view release on metacpan or  search on metacpan

ext/xxHash/cli/xsum_bench.c  view on Meta::CPAN

    { "XXH128_stream w/seed",&localXXH128_stream_seeded },
};
#define NB_HASHFUNC (sizeof(g_hashesToBench) / sizeof(*g_hashesToBench))

#define NB_TESTFUNC (1 + 2 * NB_HASHFUNC)
int const g_nbTestFunctions = NB_TESTFUNC;
char g_testIDs[NB_TESTFUNC] = { 0 };
const char k_testIDs_default[NB_TESTFUNC] = { 0,
        1 /*XXH32*/, 0,
        1 /*XXH64*/, 0,
        1 /*XXH3*/, 0, 0, 0, 0, 0,
        1 /*XXH128*/ };

int g_nbIterations = NBLOOPS_DEFAULT;
#define HASHNAME_MAX 29
static void XSUM_benchHash(hashFunction h, const char* hName, int testID,
                           const void* buffer, size_t bufferSize)
{
    XSUM_U32 nbh_perIteration = (XSUM_U32)((XXH_1ST_SPEED_TARGET MB) / (bufferSize+1)) + 1;
    int iterationNb, nbIterations = g_nbIterations + !g_nbIterations /* min 1 */;
    double fastestH = 100000000.;
    assert(HASHNAME_MAX > 2);
    XSUM_logVerbose(2, "\r%80s\r", "");       /* Clean display line */

    for (iterationNb = 1; iterationNb <= nbIterations; iterationNb++) {
        XSUM_U32 r=0;
        clock_t cStart;

        XSUM_logVerbose(2, "%2i-%-*.*s : %10u ->\r",
                        iterationNb,
                        HASHNAME_MAX, HASHNAME_MAX, hName,
                        (unsigned)bufferSize);
        cStart = clock();
        while (clock() == cStart);   /* starts clock() at its exact beginning */
        cStart = clock();

        {   XSUM_U32 u;
            for (u=0; u<nbh_perIteration; u++)
                r += h(buffer, bufferSize, u);
        }
        if (r==0) XSUM_logVerbose(3,".\r");  /* do something with r to defeat compiler "optimizing" hash away */

        {   clock_t const nbTicks = XSUM_clockSpan(cStart);
            double const ticksPerHash = ((double)nbTicks / TIMELOOP) / nbh_perIteration;
            /*
             * clock() is the only decent portable timer, but it isn't very
             * precise.
             *
             * Sometimes, this lack of precision is enough that the benchmark
             * finishes before there are enough ticks to get a meaningful result.
             *
             * For example, on a Core 2 Duo (without any sort of Turbo Boost),
             * the imprecise timer caused peculiar results like so:
             *
             *    XXH3_64b                   4800.0 MB/s // conveniently even
             *    XXH3_64b unaligned         4800.0 MB/s
             *    XXH3_64b seeded            9600.0 MB/s // magical 2x speedup?!
             *    XXH3_64b seeded unaligned  4800.0 MB/s
             *
             * If we sense a suspiciously low number of ticks, we increase the
             * iterations until we can get something meaningful.
             */
            if (nbTicks < TIMELOOP_MIN) {
                /* Not enough time spent in benchmarking, risk of rounding bias */
                if (nbTicks == 0) { /* faster than resolution timer */
                    nbh_perIteration *= 100;
                } else {
                    /*
                     * update nbh_perIteration so that the next round lasts
                     * approximately 1 second.
                     */
                    double nbh_perSecond = (1 / ticksPerHash) + 1;
                    if (nbh_perSecond > (double)(4000U<<20)) nbh_perSecond = (double)(4000U<<20);   /* avoid overflow */
                    nbh_perIteration = (XSUM_U32)nbh_perSecond;
                }
                /* g_nbIterations==0 => quick evaluation, no claim of accuracy */
                if (g_nbIterations>0) {
                    iterationNb--;   /* new round for a more accurate speed evaluation */
                    continue;
                }
            }
            if (ticksPerHash < fastestH) fastestH = ticksPerHash;
            if (fastestH>0.) { /* avoid div by zero */
                XSUM_logVerbose(2, "%2i-%-*.*s : %10u -> %8.0f it/s (%7.1f MB/s) \r",
                            iterationNb,
                            HASHNAME_MAX, HASHNAME_MAX, hName,
                            (unsigned)bufferSize,
                            (double)1 / fastestH,
                            ((double)bufferSize / (1 MB)) / fastestH);
        }   }
        {   double nbh_perSecond = (1 / fastestH) + 1;
            if (nbh_perSecond > (double)(4000U<<20)) nbh_perSecond = (double)(4000U<<20);   /* avoid overflow */
            nbh_perIteration = (XSUM_U32)nbh_perSecond;
        }
    }
    XSUM_logVerbose(1, "%2i#%-*.*s : %10u -> %8.0f it/s (%7.1f MB/s) \n",
                    testID,
                    HASHNAME_MAX, HASHNAME_MAX, hName,
                    (unsigned)bufferSize,
                    (double)1 / fastestH,
                    ((double)bufferSize / (1 MB)) / fastestH);
    if (XSUM_logLevel<1)
        XSUM_logVerbose(0, "%u, ", (unsigned)((double)1 / fastestH));
}


/*
 * Allocates a string containing s1 and s2 concatenated. Acts like strdup.
 * The result must be freed.
 */
static char* XSUM_strcatDup(const char* s1, const char* s2)
{
    assert(s1 != NULL);
    assert(s2 != NULL);
    {   size_t len1 = strlen(s1);
        size_t len2 = strlen(s2);
        char* buf = (char*)malloc(len1 + len2 + 1);
        if (buf != NULL) {
            /* strcpy(buf, s1) */
            memcpy(buf, s1, len1);
            /* strcat(buf, s2) */