FLA_Error FLA_Herk_ln_blk_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
dim_t b;
FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
FLA_Cntl_sub_scalr( cntl ) );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ) {
b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C = C + A1 * A1' */
FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE,
alpha, A1, FLA_ONE, C,
FLA_Cntl_sub_herk( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syr2k_ut_blk_var9( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t* cntl )
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj BT, B0,
BB, B1,
B2;
dim_t b;
FLA_Scalr_internal( FLA_UPPER_TRIANGULAR, beta, C,
FLA_Cntl_sub_scalr( cntl ) );
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
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_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C = C + A1' * B1 + B1' * A1 */
FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C,
FLA_Cntl_sub_syr2k( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Scalr_u_blk_var4( FLA_Obj alpha, FLA_Obj A, fla_scalr_t* cntl )
{
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_BR );
while ( FLA_Obj_min_dim( ATL ) > 0 ){
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 );
/*------------------------------------------------------------*/
// A11 = alpha * triu( A11 );
FLA_Scalr_internal( FLA_UPPER_TRIANGULAR, alpha, A11,
FLA_Cntl_sub_scalr( cntl ) );
// A01 = alpha * A01;
FLA_Scal_internal( alpha, A01,
FLA_Cntl_sub_scal( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
/* ************** */ /* ****************** */
A10, /**/ A11, A12,
&ABL, /**/ &ABR, A20, /**/ A21, A22,
FLA_BR );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Herk_uh_blk_var6( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t* cntl )
{
FLA_Obj AT, A0,
AB, A1,
A2;
dim_t b;
FLA_Scalr_internal( FLA_UPPER_TRIANGULAR, beta, C,
FLA_Cntl_sub_scalr( cntl ) );
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_BOTTOM );
while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x1_to_3x1( AT, &A0,
&A1,
/* ** */ /* ** */
AB, &A2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C = C + A1' * A1 */
FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
alpha, A1, FLA_ONE, C,
FLA_Cntl_sub_herk( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
/* ** */ /* ** */
A1,
&AB, A2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Her2k_uh_blk_var7( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj BL, BR, B0, B1, B2;
FLA_Obj CTL, CTR, C00, C01, C02,
CBL, CBR, C10, C11, C12,
C20, C21, C22;
dim_t b;
FLA_Scalr_internal( FLA_UPPER_TRIANGULAR, beta, C,
FLA_Cntl_sub_scalr( cntl ) );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_BR );
while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
b, FLA_LEFT );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
b, FLA_LEFT );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
&C10, &C11, /**/ &C12,
/* ************* */ /* ******************** */
CBL, /**/ CBR, &C20, &C21, /**/ &C22,
b, b, FLA_TL );
/*------------------------------------------------------------*/
/* C01 = C01 + B0' * A1 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, B0, A1, FLA_ONE, C01,
FLA_Cntl_sub_gemm1( cntl ) );
/* C12 = C12 + A1' * B2 */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A1, B2, FLA_ONE, C12,
FLA_Cntl_sub_gemm2( cntl ) );
/* C11 = C11 + A1' * B1 + B1' * A1 */
FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C11,
FLA_Cntl_sub_her2k( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
FLA_RIGHT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
FLA_RIGHT );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
/* ************** */ /* ****************** */
C10, /**/ C11, C12,
&CBL, /**/ &CBR, C20, /**/ C21, C22,
FLA_BR );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syr2k_ln_blk_var4( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t* cntl )
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CTL, CTR, C00, C01, C02,
CBL, CBR, C10, C11, C12,
C20, C21, C22;
dim_t b;
FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
FLA_Cntl_sub_scalr( cntl ) );
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_2x1( B, &BT,
&BB, 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_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, 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 );
/*------------------------------------------------------------*/
/* C21 = C21 + A2 * B1' */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
alpha, A2, B1, FLA_ONE, C21,
FLA_Cntl_sub_gemm1( cntl ) );
/* C21 = C21 + B2 * A1' */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
alpha, B2, A1, FLA_ONE, C21,
FLA_Cntl_sub_gemm2( cntl ) );
/* C11 = C11 + A1 * B1' + B1 * A1' */
FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C11,
FLA_Cntl_sub_syr2k( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, 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;
}