|
libflame
12600
|
Go to the source code of this file.
| FLA_Error FLA_Hemm_ll_blk_var1 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
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 );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&C1,
CB, &C2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C0 = C0 + A10' * B1 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B1, FLA_ONE, C0,
FLA_Cntl_sub_gemm1( cntl ) );
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C1 = C1 + A10 * B0 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B0, FLA_ONE, C1,
FLA_Cntl_sub_gemm2( 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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
C1,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var10 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_ll().
{
FLA_Obj BL, BR, B0, B1, B2;
FLA_Obj CL, CR, C0, C1, C2;
dim_t b;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
FLA_Part_1x2( C, &CL, &CR, 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 );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
b, FLA_LEFT );
/*------------------------------------------------------------*/
/* C1 = C1 + A * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A, B1, beta, C1,
FLA_Cntl_sub_hemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
FLA_RIGHT );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var2 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
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 );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&C1,
CB, &C2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C1 = C1 + A10 * B0 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B0, FLA_ONE, C1,
FLA_Cntl_sub_gemm1( cntl ) );
/* C1 = C1 + A21' * B2 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B2, FLA_ONE, C1,
FLA_Cntl_sub_gemm2( 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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
C1,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var3 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
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 );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&C1,
CB, &C2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C1 = C1 + A21' * B2 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B2, FLA_ONE, C1,
FLA_Cntl_sub_gemm1( cntl ) );
/* C2 = C2 + A21 * B1 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B1, FLA_ONE, C2,
FLA_Cntl_sub_gemm2( 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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
C1,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var4 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
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 );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&C1,
CB, &C2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C0 = C0 + A10' * B1 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B1, FLA_ONE, C0,
FLA_Cntl_sub_gemm1( cntl ) );
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C2 = C2 + A21 * B1 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B1, FLA_ONE, C2,
FLA_Cntl_sub_gemm2( 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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
C1,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var5 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ATL, FLA_TL, 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&B1,
/* ** */ /* ** */
BB, &B2, b, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&C1,
/* ** */ /* ** */
CB, &C2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C1 = C1 + A21' * B2 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B2, FLA_ONE, C1,
FLA_Cntl_sub_gemm1( cntl ) );
/* C2 = C2 + A21 * B1 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B1, FLA_ONE, C2,
FLA_Cntl_sub_gemm2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
/* ************** */ /* ****************** */
A10, /**/ A11, A12,
&ABL, /**/ &ABR, A20, /**/ A21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
B1,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
C1,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var6 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ATL, FLA_TL, 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&B1,
/* ** */ /* ** */
BB, &B2, b, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&C1,
/* ** */ /* ** */
CB, &C2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C1 = C1 + A10 * B0 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B0, FLA_ONE, C1,
FLA_Cntl_sub_gemm1( cntl ) );
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C1 = C1 + A21' * B2 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B2, FLA_ONE, C1,
FLA_Cntl_sub_gemm2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
/* ************** */ /* ****************** */
A10, /**/ A11, A12,
&ABL, /**/ &ABR, A20, /**/ A21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
B1,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
C1,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var7 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ATL, FLA_TL, 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&B1,
/* ** */ /* ** */
BB, &B2, b, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&C1,
/* ** */ /* ** */
CB, &C2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C0 = C0 + A10' * B1 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B1, FLA_ONE, C0,
FLA_Cntl_sub_gemm1( cntl ) );
/* C1 = C1 + A10 * B0 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B0, FLA_ONE, C1,
FLA_Cntl_sub_gemm2( cntl ) );
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
/* ************** */ /* ****************** */
A10, /**/ A11, A12,
&ABL, /**/ &ABR, A20, /**/ A21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
B1,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
C1,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var8 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ATL, FLA_TL, 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&B1,
/* ** */ /* ** */
BB, &B2, b, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&C1,
/* ** */ /* ** */
CB, &C2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C0 = C0 + A10' * B1 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A10, B1, FLA_ONE, C0,
FLA_Cntl_sub_gemm1( cntl ) );
/* C1 = C1 + A11 * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A11, B1, FLA_ONE, C1,
FLA_Cntl_sub_hemm( cntl ) );
/* C2 = C2 + A21 * B1 */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A21, B1, FLA_ONE, C2,
FLA_Cntl_sub_gemm2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
/* ************** */ /* ****************** */
A10, /**/ A11, A12,
&ABL, /**/ &ABR, A20, /**/ A21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
B1,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
C1,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Hemm_ll_blk_var9 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_hemm_t * | cntl | ||
| ) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_ll().
{
FLA_Obj BL, BR, B0, B1, B2;
FLA_Obj CL, CR, C0, C1, C2;
dim_t b;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
FLA_Part_1x2( C, &CL, &CR, 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 );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C1 = C1 + A * B1 */
FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
alpha, A, B1, beta, C1,
FLA_Cntl_sub_hemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&c1t,
CB, &C2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C0 = C0 + a10t' * b1t */
FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a10t, b1t, C0 );
/* c1t = c1t + a10t * B0 */
/* c1t' = c1t' + B0' * a10t' */
FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a10t, FLA_ONE, c1t );
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/*------------------------------------------------------------*/
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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
c1t,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_1x3_to_1x2(), FLA_Hemv_external(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_ll().
{
FLA_Obj BL, BR, B0, b1t, B2;
FLA_Obj CL, CR, C0, c1t, C2;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
1, FLA_LEFT );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2,
1, FLA_LEFT );
/*------------------------------------------------------------*/
/* c1t = c1t + A * b1t */
FLA_Hemv_external( FLA_LOWER_TRIANGULAR, alpha, A, b1t, beta, c1t );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
FLA_RIGHT );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&c1t,
CB, &C2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* c1t = c1t + a10t * B0 */
/* c1t' = c1t' + B0' * a10t' */
FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a10t, FLA_ONE, c1t );
/* c1t = c1t + a21' * B2 */
/* c1t' = c1t' + B2' * a21 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a21, FLA_ONE, c1t );
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/*------------------------------------------------------------*/
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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
c1t,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), 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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&c1t,
CB, &C2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/* c1t = c1t + a21' * B2 */
/* c1t' = c1t' + B2' * a21 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a21, FLA_ONE, c1t );
/* C2 = C2 + a21 * b1t */
FLA_Ger_external( alpha, a21, b1t, C2 );
/*------------------------------------------------------------*/
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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
c1t,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
FLA_Part_2x1( C, &CT,
&CB, 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 );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&c1t,
CB, &C2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C0 = C0 + a10t' * b1t */
FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a10t, b1t, C0 );
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/* C2 = C2 + a21 * b1t */
FLA_Ger_external( alpha, a21, b1t, C2 );
/*------------------------------------------------------------*/
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 );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
c1t,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), 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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&b1t,
/* ** */ /* ** */
BB, &B2, 1, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&c1t,
/* ** */ /* ** */
CB, &C2, 1, FLA_TOP );
/*------------------------------------------------------------*/
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/* c1t = c1t + a21' * B2 */
/* c1t' = c1t' + B2' * a21 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a21, FLA_ONE, c1t );
/* C2 = C2 + a21 * b1t */
FLA_Ger_external( alpha, a21, b1t, C2 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
/* ************** */ /* ************************ */
a10t, /**/ alpha11, a12t,
&ABL, /**/ &ABR, A20, /**/ a21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
b1t,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
c1t,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&b1t,
/* ** */ /* ** */
BB, &B2, 1, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&c1t,
/* ** */ /* ** */
CB, &C2, 1, FLA_TOP );
/*------------------------------------------------------------*/
/* c1t = c1t + a10t * B0 */
/* c1t' = c1t' + B0' * a10t' */
FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a10t, FLA_ONE, c1t );
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/* c1t = c1t + a21' * B2 */
/* c1t' = c1t' + B2' * a21 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a21, FLA_ONE, c1t );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
/* ************** */ /* ************************ */
a10t, /**/ alpha11, a12t,
&ABL, /**/ &ABR, A20, /**/ a21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
b1t,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
c1t,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&b1t,
/* ** */ /* ** */
BB, &B2, 1, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&c1t,
/* ** */ /* ** */
CB, &C2, 1, FLA_TOP );
/*------------------------------------------------------------*/
/* C0 = C0 + a10t' * b1t */
FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a10t, b1t, C0 );
/* c1t = c1t + a10t * B0 */
/* c1t' = c1t' + B0' * a10t' */
FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a10t, FLA_ONE, c1t );
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
/* ************** */ /* ************************ */
a10t, /**/ alpha11, a12t,
&ABL, /**/ &ABR, A20, /**/ a21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
b1t,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
c1t,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_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_Hemm_ll().
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( ABR ) < 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_TL );
FLA_Repart_2x1_to_3x1( BT, &B0,
&b1t,
/* ** */ /* ** */
BB, &B2, 1, FLA_TOP );
FLA_Repart_2x1_to_3x1( CT, &C0,
&c1t,
/* ** */ /* ** */
CB, &C2, 1, FLA_TOP );
/*------------------------------------------------------------*/
/* C0 = C0 + a10t' * b1t */
FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a10t, b1t, C0 );
/* c1t = c1t + alpha11 * b1t */
FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
/* C2 = C2 + a21 * b1t */
FLA_Ger_external( alpha, a21, b1t, C2 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
/* ************** */ /* ************************ */
a10t, /**/ alpha11, a12t,
&ABL, /**/ &ABR, A20, /**/ a21, A22,
FLA_BR );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
b1t,
&BB, B2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
c1t,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
References FLA_Cont_with_1x3_to_1x2(), FLA_Hemv_external(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_ll().
{
FLA_Obj BL, BR, B0, b1t, B2;
FLA_Obj CL, CR, C0, c1t, C2;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
1, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* c1t = c1t + A * b1t */
FLA_Hemv_external( FLA_LOWER_TRIANGULAR, alpha, A, b1t, beta, c1t );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
1.7.6.1