core_stslqt.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_stslqt = PCORE_stslqt
#define CORE_stslqt PCORE_stslqt
#define CORE_stsmlq PCORE_stsmlq
int  CORE_stsmlq(int side, int trans,
                 int M1, int N1, int M2, int N2, int K, int IB,
                 float *A1, int LDA1,
                 float *A2, int LDA2,
                 float *V, int LDV,
                 float *T, int LDT,
                 float *WORK, int LDWORK);
#endif
int CORE_stslqt(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;

    float alpha;
    int i, ii, sb;

    /* 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(8, "Illegal value of LDA2");
        return -8;
    }

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

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

            alpha = -(TAU[ii+i]);
            if (ii+i+1 < M) {
                /*
                 * Apply H( II+I-1 ) to A( II+I:II+IB-1, II+I-1:N  ) from the right.
                 */
                cblas_scopy(
                    sb-i-1,
                    &A1[LDA1*(ii+i)+(ii+i+1)], 1,
                    WORK, 1);

                cblas_sgemv(
                    CblasColMajor, (CBLAS_TRANSPOSE)PlasmaNoTrans,
                    sb-i-1, N,
                    (zone), &A2[ii+i+1], LDA2,
                    &A2[ii+i], LDA2,
                    (zone), WORK, 1);

                cblas_saxpy(
                    sb-i-1, (alpha),
                    WORK, 1,
                    &A1[LDA1*(ii+i)+ii+i+1], 1);

                cblas_sger(
                    CblasColMajor, sb-i-1, N,
                    (alpha), WORK, 1,
                    &A2[ii+i], LDA2,
                    &A2[ii+i+1], LDA2);
            }
            /*
             * Calculate T.
             */
            cblas_sgemv(
                CblasColMajor, (CBLAS_TRANSPOSE)PlasmaNoTrans, i, N,
                (alpha), &A2[ii], LDA2,
                &A2[ii+i], LDA2,
                (zzero), &T[LDT*(ii+i)], 1);
#ifdef COMPLEX
            LAPACKE_slacgv_work(N, &A2[ii+i], LDA2 );
            LAPACKE_slacgv_work(1, &A1[LDA1*(ii+i)+ii+i], LDA1 );
#endif
            cblas_strmv(
                CblasColMajor, (CBLAS_UPLO)PlasmaUpper,
                (CBLAS_TRANSPOSE)PlasmaNoTrans, (CBLAS_DIAG)PlasmaNonUnit, i,
                &T[LDT*ii], LDT,
                &T[LDT*(ii+i)], 1);

            T[LDT*(ii+i)+i] = TAU[ii+i];
        }
        if (M > ii+sb) {
            CORE_stsmlq(
                PlasmaRight, PlasmaTrans,
                M-(ii+sb), sb, M-(ii+sb), N, IB, IB,
                &A1[LDA1*ii+ii+sb], LDA1,
                &A2[ii+sb], LDA2,
                &A2[ii], LDA2,
                &T[LDT*ii], LDT,
                WORK, LDA1);
        }
    }
    return PLASMA_SUCCESS;
}

/***************************************************************************/
void QUARK_CORE_stslqt(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_TSLQT;
    QUARK_Insert_Task(quark, CORE_stslqt_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 | 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_stslqt_quark = PCORE_stslqt_quark
#define CORE_stslqt_quark PCORE_stslqt_quark
#endif
void CORE_stslqt_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_stslqt(m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
}