starpu-max/tests/perfmodels/max_riffa.c

#include <starpu.h>
#include <stdlib.h>
#include <stdio.h>
#include <fpga.h>
#include <starpu_scheduler.h>
#include "../helper.h"


void cpu_func(void *buffers[], void *cl_arg)
{
    (void)buffers;
    (void)cl_arg;

    printf("Hello world\n");
}

void fpga_mult(void *d[])
{
    /* Ask Fpga for a channel, or
    * equivalently for a hardware task
    */
    int chnl = fpga_reserve_a_chanel();

    /* Get inputs from STARPU */
    int* subA = STARPU_MATRIX_GET_PTR(d[0]);
    int* subB = STARPU_MATRIX_GET_PTR(d[1]);
    int* subC = STARPU_MATRIX_GET_PTR(d[2]);

    /* Get info on which part of the
    * inputs the task must operate
    */
    uint32_t nyA= STARPU_MATRIX_GET_NY(d[0]);
    uint32_t ldA= STARPU_MATRIX_GET_LD(d[0]);

    uint32_t nyB= STARPU_MATRIX_GET_NY(d[1]);
    uint32_t ldB= STARPU_MATRIX_GET_LD(d[1]);

    uint32_t nyC= STARPU_MATRIX_GET_NY(d[2]);
    uint32_t ldC= STARPU_MATRIX_GET_LD(d[2]);
    uint32_t nxC= STARPU_MATRIX_GET_NX(d[2]);


    /* Send A and B */
    int buf_s[nyA], buf_r[nxC*nyC];

    fpga_trans sent, recv;

    for (uint32_t j = 0; j < nxC; j++)
    {
        for (uint32_t k = 0; k < nyA; k++)

            buf_s[k] = subA[j+k*ldA];
        fpga_data_send(chnl, buf_s, nyA);
    }
    for (uint32_t i = 0; i < nyC; i++)
    {
        for (uint32_t k = 0; k < nyA; k++)
            buf_s[k] = subB[k+i*ldB];
        fpga_data_send(chnl, buf_s, nyA);
    }

    /* Receive C. This is blocking */
    fpga_data_recv(chnl, buf_r, nxC*nyC);
    for (uint32_t i = 0; i < nxC; i++)
    {
        for (uint32_t j = 0; j < nyC; j++)
            subC[j + i*ldC] = buf_r[i*nyC+j];
    }
    fpga_release_chanel(chnl);
}

static struct starpu_codelet cl =
{
    .cpu_funcs = {cpu_func},
    .cpu_funcs_name = {"cpu_func"},
    //.fpga_funcs = {fpga_mult},
    .fpga_funcs = {fpga_mult},
    .fpga_funcs_name={"fpga_mult"},
    .nbuffers = 3,
    .modes = {STARPU_R, STARPU_R, STARPU_W}
};


int main(int argc, char **argv)
{

    starpu_profiling_status_set(1);

    struct starpu_conf conf;
    starpu_data_handle_t A_handle, B_handle, C_handle;
    int ret;

    starpu_conf_init(&conf);

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


    /* initialize StarPU */
    starpu_init(NULL);

    for (uint32_t x = 0; x < 9; x++)
    {
        for (uint32_t y = 0; y < 9; y++)
        {
            struct starpu_task *task = starpu_task_create();
            task->cl = &cl; /* Pointer to the codelet defined above */
            /* Get handlers for each block */
            task->handles[0] = starpu_data_get_sub_data( A_handle, 1, y);
            task->handles[1] = starpu_data_get_sub_data( B_handle, 1, x);
            task->handles[2] = starpu_data_get_sub_data( C_handle, 2, x, y);
            /* submit the task to StarPU */
            starpu_task_submit(task);
        }
    }

    starpu_data_unregister(A_handle);
    starpu_data_unregister(B_handle);
    starpu_data_unregister(C_handle);

    starpu_task_wait_for_all();


    /* terminate StarPU */
    starpu_shutdown();
    return 0;
}