core_ctsqrt.c

Go to the documentation of this file.

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

/***************************************************************************/
#if defined(PLASMA_HAVE_WEAK)
#pragma weak CORE_ctsqrt = PCORE_ctsqrt
#define CORE_ctsqrt PCORE_ctsqrt
#define CORE_ctsmqr PCORE_ctsmqr
int  CORE_ctsmqr(int side, int trans,
                 int M1, int N1, int M2, int N2, int K, int IB,
                 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);
#endif

int CORE_ctsqrt(int M, int N, int IB,
                PLASMA_Complex32_t *A1, int LDA1,
                PLASMA_Complex32_t *A2, int LDA2,
                PLASMA_Complex32_t *T, int LDT,
                PLASMA_Complex32_t *TAU, PLASMA_Complex32_t *WORK)
{
    static PLASMA_Complex32_t zone  = 1.0;
    static PLASMA_Complex32_t zzero = 0.0;

    PLASMA_Complex32_t 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 < N; ii += IB) {
        sb = min(N-ii, IB);
        for(i = 0; i < sb; i++) {
            /*
             * Generate elementary reflector H( II*IB+I ) to annihilate
             * A( II*IB+I:M, II*IB+I )
             */
            LAPACKE_clarfg_work(M+1, &A1[LDA1*(ii+i)+ii+i], &A2[LDA2*(ii+i)], 1, &TAU[ii+i]);

            if (ii+i+1 < N) {
                /*
                 * Apply H( II*IB+I ) to A( II*IB+I:M, II*IB+I+1:II*IB+IB ) from the left
                 */
                alpha = -conjf(TAU[ii+i]);
                cblas_ccopy(
                    sb-i-1,
                    &A1[LDA1*(ii+i+1)+(ii+i)], LDA1,
                    WORK, 1);
#ifdef COMPLEX
                LAPACKE_clacgv_work(sb-i-1, WORK, 1);
#endif
                cblas_cgemv(
                    CblasColMajor, (CBLAS_TRANSPOSE)PlasmaConjTrans,
                    M, sb-i-1,
                    CBLAS_SADDR(zone), &A2[LDA2*(ii+i+1)], LDA2,
                    &A2[LDA2*(ii+i)], 1,
                    CBLAS_SADDR(zone), WORK, 1);
#ifdef COMPLEX
                LAPACKE_clacgv_work(sb-i-1, WORK, 1 );
#endif
                cblas_caxpy(
                    sb-i-1, CBLAS_SADDR(alpha),
                    WORK, 1,
                    &A1[LDA1*(ii+i+1)+ii+i], LDA1);
#ifdef COMPLEX
                LAPACKE_clacgv_work(sb-i-1, WORK, 1 );
#endif
                cblas_cgerc(
                    CblasColMajor, M, sb-i-1, CBLAS_SADDR(alpha),
                    &A2[LDA2*(ii+i)], 1,
                    WORK, 1,
                    &A2[LDA2*(ii+i+1)], LDA2);
            }
            /*
             * Calculate T
             */
            alpha = -TAU[ii+i];
            cblas_cgemv(
                CblasColMajor, (CBLAS_TRANSPOSE)PlasmaConjTrans, M, i,
                CBLAS_SADDR(alpha), &A2[LDA2*ii], LDA2,
                &A2[LDA2*(ii+i)], 1,
                CBLAS_SADDR(zzero), &T[LDT*(ii+i)], 1);

            cblas_ctrmv(
                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 (N > ii+sb) {
            CORE_ctsmqr(
                PlasmaLeft, PlasmaConjTrans,
                sb, N-(ii+sb), M, N-(ii+sb), IB, IB,
                &A1[LDA1*(ii+sb)+ii], LDA1,
                &A2[LDA2*(ii+sb)], LDA2,
                &A2[LDA2*ii], LDA2,
                &T[LDT*ii], LDT,
                WORK, sb);
        }
    }
    return PLASMA_SUCCESS;
}

/***************************************************************************/
void QUARK_CORE_ctsqrt(Quark *quark, Quark_Task_Flags *task_flags,
                       int m, int n, int ib, int nb,
                       PLASMA_Complex32_t *A1, int lda1,
                       PLASMA_Complex32_t *A2, int lda2,
                       PLASMA_Complex32_t *T, int ldt)
{
    DAG_CORE_TSQRT;
    QUARK_Insert_Task(quark, CORE_ctsqrt_quark, task_flags,
        sizeof(int),                        &m,     VALUE,
        sizeof(int),                        &n,     VALUE,
        sizeof(int),                        &ib,    VALUE,
        sizeof(PLASMA_Complex32_t)*nb*nb,    A1,            INOUT | QUARK_REGION_D | QUARK_REGION_U,
        sizeof(int),                        &lda1,  VALUE,
        sizeof(PLASMA_Complex32_t)*nb*nb,    A2,            INOUT | LOCALITY,
        sizeof(int),                        &lda2,  VALUE,
        sizeof(PLASMA_Complex32_t)*ib*nb,    T,             OUTPUT,
        sizeof(int),                        &ldt,   VALUE,
        sizeof(PLASMA_Complex32_t)*nb,       NULL,          SCRATCH,
        sizeof(PLASMA_Complex32_t)*ib*nb,    NULL,          SCRATCH,
        0);
}

/***************************************************************************/
#if defined(PLASMA_HAVE_WEAK)
#pragma weak CORE_ctsqrt_quark = PCORE_ctsqrt_quark
#define CORE_ctsqrt_quark PCORE_ctsqrt_quark
#endif
void CORE_ctsqrt_quark(Quark *quark)
{
    int m;
    int n;
    int ib;
    PLASMA_Complex32_t *A1;
    int lda1;
    PLASMA_Complex32_t *A2;
    int lda2;
    PLASMA_Complex32_t *T;
    int ldt;
    PLASMA_Complex32_t *TAU;
    PLASMA_Complex32_t *WORK;

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