int main(void)
{
static const size_t default_initial_size = 655360;
struct p_bench_specification spec;
char *raw_mem = NULL;
spec.current_size = default_initial_size;
setup_memory(&spec.mem, &raw_mem, spec.current_size);
for (const struct p_bench_item *item = benchmark_items; item->name != NULL;
++item) {
struct item_data data;
item_preface(&data, item);
item->benchmark(&spec);
item_done(&data, &spec, item->name);
}
return EXIT_SUCCESS;
}
struct item_data item_bench(const struct p_bench_item *item,
struct p_bench_specification *spec)
{
struct item_data data, best = { .end = ~(0ULL) };
uint64_t item_start_time = platform_clock();
/* Warm up caches, branch predictors etc. */
/* Calculate inner loop. */
int inner_loop;
data.start = platform_clock();
for (int i = 0; i < 50; i++)
item->benchmark(spec);
data.end = platform_clock();
{
float tmp = data.end - data.start;
tmp /= 50.0f;
/* 50k us seems to work */
tmp = 50000.0f / tmp;
inner_loop = ceilf(tmp);
}
/* Repeat tests to get more stable results between runs */
while (true) {
/* Measure 10 iterations so the clocksource's resolution doesn't
* play tricks on us */
data.start = platform_clock();
for (int j = 0; j < inner_loop; j++)
item->benchmark(spec);
data.end = platform_clock();
/* Use best measurement */
if (best.end - best.start > data.end - data.start)
best = data;
/* Test each function for 1/2 second */
if (data.end - item_start_time >= 500000000ULL)
break;
}
{
/* Adjust for iterations in inner loop above */
float tmp = best.end - best.start;
tmp /= (float) inner_loop;
best.end = best.start + tmp;
}
return best;
}
int main(void)
{
struct p_bench_specification spec = { 0 };
char *raw_mem = NULL;
spec.current_size = MAX_ELEMS;
setup_memory(&spec.mem, &raw_mem, spec.current_size);
bench_printf(";name, size, duration (ns)\n");
for (const struct p_bench_item *item = benchmark_items; item->name != NULL;
++item) {
struct item_data best;
bool consistent = false;
best = item_bench(item, &spec);
for (int tries = 0; tries < 50; tries++) {
struct item_data snd;
float fst_time, snd_time;
/* Benchmark again ... */
snd = item_bench(item, &spec);
fst_time = best.end - best.start;
snd_time = snd.end - snd.start;
/* ... and start over if results deviate too much */
if (fst_time / snd_time < 0.995 || snd_time / fst_time < 0.995) {
/* Take average so abnormally low results converge over time */
best.end += (snd_time - fst_time) / 2.0f;
usleep(100000);
continue;
}
if (fst_time > snd_time)
best = snd;
consistent = true;
break;
}
if (!consistent) {
fprintf(stderr, ";WARNING: %s not consistent\n", item->name);
fflush(stderr);
}
item_done(&best, &spec, item->name);
}
return EXIT_SUCCESS;
}
#else /* __epiphany__ */
int main(void)
{
struct p_bench_specification spec = { 0 };
char *raw_mem = NULL;
spec.current_size = MAX_ELEMS;
uint32_t nbench = 0;
setup_memory(&spec.mem, &raw_mem, spec.current_size);
bench_printf(";name, size, duration (ns)\n");
for (const struct p_bench_item *item = benchmark_items; item->name != NULL;
++item) {
struct item_data data;
data.start = platform_clock();
item->benchmark(&spec);
data.end = platform_clock();
strcpy(epiphany_results[nbench].name, item->name);
epiphany_results[nbench].ns = data.end - data.start;
epiphany_results[nbench].size = (uint64_t) spec.current_size;
nbench++;
epiphany_status->nbench = nbench;
}
epiphany_status->done = 1;
return EXIT_SUCCESS;
}
#endif
static void setup_output_pointers(struct p_bench_raw_memory *mem, void *p)
{
/* Assume largest type is 64 bits */
/* TODO: All pointers point to same memory region so output will be bogus */
mem->o1.p_u64 = p;
mem->o2.p_u64 = p;
mem->o3.p_u64 = p;
mem->o4.p_u64 = p;
}
