bl1_copymt.c File Reference

(r)

Functions
void	bl1_icopymt (trans1_t trans, int m, int n, int *a, int a_rs, int a_cs, int *b, int b_rs, int b_cs)
void	bl1_scopymt (trans1_t trans, int m, int n, float *a, int a_rs, int a_cs, float *b, int b_rs, int b_cs)
void	bl1_dcopymt (trans1_t trans, int m, int n, double *a, int a_rs, int a_cs, double *b, int b_rs, int b_cs)
void	bl1_ccopymt (trans1_t trans, int m, int n, scomplex *a, int a_rs, int a_cs, scomplex *b, int b_rs, int b_cs)
void	bl1_zcopymt (trans1_t trans, int m, int n, dcomplex *a, int a_rs, int a_cs, dcomplex *b, int b_rs, int b_cs)
void	bl1_sscopymt (trans1_t trans, int m, int n, float *a, int a_rs, int a_cs, float *b, int b_rs, int b_cs)
void	bl1_sdcopymt (trans1_t trans, int m, int n, float *a, int a_rs, int a_cs, double *b, int b_rs, int b_cs)
void	bl1_dscopymt (trans1_t trans, int m, int n, double *a, int a_rs, int a_cs, float *b, int b_rs, int b_cs)
void	bl1_sccopymt (trans1_t trans, int m, int n, float *a, int a_rs, int a_cs, scomplex *b, int b_rs, int b_cs)
void	bl1_cscopymt (trans1_t trans, int m, int n, scomplex *a, int a_rs, int a_cs, float *b, int b_rs, int b_cs)
void	bl1_szcopymt (trans1_t trans, int m, int n, float *a, int a_rs, int a_cs, dcomplex *b, int b_rs, int b_cs)
void	bl1_zscopymt (trans1_t trans, int m, int n, dcomplex *a, int a_rs, int a_cs, float *b, int b_rs, int b_cs)
void	bl1_ddcopymt (trans1_t trans, int m, int n, double *a, int a_rs, int a_cs, double *b, int b_rs, int b_cs)
void	bl1_dccopymt (trans1_t trans, int m, int n, double *a, int a_rs, int a_cs, scomplex *b, int b_rs, int b_cs)
void	bl1_cdcopymt (trans1_t trans, int m, int n, scomplex *a, int a_rs, int a_cs, double *b, int b_rs, int b_cs)
void	bl1_dzcopymt (trans1_t trans, int m, int n, double *a, int a_rs, int a_cs, dcomplex *b, int b_rs, int b_cs)
void	bl1_zdcopymt (trans1_t trans, int m, int n, dcomplex *a, int a_rs, int a_cs, double *b, int b_rs, int b_cs)
void	bl1_cccopymt (trans1_t trans, int m, int n, scomplex *a, int a_rs, int a_cs, scomplex *b, int b_rs, int b_cs)
void	bl1_czcopymt (trans1_t trans, int m, int n, scomplex *a, int a_rs, int a_cs, dcomplex *b, int b_rs, int b_cs)
void	bl1_zccopymt (trans1_t trans, int m, int n, dcomplex *a, int a_rs, int a_cs, scomplex *b, int b_rs, int b_cs)
void	bl1_zzcopymt (trans1_t trans, int m, int n, dcomplex *a, int a_rs, int a_cs, dcomplex *b, int b_rs, int b_cs)

Function Documentation

bl1_cccopymt()

void bl1_cccopymt	(	trans1_t	trans,
		int	m,
		int	n,
		scomplex *	a,
		int	a_rs,
		int	a_cs,
		scomplex *	b,
		int	b_rs,
		int	b_cs )

{
    scomplex* a_begin;
    scomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_ccopyv( conj,
                    n_elem,
                    a_begin, inca,
                    b_begin, incb );
    }
}

References bl1_ccopyv(), bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

bl1_ccopymt()

void bl1_ccopymt	(	trans1_t	trans,
		int	m,
		int	n,
		scomplex *	a,
		int	a_rs,
		int	a_cs,
		scomplex *	b,
		int	b_rs,
		int	b_cs )

{
    scomplex* a_begin;
    scomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_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 ( bl1_does_trans( trans ) )
        {
            bl1_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 matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bl1_is_col_storage( a_rs, a_cs ) && bl1_does_trans( trans ) ) ||
                 ( bl1_is_row_storage( a_rs, a_cs ) && bl1_does_notrans( trans ) ) )
            {
                bl1_swap_ints( n_iter, n_elem );
                bl1_swap_ints( lda, inca );
                bl1_swap_ints( ldb, incb );
            }
        }
    }
 
    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
        
        bl1_ccopy( n_elem,
                   a_begin, inca, 
                   b_begin, incb );
 
        if ( bl1_does_conj( trans ) )
            bl1_cconjv( n_elem,
                        b_begin, incb );
    }
}

References bl1_cconjv(), bl1_ccopy(), bl1_does_conj(), bl1_does_notrans(), bl1_does_trans(), bl1_is_col_storage(), bl1_is_row_storage(), bl1_is_vector(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by bl1_ccreate_contigm(), bl1_ccreate_contigmt(), bl1_cfree_saved_contigm(), bl1_cfree_saved_contigmsr(), bl1_cgemm(), bl1_chemm(), bl1_cher2k(), bl1_csymm(), bl1_csyr2k(), bl1_ctrmmsx(), bl1_ctrsmsx(), FLA_Copy_external(), and FLA_Copyt_external().

bl1_cdcopymt()

void bl1_cdcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		scomplex *	a,
		int	a_rs,
		int	a_cs,
		double *	b,
		int	b_rs,
		int	b_cs )

{
    scomplex* a_begin;
    double*   b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_cdcopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_cdcopyv(), bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_cscopymt()

void bl1_cscopymt	(	trans1_t	trans,
		int	m,
		int	n,
		scomplex *	a,
		int	a_rs,
		int	a_cs,
		float *	b,
		int	b_rs,
		int	b_cs )

{
    scomplex* a_begin;
    float*    b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_cscopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_cscopyv(), bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_czcopymt()

void bl1_czcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		scomplex *	a,
		int	a_rs,
		int	a_cs,
		dcomplex *	b,
		int	b_rs,
		int	b_cs )

{
    scomplex* a_begin;
    dcomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_czcopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_czcopyv(), bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_dccopymt()

void bl1_dccopymt	(	trans1_t	trans,
		int	m,
		int	n,
		double *	a,
		int	a_rs,
		int	a_cs,
		scomplex *	b,
		int	b_rs,
		int	b_cs )

{
    double*   a_begin;
    scomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_dccopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_dccopyv(), bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_dcopymt()

void bl1_dcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		double *	a,
		int	a_rs,
		int	a_cs,
		double *	b,
		int	b_rs,
		int	b_cs )

{
    double*   a_begin;
    double*   b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_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 ( bl1_does_trans( trans ) )
        {
            bl1_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 matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bl1_is_col_storage( a_rs, a_cs ) && bl1_does_trans( trans ) ) ||
                 ( bl1_is_row_storage( a_rs, a_cs ) && bl1_does_notrans( trans ) ) )
            {
                bl1_swap_ints( n_iter, n_elem );
                bl1_swap_ints( lda, inca );
                bl1_swap_ints( ldb, incb );
            }
        }
    }
 
    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
        
        bl1_dcopy( n_elem,
                   a_begin, inca, 
                   b_begin, incb );
    }
}

References bl1_dcopy(), bl1_does_notrans(), bl1_does_trans(), bl1_is_col_storage(), bl1_is_row_storage(), bl1_is_vector(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by bl1_dcreate_contigm(), bl1_dcreate_contigmt(), bl1_dfree_saved_contigm(), bl1_dfree_saved_contigmsr(), bl1_dsymm(), bl1_dsyr2k(), bl1_dtrmmsx(), bl1_dtrsmsx(), FLA_Bsvd_v_opd_var2(), FLA_Copy_external(), FLA_Copyt_external(), and FLA_Tevd_v_opd_var2().

bl1_ddcopymt()

void bl1_ddcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		double *	a,
		int	a_rs,
		int	a_cs,
		double *	b,
		int	b_rs,
		int	b_cs )

{
    double*   a_begin;
    double*   b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_dcopyv( conj,
                    n_elem,
                    a_begin, inca,
                    b_begin, incb );
    }
}

References bl1_dcopyv(), bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

bl1_dscopymt()

void bl1_dscopymt	(	trans1_t	trans,
		int	m,
		int	n,
		double *	a,
		int	a_rs,
		int	a_cs,
		float *	b,
		int	b_rs,
		int	b_cs )

{
    double*   a_begin;
    float*    b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_dscopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_dscopyv(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_dzcopymt()

void bl1_dzcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		double *	a,
		int	a_rs,
		int	a_cs,
		dcomplex *	b,
		int	b_rs,
		int	b_cs )

{
    double*   a_begin;
    dcomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_dzcopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_dzcopyv(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_icopymt()

void bl1_icopymt	(	trans1_t	trans,
		int	m,
		int	n,
		int *	a,
		int	a_rs,
		int	a_cs,
		int *	b,
		int	b_rs,
		int	b_cs )

{
    int*      a_begin;
    int*      b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_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 ( bl1_does_trans( trans ) )
        {
            bl1_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 matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bl1_is_col_storage( a_rs, a_cs ) && bl1_does_trans( trans ) ) ||
                 ( bl1_is_row_storage( a_rs, a_cs ) && bl1_does_notrans( trans ) ) )
            {
                bl1_swap_ints( n_iter, n_elem );
                bl1_swap_ints( lda, inca );
                bl1_swap_ints( ldb, incb );
            }
        }
    }
 
    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
        
        bl1_icopyv( bl1_proj_trans1_to_conj( trans ),
                    n_elem,
                    a_begin, inca, 
                    b_begin, incb );
    }
}

References bl1_does_notrans(), bl1_does_trans(), bl1_icopyv(), bl1_is_col_storage(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_sccopymt()

void bl1_sccopymt	(	trans1_t	trans,
		int	m,
		int	n,
		float *	a,
		int	a_rs,
		int	a_cs,
		scomplex *	b,
		int	b_rs,
		int	b_cs )

{
    float*    a_begin;
    scomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_sccopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_sccopyv(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_scopymt()

void bl1_scopymt	(	trans1_t	trans,
		int	m,
		int	n,
		float *	a,
		int	a_rs,
		int	a_cs,
		float *	b,
		int	b_rs,
		int	b_cs )

{
    float*    a_begin;
    float*    b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_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 ( bl1_does_trans( trans ) )
        {
            bl1_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 matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bl1_is_col_storage( a_rs, a_cs ) && bl1_does_trans( trans ) ) ||
                 ( bl1_is_row_storage( a_rs, a_cs ) && bl1_does_notrans( trans ) ) )
            {
                bl1_swap_ints( n_iter, n_elem );
                bl1_swap_ints( lda, inca );
                bl1_swap_ints( ldb, incb );
            }
        }
    }
 
    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
        
        bl1_scopy( n_elem,
                   a_begin, inca, 
                   b_begin, incb );
    }
}

References bl1_does_notrans(), bl1_does_trans(), bl1_is_col_storage(), bl1_is_row_storage(), bl1_is_vector(), bl1_scopy(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by bl1_screate_contigm(), bl1_screate_contigmt(), bl1_sfree_saved_contigm(), bl1_sfree_saved_contigmsr(), bl1_ssymm(), bl1_ssyr2k(), bl1_strmmsx(), bl1_strsmsx(), FLA_Copy_external(), and FLA_Copyt_external().

bl1_sdcopymt()

void bl1_sdcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		float *	a,
		int	a_rs,
		int	a_cs,
		double *	b,
		int	b_rs,
		int	b_cs )

{
    float*    a_begin;
    double*   b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_sdcopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_sdcopyv(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_sscopymt()

void bl1_sscopymt	(	trans1_t	trans,
		int	m,
		int	n,
		float *	a,
		int	a_rs,
		int	a_cs,
		float *	b,
		int	b_rs,
		int	b_cs )

{
    float*    a_begin;
    float*    b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_scopyv( conj,
                    n_elem,
                    a_begin, inca,
                    b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_scopyv(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

bl1_szcopymt()

void bl1_szcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		float *	a,
		int	a_rs,
		int	a_cs,
		dcomplex *	b,
		int	b_rs,
		int	b_cs )

{
    float*    a_begin;
    dcomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_szcopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_szcopyv(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_zccopymt()

void bl1_zccopymt	(	trans1_t	trans,
		int	m,
		int	n,
		dcomplex *	a,
		int	a_rs,
		int	a_cs,
		scomplex *	b,
		int	b_rs,
		int	b_cs )

{
    dcomplex* a_begin;
    scomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_zccopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zccopyv(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_zcopymt()

void bl1_zcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		dcomplex *	a,
		int	a_rs,
		int	a_cs,
		dcomplex *	b,
		int	b_rs,
		int	b_cs )

{
    dcomplex* a_begin;
    dcomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_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 ( bl1_does_trans( trans ) )
        {
            bl1_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 matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            if ( ( bl1_is_col_storage( a_rs, a_cs ) && bl1_does_trans( trans ) ) ||
                 ( bl1_is_row_storage( a_rs, a_cs ) && bl1_does_notrans( trans ) ) )
            {
                bl1_swap_ints( n_iter, n_elem );
                bl1_swap_ints( lda, inca );
                bl1_swap_ints( ldb, incb );
            }
        }
    }
 
    for ( j = 0; j < n_iter; j++ )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
        
        bl1_zcopy( n_elem,
                   a_begin, inca, 
                   b_begin, incb );
 
        if ( bl1_does_conj( trans ) )
            bl1_zconjv( n_elem,
                        b_begin, incb );
    }
}

References bl1_does_conj(), bl1_does_notrans(), bl1_does_trans(), bl1_is_col_storage(), bl1_is_row_storage(), bl1_is_vector(), bl1_vector_dim(), bl1_vector_inc(), bl1_zconjv(), bl1_zcopy(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by bl1_zcreate_contigm(), bl1_zcreate_contigmt(), bl1_zfree_saved_contigm(), bl1_zgemm(), bl1_zhemm(), bl1_zher2k(), bl1_zsymm(), bl1_zsyr2k(), bl1_ztrmmsx(), bl1_ztrsmsx(), FLA_Bsvd_v_opz_var2(), FLA_Copy_external(), FLA_Copyt_external(), and FLA_Tevd_v_opz_var2().

bl1_zdcopymt()

void bl1_zdcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		dcomplex *	a,
		int	a_rs,
		int	a_cs,
		double *	b,
		int	b_rs,
		int	b_cs )

{
    dcomplex* a_begin;
    double*   b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_zdcopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zdcopyv(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_zscopymt()

void bl1_zscopymt	(	trans1_t	trans,
		int	m,
		int	n,
		dcomplex *	a,
		int	a_rs,
		int	a_cs,
		float *	b,
		int	b_rs,
		int	b_cs )

{
    dcomplex* a_begin;
    float*    b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_zscopyv( conj,
                     n_elem,
                     a_begin, inca,
                     b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zero_dim2(), bl1_zscopyv(), and BLIS1_NO_TRANSPOSE.

Referenced by FLA_Copy_external(), and FLA_Copyt_external().

bl1_zzcopymt()

void bl1_zzcopymt	(	trans1_t	trans,
		int	m,
		int	n,
		dcomplex *	a,
		int	a_rs,
		int	a_cs,
		dcomplex *	b,
		int	b_rs,
		int	b_cs )

{
    dcomplex* a_begin;
    dcomplex* b_begin;
    int       lda, inca;
    int       ldb, incb;
    int       n_iter;
    int       n_elem;
    int       j;
    conj1_t    conj;
 
    // Return early if possible.
    if ( bl1_zero_dim2( m, n ) ) return;
 
    // Handle cases where A and B are vectors to ensure that the underlying copy
    // gets invoked only once.
    if ( bl1_is_vector( m, n ) )
    {
        // Initialize with values appropriate for vectors.
        n_iter = 1;
        n_elem = bl1_vector_dim( m, n );
        lda    = 1; // multiplied by zero when n_iter == 1; not needed.
        inca   = bl1_vector_inc( trans,             m, n, a_rs, a_cs );
        ldb    = 1; // multiplied by zero when n_iter == 1; not needed.
        incb   = bl1_vector_inc( BLIS1_NO_TRANSPOSE, m, n, b_rs, b_cs );
    }
    else // matrix case
    {
        // Initialize with optimal values for column-major storage of B.
        n_iter = n;
        n_elem = m;
        lda    = a_cs;
        inca   = a_rs;
        ldb    = b_cs;
        incb   = b_rs;
        
        // Handle the transposition of A.
        if ( bl1_does_trans( trans ) )
        {
            bl1_swap_ints( lda, inca );
        }
 
        // An optimization: if B is row-major, then let's access the matrix by rows
        // instead of by columns for increased spatial locality.
        if ( bl1_is_row_storage( b_rs, b_cs ) )
        {
            bl1_swap_ints( n_iter, n_elem );
            bl1_swap_ints( lda, inca );
            bl1_swap_ints( ldb, incb );
        }
    }
 
    // Extract conj component from trans parameter.
    conj = bl1_proj_trans1_to_conj( trans );
 
    for ( j = 0; j < n_iter; ++j )
    {
        a_begin = a + j*lda;
        b_begin = b + j*ldb;
 
        bl1_zcopyv( conj,
                    n_elem,
                    a_begin, inca,
                    b_begin, incb );
    }
}

References bl1_does_trans(), bl1_is_row_storage(), bl1_is_vector(), bl1_proj_trans1_to_conj(), bl1_vector_dim(), bl1_vector_inc(), bl1_zcopyv(), bl1_zero_dim2(), and BLIS1_NO_TRANSPOSE.