contrib/zstd/programs/bench.c

   1 /**
   2  * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
   3  * All rights reserved.
   4  *
   5  * This source code is licensed under the BSD-style license found in the
   6  * LICENSE file in the root directory of this source tree. An additional grant
   7  * of patent rights can be found in the PATENTS file in the same directory.
   8  */
   9
  10
  11
  12 /* **************************************
  13 *  Tuning parameters
  14 ****************************************/
  15 #ifndef BMK_TIMETEST_DEFAULT_S   /* default minimum time per test */
  16 #define BMK_TIMETEST_DEFAULT_S 3
  17 #endif
  18
  19
  20 /* **************************************
  21 *  Compiler Warnings
  22 ****************************************/
  23 #ifdef _MSC_VER
  24 #  pragma warning(disable : 4127)                /* disable: C4127: conditional expression is constant */
  25 #endif
  26
  27
  28 /* *************************************
  29 *  Includes
  30 ***************************************/
  31 #include "platform.h"    /* Large Files support */
  32 #include "util.h"        /* UTIL_getFileSize, UTIL_sleep */
  33 #include <stdlib.h>      /* malloc, free */
  34 #include <string.h>      /* memset */
  35 #include <stdio.h>       /* fprintf, fopen */
  36 #include <time.h>        /* clock_t, clock, CLOCKS_PER_SEC */
  37
  38 #include "mem.h"
  39 #define ZSTD_STATIC_LINKING_ONLY
  40 #include "zstd.h"
  41 #include "datagen.h"     /* RDG_genBuffer */
  42 #include "xxhash.h"
  43 #include "zstdmt_compress.h"
  44
  45
  46 /* *************************************
  47 *  Constants
  48 ***************************************/
  49 #ifndef ZSTD_GIT_COMMIT
  50 #  define ZSTD_GIT_COMMIT_STRING ""
  51 #else
  52 #  define ZSTD_GIT_COMMIT_STRING ZSTD_EXPAND_AND_QUOTE(ZSTD_GIT_COMMIT)
  53 #endif
  54
  55 #define TIMELOOP_MICROSEC     1*1000000ULL /* 1 second */
  56 #define ACTIVEPERIOD_MICROSEC 70*1000000ULL /* 70 seconds */
  57 #define COOLPERIOD_SEC        10
  58
  59 #define KB *(1 <<10)
  60 #define MB *(1 <<20)
  61 #define GB *(1U<<30)
  62
  63 static const size_t maxMemory = (sizeof(size_t)==4)  ?  (2 GB - 64 MB) : (size_t)(1ULL << ((sizeof(size_t)*8)-31));
  64
  65 static U32 g_compressibilityDefault = 50;
  66
  67
  68 /* *************************************
  69 *  console display
  70 ***************************************/
  71 #define DISPLAY(...)         fprintf(stderr, __VA_ARGS__)
  72 #define DISPLAYLEVEL(l, ...) if (g_displayLevel>=l) { DISPLAY(__VA_ARGS__); }
  73 static U32 g_displayLevel = 2;   /* 0 : no display;   1: errors;   2 : + result + interaction + warnings;   3 : + progression;   4 : + information */
  74
  75 #define DISPLAYUPDATE(l, ...) if (g_displayLevel>=l) { \
  76             if ((clock() - g_time > refreshRate) || (g_displayLevel>=4)) \
  77             { g_time = clock(); DISPLAY(__VA_ARGS__); \
  78             if (g_displayLevel>=4) fflush(stdout); } }
  79 static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
  80 static clock_t g_time = 0;
  81
  82
  83 /* *************************************
  84 *  Exceptions
  85 ***************************************/
  86 #ifndef DEBUG
  87 #  define DEBUG 0
  88 #endif
  89 #define DEBUGOUTPUT(...) if (DEBUG) DISPLAY(__VA_ARGS__);
  90 #define EXM_THROW(error, ...)                                             \
  91 {                                                                         \
  92     DEBUGOUTPUT("Error defined at %s, line %i : \n", __FILE__, __LINE__); \
  93     DISPLAYLEVEL(1, "Error %i : ", error);                                \
  94     DISPLAYLEVEL(1, __VA_ARGS__);                                         \
  95     DISPLAYLEVEL(1, " \n");                                               \
  96     exit(error);                                                          \
  97 }
  98
  99
 100 /* *************************************
 101 *  Benchmark Parameters
 102 ***************************************/
 103 static int g_additionalParam = 0;
 104 static U32 g_decodeOnly = 0;
 105
 106 void BMK_setNotificationLevel(unsigned level) { g_displayLevel=level; }
 107
 108 void BMK_setAdditionalParam(int additionalParam) { g_additionalParam=additionalParam; }
 109
 110 static U32 g_nbSeconds = BMK_TIMETEST_DEFAULT_S;
 111 void BMK_setNbSeconds(unsigned nbSeconds)
 112 {
 113     g_nbSeconds = nbSeconds;
 114     DISPLAYLEVEL(3, "- test >= %u seconds per compression / decompression - \n", g_nbSeconds);
 115 }
 116
 117 static size_t g_blockSize = 0;
 118 void BMK_setBlockSize(size_t blockSize)
 119 {
 120     g_blockSize = blockSize;
 121     if (g_blockSize) DISPLAYLEVEL(2, "using blocks of size %u KB \n", (U32)(blockSize>>10));
 122 }
 123
 124 void BMK_setDecodeOnlyMode(unsigned decodeFlag) { g_decodeOnly = (decodeFlag>0); }
 125
 126 static U32 g_nbThreads = 1;
 127 void BMK_setNbThreads(unsigned nbThreads) {
 128 #ifndef ZSTD_MULTITHREAD
 129     if (nbThreads > 1) DISPLAYLEVEL(2, "Note : multi-threading is disabled \n");
 130 #endif
 131     g_nbThreads = nbThreads;
 132 }
 133
 134
 135 /* ********************************************************
 136 *  Bench functions
 137 **********************************************************/
 138 typedef struct {
 139     const void* srcPtr;
 140     size_t srcSize;
 141     void*  cPtr;
 142     size_t cRoom;
 143     size_t cSize;
 144     void*  resPtr;
 145     size_t resSize;
 146 } blockParam_t;
 147
 148
 149 #define MIN(a,b) ((a)<(b) ? (a) : (b))
 150 #define MAX(a,b) ((a)>(b) ? (a) : (b))
 151
 152 static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
 153                         const char* displayName, int cLevel,
 154                         const size_t* fileSizes, U32 nbFiles,
 155                         const void* dictBuffer, size_t dictBufferSize,
 156                         ZSTD_compressionParameters *comprParams)
 157 {
 158     size_t const blockSize = ((g_blockSize>=32 && !g_decodeOnly) ? g_blockSize : srcSize) + (!srcSize) /* avoid div by 0 */ ;
 159     size_t const avgSize = MIN(g_blockSize, (srcSize / nbFiles));
 160     U32 const maxNbBlocks = (U32) ((srcSize + (blockSize-1)) / blockSize) + nbFiles;
 161     blockParam_t* const blockTable = (blockParam_t*) malloc(maxNbBlocks * sizeof(blockParam_t));
 162     size_t const maxCompressedSize = ZSTD_compressBound(srcSize) + (maxNbBlocks * 1024);   /* add some room for safety */
 163     void* const compressedBuffer = malloc(maxCompressedSize);
 164     void* resultBuffer = malloc(srcSize);
 165     ZSTDMT_CCtx* const mtctx = ZSTDMT_createCCtx(g_nbThreads);
 166     ZSTD_CCtx* const ctx = ZSTD_createCCtx();
 167     ZSTD_DCtx* const dctx = ZSTD_createDCtx();
 168     size_t const loadedCompressedSize = srcSize;
 169     size_t cSize = 0;
 170     double ratio = 0.;
 171     U32 nbBlocks;
 172     UTIL_freq_t ticksPerSecond;
 173
 174     /* checks */
 175     if (!compressedBuffer || !resultBuffer || !blockTable || !ctx || !dctx)
 176         EXM_THROW(31, "allocation error : not enough memory");
 177
 178     /* init */
 179     if (strlen(displayName)>17) displayName += strlen(displayName)-17;   /* can only display 17 characters */
 180     UTIL_initTimer(&ticksPerSecond);
 181
 182     if (g_decodeOnly) {
 183         const char* srcPtr = (const char*) srcBuffer;
 184         U64 dSize64 = 0;
 185         U32 fileNb;
 186         for (fileNb=0; fileNb<nbFiles; fileNb++) {
 187             U64 const fSize64 = ZSTD_findDecompressedSize(srcPtr, fileSizes[fileNb]);
 188             if (fSize64==0) EXM_THROW(32, "Impossible to determine original size ");
 189             dSize64 += fSize64;
 190             srcPtr += fileSizes[fileNb];
 191         }
 192         {   size_t const decodedSize = (size_t)dSize64;
 193             if (dSize64 > decodedSize) EXM_THROW(32, "original size is too large");
 194             if (decodedSize==0) EXM_THROW(32, "Impossible to determine original size ");
 195             free(resultBuffer);
 196             resultBuffer = malloc(decodedSize);
 197             if (!resultBuffer) EXM_THROW(33, "not enough memory");
 198             cSize = srcSize;
 199             srcSize = decodedSize;
 200             ratio = (double)srcSize / (double)cSize;
 201     }   }
 202
 203     /* Init blockTable data */
 204     {   const char* srcPtr = (const char*)srcBuffer;
 205         char* cPtr = (char*)compressedBuffer;
 206         char* resPtr = (char*)resultBuffer;
 207         U32 fileNb;
 208         for (nbBlocks=0, fileNb=0; fileNb<nbFiles; fileNb++) {
 209             size_t remaining = fileSizes[fileNb];
 210             U32 const nbBlocksforThisFile = g_decodeOnly ? 1 : (U32)((remaining + (blockSize-1)) / blockSize);
 211             U32 const blockEnd = nbBlocks + nbBlocksforThisFile;
 212             for ( ; nbBlocks<blockEnd; nbBlocks++) {
 213                 size_t const thisBlockSize = MIN(remaining, blockSize);
 214                 blockTable[nbBlocks].srcPtr = (const void*)srcPtr;
 215                 blockTable[nbBlocks].srcSize = thisBlockSize;
 216                 blockTable[nbBlocks].cPtr = (void*)cPtr;
 217                 blockTable[nbBlocks].cRoom = g_decodeOnly ? thisBlockSize : ZSTD_compressBound(thisBlockSize);
 218                 blockTable[nbBlocks].cSize = blockTable[nbBlocks].cRoom;
 219                 blockTable[nbBlocks].resPtr = (void*)resPtr;
 220                 blockTable[nbBlocks].resSize = g_decodeOnly ? (size_t) ZSTD_findDecompressedSize(srcPtr, thisBlockSize) : thisBlockSize;
 221                 srcPtr += thisBlockSize;
 222                 cPtr += blockTable[nbBlocks].cRoom;
 223                 resPtr += thisBlockSize;
 224                 remaining -= thisBlockSize;
 225     }   }   }
 226
 227     /* warmimg up memory */
 228     RDG_genBuffer(compressedBuffer, maxCompressedSize, 0.10, 0.50, 1);
 229
 230     /* Bench */
 231     {   U64 fastestC = (U64)(-1LL), fastestD = (U64)(-1LL);
 232         U64 const crcOrig = g_decodeOnly ? 0 : XXH64(srcBuffer, srcSize, 0);
 233         UTIL_time_t coolTime;
 234         U64 const maxTime = (g_nbSeconds * TIMELOOP_MICROSEC) + 1;
 235         U64 totalCTime=0, totalDTime=0;
 236         U32 cCompleted=g_decodeOnly, dCompleted=0;
 237 #       define NB_MARKS 4
 238         const char* const marks[NB_MARKS] = { " |", " /", " =",  "\\" };
 239         U32 markNb = 0;
 240
 241         UTIL_getTime(&coolTime);
 242         DISPLAYLEVEL(2, "\r%79s\r", "");
 243         while (!cCompleted || !dCompleted) {
 244
 245             /* overheat protection */
 246             if (UTIL_clockSpanMicro(coolTime, ticksPerSecond) > ACTIVEPERIOD_MICROSEC) {
 247                 DISPLAYLEVEL(2, "\rcooling down ...    \r");
 248                 UTIL_sleep(COOLPERIOD_SEC);
 249                 UTIL_getTime(&coolTime);
 250             }
 251
 252             if (!g_decodeOnly) {
 253                 UTIL_time_t clockStart;
 254                 /* Compression */
 255                 DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->\r", marks[markNb], displayName, (U32)srcSize);
 256                 if (!cCompleted) memset(compressedBuffer, 0xE5, maxCompressedSize);  /* warm up and erase result buffer */
 257
 258                 UTIL_sleepMilli(1);  /* give processor time to other processes */
 259                 UTIL_waitForNextTick(ticksPerSecond);
 260                 UTIL_getTime(&clockStart);
 261
 262                 if (!cCompleted) {   /* still some time to do compression tests */
 263                     ZSTD_parameters zparams = ZSTD_getParams(cLevel, avgSize, dictBufferSize);
 264                     ZSTD_customMem const cmem = { NULL, NULL, NULL };
 265                     U64 clockLoop = g_nbSeconds ? TIMELOOP_MICROSEC : 1;
 266                     U32 nbLoops = 0;
 267                     ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dictBuffer, dictBufferSize, 1, zparams, cmem);
 268                     if (cdict==NULL) EXM_THROW(1, "ZSTD_createCDict_advanced() allocation failure");
 269                     if (comprParams->windowLog) zparams.cParams.windowLog = comprParams->windowLog;
 270                     if (comprParams->chainLog) zparams.cParams.chainLog = comprParams->chainLog;
 271                     if (comprParams->hashLog) zparams.cParams.hashLog = comprParams->hashLog;
 272                     if (comprParams->searchLog) zparams.cParams.searchLog = comprParams->searchLog;
 273                     if (comprParams->searchLength) zparams.cParams.searchLength = comprParams->searchLength;
 274                     if (comprParams->targetLength) zparams.cParams.targetLength = comprParams->targetLength;
 275                     if (comprParams->strategy) zparams.cParams.strategy = (ZSTD_strategy)(comprParams->strategy - 1);
 276                     do {
 277                         U32 blockNb;
 278                         size_t rSize;
 279                         for (blockNb=0; blockNb<nbBlocks; blockNb++) {
 280                             if (dictBufferSize) {
 281                                 rSize = ZSTD_compress_usingCDict(ctx,
 282                                                 blockTable[blockNb].cPtr,  blockTable[blockNb].cRoom,
 283                                                 blockTable[blockNb].srcPtr,blockTable[blockNb].srcSize,
 284                                                 cdict);
 285                             } else {
 286 #ifdef ZSTD_MULTITHREAD         /* note : limitation : MT single-pass does not support compression with dictionary */
 287                                 rSize = ZSTDMT_compressCCtx(mtctx,
 288                                                 blockTable[blockNb].cPtr,  blockTable[blockNb].cRoom,
 289                                                 blockTable[blockNb].srcPtr,blockTable[blockNb].srcSize,
 290                                                 cLevel);
 291 #else
 292                                 rSize = ZSTD_compress_advanced (ctx,
 293                                                 blockTable[blockNb].cPtr,  blockTable[blockNb].cRoom,
 294                                                 blockTable[blockNb].srcPtr,blockTable[blockNb].srcSize, NULL, 0, zparams);
 295 #endif
 296                             }
 297                             if (ZSTD_isError(rSize)) EXM_THROW(1, "ZSTD_compress_usingCDict() failed : %s", ZSTD_getErrorName(rSize));
 298                             blockTable[blockNb].cSize = rSize;
 299                         }
 300                         nbLoops++;
 301                     } while (UTIL_clockSpanMicro(clockStart, ticksPerSecond) < clockLoop);
 302                     ZSTD_freeCDict(cdict);
 303                     {   U64 const clockSpanMicro = UTIL_clockSpanMicro(clockStart, ticksPerSecond);
 304                         if (clockSpanMicro < fastestC*nbLoops) fastestC = clockSpanMicro / nbLoops;
 305                         totalCTime += clockSpanMicro;
 306                         cCompleted = (totalCTime >= maxTime);
 307                 }   }
 308
 309                 cSize = 0;
 310                 { U32 blockNb; for (blockNb=0; blockNb<nbBlocks; blockNb++) cSize += blockTable[blockNb].cSize; }
 311                 ratio = (double)srcSize / (double)cSize;
 312                 markNb = (markNb+1) % NB_MARKS;
 313                 DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s\r",
 314                         marks[markNb], displayName, (U32)srcSize, (U32)cSize, ratio,
 315                         (double)srcSize / fastestC );
 316             } else {   /* g_decodeOnly */
 317                 memcpy(compressedBuffer, srcBuffer, loadedCompressedSize);
 318             }
 319
 320 #if 0       /* disable decompression test */
 321             dCompleted=1;
 322             (void)totalDTime; (void)fastestD; (void)crcOrig;   /*  unused when decompression disabled */
 323 #else
 324             /* Decompression */
 325             if (!dCompleted) memset(resultBuffer, 0xD6, srcSize);  /* warm result buffer */
 326
 327             UTIL_sleepMilli(1); /* give processor time to other processes */
 328             UTIL_waitForNextTick(ticksPerSecond);
 329
 330             if (!dCompleted) {
 331                 U64 clockLoop = g_nbSeconds ? TIMELOOP_MICROSEC : 1;
 332                 U32 nbLoops = 0;
 333                 UTIL_time_t clockStart;
 334                 ZSTD_DDict* const ddict = ZSTD_createDDict(dictBuffer, dictBufferSize);
 335                 if (!ddict) EXM_THROW(2, "ZSTD_createDDict() allocation failure");
 336                 UTIL_getTime(&clockStart);
 337                 do {
 338                     U32 blockNb;
 339                     for (blockNb=0; blockNb<nbBlocks; blockNb++) {
 340                         size_t const regenSize = ZSTD_decompress_usingDDict(dctx,
 341                             blockTable[blockNb].resPtr, blockTable[blockNb].resSize,
 342                             blockTable[blockNb].cPtr, blockTable[blockNb].cSize,
 343                             ddict);
 344                         if (ZSTD_isError(regenSize)) {
 345                             DISPLAY("ZSTD_decompress_usingDDict() failed on block %u of size %u : %s  \n",
 346                                       blockNb, (U32)blockTable[blockNb].cSize, ZSTD_getErrorName(regenSize));
 347                             clockLoop = 0;   /* force immediate test end */
 348                             break;
 349                         }
 350                         blockTable[blockNb].resSize = regenSize;
 351                     }
 352                     nbLoops++;
 353                 } while (UTIL_clockSpanMicro(clockStart, ticksPerSecond) < clockLoop);
 354                 ZSTD_freeDDict(ddict);
 355                 {   U64 const clockSpanMicro = UTIL_clockSpanMicro(clockStart, ticksPerSecond);
 356                     if (clockSpanMicro < fastestD*nbLoops) fastestD = clockSpanMicro / nbLoops;
 357                     totalDTime += clockSpanMicro;
 358                     dCompleted = (totalDTime >= maxTime);
 359             }   }
 360
 361             markNb = (markNb+1) % NB_MARKS;
 362             DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s ,%6.1f MB/s\r",
 363                     marks[markNb], displayName, (U32)srcSize, (U32)cSize, ratio,
 364                     (double)srcSize / fastestC,
 365                     (double)srcSize / fastestD );
 366
 367             /* CRC Checking */
 368             {   U64 const crcCheck = XXH64(resultBuffer, srcSize, 0);
 369                 if (!g_decodeOnly && (crcOrig!=crcCheck)) {
 370                     size_t u;
 371                     DISPLAY("!!! WARNING !!! %14s : Invalid Checksum : %x != %x   \n", displayName, (unsigned)crcOrig, (unsigned)crcCheck);
 372                     for (u=0; u<srcSize; u++) {
 373                         if (((const BYTE*)srcBuffer)[u] != ((const BYTE*)resultBuffer)[u]) {
 374                             U32 segNb, bNb, pos;
 375                             size_t bacc = 0;
 376                             DISPLAY("Decoding error at pos %u ", (U32)u);
 377                             for (segNb = 0; segNb < nbBlocks; segNb++) {
 378                                 if (bacc + blockTable[segNb].srcSize > u) break;
 379                                 bacc += blockTable[segNb].srcSize;
 380                             }
 381                             pos = (U32)(u - bacc);
 382                             bNb = pos / (128 KB);
 383                             DISPLAY("(block %u, sub %u, pos %u) \n", segNb, bNb, pos);
 384                             if (u>5) {
 385                                 int n;
 386                                 for (n=-5; n<0; n++) DISPLAY("%02X ", ((const BYTE*)srcBuffer)[u+n]);
 387                                 DISPLAY(" :%02X:  ", ((const BYTE*)srcBuffer)[u]);
 388                                 for (n=1; n<3; n++) DISPLAY("%02X ", ((const BYTE*)srcBuffer)[u+n]);
 389                                 DISPLAY(" \n");
 390                                 for (n=-5; n<0; n++) DISPLAY("%02X ", ((const BYTE*)resultBuffer)[u+n]);
 391                                 DISPLAY(" :%02X:  ", ((const BYTE*)resultBuffer)[u]);
 392                                 for (n=1; n<3; n++) DISPLAY("%02X ", ((const BYTE*)resultBuffer)[u+n]);
 393                                 DISPLAY(" \n");
 394                             }
 395                             break;
 396                         }
 397                         if (u==srcSize-1) {  /* should never happen */
 398                             DISPLAY("no difference detected\n");
 399                     }   }
 400                     break;
 401             }   }   /* CRC Checking */
 402 #endif
 403         }   /* for (testNb = 1; testNb <= (g_nbSeconds + !g_nbSeconds); testNb++) */
 404
 405         if (g_displayLevel == 1) {
 406             double cSpeed = (double)srcSize / fastestC;
 407             double dSpeed = (double)srcSize / fastestD;
 408             if (g_additionalParam)
 409                 DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s  %s (param=%d)\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName, g_additionalParam);
 410             else
 411                 DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s  %s\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName);
 412         }
 413         DISPLAYLEVEL(2, "%2i#\n", cLevel);
 414     }   /* Bench */
 415
 416     /* clean up */
 417     free(blockTable);
 418     free(compressedBuffer);
 419     free(resultBuffer);
 420     ZSTDMT_freeCCtx(mtctx);
 421     ZSTD_freeCCtx(ctx);
 422     ZSTD_freeDCtx(dctx);
 423     return 0;
 424 }
 425
 426
 427 static size_t BMK_findMaxMem(U64 requiredMem)
 428 {
 429     size_t const step = 64 MB;
 430     BYTE* testmem = NULL;
 431
 432     requiredMem = (((requiredMem >> 26) + 1) << 26);
 433     requiredMem += step;
 434     if (requiredMem > maxMemory) requiredMem = maxMemory;
 435
 436     do {
 437         testmem = (BYTE*)malloc((size_t)requiredMem);
 438         requiredMem -= step;
 439     } while (!testmem);
 440
 441     free(testmem);
 442     return (size_t)(requiredMem);
 443 }
 444
 445 static void BMK_benchCLevel(void* srcBuffer, size_t benchedSize,
 446                             const char* displayName, int cLevel, int cLevelLast,
 447                             const size_t* fileSizes, unsigned nbFiles,
 448                             const void* dictBuffer, size_t dictBufferSize,
 449                             ZSTD_compressionParameters *compressionParams, int setRealTimePrio)
 450 {
 451     int l;
 452
 453     const char* pch = strrchr(displayName, '\\'); /* Windows */
 454     if (!pch) pch = strrchr(displayName, '/'); /* Linux */
 455     if (pch) displayName = pch+1;
 456
 457     if (setRealTimePrio) {
 458         DISPLAYLEVEL(2, "Note : switching to a real-time priority \n");
 459         SET_REALTIME_PRIORITY;
 460     }
 461
 462     if (g_displayLevel == 1 && !g_additionalParam)
 463         DISPLAY("bench %s %s: input %u bytes, %u seconds, %u KB blocks\n", ZSTD_VERSION_STRING, ZSTD_GIT_COMMIT_STRING, (U32)benchedSize, g_nbSeconds, (U32)(g_blockSize>>10));
 464
 465     if (cLevelLast < cLevel) cLevelLast = cLevel;
 466
 467     for (l=cLevel; l <= cLevelLast; l++) {
 468         BMK_benchMem(srcBuffer, benchedSize,
 469                      displayName, l,
 470                      fileSizes, nbFiles,
 471                      dictBuffer, dictBufferSize, compressionParams);
 472     }
 473 }
 474
 475
 476 /*! BMK_loadFiles() :
 477     Loads `buffer` with content of files listed within `fileNamesTable`.
 478     At most, fills `buffer` entirely */
 479 static void BMK_loadFiles(void* buffer, size_t bufferSize,
 480                           size_t* fileSizes,
 481                           const char** fileNamesTable, unsigned nbFiles)
 482 {
 483     size_t pos = 0, totalSize = 0;
 484     unsigned n;
 485     for (n=0; n<nbFiles; n++) {
 486         FILE* f;
 487         U64 fileSize = UTIL_getFileSize(fileNamesTable[n]);
 488         if (UTIL_isDirectory(fileNamesTable[n])) {
 489             DISPLAYLEVEL(2, "Ignoring %s directory...       \n", fileNamesTable[n]);
 490             fileSizes[n] = 0;
 491             continue;
 492         }
 493         f = fopen(fileNamesTable[n], "rb");
 494         if (f==NULL) EXM_THROW(10, "impossible to open file %s", fileNamesTable[n]);
 495         DISPLAYUPDATE(2, "Loading %s...       \r", fileNamesTable[n]);
 496         if (fileSize > bufferSize-pos) fileSize = bufferSize-pos, nbFiles=n;   /* buffer too small - stop after this file */
 497         { size_t const readSize = fread(((char*)buffer)+pos, 1, (size_t)fileSize, f);
 498           if (readSize != (size_t)fileSize) EXM_THROW(11, "could not read %s", fileNamesTable[n]);
 499           pos += readSize; }
 500         fileSizes[n] = (size_t)fileSize;
 501         totalSize += (size_t)fileSize;
 502         fclose(f);
 503     }
 504
 505     if (totalSize == 0) EXM_THROW(12, "no data to bench");
 506 }
 507
 508 static void BMK_benchFileTable(const char** fileNamesTable, unsigned nbFiles, const char* dictFileName, int cLevel,
 509                                int cLevelLast, ZSTD_compressionParameters *compressionParams, int setRealTimePrio)
 510 {
 511     void* srcBuffer;
 512     size_t benchedSize;
 513     void* dictBuffer = NULL;
 514     size_t dictBufferSize = 0;
 515     size_t* fileSizes = (size_t*)malloc(nbFiles * sizeof(size_t));
 516     U64 const totalSizeToLoad = UTIL_getTotalFileSize(fileNamesTable, nbFiles);
 517     char mfName[20] = {0};
 518
 519     if (!fileSizes) EXM_THROW(12, "not enough memory for fileSizes");
 520
 521     /* Load dictionary */
 522     if (dictFileName != NULL) {
 523         U64 dictFileSize = UTIL_getFileSize(dictFileName);
 524         if (dictFileSize > 64 MB) EXM_THROW(10, "dictionary file %s too large", dictFileName);
 525         dictBufferSize = (size_t)dictFileSize;
 526         dictBuffer = malloc(dictBufferSize);
 527         if (dictBuffer==NULL) EXM_THROW(11, "not enough memory for dictionary (%u bytes)", (U32)dictBufferSize);
 528         BMK_loadFiles(dictBuffer, dictBufferSize, fileSizes, &dictFileName, 1);
 529     }
 530
 531     /* Memory allocation & restrictions */
 532     benchedSize = BMK_findMaxMem(totalSizeToLoad * 3) / 3;
 533     if ((U64)benchedSize > totalSizeToLoad) benchedSize = (size_t)totalSizeToLoad;
 534     if (benchedSize < totalSizeToLoad)
 535         DISPLAY("Not enough memory; testing %u MB only...\n", (U32)(benchedSize >> 20));
 536     srcBuffer = malloc(benchedSize);
 537     if (!srcBuffer) EXM_THROW(12, "not enough memory");
 538
 539     /* Load input buffer */
 540     BMK_loadFiles(srcBuffer, benchedSize, fileSizes, fileNamesTable, nbFiles);
 541
 542     /* Bench */
 543     snprintf (mfName, sizeof(mfName), " %u files", nbFiles);
 544     {   const char* displayName = (nbFiles > 1) ? mfName : fileNamesTable[0];
 545         BMK_benchCLevel(srcBuffer, benchedSize,
 546                         displayName, cLevel, cLevelLast,
 547                         fileSizes, nbFiles,
 548                         dictBuffer, dictBufferSize, compressionParams, setRealTimePrio);
 549     }
 550
 551     /* clean up */
 552     free(srcBuffer);
 553     free(dictBuffer);
 554     free(fileSizes);
 555 }
 556
 557
 558 static void BMK_syntheticTest(int cLevel, int cLevelLast, double compressibility, ZSTD_compressionParameters* compressionParams, int setRealTimePrio)
 559 {
 560     char name[20] = {0};
 561     size_t benchedSize = 10000000;
 562     void* const srcBuffer = malloc(benchedSize);
 563
 564     /* Memory allocation */
 565     if (!srcBuffer) EXM_THROW(21, "not enough memory");
 566
 567     /* Fill input buffer */
 568     RDG_genBuffer(srcBuffer, benchedSize, compressibility, 0.0, 0);
 569
 570     /* Bench */
 571     snprintf (name, sizeof(name), "Synthetic %2u%%", (unsigned)(compressibility*100));
 572     BMK_benchCLevel(srcBuffer, benchedSize, name, cLevel, cLevelLast, &benchedSize, 1, NULL, 0, compressionParams, setRealTimePrio);
 573
 574     /* clean up */
 575     free(srcBuffer);
 576 }
 577
 578
 579 int BMK_benchFiles(const char** fileNamesTable, unsigned nbFiles, const char* dictFileName,
 580                    int cLevel, int cLevelLast, ZSTD_compressionParameters* compressionParams, int setRealTimePrio)
 581 {
 582     double const compressibility = (double)g_compressibilityDefault / 100;
 583
 584     if (cLevel < 1) cLevel = 1;   /* minimum compression level */
 585     if (cLevel > ZSTD_maxCLevel()) cLevel = ZSTD_maxCLevel();
 586     if (cLevelLast > ZSTD_maxCLevel()) cLevelLast = ZSTD_maxCLevel();
 587     if (cLevelLast < cLevel) cLevelLast = cLevel;
 588     if (cLevelLast > cLevel) DISPLAYLEVEL(2, "Benchmarking levels from %d to %d\n", cLevel, cLevelLast);
 589
 590     if (nbFiles == 0)
 591         BMK_syntheticTest(cLevel, cLevelLast, compressibility, compressionParams, setRealTimePrio);
 592     else
 593         BMK_benchFileTable(fileNamesTable, nbFiles, dictFileName, cLevel, cLevelLast, compressionParams, setRealTimePrio);
 594     return 0;
 595 }