238 | |
| 239 | ARM_COMPUTE_ERROR_ON_MSG(max_val >= 256, "2 bytes per colour channel not supported in file %s", ppm_filename.c_str()); |
| 240 | } |
| 241 | catch(const std::ifstream::failure &e) |
| 242 | { |
| 243 | ARM_COMPUTE_ERROR("Accessing %s: %s", ppm_filename.c_str(), e.what()); |
| 244 | } |
| 245 | } |
| 246 | /** Return true if a PPM file is currently open |
| 247 | */ |
| 248 | bool is_open() |
| 249 | { |
| 250 | return _fs.is_open(); |
| 251 | } |
| 252 | |
| 253 | /** Initialise an image's metadata with the dimensions of the PPM file currently open |
| 254 | * |
| 255 | * @param[out] image Image to initialise |
| 256 | * @param[in] format Format to use for the image (Must be RGB888 or U8) |
| 257 | */ |
| 258 | template <typename T> |
| 259 | void init_image(T &image, arm_compute::Format format) |
| 260 | { |
| 261 | ARM_COMPUTE_ERROR_ON(!is_open()); |
| 262 | ARM_COMPUTE_ERROR_ON(format != arm_compute::Format::RGB888 && format != arm_compute::Format::U8); |
| 263 | |
| 264 | // Use the size of the input PPM image |
| 265 | arm_compute::TensorInfo image_info(_width, _height, format); |
| 266 | image.allocator()->init(image_info); |
| 267 | } |
| 268 | |
| 269 | /** Fill an image with the content of the currently open PPM file. |
| 270 | * |
| 271 | * @note If the image is a CLImage, the function maps and unmaps the image |
| 272 | * |
| 273 | * @param[in,out] image Image to fill (Must be allocated, and of matching dimensions with the opened PPM). |
| 274 | */ |
| 275 | template <typename T> |
| 276 | void fill_image(T &image) |
| 277 | { |
| 278 | ARM_COMPUTE_ERROR_ON(!is_open()); |
| 279 | ARM_COMPUTE_ERROR_ON(image.info()->dimension(0) != _width || image.info()->dimension(1) != _height); |
| 280 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(&image, arm_compute::Format::U8, arm_compute::Format::RGB888); |
| 281 | try |
| 282 | { |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 283 | // Map buffer if creating a CLTensor/GCTensor |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 284 | map(image, true); |
| 285 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 286 | // Check if the file is large enough to fill the image |
| 287 | const size_t current_position = _fs.tellg(); |
| 288 | _fs.seekg(0, std::ios_base::end); |
| 289 | const size_t end_position = _fs.tellg(); |
| 290 | _fs.seekg(current_position, std::ios_base::beg); |
| 291 | |
| 292 | ARM_COMPUTE_ERROR_ON_MSG((end_position - current_position) < image.info()->tensor_shape().total_size() * image.info()->element_size(), |
| 293 | "Not enough data in file"); |
| 294 | ARM_COMPUTE_UNUSED(end_position); |
| 295 | |
| 296 | switch(image.info()->format()) |
| 297 | { |
| 298 | case arm_compute::Format::U8: |
| 299 | { |
| 300 | // We need to convert the data from RGB to grayscale: |
| 301 | // Iterate through every pixel of the image |
| 302 | arm_compute::Window window; |
| 303 | window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, _width, 1)); |
| 304 | window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, _height, 1)); |
| 305 | |
| 306 | arm_compute::Iterator out(&image, window); |
| 307 | |
| 308 | unsigned char red = 0; |
| 309 | unsigned char green = 0; |
| 310 | unsigned char blue = 0; |
| 311 | |
| 312 | arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id) |
| 313 | { |
| 314 | red = _fs.get(); |
| 315 | green = _fs.get(); |
| 316 | blue = _fs.get(); |
| 317 | |
| 318 | *out.ptr() = 0.2126f * red + 0.7152f * green + 0.0722f * blue; |
| 319 | }, |
| 320 | out); |
| 321 | |
| 322 | break; |
| 323 | } |
| 324 | case arm_compute::Format::RGB888: |
| 325 | { |
| 326 | // There is no format conversion needed: we can simply copy the content of the input file to the image one row at the time. |
| 327 | // Create a vertical window to iterate through the image's rows: |
| 328 | arm_compute::Window window; |
| 329 | window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, _height, 1)); |
| 330 | |
| 331 | arm_compute::Iterator out(&image, window); |
| 332 | |
| 333 | arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id) |
| 334 | { |
| 335 | // Copy one row from the input file to the current row of the image: |
| 336 | _fs.read(reinterpret_cast<std::fstream::char_type *>(out.ptr()), _width * image.info()->element_size()); |
| 337 | }, |
| 338 | out); |
| 339 | |
| 340 | break; |
| 341 | } |
| 342 | default: |
| 343 | ARM_COMPUTE_ERROR("Unsupported format"); |
| 344 | } |
| 345 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 346 | // Unmap buffer if creating a CLTensor/GCTensor |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 347 | unmap(image); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 348 | } |
| 349 | catch(const std::ifstream::failure &e) |
| 350 | { |
| 351 | ARM_COMPUTE_ERROR("Loading PPM file: %s", e.what()); |
| 352 | } |
| 353 | } |
| 354 | |
Gian Marco | 44ec2e7 | 2017-10-19 14:13:38 +0100 | [diff] [blame] | 355 | /** Fill a tensor with 3 planes (one for each channel) with the content of the currently open PPM file. |
| 356 | * |
| 357 | * @note If the image is a CLImage, the function maps and unmaps the image |
| 358 | * |
| 359 | * @param[in,out] tensor Tensor with 3 planes to fill (Must be allocated, and of matching dimensions with the opened PPM). Data types supported: U8/F32 |
| 360 | * @param[in] bgr (Optional) Fill the first plane with blue channel (default = false) |
| 361 | */ |
| 362 | template <typename T> |
| 363 | void fill_planar_tensor(T &tensor, bool bgr = false) |
| 364 | { |
| 365 | ARM_COMPUTE_ERROR_ON(!is_open()); |
| 366 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::U8, DataType::F32); |
| 367 | ARM_COMPUTE_ERROR_ON(tensor.info()->dimension(0) != _width || tensor.info()->dimension(1) != _height || tensor.info()->dimension(2) != 3); |
| 368 | |
| 369 | try |
| 370 | { |
| 371 | // Map buffer if creating a CLTensor |
| 372 | map(tensor, true); |
| 373 | |
| 374 | // Check if the file is large enough to fill the image |
| 375 | const size_t current_position = _fs.tellg(); |
| 376 | _fs.seekg(0, std::ios_base::end); |
| 377 | const size_t end_position = _fs.tellg(); |
| 378 | _fs.seekg(current_position, std::ios_base::beg); |
| 379 | |
| 380 | ARM_COMPUTE_ERROR_ON_MSG((end_position - current_position) < tensor.info()->tensor_shape().total_size(), |
| 381 | "Not enough data in file"); |
| 382 | ARM_COMPUTE_UNUSED(end_position); |
| 383 | |
| 384 | // Iterate through every pixel of the image |
| 385 | arm_compute::Window window; |
| 386 | window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, _width, 1)); |
| 387 | window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, _height, 1)); |
| 388 | window.set(arm_compute::Window::DimZ, arm_compute::Window::Dimension(0, 1, 1)); |
| 389 | |
| 390 | arm_compute::Iterator out(&tensor, window); |
| 391 | |
| 392 | unsigned char red = 0; |
| 393 | unsigned char green = 0; |
| 394 | unsigned char blue = 0; |
| 395 | |
| 396 | size_t stride_z = tensor.info()->strides_in_bytes()[2]; |
| 397 | |
| 398 | arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id) |
| 399 | { |
| 400 | red = _fs.get(); |
| 401 | green = _fs.get(); |
| 402 | blue = _fs.get(); |
| 403 | |
| 404 | switch(tensor.info()->data_type()) |
| 405 | { |
| 406 | case arm_compute::DataType::U8: |
| 407 | { |
| 408 | *(out.ptr() + 0 * stride_z) = bgr ? blue : red; |
| 409 | *(out.ptr() + 1 * stride_z) = green; |
| 410 | *(out.ptr() + 2 * stride_z) = bgr ? red : blue; |
| 411 | break; |
| 412 | } |
| 413 | case arm_compute::DataType::F32: |
| 414 | { |
| 415 | *reinterpret_cast<float *>(out.ptr() + 0 * stride_z) = static_cast<float>(bgr ? blue : red); |
| 416 | *reinterpret_cast<float *>(out.ptr() + 1 * stride_z) = static_cast<float>(green); |
| 417 | *reinterpret_cast<float *>(out.ptr() + 2 * stride_z) = static_cast<float>(bgr ? red : blue); |
| 418 | break; |
| 419 | } |
| 420 | default: |
| 421 | { |
| 422 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 423 | } |
| 424 | } |
| 425 | }, |
| 426 | out); |
| 427 | |
| 428 | // Unmap buffer if creating a CLTensor |
| 429 | unmap(tensor); |
| 430 | } |
| 431 | catch(const std::ifstream::failure &e) |
| 432 | { |
| 433 | ARM_COMPUTE_ERROR("Loading PPM file: %s", e.what()); |
| 434 | } |
| 435 | } |
| 436 | |
Isabella Gottardi | a4c6188 | 2017-11-03 12:11:55 +0000 | [diff] [blame] | 437 | /** Return the width of the currently open PPM file. |
| 438 | */ |
| 439 | unsigned int width() const |
| 440 | { |
| 441 | return _width; |
| 442 | } |
| 443 | |
| 444 | /** Return the height of the currently open PPM file. |
| 445 | */ |
| 446 | unsigned int height() const |
| 447 | { |
| 448 | return _height; |
| 449 | } |
| 450 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 451 | private: |
| 452 | std::ifstream _fs; |
| 453 | unsigned int _width, _height; |
| 454 | }; |
| 455 | |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 456 | class NPYLoader |
| 457 | { |
| 458 | public: |
| 459 | NPYLoader() |
| 460 | : _fs(), _shape(), _fortran_order(false), _typestring() |
| 461 | { |
| 462 | } |
| 463 | |
| 464 | /** Open a NPY file and reads its metadata |
| 465 | * |
| 466 | * @param[in] npy_filename File to open |
| 467 | */ |
| 468 | void open(const std::string &npy_filename) |
| 469 | { |
| 470 | ARM_COMPUTE_ERROR_ON(is_open()); |
| 471 | try |
| 472 | { |
| 473 | _fs.exceptions(std::ifstream::failbit | std::ifstream::badbit); |
| 474 | _fs.open(npy_filename, std::ios::in | std::ios::binary); |
| 475 | |
| 476 | std::tie(_shape, _fortran_order, _typestring) = parse_npy_header(_fs); |
| 477 | } |
| 478 | catch(const std::ifstream::failure &e) |
| 479 | { |
| 480 | ARM_COMPUTE_ERROR("Accessing %s: %s", npy_filename.c_str(), e.what()); |
| 481 | } |
| 482 | } |
| 483 | /** Return true if a NPY file is currently open */ |
| 484 | bool is_open() |
| 485 | { |
| 486 | return _fs.is_open(); |
| 487 | } |
| 488 | |
| 489 | /** Return true if a NPY file is in fortran order */ |
| 490 | bool is_fortran() |
| 491 | { |
| 492 | return _fortran_order; |
| 493 | } |
| 494 | |
| 495 | /** Initialise an image's metadata with the dimensions of the NPY file currently open |
| 496 | * |
| 497 | * @param[out] tensor Tensor to initialise |
| 498 | * @param[in] format Format to use for the image |
| 499 | */ |
| 500 | template <typename T> |
| 501 | void init_tensor(T &tensor, arm_compute::Format format) |
| 502 | { |
| 503 | ARM_COMPUTE_ERROR_ON(!is_open()); |
| 504 | ARM_COMPUTE_ERROR_ON(format != arm_compute::Format::F32); |
| 505 | |
| 506 | // Use the size of the input NPY tensor |
| 507 | TensorShape shape; |
| 508 | shape.set_num_dimensions(_shape.size()); |
| 509 | for(size_t i = 0; i < _shape.size(); ++i) |
| 510 | { |
| 511 | shape.set(i, _shape.at(i)); |
| 512 | } |
| 513 | |
| 514 | arm_compute::TensorInfo tensor_info(shape, format); |
| 515 | tensor.allocator()->init(tensor_info); |
| 516 | } |
| 517 | |
| 518 | /** Fill a tensor with the content of the currently open NPY file. |
| 519 | * |
| 520 | * @note If the tensor is a CLTensor, the function maps and unmaps the tensor |
| 521 | * |
| 522 | * @param[in,out] tensor Tensor to fill (Must be allocated, and of matching dimensions with the opened NPY). |
| 523 | */ |
| 524 | template <typename T> |
| 525 | void fill_tensor(T &tensor) |
| 526 | { |
| 527 | ARM_COMPUTE_ERROR_ON(!is_open()); |
| 528 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(&tensor, arm_compute::Format::F32); |
| 529 | try |
| 530 | { |
| 531 | // Map buffer if creating a CLTensor |
| 532 | map(tensor, true); |
| 533 | |
| 534 | // Check if the file is large enough to fill the tensor |
| 535 | const size_t current_position = _fs.tellg(); |
| 536 | _fs.seekg(0, std::ios_base::end); |
| 537 | const size_t end_position = _fs.tellg(); |
| 538 | _fs.seekg(current_position, std::ios_base::beg); |
| 539 | |
| 540 | ARM_COMPUTE_ERROR_ON_MSG((end_position - current_position) < tensor.info()->tensor_shape().total_size() * tensor.info()->element_size(), |
| 541 | "Not enough data in file"); |
| 542 | ARM_COMPUTE_UNUSED(end_position); |
| 543 | |
| 544 | // Check if the typestring matches the given one |
| 545 | std::string expect_typestr = get_typestring(tensor.info()->data_type()); |
| 546 | ARM_COMPUTE_ERROR_ON_MSG(_typestring != expect_typestr, "Typestrings mismatch"); |
| 547 | |
| 548 | // Validate tensor shape |
| 549 | ARM_COMPUTE_ERROR_ON_MSG(_shape.size() != tensor.shape().num_dimensions(), "Tensor ranks mismatch"); |
| 550 | if(_fortran_order) |
| 551 | { |
| 552 | for(size_t i = 0; i < _shape.size(); ++i) |
| 553 | { |
| 554 | ARM_COMPUTE_ERROR_ON_MSG(tensor.shape()[i] != _shape[i], "Tensor dimensions mismatch"); |
| 555 | } |
| 556 | } |
| 557 | else |
| 558 | { |
| 559 | for(size_t i = 0; i < _shape.size(); ++i) |
| 560 | { |
| 561 | ARM_COMPUTE_ERROR_ON_MSG(tensor.shape()[i] != _shape[_shape.size() - i - 1], "Tensor dimensions mismatch"); |
| 562 | } |
| 563 | } |
| 564 | |
| 565 | switch(tensor.info()->format()) |
| 566 | { |
| 567 | case arm_compute::Format::F32: |
| 568 | { |
| 569 | // Read data |
| 570 | if(tensor.info()->padding().empty()) |
| 571 | { |
| 572 | // If tensor has no padding read directly from stream. |
| 573 | _fs.read(reinterpret_cast<char *>(tensor.buffer()), tensor.info()->total_size()); |
| 574 | } |
| 575 | else |
| 576 | { |
| 577 | // If tensor has padding accessing tensor elements through execution window. |
| 578 | Window window; |
| 579 | window.use_tensor_dimensions(tensor.info()->tensor_shape()); |
| 580 | |
| 581 | execute_window_loop(window, [&](const Coordinates & id) |
| 582 | { |
| 583 | _fs.read(reinterpret_cast<char *>(tensor.ptr_to_element(id)), tensor.info()->element_size()); |
| 584 | }); |
| 585 | } |
| 586 | |
| 587 | break; |
| 588 | } |
| 589 | default: |
| 590 | ARM_COMPUTE_ERROR("Unsupported format"); |
| 591 | } |
| 592 | |
| 593 | // Unmap buffer if creating a CLTensor |
| 594 | unmap(tensor); |
| 595 | } |
| 596 | catch(const std::ifstream::failure &e) |
| 597 | { |
| 598 | ARM_COMPUTE_ERROR("Loading NPY file: %s", e.what()); |
| 599 | } |
| 600 | } |
| 601 | |
| 602 | private: |
| 603 | std::ifstream _fs; |
| 604 | std::vector<unsigned long> _shape; |
| 605 | bool _fortran_order; |
| 606 | std::string _typestring; |
| 607 | }; |
| 608 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 609 | /** Template helper function to save a tensor image to a PPM file. |
| 610 | * |
| 611 | * @note Only U8 and RGB888 formats supported. |
| 612 | * @note Only works with 2D tensors. |
| 613 | * @note If the input tensor is a CLTensor, the function maps and unmaps the image |
| 614 | * |
| 615 | * @param[in] tensor The tensor to save as PPM file |
| 616 | * @param[in] ppm_filename Filename of the file to create. |
| 617 | */ |
| 618 | template <typename T> |
| 619 | void save_to_ppm(T &tensor, const std::string &ppm_filename) |
| 620 | { |
| 621 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(&tensor, arm_compute::Format::RGB888, arm_compute::Format::U8); |
| 622 | ARM_COMPUTE_ERROR_ON(tensor.info()->num_dimensions() > 2); |
| 623 | |
| 624 | std::ofstream fs; |
| 625 | |
| 626 | try |
| 627 | { |
| 628 | fs.exceptions(std::ofstream::failbit | std::ofstream::badbit | std::ofstream::eofbit); |
| 629 | fs.open(ppm_filename, std::ios::out | std::ios::binary); |
| 630 | |
| 631 | const unsigned int width = tensor.info()->tensor_shape()[0]; |
| 632 | const unsigned int height = tensor.info()->tensor_shape()[1]; |
| 633 | |
| 634 | fs << "P6\n" |
| 635 | << width << " " << height << " 255\n"; |
| 636 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 637 | // Map buffer if creating a CLTensor/GCTensor |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 638 | map(tensor, true); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 639 | |
| 640 | switch(tensor.info()->format()) |
| 641 | { |
| 642 | case arm_compute::Format::U8: |
| 643 | { |
| 644 | arm_compute::Window window; |
| 645 | window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, width, 1)); |
| 646 | window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, height, 1)); |
| 647 | |
| 648 | arm_compute::Iterator in(&tensor, window); |
| 649 | |
| 650 | arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id) |
| 651 | { |
| 652 | const unsigned char value = *in.ptr(); |
| 653 | |
| 654 | fs << value << value << value; |
| 655 | }, |
| 656 | in); |
| 657 | |
| 658 | break; |
| 659 | } |
| 660 | case arm_compute::Format::RGB888: |
| 661 | { |
| 662 | arm_compute::Window window; |
| 663 | window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, width, width)); |
| 664 | window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, height, 1)); |
| 665 | |
| 666 | arm_compute::Iterator in(&tensor, window); |
| 667 | |
| 668 | arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id) |
| 669 | { |
| 670 | fs.write(reinterpret_cast<std::fstream::char_type *>(in.ptr()), width * tensor.info()->element_size()); |
| 671 | }, |
| 672 | in); |
| 673 | |
| 674 | break; |
| 675 | } |
| 676 | default: |
| 677 | ARM_COMPUTE_ERROR("Unsupported format"); |
| 678 | } |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 679 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 680 | // Unmap buffer if creating a CLTensor/GCTensor |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 681 | unmap(tensor); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 682 | } |
| 683 | catch(const std::ofstream::failure &e) |
| 684 | { |
| 685 | ARM_COMPUTE_ERROR("Writing %s: (%s)", ppm_filename.c_str(), e.what()); |
| 686 | } |
| 687 | } |
steniu01 | bee466b | 2017-06-21 16:45:41 +0100 | [diff] [blame] | 688 | |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 689 | /** Template helper function to save a tensor image to a NPY file. |
| 690 | * |
| 691 | * @note Only F32 format supported. |
| 692 | * @note Only works with 2D tensors. |
| 693 | * @note If the input tensor is a CLTensor, the function maps and unmaps the image |
| 694 | * |
| 695 | * @param[in] tensor The tensor to save as NPY file |
| 696 | * @param[in] npy_filename Filename of the file to create. |
| 697 | * @param[in] fortran_order If true, save matrix in fortran order. |
| 698 | */ |
| 699 | template <typename T> |
| 700 | void save_to_npy(T &tensor, const std::string &npy_filename, bool fortran_order) |
| 701 | { |
| 702 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(&tensor, arm_compute::Format::F32); |
| 703 | ARM_COMPUTE_ERROR_ON(tensor.info()->num_dimensions() > 2); |
| 704 | |
| 705 | std::ofstream fs; |
| 706 | |
| 707 | try |
| 708 | { |
| 709 | fs.exceptions(std::ofstream::failbit | std::ofstream::badbit | std::ofstream::eofbit); |
| 710 | fs.open(npy_filename, std::ios::out | std::ios::binary); |
| 711 | |
| 712 | const unsigned int width = tensor.info()->tensor_shape()[0]; |
| 713 | const unsigned int height = tensor.info()->tensor_shape()[1]; |
| 714 | unsigned long shape[2]; |
| 715 | |
| 716 | if(!fortran_order) |
| 717 | { |
| 718 | shape[0] = height, shape[1] = width; |
| 719 | } |
| 720 | else |
| 721 | { |
| 722 | shape[0] = width, shape[1] = height; |
| 723 | } |
| 724 | |
| 725 | // Map buffer if creating a CLTensor |
| 726 | map(tensor, true); |
| 727 | |
| 728 | switch(tensor.info()->format()) |
| 729 | { |
| 730 | case arm_compute::Format::F32: |
| 731 | { |
| 732 | std::vector<float> tmp; /* Used only to get the typestring */ |
| 733 | npy::Typestring typestring_o{ tmp }; |
| 734 | std::string typestring = typestring_o.str(); |
| 735 | |
| 736 | std::ofstream stream(npy_filename, std::ofstream::binary); |
| 737 | npy::WriteHeader(stream, typestring, fortran_order, 2, shape); |
| 738 | |
| 739 | arm_compute::Window window; |
| 740 | window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, width, 1)); |
| 741 | window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, height, 1)); |
| 742 | |
| 743 | arm_compute::Iterator in(&tensor, window); |
| 744 | |
| 745 | arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id) |
| 746 | { |
| 747 | stream.write(reinterpret_cast<const char *>(in.ptr()), sizeof(float)); |
| 748 | }, |
| 749 | in); |
| 750 | |
| 751 | break; |
| 752 | } |
| 753 | default: |
| 754 | ARM_COMPUTE_ERROR("Unsupported format"); |
| 755 | } |
| 756 | |
| 757 | // Unmap buffer if creating a CLTensor |
| 758 | unmap(tensor); |
| 759 | } |
| 760 | catch(const std::ofstream::failure &e) |
| 761 | { |
| 762 | ARM_COMPUTE_ERROR("Writing %s: (%s)", npy_filename.c_str(), e.what()); |
| 763 | } |
| 764 | } |
| 765 | |
steniu01 | bee466b | 2017-06-21 16:45:41 +0100 | [diff] [blame] | 766 | /** Load the tensor with pre-trained data from a binary file |
| 767 | * |
| 768 | * @param[in] tensor The tensor to be filled. Data type supported: F32. |
| 769 | * @param[in] filename Filename of the binary file to load from. |
| 770 | */ |
| 771 | template <typename T> |
| 772 | void load_trained_data(T &tensor, const std::string &filename) |
| 773 | { |
| 774 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::F32); |
| 775 | |
| 776 | std::ifstream fs; |
| 777 | |
| 778 | try |
| 779 | { |
| 780 | fs.exceptions(std::ofstream::failbit | std::ofstream::badbit | std::ofstream::eofbit); |
| 781 | // Open file |
| 782 | fs.open(filename, std::ios::in | std::ios::binary); |
| 783 | |
| 784 | if(!fs.good()) |
| 785 | { |
| 786 | throw std::runtime_error("Could not load binary data: " + filename); |
| 787 | } |
| 788 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 789 | // Map buffer if creating a CLTensor/GCTensor |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 790 | map(tensor, true); |
| 791 | |
steniu01 | bee466b | 2017-06-21 16:45:41 +0100 | [diff] [blame] | 792 | Window window; |
| 793 | |
| 794 | window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, 1, 1)); |
| 795 | |
| 796 | for(unsigned int d = 1; d < tensor.info()->num_dimensions(); ++d) |
| 797 | { |
| 798 | window.set(d, Window::Dimension(0, tensor.info()->tensor_shape()[d], 1)); |
| 799 | } |
| 800 | |
| 801 | arm_compute::Iterator in(&tensor, window); |
| 802 | |
| 803 | execute_window_loop(window, [&](const Coordinates & id) |
| 804 | { |
| 805 | fs.read(reinterpret_cast<std::fstream::char_type *>(in.ptr()), tensor.info()->tensor_shape()[0] * tensor.info()->element_size()); |
| 806 | }, |
| 807 | in); |
| 808 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 809 | // Unmap buffer if creating a CLTensor/GCTensor |
Georgios Pinitas | dc836b6 | 2017-09-20 14:02:37 +0100 | [diff] [blame] | 810 | unmap(tensor); |
steniu01 | bee466b | 2017-06-21 16:45:41 +0100 | [diff] [blame] | 811 | } |
| 812 | catch(const std::ofstream::failure &e) |
| 813 | { |
| 814 | ARM_COMPUTE_ERROR("Writing %s: (%s)", filename.c_str(), e.what()); |
| 815 | } |
| 816 | } |
| 817 | |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 818 | template <typename T> |
| 819 | void fill_random_tensor(T &tensor, float lower_bound, float upper_bound) |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 820 | { |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 821 | std::random_device rd; |
| 822 | std::mt19937 gen(rd()); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 823 | |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 824 | TensorShape shape(tensor.info()->dimension(0), tensor.info()->dimension(1)); |
| 825 | |
| 826 | Window window; |
| 827 | window.set(Window::DimX, Window::Dimension(0, shape.x(), 1)); |
| 828 | window.set(Window::DimY, Window::Dimension(0, shape.y(), 1)); |
| 829 | |
| 830 | map(tensor, true); |
| 831 | |
| 832 | Iterator it(&tensor, window); |
| 833 | |
| 834 | switch(tensor.info()->format()) |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 835 | { |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 836 | case arm_compute::Format::F32: |
| 837 | { |
| 838 | std::uniform_real_distribution<float> dist(lower_bound, upper_bound); |
| 839 | |
| 840 | execute_window_loop(window, [&](const Coordinates & id) |
| 841 | { |
| 842 | *reinterpret_cast<float *>(it.ptr()) = dist(gen); |
| 843 | }, |
| 844 | it); |
| 845 | |
| 846 | break; |
| 847 | } |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 848 | default: |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 849 | { |
| 850 | ARM_COMPUTE_ERROR("Unsupported format"); |
| 851 | } |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 852 | } |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 853 | |
| 854 | unmap(tensor); |
Anthony Barbier | 2a07e18 | 2017-08-04 18:20:27 +0100 | [diff] [blame] | 855 | } |
Giorgio Arena | cf3935f | 2017-10-26 17:14:13 +0100 | [diff] [blame] | 856 | |
| 857 | template <typename T> |
| 858 | void init_sgemm_output(T &dst, T &src0, T &src1, arm_compute::Format format) |
| 859 | { |
| 860 | dst.allocator()->init(TensorInfo(src1.info()->dimension(0), src0.info()->dimension(1), format)); |
| 861 | } |
| 862 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 863 | } // namespace utils |
| 864 | } // namespace arm_compute |
| 865 | #endif /* __UTILS_UTILS_H__*/ |