FLA_Blocksize.c File Reference

(r)

Functions
void	blas_set_parameter (void)
fla_blocksize_t *	FLA_Blocksize_create (dim_t b_s, dim_t b_d, dim_t b_c, dim_t b_z)
void	FLA_Blocksize_set (fla_blocksize_t *bp, dim_t b_s, dim_t b_d, dim_t b_c, dim_t b_z)
void	FLA_Blocksize_scale (fla_blocksize_t *bp, double factor)
fla_blocksize_t *	FLA_Blocksize_create_copy (fla_blocksize_t *bp)
void	FLA_Blocksize_free (fla_blocksize_t *bp)
dim_t	FLA_Blocksize_extract (FLA_Datatype dt, fla_blocksize_t *bp)
fla_blocksize_t *	FLA_Query_blocksizes (FLA_Dimension dim)
dim_t	FLA_Query_blocksize (FLA_Datatype dt, FLA_Dimension dim)
dim_t	FLA_Determine_blocksize (FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *bp)
dim_t	FLA_determine_matrix_size (FLA_Obj A_unproc, FLA_Quadrant to_dir)

Variables
long	sgemm_p
long	sgemm_q
long	sgemm_r
long	dgemm_p
long	dgemm_q
long	dgemm_r
long	cgemm_p
long	cgemm_q
long	cgemm_r
long	zgemm_p
long	zgemm_q
long	zgemm_r
long	fla_goto_gemm_blocksize [4][4]

Function Documentation

blas_set_parameter()

void blas_set_parameter ( void  )

extern

FLA_Blocksize_create()

fla_blocksize_t * FLA_Blocksize_create	(	dim_t	b_s,
		dim_t	b_d,
		dim_t	b_c,
		dim_t	b_z
	)

{
    fla_blocksize_t* bp;
    
    // Allocate memory for the blocksize structure.
    bp = ( fla_blocksize_t* ) FLA_malloc( sizeof(fla_blocksize_t) );
    
    // Assign the provided blocksize values into the corresponding fields.
    bp->s = b_s;
    bp->d = b_d;
    bp->c = b_c;
    bp->z = b_z;
    
    // Return a pointer to the structure.
    return bp;
}

References FLA_Blocksize_s::c, FLA_Blocksize_s::d, FLA_malloc(), FLA_Blocksize_s::s, and FLA_Blocksize_s::z.

Referenced by FLA_Query_blocksizes(), FLASH_Apply_CAQ2_UT_cntl_init(), FLASH_Apply_CAQ_UT_inc_cntl_init(), FLASH_Apply_pivots_cntl_init(), FLASH_Apply_Q2_UT_cntl_init(), FLASH_Apply_Q_UT_cntl_init(), FLASH_Apply_Q_UT_inc_cntl_init(), FLASH_Apply_QUD_UT_cntl_init(), FLASH_Apply_QUD_UT_inc_cntl_init(), FLASH_Axpy_cntl_init(), FLASH_Axpyt_cntl_init(), FLASH_CAQR2_UT_cntl_init(), FLASH_CAQR_UT_inc_cntl_init(), FLASH_Chol_cntl_init(), FLASH_Copy_cntl_init(), FLASH_Copyr_cntl_init(), FLASH_Copyt_cntl_init(), FLASH_Eig_gest_cntl_init(), FLASH_Gemm_cntl_init(), FLASH_Gemv_cntl_init(), FLASH_Hemm_cntl_init(), FLASH_Her2k_cntl_init(), FLASH_Herk_cntl_init(), FLASH_LQ_UT_cntl_init(), FLASH_LU_incpiv_cntl_init(), FLASH_LU_nopiv_cntl_init(), FLASH_LU_piv_cntl_init(), FLASH_Lyap_cntl_init(), FLASH_QR2_UT_cntl_init(), FLASH_QR_UT_cntl_init(), FLASH_QR_UT_inc_cntl_init(), FLASH_Scal_cntl_init(), FLASH_Scalr_cntl_init(), FLASH_Sylv_cntl_init(), FLASH_Symm_cntl_init(), FLASH_Syr2k_cntl_init(), FLASH_Syrk_cntl_init(), FLASH_Trinv_cntl_init(), FLASH_Trmm_cntl_init(), FLASH_Trsm_cntl_init(), FLASH_Trsv_cntl_init(), FLASH_Ttmm_cntl_init(), FLASH_UDdate_UT_cntl_init(), and FLASH_UDdate_UT_inc_cntl_init().

FLA_Blocksize_create_copy()

fla_blocksize_t * FLA_Blocksize_create_copy

(

fla_blocksize_t *

bp

)

{
    fla_blocksize_t* bp_copy;
    FLA_Error        e_val;
    
    // Verify that the given blocksize pointer is valid.
    if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
    {
        e_val = FLA_Check_null_pointer( bp );
        FLA_Check_error_code( e_val );
    }
 
    // Allocate memory for the blocksize structure.
    bp_copy = ( fla_blocksize_t* ) FLA_malloc( sizeof(fla_blocksize_t) );
    
    // Assign the provided blocksize object's values into the corresponding
    // fields of the new object.
    bp_copy->s = bp->s;
    bp_copy->d = bp->d;
    bp_copy->c = bp->c;
    bp_copy->z = bp->z;
    
    // Return a pointer to the structure.
    return bp_copy;
}

References FLA_Blocksize_s::c, FLA_Blocksize_s::d, FLA_Check_error_level(), FLA_Check_null_pointer(), FLA_malloc(), FLA_Blocksize_s::s, and FLA_Blocksize_s::z.

Referenced by FLA_Chol_cntl_init(), FLA_LU_nopiv_cntl_init(), and FLA_LU_piv_cntl_init().

FLA_Blocksize_extract()

dim_t FLA_Blocksize_extract	(	FLA_Datatype	dt,
		fla_blocksize_t *	bp
	)

{
    dim_t     b = 0;
    FLA_Error e_val;
 
    // Verify that the given blocksize pointer is valid.
    if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
    {
        e_val = FLA_Check_null_pointer( bp );
        FLA_Check_error_code( e_val );
    }
 
    if      ( dt == FLA_FLOAT )
        b = bp->s;
    else if ( dt == FLA_DOUBLE )
        b = bp->d;
    else if ( dt == FLA_COMPLEX )
        b = bp->c;
    else if ( dt == FLA_DOUBLE_COMPLEX )
        b = bp->z;
    
    // Return the blocksize corresponding with the datatype.
    return b;
}

References FLA_Blocksize_s::c, FLA_Blocksize_s::d, FLA_Check_error_level(), FLA_Check_null_pointer(), FLA_Blocksize_s::s, and FLA_Blocksize_s::z.

Referenced by FLA_Check_blocksize_object(), FLA_Determine_blocksize(), FLA_Trinv(), and FLA_Ttmm().

FLA_Blocksize_free()

void FLA_Blocksize_free

(

fla_blocksize_t *

bp

)

{
    FLA_free( bp );
}

References FLA_free().

Referenced by FLA_Apply_CAQ2_UT_cntl_finalize(), FLA_Apply_Q2_UT_cntl_finalize(), FLA_Apply_Q_UT_cntl_finalize(), FLA_Apply_QUD_UT_cntl_finalize(), FLA_Bidiag_UT_cntl_finalize(), FLA_CAQR2_UT_cntl_finalize(), FLA_Chol_cntl_finalize(), FLA_Eig_gest_cntl_finalize(), FLA_Gemm_cntl_finalize(), FLA_Hemm_cntl_finalize(), FLA_Her2k_cntl_finalize(), FLA_Herk_cntl_finalize(), FLA_Hess_UT_cntl_finalize(), FLA_LQ_UT_cntl_finalize(), FLA_LU_nopiv_cntl_finalize(), FLA_LU_piv_cntl_finalize(), FLA_Lyap_cntl_finalize(), FLA_QR2_UT_cntl_finalize(), FLA_QR_UT_cntl_finalize(), FLA_SPDinv_cntl_finalize(), FLA_Sylv_cntl_finalize(), FLA_Symm_cntl_finalize(), FLA_Syr2k_cntl_finalize(), FLA_Syrk_cntl_finalize(), FLA_Transpose_cntl_finalize(), FLA_Tridiag_UT_cntl_finalize(), FLA_Trinv_cntl_finalize(), FLA_Trmm_cntl_finalize(), FLA_Trsm_cntl_finalize(), FLA_Ttmm_cntl_finalize(), FLA_UDdate_UT_cntl_finalize(), FLASH_Apply_CAQ2_UT_cntl_finalize(), FLASH_Apply_CAQ_UT_inc_cntl_finalize(), FLASH_Apply_pivots_cntl_finalize(), FLASH_Apply_Q2_UT_cntl_finalize(), FLASH_Apply_Q_UT_cntl_finalize(), FLASH_Apply_Q_UT_inc_cntl_finalize(), FLASH_Apply_QUD_UT_cntl_finalize(), FLASH_Apply_QUD_UT_inc_cntl_finalize(), FLASH_Axpy_cntl_finalize(), FLASH_Axpyt_cntl_finalize(), FLASH_CAQR2_UT_cntl_finalize(), FLASH_CAQR_UT_inc_cntl_finalize(), FLASH_Chol_cntl_finalize(), FLASH_Copy_cntl_finalize(), FLASH_Copyr_cntl_finalize(), FLASH_Copyt_cntl_finalize(), FLASH_Eig_gest_cntl_finalize(), FLASH_Gemm_cntl_finalize(), FLASH_Gemv_cntl_finalize(), FLASH_Hemm_cntl_finalize(), FLASH_Her2k_cntl_finalize(), FLASH_Herk_cntl_finalize(), FLASH_LQ_UT_cntl_finalize(), FLASH_LU_incpiv_cntl_finalize(), FLASH_LU_nopiv_cntl_finalize(), FLASH_LU_piv_cntl_finalize(), FLASH_Lyap_cntl_finalize(), FLASH_QR2_UT_cntl_finalize(), FLASH_QR_UT_cntl_finalize(), FLASH_QR_UT_inc_cntl_finalize(), FLASH_Scal_cntl_finalize(), FLASH_Scalr_cntl_finalize(), FLASH_SPDinv_cntl_finalize(), FLASH_Sylv_cntl_finalize(), FLASH_Symm_cntl_finalize(), FLASH_Syr2k_cntl_finalize(), FLASH_Syrk_cntl_finalize(), FLASH_Trinv_cntl_finalize(), FLASH_Trmm_cntl_finalize(), FLASH_Trsm_cntl_finalize(), FLASH_Trsv_cntl_finalize(), FLASH_Ttmm_cntl_finalize(), FLASH_UDdate_UT_cntl_finalize(), and FLASH_UDdate_UT_inc_cntl_finalize().

FLA_Blocksize_scale()

void FLA_Blocksize_scale	(	fla_blocksize_t *	bp,
		double	factor
	)

{
    FLA_Error e_val;
    
    // Verify that the given blocksize pointer is valid.
    if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
    {
        e_val = FLA_Check_null_pointer( bp );
        FLA_Check_error_code( e_val );
    }
    
    // Assign the provided blocksize values into the corresponding fields.
    bp->s = ( dim_t )( ( double ) bp->s * factor );
    bp->d = ( dim_t )( ( double ) bp->d * factor );
    bp->c = ( dim_t )( ( double ) bp->c * factor );
    bp->z = ( dim_t )( ( double ) bp->z * factor );
}

References FLA_Blocksize_s::c, FLA_Blocksize_s::d, FLA_Check_error_level(), FLA_Check_null_pointer(), FLA_Blocksize_s::s, and FLA_Blocksize_s::z.

Referenced by FLA_Apply_CAQ2_UT_cntl_init(), FLA_Apply_Q2_UT_cntl_init(), FLA_Apply_Q_UT_cntl_init(), FLA_Apply_QUD_UT_cntl_init(), FLA_Bidiag_UT_cntl_init(), FLA_CAQR2_UT_cntl_init(), FLA_Chol_cntl_init(), FLA_Hess_UT_cntl_init(), FLA_LQ_UT_cntl_init(), FLA_LU_nopiv_cntl_init(), FLA_LU_piv_cntl_init(), FLA_QR2_UT_cntl_init(), FLA_QR_UT_cntl_init(), FLA_Tridiag_UT_cntl_init(), and FLA_UDdate_UT_cntl_init().

FLA_Blocksize_set()

void FLA_Blocksize_set	(	fla_blocksize_t *	bp,
		dim_t	b_s,
		dim_t	b_d,
		dim_t	b_c,
		dim_t	b_z
	)

{
    // Assign the provided blocksize values into the corresponding fields.
    if ( b_s != 0 ) bp->s = b_s;
    if ( b_d != 0 ) bp->d = b_d;
    if ( b_c != 0 ) bp->c = b_c;
    if ( b_z != 0 ) bp->z = b_z;
}

References FLA_Blocksize_s::c, FLA_Blocksize_s::d, FLA_Blocksize_s::s, and FLA_Blocksize_s::z.

Referenced by FLASH_Apply_Q_UT().

FLA_Determine_blocksize()

dim_t FLA_Determine_blocksize	(	FLA_Obj	A_unproc,
		FLA_Quadrant	to_dir,
		fla_blocksize_t *	bp
	)

{
    FLA_Error    e_val;
    FLA_Datatype datatype;
    dim_t        A_unproc_size;
    dim_t        typed_blocksize;
    dim_t        b;
 
    // Determine the size of the remaining portion of the matrix.
    A_unproc_size = FLA_determine_matrix_size( A_unproc, to_dir );
    
    // Determine the datatype of the matrix.
    datatype = FLA_Obj_datatype( A_unproc );
    
    // Determine the raw blocksize value from the blocksize structure.
    typed_blocksize = FLA_Blocksize_extract( datatype, bp );
 
    // Check blocksize for zero value.
    if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
    {
        e_val = FLA_Check_blocksize_value( typed_blocksize );
        FLA_Check_error_code( e_val );
    }
 
    // If the unprocessed partition is smaller than our blocksize allows,
    // we have to use it's length/width instead.
    b = min( A_unproc_size, typed_blocksize );
    
    // Return the computed blocksize.
    return b;
}

References FLA_Blocksize_extract(), FLA_Check_blocksize_value(), FLA_Check_error_level(), FLA_determine_matrix_size(), and FLA_Obj_datatype().

Referenced by FLA_Apply_CAQ2_UT_lhfc_blk_var2(), FLA_Apply_CAQ2_UT_lhfc_blk_var3(), FLA_Apply_CAQ_UT_inc_lhfc_blk_var1(), FLA_Apply_pivots_ln_blk_var1(), FLA_Apply_pivots_ln_blk_var2(), FLA_Apply_Q2_UT_lhfc_blk_var2(), FLA_Apply_Q2_UT_lhfc_blk_var3(), FLA_Apply_Q2_UT_lnfc_blk_var2(), FLA_Apply_Q2_UT_lnfc_blk_var3(), FLA_Apply_Q_UT_inc_lhfc_blk_var1(), FLA_Apply_Q_UT_inc_lnfc_blk_var1(), FLA_Apply_Q_UT_lhbc_blk_var2(), FLA_Apply_Q_UT_lhbc_blk_var3(), FLA_Apply_Q_UT_lhbr_blk_var2(), FLA_Apply_Q_UT_lhbr_blk_var3(), FLA_Apply_Q_UT_lhfc_blk_var2(), FLA_Apply_Q_UT_lhfc_blk_var3(), FLA_Apply_Q_UT_lhfr_blk_var2(), FLA_Apply_Q_UT_lhfr_blk_var3(), FLA_Apply_Q_UT_lnbc_blk_var2(), FLA_Apply_Q_UT_lnbc_blk_var3(), FLA_Apply_Q_UT_lnbr_blk_var2(), FLA_Apply_Q_UT_lnbr_blk_var3(), FLA_Apply_Q_UT_lnfc_blk_var2(), FLA_Apply_Q_UT_lnfc_blk_var3(), FLA_Apply_Q_UT_lnfr_blk_var2(), FLA_Apply_Q_UT_lnfr_blk_var3(), FLA_Apply_Q_UT_rhbc_blk_var2(), FLA_Apply_Q_UT_rhbc_blk_var3(), FLA_Apply_Q_UT_rhbr_blk_var2(), FLA_Apply_Q_UT_rhbr_blk_var3(), FLA_Apply_Q_UT_rhfc_blk_var2(), FLA_Apply_Q_UT_rhfc_blk_var3(), FLA_Apply_Q_UT_rhfr_blk_var2(), FLA_Apply_Q_UT_rhfr_blk_var3(), FLA_Apply_Q_UT_rnbc_blk_var2(), FLA_Apply_Q_UT_rnbc_blk_var3(), FLA_Apply_Q_UT_rnbr_blk_var2(), FLA_Apply_Q_UT_rnbr_blk_var3(), FLA_Apply_Q_UT_rnfc_blk_var2(), FLA_Apply_Q_UT_rnfc_blk_var3(), FLA_Apply_Q_UT_rnfr_blk_var2(), FLA_Apply_Q_UT_rnfr_blk_var3(), FLA_Apply_QUD_UT_inc_lhfc_blk_var1(), FLA_Apply_QUD_UT_lhfc_blk_var2(), FLA_Apply_QUD_UT_lhfc_blk_var3(), FLA_Axpy_blk_var1(), FLA_Axpy_blk_var2(), FLA_Axpy_blk_var3(), FLA_Axpy_blk_var4(), FLA_Axpyt_c_blk_var1(), FLA_Axpyt_c_blk_var2(), FLA_Axpyt_c_blk_var3(), FLA_Axpyt_c_blk_var4(), FLA_Axpyt_h_blk_var1(), FLA_Axpyt_h_blk_var2(), FLA_Axpyt_h_blk_var3(), FLA_Axpyt_h_blk_var4(), FLA_Axpyt_n_blk_var1(), FLA_Axpyt_n_blk_var2(), FLA_Axpyt_n_blk_var3(), FLA_Axpyt_n_blk_var4(), FLA_Axpyt_t_blk_var1(), FLA_Axpyt_t_blk_var2(), FLA_Axpyt_t_blk_var3(), FLA_Axpyt_t_blk_var4(), FLA_CAQR2_UT_blk_var2(), FLA_CAQR_UT_inc_blk_var1(), FLA_Chol_l_blk_var1(), FLA_Chol_l_blk_var2(), FLA_Chol_l_blk_var3(), FLA_Chol_u_blk_var1(), FLA_Chol_u_blk_var2(), FLA_Chol_u_blk_var3(), FLA_Copy_blk_var1(), FLA_Copy_blk_var2(), FLA_Copy_blk_var3(), FLA_Copy_blk_var4(), FLA_Copyr_l_blk_var1(), FLA_Copyr_l_blk_var2(), FLA_Copyr_l_blk_var3(), FLA_Copyr_l_blk_var4(), FLA_Copyr_u_blk_var1(), FLA_Copyr_u_blk_var2(), FLA_Copyr_u_blk_var3(), FLA_Copyr_u_blk_var4(), FLA_Copyt_c_blk_var1(), FLA_Copyt_c_blk_var2(), FLA_Copyt_c_blk_var3(), FLA_Copyt_c_blk_var4(), FLA_Copyt_h_blk_var1(), FLA_Copyt_h_blk_var2(), FLA_Copyt_h_blk_var3(), FLA_Copyt_h_blk_var4(), FLA_Copyt_n_blk_var1(), FLA_Copyt_n_blk_var2(), FLA_Copyt_n_blk_var3(), FLA_Copyt_n_blk_var4(), FLA_Copyt_t_blk_var1(), FLA_Copyt_t_blk_var2(), FLA_Copyt_t_blk_var3(), FLA_Copyt_t_blk_var4(), FLA_Eig_gest_il_blk_var1(), FLA_Eig_gest_il_blk_var2(), FLA_Eig_gest_il_blk_var3(), FLA_Eig_gest_il_blk_var4(), FLA_Eig_gest_il_blk_var5(), FLA_Eig_gest_iu_blk_var1(), FLA_Eig_gest_iu_blk_var2(), FLA_Eig_gest_iu_blk_var3(), FLA_Eig_gest_iu_blk_var4(), FLA_Eig_gest_iu_blk_var5(), FLA_Eig_gest_nl_blk_var1(), FLA_Eig_gest_nl_blk_var2(), FLA_Eig_gest_nl_blk_var4(), FLA_Eig_gest_nl_blk_var5(), FLA_Eig_gest_nu_blk_var1(), FLA_Eig_gest_nu_blk_var2(), FLA_Eig_gest_nu_blk_var4(), FLA_Eig_gest_nu_blk_var5(), FLA_Gemm_cc_blk_var1(), FLA_Gemm_cc_blk_var2(), FLA_Gemm_cc_blk_var3(), FLA_Gemm_cc_blk_var4(), FLA_Gemm_cc_blk_var5(), FLA_Gemm_cc_blk_var6(), FLA_Gemm_ch_blk_var1(), FLA_Gemm_ch_blk_var2(), FLA_Gemm_ch_blk_var3(), FLA_Gemm_ch_blk_var4(), FLA_Gemm_ch_blk_var5(), FLA_Gemm_ch_blk_var6(), FLA_Gemm_cn_blk_var1(), FLA_Gemm_cn_blk_var2(), FLA_Gemm_cn_blk_var3(), FLA_Gemm_cn_blk_var4(), FLA_Gemm_cn_blk_var5(), FLA_Gemm_cn_blk_var6(), FLA_Gemm_ct_blk_var1(), FLA_Gemm_ct_blk_var2(), FLA_Gemm_ct_blk_var3(), FLA_Gemm_ct_blk_var4(), FLA_Gemm_ct_blk_var5(), FLA_Gemm_ct_blk_var6(), FLA_Gemm_hc_blk_var1(), FLA_Gemm_hc_blk_var2(), FLA_Gemm_hc_blk_var3(), FLA_Gemm_hc_blk_var4(), FLA_Gemm_hc_blk_var5(), FLA_Gemm_hc_blk_var6(), FLA_Gemm_hh_blk_var1(), FLA_Gemm_hh_blk_var2(), FLA_Gemm_hh_blk_var3(), FLA_Gemm_hh_blk_var4(), FLA_Gemm_hh_blk_var5(), FLA_Gemm_hh_blk_var6(), FLA_Gemm_hn_blk_var1(), FLA_Gemm_hn_blk_var2(), FLA_Gemm_hn_blk_var3(), FLA_Gemm_hn_blk_var4(), FLA_Gemm_hn_blk_var5(), FLA_Gemm_hn_blk_var6(), FLA_Gemm_ht_blk_var1(), FLA_Gemm_ht_blk_var2(), FLA_Gemm_ht_blk_var3(), FLA_Gemm_ht_blk_var4(), FLA_Gemm_ht_blk_var5(), FLA_Gemm_ht_blk_var6(), FLA_Gemm_nc_blk_var1(), FLA_Gemm_nc_blk_var2(), FLA_Gemm_nc_blk_var3(), FLA_Gemm_nc_blk_var4(), FLA_Gemm_nc_blk_var5(), FLA_Gemm_nc_blk_var6(), FLA_Gemm_nh_blk_var1(), FLA_Gemm_nh_blk_var2(), FLA_Gemm_nh_blk_var3(), FLA_Gemm_nh_blk_var4(), FLA_Gemm_nh_blk_var5(), FLA_Gemm_nh_blk_var6(), FLA_Gemm_nn_blk_var1(), FLA_Gemm_nn_blk_var2(), FLA_Gemm_nn_blk_var3(), FLA_Gemm_nn_blk_var4(), FLA_Gemm_nn_blk_var5(), FLA_Gemm_nn_blk_var6(), FLA_Gemm_nt_blk_var1(), FLA_Gemm_nt_blk_var2(), FLA_Gemm_nt_blk_var3(), FLA_Gemm_nt_blk_var4(), FLA_Gemm_nt_blk_var5(), FLA_Gemm_nt_blk_var6(), FLA_Gemm_tc_blk_var1(), FLA_Gemm_tc_blk_var2(), FLA_Gemm_tc_blk_var3(), FLA_Gemm_tc_blk_var4(), FLA_Gemm_tc_blk_var5(), FLA_Gemm_tc_blk_var6(), FLA_Gemm_th_blk_var1(), FLA_Gemm_th_blk_var2(), FLA_Gemm_th_blk_var3(), FLA_Gemm_th_blk_var4(), FLA_Gemm_th_blk_var5(), FLA_Gemm_th_blk_var6(), FLA_Gemm_tn_blk_var1(), FLA_Gemm_tn_blk_var2(), FLA_Gemm_tn_blk_var3(), FLA_Gemm_tn_blk_var4(), FLA_Gemm_tn_blk_var5(), FLA_Gemm_tn_blk_var6(), FLA_Gemm_tt_blk_var1(), FLA_Gemm_tt_blk_var2(), FLA_Gemm_tt_blk_var3(), FLA_Gemm_tt_blk_var4(), FLA_Gemm_tt_blk_var5(), FLA_Gemm_tt_blk_var6(), FLA_Gemv_h_blk_var1(), FLA_Gemv_h_blk_var2(), FLA_Gemv_h_blk_var5(), FLA_Gemv_h_blk_var6(), FLA_Gemv_n_blk_var1(), FLA_Gemv_n_blk_var2(), FLA_Gemv_n_blk_var5(), FLA_Gemv_n_blk_var6(), FLA_Gemv_t_blk_var1(), FLA_Gemv_t_blk_var2(), FLA_Gemv_t_blk_var5(), FLA_Gemv_t_blk_var6(), FLA_Hemm_ll_blk_var1(), FLA_Hemm_ll_blk_var10(), FLA_Hemm_ll_blk_var2(), FLA_Hemm_ll_blk_var3(), FLA_Hemm_ll_blk_var4(), FLA_Hemm_ll_blk_var5(), FLA_Hemm_ll_blk_var6(), FLA_Hemm_ll_blk_var7(), FLA_Hemm_ll_blk_var8(), FLA_Hemm_ll_blk_var9(), FLA_Hemm_lu_blk_var1(), FLA_Hemm_lu_blk_var10(), FLA_Hemm_lu_blk_var2(), FLA_Hemm_lu_blk_var3(), FLA_Hemm_lu_blk_var4(), FLA_Hemm_lu_blk_var5(), FLA_Hemm_lu_blk_var6(), FLA_Hemm_lu_blk_var7(), FLA_Hemm_lu_blk_var8(), FLA_Hemm_lu_blk_var9(), FLA_Hemm_rl_blk_var1(), FLA_Hemm_rl_blk_var10(), FLA_Hemm_rl_blk_var2(), FLA_Hemm_rl_blk_var3(), FLA_Hemm_rl_blk_var4(), FLA_Hemm_rl_blk_var5(), FLA_Hemm_rl_blk_var6(), FLA_Hemm_rl_blk_var7(), FLA_Hemm_rl_blk_var8(), FLA_Hemm_rl_blk_var9(), FLA_Hemm_ru_blk_var1(), FLA_Hemm_ru_blk_var10(), FLA_Hemm_ru_blk_var2(), FLA_Hemm_ru_blk_var3(), FLA_Hemm_ru_blk_var4(), FLA_Hemm_ru_blk_var5(), FLA_Hemm_ru_blk_var6(), FLA_Hemm_ru_blk_var7(), FLA_Hemm_ru_blk_var8(), FLA_Hemm_ru_blk_var9(), FLA_Her2k_lh_blk_var1(), FLA_Her2k_lh_blk_var10(), FLA_Her2k_lh_blk_var2(), FLA_Her2k_lh_blk_var3(), FLA_Her2k_lh_blk_var4(), FLA_Her2k_lh_blk_var5(), FLA_Her2k_lh_blk_var6(), FLA_Her2k_lh_blk_var7(), FLA_Her2k_lh_blk_var8(), FLA_Her2k_lh_blk_var9(), FLA_Her2k_ln_blk_var1(), FLA_Her2k_ln_blk_var10(), FLA_Her2k_ln_blk_var2(), FLA_Her2k_ln_blk_var3(), FLA_Her2k_ln_blk_var4(), FLA_Her2k_ln_blk_var5(), FLA_Her2k_ln_blk_var6(), FLA_Her2k_ln_blk_var7(), FLA_Her2k_ln_blk_var8(), FLA_Her2k_ln_blk_var9(), FLA_Her2k_uh_blk_var1(), FLA_Her2k_uh_blk_var10(), FLA_Her2k_uh_blk_var2(), FLA_Her2k_uh_blk_var3(), FLA_Her2k_uh_blk_var4(), FLA_Her2k_uh_blk_var5(), FLA_Her2k_uh_blk_var6(), FLA_Her2k_uh_blk_var7(), FLA_Her2k_uh_blk_var8(), FLA_Her2k_uh_blk_var9(), FLA_Her2k_un_blk_var1(), FLA_Her2k_un_blk_var10(), FLA_Her2k_un_blk_var2(), FLA_Her2k_un_blk_var3(), FLA_Her2k_un_blk_var4(), FLA_Her2k_un_blk_var5(), FLA_Her2k_un_blk_var6(), FLA_Her2k_un_blk_var7(), FLA_Her2k_un_blk_var8(), FLA_Her2k_un_blk_var9(), FLA_Herk_lh_blk_var1(), FLA_Herk_lh_blk_var2(), FLA_Herk_lh_blk_var3(), FLA_Herk_lh_blk_var4(), FLA_Herk_lh_blk_var5(), FLA_Herk_lh_blk_var6(), FLA_Herk_ln_blk_var1(), FLA_Herk_ln_blk_var2(), FLA_Herk_ln_blk_var3(), FLA_Herk_ln_blk_var4(), FLA_Herk_ln_blk_var5(), FLA_Herk_ln_blk_var6(), FLA_Herk_uh_blk_var1(), FLA_Herk_uh_blk_var2(), FLA_Herk_uh_blk_var3(), FLA_Herk_uh_blk_var4(), FLA_Herk_uh_blk_var5(), FLA_Herk_uh_blk_var6(), FLA_Herk_un_blk_var1(), FLA_Herk_un_blk_var2(), FLA_Herk_un_blk_var3(), FLA_Herk_un_blk_var4(), FLA_Herk_un_blk_var5(), FLA_Herk_un_blk_var6(), FLA_LQ_UT_blk_var2(), FLA_LQ_UT_blk_var3(), FLA_LU_nopiv_blk_var1(), FLA_LU_nopiv_blk_var2(), FLA_LU_nopiv_blk_var3(), FLA_LU_nopiv_blk_var4(), FLA_LU_nopiv_blk_var5(), FLA_LU_piv_blk_var3(), FLA_LU_piv_blk_var4(), FLA_LU_piv_blk_var5(), FLA_Lyap_h_blk_var1(), FLA_Lyap_h_blk_var2(), FLA_Lyap_h_blk_var3(), FLA_Lyap_h_blk_var4(), FLA_Lyap_n_blk_var1(), FLA_Lyap_n_blk_var2(), FLA_Lyap_n_blk_var3(), FLA_Lyap_n_blk_var4(), FLA_QR2_UT_blk_var2(), FLA_QR_UT_blk_var2(), FLA_QR_UT_blk_var3(), FLA_QR_UT_inc_blk_var1(), FLA_QR_UT_inc_blk_var2(), FLA_Scal_blk_var1(), FLA_Scal_blk_var2(), FLA_Scal_blk_var3(), FLA_Scal_blk_var4(), FLA_Scalr_l_blk_var1(), FLA_Scalr_l_blk_var2(), FLA_Scalr_l_blk_var3(), FLA_Scalr_l_blk_var4(), FLA_Scalr_u_blk_var1(), FLA_Scalr_u_blk_var2(), FLA_Scalr_u_blk_var3(), FLA_Scalr_u_blk_var4(), FLA_Swap_t_blk_var1(), FLA_Swap_t_blk_var2(), FLA_Sylv_hh_blk_var1(), FLA_Sylv_hh_blk_var10(), FLA_Sylv_hh_blk_var11(), FLA_Sylv_hh_blk_var12(), FLA_Sylv_hh_blk_var13(), FLA_Sylv_hh_blk_var14(), FLA_Sylv_hh_blk_var15(), FLA_Sylv_hh_blk_var16(), FLA_Sylv_hh_blk_var17(), FLA_Sylv_hh_blk_var18(), FLA_Sylv_hh_blk_var2(), FLA_Sylv_hh_blk_var3(), FLA_Sylv_hh_blk_var4(), FLA_Sylv_hh_blk_var5(), FLA_Sylv_hh_blk_var6(), FLA_Sylv_hh_blk_var7(), FLA_Sylv_hh_blk_var8(), FLA_Sylv_hh_blk_var9(), FLA_Sylv_hn_blk_var1(), FLA_Sylv_hn_blk_var10(), FLA_Sylv_hn_blk_var11(), FLA_Sylv_hn_blk_var12(), FLA_Sylv_hn_blk_var13(), FLA_Sylv_hn_blk_var14(), FLA_Sylv_hn_blk_var15(), FLA_Sylv_hn_blk_var16(), FLA_Sylv_hn_blk_var17(), FLA_Sylv_hn_blk_var18(), FLA_Sylv_hn_blk_var2(), FLA_Sylv_hn_blk_var3(), FLA_Sylv_hn_blk_var4(), FLA_Sylv_hn_blk_var5(), FLA_Sylv_hn_blk_var6(), FLA_Sylv_hn_blk_var7(), FLA_Sylv_hn_blk_var8(), FLA_Sylv_hn_blk_var9(), FLA_Sylv_nh_blk_var1(), FLA_Sylv_nh_blk_var10(), FLA_Sylv_nh_blk_var11(), FLA_Sylv_nh_blk_var12(), FLA_Sylv_nh_blk_var13(), FLA_Sylv_nh_blk_var14(), FLA_Sylv_nh_blk_var15(), FLA_Sylv_nh_blk_var16(), FLA_Sylv_nh_blk_var17(), FLA_Sylv_nh_blk_var18(), FLA_Sylv_nh_blk_var2(), FLA_Sylv_nh_blk_var3(), FLA_Sylv_nh_blk_var4(), FLA_Sylv_nh_blk_var5(), FLA_Sylv_nh_blk_var6(), FLA_Sylv_nh_blk_var7(), FLA_Sylv_nh_blk_var8(), FLA_Sylv_nh_blk_var9(), FLA_Sylv_nn_blk_var1(), FLA_Sylv_nn_blk_var10(), FLA_Sylv_nn_blk_var11(), FLA_Sylv_nn_blk_var12(), FLA_Sylv_nn_blk_var13(), FLA_Sylv_nn_blk_var14(), FLA_Sylv_nn_blk_var15(), FLA_Sylv_nn_blk_var16(), FLA_Sylv_nn_blk_var17(), FLA_Sylv_nn_blk_var18(), FLA_Sylv_nn_blk_var2(), FLA_Sylv_nn_blk_var3(), FLA_Sylv_nn_blk_var4(), FLA_Sylv_nn_blk_var5(), FLA_Sylv_nn_blk_var6(), FLA_Sylv_nn_blk_var7(), FLA_Sylv_nn_blk_var8(), FLA_Sylv_nn_blk_var9(), FLA_Symm_ll_blk_var1(), FLA_Symm_ll_blk_var10(), FLA_Symm_ll_blk_var2(), FLA_Symm_ll_blk_var3(), FLA_Symm_ll_blk_var4(), FLA_Symm_ll_blk_var5(), FLA_Symm_ll_blk_var6(), FLA_Symm_ll_blk_var7(), FLA_Symm_ll_blk_var8(), FLA_Symm_ll_blk_var9(), FLA_Symm_lu_blk_var1(), FLA_Symm_lu_blk_var10(), FLA_Symm_lu_blk_var2(), FLA_Symm_lu_blk_var3(), FLA_Symm_lu_blk_var4(), FLA_Symm_lu_blk_var5(), FLA_Symm_lu_blk_var6(), FLA_Symm_lu_blk_var7(), FLA_Symm_lu_blk_var8(), FLA_Symm_lu_blk_var9(), FLA_Symm_rl_blk_var1(), FLA_Symm_rl_blk_var10(), FLA_Symm_rl_blk_var2(), FLA_Symm_rl_blk_var3(), FLA_Symm_rl_blk_var4(), FLA_Symm_rl_blk_var5(), FLA_Symm_rl_blk_var6(), FLA_Symm_rl_blk_var7(), FLA_Symm_rl_blk_var8(), FLA_Symm_rl_blk_var9(), FLA_Symm_ru_blk_var1(), FLA_Symm_ru_blk_var10(), FLA_Symm_ru_blk_var2(), FLA_Symm_ru_blk_var3(), FLA_Symm_ru_blk_var4(), FLA_Symm_ru_blk_var5(), FLA_Symm_ru_blk_var6(), FLA_Symm_ru_blk_var7(), FLA_Symm_ru_blk_var8(), FLA_Symm_ru_blk_var9(), FLA_Syr2k_ln_blk_var1(), FLA_Syr2k_ln_blk_var10(), FLA_Syr2k_ln_blk_var2(), FLA_Syr2k_ln_blk_var3(), FLA_Syr2k_ln_blk_var4(), FLA_Syr2k_ln_blk_var5(), FLA_Syr2k_ln_blk_var6(), FLA_Syr2k_ln_blk_var7(), FLA_Syr2k_ln_blk_var8(), FLA_Syr2k_ln_blk_var9(), FLA_Syr2k_lt_blk_var1(), FLA_Syr2k_lt_blk_var10(), FLA_Syr2k_lt_blk_var2(), FLA_Syr2k_lt_blk_var3(), FLA_Syr2k_lt_blk_var4(), FLA_Syr2k_lt_blk_var5(), FLA_Syr2k_lt_blk_var6(), FLA_Syr2k_lt_blk_var7(), FLA_Syr2k_lt_blk_var8(), FLA_Syr2k_lt_blk_var9(), FLA_Syr2k_un_blk_var1(), FLA_Syr2k_un_blk_var10(), FLA_Syr2k_un_blk_var2(), FLA_Syr2k_un_blk_var3(), FLA_Syr2k_un_blk_var4(), FLA_Syr2k_un_blk_var5(), FLA_Syr2k_un_blk_var6(), FLA_Syr2k_un_blk_var7(), FLA_Syr2k_un_blk_var8(), FLA_Syr2k_un_blk_var9(), FLA_Syr2k_ut_blk_var1(), FLA_Syr2k_ut_blk_var10(), FLA_Syr2k_ut_blk_var2(), FLA_Syr2k_ut_blk_var3(), FLA_Syr2k_ut_blk_var4(), FLA_Syr2k_ut_blk_var5(), FLA_Syr2k_ut_blk_var6(), FLA_Syr2k_ut_blk_var7(), FLA_Syr2k_ut_blk_var8(), FLA_Syr2k_ut_blk_var9(), FLA_Syrk_ln_blk_var1(), FLA_Syrk_ln_blk_var2(), FLA_Syrk_ln_blk_var3(), FLA_Syrk_ln_blk_var4(), FLA_Syrk_ln_blk_var5(), FLA_Syrk_ln_blk_var6(), FLA_Syrk_lt_blk_var1(), FLA_Syrk_lt_blk_var2(), FLA_Syrk_lt_blk_var3(), FLA_Syrk_lt_blk_var4(), FLA_Syrk_lt_blk_var5(), FLA_Syrk_lt_blk_var6(), FLA_Syrk_un_blk_var1(), FLA_Syrk_un_blk_var2(), FLA_Syrk_un_blk_var3(), FLA_Syrk_un_blk_var4(), FLA_Syrk_un_blk_var5(), FLA_Syrk_un_blk_var6(), FLA_Syrk_ut_blk_var1(), FLA_Syrk_ut_blk_var2(), FLA_Syrk_ut_blk_var3(), FLA_Syrk_ut_blk_var4(), FLA_Syrk_ut_blk_var5(), FLA_Syrk_ut_blk_var6(), FLA_Transpose_blk_var1(), FLA_Transpose_blk_var2(), FLA_Trinv_ln_blk_var1(), FLA_Trinv_ln_blk_var2(), FLA_Trinv_ln_blk_var3(), FLA_Trinv_ln_blk_var4(), FLA_Trinv_lu_blk_var1(), FLA_Trinv_lu_blk_var2(), FLA_Trinv_lu_blk_var3(), FLA_Trinv_lu_blk_var4(), FLA_Trinv_un_blk_var1(), FLA_Trinv_un_blk_var2(), FLA_Trinv_un_blk_var3(), FLA_Trinv_un_blk_var4(), FLA_Trinv_uu_blk_var1(), FLA_Trinv_uu_blk_var2(), FLA_Trinv_uu_blk_var3(), FLA_Trinv_uu_blk_var4(), FLA_Trmm_llc_blk_var1(), FLA_Trmm_llc_blk_var2(), FLA_Trmm_llc_blk_var3(), FLA_Trmm_llc_blk_var4(), FLA_Trmm_llh_blk_var1(), FLA_Trmm_llh_blk_var2(), FLA_Trmm_llh_blk_var3(), FLA_Trmm_llh_blk_var4(), FLA_Trmm_lln_blk_var1(), FLA_Trmm_lln_blk_var2(), FLA_Trmm_lln_blk_var3(), FLA_Trmm_lln_blk_var4(), FLA_Trmm_llt_blk_var1(), FLA_Trmm_llt_blk_var2(), FLA_Trmm_llt_blk_var3(), FLA_Trmm_llt_blk_var4(), FLA_Trmm_luc_blk_var1(), FLA_Trmm_luc_blk_var2(), FLA_Trmm_luc_blk_var3(), FLA_Trmm_luc_blk_var4(), FLA_Trmm_luh_blk_var1(), FLA_Trmm_luh_blk_var2(), FLA_Trmm_luh_blk_var3(), FLA_Trmm_luh_blk_var4(), FLA_Trmm_lun_blk_var1(), FLA_Trmm_lun_blk_var2(), FLA_Trmm_lun_blk_var3(), FLA_Trmm_lun_blk_var4(), FLA_Trmm_lut_blk_var1(), FLA_Trmm_lut_blk_var2(), FLA_Trmm_lut_blk_var3(), FLA_Trmm_lut_blk_var4(), FLA_Trmm_rlc_blk_var1(), FLA_Trmm_rlc_blk_var2(), FLA_Trmm_rlc_blk_var3(), FLA_Trmm_rlc_blk_var4(), FLA_Trmm_rlh_blk_var1(), FLA_Trmm_rlh_blk_var2(), FLA_Trmm_rlh_blk_var3(), FLA_Trmm_rlh_blk_var4(), FLA_Trmm_rln_blk_var1(), FLA_Trmm_rln_blk_var2(), FLA_Trmm_rln_blk_var3(), FLA_Trmm_rln_blk_var4(), FLA_Trmm_rlt_blk_var1(), FLA_Trmm_rlt_blk_var2(), FLA_Trmm_rlt_blk_var3(), FLA_Trmm_rlt_blk_var4(), FLA_Trmm_ruc_blk_var1(), FLA_Trmm_ruc_blk_var2(), FLA_Trmm_ruc_blk_var3(), FLA_Trmm_ruc_blk_var4(), FLA_Trmm_ruh_blk_var1(), FLA_Trmm_ruh_blk_var2(), FLA_Trmm_ruh_blk_var3(), FLA_Trmm_ruh_blk_var4(), FLA_Trmm_run_blk_var1(), FLA_Trmm_run_blk_var2(), FLA_Trmm_run_blk_var3(), FLA_Trmm_run_blk_var4(), FLA_Trmm_rut_blk_var1(), FLA_Trmm_rut_blk_var2(), FLA_Trmm_rut_blk_var3(), FLA_Trmm_rut_blk_var4(), FLA_Trsm_llc_blk_var1(), FLA_Trsm_llc_blk_var2(), FLA_Trsm_llc_blk_var3(), FLA_Trsm_llc_blk_var4(), FLA_Trsm_llh_blk_var1(), FLA_Trsm_llh_blk_var2(), FLA_Trsm_llh_blk_var3(), FLA_Trsm_llh_blk_var4(), FLA_Trsm_lln_blk_var1(), FLA_Trsm_lln_blk_var2(), FLA_Trsm_lln_blk_var3(), FLA_Trsm_lln_blk_var4(), FLA_Trsm_llt_blk_var1(), FLA_Trsm_llt_blk_var2(), FLA_Trsm_llt_blk_var3(), FLA_Trsm_llt_blk_var4(), FLA_Trsm_luc_blk_var1(), FLA_Trsm_luc_blk_var2(), FLA_Trsm_luc_blk_var3(), FLA_Trsm_luc_blk_var4(), FLA_Trsm_luh_blk_var1(), FLA_Trsm_luh_blk_var2(), FLA_Trsm_luh_blk_var3(), FLA_Trsm_luh_blk_var4(), FLA_Trsm_lun_blk_var1(), FLA_Trsm_lun_blk_var2(), FLA_Trsm_lun_blk_var3(), FLA_Trsm_lun_blk_var4(), FLA_Trsm_lut_blk_var1(), FLA_Trsm_lut_blk_var2(), FLA_Trsm_lut_blk_var3(), FLA_Trsm_lut_blk_var4(), FLA_Trsm_rlc_blk_var1(), FLA_Trsm_rlc_blk_var2(), FLA_Trsm_rlc_blk_var3(), FLA_Trsm_rlc_blk_var4(), FLA_Trsm_rlh_blk_var1(), FLA_Trsm_rlh_blk_var2(), FLA_Trsm_rlh_blk_var3(), FLA_Trsm_rlh_blk_var4(), FLA_Trsm_rln_blk_var1(), FLA_Trsm_rln_blk_var2(), FLA_Trsm_rln_blk_var3(), FLA_Trsm_rln_blk_var4(), FLA_Trsm_rlt_blk_var1(), FLA_Trsm_rlt_blk_var2(), FLA_Trsm_rlt_blk_var3(), FLA_Trsm_rlt_blk_var4(), FLA_Trsm_ruc_blk_var1(), FLA_Trsm_ruc_blk_var2(), FLA_Trsm_ruc_blk_var3(), FLA_Trsm_ruc_blk_var4(), FLA_Trsm_ruh_blk_var1(), FLA_Trsm_ruh_blk_var2(), FLA_Trsm_ruh_blk_var3(), FLA_Trsm_ruh_blk_var4(), FLA_Trsm_run_blk_var1(), FLA_Trsm_run_blk_var2(), FLA_Trsm_run_blk_var3(), FLA_Trsm_run_blk_var4(), FLA_Trsm_rut_blk_var1(), FLA_Trsm_rut_blk_var2(), FLA_Trsm_rut_blk_var3(), FLA_Trsm_rut_blk_var4(), FLA_Trsv_lc_blk_var1(), FLA_Trsv_lc_blk_var2(), FLA_Trsv_ln_blk_var1(), FLA_Trsv_ln_blk_var2(), FLA_Trsv_lt_blk_var1(), FLA_Trsv_lt_blk_var2(), FLA_Trsv_uc_blk_var1(), FLA_Trsv_uc_blk_var2(), FLA_Trsv_un_blk_var1(), FLA_Trsv_un_blk_var2(), FLA_Trsv_ut_blk_var1(), FLA_Trsv_ut_blk_var2(), FLA_Ttmm_l_blk_var1(), FLA_Ttmm_l_blk_var2(), FLA_Ttmm_l_blk_var3(), FLA_Ttmm_u_blk_var1(), FLA_Ttmm_u_blk_var2(), FLA_Ttmm_u_blk_var3(), FLA_UDdate_UT_blk_var2(), and FLA_UDdate_UT_inc_blk_var1().

FLA_determine_matrix_size()

dim_t FLA_determine_matrix_size	(	FLA_Obj	A_unproc,
		FLA_Quadrant	to_dir
	)

{
    dim_t r_val = 0;
 
    // Determine the size of the matrix dimension along which we are moving.
    switch( to_dir )
    {
        case FLA_TOP:
        case FLA_BOTTOM:
        {
            r_val = FLA_Obj_length( A_unproc );
            break;
        }
        case FLA_LEFT:
        case FLA_RIGHT:
        {
            r_val = FLA_Obj_width( A_unproc );
            break;
        }
        case FLA_TL:
        case FLA_TR:
        case FLA_BL:
        case FLA_BR:
        {
            // We need to use min_dim() here because the matrix might be
            // rectangular.
            r_val = FLA_Obj_min_dim( A_unproc );
            break;
        }
    }
    
    return r_val;
}

References FLA_Obj_length(), FLA_Obj_min_dim(), and FLA_Obj_width().

Referenced by FLA_Determine_blocksize().

FLA_Query_blocksize()

dim_t FLA_Query_blocksize	(	FLA_Datatype	dt,
		FLA_Dimension	dim
	)

{
    dim_t b_val = 0;
 
#ifdef FLA_ENABLE_GOTO_INTERFACES
 
    int dt_index;
    int dim_index;
 
    if ( first_time )
    {
        long sgemm_min, dgemm_min, cgemm_min, zgemm_min;
        
        // Find the blocksizes associated with FLA_DIMENSION_MIN.
        sgemm_min = min( sgemm_p, sgemm_q );
        dgemm_min = min( dgemm_p, dgemm_q );
        cgemm_min = min( cgemm_p, cgemm_q );
        zgemm_min = min( zgemm_p, zgemm_q );
 
        // Set the values for each datatype and dimension constant.
        fla_goto_gemm_blocksize[FLA_S_INDEX][FLA_DIM_M_INDEX]   = sgemm_p;
        fla_goto_gemm_blocksize[FLA_S_INDEX][FLA_DIM_K_INDEX]   = sgemm_q;
        fla_goto_gemm_blocksize[FLA_S_INDEX][FLA_DIM_N_INDEX]   = sgemm_r;
        fla_goto_gemm_blocksize[FLA_S_INDEX][FLA_DIM_MIN_INDEX] = sgemm_min;
 
        fla_goto_gemm_blocksize[FLA_D_INDEX][FLA_DIM_M_INDEX]   = dgemm_p;
        fla_goto_gemm_blocksize[FLA_D_INDEX][FLA_DIM_K_INDEX]   = dgemm_q;
        fla_goto_gemm_blocksize[FLA_D_INDEX][FLA_DIM_N_INDEX]   = dgemm_r;
        fla_goto_gemm_blocksize[FLA_D_INDEX][FLA_DIM_MIN_INDEX] = dgemm_min;
 
        fla_goto_gemm_blocksize[FLA_C_INDEX][FLA_DIM_M_INDEX]   = cgemm_p;
        fla_goto_gemm_blocksize[FLA_C_INDEX][FLA_DIM_K_INDEX]   = cgemm_q;
        fla_goto_gemm_blocksize[FLA_C_INDEX][FLA_DIM_N_INDEX]   = cgemm_r;
        fla_goto_gemm_blocksize[FLA_C_INDEX][FLA_DIM_MIN_INDEX] = cgemm_min;
 
        fla_goto_gemm_blocksize[FLA_Z_INDEX][FLA_DIM_M_INDEX]   = zgemm_p;
        fla_goto_gemm_blocksize[FLA_Z_INDEX][FLA_DIM_K_INDEX]   = zgemm_q;
        fla_goto_gemm_blocksize[FLA_Z_INDEX][FLA_DIM_N_INDEX]   = zgemm_r;
        fla_goto_gemm_blocksize[FLA_Z_INDEX][FLA_DIM_MIN_INDEX] = zgemm_min;
 
        first_time = FALSE;
    }
 
    // Compute the index of the requested datatype.
    dt_index  = dt  & FLA_DTYPE_INDEX_MASK;
    dim_index = dim & FLA_DIM_INDEX_MASK;
 
    // Index into the array and choose the appropriate blocksize.
    b_val = ( dim_t ) fla_goto_gemm_blocksize[dt_index][dim_index];
 
#else
 
    // Assign the return value to a default sane blocksize in case
    // we cannot access the libgoto symbols.
    if      ( dim == FLA_DIMENSION_M )
        b_val = FLA_DEFAULT_M_BLOCKSIZE;
    else if ( dim == FLA_DIMENSION_K )
        b_val = FLA_DEFAULT_K_BLOCKSIZE;
    else if ( dim == FLA_DIMENSION_N )
        b_val = FLA_DEFAULT_N_BLOCKSIZE;
    else if ( dim == FLA_DIMENSION_MIN )
    {
        b_val = min( FLA_DEFAULT_M_BLOCKSIZE, FLA_DEFAULT_K_BLOCKSIZE );
        b_val = min( b_val, FLA_DEFAULT_N_BLOCKSIZE );
    }
 
#endif
 
    // Return the blocksize.
    return b_val;
}

References cgemm_p, cgemm_q, cgemm_r, dgemm_p, dgemm_q, dgemm_r, fla_goto_gemm_blocksize, sgemm_p, sgemm_q, sgemm_r, zgemm_p, zgemm_q, and zgemm_r.

Referenced by FLA_Apply_Q_blk_external(), FLA_Bidiag_blk_external(), FLA_Bidiag_UT_create_T(), FLA_Check_householder_panel_dims(), FLA_Hess_blk_external(), FLA_Hess_UT_create_T(), FLA_LQ_blk_external(), FLA_LQ_UT_create_T(), FLA_QR_blk_external(), FLA_QR_UT_create_T(), FLA_Query_blocksizes(), FLA_Tridiag_blk_external(), FLA_Tridiag_form_Q_external(), FLA_Tridiag_UT_create_T(), and FLA_UDdate_UT_create_T().

FLA_Query_blocksizes()

fla_blocksize_t * FLA_Query_blocksizes

(

FLA_Dimension

dim

)

{
    fla_blocksize_t* bp;
    
    // Create an FLA_Blocksize_t object.
    bp = FLA_Blocksize_create( 0, 0, 0, 0 );
    
    // Query the requested blocksize (m, k, or n dimension) for all of the
    // datatypes and package the results in an FLA_Blocksize_t structure.
    bp->s = FLA_Query_blocksize( FLA_FLOAT, dim );
    bp->d = FLA_Query_blocksize( FLA_DOUBLE, dim );
    bp->c = FLA_Query_blocksize( FLA_COMPLEX, dim );
    bp->z = FLA_Query_blocksize( FLA_DOUBLE_COMPLEX, dim );
    
    // Return a pointer to the structure.
    return bp;
}

References FLA_Blocksize_s::c, FLA_Blocksize_s::d, FLA_Blocksize_create(), FLA_Query_blocksize(), FLA_Blocksize_s::s, and FLA_Blocksize_s::z.

Referenced by FLA_Apply_CAQ2_UT_cntl_init(), FLA_Apply_Q2_UT_cntl_init(), FLA_Apply_Q_UT_cntl_init(), FLA_Apply_QUD_UT_cntl_init(), FLA_Bidiag_UT_cntl_init(), FLA_CAQR2_UT_cntl_init(), FLA_Chol_cntl_init(), FLA_Eig_gest_cntl_init(), FLA_Gemm_cntl_init(), FLA_Hemm_cntl_init(), FLA_Her2k_cntl_init(), FLA_Herk_cntl_init(), FLA_Hess_UT_cntl_init(), FLA_LQ_UT_cntl_init(), FLA_LU_nopiv_cntl_init(), FLA_LU_piv_cntl_init(), FLA_Lyap_cntl_init(), FLA_QR2_UT_cntl_init(), FLA_QR_UT_cntl_init(), FLA_SPDinv_cntl_init(), FLA_Sylv_cntl_init(), FLA_Symm_cntl_init(), FLA_Syr2k_cntl_init(), FLA_Syrk_cntl_init(), FLA_Transpose_cntl_init(), FLA_Tridiag_UT_cntl_init(), FLA_Trinv_cntl_init(), FLA_Trmm_cntl_init(), FLA_Trsm_cntl_init(), FLA_Ttmm_cntl_init(), FLA_UDdate_UT_cntl_init(), and FLASH_SPDinv_cntl_init().

Variable Documentation

cgemm_p

long cgemm_p

extern

Referenced by FLA_Query_blocksize().

cgemm_q

long cgemm_q

Referenced by FLA_Query_blocksize().

cgemm_r

long cgemm_r

Referenced by FLA_Query_blocksize().

dgemm_p

long dgemm_p

extern

Referenced by FLA_Query_blocksize().

dgemm_q

long dgemm_q

Referenced by FLA_Query_blocksize().

dgemm_r

long dgemm_r

Referenced by FLA_Query_blocksize().

fla_goto_gemm_blocksize

long fla_goto_gemm_blocksize[4][4]

Referenced by FLA_Query_blocksize().

sgemm_p

long sgemm_p

extern

Referenced by FLA_Query_blocksize().

sgemm_q

long sgemm_q

Referenced by FLA_Query_blocksize().

sgemm_r

long sgemm_r

Referenced by FLA_Query_blocksize().

zgemm_p

long zgemm_p

extern

Referenced by FLA_Query_blocksize().

zgemm_q

long zgemm_q

Referenced by FLA_Query_blocksize().

zgemm_r

long zgemm_r

Referenced by FLA_Query_blocksize().