FLA_Error FLA_Trmm_ruc_blk_var2( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t* cntl )
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BL, BR, B0, 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_BR );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
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_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
b, FLA_LEFT );
/*------------------------------------------------------------*/
/* B2 = B2 + B1 * A12; */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE,
FLA_ONE, B1, A12, FLA_ONE, B2,
FLA_Cntl_sub_gemm( cntl ) );
/* B1 = B1 * triu( A11 ); */
FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_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_BR );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Sylv_nn_blk_var17( FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t* cntl )
{
FLA_Obj BTL, BTR, B00, B01, B02,
BBL, BBR, B10, B11, B12,
B20, B21, B22;
FLA_Obj CL, CR, C0, C1, C2;
dim_t b;
FLA_Part_2x2( B, &BTL, &BTR,
&BBL, &BBR, 0, 0, FLA_TL );
FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
while ( FLA_Obj_length( BTL ) < FLA_Obj_length( B ) ){
b = FLA_Determine_blocksize( CR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
/* ************* */ /* ******************** */
&B10, /**/ &B11, &B12,
BBL, /**/ BBR, &B20, /**/ &B21, &B22,
b, b, FLA_BR );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
b, FLA_RIGHT );
// Loop Invariant:
// CL =
// CR =
/*------------------------------------------------------------*/
// C1 = sylv( A, B11, C1 -/+ C0 * B01 );
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_NEGATE( isgn ), C0, B01, FLA_ONE, C1,
FLA_Cntl_sub_gemm1( cntl ) );
FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
isgn, A, B11, C1, scale,
FLA_Cntl_sub_sylv1( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
B10, B11, /**/ B12,
/* ************** */ /* ****************** */
&BBL, /**/ &BBR, B20, B21, /**/ B22,
FLA_TL );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nt_blk_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj BL, BR, B0, B1, B2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_1x2( B, &BL, &BR, 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 );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C = alpha * A1 * B1' + C; */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C,
FLA_Cntl_sub_gemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Her2k_ln_blk_var9( 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;
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 );
FLA_Part_1x2( B, &BL, &BR, 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 );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C = C + A1 * B1' + B1 * A1' */
FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C,
FLA_Cntl_sub_her2k( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_QR_UT_unb_var1( FLA_Obj A, FLA_Obj t )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj tLt, tRt, t0t, tau1, t2t;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_1x2( t, &tLt, &tRt, 0, FLA_LEFT );
while ( FLA_Obj_min_dim( ABR ) > 0 ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
FLA_Repart_1x2_to_1x3( tLt, /**/ tRt, &t0t, /**/ &tau1, &t2t,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
// Compute tau11 and u21 from alpha11 and a21 such that tau11 and u21
// determine a Householder transform H such that applying H from the
// left to the column vector consisting of alpha11 and a21 annihilates
// the entries in a21 (and updates alpha11).
FLA_Househ2_UT( FLA_LEFT,
alpha11,
a21, tau1 );
// / a12t \ = H / a12t \
// \ A22 / \ A22 /
//
// where H is formed from tau11 and u21.
FLA_Apply_H2_UT( FLA_LEFT, tau1, a21, a12t,
A22 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
FLA_Cont_with_1x3_to_1x2( &tLt, /**/ &tRt, t0t, tau1, /**/ t2t,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syrk_ut_blk_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj CTL, CTR, C00, C01, C02,
CBL, CBR, C10, C11, C12,
C20, C21, C22;
dim_t b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_TL );
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 );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
/* ************* */ /* ******************** */
&C10, /**/ &C11, &C12,
CBL, /**/ CBR, &C20, /**/ &C21, &C22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
/* C12 = C12 + A1' * A2 */
FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A1, A2, beta, C12,
FLA_Cntl_sub_gemm( cntl ) );
/* C11 = C11 + A1' * A1 */
FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE,
alpha, A1, beta, C11,
FLA_Cntl_sub_syrk( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
C10, C11, /**/ C12,
/* ************** */ /* ****************** */
&CBL, /**/ &CBR, C20, C21, /**/ C22,
FLA_TL );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Hemm_lu_blk_var9( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t* cntl )
{
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_UPPER_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;
}
FLA_Error FLA_Gemm_tc_blk_var4( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
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 = alpha * A' * B1 + C1; */
FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE,
alpha, A, B1, beta, C1,
FLA_Cntl_sub_gemm( 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_Apply_Q_UT_lhfr_blk_var2( FLA_Obj A, FLA_Obj T, FLA_Obj W, FLA_Obj B, fla_apqut_t* cntl )
{
FLA_Obj BL, BR, B0, B1, B2;
FLA_Obj WL, WR, W0, W1, W2;
dim_t b;
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
FLA_Part_1x2( W, &WL, &WR, 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( WL, /**/ WR, &W0, /**/ &W1, &W2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
// B1 = Q' * B1;
FLA_Apply_Q_UT_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_ROWWISE,
A, T, W1, B1,
FLA_Cntl_sub_apqut( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &WL, /**/ &WR, W0, W1, /**/ W2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_hh_blk_var6( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj BL, BR, B0, B1, B2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_BOTTOM );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
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 );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
b, FLA_LEFT );
/*------------------------------------------------------------*/
/* C = alpha * A1' * B1' + C; */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C,
FLA_Cntl_sub_gemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
/* ** */ /* ** */
A1,
&AB, A2, FLA_BOTTOM );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nc_blk_var6( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj BT, B0,
BB, B1,
B2;
dim_t b;
FLA_Scal_internal( beta, C,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_BOTTOM );
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_2x1_to_3x1( BT, &B0,
&B1,
/* ** */ /* ** */
BB, &B2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C = alpha * A1 * B1 + C; */
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C,
FLA_Cntl_sub_gemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
FLA_RIGHT );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
B1,
&BB, B2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syr2k_un_unb_var9( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1t, A2;
FLA_Obj BL, BR, B0, b1t, B2;
FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1t, &A2,
1, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C = C + a1t * b1t' + b1t * a1t' */
FLA_Syr2_external( FLA_UPPER_TRIANGULAR, alpha, a1t, b1t, C );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1t, /**/ A2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Copyt_c_blk_var3( FLA_Obj A, FLA_Obj B, fla_copyt_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj BL, BR, B0, B1, B2;
dim_t b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_1x2( B, &BL, &BR, 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 );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
FLA_Copyt_internal( FLA_CONJ_NO_TRANSPOSE, A1, B1,
FLA_Cntl_sub_copyt( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Herk_lh_unb_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj CTL, CTR, C00, c01, C02,
CBL, CBR, c10t, gamma11, c12t,
C20, c21, C22;
FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_TL );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
1, FLA_RIGHT );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
/* ************* */ /* ************************** */
&c10t, /**/ &gamma11, &c12t,
CBL, /**/ CBR, &C20, /**/ &c21, &C22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
/* C10 = C10 + A1' * A0 */
/* c10t = c10t + A0' * a1 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, A0, a1, FLA_ONE, c10t );
/* gamma11 = gamma11 + a1' * a1 */
FLA_Dotcs_external( FLA_CONJUGATE, alpha, a1, a1, FLA_ONE, gamma11 );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
c10t, gamma11, /**/ c12t,
/* ************** */ /* ************************ */
&CBL, /**/ &CBR, C20, c21, /**/ C22,
FLA_TL );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Trsm_rlc_unb_var1( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BL, BR, B0, b1, B2;
FLA_Scal_external( alpha, B );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
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_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
1, FLA_LEFT );
/*------------------------------------------------------------*/
/* b1 = b1 - B2 * a21; */
FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, B2, a21, FLA_ONE, b1 );
/* b1 = b1 / alpha11; */
if ( diagA != FLA_UNIT_DIAG )
FLA_Inv_scalc_external( FLA_CONJUGATE, alpha11, b1 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
/* ************** */ /* ************************ */
a10t, /**/ alpha11, a12t,
&ABL, /**/ &ABR, A20, /**/ a21, A22,
FLA_BR );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Trmm_run_unb_var2( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BL, BR, B0, b1, B2;
FLA_Scal_external( alpha, B );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_BR );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
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_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
1, FLA_LEFT );
/*------------------------------------------------------------*/
/* B2 = B2 + b1 * a12t; */
FLA_Ger_external( FLA_ONE, b1, a12t, B2 );
/* b1 = b1 * alpha11; */
if ( diagA != FLA_UNIT_DIAG )
FLA_Scal_external( alpha11, b1 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
/* ************** */ /* ************************ */
a10t, /**/ alpha11, a12t,
&ABL, /**/ &ABR, A20, /**/ a21, A22,
FLA_BR );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nn_unb_var3( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj BL, BR, B0, b1, B2;
FLA_Obj CL, CR, C0, c1, C2;
FLA_Scal_external( beta, C );
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, /**/ &b1, &B2,
1, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1, &C2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* c1 = A * b1 + c1 */
FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A, b1, FLA_ONE, c1 );
/*------------------------------------------------------------*/
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;
}
FLA_Error FLA_Gemm_ct_unb_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj BL, BR, B0, b1, B2;
FLA_Scal_external( beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ) {
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
1, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C = a1 * b1' + C */
FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a1, b1, C );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syrk_ut_unb_var3( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj CTL, CTR, C00, c01, C02,
CBL, CBR, c10t, gamma11, c12t,
C20, c21, C22;
FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_BR );
while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
1, FLA_LEFT );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
&c10t, &gamma11, /**/ &c12t,
/* ************* */ /* ************************** */
CBL, /**/ CBR, &C20, &c21, /**/ &C22,
1, 1, FLA_TL );
/*------------------------------------------------------------*/
/* c12t = c12t + A2' * a1 */
FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, a1, FLA_ONE, c12t );
/* gamma11 = gamma11 + a1 * a1' */
FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
FLA_RIGHT );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
/* ************** */ /* ************************ */
c10t, /**/ gamma11, c12t,
&CBL, /**/ &CBR, C20, /**/ c21, C22,
FLA_BR );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_hh_blk_var3( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj CL, CR, C0, C1, C2;
dim_t b;
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){
b = FLA_Determine_blocksize( BB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C1 = alpha * A' * B1' + C1; */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
alpha, A, B1, beta, C1,
FLA_Cntl_sub_gemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemv_t_blk_var2( FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj yT, y0,
yB, y1,
y2;
dim_t b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
FLA_Part_2x1( y, &yT,
&yB, 0, FLA_BOTTOM );
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_2x1_to_3x1( yT, &y0,
&y1,
/* ** */ /* ** */
yB, &y2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* y1 = alpha * A1' * x + y1 */
FLA_Gemv_internal( FLA_TRANSPOSE,
alpha, A1, x, beta, y1,
FLA_Cntl_sub_gemv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
FLA_RIGHT );
FLA_Cont_with_3x1_to_2x1( &yT, y0,
/* ** */ /* ** */
y1,
&yB, y2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_hn_blk_var2( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj CT, C0,
CB, C1,
C2;
dim_t b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_BOTTOM );
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_2x1_to_3x1( CT, &C0,
&C1,
/* ** */ /* ** */
CB, &C2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* C1 = alpha * A1' * B + C1; */
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A1, B, beta, C1,
FLA_Cntl_sub_gemm( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
FLA_RIGHT );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
/* ** */ /* ** */
C1,
&CB, C2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Axpyt_h_blk_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpyt_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj BT, B0,
BB, B1,
B2;
dim_t b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
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 );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
FLA_Axpyt_internal( FLA_CONJ_TRANSPOSE, alpha, A1, B1,
FLA_Cntl_sub_axpyt( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Swap_t_blk_var2( FLA_Obj A, FLA_Obj B, fla_swap_t* cntl )
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj BL, BR, B0, B1, B2;
dim_t b;
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
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_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
FLA_Swapt_external( FLA_TRANSPOSE, A1, B1 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_hh_unb_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj CT, C0,
CB, c1t,
C2;
FLA_Scal_external( beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_TOP );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
1, FLA_RIGHT );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* *** */
&c1t,
CB, &C2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* c1t = a1' * B' + c1t */
/* c1t' = B * a1 + c1t' */
FLA_Gemvc_external( FLA_CONJ_NO_TRANSPOSE, FLA_CONJUGATE, alpha, B, a1, FLA_ONE, c1t );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
c1t,
/* ** */ /* *** */
&CB, C2, FLA_TOP );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nt_unb_var3( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Obj CL, CR, C0, c1, C2;
FLA_Scal_external( beta, C );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* *** */
&b1t,
BB, &B2, 1, FLA_BOTTOM );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1, &C2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* c1 = A * b1t + c1 */
FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A, b1t, FLA_ONE, c1 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &BT, B0,
b1t,
/* ** */ /* *** */
&BB, B2, FLA_TOP );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1, /**/ C2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syrk_ln_omp1t_var5_fc( FLA_Obj A, FLA_Obj C, int nb_alg )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj MyC;
int b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
#pragma intel omp parallel taskq
{
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
nb_alg = FLA_Obj_width( A )/omp_get_num_threads() + 1;
b = min( FLA_Obj_width( AR ), nb_alg );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
#pragma intel omp task captureprivate(A1) private(MyC)
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &MyC );
FLA_Obj_set_to_zero( MyC );
/* MyC := A1 * A1' */
FLA_Syrk( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A1, FLA_ZERO, MyC );
/* C := MyC */
FLA_Axpy_sync_circular( FLA_ONE, MyC, C );
FLA_Obj_free( &MyC );
}
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
}
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nc_unb_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj BT, B0,
BB, b1t,
B2;
FLA_Scal_external( beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
1, FLA_RIGHT );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* *** */
&b1t,
BB, &B2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C = a1 * b1t + C */
FLA_Gerc_external( FLA_NO_CONJUGATE, FLA_CONJUGATE, alpha, a1, b1t, C );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
b1t,
/* ** */ /* *** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_tt_unb_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AT, A0,
AB, a1t,
A2;
FLA_Obj BL, BR, B0, b1, B2;
FLA_Scal_external( beta, C );
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
FLA_Repart_2x1_to_3x1( AT, &A0,
/* ** */ /* *** */
&a1t,
AB, &A2, 1, FLA_BOTTOM );
FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
1, FLA_RIGHT );
/*------------------------------------------------------------*/
/* C = a1t' * b1' + C */
FLA_Ger_external( alpha, a1t, b1, C );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
a1t,
/* ** */ /* *** */
&AB, A2, FLA_TOP );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syrk_ln_omp1t_var5( FLA_Obj A, FLA_Obj C )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj MyC;
int b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
#pragma intel omp parallel taskq
{
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
b = FLA_Task_compute_blocksize( 0, A, AL, FLA_LEFT );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
#pragma intel omp task captureprivate(A1) private(MyC)
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &MyC );
FLA_Obj_set_to_zero( MyC );
/* MyC := A1 * A1' */
FLA_Syrk_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A1, FLA_ZERO, MyC );
/* C := MyC */
FLA_Axpy_sync_pipeline( FLA_ONE, MyC, C );
FLA_Obj_free( &MyC );
}
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
}
}
return FLA_SUCCESS;
}
