118 lines
2.7 KiB
C
118 lines
2.7 KiB
C
/* Copyright (c) 2025 Barcelona Supercomputing Center (BSC)
|
|
* SPDX-License-Identifier: GPL-3.0-or-later */
|
|
|
|
/*
|
|
* This test creates several tasks that all perform the computation with the
|
|
* same instructions. However, the access to the memory is done differently. The
|
|
* first set of tasks use a stride 1, the next 2, the next 4 and so on until
|
|
* 2^(NSTRIDE-1). This access causes more L3 cache misses, which increases the
|
|
* execution time, typically directly proportional to the stride number.
|
|
*
|
|
* The number of instructions given by PAPI_TOT_INS should remain constant
|
|
* across tasks, but it is expected that PAPI_L3_TCM increases with the
|
|
* stride.
|
|
*/
|
|
|
|
#include <math.h>
|
|
#include <nosv.h>
|
|
#include <stdatomic.h>
|
|
#include <stdlib.h>
|
|
#include <unistd.h>
|
|
|
|
#include "common.h"
|
|
#include "compat.h"
|
|
|
|
#define NTASKS 200
|
|
atomic_int ncompleted = 0;
|
|
|
|
nosv_task_t tasks[NTASKS];
|
|
|
|
#define NRUNS 2
|
|
#define NSTRIDE 4
|
|
#define MAXN (256L * 1024L) /* Adjust this for larger L3 */
|
|
|
|
struct meta {
|
|
long n;
|
|
long stride;
|
|
double *vec;
|
|
};
|
|
|
|
static double
|
|
get_time(void)
|
|
{
|
|
struct timespec ts;
|
|
clock_gettime(CLOCK_MONOTONIC, &ts);
|
|
return (double) ts.tv_sec + (double) ts.tv_nsec * 1.0e-9;
|
|
}
|
|
|
|
static void
|
|
task_body(nosv_task_t task)
|
|
{
|
|
struct meta *meta = nosv_get_task_metadata(task);
|
|
|
|
long stride = meta->stride;
|
|
|
|
/* Stride access, some computation */
|
|
for (long i = 0; i < stride; i++)
|
|
for (long j = i; j < meta->n; j += stride)
|
|
meta->vec[j] = sqrt(meta->vec[j]);
|
|
|
|
atomic_fetch_add(&ncompleted, 1);
|
|
}
|
|
|
|
int
|
|
main(void)
|
|
{
|
|
nosv_init();
|
|
|
|
nosv_task_type_t task_type;
|
|
nosv_type_init(&task_type, task_body, NULL, NULL, "task", NULL, NULL, 0);
|
|
|
|
for (int i = 0; i < NTASKS; i++) {
|
|
nosv_create(&tasks[i], task_type, sizeof(struct meta), 0);
|
|
struct meta *meta = nosv_get_task_metadata(tasks[i]);
|
|
meta->n = MAXN;
|
|
meta->vec = calloc(MAXN, sizeof(double));
|
|
for (long i = 0; i < MAXN; i++)
|
|
meta->vec[i] = (double) i;
|
|
}
|
|
|
|
fprintf(stderr, "%8s %8s %8s\n", "run", "stride", "time");
|
|
|
|
/* Repeat for warmup */
|
|
for (int run = 0; run < NRUNS; run++) {
|
|
for (int s = 0; s < NSTRIDE; s++) {
|
|
long stride = 1L << s;
|
|
|
|
atomic_store(&ncompleted, 0); /* reset */
|
|
|
|
double t0 = get_time();
|
|
|
|
for (int i = 0; i < NTASKS; i++) {
|
|
struct meta *meta = nosv_get_task_metadata(tasks[i]);
|
|
meta->stride = stride;
|
|
nosv_submit(tasks[i], 0);
|
|
}
|
|
|
|
while (atomic_load(&ncompleted) != NTASKS)
|
|
sleep_us(1000);
|
|
|
|
double t1 = get_time();
|
|
|
|
printf("%8d %8ld %8.3f\n", run, stride, t1 - t0);
|
|
}
|
|
}
|
|
|
|
for (int i = 0; i < NTASKS; i++) {
|
|
struct meta *meta = nosv_get_task_metadata(tasks[i]);
|
|
free(meta->vec);
|
|
nosv_destroy(tasks[i], 0);
|
|
}
|
|
|
|
nosv_type_destroy(task_type, 0);
|
|
|
|
nosv_shutdown();
|
|
|
|
return 0;
|
|
}
|