|
libflame
revision_anchor
|
Functions | |
| FLA_Error | FLA_Herk_un_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t *cntl) |
| FLA_Error FLA_Herk_un_blk_var2 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_herk_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Herk_un().
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj CTL, CTR, C00, C01, C02,
CBL, CBR, C10, C11, C12,
C20, C21, C22;
dim_t b;
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_TL );
while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x1_to_3x1( AT, &A0,
/* ** */ /* ** */
&A1,
AB, &A2, b, FLA_BOTTOM );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
/* ************* */ /* ******************** */
&C10, /**/ &C11, &C12,
CBL, /**/ CBR, &C20, /**/ &C21, &C22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
/* C01 = C01 + A0 * A1' */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
alpha, A0, A1, beta, C01,
FLA_Cntl_sub_gemm( cntl ) );
/* C11 = C11 + A1 * A1' */
FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
alpha, A1, beta, C11,
FLA_Cntl_sub_herk( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
C10, C11, /**/ C12,
/* ************** */ /* ****************** */
&CBL, /**/ &CBR, C20, C21, /**/ C22,
FLA_TL );
}
return FLA_SUCCESS;
}
1.7.6.1