ceph/src/zstd/tests/fullbench.c

   1 /*
   2  * Copyright (c) 2015-present, Yann Collet, Facebook, Inc.
   3  * All rights reserved.
   4  *
   5  * This source code is licensed under both the BSD-style license (found in the
   6  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
   7  * in the COPYING file in the root directory of this source tree).
   8  * You may select, at your option, one of the above-listed licenses.
   9  */
  10
  11
  12 /*_************************************
  13 *  Includes
  14 **************************************/
  15 #include "util.h"        /* Compiler options, UTIL_GetFileSize */
  16 #include <stdlib.h>      /* malloc */
  17 #include <stdio.h>       /* fprintf, fopen, ftello64 */
  18 #include <assert.h>      /* assert */
  19
  20 #include "timefn.h"      /* UTIL_clockSpanNano, UTIL_getTime */
  21 #include "mem.h"         /* U32 */
  22 #ifndef ZSTD_DLL_IMPORT
  23     #include "zstd_internal.h"   /* ZSTD_decodeSeqHeaders, ZSTD_blockHeaderSize, blockType_e, KB, MB */
  24 #else
  25     #define KB *(1 <<10)
  26     #define MB *(1 <<20)
  27     #define GB *(1U<<30)
  28     typedef enum { bt_raw, bt_rle, bt_compressed, bt_reserved } blockType_e;
  29 #endif
  30 #define ZSTD_STATIC_LINKING_ONLY  /* ZSTD_compressBegin, ZSTD_compressContinue, etc. */
  31 #include "zstd.h"        /* ZSTD_versionString */
  32 #include "util.h"        /* time functions */
  33 #include "datagen.h"
  34 #include "benchfn.h"       /* CustomBench*/
  35 #include "benchzstd.h"     /* MB_UNIT */
  36
  37
  38 /*_************************************
  39 *  Constants
  40 **************************************/
  41 #define PROGRAM_DESCRIPTION "Zstandard speed analyzer"
  42 #define AUTHOR "Yann Collet"
  43 #define WELCOME_MESSAGE "*** %s %s %i-bits, by %s (%s) ***\n", PROGRAM_DESCRIPTION, ZSTD_versionString(), (int)(sizeof(void*)*8), AUTHOR, __DATE__
  44
  45 #define NBLOOPS    6
  46 #define TIMELOOP_S 2
  47
  48 #define KNUTH      2654435761U
  49 #define MAX_MEM    (1984 MB)
  50
  51 #define DEFAULT_CLEVEL 1
  52
  53 #define COMPRESSIBILITY_DEFAULT 0.50
  54 static const size_t g_sampleSize = 10000000;
  55
  56 #define TIMELOOP_NANOSEC      (1*1000000000ULL) /* 1 second */
  57
  58
  59 /*_************************************
  60 *  Macros
  61 **************************************/
  62 #define DISPLAY(...)  fprintf(stderr, __VA_ARGS__)
  63
  64
  65 /*_************************************
  66 *  Benchmark Parameters
  67 **************************************/
  68 static unsigned g_nbIterations = NBLOOPS;
  69 static double g_compressibility = COMPRESSIBILITY_DEFAULT;
  70
  71
  72 /*_*******************************************************
  73 *  Private functions
  74 *********************************************************/
  75 static size_t BMK_findMaxMem(U64 requiredMem)
  76 {
  77     size_t const step = 64 MB;
  78     void* testmem = NULL;
  79
  80     requiredMem = (((requiredMem >> 26) + 1) << 26);
  81     if (requiredMem > MAX_MEM) requiredMem = MAX_MEM;
  82
  83     requiredMem += step;
  84     do {
  85         testmem = malloc ((size_t)requiredMem);
  86         requiredMem -= step;
  87     } while (!testmem);
  88
  89     free (testmem);
  90     return (size_t) requiredMem;
  91 }
  92
  93
  94 /*_*******************************************************
  95 *  Benchmark wrappers
  96 *********************************************************/
  97
  98 static ZSTD_CCtx* g_zcc = NULL;
  99
 100 static size_t
 101 local_ZSTD_compress(const void* src, size_t srcSize,
 102                     void* dst, size_t dstSize,
 103                     void* buff2)
 104 {
 105     ZSTD_parameters p;
 106     ZSTD_frameParameters f = { 1 /* contentSizeHeader*/, 0, 0 };
 107     p.fParams = f;
 108     p.cParams = *(ZSTD_compressionParameters*)buff2;
 109     return ZSTD_compress_advanced (g_zcc, dst, dstSize, src, srcSize, NULL ,0, p);
 110     //return ZSTD_compress(dst, dstSize, src, srcSize, cLevel);
 111 }
 112
 113 static size_t g_cSize = 0;
 114 static size_t local_ZSTD_decompress(const void* src, size_t srcSize,
 115                                     void* dst, size_t dstSize,
 116                                     void* buff2)
 117 {
 118     (void)src; (void)srcSize;
 119     return ZSTD_decompress(dst, dstSize, buff2, g_cSize);
 120 }
 121
 122 static ZSTD_DCtx* g_zdc = NULL;
 123
 124 #ifndef ZSTD_DLL_IMPORT
 125 extern size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* ctx, const void* src, size_t srcSize);
 126 static size_t local_ZSTD_decodeLiteralsBlock(const void* src, size_t srcSize, void* dst, size_t dstSize, void* buff2)
 127 {
 128     (void)src; (void)srcSize; (void)dst; (void)dstSize;
 129     return ZSTD_decodeLiteralsBlock((ZSTD_DCtx*)g_zdc, buff2, g_cSize);
 130 }
 131
 132 static size_t local_ZSTD_decodeSeqHeaders(const void* src, size_t srcSize, void* dst, size_t dstSize, void* buff2)
 133 {
 134     int nbSeq;
 135     (void)src; (void)srcSize; (void)dst; (void)dstSize;
 136     return ZSTD_decodeSeqHeaders(g_zdc, &nbSeq, buff2, g_cSize);
 137 }
 138 #endif
 139
 140 static ZSTD_CStream* g_cstream= NULL;
 141 static size_t
 142 local_ZSTD_compressStream(const void* src, size_t srcSize,
 143                           void* dst, size_t dstCapacity,
 144                           void* buff2)
 145 {
 146     ZSTD_outBuffer buffOut;
 147     ZSTD_inBuffer buffIn;
 148     ZSTD_parameters p;
 149     ZSTD_frameParameters f = {1 /* contentSizeHeader*/, 0, 0};
 150     p.fParams = f;
 151     p.cParams = *(ZSTD_compressionParameters*)buff2;
 152     ZSTD_initCStream_advanced(g_cstream, NULL, 0, p, ZSTD_CONTENTSIZE_UNKNOWN);
 153     buffOut.dst = dst;
 154     buffOut.size = dstCapacity;
 155     buffOut.pos = 0;
 156     buffIn.src = src;
 157     buffIn.size = srcSize;
 158     buffIn.pos = 0;
 159     ZSTD_compressStream(g_cstream, &buffOut, &buffIn);
 160     ZSTD_endStream(g_cstream, &buffOut);
 161     return buffOut.pos;
 162 }
 163
 164 static size_t
 165 local_ZSTD_compress_generic_end(const void* src, size_t srcSize,
 166                                 void* dst, size_t dstCapacity,
 167                                 void* buff2)
 168 {
 169     (void)buff2;
 170     return ZSTD_compress2(g_cstream, dst, dstCapacity, src, srcSize);
 171 }
 172
 173 static size_t
 174 local_ZSTD_compress_generic_continue(const void* src, size_t srcSize,
 175                                      void* dst, size_t dstCapacity,
 176                                      void* buff2)
 177 {
 178     ZSTD_outBuffer buffOut;
 179     ZSTD_inBuffer buffIn;
 180     (void)buff2;
 181     buffOut.dst = dst;
 182     buffOut.size = dstCapacity;
 183     buffOut.pos = 0;
 184     buffIn.src = src;
 185     buffIn.size = srcSize;
 186     buffIn.pos = 0;
 187     ZSTD_compressStream2(g_cstream, &buffOut, &buffIn, ZSTD_e_continue);
 188     ZSTD_compressStream2(g_cstream, &buffOut, &buffIn, ZSTD_e_end);
 189     return buffOut.pos;
 190 }
 191
 192 static size_t
 193 local_ZSTD_compress_generic_T2_end(const void* src, size_t srcSize,
 194                                    void* dst, size_t dstCapacity,
 195                                    void* buff2)
 196 {
 197     (void)buff2;
 198     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_nbWorkers, 2);
 199     return ZSTD_compress2(g_cstream, dst, dstCapacity, src, srcSize);
 200 }
 201
 202 static size_t
 203 local_ZSTD_compress_generic_T2_continue(const void* src, size_t srcSize,
 204                                         void* dst, size_t dstCapacity,
 205                                         void* buff2)
 206 {
 207     ZSTD_outBuffer buffOut;
 208     ZSTD_inBuffer buffIn;
 209     (void)buff2;
 210     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_nbWorkers, 2);
 211     buffOut.dst = dst;
 212     buffOut.size = dstCapacity;
 213     buffOut.pos = 0;
 214     buffIn.src = src;
 215     buffIn.size = srcSize;
 216     buffIn.pos = 0;
 217     ZSTD_compressStream2(g_cstream, &buffOut, &buffIn, ZSTD_e_continue);
 218     while(ZSTD_compressStream2(g_cstream, &buffOut, &buffIn, ZSTD_e_end)) {}
 219     return buffOut.pos;
 220 }
 221
 222 static ZSTD_DStream* g_dstream= NULL;
 223 static size_t
 224 local_ZSTD_decompressStream(const void* src, size_t srcSize,
 225                             void* dst, size_t dstCapacity,
 226                             void* buff2)
 227 {
 228     ZSTD_outBuffer buffOut;
 229     ZSTD_inBuffer buffIn;
 230     (void)src; (void)srcSize;
 231     ZSTD_initDStream(g_dstream);
 232     buffOut.dst = dst;
 233     buffOut.size = dstCapacity;
 234     buffOut.pos = 0;
 235     buffIn.src = buff2;
 236     buffIn.size = g_cSize;
 237     buffIn.pos = 0;
 238     ZSTD_decompressStream(g_dstream, &buffOut, &buffIn);
 239     return buffOut.pos;
 240 }
 241
 242 #ifndef ZSTD_DLL_IMPORT
 243 static size_t local_ZSTD_compressContinue(const void* src, size_t srcSize,
 244                                           void* dst, size_t dstCapacity,
 245                                           void* buff2)
 246 {
 247     ZSTD_parameters p;
 248     ZSTD_frameParameters f = { 1 /* contentSizeHeader*/, 0, 0 };
 249     p.fParams = f;
 250     p.cParams = *(ZSTD_compressionParameters*)buff2;
 251     ZSTD_compressBegin_advanced(g_zcc, NULL, 0, p, srcSize);
 252     return ZSTD_compressEnd(g_zcc, dst, dstCapacity, src, srcSize);
 253 }
 254
 255 #define FIRST_BLOCK_SIZE 8
 256 static size_t local_ZSTD_compressContinue_extDict(const void* src, size_t srcSize,
 257                                                   void* dst, size_t dstCapacity,
 258                                                   void* buff2)
 259 {
 260     BYTE firstBlockBuf[FIRST_BLOCK_SIZE];
 261
 262     ZSTD_parameters p;
 263     ZSTD_frameParameters f = { 1, 0, 0 };
 264     p.fParams = f;
 265     p.cParams = *(ZSTD_compressionParameters*)buff2;
 266     ZSTD_compressBegin_advanced(g_zcc, NULL, 0, p, srcSize);
 267     memcpy(firstBlockBuf, src, FIRST_BLOCK_SIZE);
 268
 269     {   size_t const compressResult = ZSTD_compressContinue(g_zcc,
 270                                             dst, dstCapacity,
 271                                             firstBlockBuf, FIRST_BLOCK_SIZE);
 272         if (ZSTD_isError(compressResult)) {
 273             DISPLAY("local_ZSTD_compressContinue_extDict error : %s\n",
 274                     ZSTD_getErrorName(compressResult));
 275             return compressResult;
 276         }
 277         dst = (BYTE*)dst + compressResult;
 278         dstCapacity -= compressResult;
 279     }
 280     return ZSTD_compressEnd(g_zcc, dst, dstCapacity,
 281                             (const BYTE*)src + FIRST_BLOCK_SIZE,
 282                             srcSize - FIRST_BLOCK_SIZE);
 283 }
 284
 285 static size_t local_ZSTD_decompressContinue(const void* src, size_t srcSize,
 286                                             void* dst, size_t dstCapacity,
 287                                             void* buff2)
 288 {
 289     size_t regeneratedSize = 0;
 290     const BYTE* ip = (const BYTE*)buff2;
 291     const BYTE* const iend = ip + g_cSize;
 292     BYTE* op = (BYTE*)dst;
 293     size_t remainingCapacity = dstCapacity;
 294
 295     (void)src; (void)srcSize;  /* unused */
 296     ZSTD_decompressBegin(g_zdc);
 297     while (ip < iend) {
 298         size_t const iSize = ZSTD_nextSrcSizeToDecompress(g_zdc);
 299         size_t const decodedSize = ZSTD_decompressContinue(g_zdc, op, remainingCapacity, ip, iSize);
 300         ip += iSize;
 301         regeneratedSize += decodedSize;
 302         op += decodedSize;
 303         remainingCapacity -= decodedSize;
 304     }
 305
 306     return regeneratedSize;
 307 }
 308 #endif
 309
 310
 311 /*_*******************************************************
 312 *  Bench functions
 313 *********************************************************/
 314 static int benchMem(unsigned benchNb,
 315                     const void* src, size_t srcSize,
 316                     int cLevel, ZSTD_compressionParameters cparams)
 317 {
 318     size_t dstBuffSize = ZSTD_compressBound(srcSize);
 319     BYTE*  dstBuff;
 320     void*  dstBuff2;
 321     void*  buff2;
 322     const char* benchName;
 323     BMK_benchFn_t benchFunction;
 324     int errorcode = 0;
 325
 326     /* Selection */
 327     switch(benchNb)
 328     {
 329     case 1:
 330         benchFunction = local_ZSTD_compress; benchName = "compress";
 331         break;
 332     case 2:
 333         benchFunction = local_ZSTD_decompress; benchName = "decompress";
 334         break;
 335 #ifndef ZSTD_DLL_IMPORT
 336     case 11:
 337         benchFunction = local_ZSTD_compressContinue; benchName = "compressContinue";
 338         break;
 339     case 12:
 340         benchFunction = local_ZSTD_compressContinue_extDict; benchName = "compressContinue_extDict";
 341         break;
 342     case 13:
 343         benchFunction = local_ZSTD_decompressContinue; benchName = "decompressContinue";
 344         break;
 345     case 31:
 346         benchFunction = local_ZSTD_decodeLiteralsBlock; benchName = "decodeLiteralsBlock";
 347         break;
 348     case 32:
 349         benchFunction = local_ZSTD_decodeSeqHeaders; benchName = "decodeSeqHeaders";
 350         break;
 351 #endif
 352     case 41:
 353         benchFunction = local_ZSTD_compressStream; benchName = "compressStream";
 354         break;
 355     case 42:
 356         benchFunction = local_ZSTD_decompressStream; benchName = "decompressStream";
 357         break;
 358     case 51:
 359         benchFunction = local_ZSTD_compress_generic_continue; benchName = "compress_generic, continue";
 360         break;
 361     case 52:
 362         benchFunction = local_ZSTD_compress_generic_end; benchName = "compress_generic, end";
 363         break;
 364     case 61:
 365         benchFunction = local_ZSTD_compress_generic_T2_continue; benchName = "compress_generic, -T2, continue";
 366         break;
 367     case 62:
 368         benchFunction = local_ZSTD_compress_generic_T2_end; benchName = "compress_generic, -T2, end";
 369         break;
 370     default :
 371         return 0;
 372     }
 373
 374     /* Allocation */
 375     dstBuff = (BYTE*)malloc(dstBuffSize);
 376     dstBuff2 = malloc(dstBuffSize);
 377     if ((!dstBuff) || (!dstBuff2)) {
 378         DISPLAY("\nError: not enough memory!\n");
 379         free(dstBuff); free(dstBuff2);
 380         return 12;
 381     }
 382     buff2 = dstBuff2;
 383     if (g_zcc==NULL) g_zcc = ZSTD_createCCtx();
 384     if (g_zdc==NULL) g_zdc = ZSTD_createDCtx();
 385     if (g_cstream==NULL) g_cstream = ZSTD_createCStream();
 386     if (g_dstream==NULL) g_dstream = ZSTD_createDStream();
 387
 388     /* DISPLAY("params: cLevel %d, wlog %d hlog %d clog %d slog %d mml %d tlen %d strat %d \n",
 389           cLevel, cparams->windowLog, cparams->hashLog, cparams->chainLog, cparams->searchLog,
 390           cparams->minMatch, cparams->targetLength, cparams->strategy); */
 391
 392     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_compressionLevel, cLevel);
 393     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_windowLog, (int)cparams.windowLog);
 394     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_hashLog, (int)cparams.hashLog);
 395     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_chainLog, (int)cparams.chainLog);
 396     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_searchLog, (int)cparams.searchLog);
 397     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_minMatch, (int)cparams.minMatch);
 398     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_targetLength, (int)cparams.targetLength);
 399     ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_strategy, cparams.strategy);
 400
 401
 402     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_compressionLevel, cLevel);
 403     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_windowLog, (int)cparams.windowLog);
 404     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_hashLog, (int)cparams.hashLog);
 405     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_chainLog, (int)cparams.chainLog);
 406     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_searchLog, (int)cparams.searchLog);
 407     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_minMatch, (int)cparams.minMatch);
 408     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_targetLength, (int)cparams.targetLength);
 409     ZSTD_CCtx_setParameter(g_cstream, ZSTD_c_strategy, cparams.strategy);
 410
 411     /* Preparation */
 412     switch(benchNb)
 413     {
 414     case 1:
 415         buff2 = &cparams;
 416         break;
 417     case 2:
 418         g_cSize = ZSTD_compress(buff2, dstBuffSize, src, srcSize, cLevel);
 419         break;
 420 #ifndef ZSTD_DLL_IMPORT
 421     case 11:
 422         buff2 = &cparams;
 423         break;
 424     case 12:
 425         buff2 = &cparams;
 426         break;
 427     case 13 :
 428         g_cSize = ZSTD_compress(buff2, dstBuffSize, src, srcSize, cLevel);
 429         break;
 430     case 31:  /* ZSTD_decodeLiteralsBlock */
 431         {   blockProperties_t bp;
 432             ZSTD_frameHeader zfp;
 433             size_t frameHeaderSize, skippedSize;
 434             g_cSize = ZSTD_compress(dstBuff, dstBuffSize, src, srcSize, cLevel);
 435             frameHeaderSize = ZSTD_getFrameHeader(&zfp, dstBuff, ZSTD_FRAMEHEADERSIZE_MIN);
 436             if (frameHeaderSize==0) frameHeaderSize = ZSTD_FRAMEHEADERSIZE_MIN;
 437             ZSTD_getcBlockSize(dstBuff+frameHeaderSize, dstBuffSize, &bp);  /* Get 1st block type */
 438             if (bp.blockType != bt_compressed) {
 439                 DISPLAY("ZSTD_decodeLiteralsBlock : impossible to test on this sample (not compressible)\n");
 440                 goto _cleanOut;
 441             }
 442             skippedSize = frameHeaderSize + ZSTD_blockHeaderSize;
 443             memcpy(buff2, dstBuff+skippedSize, g_cSize-skippedSize);
 444             srcSize = srcSize > 128 KB ? 128 KB : srcSize;    /* speed relative to block */
 445             ZSTD_decompressBegin(g_zdc);
 446             break;
 447         }
 448     case 32:   /* ZSTD_decodeSeqHeaders */
 449         {   blockProperties_t bp;
 450             ZSTD_frameHeader zfp;
 451             const BYTE* ip = dstBuff;
 452             const BYTE* iend;
 453             size_t frameHeaderSize, cBlockSize;
 454             ZSTD_compress(dstBuff, dstBuffSize, src, srcSize, cLevel);   /* it would be better to use direct block compression here */
 455             g_cSize = ZSTD_compress(dstBuff, dstBuffSize, src, srcSize, cLevel);
 456             frameHeaderSize = ZSTD_getFrameHeader(&zfp, dstBuff, ZSTD_FRAMEHEADERSIZE_MIN);
 457             if (frameHeaderSize==0) frameHeaderSize = ZSTD_FRAMEHEADERSIZE_MIN;
 458             ip += frameHeaderSize;   /* Skip frame Header */
 459             cBlockSize = ZSTD_getcBlockSize(ip, dstBuffSize, &bp);   /* Get 1st block type */
 460             if (bp.blockType != bt_compressed) {
 461                 DISPLAY("ZSTD_decodeSeqHeaders : impossible to test on this sample (not compressible)\n");
 462                 goto _cleanOut;
 463             }
 464             iend = ip + ZSTD_blockHeaderSize + cBlockSize;   /* End of first block */
 465             ip += ZSTD_blockHeaderSize;                      /* skip block header */
 466             ZSTD_decompressBegin(g_zdc);
 467             assert(iend > ip);
 468             ip += ZSTD_decodeLiteralsBlock(g_zdc, ip, (size_t)(iend-ip));   /* skip literal segment */
 469             g_cSize = (size_t)(iend-ip);
 470             memcpy(buff2, ip, g_cSize);   /* copy rest of block (it starts by SeqHeader) */
 471             srcSize = srcSize > 128 KB ? 128 KB : srcSize;   /* speed relative to block */
 472             break;
 473         }
 474 #else
 475     case 31:
 476         goto _cleanOut;
 477 #endif
 478     case 41 :
 479         buff2 = &cparams;
 480         break;
 481     case 42 :
 482         g_cSize = ZSTD_compress(buff2, dstBuffSize, src, srcSize, cLevel);
 483         break;
 484
 485     /* test functions */
 486     /* convention: test functions have ID > 100 */
 487
 488     default : ;
 489     }
 490
 491      /* warming up dstBuff */
 492     { size_t i; for (i=0; i<dstBuffSize; i++) dstBuff[i]=(BYTE)i; }
 493
 494     /* benchmark loop */
 495     {   BMK_timedFnState_t* const tfs = BMK_createTimedFnState(g_nbIterations * 1000, 1000);
 496         void* const avoidStrictAliasingPtr = &dstBuff;
 497         BMK_benchParams_t bp;
 498         BMK_runTime_t bestResult;
 499         bestResult.sumOfReturn = 0;
 500         bestResult.nanoSecPerRun = (double)TIMELOOP_NANOSEC * 2000000000;  /* hopefully large enough : must be larger than any potential measurement */
 501         assert(tfs != NULL);
 502
 503         bp.benchFn = benchFunction;
 504         bp.benchPayload = buff2;
 505         bp.initFn = NULL;
 506         bp.initPayload = NULL;
 507         bp.errorFn = ZSTD_isError;
 508         bp.blockCount = 1;
 509         bp.srcBuffers = &src;
 510         bp.srcSizes = &srcSize;
 511         bp.dstBuffers = (void* const*) avoidStrictAliasingPtr;  /* circumvent strict aliasing warning on gcc-8,
 512                                                                  * because gcc considers that `void* const *`  and `void**` are 2 different types */
 513         bp.dstCapacities = &dstBuffSize;
 514         bp.blockResults = NULL;
 515
 516         for (;;) {
 517             BMK_runOutcome_t const bOutcome = BMK_benchTimedFn(tfs, bp);
 518
 519             if (!BMK_isSuccessful_runOutcome(bOutcome)) {
 520                 DISPLAY("ERROR benchmarking function ! ! \n");
 521                 errorcode = 1;
 522                 goto _cleanOut;
 523             }
 524
 525             {   BMK_runTime_t const newResult = BMK_extract_runTime(bOutcome);
 526                 if (newResult.nanoSecPerRun < bestResult.nanoSecPerRun )
 527                     bestResult.nanoSecPerRun = newResult.nanoSecPerRun;
 528                 DISPLAY("\r%2u#%-29.29s:%8.1f MB/s  (%8u) ",
 529                         benchNb, benchName,
 530                         (double)srcSize * TIMELOOP_NANOSEC / bestResult.nanoSecPerRun / MB_UNIT,
 531                         (unsigned)newResult.sumOfReturn );
 532             }
 533
 534             if ( BMK_isCompleted_TimedFn(tfs) ) break;
 535         }
 536         BMK_freeTimedFnState(tfs);
 537     }
 538     DISPLAY("\n");
 539
 540 _cleanOut:
 541     free(dstBuff);
 542     free(dstBuff2);
 543     ZSTD_freeCCtx(g_zcc); g_zcc=NULL;
 544     ZSTD_freeDCtx(g_zdc); g_zdc=NULL;
 545     ZSTD_freeCStream(g_cstream); g_cstream=NULL;
 546     ZSTD_freeDStream(g_dstream); g_dstream=NULL;
 547     return errorcode;
 548 }
 549
 550
 551 static int benchSample(U32 benchNb,
 552                        int cLevel, ZSTD_compressionParameters cparams)
 553 {
 554     size_t const benchedSize = g_sampleSize;
 555     const char* const name = "Sample 10MiB";
 556
 557     /* Allocation */
 558     void* const origBuff = malloc(benchedSize);
 559     if (!origBuff) { DISPLAY("\nError: not enough memory!\n"); return 12; }
 560
 561     /* Fill buffer */
 562     RDG_genBuffer(origBuff, benchedSize, g_compressibility, 0.0, 0);
 563
 564     /* bench */
 565     DISPLAY("\r%70s\r", "");
 566     DISPLAY(" %s : \n", name);
 567     if (benchNb) {
 568         benchMem(benchNb, origBuff, benchedSize, cLevel, cparams);
 569     } else {  /* 0 == run all tests */
 570         for (benchNb=0; benchNb<100; benchNb++) {
 571             benchMem(benchNb, origBuff, benchedSize, cLevel, cparams);
 572     }   }
 573
 574     free(origBuff);
 575     return 0;
 576 }
 577
 578
 579 static int benchFiles(U32 benchNb,
 580                       const char** fileNamesTable, const int nbFiles,
 581                       int cLevel, ZSTD_compressionParameters cparams)
 582 {
 583     /* Loop for each file */
 584     int fileIdx;
 585     for (fileIdx=0; fileIdx<nbFiles; fileIdx++) {
 586         const char* const inFileName = fileNamesTable[fileIdx];
 587         FILE* const inFile = fopen( inFileName, "rb" );
 588         size_t benchedSize;
 589
 590         /* Check file existence */
 591         if (inFile==NULL) { DISPLAY( "Pb opening %s\n", inFileName); return 11; }
 592
 593         /* Memory allocation & restrictions */
 594         {   U64 const inFileSize = UTIL_getFileSize(inFileName);
 595             if (inFileSize == UTIL_FILESIZE_UNKNOWN) {
 596                 DISPLAY( "Cannot measure size of %s\n", inFileName);
 597                 fclose(inFile);
 598                 return 11;
 599             }
 600             benchedSize = BMK_findMaxMem(inFileSize*3) / 3;
 601             if ((U64)benchedSize > inFileSize)
 602                 benchedSize = (size_t)inFileSize;
 603             if ((U64)benchedSize < inFileSize) {
 604                 DISPLAY("Not enough memory for '%s' full size; testing %u MB only... \n",
 605                         inFileName, (unsigned)(benchedSize>>20));
 606         }   }
 607
 608         /* Alloc */
 609         {   void* const origBuff = malloc(benchedSize);
 610             if (!origBuff) { DISPLAY("\nError: not enough memory!\n"); fclose(inFile); return 12; }
 611
 612             /* Fill input buffer */
 613             DISPLAY("Loading %s...       \r", inFileName);
 614             {   size_t const readSize = fread(origBuff, 1, benchedSize, inFile);
 615                 fclose(inFile);
 616                 if (readSize != benchedSize) {
 617                     DISPLAY("\nError: problem reading file '%s' !!    \n", inFileName);
 618                     free(origBuff);
 619                     return 13;
 620             }   }
 621
 622             /* bench */
 623             DISPLAY("\r%70s\r", "");   /* blank line */
 624             DISPLAY(" %s : \n", inFileName);
 625             if (benchNb) {
 626                 benchMem(benchNb, origBuff, benchedSize, cLevel, cparams);
 627             } else {
 628                 for (benchNb=0; benchNb<100; benchNb++) {
 629                     benchMem(benchNb, origBuff, benchedSize, cLevel, cparams);
 630             }   }
 631
 632             free(origBuff);
 633     }   }
 634
 635     return 0;
 636 }
 637
 638
 639
 640 /*_*******************************************************
 641 *  Argument Parsing
 642 *********************************************************/
 643
 644 #define ERROR_OUT(msg) { DISPLAY("%s \n", msg); exit(1); }
 645
 646 static unsigned readU32FromChar(const char** stringPtr)
 647 {
 648     const char errorMsg[] = "error: numeric value too large";
 649     unsigned result = 0;
 650     while ((**stringPtr >='0') && (**stringPtr <='9')) {
 651         unsigned const max = (((unsigned)(-1)) / 10) - 1;
 652         if (result > max) ERROR_OUT(errorMsg);
 653         result *= 10;
 654         result += (unsigned)(**stringPtr - '0');
 655         (*stringPtr)++ ;
 656     }
 657     if ((**stringPtr=='K') || (**stringPtr=='M')) {
 658         unsigned const maxK = ((unsigned)(-1)) >> 10;
 659         if (result > maxK) ERROR_OUT(errorMsg);
 660         result <<= 10;
 661         if (**stringPtr=='M') {
 662             if (result > maxK) ERROR_OUT(errorMsg);
 663             result <<= 10;
 664         }
 665         (*stringPtr)++;  /* skip `K` or `M` */
 666         if (**stringPtr=='i') (*stringPtr)++;
 667         if (**stringPtr=='B') (*stringPtr)++;
 668     }
 669     return result;
 670 }
 671
 672 static int longCommandWArg(const char** stringPtr, const char* longCommand)
 673 {
 674     size_t const comSize = strlen(longCommand);
 675     int const result = !strncmp(*stringPtr, longCommand, comSize);
 676     if (result) *stringPtr += comSize;
 677     return result;
 678 }
 679
 680
 681 /*_*******************************************************
 682 *  Command line
 683 *********************************************************/
 684
 685 static int usage(const char* exename)
 686 {
 687     DISPLAY( "Usage :\n");
 688     DISPLAY( "      %s [arg] file1 file2 ... fileX\n", exename);
 689     DISPLAY( "Arguments :\n");
 690     DISPLAY( " -H/-h  : Help (this text + advanced options)\n");
 691     return 0;
 692 }
 693
 694 static int usage_advanced(const char* exename)
 695 {
 696     usage(exename);
 697     DISPLAY( "\nAdvanced options :\n");
 698     DISPLAY( " -b#    : test only function # \n");
 699     DISPLAY( " -i#    : iteration loops [1-9](default : %i)\n", NBLOOPS);
 700     DISPLAY( " -P#    : sample compressibility (default : %.1f%%)\n", COMPRESSIBILITY_DEFAULT * 100);
 701     DISPLAY( " -l#    : benchmark functions at that compression level (default : %i)\n", DEFAULT_CLEVEL);
 702     DISPLAY( " --zstd : custom parameter selection. Format same as zstdcli \n");
 703     return 0;
 704 }
 705
 706 static int badusage(const char* exename)
 707 {
 708     DISPLAY("Wrong parameters\n");
 709     usage(exename);
 710     return 1;
 711 }
 712
 713 int main(int argc, const char** argv)
 714 {
 715     int argNb, filenamesStart=0, result;
 716     const char* const exename = argv[0];
 717     const char* input_filename = NULL;
 718     U32 benchNb = 0, main_pause = 0;
 719     int cLevel = DEFAULT_CLEVEL;
 720     ZSTD_compressionParameters cparams = ZSTD_getCParams(cLevel, 0, 0);
 721
 722     DISPLAY(WELCOME_MESSAGE);
 723     if (argc<1) return badusage(exename);
 724
 725     for (argNb=1; argNb<argc; argNb++) {
 726         const char* argument = argv[argNb];
 727         assert(argument != NULL);
 728
 729         if (longCommandWArg(&argument, "--zstd=")) {
 730             for ( ; ;) {
 731                 if (longCommandWArg(&argument, "windowLog=") || longCommandWArg(&argument, "wlog=")) { cparams.windowLog = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; }
 732                 if (longCommandWArg(&argument, "chainLog=") || longCommandWArg(&argument, "clog=")) { cparams.chainLog = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; }
 733                 if (longCommandWArg(&argument, "hashLog=") || longCommandWArg(&argument, "hlog=")) { cparams.hashLog = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; }
 734                 if (longCommandWArg(&argument, "searchLog=") || longCommandWArg(&argument, "slog=")) { cparams.searchLog = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; }
 735                 if (longCommandWArg(&argument, "minMatch=") || longCommandWArg(&argument, "mml=")) { cparams.minMatch = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; }
 736                 if (longCommandWArg(&argument, "targetLength=") || longCommandWArg(&argument, "tlen=")) { cparams.targetLength = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; }
 737                 if (longCommandWArg(&argument, "strategy=") || longCommandWArg(&argument, "strat=")) { cparams.strategy = (ZSTD_strategy)(readU32FromChar(&argument)); if (argument[0]==',') { argument++; continue; } else break; }
 738                 if (longCommandWArg(&argument, "level=") || longCommandWArg(&argument, "lvl=")) { cLevel = (int)readU32FromChar(&argument); cparams = ZSTD_getCParams(cLevel, 0, 0); if (argument[0]==',') { argument++; continue; } else break; }
 739                 DISPLAY("invalid compression parameter \n");
 740                 return 1;
 741             }
 742
 743             /* check end of string */
 744             if (argument[0] != 0) {
 745                 DISPLAY("invalid --zstd= format \n");
 746                 return 1;
 747             } else {
 748                 continue;
 749             }
 750
 751         } else if (argument[0]=='-') { /* Commands (note : aggregated commands are allowed) */
 752             argument++;
 753             while (argument[0]!=0) {
 754
 755                 switch(argument[0])
 756                 {
 757                     /* Display help on usage */
 758                 case 'h':
 759                 case 'H': return usage_advanced(exename);
 760
 761                     /* Pause at the end (hidden option) */
 762                 case 'p': main_pause = 1; break;
 763
 764                     /* Select specific algorithm to bench */
 765                 case 'b':
 766                     argument++;
 767                     benchNb = readU32FromChar(&argument);
 768                     break;
 769
 770                     /* Modify Nb Iterations */
 771                 case 'i':
 772                     argument++;
 773                     g_nbIterations = readU32FromChar(&argument);
 774                     break;
 775
 776                     /* Select compressibility of synthetic sample */
 777                 case 'P':
 778                     argument++;
 779                     g_compressibility = (double)readU32FromChar(&argument) / 100.;
 780                     break;
 781                 case 'l':
 782                     argument++;
 783                     cLevel = (int)readU32FromChar(&argument);
 784                     cparams = ZSTD_getCParams(cLevel, 0, 0);
 785                     break;
 786
 787                     /* Unknown command */
 788                 default : return badusage(exename);
 789                 }
 790             }
 791             continue;
 792         }
 793
 794         /* first provided filename is input */
 795         if (!input_filename) { input_filename=argument; filenamesStart=argNb; continue; }
 796     }
 797
 798
 799
 800     if (filenamesStart==0)   /* no input file */
 801         result = benchSample(benchNb, cLevel, cparams);
 802     else
 803         result = benchFiles(benchNb, argv+filenamesStart, argc-filenamesStart, cLevel, cparams);
 804
 805     if (main_pause) { int unused; printf("press enter...\n"); unused = getchar(); (void)unused; }
 806
 807     return result;
 808 }