|
libflame
revision_anchor
|
Go to the source code of this file.
Functions | |
| FLA_Error | FLA_Trmm_lun_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
| FLA_Error | FLA_Trmm_lun_blk_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
| FLA_Error | FLA_Trmm_lun_blk_var3 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
| FLA_Error | FLA_Trmm_lun_blk_var4 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
| FLA_Error | FLA_Trmm_lun_unb_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
| FLA_Error | FLA_Trmm_lun_unb_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
| FLA_Error | FLA_Trmm_lun_unb_var3 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
| FLA_Error | FLA_Trmm_lun_unb_var4 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
| FLA_Error FLA_Trmm_lun_blk_var1 | ( | FLA_Diag | diagA, |
| FLA_Obj | alpha, | ||
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| fla_trmm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_lun().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
dim_t b;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
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 );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* B1 = triu( A11 ) * B1; */
FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA,
alpha, A11, B1,
FLA_Cntl_sub_trmm( cntl ) );
/* B1 = B1 + A12 * B2; */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A12, B2, FLA_ONE, B1,
FLA_Cntl_sub_gemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trmm_lun_blk_var2 | ( | FLA_Diag | diagA, |
| FLA_Obj | alpha, | ||
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| fla_trmm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_lun().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
dim_t b;
FLA_Scal_internal( alpha, B,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
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 );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* B0 = B0 + A01 * B1; */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, A01, B1, FLA_ONE, B0,
FLA_Cntl_sub_gemm( cntl ) );
/* B1 = triu( A11 ) * B1; */
FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA,
FLA_ONE, A11, B1,
FLA_Cntl_sub_trmm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trmm_lun_blk_var3 | ( | FLA_Diag | diagA, |
| FLA_Obj | alpha, | ||
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| fla_trmm_t * | cntl | ||
| ) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_lun().
{
FLA_Obj BL, BR, B0, B1, B2;
dim_t b;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
b = FLA_Determine_blocksize( BR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
/* B1 = triu( A ) * B1 */
FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA,
alpha, A, B1,
FLA_Cntl_sub_trmm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Trmm_lun_blk_var4 | ( | FLA_Diag | diagA, |
| FLA_Obj | alpha, | ||
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| fla_trmm_t * | cntl | ||
| ) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_lun().
{
FLA_Obj BL, BR, B0, B1, B2;
dim_t b;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
b, FLA_LEFT );
/*------------------------------------------------------------*/
/* B1 = triu( A ) * B1 */
FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA,
alpha, A, B1,
FLA_Cntl_sub_trmm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().
Referenced by FLA_Trmm_lun().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Scal_external( alpha, B );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
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 );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* *** */
&b1t,
BB, &B2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* b1t = b1t * alpha11; */
if ( diagA != FLA_UNIT_DIAG )
FLA_Scal_external( alpha11, b1t );
/* b1t = b1t + a12t * B2; */
FLA_Gemv_external( FLA_TRANSPOSE, FLA_ONE, B2, a12t, FLA_ONE, b1t );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
b1t,
/* ** */ /* *** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().
Referenced by FLA_Trmm_lun().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Scal_external( alpha, B );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
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 );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* *** */
&b1t,
BB, &B2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* B0 = B0 + a01 * b1t; */
FLA_Ger_external( FLA_ONE, a01, b1t, B0 );
/* b1t = alpha11 * b1t; */
if ( diagA != FLA_UNIT_DIAG )
FLA_Scal_external( alpha11, b1t );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
b1t,
/* ** */ /* *** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_1x3_to_1x2(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scal_external(), and FLA_Trmv_external().
Referenced by FLA_Trmm_lun().
{
FLA_Obj BL, BR, B0, b1, B2;
FLA_Scal_external( alpha, B );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* b1 = triu( A ) * b1 */
FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, A, b1 );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_1x3_to_1x2(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scal_external(), and FLA_Trmv_external().
Referenced by FLA_Trmm_lun().
{
FLA_Obj BL, BR, B0, b1, B2;
FLA_Scal_external( alpha, B );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
1, FLA_LEFT );
/*------------------------------------------------------------*/
/* b1 = triu( A ) * b1 */
FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, A, b1 );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
1.7.6.1