libflame  revision_anchor
Functions
bl1_axmyv2.c File Reference

(r)

Functions

void bl1_saxmyv2 (conj1_t conjx, int n, float *alpha, float *beta, float *x, int inc_x, float *y, int inc_y, float *z, int inc_z)
void bl1_daxmyv2 (conj1_t conjx, int n, double *alpha, double *beta, double *x, int inc_x, double *y, int inc_y, double *z, int inc_z)

Function Documentation

void bl1_daxmyv2 ( conj1_t  conjx,
int  n,
double *  alpha,
double *  beta,
double *  x,
int  inc_x,
double *  y,
int  inc_y,
double *  z,
int  inc_z 
)

References bl1_abort(), and v2df_t::v.

Referenced by FLA_Fused_UYx_ZVx_opd_var1().

{
    double*   restrict chi1;
    double*   restrict psi1;
    double*   restrict zeta1;
    int       i;

    int       n_pre;
    int       n_run;
    int       n_left;

    v2df_t    a1v, b1v;
    v2df_t    x1v, y1v, z1v;
    v2df_t    x2v, y2v, z2v;

    if ( inc_x != 1 ||
         inc_y != 1 ||
         inc_z != 1 ) bl1_abort();

    n_pre = 0;
    if ( ( unsigned long ) z % 16 != 0 )
    {
        if ( ( unsigned long ) x % 16 == 0 ||
             ( unsigned long ) y % 16 == 0 ) bl1_abort();

        n_pre = 1;
    }

    n_run       = ( n - n_pre ) / 4;
    n_left      = ( n - n_pre ) % 4;

    chi1  = x;
    psi1  = y;
    zeta1 = z;

    if ( n_pre == 1 )
    {
        double   alpha_c = *alpha;
        double   beta_c  = *beta;
        double   chi1_c  = *chi1;

        *psi1  -= alpha_c * chi1_c;
        *zeta1 -= beta_c  * chi1_c;

        chi1  += inc_x;
        psi1  += inc_y;
        zeta1 += inc_z;
    }

    a1v.v = _mm_loaddup_pd( ( double* )alpha );
    b1v.v = _mm_loaddup_pd( ( double* )beta );

    for ( i = 0; i < n_run; ++i )
    {
        x1v.v = _mm_load_pd( ( double* )chi1 );
        y1v.v = _mm_load_pd( ( double* )psi1 );
        z1v.v = _mm_load_pd( ( double* )zeta1 );

        x2v.v = _mm_load_pd( ( double* )(chi1 + 2) );
        y2v.v = _mm_load_pd( ( double* )(psi1 + 2) );
        z2v.v = _mm_load_pd( ( double* )(zeta1 + 2) );

        y1v.v = y1v.v - a1v.v * x1v.v;
        z1v.v = z1v.v - b1v.v * x1v.v;

        _mm_store_pd( ( double* )psi1,  y1v.v );
        _mm_store_pd( ( double* )zeta1, z1v.v );

        y2v.v = y2v.v - a1v.v * x2v.v;
        z2v.v = z2v.v - b1v.v * x2v.v;

        _mm_store_pd( ( double* )(psi1 + 2),  y2v.v );
        _mm_store_pd( ( double* )(zeta1 + 2), z2v.v );

        chi1  += 4;
        psi1  += 4;
        zeta1 += 4;
    }

    if ( n_left > 0 )
    {
        double   alpha_c = *alpha;
        double   beta_c  = *beta;

        for( i = 0; i < n_left; ++i )
        {
            double   chi1_c = *chi1;

            *psi1  -= alpha_c * chi1_c;
            *zeta1 -= beta_c  * chi1_c;

            chi1  += inc_x;
            psi1  += inc_y;
            zeta1 += inc_z;
        }
    }
}
void bl1_saxmyv2 ( conj1_t  conjx,
int  n,
float *  alpha,
float *  beta,
float *  x,
int  inc_x,
float *  y,
int  inc_y,
float *  z,
int  inc_z 
)

References bl1_abort().

{
    bl1_abort();
}