| 1 | /********************************************************************* |
|---|
| 2 | Blosc - Blocked Suffling and Compression Library |
|---|
| 3 | |
|---|
| 4 | Author: Francesc Alted (faltet@pytables.org) |
|---|
| 5 | Creation date: 2009-05-20 |
|---|
| 6 | |
|---|
| 7 | See LICENSES/BLOSC.txt for details about copyright and rights to use. |
|---|
| 8 | **********************************************************************/ |
|---|
| 9 | |
|---|
| 10 | |
|---|
| 11 | #include <stdlib.h> |
|---|
| 12 | #include <stdio.h> |
|---|
| 13 | #include <string.h> |
|---|
| 14 | #include <sys/types.h> |
|---|
| 15 | #include <sys/stat.h> |
|---|
| 16 | #include "blosc.h" |
|---|
| 17 | #include "blosclz.h" |
|---|
| 18 | #include "shuffle.h" |
|---|
| 19 | #ifdef _WIN32 |
|---|
| 20 | #include <windows.h> |
|---|
| 21 | #else |
|---|
| 22 | #include <stdint.h> |
|---|
| 23 | #include <unistd.h> |
|---|
| 24 | #endif /* _WIN32 */ |
|---|
| 25 | |
|---|
| 26 | |
|---|
| 27 | /* Starting point for the blocksize computation */ |
|---|
| 28 | #define BLOCKSIZE (4*1024) /* 4 KB (page size) */ |
|---|
| 29 | |
|---|
| 30 | /* Maximum typesize before considering buffer as a stream of bytes. */ |
|---|
| 31 | #define MAXTYPESIZE 256 /* Cannot be larger than 256 */ |
|---|
| 32 | |
|---|
| 33 | /* The maximum number of splits in a block for compression */ |
|---|
| 34 | #define MAXSPLITS 16 /* Cannot be larger than 128 */ |
|---|
| 35 | |
|---|
| 36 | |
|---|
| 37 | /* Global variables for compressing/shuffling actions */ |
|---|
| 38 | int clevel; /* Compression level */ |
|---|
| 39 | int do_shuffle; /* Shuffle is active? */ |
|---|
| 40 | |
|---|
| 41 | |
|---|
| 42 | |
|---|
| 43 | /* Shuffle & Compress a single block */ |
|---|
| 44 | static size_t |
|---|
| 45 | _blosc_c(int clevel, int doshuffle, size_t typesize, size_t blocksize, |
|---|
| 46 | int leftoverblock, unsigned int ctbytes, unsigned int nbytes, |
|---|
| 47 | unsigned char* _src, unsigned char* _dest, unsigned char *tmp) { |
|---|
| 48 | size_t j, neblock, nsplits; |
|---|
| 49 | int cbytes, maxout; |
|---|
| 50 | unsigned int btbytes = 0; |
|---|
| 51 | unsigned char* _tmp; |
|---|
| 52 | |
|---|
| 53 | if (doshuffle && (typesize > 1)) { |
|---|
| 54 | /* Shuffle this block (this makes sense only if typesize > 1) */ |
|---|
| 55 | shuffle(typesize, blocksize, _src, tmp); |
|---|
| 56 | _tmp = tmp; |
|---|
| 57 | } |
|---|
| 58 | else { |
|---|
| 59 | _tmp = _src; |
|---|
| 60 | } |
|---|
| 61 | |
|---|
| 62 | /* Compress for each shuffled slice split for this block. */ |
|---|
| 63 | /* If the number of bytes is too large, or we are in a leftover |
|---|
| 64 | block, do not split all. */ |
|---|
| 65 | if ((typesize <= MAXSPLITS) && (!leftoverblock)) { |
|---|
| 66 | nsplits = typesize; |
|---|
| 67 | } |
|---|
| 68 | else { |
|---|
| 69 | nsplits = 1; |
|---|
| 70 | } |
|---|
| 71 | neblock = blocksize / nsplits; |
|---|
| 72 | for (j = 0; j < nsplits; j++) { |
|---|
| 73 | _dest += sizeof(int); |
|---|
| 74 | btbytes += sizeof(int); |
|---|
| 75 | ctbytes += sizeof(int); |
|---|
| 76 | maxout = neblock; |
|---|
| 77 | if (ctbytes+maxout > nbytes) { |
|---|
| 78 | maxout = nbytes - ctbytes; /* avoid buffer overrun */ |
|---|
| 79 | if (maxout <= 0) { |
|---|
| 80 | return 0; /* non-compressible block */ |
|---|
| 81 | } |
|---|
| 82 | } |
|---|
| 83 | cbytes = blosclz_compress(clevel, _tmp+j*neblock, neblock, |
|---|
| 84 | _dest, maxout); |
|---|
| 85 | if (cbytes > maxout) { |
|---|
| 86 | /* Buffer overrun caused by blosclz_compress (should never happen) */ |
|---|
| 87 | return -1; |
|---|
| 88 | } |
|---|
| 89 | else if (cbytes < 0) { |
|---|
| 90 | /* cbytes should never be negative */ |
|---|
| 91 | return -2; |
|---|
| 92 | } |
|---|
| 93 | else if ((cbytes == 0) || (cbytes == (int) neblock)) { |
|---|
| 94 | /* The compressor has been unable to compress data |
|---|
| 95 | significantly. Also, it may happen that the compressed |
|---|
| 96 | buffer has exactly the same length than the buffer size, but |
|---|
| 97 | this means uncompressible data. Before doing the copy, check |
|---|
| 98 | that we are not running into a buffer overflow. */ |
|---|
| 99 | if ((ctbytes+neblock) > nbytes) { |
|---|
| 100 | return 0; /* Non-compressible data */ |
|---|
| 101 | } |
|---|
| 102 | memcpy(_dest, _tmp+j*neblock, neblock); |
|---|
| 103 | cbytes = neblock; |
|---|
| 104 | } |
|---|
| 105 | ((unsigned int *)(_dest))[-1] = cbytes; |
|---|
| 106 | _dest += cbytes; |
|---|
| 107 | btbytes += cbytes; |
|---|
| 108 | ctbytes += cbytes; |
|---|
| 109 | } /* Closes j < nsplits */ |
|---|
| 110 | |
|---|
| 111 | return btbytes; |
|---|
| 112 | } |
|---|
| 113 | |
|---|
| 114 | |
|---|
| 115 | unsigned int |
|---|
| 116 | blosc_compress(int clevel, int doshuffle, size_t typesize, size_t nbytes, |
|---|
| 117 | const void *src, void *dest) |
|---|
| 118 | { |
|---|
| 119 | unsigned char *_src=NULL; /* alias for source buffer */ |
|---|
| 120 | unsigned char *_dest=NULL; /* alias for destination buffer */ |
|---|
| 121 | unsigned char *flags; /* flags for header */ |
|---|
| 122 | unsigned int *starts; /* start pointers for each block */ |
|---|
| 123 | size_t nblocks; /* number of complete blocks in buffer */ |
|---|
| 124 | size_t tblocks; /* number of total blocks in buffer */ |
|---|
| 125 | size_t leftover; /* extra bytes at end of buffer */ |
|---|
| 126 | size_t blocksize; /* length of the block in bytes */ |
|---|
| 127 | size_t bsize; /* corrected blocksize for last block */ |
|---|
| 128 | unsigned int ctbytes = 0; /* the number of bytes in output buffer */ |
|---|
| 129 | unsigned int *ctbytes_; /* the number of bytes in output buffer */ |
|---|
| 130 | unsigned char *tmp; /* temporary buffer for data block */ |
|---|
| 131 | int cbytes; /* temporary compressed buffer length */ |
|---|
| 132 | int leftoverblock; /* left over block? */ |
|---|
| 133 | unsigned int i, j; /* local index variables */ |
|---|
| 134 | const char *too_long_message = "The impossible happened: buffer overflow!\n"; |
|---|
| 135 | |
|---|
| 136 | /* Compression level */ |
|---|
| 137 | if (clevel < 0 || clevel > 9) { |
|---|
| 138 | /* If clevel not in 0..9, print an error */ |
|---|
| 139 | fprintf(stderr, "`clevel` parameter must be between 0 and 9!\n"); |
|---|
| 140 | return -10; |
|---|
| 141 | } |
|---|
| 142 | else if (clevel == 0) { |
|---|
| 143 | /* No compression wanted. Just return without doing anything else. */ |
|---|
| 144 | return 0; |
|---|
| 145 | } |
|---|
| 146 | |
|---|
| 147 | /* Shuffle */ |
|---|
| 148 | if (doshuffle != 0 && doshuffle != 1) { |
|---|
| 149 | /* If shuffle not in 0,1, print an error */ |
|---|
| 150 | fprintf(stderr, "`doshuffle` parameter must be either 0 or 1!\n"); |
|---|
| 151 | return -10; |
|---|
| 152 | } |
|---|
| 153 | |
|---|
| 154 | /* Compute a blocksize depending on the optimization level */ |
|---|
| 155 | blocksize = BLOCKSIZE; |
|---|
| 156 | /* 3 first optimization levels will not change blocksize */ |
|---|
| 157 | for (i=4; i<=(unsigned int)clevel; i++) { |
|---|
| 158 | /* Escape if blocksize grows more than nbytes */ |
|---|
| 159 | if (blocksize*2 > nbytes) break; |
|---|
| 160 | blocksize *= 2; |
|---|
| 161 | } |
|---|
| 162 | |
|---|
| 163 | /* blocksize must be a multiple of the typesize */ |
|---|
| 164 | blocksize = blocksize / typesize * typesize; |
|---|
| 165 | |
|---|
| 166 | /* Create temporary area */ |
|---|
| 167 | #ifdef _WIN32 |
|---|
| 168 | tmp = (unsigned char *)_aligned_malloc(blocksize, 16); |
|---|
| 169 | #elif defined __APPLE__ |
|---|
| 170 | /* Mac OS X guarantees 16-byte alignment in small allocs */ |
|---|
| 171 | tmp = (unsigned char *)(malloc(blocksize)); |
|---|
| 172 | #else |
|---|
| 173 | posix_memalign((void **)&tmp, 16, blocksize); |
|---|
| 174 | #endif /* _WIN32 */ |
|---|
| 175 | |
|---|
| 176 | nblocks = nbytes / blocksize; |
|---|
| 177 | leftover = nbytes % blocksize; |
|---|
| 178 | _src = (unsigned char *)(src); |
|---|
| 179 | _dest = (unsigned char *)(dest); |
|---|
| 180 | tblocks = (leftover>0)? nblocks+1: nblocks; |
|---|
| 181 | |
|---|
| 182 | /* Check typesize limits */ |
|---|
| 183 | if (typesize == MAXTYPESIZE) { |
|---|
| 184 | typesize = 0; /* zero means MAXTYPESIZE */ |
|---|
| 185 | } |
|---|
| 186 | else if (typesize > MAXTYPESIZE) { |
|---|
| 187 | /* If typesize is too large, treat buffer as an 1-byte stream. */ |
|---|
| 188 | typesize = 1; |
|---|
| 189 | } |
|---|
| 190 | |
|---|
| 191 | /* Write header for this block */ |
|---|
| 192 | _dest[0] = BLOSC_VERSION_FORMAT; /* blosc format version */ |
|---|
| 193 | _dest[1] = BLOSCLZ_VERSION_FORMAT; /* blosclz format version */ |
|---|
| 194 | flags = _dest+2; /* flags */ |
|---|
| 195 | _dest[2] = 0; /* zeroes flags */ |
|---|
| 196 | _dest[3] = (unsigned char)typesize; /* type size */ |
|---|
| 197 | _dest += 4; |
|---|
| 198 | ctbytes += 4; |
|---|
| 199 | ((unsigned int *)_dest)[0] = nbytes; /* size of the chunk */ |
|---|
| 200 | ((unsigned int *)_dest)[1] = blocksize; /* block size */ |
|---|
| 201 | ctbytes_ = (unsigned int *)(_dest+8); /* compressed chunk size (pointer) */ |
|---|
| 202 | _dest += sizeof(int)*3; |
|---|
| 203 | ctbytes += sizeof(int)*3; |
|---|
| 204 | starts = (unsigned int *)_dest; /* starts for every block */ |
|---|
| 205 | _dest += sizeof(int)*tblocks; /* book space for pointers to */ |
|---|
| 206 | ctbytes += sizeof(int)*tblocks; /* every block in output */ |
|---|
| 207 | |
|---|
| 208 | if (doshuffle == 1) { |
|---|
| 209 | /* Shuffle is active */ |
|---|
| 210 | *flags |= 0x1; /* bit 0 set to one in flags */ |
|---|
| 211 | } |
|---|
| 212 | |
|---|
| 213 | for (j = 0; j < tblocks; j++) { |
|---|
| 214 | starts[j] = ctbytes; |
|---|
| 215 | bsize = blocksize; |
|---|
| 216 | leftoverblock = 0; |
|---|
| 217 | if ((j == tblocks - 1) && (leftover > 0)) { |
|---|
| 218 | bsize = leftover; |
|---|
| 219 | leftoverblock = 1; |
|---|
| 220 | } |
|---|
| 221 | cbytes = _blosc_c(clevel, doshuffle, typesize, bsize, leftoverblock, |
|---|
| 222 | ctbytes, nbytes, _src, _dest, tmp); |
|---|
| 223 | if (cbytes < 0) { |
|---|
| 224 | fprintf(stderr, too_long_message); |
|---|
| 225 | ctbytes = cbytes; /* error in _blosc_c */ |
|---|
| 226 | goto out; |
|---|
| 227 | } |
|---|
| 228 | if (cbytes == 0) { |
|---|
| 229 | ctbytes = 0; /* uncompressible data */ |
|---|
| 230 | goto out; |
|---|
| 231 | } |
|---|
| 232 | _dest += cbytes; |
|---|
| 233 | _src += blocksize; |
|---|
| 234 | ctbytes += cbytes; |
|---|
| 235 | } /* Close j < tblocks */ |
|---|
| 236 | |
|---|
| 237 | if (ctbytes == nbytes) { |
|---|
| 238 | ctbytes = 0; /* non-compressible data */ |
|---|
| 239 | goto out; |
|---|
| 240 | } |
|---|
| 241 | else if (ctbytes > nbytes) { |
|---|
| 242 | fprintf(stderr, too_long_message); |
|---|
| 243 | ctbytes = -5; /* too large buffer */ |
|---|
| 244 | goto out; |
|---|
| 245 | } |
|---|
| 246 | |
|---|
| 247 | out: |
|---|
| 248 | #ifdef _WIN32 |
|---|
| 249 | _aligned_free(tmp); |
|---|
| 250 | #else |
|---|
| 251 | free(tmp); |
|---|
| 252 | #endif /* _WIN32 */ |
|---|
| 253 | |
|---|
| 254 | *ctbytes_ = ctbytes; /* set the number of compressed bytes in header */ |
|---|
| 255 | return ctbytes; |
|---|
| 256 | |
|---|
| 257 | } |
|---|
| 258 | |
|---|
| 259 | |
|---|
| 260 | /* Decompress & unshuffle a single block */ |
|---|
| 261 | static size_t |
|---|
| 262 | _blosc_d(int dounshuffle, size_t typesize, size_t blocksize, int leftoverblock, |
|---|
| 263 | unsigned char* _src, unsigned char* _dest, |
|---|
| 264 | unsigned char *tmp, unsigned char *tmp2) |
|---|
| 265 | { |
|---|
| 266 | size_t j, neblock, nsplits; |
|---|
| 267 | size_t nbytes, cbytes, ctbytes = 0, ntbytes = 0; |
|---|
| 268 | unsigned char* _tmp; |
|---|
| 269 | |
|---|
| 270 | if (dounshuffle && (typesize > 1)) { |
|---|
| 271 | _tmp = tmp; |
|---|
| 272 | } |
|---|
| 273 | else { |
|---|
| 274 | _tmp = _dest; |
|---|
| 275 | } |
|---|
| 276 | |
|---|
| 277 | /* Compress for each shuffled slice split for this block. */ |
|---|
| 278 | /* If the number of bytes is too large, do not split all. */ |
|---|
| 279 | if ((typesize <= MAXSPLITS) && (!leftoverblock)) { |
|---|
| 280 | nsplits = typesize; |
|---|
| 281 | } |
|---|
| 282 | else { |
|---|
| 283 | nsplits = 1; |
|---|
| 284 | } |
|---|
| 285 | neblock = blocksize / nsplits; |
|---|
| 286 | for (j = 0; j < nsplits; j++) { |
|---|
| 287 | cbytes = ((unsigned int *)(_src))[0]; /* amount of compressed bytes */ |
|---|
| 288 | _src += sizeof(int); |
|---|
| 289 | ctbytes += sizeof(int); |
|---|
| 290 | /* Uncompress */ |
|---|
| 291 | if (cbytes == neblock) { |
|---|
| 292 | memcpy(_tmp, _src, neblock); |
|---|
| 293 | nbytes = neblock; |
|---|
| 294 | } |
|---|
| 295 | else { |
|---|
| 296 | nbytes = blosclz_decompress(_src, cbytes, _tmp, neblock); |
|---|
| 297 | if (nbytes != neblock) { |
|---|
| 298 | return -2; |
|---|
| 299 | } |
|---|
| 300 | } |
|---|
| 301 | _src += cbytes; |
|---|
| 302 | ctbytes += cbytes; |
|---|
| 303 | _tmp += neblock; |
|---|
| 304 | ntbytes += nbytes; |
|---|
| 305 | } /* Closes j < nsplits */ |
|---|
| 306 | |
|---|
| 307 | if (dounshuffle && (typesize > 1)) { |
|---|
| 308 | if ((uintptr_t)_dest % 16 == 0) { |
|---|
| 309 | /* 16-bytes aligned _dest. SSE2 unshuffle will work. */ |
|---|
| 310 | unshuffle(typesize, blocksize, tmp, _dest); |
|---|
| 311 | } |
|---|
| 312 | else { |
|---|
| 313 | /* _dest is not aligned. Use tmp2, which is aligned, and copy. */ |
|---|
| 314 | unshuffle(typesize, blocksize, tmp, tmp2); |
|---|
| 315 | memcpy(_dest, tmp2, blocksize); |
|---|
| 316 | } |
|---|
| 317 | } |
|---|
| 318 | |
|---|
| 319 | return ctbytes; |
|---|
| 320 | } |
|---|
| 321 | |
|---|
| 322 | |
|---|
| 323 | unsigned int |
|---|
| 324 | blosc_decompress(const void *src, void *dest, size_t dest_size) |
|---|
| 325 | { |
|---|
| 326 | unsigned char *_src=NULL; /* alias for source buffer */ |
|---|
| 327 | unsigned char *_dest=NULL; /* alias for destination buffer */ |
|---|
| 328 | unsigned char version, versionlz; /* versions for compressed header */ |
|---|
| 329 | unsigned char flags; /* flags for header */ |
|---|
| 330 | size_t leftover; /* extra bytes at end of buffer */ |
|---|
| 331 | size_t nblocks; /* number of complete blocks in buffer */ |
|---|
| 332 | size_t tblocks; /* number of total blocks in buffer */ |
|---|
| 333 | size_t j; |
|---|
| 334 | size_t nbytes, ntbytes = 0; |
|---|
| 335 | int cbytes; |
|---|
| 336 | unsigned char *tmp, *tmp2; |
|---|
| 337 | int dounshuffle = 0; |
|---|
| 338 | unsigned int typesize, blocksize, bsize, ctbytes_; |
|---|
| 339 | int leftoverblock; /* left over block? */ |
|---|
| 340 | |
|---|
| 341 | _src = (unsigned char *)(src); |
|---|
| 342 | _dest = (unsigned char *)(dest); |
|---|
| 343 | |
|---|
| 344 | /* Read the header block */ |
|---|
| 345 | version = _src[0]; /* blosc format version */ |
|---|
| 346 | versionlz = _src[1]; /* blosclz format version */ |
|---|
| 347 | flags = _src[2]; /* flags */ |
|---|
| 348 | typesize = (unsigned int)_src[3]; /* typesize */ |
|---|
| 349 | _src += 4; |
|---|
| 350 | nbytes = ((unsigned int *)_src)[0]; /* chunk size */ |
|---|
| 351 | blocksize = ((unsigned int *)_src)[1]; /* block size */ |
|---|
| 352 | ctbytes_ = ((unsigned int *)_src)[2]; /* compressed chunk size */ |
|---|
| 353 | _src += sizeof(int)*3; |
|---|
| 354 | /* Compute some params */ |
|---|
| 355 | nblocks = nbytes / blocksize; |
|---|
| 356 | leftover = nbytes % blocksize; |
|---|
| 357 | tblocks = (leftover>0)? nblocks+1: nblocks; |
|---|
| 358 | _src += sizeof(int)*tblocks; /* skip starts of blocks */ |
|---|
| 359 | |
|---|
| 360 | /* Check zero typesizes */ |
|---|
| 361 | if (typesize == 0) { |
|---|
| 362 | typesize = MAXTYPESIZE; |
|---|
| 363 | } |
|---|
| 364 | |
|---|
| 365 | if (nbytes > dest_size) { |
|---|
| 366 | /* This should never happen, but just in case. */ |
|---|
| 367 | return -1; |
|---|
| 368 | } |
|---|
| 369 | |
|---|
| 370 | if ((flags & 0x1) == 1) { |
|---|
| 371 | /* Input is shuffled. Unshuffle it. */ |
|---|
| 372 | dounshuffle = 1; |
|---|
| 373 | } |
|---|
| 374 | |
|---|
| 375 | /* Create temporary area */ |
|---|
| 376 | #ifdef _WIN32 |
|---|
| 377 | tmp = (unsigned char*)_aligned_malloc(blocksize, 16); |
|---|
| 378 | tmp2 = (unsigned char*)_aligned_malloc(blocksize, 16); |
|---|
| 379 | #elif defined __APPLE__ |
|---|
| 380 | /* Mac OS X guarantees 16-byte alignment in small allocs */ |
|---|
| 381 | tmp = (unsigned char *)(malloc(blocksize)); |
|---|
| 382 | tmp2 = (unsigned char *)(malloc(blocksize)); |
|---|
| 383 | #else |
|---|
| 384 | posix_memalign((void **)&tmp, 16, blocksize); |
|---|
| 385 | posix_memalign((void **)&tmp2, 16, blocksize); |
|---|
| 386 | #endif /* _WIN32 */ |
|---|
| 387 | |
|---|
| 388 | for (j = 0; j < tblocks; j++) { |
|---|
| 389 | bsize = blocksize; |
|---|
| 390 | leftoverblock = 0; |
|---|
| 391 | if ((j == tblocks - 1) && (leftover > 0)) { |
|---|
| 392 | bsize = leftover; |
|---|
| 393 | leftoverblock = 1; |
|---|
| 394 | } |
|---|
| 395 | cbytes = _blosc_d(dounshuffle, typesize, bsize, leftoverblock, |
|---|
| 396 | _src, _dest, tmp, tmp2); |
|---|
| 397 | if (cbytes < 0) { |
|---|
| 398 | nbytes = cbytes; /* _blosc_d failure */ |
|---|
| 399 | goto out; |
|---|
| 400 | } |
|---|
| 401 | _src += cbytes; |
|---|
| 402 | _dest += blocksize; |
|---|
| 403 | ntbytes += blocksize; |
|---|
| 404 | } |
|---|
| 405 | |
|---|
| 406 | out: |
|---|
| 407 | #ifdef _WIN32 |
|---|
| 408 | _aligned_free(tmp); |
|---|
| 409 | _aligned_free(tmp2); |
|---|
| 410 | #else |
|---|
| 411 | free(tmp); |
|---|
| 412 | free(tmp2); |
|---|
| 413 | #endif /* _WIN32 */ |
|---|
| 414 | return nbytes; |
|---|
| 415 | } |
|---|
| 416 | |
|---|
| 417 | |
|---|