double
max_parallelism(benchmp_f* benchmarks,
int warmup, int repetitions, void* cookie)
{
int i, j, k;
double baseline, max_load_parallelism, load_parallelism;
result_t *results, *r_save;
max_load_parallelism = 1.;
for (i = 0; i < MAX_LOAD_PARALLELISM; ++i) {
benchmp(initialize, benchmarks[i], cleanup,
0, 1, warmup, repetitions, cookie);
save_minimum();
if (gettime() == 0)
return -1.;
if (i == 0) {
baseline = (double)gettime() / (double)get_n();
} else {
load_parallelism = baseline;
load_parallelism /= (double)gettime();
load_parallelism *= (double)((i + 1) * get_n());
if (load_parallelism > max_load_parallelism) {
max_load_parallelism = load_parallelism;
}
}
}
return max_load_parallelism;
}
void
loads(size_t len, size_t range, size_t stride,
int parallel, int warmup, int repetitions)
{
double result;
size_t count;
struct mem_state state;
if (range < stride) return;
state.width = 1;
state.len = range;
state.maxlen = len;
state.line = stride;
state.pagesize = getpagesize();
count = 100 * (state.len / (state.line * 100) + 1);
#if 0
(*fpInit)(0, &state);
fprintf(stderr, "loads: after init\n");
(*benchmark_loads)(2, &state);
fprintf(stderr, "loads: after benchmark\n");
mem_cleanup(0, &state);
fprintf(stderr, "loads: after cleanup\n");
settime(1);
save_n(1);
#else
/*
* Now walk them and time it.
*/
benchmp(fpInit, benchmark_loads, mem_cleanup,
100000, parallel, warmup, repetitions, &state);
#endif
/* We want to get to nanoseconds / load. */
save_minimum();
result = (1000. * (double)gettime()) / (double)(count * get_n());
fprintf(stderr, "%.5f %.3f\n", range / (1024. * 1024.), result);
}
