core_cgeadd.c

Go to the documentation of this file.

#include "common.h"

/***************************************************************************/
#if defined(PLASMA_HAVE_WEAK)
#pragma weak CORE_cgeadd = PCORE_cgeadd
#define CORE_cgeadd PCORE_cgeadd
#endif
void CORE_cgeadd(int M, int N,  PLASMA_Complex32_t alpha,
                PLASMA_Complex32_t *A, int LDA,
                PLASMA_Complex32_t *B, int LDB)
{
    int j;

    if (M == LDA && M == LDB)
        cblas_caxpy(M*N, CBLAS_SADDR(alpha), A, 1, B, 1);
    else {
        for (j = 0; j < N; j++)
            cblas_caxpy(M, CBLAS_SADDR(alpha), &A[j*LDA], 1, &B[j*LDB], 1);
    }
}

/***************************************************************************/
void QUARK_CORE_cgeadd(Quark *quark, Quark_Task_Flags *task_flags,
                      int m, int n, int nb, PLASMA_Complex32_t alpha,
                      PLASMA_Complex32_t *A, int lda,
                      PLASMA_Complex32_t *B, int ldb)
{
    DAG_CORE_GEADD;
    QUARK_Insert_Task(quark, CORE_cgeadd_quark, task_flags,
        sizeof(int),                        &m,     VALUE,
        sizeof(int),                        &n,     VALUE,
        sizeof(PLASMA_Complex32_t),         &alpha, VALUE,
        sizeof(PLASMA_Complex32_t)*nb*nb,    A,             INPUT,
        sizeof(int),                        &lda,   VALUE,
        sizeof(PLASMA_Complex32_t)*nb*nb,    B,             INOUT,
        sizeof(int),                        &ldb,   VALUE,
        0);
}

/***************************************************************************/
#if defined(PLASMA_HAVE_WEAK)
#pragma weak CORE_cgeadd_quark = PCORE_cgeadd_quark
#define CORE_cgeadd_quark PCORE_cgeadd_quark
#endif
void CORE_cgeadd_quark(Quark *quark)
{
    int M;
    int N;
    PLASMA_Complex32_t alpha;
    PLASMA_Complex32_t *A;
    int LDA;
    PLASMA_Complex32_t *B;
    int LDB;

    int j;

    quark_unpack_args_7(quark, M, N, alpha, A, LDA, B, LDB);
    if (M == LDA && M == LDB)
        cblas_caxpy(M*N, CBLAS_SADDR(alpha), A, 1, B, 1);
    else {
        for (j = 0; j < N; j++)
            cblas_caxpy(M, CBLAS_SADDR(alpha), &A[j*LDA], 1, &B[j*LDB], 1);
    }
}