starpu-max/tests/perfmodels/regression_based_01.c

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2012,2016                                Inria
 * Copyright (C) 2010-2015,2017                           Université de Bordeaux
 * Copyright (C) 2010-2013,2015,2017                      CNRS
 *
 * StarPU is free software; you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published by
 * the Free Software Foundation; either version 2.1 of the License, or (at
 * your option) any later version.
 *
 * StarPU is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * See the GNU Lesser General Public License in COPYING.LGPL for more details.
 */

/*
 * This examplifies how to get task execution profiling from the application.
 */

#include <starpu.h>
#include <assert.h>
#include <starpu_scheduler.h>
#include <unistd.h>
#include "../helper.h"

#define FPRINTF(ofile, fmt, ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ## __VA_ARGS__); }} while(0)

#define START 1024
#ifdef STARPU_QUICK_CHECK
#define END 1048576
#else
#define END 16777216
#endif

//static unsigned niter = 500;

int ret;


#ifdef STARPU_USE_CUDA
static void memset_cuda(void *descr[], void *arg)
{
    (void)arg;
    STARPU_SKIP_IF_VALGRIND;

    int *ptr = (int *)STARPU_VECTOR_GET_PTR(descr[0]);
    unsigned n = STARPU_VECTOR_GET_NX(descr[0]);

    cudaMemsetAsync(ptr, 42, n * sizeof(*ptr), starpu_cuda_get_local_stream());
}
#endif

#ifdef STARPU_USE_OPENCL
extern void memset_opencl(void *buffers[], void *args);
#endif


void memset_cpu(void *descr[], void *arg)
{
    (void)arg;
    STARPU_SKIP_IF_VALGRIND;

    int *ptr = (int *)STARPU_VECTOR_GET_PTR(descr[0]);
    unsigned n = STARPU_VECTOR_GET_NX(descr[0]);

    usleep(1000);

    for (int i=0; i<n ; i++)
    {

        ptr[0] += i;
    }
}


static struct starpu_perfmodel model =
{
    .type = STARPU_REGRESSION_BASED,
    .symbol = "memset_regression_based"
};

static struct starpu_perfmodel nl_model =
{
    .type = STARPU_NL_REGRESSION_BASED,
    .symbol = "non_linear_memset_regression_based"
};

static struct starpu_codelet memset_cl =
{
#ifdef STARPU_USE_CUDA
    .cuda_funcs = {memset_cuda},
    .cuda_flags = {STARPU_CUDA_ASYNC},
#endif
#ifdef STARPU_USE_OPENCL
    .opencl_funcs = {memset_opencl},
    .opencl_flags = {STARPU_OPENCL_ASYNC},
#endif
    .cpu_funcs = {memset_cpu},
    .cpu_funcs_name = {"memset_cpu"},
    .model = &model,
    .nbuffers = 1,
    .modes = {STARPU_W}
};

static struct starpu_codelet nl_memset_cl =
{
#ifdef STARPU_USE_CUDA
    .cuda_funcs = {memset_cuda},
    .cuda_flags = {STARPU_CUDA_ASYNC},
#endif
#ifdef STARPU_USE_OPENCL
    .opencl_funcs = {memset_opencl},
    .opencl_flags = {STARPU_OPENCL_ASYNC},
#endif
    .cpu_funcs = {memset_cpu},
    .cpu_funcs_name = {"memset_cpu"},
    .model = &nl_model,
    .nbuffers = 1,
    .modes = {STARPU_W}
};


static void test_memset(int nelems, struct starpu_codelet *codelet)
{
    int nloops = 100;
    int loop;
    starpu_data_handle_t handle;

    starpu_vector_data_register(&handle, -1, (uintptr_t)NULL, nelems, sizeof(int));
    for (loop = 0; loop < nloops; loop++)
    {
        struct starpu_task *task = starpu_task_create();

        task->cl = codelet;
        task->handles[0] = handle;

        int ret = starpu_task_submit(task);
        if (ret == -ENODEV)
            exit(STARPU_TEST_SKIPPED);
        STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
    }

    starpu_data_unregister(handle);
}


static void compare_performance(int size, struct starpu_codelet *codelet, struct starpu_task *task)
{
    unsigned i;
    int niter = 100;
    starpu_data_handle_t handle;

    starpu_vector_data_register(&handle, -1, (uintptr_t)NULL, size, sizeof(int));

    struct starpu_task **tasks = (struct starpu_task **) malloc(niter*sizeof(struct starpu_task *));
    assert(tasks);

    for (i = 0; i < niter; i++)
    {
        struct starpu_task *task = starpu_task_create();

        task->cl = codelet;
        task->handles[0] = handle;

        task->synchronous = 1;

        /* We will destroy the task structure by hand so that we can
         * query the profiling info before the task is destroyed. */
        task->destroy = 0;

        tasks[i] = task;

        ret = starpu_task_submit(task);

        if (STARPU_UNLIKELY(ret == -ENODEV))
        {
            FPRINTF(stderr, "No worker may execute this task\n");
            exit(0);
        }
    }

    starpu_data_unregister(handle);

    starpu_task_wait_for_all();

    double length_sum = 0.0;

    for (i = 0; i < niter; i++)
    {
        struct starpu_task *task = tasks[i];
        struct starpu_profiling_task_info *info = task->profiling_info;


        /* How long was the task execution ? */
        length_sum += starpu_timing_timespec_delay_us(&info->start_time, &info->end_time);

        /* We don't need the task structure anymore */
        starpu_task_destroy(task);
    }


    /* Display the occupancy of all workers during the test */
    unsigned worker;
    for (worker = 0; worker < starpu_worker_get_count(); worker++)
    {
        struct starpu_profiling_worker_info worker_info;
        ret = starpu_profiling_worker_get_info(worker, &worker_info);
        STARPU_ASSERT(!ret);

        char workername[128];
        starpu_worker_get_name(worker, workername, sizeof(workername));
        unsigned nimpl;

        FPRINTF(stdout, "\n Worker :%s ::::::::::\n\n", workername);

        for (nimpl = 0; nimpl < STARPU_MAXIMPLEMENTATIONS; nimpl++)
        {

            FPRINTF(stdout, "Expected time for %d on %s (impl %u): %f, Measured time: %f\n",
                    size, workername, nimpl,starpu_task_expected_length(task, starpu_worker_get_perf_archtype(worker, task->sched_ctx), nimpl), ((length_sum)/niter));

        }

    }


}

#ifdef STARPU_USE_OPENCL
struct starpu_opencl_program opencl_program;
#endif
int main(int argc, char **argv)
{
    /* Enable profiling */
    starpu_profiling_status_set(STARPU_PROFILING_ENABLE);

    struct starpu_conf conf;
    starpu_data_handle_t handle;
    int ret;

    starpu_conf_init(&conf);

    conf.sched_policy_name = "eager";
    conf.calibrate = 2;

    ret = starpu_initialize(&conf, &argc, &argv);
    if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
    STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");

#ifdef STARPU_USE_OPENCL
    ret = starpu_opencl_load_opencl_from_file("tests/perfmodels/opencl_memset_kernel.cl",
            &opencl_program, NULL);
    STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");
#endif
    int size;
    for (size = START; size < END; size *= 2)
    {
        /* Use a linear regression */
        test_memset(size, &memset_cl);

        /* Use a non-linear regression */
        test_memset(size, &nl_memset_cl);
    }
    ret = starpu_task_wait_for_all();
    STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait_for_all");

    starpu_shutdown();


    /* Test Phase */
    starpu_conf_init(&conf);

    conf.sched_policy_name = "eager";
    conf.calibrate = 0;

    ret = starpu_initialize(&conf, &argc, &argv);
    if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
    STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");

    /* Now create a dummy task just to estimate its duration according to the regression */

    size = 12345;

    starpu_vector_data_register(&handle, -1, (uintptr_t)NULL, size, sizeof(int));

    struct starpu_task *task = starpu_task_create();
    task->cl = &memset_cl;
    task->handles[0] = handle;
    task->destroy = 0;

    FPRINTF(stdout, "\n ////linear regression results////\n");
    compare_performance(size, &memset_cl,task);


    task->cl = &nl_memset_cl;

    FPRINTF(stdout, "\n ////non linear regression results////\n");

    compare_performance(size, &nl_memset_cl,task);


    starpu_task_destroy(task);

    starpu_data_unregister(handle);

#ifdef STARPU_USE_OPENCL
    ret = starpu_opencl_unload_opencl(&opencl_program);
    STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_unload_opencl");
#endif
    starpu_shutdown();

    return 0;
}