core_cparfb.c

Go to the documentation of this file.

#include <cblas.h>
#include <lapacke.h>
#include "common.h"

/***************************************************************************/
int 
CORE_cparfb(int side, int trans, int direct, int storev,
            int M1, int N1, int M2, int N2, int K, int L,
            PLASMA_Complex32_t *A1, int LDA1,
            PLASMA_Complex32_t *A2, int LDA2,
            PLASMA_Complex32_t *V, int LDV,
            PLASMA_Complex32_t *T, int LDT,
            PLASMA_Complex32_t *WORK, int LDWORK)
{
    static PLASMA_Complex32_t zone  =  1.0;
    static PLASMA_Complex32_t mzone = -1.0;

    int j;

    /* Check input arguments */
    if ((side != PlasmaLeft) && (side != PlasmaRight)) {
        coreblas_error(1, "Illegal value of side");
        return -1;
    }
    if ((trans != PlasmaNoTrans) && (trans != PlasmaConjTrans)) {
        coreblas_error(2, "Illegal value of trans");
        return -2;
    }
    if ((direct != PlasmaForward) && (direct != PlasmaBackward)) {
        coreblas_error(3, "Illegal value of direct");
        return -3;
    }
    if ((storev != PlasmaColumnwise) && (storev != PlasmaRowwise)) {
        coreblas_error(4, "Illegal value of storev");
        return -4;
    }
    if (M1 < 0) {
        coreblas_error(5, "Illegal value of M1");
        return -5;
    }
    if (N1 < 0) {
        coreblas_error(6, "Illegal value of N1");
        return -6;
    }
    if ((M2 < 0) ||
        ( (side == PlasmaRight) && (M1 != M2) ) ) {
        coreblas_error(7, "Illegal value of M2");
        return -7;
    }
    if ((N2 < 0) ||
        ( (side == PlasmaLeft) && (N1 != N2) ) ) {
        coreblas_error(8, "Illegal value of N2");
        return -8;
    }
    if (K < 0) {
        coreblas_error(9, "Illegal value of K");
        return -9;
    }

    /* Quick return */
    if ((M1 == 0) || (N1 == 0) || (M2 == 0) || (N2 == 0) || (K == 0))
        return PLASMA_SUCCESS;

    if (direct == PlasmaForward) {

        if (side == PlasmaLeft) {

            /*
             * Column or Rowwise / Forward / Left
             * ----------------------------------
             *
             * Form  H * A  or  H' * A  where  A = ( A1 )
             *                                     ( A2 )
             */

            /* W = A1 + op(V) * A2 */
            CORE_cpamm(
                    PlasmaW, PlasmaLeft, storev,
                    K, N1, M2, L,
                    A1, LDA1,
                    A2, LDA2,
                    V, LDV,
                    WORK, LDWORK);

            /* W = op(T) * W */
            cblas_ctrmm(
                CblasColMajor, CblasLeft, CblasUpper,
                (CBLAS_TRANSPOSE)trans, CblasNonUnit, K, N2,
                CBLAS_SADDR(zone), T, LDT, WORK, LDWORK);

            /* A1 = A1 - W */
            for(j = 0; j < N1; j++) {
                cblas_caxpy(
                        K, CBLAS_SADDR(mzone),
                        &WORK[LDWORK*j], 1,
                        &A1[LDA1*j], 1);
            }

            /* A2 = A2 - op(V) * W  */
            /* W also changes: W = V * W, A2 = A2 - W */
            CORE_cpamm(
                    PlasmaA2, PlasmaLeft, storev,
                    M2, N2, K, L,
                    A1, LDA1,
                    A2, LDA2,
                    V, LDV,
                    WORK, LDWORK);
        }
        else {
            /* 
             * Column or Rowwise / Forward / Right
             * -----------------------------------
             *
             * Form  H * A  or  H' * A  where A  = ( A1 A2 )
             *
             */

            /* W = A1 + A2 * op(V) */
            CORE_cpamm(
                    PlasmaW, PlasmaRight, storev,
                    M1, K, N2, L,
                    A1, LDA1,
                    A2, LDA2,
                    V, LDV,
                    WORK, LDWORK);

            /* W = W * op(T) */
            cblas_ctrmm(
                CblasColMajor, CblasRight, CblasUpper,
                (CBLAS_TRANSPOSE)trans, CblasNonUnit, M2, K,
                CBLAS_SADDR(zone), T, LDT, WORK, LDWORK);

            /* A1 = A1 - W */
            for(j = 0; j < K; j++) {
                cblas_caxpy(
                        M1, CBLAS_SADDR(mzone),
                        &WORK[LDWORK*j], 1,
                        &A1[LDA1*j], 1);
            }

            /* A2 = A2 - W * op(V) */
            /* W also changes: W = W * V', A2 = A2 - W */
            CORE_cpamm(
                    PlasmaA2, PlasmaRight, storev,
                    M2, N2, K, L,
                    A1, LDA1,
                    A2, LDA2,
                    V, LDV,
                    WORK, LDWORK);
        }
    }
    else {
        coreblas_error(3, "Not implemented (Backward / Left or Right)");
        return PLASMA_ERR_NOT_SUPPORTED;
    }

    return PLASMA_SUCCESS;
}