pdpotrf_diag.c

Go to the documentation of this file.

#include "../include/diag.h"

#define A(m,n) A,  m,  n
/***************************************************************************/
void morse_pdpotrf_diag(MORSE_enum uplo, MORSE_desc_t *A, int diag_thick,
                   MORSE_sequence_t *sequence, MORSE_request_t *request)
{
    MORSE_context_t *morse;
    MORSE_option_t options;

    int k, m, n;
    int ldak, ldam, ldan;
    int tempkm, tempmm, tempnn;
    size_t ws_host   = 0;

    double zone  = (double) 1.0;
    double mzone = (double)-1.0;

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

    RUNTIME_options_ws_alloc( &options, 0, ws_host );

    /*
     *  MorseLower
     */
    if (uplo == MorseLower) {
        for (k = 0; k < A->mt; k++) {
            RUNTIME_iteration_push(morse, k);

            tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
            ldak = BLKLDD(A, k);

            options.priority = 2*A->mt - 2*k;
            MORSE_TASK_dpotrf(
                &options,
                MorseLower, tempkm, A->mb,
                A(k, k), ldak, A->nb*k);

            for (m = k+1; m < A->mt && m < k+diag_thick; m++) {
                tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
                ldam = BLKLDD(A, m);

                options.priority = 2*A->mt - 2*k - m;
                MORSE_TASK_dtrsm(
                    &options,
                    MorseRight, MorseLower, MorseTrans, MorseNonUnit,
                    tempmm, A->mb, A->mb,
                    zone, A(k, k), ldak,
                          A(m, k), ldam);
            }
            RUNTIME_data_flush( sequence, A(k, k) );

            for (n = k+1; n < A->nt && n < k+diag_thick; n++) {
                tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
                ldan = BLKLDD(A, n);

                options.priority = 2*A->mt - 2*k - n;
                MORSE_TASK_dsyrk(
                    &options,
                    MorseLower, MorseNoTrans,
                    tempnn, A->nb, A->mb,
                    -1.0, A(n, k), ldan,
                     1.0, A(n, n), ldan);

                for (m = n+1; m < A->mt && m < n+diag_thick; m++) {
                    tempmm = m == A->mt-1 ? A->m - m*A->mb : A->mb;
                    ldam = BLKLDD(A, m);

                    options.priority = 2*A->mt - 2*k - n - m;
                    MORSE_TASK_dgemm(
                        &options,
                        MorseNoTrans, MorseTrans,
                        tempmm, tempnn, A->mb, A->mb,
                        mzone, A(m, k), ldam,
                               A(n, k), ldan,
                        zone,  A(m, n), ldam);
                }
                RUNTIME_data_flush( sequence, A(n, k) );
            }
            RUNTIME_iteration_pop(morse);
        }
    }
    /*
     *  MorseUpper
     */
    else {
        for (k = 0; k < A->nt; k++) {
            RUNTIME_iteration_push(morse, k);

            tempkm = k == A->nt-1 ? A->n-k*A->nb : A->nb;
            ldak = BLKLDD(A, k);

            options.priority = 2*A->nt - 2*k;
            MORSE_TASK_dpotrf(
                &options,
                MorseUpper,
                tempkm, A->mb,
                A(k, k), ldak, A->nb*k);

            for (n = k+1; n < A->nt; n++) {
                tempnn = n == A->nt-1 ? A->n - n*A->nb : A->nb;

                options.priority = 2*A->nt - 2*k - n;
                MORSE_TASK_dtrsm(
                    &options,
                    MorseLeft, MorseUpper, MorseTrans, MorseNonUnit,
                    A->mb, tempnn, A->mb,
                    zone, A(k, k), ldak,
                          A(k, n), ldak);
            }
            RUNTIME_data_flush( sequence, A(k, k) );

            for (m = k+1; m < A->mt; m++) {
                tempmm = m == A->mt-1 ? A->m - m*A->mb : A->mb;
                ldam = BLKLDD(A, m);

                options.priority = 2*A->nt - 2*k  - m;
                MORSE_TASK_dsyrk(
                    &options,
                    MorseUpper, MorseTrans,
                    tempmm, A->mb, A->mb,
                    -1.0, A(k, m), ldak,
                     1.0, A(m, m), ldam);

                for (n = m+1; n < A->nt; n++) {
                    tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;

                    options.priority = 2*A->nt - 2*k - n - m;
                    MORSE_TASK_dgemm(
                        &options,
                        MorseTrans, MorseNoTrans,
                        tempmm, tempnn, A->mb, A->mb,
                        mzone, A(k, m), ldak,
                               A(k, n), ldak,
                        zone,  A(m, n), ldam);
                }
                RUNTIME_data_flush( sequence, A(k, m) );
            }

            RUNTIME_iteration_pop(morse);
        }
    }

    RUNTIME_options_ws_free(&options);
    RUNTIME_options_finalize(&options, morse);
}