libflame  revision_anchor
Functions
bli_axpysmt.c File Reference

(r)

Functions

void bli_saxpysmt (trans_t trans, int m, int n, float *alpha0, float *alpha1, float *a, int a_rs, int a_cs, float *beta, float *b, int b_rs, int b_cs)
void bli_daxpysmt (trans_t trans, int m, int n, double *alpha0, double *alpha1, double *a, int a_rs, int a_cs, double *beta, double *b, int b_rs, int b_cs)
void bli_caxpysmt (trans_t trans, int m, int n, scomplex *alpha0, scomplex *alpha1, scomplex *a, int a_rs, int a_cs, scomplex *beta, scomplex *b, int b_rs, int b_cs)
void bli_zaxpysmt (trans_t trans, int m, int n, dcomplex *alpha0, dcomplex *alpha1, dcomplex *a, int a_rs, int a_cs, dcomplex *beta, dcomplex *b, int b_rs, int b_cs)

Function Documentation

void bli_caxpysmt ( trans_t  trans,
int  m,
int  n,
scomplex alpha0,
scomplex alpha1,
scomplex a,
int  a_rs,
int  a_cs,
scomplex beta,
scomplex b,
int  b_rs,
int  b_cs 
)

References bli_callocv(), bli_caxpy(), bli_ccopyv(), bli_cfree(), bli_cscal(), bli_does_conj(), bli_does_notrans(), bli_does_trans(), bli_is_col_storage(), bli_is_row_storage(), bli_is_vector(), bli_proj_trans_to_conj(), bli_vector_dim(), bli_vector_inc(), bli_zero_dim2(), BLIS_NO_TRANSPOSE, scomplex::imag, and scomplex::real.

Referenced by FLA_Axpys_external().

{
    scomplex* a_begin;
    scomplex* b_begin;
    scomplex* a_temp;
    scomplex  alpha_prod;
    int       inca_temp;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;

    // Return early if possible.
    if ( bli_zero_dim2( m, n ) ) return;

    alpha_prod.real = alpha0->real * alpha1->real - alpha0->imag * alpha1->imag;
    alpha_prod.imag = alpha0->real * alpha1->imag + alpha0->imag * alpha1->real;

    // Handle cases where A and B are vectors to ensure that the underlying axpy
    // gets invoked only once.
    if ( bli_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bli_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bli_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bli_vector_inc( BLIS_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bli_does_trans( trans ) )
        {
            bli_swap_ints( lda, inca );
        }

        // An optimization: if B is row-major and if A is effectively row-major
        // after a possible transposition, then let's access the matrices by rows
        // instead of by columns for increased spatial locality.
        if ( bli_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bli_is_col_storage( a_rs, a_cs ) && bli_does_trans( trans ) ) ||
                 ( bli_is_row_storage( a_rs, a_cs ) && bli_does_notrans( trans ) ) )
            {
                bli_swap_ints( n_iter, n_elem );
                bli_swap_ints( lda, inca );
                bli_swap_ints( ldb, incb );
            }
        }
    }

    if ( bli_does_conj( trans ) )
    {
        conj_t conj = bli_proj_trans_to_conj( trans );

        a_temp = bli_callocv( n_elem );
        inca_temp = 1;

        for ( j = 0; j < n_iter; j++ )
        {
            a_begin = a + j*lda;
            b_begin = b + j*ldb;

            bli_ccopyv( conj,
                        n_elem,
                        a_begin, inca,
                        a_temp,  inca_temp );

            bli_cscal( n_elem,
                       beta,
                       b_begin, incb );

            bli_caxpy( n_elem,
                       &alpha_prod,
                       a_temp,  inca_temp, 
                       b_begin, incb );
        }
    
        bli_cfree( a_temp );
    }
    else // if ( !bli_does_conj( trans ) )
    {
        for ( j = 0; j < n_iter; j++ )
        {
            a_begin = a + j*lda;
            b_begin = b + j*ldb;

            bli_cscal( n_elem,
                       beta,
                       b_begin, incb );

            bli_caxpy( n_elem,
                       &alpha_prod,
                       a_begin, inca, 
                       b_begin, incb );
        }
    }
}
void bli_daxpysmt ( trans_t  trans,
int  m,
int  n,
double *  alpha0,
double *  alpha1,
double *  a,
int  a_rs,
int  a_cs,
double *  beta,
double *  b,
int  b_rs,
int  b_cs 
)

References bli_daxpy(), bli_does_notrans(), bli_does_trans(), bli_dscal(), bli_is_col_storage(), bli_is_row_storage(), bli_is_vector(), bli_vector_dim(), bli_vector_inc(), bli_zero_dim2(), and BLIS_NO_TRANSPOSE.

Referenced by FLA_Axpys_external().

{
    double*   a_begin;
    double*   b_begin;
    double    alpha_prod;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;

    // Return early if possible.
    if ( bli_zero_dim2( m, n ) ) return;

    alpha_prod = (*alpha0) * (*alpha1);

    // Handle cases where A and B are vectors to ensure that the underlying axpy
    // gets invoked only once.
    if ( bli_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bli_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bli_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bli_vector_inc( BLIS_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bli_does_trans( trans ) )
        {
            bli_swap_ints( lda, inca );
        }

        // An optimization: if B is row-major and if A is effectively row-major
        // after a possible transposition, then let's access the matrices by rows
        // instead of by columns for increased spatial locality.
        if ( bli_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bli_is_col_storage( a_rs, a_cs ) && bli_does_trans( trans ) ) ||
                 ( bli_is_row_storage( a_rs, a_cs ) && bli_does_notrans( trans ) ) )
            {
                bli_swap_ints( n_iter, n_elem );
                bli_swap_ints( lda, inca );
                bli_swap_ints( ldb, incb );
            }
        }
    }

    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;

        bli_dscal( n_elem,
                   beta,
                   b_begin, incb );

        bli_daxpy( n_elem,
                   &alpha_prod,
                   a_begin, inca, 
                   b_begin, incb );
    }
}
void bli_saxpysmt ( trans_t  trans,
int  m,
int  n,
float *  alpha0,
float *  alpha1,
float *  a,
int  a_rs,
int  a_cs,
float *  beta,
float *  b,
int  b_rs,
int  b_cs 
)

References bli_does_notrans(), bli_does_trans(), bli_is_col_storage(), bli_is_row_storage(), bli_is_vector(), bli_saxpy(), bli_sscal(), bli_vector_dim(), bli_vector_inc(), bli_zero_dim2(), and BLIS_NO_TRANSPOSE.

Referenced by FLA_Axpys_external().

{
    float*    a_begin;
    float*    b_begin;
    float     alpha_prod;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;

    // Return early if possible.
    if ( bli_zero_dim2( m, n ) ) return;

    alpha_prod = (*alpha0) * (*alpha1);

    // Handle cases where A and B are vectors to ensure that the underlying axpy
    // gets invoked only once.
    if ( bli_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bli_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bli_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bli_vector_inc( BLIS_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bli_does_trans( trans ) )
        {
            bli_swap_ints( lda, inca );
        }

        // An optimization: if B is row-major and if A is effectively row-major
        // after a possible transposition, then let's access the matrices by rows
        // instead of by columns for increased spatial locality.
        if ( bli_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bli_is_col_storage( a_rs, a_cs ) && bli_does_trans( trans ) ) ||
                 ( bli_is_row_storage( a_rs, a_cs ) && bli_does_notrans( trans ) ) )
            {
                bli_swap_ints( n_iter, n_elem );
                bli_swap_ints( lda, inca );
                bli_swap_ints( ldb, incb );
            }
        }
    }

    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;

        bli_sscal( n_elem,
                   beta,
                   b_begin, incb );

        bli_saxpy( n_elem,
                   &alpha_prod,
                   a_begin, inca, 
                   b_begin, incb );
    }
}
void bli_zaxpysmt ( trans_t  trans,
int  m,
int  n,
dcomplex alpha0,
dcomplex alpha1,
dcomplex a,
int  a_rs,
int  a_cs,
dcomplex beta,
dcomplex b,
int  b_rs,
int  b_cs 
)

References bli_does_conj(), bli_does_notrans(), bli_does_trans(), bli_is_col_storage(), bli_is_row_storage(), bli_is_vector(), bli_proj_trans_to_conj(), bli_vector_dim(), bli_vector_inc(), bli_zallocv(), bli_zaxpy(), bli_zcopyv(), bli_zero_dim2(), bli_zfree(), bli_zscal(), BLIS_NO_TRANSPOSE, dcomplex::imag, and dcomplex::real.

Referenced by FLA_Axpys_external().

{
    dcomplex* a_begin;
    dcomplex* b_begin;
    dcomplex* a_temp;
    dcomplex  alpha_prod;
    int       inca_temp;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;

    // Return early if possible.
    if ( bli_zero_dim2( m, n ) ) return;

    alpha_prod.real = alpha0->real * alpha1->real - alpha0->imag * alpha1->imag;
    alpha_prod.imag = alpha0->real * alpha1->imag + alpha0->imag * alpha1->real;

    // Handle cases where A and B are vectors to ensure that the underlying axpy
    // gets invoked only once.
    if ( bli_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bli_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bli_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bli_vector_inc( BLIS_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bli_does_trans( trans ) )
        {
            bli_swap_ints( lda, inca );
        }

        // An optimization: if B is row-major and if A is effectively row-major
        // after a possible transposition, then let's access the matrices by rows
        // instead of by columns for increased spatial locality.
        if ( bli_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bli_is_col_storage( a_rs, a_cs ) && bli_does_trans( trans ) ) ||
                 ( bli_is_row_storage( a_rs, a_cs ) && bli_does_notrans( trans ) ) )
            {
                bli_swap_ints( n_iter, n_elem );
                bli_swap_ints( lda, inca );
                bli_swap_ints( ldb, incb );
            }
        }
    }

    if ( bli_does_conj( trans ) )
    {
        conj_t conj = bli_proj_trans_to_conj( trans );

        a_temp = bli_zallocv( n_elem );
        inca_temp = 1;

        for ( j = 0; j < n_iter; j++ )
        {
            a_begin = a + j*lda;
            b_begin = b + j*ldb;

            bli_zcopyv( conj,
                        n_elem,
                        a_begin, inca,
                        a_temp,  inca_temp );

            bli_zscal( n_elem,
                       beta,
                       b_begin, incb );

            bli_zaxpy( n_elem,
                       &alpha_prod,
                       a_temp,  inca_temp, 
                       b_begin, incb );
        }
    
        bli_zfree( a_temp );
    }
    else // if ( !bli_does_conj( trans ) )
    {
        for ( j = 0; j < n_iter; j++ )
        {
            a_begin = a + j*lda;
            b_begin = b + j*ldb;

            bli_zscal( n_elem,
                       beta,
                       b_begin, incb );

            bli_zaxpy( n_elem,
                       &alpha_prod,
                       a_begin, inca, 
                       b_begin, incb );
        }
    }
}