int main(int argc, char** argv)
{
if (argc < 2) {
fprintf(stderr, "Usage: %s n", argv[0]);
return 1;
}
const uint32_t n = atol(argv[1]);
float *buf;
posix_memalign((void**) &buf, 16, sizeof(float)*n);
printf("Initialize a random buffer of %u floats...\n", n);
srand(time(NULL));
for (uint32_t i = 0; i < n; i++) {
buf[i] = (float) rand();
}
printf("Done!\n");
{
float min, max;
uint32_t min_idx, max_idx;
BENCH_START(org);
minmax(n, buf, &min_idx, &max_idx, &min, &max);
BENCH_END(org, "org", sizeof(float), n, 1, 1);
printf("Min (idx): %0.4f (%u)\n", min, min_idx);
printf("Max (idx): %0.4f (%u)\n", max, max_idx);
}
{
float min, max;
uint32_t min_idx, max_idx;
BENCH_START(sse);
minmax_vec(n, buf, &min_idx, &max_idx, &min, &max);
BENCH_END(sse, "sse", sizeof(float), n, 1, 1);
printf("Min (idx): %0.4f (%u)\n", min, min_idx);
printf("Max (idx): %0.4f (%u)\n", max, max_idx);
}
{
float min, max;
uint32_t min_idx, max_idx;
BENCH_START(sse);
minmax_vec2(n, buf, &min_idx, &max_idx, &min, &max);
BENCH_END(sse, "sse2", sizeof(float), n, 1, 1);
printf("Min (idx): %0.4f (%u)\n", min, min_idx);
printf("Max (idx): %0.4f (%u)\n", max, max_idx);
}
return 0;
}
uint32_t* compute(float const* pts, const size_t n, const uint32_t size_interval_u32, const float size_interval)
{
std::vector<uint32_t*> hist_locals;
hist_locals.resize(omp_get_num_threads(), NULL);
BENCH_START(b);
#pragma omp parallel num_threads(4)
{
uint32_t* hist_local;
posix_memalign((void**) &hist_local, 16, sizeof(uint32_t)*size_interval_u32);
memset(hist_local, 0, sizeof(uint32_t)*size_interval_u32);
hist_locals[omp_get_thread_num()] = hist_local;
#pragma omp for
for (size_t i = 0; i < n; i++) {
const size_t idx = floor(pts[i]/size_interval);
hist_local[idx]++;
}
}
std::cerr << "Final reduction..." << std::endl;
uint32_t* final_hist = NULL;
// Find the first non-null partial histogram
std::vector<uint32_t*>::const_iterator it;
for (it = hist_locals.begin(); it != hist_locals.end(); it++) {
if (*it != NULL) {
final_hist = *it;
it++;
break;
}
}
if (final_hist == NULL) {
return NULL;
}
// Final reduction
for (; it != hist_locals.end(); it++) {
uint32_t* hist_local = *it;
if (hist_local == NULL) {
continue;
}
for (uint32_t i = 0; i < size_interval_u32; i++) {
final_hist[i] += hist_local[i];
}
}
BENCH_END(b, "compute_omp", sizeof(float), n, sizeof(uint32_t), n);
return final_hist;
}
