codelet_dgemv.c

Go to the documentation of this file.

#include "../include/starpu_exageostat.h"

static void CORE_dgemv_starpu(void *buffers[], void *cl_arg){
        int m, n, ldam;
        int m0, n0;
        double * A;
        double * Z;
        double * Zout;

        A   = (double *)STARPU_MATRIX_GET_PTR(buffers[0]);
        Z   = (double *)STARPU_MATRIX_GET_PTR(buffers[1]);
        Zout    = (double *)STARPU_MATRIX_GET_PTR(buffers[2]);
        starpu_codelet_unpack_args(cl_arg, &m, &n, &ldam, &m0, &n0);

        cblas_dgemv(CblasColMajor , CblasNoTrans , m, n, 1 , A, m, Z, 1, 0, Zout, 1);

}

static struct starpu_codelet cl_dgemv =
{
                .where       = STARPU_CPU,
                .cpu_funcs   = {CORE_dgemv_starpu},
                .nbuffers    = 3,
                .modes       = {STARPU_R,STARPU_R,STARPU_RW},
        .name        = "dgemv"
};

/***************************************************************************/
int MORSE_MLE_dgemv_Tile_Async(MORSE_desc_t *descA, MORSE_desc_t *descZ, MORSE_desc_t *descZout, MORSE_sequence_t *sequence, MORSE_request_t  *request) {

        MORSE_context_t *morse;
        MORSE_option_t options;
        morse = morse_context_self();

        if (sequence->status != MORSE_SUCCESS)
                return -2;
        RUNTIME_options_init(&options, morse, sequence, request);

        int m, n, ldam, m0, n0;
        int tempmm, tempnn;
        MORSE_desc_t A = *descA;

        struct starpu_codelet *cl=&cl_dgemv;
        for (n = 0; n < A.nt; n++) {
                 tempnn = n == A.nt -1 ? A.n - n * A.nb : A.nb;
                for(m = 0; m < A.mt; m++)
                {
                        tempmm = m == A.mt - 1 ? A.m - m* A.mb : A.mb;
                        ldam= mBLKLDD(descA, m);
                        m0= m * A.mb;
                        n0= n * A.nb;
                        starpu_insert_task(starpu_mpi_codelet(cl),
                                        STARPU_VALUE, &tempmm,  sizeof(int),
                                        STARPU_VALUE, &tempnn, sizeof(int),
                                        STARPU_VALUE, &ldam, sizeof(int),
                                        STARPU_VALUE, &m0,   sizeof(int),
                                        STARPU_VALUE, &n0,   sizeof(int),
                                        STARPU_R, RTBLKADDR(descA, sizeof(double)*A.mt, A.mt, n),
                                        STARPU_R, RTBLKADDR(descZ, sizeof(double)*tempnn, 0, n),
                                        STARPU_RW, RTBLKADDR(descZout, sizeof(double)*tempnn, 0, n),
                                        #if defined(CHAMELEON_CODELETS_HAVE_NAME)
                                        STARPU_NAME, "dgemv",
                                        #endif
                         0);
                }

        }

    //MORSE_TASK_flush_desc( &options, MorseUpperLower, descA);
        //MORSE_TASK_flush_desc( &options, MorseUpperLower, descZ);
        //MORSE_TASK_flush_desc( &options, MorseUpperLower, descZout);
        RUNTIME_options_ws_free(&options);
    RUNTIME_options_finalize(&options, morse);
        //MORSE_TASK_dataflush_all();
    MORSE_Sequence_Wait(sequence);
        return MORSE_SUCCESS;
}