Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 1 | |
| 2 | // Copyright (c) 2020-2021, ARM Limited. |
| 3 | // |
| 4 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | // you may not use this file except in compliance with the License. |
| 6 | // You may obtain a copy of the License at |
| 7 | // |
| 8 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | // |
| 10 | // Unless required by applicable law or agreed to in writing, software |
| 11 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | // See the License for the specific language governing permissions and |
| 14 | // limitations under the License. |
| 15 | |
| 16 | #include "numpy_utils.h" |
James Ward | 485a11d | 2022-08-05 13:48:37 +0100 | [diff] [blame^] | 17 | #include "half.hpp" |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 18 | |
| 19 | // Magic NUMPY header |
| 20 | static const char NUMPY_HEADER_STR[] = "\x93NUMPY\x1\x0\x76\x0{"; |
| 21 | static const int NUMPY_HEADER_SZ = 128; |
Jeremy Johnson | 82dbb32 | 2021-07-08 11:53:04 +0100 | [diff] [blame] | 22 | // Maximum shape dimensions supported |
| 23 | static const int NUMPY_MAX_DIMS_SUPPORTED = 10; |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 24 | |
| 25 | NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, bool* databuf) |
| 26 | { |
| 27 | const char dtype_str[] = "'|b1'"; |
| 28 | return readFromNpyFileCommon(filename, dtype_str, 1, elems, databuf, true); |
| 29 | } |
| 30 | |
| 31 | NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, int32_t* databuf) |
| 32 | { |
| 33 | const char dtype_str[] = "'<i4'"; |
| 34 | return readFromNpyFileCommon(filename, dtype_str, sizeof(int32_t), elems, databuf, false); |
| 35 | } |
| 36 | |
| 37 | NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, int64_t* databuf) |
| 38 | { |
| 39 | const char dtype_str[] = "'<i8'"; |
| 40 | return readFromNpyFileCommon(filename, dtype_str, sizeof(int64_t), elems, databuf, false); |
| 41 | } |
| 42 | |
| 43 | NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, float* databuf) |
| 44 | { |
| 45 | const char dtype_str[] = "'<f4'"; |
| 46 | return readFromNpyFileCommon(filename, dtype_str, sizeof(float), elems, databuf, false); |
| 47 | } |
| 48 | |
James Ward | 485a11d | 2022-08-05 13:48:37 +0100 | [diff] [blame^] | 49 | NumpyUtilities::NPError |
| 50 | NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, half_float::half* databuf) |
| 51 | { |
| 52 | const char dtype_str[] = "'<f2'"; |
| 53 | return readFromNpyFileCommon(filename, dtype_str, sizeof(half_float::half), elems, databuf, false); |
| 54 | } |
| 55 | |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 56 | NumpyUtilities::NPError NumpyUtilities::readFromNpyFileCommon(const char* filename, |
| 57 | const char* dtype_str, |
| 58 | const size_t elementsize, |
| 59 | const uint32_t elems, |
| 60 | void* databuf, |
| 61 | bool bool_translate) |
| 62 | { |
| 63 | FILE* infile = nullptr; |
| 64 | NPError rc = NO_ERROR; |
| 65 | |
| 66 | assert(filename); |
| 67 | assert(databuf); |
| 68 | |
| 69 | infile = fopen(filename, "rb"); |
| 70 | if (!infile) |
| 71 | { |
| 72 | return FILE_NOT_FOUND; |
| 73 | } |
| 74 | |
| 75 | rc = checkNpyHeader(infile, elems, dtype_str); |
| 76 | if (rc == NO_ERROR) |
| 77 | { |
| 78 | if (bool_translate) |
| 79 | { |
| 80 | // Read in the data from numpy byte array to native bool |
| 81 | // array format |
| 82 | bool* buf = reinterpret_cast<bool*>(databuf); |
| 83 | for (uint32_t i = 0; i < elems; i++) |
| 84 | { |
| 85 | int val = fgetc(infile); |
| 86 | |
| 87 | if (val == EOF) |
| 88 | { |
| 89 | rc = FILE_IO_ERROR; |
| 90 | } |
| 91 | |
| 92 | buf[i] = val; |
| 93 | } |
| 94 | } |
| 95 | else |
| 96 | { |
| 97 | // Now we are at the beginning of the data |
| 98 | // Parse based on the datatype and number of dimensions |
| 99 | if (fread(databuf, elementsize, elems, infile) != elems) |
| 100 | { |
| 101 | rc = FILE_IO_ERROR; |
| 102 | } |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | if (infile) |
| 107 | fclose(infile); |
| 108 | |
| 109 | return rc; |
| 110 | } |
| 111 | |
| 112 | NumpyUtilities::NPError NumpyUtilities::checkNpyHeader(FILE* infile, const uint32_t elems, const char* dtype_str) |
| 113 | { |
| 114 | char buf[NUMPY_HEADER_SZ + 1]; |
| 115 | char* ptr = nullptr; |
| 116 | NPError rc = NO_ERROR; |
| 117 | bool foundFormat = false; |
| 118 | bool foundOrder = false; |
| 119 | bool foundShape = false; |
| 120 | bool fortranOrder = false; |
| 121 | std::vector<int> shape; |
| 122 | uint32_t totalElems = 1; |
| 123 | char* outer_end = NULL; |
| 124 | |
| 125 | assert(infile); |
| 126 | assert(elems > 0); |
| 127 | |
| 128 | if (fread(buf, NUMPY_HEADER_SZ, 1, infile) != 1) |
| 129 | { |
| 130 | return HEADER_PARSE_ERROR; |
| 131 | } |
| 132 | |
| 133 | if (memcmp(buf, NUMPY_HEADER_STR, sizeof(NUMPY_HEADER_STR) - 1)) |
| 134 | { |
| 135 | return HEADER_PARSE_ERROR; |
| 136 | } |
| 137 | |
| 138 | ptr = strtok_r(buf + sizeof(NUMPY_HEADER_STR) - 1, ":", &outer_end); |
| 139 | |
| 140 | // Read in the data type, order, and shape |
| 141 | while (ptr && (!foundFormat || !foundOrder || !foundShape)) |
| 142 | { |
| 143 | |
| 144 | // End of string? |
| 145 | if (!ptr) |
| 146 | break; |
| 147 | |
| 148 | // Skip whitespace |
| 149 | while (isspace(*ptr)) |
| 150 | ptr++; |
| 151 | |
| 152 | // Parse the dictionary field name |
| 153 | if (!strcmp(ptr, "'descr'")) |
| 154 | { |
| 155 | ptr = strtok_r(NULL, ",", &outer_end); |
| 156 | if (!ptr) |
| 157 | break; |
| 158 | |
| 159 | while (isspace(*ptr)) |
| 160 | ptr++; |
| 161 | |
| 162 | if (strcmp(ptr, dtype_str)) |
| 163 | { |
| 164 | return FILE_TYPE_MISMATCH; |
| 165 | } |
| 166 | |
| 167 | foundFormat = true; |
| 168 | } |
| 169 | else if (!strcmp(ptr, "'fortran_order'")) |
| 170 | { |
| 171 | ptr = strtok_r(NULL, ",", &outer_end); |
| 172 | if (!ptr) |
| 173 | break; |
| 174 | |
| 175 | while (isspace(*ptr)) |
| 176 | ptr++; |
| 177 | |
| 178 | if (!strcmp(ptr, "False")) |
| 179 | { |
| 180 | fortranOrder = false; |
| 181 | } |
| 182 | else |
| 183 | { |
| 184 | return FILE_TYPE_MISMATCH; |
| 185 | } |
| 186 | |
| 187 | foundOrder = true; |
| 188 | } |
| 189 | else if (!strcmp(ptr, "'shape'")) |
| 190 | { |
| 191 | |
| 192 | ptr = strtok_r(NULL, "(", &outer_end); |
| 193 | if (!ptr) |
| 194 | break; |
| 195 | ptr = strtok_r(NULL, ")", &outer_end); |
| 196 | if (!ptr) |
| 197 | break; |
| 198 | |
| 199 | while (isspace(*ptr)) |
| 200 | ptr++; |
| 201 | |
Jeremy Johnson | 82dbb32 | 2021-07-08 11:53:04 +0100 | [diff] [blame] | 202 | // The shape contains N comma-separated integers. Read up to MAX_DIMS. |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 203 | char* end = NULL; |
| 204 | |
| 205 | ptr = strtok_r(ptr, ",", &end); |
Jeremy Johnson | 82dbb32 | 2021-07-08 11:53:04 +0100 | [diff] [blame] | 206 | for (int i = 0; i < NUMPY_MAX_DIMS_SUPPORTED; i++) |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 207 | { |
| 208 | // Out of dimensions |
| 209 | if (!ptr) |
| 210 | break; |
| 211 | |
| 212 | int dim = atoi(ptr); |
| 213 | |
| 214 | // Dimension is 0 |
| 215 | if (dim == 0) |
| 216 | break; |
| 217 | |
| 218 | shape.push_back(dim); |
| 219 | totalElems *= dim; |
| 220 | ptr = strtok_r(NULL, ",", &end); |
| 221 | } |
| 222 | |
| 223 | foundShape = true; |
| 224 | } |
| 225 | else |
| 226 | { |
| 227 | return HEADER_PARSE_ERROR; |
| 228 | } |
| 229 | |
| 230 | if (!ptr) |
| 231 | break; |
| 232 | |
| 233 | ptr = strtok_r(NULL, ":", &outer_end); |
| 234 | } |
| 235 | |
| 236 | if (!foundShape || !foundFormat || !foundOrder) |
| 237 | { |
| 238 | return HEADER_PARSE_ERROR; |
| 239 | } |
| 240 | |
| 241 | // Validate header |
| 242 | if (fortranOrder) |
| 243 | { |
| 244 | return FILE_TYPE_MISMATCH; |
| 245 | } |
| 246 | |
| 247 | if (totalElems != elems) |
| 248 | { |
| 249 | return BUFFER_SIZE_MISMATCH; |
| 250 | } |
| 251 | |
| 252 | // Go back to the begininng and read until the end of the header dictionary |
| 253 | rewind(infile); |
| 254 | int val; |
| 255 | |
| 256 | do |
| 257 | { |
| 258 | val = fgetc(infile); |
| 259 | } while (val != EOF && val != '\n'); |
| 260 | |
| 261 | return rc; |
| 262 | } |
| 263 | |
| 264 | NumpyUtilities::NPError NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const bool* databuf) |
| 265 | { |
| 266 | std::vector<int32_t> shape = { (int32_t)elems }; |
| 267 | return writeToNpyFile(filename, shape, databuf); |
| 268 | } |
| 269 | |
| 270 | NumpyUtilities::NPError |
| 271 | NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const bool* databuf) |
| 272 | { |
| 273 | const char dtype_str[] = "'|b1'"; |
| 274 | return writeToNpyFileCommon(filename, dtype_str, 1, shape, databuf, true); // bools written as size 1 |
| 275 | } |
| 276 | |
| 277 | NumpyUtilities::NPError |
| 278 | NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const int32_t* databuf) |
| 279 | { |
| 280 | std::vector<int32_t> shape = { (int32_t)elems }; |
| 281 | return writeToNpyFile(filename, shape, databuf); |
| 282 | } |
| 283 | |
| 284 | NumpyUtilities::NPError |
| 285 | NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const int32_t* databuf) |
| 286 | { |
| 287 | const char dtype_str[] = "'<i4'"; |
| 288 | return writeToNpyFileCommon(filename, dtype_str, sizeof(int32_t), shape, databuf, false); |
| 289 | } |
| 290 | |
| 291 | NumpyUtilities::NPError |
| 292 | NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const int64_t* databuf) |
| 293 | { |
| 294 | std::vector<int32_t> shape = { (int32_t)elems }; |
| 295 | return writeToNpyFile(filename, shape, databuf); |
| 296 | } |
| 297 | |
| 298 | NumpyUtilities::NPError |
| 299 | NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const int64_t* databuf) |
| 300 | { |
| 301 | const char dtype_str[] = "'<i8'"; |
| 302 | return writeToNpyFileCommon(filename, dtype_str, sizeof(int64_t), shape, databuf, false); |
| 303 | } |
| 304 | |
| 305 | NumpyUtilities::NPError NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const float* databuf) |
| 306 | { |
| 307 | std::vector<int32_t> shape = { (int32_t)elems }; |
| 308 | return writeToNpyFile(filename, shape, databuf); |
| 309 | } |
| 310 | |
| 311 | NumpyUtilities::NPError |
| 312 | NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const float* databuf) |
| 313 | { |
| 314 | const char dtype_str[] = "'<f4'"; |
| 315 | return writeToNpyFileCommon(filename, dtype_str, sizeof(float), shape, databuf, false); |
| 316 | } |
| 317 | |
James Ward | 485a11d | 2022-08-05 13:48:37 +0100 | [diff] [blame^] | 318 | NumpyUtilities::NPError NumpyUtilities::writeToNpyFile(const char* filename, |
| 319 | const std::vector<int32_t>& shape, |
| 320 | const half_float::half* databuf) |
| 321 | { |
| 322 | const char dtype_str[] = "'<f2'"; |
| 323 | return writeToNpyFileCommon(filename, dtype_str, sizeof(half_float::half), shape, databuf, false); |
| 324 | } |
| 325 | |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 326 | NumpyUtilities::NPError NumpyUtilities::writeToNpyFileCommon(const char* filename, |
| 327 | const char* dtype_str, |
| 328 | const size_t elementsize, |
| 329 | const std::vector<int32_t>& shape, |
| 330 | const void* databuf, |
| 331 | bool bool_translate) |
| 332 | { |
| 333 | FILE* outfile = nullptr; |
| 334 | NPError rc = NO_ERROR; |
| 335 | uint32_t totalElems = 1; |
| 336 | |
| 337 | assert(filename); |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 338 | assert(databuf); |
| 339 | |
| 340 | outfile = fopen(filename, "wb"); |
| 341 | |
| 342 | if (!outfile) |
| 343 | { |
| 344 | return FILE_NOT_FOUND; |
| 345 | } |
| 346 | |
| 347 | for (uint32_t i = 0; i < shape.size(); i++) |
| 348 | { |
| 349 | totalElems *= shape[i]; |
| 350 | } |
| 351 | |
| 352 | rc = writeNpyHeader(outfile, shape, dtype_str); |
| 353 | |
| 354 | if (rc == NO_ERROR) |
| 355 | { |
| 356 | if (bool_translate) |
| 357 | { |
| 358 | // Numpy save format stores booleans as a byte array |
| 359 | // with one byte per boolean. This somewhat inefficiently |
| 360 | // remaps from system bool[] to this format. |
| 361 | const bool* buf = reinterpret_cast<const bool*>(databuf); |
| 362 | for (uint32_t i = 0; i < totalElems; i++) |
| 363 | { |
| 364 | int val = buf[i] ? 1 : 0; |
| 365 | if (fputc(val, outfile) == EOF) |
| 366 | { |
| 367 | rc = FILE_IO_ERROR; |
| 368 | } |
| 369 | } |
| 370 | } |
| 371 | else |
| 372 | { |
| 373 | if (fwrite(databuf, elementsize, totalElems, outfile) != totalElems) |
| 374 | { |
| 375 | rc = FILE_IO_ERROR; |
| 376 | } |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | if (outfile) |
| 381 | fclose(outfile); |
| 382 | |
| 383 | return rc; |
| 384 | } |
| 385 | |
| 386 | NumpyUtilities::NPError |
| 387 | NumpyUtilities::writeNpyHeader(FILE* outfile, const std::vector<int32_t>& shape, const char* dtype_str) |
| 388 | { |
| 389 | NPError rc = NO_ERROR; |
| 390 | uint32_t i; |
| 391 | char header[NUMPY_HEADER_SZ + 1]; |
| 392 | int headerPos = 0; |
| 393 | |
| 394 | assert(outfile); |
Eric Kunze | 2364dcd | 2021-04-26 11:06:57 -0700 | [diff] [blame] | 395 | |
| 396 | // Space-fill the header and end with a newline to start per numpy spec |
| 397 | memset(header, 0x20, NUMPY_HEADER_SZ); |
| 398 | header[NUMPY_HEADER_SZ - 1] = '\n'; |
| 399 | header[NUMPY_HEADER_SZ] = 0; |
| 400 | |
| 401 | // Write out the hard-coded header. We only support a 128-byte 1.0 header |
| 402 | // for now, which should be sufficient for simple tensor types of any |
| 403 | // reasonable rank. |
| 404 | memcpy(header, NUMPY_HEADER_STR, sizeof(NUMPY_HEADER_STR) - 1); |
| 405 | headerPos += sizeof(NUMPY_HEADER_STR) - 1; |
| 406 | |
| 407 | // Output the format dictionary |
| 408 | // Hard-coded for I32 for now |
| 409 | headerPos += |
| 410 | snprintf(header + headerPos, NUMPY_HEADER_SZ - headerPos, "'descr': %s, 'fortran_order': False, 'shape': (%d,", |
| 411 | dtype_str, shape.empty() ? 1 : shape[0]); |
| 412 | |
| 413 | // Remainder of shape array |
| 414 | for (i = 1; i < shape.size(); i++) |
| 415 | { |
| 416 | headerPos += snprintf(header + headerPos, NUMPY_HEADER_SZ - headerPos, " %d,", shape[i]); |
| 417 | } |
| 418 | |
| 419 | // Close off the dictionary |
| 420 | headerPos += snprintf(header + headerPos, NUMPY_HEADER_SZ - headerPos, "), }"); |
| 421 | |
| 422 | // snprintf leaves a NULL at the end. Replace with a space |
| 423 | header[headerPos] = 0x20; |
| 424 | |
| 425 | if (fwrite(header, NUMPY_HEADER_SZ, 1, outfile) != 1) |
| 426 | { |
| 427 | rc = FILE_IO_ERROR; |
| 428 | } |
| 429 | |
| 430 | return rc; |
| 431 | } |