core_sttlqt.c

Go to the documentation of this file.

#include <lapacke.h>
#include "common.h"
#undef COMPLEX
#define REAL

/***************************************************************************/
#if defined(PLASMA_HAVE_WEAK)
#pragma weak CORE_sttlqt = PCORE_sttlqt
#define CORE_sttlqt PCORE_sttlqt
#endif
int CORE_sttlqt(int M, int N, int IB,
                float *A1, int LDA1,
                float *A2, int LDA2,
                float *T, int LDT,
                float *TAU, float *WORK)
{
    static float zone  = 1.0;
    static float zzero = 0.0;
#ifdef COMPLEX
    static int                ione  = 1;
#endif

    float alpha;
    int i, j, l, ii, sb, mi, ni;

    /* Check input arguments */
    if (M < 0) {
        coreblas_error(1, "Illegal value of M");
        return -1;
    }
    if (N < 0) {
        coreblas_error(2, "Illegal value of N");
        return -2;
    }
    if (IB < 0) {
        coreblas_error(3, "Illegal value of IB");
        return -3;
    }
    if ((LDA2 < max(1,M)) && (M > 0)) {
        coreblas_error(7, "Illegal value of LDA2");
        return -7;
    }

    /* Quick return */
    if ((M == 0) || (N == 0) || (IB == 0))
        return PLASMA_SUCCESS;

    /* TODO: Need to check why some cases require 
     *  this to not have uninitialized values */
    CORE_slaset( PlasmaUpperLower, IB, N,
                 0., 0., T, LDT);

    for(ii = 0; ii < M; ii += IB) {
        sb = min(M-ii, IB);
        for(i = 0; i < sb; i++) {
            j  = ii + i;
            mi = sb-i-1;
            ni = min( j + 1, N);
            /*
             * Generate elementary reflector H( II*IB+I ) to annihilate A( II*IB+I, II*IB+I:M ).
             */
#ifdef COMPLEX
            LAPACKE_slacgv_work(ni, &A2[j], LDA2);
            LAPACKE_slacgv_work(ione, &A1[LDA1*j+j], LDA1);
#endif
            LAPACKE_slarfg_work(ni+1, &A1[LDA1*j+j], &A2[j], LDA2, &TAU[j]);

            if (mi > 0) {
                /*
                 * Apply H( j-1 ) to A( j:II+IB-1, j-1:M  ) from the right.
                 */
                cblas_scopy(
                    mi,
                    &A1[LDA1*j+(j+1)], 1,
                    WORK, 1);

                cblas_sgemv(
                    CblasColMajor, (CBLAS_TRANSPOSE)PlasmaNoTrans,
                    mi, ni,
                    (zone), &A2[j+1], LDA2,
                    &A2[j], LDA2,
                    (zone), WORK, 1);

                alpha = -(TAU[j]);
                cblas_saxpy(
                    mi, (alpha),
                    WORK, 1,
                    &A1[LDA1*j+j+1], 1);

                cblas_sger(
                    CblasColMajor, mi, ni,
                    (alpha), WORK, 1,
                    &A2[j], LDA2,
                    &A2[j+1], LDA2);
            }

            /*
             * Calculate T.
             */

            if (i > 0 ) {

                l = min(i, max(0, N-ii));
                alpha = -(TAU[j]);

                CORE_spemv(
                        PlasmaNoTrans, PlasmaRowwise,
                        i , min(j, N), l,
                        alpha, &A2[ii], LDA2,
                        &A2[j], LDA2,
                        zzero, &T[LDT*j], 1,
                        WORK);

                /* T(0:i-1, j) = T(0:i-1, ii:j-1) * T(0:i-1, j) */
                cblas_strmv(
                        CblasColMajor, (CBLAS_UPLO)PlasmaUpper,
                        (CBLAS_TRANSPOSE)PlasmaNoTrans,
                        (CBLAS_DIAG)PlasmaNonUnit,
                        i, &T[LDT*ii], LDT,
                        &T[LDT*j], 1);

            }

#ifdef COMPLEX
            LAPACKE_slacgv_work(ni, &A2[j], LDA2 );
            LAPACKE_slacgv_work(ione, &A1[LDA1*j+j], LDA1 );
#endif

            T[LDT*j+i] = TAU[j];
        }

        /* Apply Q to the rest of the matrix to the right */
        if (M > ii+sb) {
            mi = M-(ii+sb);
            ni = min(ii+sb, N);
            l  = min(sb, max(0, ni-ii));
            CORE_sparfb(
                PlasmaRight, PlasmaNoTrans,
                PlasmaForward, PlasmaRowwise,
                mi, IB, mi, ni, sb, l,
                &A1[LDA1*ii+ii+sb], LDA1,
                &A2[ii+sb], LDA2,
                &A2[ii], LDA2,
                &T[LDT*ii], LDT,
                WORK, M);

        }
    }
    return PLASMA_SUCCESS;
}

/***************************************************************************/
void QUARK_CORE_sttlqt(Quark *quark, Quark_Task_Flags *task_flags,
                       int m, int n, int ib, int nb,
                       float *A1, int lda1,
                       float *A2, int lda2,
                       float *T, int ldt)
{
    DAG_CORE_TTLQT;
    QUARK_Insert_Task(quark, CORE_sttlqt_quark, task_flags,
        sizeof(int),                        &m,     VALUE,
        sizeof(int),                        &n,     VALUE,
        sizeof(int),                        &ib,    VALUE,
        sizeof(float)*nb*nb,    A1,            INOUT|QUARK_REGION_D|QUARK_REGION_L,
        sizeof(int),                        &lda1,  VALUE,
        sizeof(float)*nb*nb,    A2,            INOUT|QUARK_REGION_D|QUARK_REGION_L|LOCALITY,
        sizeof(int),                        &lda2,  VALUE,
        sizeof(float)*ib*nb,    T,             OUTPUT,
        sizeof(int),                        &ldt,   VALUE,
        sizeof(float)*nb,       NULL,          SCRATCH,
        sizeof(float)*ib*nb,    NULL,          SCRATCH,
        0);
}

/***************************************************************************/
#if defined(PLASMA_HAVE_WEAK)
#pragma weak CORE_sttlqt_quark = PCORE_sttlqt_quark
#define CORE_sttlqt_quark PCORE_sttlqt_quark
#endif
void CORE_sttlqt_quark(Quark *quark)
{
    int m;
    int n;
    int ib;
    float *A1;
    int lda1;
    float *A2;
    int lda2;
    float *T;
    int ldt;
    float *TAU;
    float *WORK;

    quark_unpack_args_11(quark, m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
    CORE_sttlqt(m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
}