|
libflame
revision_anchor
|
Go to the source code of this file.
| FLA_Error FLA_Trinv_ln_blk_var1 | ( | FLA_Obj | A, |
| fla_trinv_t * | cntl | ||
| ) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), FLA_Trmm_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
dim_t b;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
// A10 = A10 * A00;
FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, A00, A10,
FLA_Cntl_sub_trmm( cntl ) );
// A10 = -tril( A11 ) \ A10;
FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_MINUS_ONE, A11, A10,
FLA_Cntl_sub_trsm1( cntl ) );
// A11 = inv( A11 );
FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A11,
FLA_Cntl_sub_trinv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_blk_var2 | ( | FLA_Obj | A, |
| fla_trinv_t * | cntl | ||
| ) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
dim_t b;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
// A21 = tril( A22 ) \ A21;
FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, A22, A21,
FLA_Cntl_sub_trsm1( cntl ) );
// A21 = -A21 / tril( A11 );
FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_MINUS_ONE, A11, A21,
FLA_Cntl_sub_trsm2( cntl ) );
// A11 = inv( A11 );
FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A11,
FLA_Cntl_sub_trinv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_blk_var3 | ( | FLA_Obj | A, |
| fla_trinv_t * | cntl | ||
| ) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
dim_t b;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
// A21 = -A21 / tril( A11 );
FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_MINUS_ONE, A11, A21,
FLA_Cntl_sub_trsm1( cntl ) );
// A20 = A21 * A10 + A20;
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, A21, A10, FLA_ONE, A20,
FLA_Cntl_sub_gemm( cntl ) );
// A10 = tril( A11 ) \ A10;
FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, A11, A10,
FLA_Cntl_sub_trsm2( cntl ) );
// A11 = inv( A11 );
FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A11,
FLA_Cntl_sub_trinv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_blk_var4 | ( | FLA_Obj | A, |
| fla_trinv_t * | cntl | ||
| ) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), FLA_Trmm_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
dim_t b;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
// A21 = -tril( A22 ) \ A21;
FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_MINUS_ONE, A22, A21,
FLA_Cntl_sub_trsm1( cntl ) );
// A20 = -A21 * A10 + A20;
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_MINUS_ONE, A21, A10, FLA_ONE, A20,
FLA_Cntl_sub_gemm( cntl ) );
// A10 = A10 * A00;
FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, A00, A10,
FLA_Cntl_sub_trmm( cntl ) );
// A11 = inv( A11 );
FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A11,
FLA_Cntl_sub_trinv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opc_var1 | ( | int | mn_A, |
| scomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_cinverts(), bli_cinvscalv(), bli_ctrmv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NONUNIT_DIAG, and BLIS_TRANSPOSE.
Referenced by FLA_Trinv_ln_opt_var1().
{
scomplex alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_ctrmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a10t );
// FLA_Inv_scal_external( alpha11, a10t );
bli_cneg2( alpha11, &alpha11_m1 );
bli_cinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
&alpha11_m1,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_cinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opc_var2 | ( | int | mn_A, |
| scomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_cinverts(), bli_cinvscalv(), bli_ctrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, and BLIS_NONUNIT_DIAG.
Referenced by FLA_Trinv_ln_opt_var2().
{
scomplex alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
/*------------------------------------------------------------*/
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_ctrsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_cneg2( alpha11, &alpha11_m1 );
bli_cinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_cinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opc_var3 | ( | int | mn_A, |
| scomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_cger(), bli_cinverts(), bli_cinvscalv(), BLIS_NO_CONJUGATE, and FLA_ONE.
Referenced by FLA_Trinv_ln_opt_var3().
{
scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
scomplex alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_cneg2( alpha11, &alpha11_m1 );
bli_cinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Ger_external( FLA_ONE, a21, a10t, A20 );
bli_cger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Inv_scal_external( alpha11, a10t );
bli_cinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
alpha11,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_cinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opc_var4 | ( | int | mn_A, |
| scomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_cger(), bli_cinverts(), bli_cscalv(), bli_ctrmv(), bli_ctrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{
scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < mn_A; ++i )
{
scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_cscalv( BLIS_NO_CONJUGATE,
mn_ahead,
buff_m1,
a21, rs_A );
bli_ctrsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 );
bli_cger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_m1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_ctrmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_cinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opd_var1 | ( | int | mn_A, |
| double * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_dinverts(), bli_dinvscalv(), bli_dtrmv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NONUNIT_DIAG, and BLIS_TRANSPOSE.
Referenced by FLA_Trinv_ln_opt_var1().
{
double alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_dtrmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a10t );
// FLA_Inv_scal_external( alpha11, a10t );
bli_dneg2( alpha11, &alpha11_m1 );
bli_dinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
&alpha11_m1,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_dinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opd_var2 | ( | int | mn_A, |
| double * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_dinverts(), bli_dinvscalv(), bli_dtrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, and BLIS_NONUNIT_DIAG.
Referenced by FLA_Trinv_ln_opt_var2().
{
double alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
/*------------------------------------------------------------*/
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_dtrsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_dneg2( alpha11, &alpha11_m1 );
bli_dinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_dinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opd_var3 | ( | int | mn_A, |
| double * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_dger(), bli_dinverts(), bli_dinvscalv(), BLIS_NO_CONJUGATE, and FLA_ONE.
Referenced by FLA_Trinv_ln_opt_var3().
{
double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
double alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_dneg2( alpha11, &alpha11_m1 );
bli_dinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Ger_external( FLA_ONE, a21, a10t, A20 );
bli_dger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Inv_scal_external( alpha11, a10t );
bli_dinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
alpha11,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_dinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opd_var4 | ( | int | mn_A, |
| double * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_dger(), bli_dinverts(), bli_dscalv(), bli_dtrmv(), bli_dtrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{
double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < mn_A; ++i )
{
double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_dscalv( BLIS_NO_CONJUGATE,
mn_ahead,
buff_m1,
a21, rs_A );
bli_dtrsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 );
bli_dger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_m1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_dtrmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_dinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_ops_var1 | ( | int | mn_A, |
| float * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_sinverts(), bli_sinvscalv(), bli_strmv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NONUNIT_DIAG, and BLIS_TRANSPOSE.
Referenced by FLA_Trinv_ln_opt_var1().
{
float alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_strmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a10t );
// FLA_Inv_scal_external( alpha11, a10t );
bli_sneg2( alpha11, &alpha11_m1 );
bli_sinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
&alpha11_m1,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_sinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_ops_var2 | ( | int | mn_A, |
| float * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_sinverts(), bli_sinvscalv(), bli_strsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, and BLIS_NONUNIT_DIAG.
Referenced by FLA_Trinv_ln_opt_var2().
{
float alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
/*------------------------------------------------------------*/
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_strsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_sneg2( alpha11, &alpha11_m1 );
bli_sinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_sinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_ops_var3 | ( | int | mn_A, |
| float * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_sger(), bli_sinverts(), bli_sinvscalv(), BLIS_NO_CONJUGATE, and FLA_ONE.
Referenced by FLA_Trinv_ln_opt_var3().
{
float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
float alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_sneg2( alpha11, &alpha11_m1 );
bli_sinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Ger_external( FLA_ONE, a21, a10t, A20 );
bli_sger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Inv_scal_external( alpha11, a10t );
bli_sinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
alpha11,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_sinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_ops_var4 | ( | int | mn_A, |
| float * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_sger(), bli_sinverts(), bli_sscalv(), bli_strmv(), bli_strsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{
float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < mn_A; ++i )
{
float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_sscalv( BLIS_NO_CONJUGATE,
mn_ahead,
buff_m1,
a21, rs_A );
bli_strsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 );
bli_sger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_m1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_strmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_sinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Trinv_ln_opc_var1(), FLA_Trinv_ln_opd_var1(), FLA_Trinv_ln_ops_var1(), and FLA_Trinv_ln_opz_var1().
Referenced by FLA_Trinv_ln().
{
FLA_Datatype datatype;
int mn_A;
int rs_A, cs_A;
datatype = FLA_Obj_datatype( A );
mn_A = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
FLA_Trinv_ln_ops_var1( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
FLA_Trinv_ln_opd_var1( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
FLA_Trinv_ln_opc_var1( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
FLA_Trinv_ln_opz_var1( mn_A,
buff_A, rs_A, cs_A );
break;
}
}
return FLA_SUCCESS;
}
References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Trinv_ln_opc_var2(), FLA_Trinv_ln_opd_var2(), FLA_Trinv_ln_ops_var2(), and FLA_Trinv_ln_opz_var2().
Referenced by FLA_Trinv_ln().
{
FLA_Datatype datatype;
int mn_A;
int rs_A, cs_A;
datatype = FLA_Obj_datatype( A );
mn_A = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
FLA_Trinv_ln_ops_var2( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
FLA_Trinv_ln_opd_var2( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
FLA_Trinv_ln_opc_var2( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
FLA_Trinv_ln_opz_var2( mn_A,
buff_A, rs_A, cs_A );
break;
}
}
return FLA_SUCCESS;
}
References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Trinv_ln_opc_var3(), FLA_Trinv_ln_opd_var3(), FLA_Trinv_ln_ops_var3(), and FLA_Trinv_ln_opz_var3().
Referenced by FLA_Trinv_ln().
{
FLA_Datatype datatype;
int mn_A;
int rs_A, cs_A;
datatype = FLA_Obj_datatype( A );
mn_A = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
FLA_Trinv_ln_ops_var3( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
FLA_Trinv_ln_opd_var3( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
FLA_Trinv_ln_opc_var3( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
FLA_Trinv_ln_opz_var3( mn_A,
buff_A, rs_A, cs_A );
break;
}
}
return FLA_SUCCESS;
}
References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Trinv_ln_opc_var4(), FLA_Trinv_ln_opd_var4(), FLA_Trinv_ln_ops_var4(), and FLA_Trinv_ln_opz_var4().
Referenced by FLA_Trinv_ln().
{
FLA_Datatype datatype;
int mn_A;
int rs_A, cs_A;
datatype = FLA_Obj_datatype( A );
mn_A = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
FLA_Trinv_ln_ops_var4( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
FLA_Trinv_ln_opd_var4( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
FLA_Trinv_ln_opc_var4( mn_A,
buff_A, rs_A, cs_A );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
FLA_Trinv_ln_opz_var4( mn_A,
buff_A, rs_A, cs_A );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opz_var1 | ( | int | mn_A, |
| dcomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_zinverts(), bli_zinvscalv(), bli_ztrmv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NONUNIT_DIAG, and BLIS_TRANSPOSE.
Referenced by FLA_Trinv_ln_opt_var1().
{
dcomplex alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_ztrmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a10t );
// FLA_Inv_scal_external( alpha11, a10t );
bli_zneg2( alpha11, &alpha11_m1 );
bli_zinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
&alpha11_m1,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_zinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opz_var2 | ( | int | mn_A, |
| dcomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_zinverts(), bli_zinvscalv(), bli_ztrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, and BLIS_NONUNIT_DIAG.
Referenced by FLA_Trinv_ln_opt_var2().
{
dcomplex alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
/*------------------------------------------------------------*/
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_ztrsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_zneg2( alpha11, &alpha11_m1 );
bli_zinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_zinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opz_var3 | ( | int | mn_A, |
| dcomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_zger(), bli_zinverts(), bli_zinvscalv(), BLIS_NO_CONJUGATE, and FLA_ONE.
Referenced by FLA_Trinv_ln_opt_var3().
{
dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
dcomplex alpha11_m1;
int i;
for ( i = 0; i < mn_A; ++i )
{
dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Inv_scal_external( alpha11, a21 );
bli_zneg2( alpha11, &alpha11_m1 );
bli_zinvscalv( BLIS_NO_CONJUGATE,
mn_ahead,
&alpha11_m1,
a21, rs_A );
// FLA_Ger_external( FLA_ONE, a21, a10t, A20 );
bli_zger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Inv_scal_external( alpha11, a10t );
bli_zinvscalv( BLIS_NO_CONJUGATE,
mn_behind,
alpha11,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_zinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trinv_ln_opz_var4 | ( | int | mn_A, |
| dcomplex * | A, | ||
| int | rs_A, | ||
| int | cs_A | ||
| ) |
References bli_zger(), bli_zinverts(), bli_zscalv(), bli_ztrmv(), bli_ztrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{
dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < mn_A; ++i )
{
dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
int mn_ahead = mn_A - i - 1;
int mn_behind = i;
/*------------------------------------------------------------*/
// FLA_Scal_external( FLA_MINUS_ONE, a21 );
// FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
bli_zscalv( BLIS_NO_CONJUGATE,
mn_ahead,
buff_m1,
a21, rs_A );
bli_ztrsv( BLIS_LOWER_TRIANGULAR,
BLIS_NO_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_ahead,
A22, rs_A, cs_A,
a21, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 );
bli_zger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
mn_ahead,
mn_behind,
buff_m1,
a21, rs_A,
a10t, cs_A,
A20, rs_A, cs_A );
// FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
bli_ztrmv( BLIS_LOWER_TRIANGULAR,
BLIS_TRANSPOSE,
BLIS_NONUNIT_DIAG,
mn_behind,
A00, rs_A, cs_A,
a10t, cs_A );
// FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
bli_zinverts( BLIS_NO_CONJUGATE,
alpha11 );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
References FLA_Cont_with_3x3_to_2x2(), FLA_Inv_scal_external(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Trmv_external().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// a10t' = tril( A00' ) * a10t';
FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
// a10t = -a10t / alpha11;
FLA_Scal_external( FLA_MINUS_ONE, a10t );
FLA_Inv_scal_external( alpha11, a10t );
// alpha11 = 1.0 / alpha11;
FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_3x3_to_2x2(), FLA_Inv_scal_external(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Trsv_external().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// a21 = tril( A22 ) \ a21;
FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
// a21 = -a21 / alpha11;
FLA_Scal_external( FLA_MINUS_ONE, a21 );
FLA_Inv_scal_external( alpha11, a21 );
// alpha11 = 1.0 / alpha11;
FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Inv_scal_external(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// a21 = -a21 / alpha11;
FLA_Scal_external( FLA_MINUS_ONE, a21 );
FLA_Inv_scal_external( alpha11, a21 );
// A20 = a21 * a10t + A20;
FLA_Ger_external( FLA_ONE, a21, a10t, A20 );
// a10t = a10t / alpha11;
FLA_Inv_scal_external( alpha11, a10t );
// alpha11 = 1.0 / alpha11;
FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), FLA_Trmv_external(), and FLA_Trsv_external().
Referenced by FLA_Trinv_ln().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// a21 = -tril( A22 ) \ a21;
FLA_Scal_external( FLA_MINUS_ONE, a21 );
FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 );
// A20 = -a21 * a10t + A20;
FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 );
// a10t' = tril( A00' ) * a10t';
FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
// alpha11 = 1.0 / alpha11;
FLA_Invert( FLA_NO_CONJUGATE, alpha11 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}
1.7.6.1