FLA_Error FLA_Symm_internal( FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_symm_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Symm_internal_check( side, uplo, alpha, A, B, beta, C, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Symm_internal( side,
uplo,
alpha,
*FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( B ),
beta,
*FLASH_OBJ_PTR_AT( C ),
flash_symm_cntl_mm );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Symm( side, uplo, alpha, A, B, beta, C, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = flash_symm_cntl_blas;
}
// Parameter combinations
if ( side == FLA_LEFT )
{
if ( uplo == FLA_LOWER_TRIANGULAR )
r_val = FLA_Symm_ll( alpha, A, B, beta, C, cntl );
else if ( uplo == FLA_UPPER_TRIANGULAR )
r_val = FLA_Symm_lu( alpha, A, B, beta, C, cntl );
}
else if ( side == FLA_RIGHT )
{
if ( uplo == FLA_LOWER_TRIANGULAR )
r_val = FLA_Symm_rl( alpha, A, B, beta, C, cntl );
else if ( uplo == FLA_UPPER_TRIANGULAR )
r_val = FLA_Symm_ru( alpha, A, B, beta, C, cntl );
}
}
return r_val;
}
FLA_Error FLA_Copyt_internal( FLA_Trans trans, FLA_Obj A, FLA_Obj B, fla_copyt_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Copyt_internal_check( trans, A, B, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Copyt_internal( trans,
*FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( B ),
flash_copyt_cntl );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Copyt( trans, A, B, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = flash_copyt_cntl_blas;
}
// Parameter combinations
if ( trans == FLA_NO_TRANSPOSE )
{
r_val = FLA_Copyt_n( A, B, cntl );
}
else if ( trans == FLA_TRANSPOSE )
{
r_val = FLA_Copyt_t( A, B, cntl );
}
else if ( trans == FLA_CONJ_NO_TRANSPOSE )
{
r_val = FLA_Copyt_c( A, B, cntl );
}
else if ( trans == FLA_CONJ_TRANSPOSE )
{
r_val = FLA_Copyt_h( A, B, cntl );
}
}
return r_val;
}
FLA_Error FLA_LQ_UT_internal( FLA_Obj A, FLA_Obj T, fla_lqut_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_LQ_UT_internal_check( A, T, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
if ( FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_LQ_UT_macro( A, T, cntl );
}
else
{
// Execute
r_val = FLA_LQ_UT_macro_task( A, T, cntl );
}
}
else
{
if ( FLA_Cntl_variant( cntl ) == FLA_UNBLOCKED_VARIANT1 )
{
r_val = FLA_LQ_UT_unb_var1( A, T );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_UNB_OPT_VARIANT1 )
{
r_val = FLA_LQ_UT_opt_var1( A, T );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT1 )
{
r_val = FLA_LQ_UT_blk_var1( A, T, cntl );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_UNBLOCKED_VARIANT2 )
{
r_val = FLA_LQ_UT_unb_var2( A, T );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_UNB_OPT_VARIANT2 )
{
r_val = FLA_LQ_UT_opt_var2( A, T );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT2 )
{
r_val = FLA_LQ_UT_blk_var2( A, T, cntl );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT3 )
{
r_val = FLA_LQ_UT_blk_var3( A, T, cntl );
}
else
{
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
}
return r_val;
}
FLA_Error FLA_Apply_pivots_internal( FLA_Side side, FLA_Trans trans, FLA_Obj p, FLA_Obj A, fla_appiv_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
if ( FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Apply_pivots_macro( side, trans, *FLASH_OBJ_PTR_AT( p ), A, cntl );
}
else
{
// Execute leaf
r_val = FLA_Apply_pivots_macro_task( side, trans, *FLASH_OBJ_PTR_AT( p ), A, cntl );
}
}
else
{
// Parameter combinations
if ( trans == FLA_NO_TRANSPOSE )
{
if ( side == FLA_LEFT )
{
r_val = FLA_Apply_pivots_ln( p, A, cntl );
}
else if ( side == FLA_RIGHT )
{
r_val = FLA_Apply_pivots_rn( p, A, cntl );
}
}
else if ( trans == FLA_TRANSPOSE )
{
if ( side == FLA_LEFT )
{
r_val = FLA_Apply_pivots_lt( p, A, cntl );
}
else if ( side == FLA_RIGHT )
{
r_val = FLA_Apply_pivots_rt( p, A, cntl );
}
}
}
return r_val;
}
FLA_Error FLASH_Gemv( FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y )
{
FLA_Error r_val;
FLA_Bool enable_supermatrix;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Gemv_check( transa, alpha, A, x, beta, y );
// Find the status of SuperMatrix.
enable_supermatrix = FLASH_Queue_get_enabled();
// Temporarily disable SuperMatrix.
FLASH_Queue_disable();
// Execute tasks.
r_val = FLA_Gemv_internal( transa, alpha, A, x, beta, y, flash_gemv_cntl_fm_rp );
// Restore SuperMatrix to its previous status.
if ( enable_supermatrix )
FLASH_Queue_enable();
return r_val;
}
FLA_Error FLASH_Axpy( FLA_Obj alpha, FLA_Obj A, FLA_Obj B )
{
FLA_Error r_val;
FLA_Bool enable_supermatrix;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Axpy_check( alpha, A, B );
// Find the status of SuperMatrix.
enable_supermatrix = FLASH_Queue_get_enabled();
// Temporarily disable SuperMatrix.
FLASH_Queue_disable();
// Execute tasks.
r_val = FLA_Axpy_internal( alpha, A, B, flash_axpy_cntl );
// Restore SuperMatrix to its previous status.
if ( enable_supermatrix )
FLASH_Queue_enable();
return r_val;
}
FLA_Error FLASH_Trsv( FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj A, FLA_Obj x )
{
FLA_Error r_val;
FLA_Bool enable_supermatrix;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Trsv_check( uplo, trans, diag, A, x );
// Find the status of SuperMatrix.
enable_supermatrix = FLASH_Queue_get_enabled();
// Temporarily disable SuperMatrix.
FLASH_Queue_disable();
// Execute tasks.
r_val = FLA_Trsv_internal( uplo, trans, diag, A, x, flash_trsv_cntl );
// Restore SuperMatrix to its previous status.
if ( enable_supermatrix )
FLASH_Queue_enable();
return r_val;
}
FLA_Error FLA_Trinv_internal( FLA_Uplo uplo, FLA_Diag diag, FLA_Obj A, fla_trinv_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Trinv_internal_check( uplo, diag, A, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Trinv_internal( uplo,
diag,
*FLASH_OBJ_PTR_AT( A ),
flash_trinv_cntl );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Trinv( uplo, diag, A, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = fla_trinv_cntl_leaf;
}
// Parameter combinations
if ( uplo == FLA_LOWER_TRIANGULAR )
{
if ( diag == FLA_NONUNIT_DIAG )
{
r_val = FLA_Trinv_ln( A, cntl );
}
else if ( diag == FLA_UNIT_DIAG )
{
r_val = FLA_Trinv_lu( A, cntl );
}
}
else if ( uplo == FLA_UPPER_TRIANGULAR )
{
if ( diag == FLA_NONUNIT_DIAG )
{
r_val = FLA_Trinv_un( A, cntl );
}
else if ( diag == FLA_UNIT_DIAG )
{
r_val = FLA_Trinv_uu( A, cntl );
}
}
}
return r_val;
}
FLA_Error FLA_Gemm_internal( FLA_Trans transa, FLA_Trans transb, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Gemm_internal_check( transa, transb, alpha, A, B, beta, C, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
(FLA_Obj_elemtype( A ) == FLA_TENSOR || FLA_Obj_elemtype( A ) == FLA_MATRIX) &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Gemm_internal( transa,
transb,
alpha,
*FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( B ),
beta,
*FLASH_OBJ_PTR_AT( C ),
flash_gemm_cntl_mm_op );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Gemm( transa, transb, alpha, A, B, beta, C, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = flash_gemm_cntl_blas;
}
// Parameter combinations
if ( transa == FLA_NO_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_nn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_nt( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_nc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_nh( alpha, A, B, beta, C, cntl );
}
else if ( transa == FLA_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_tn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_tt( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_tc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_th( alpha, A, B, beta, C, cntl );
}
else if ( transa == FLA_CONJ_NO_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_cn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_ct( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_cc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_ch( alpha, A, B, beta, C, cntl );
}
else if ( transa == FLA_CONJ_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_hn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_ht( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_hc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_hh( alpha, A, B, beta, C, cntl );
}
}
return r_val;
}
FLA_Error FLASH_SA_LU( FLA_Obj B, FLA_Obj C,
FLA_Obj D, FLA_Obj E, FLA_Obj p, FLA_Obj L, dim_t nb_alg, fla_lu_t* cntl )
{
FLA_Obj DT, D0,
DB, D1,
D2;
FLA_Obj ET, E0,
EB, E1,
E2;
FLA_Obj pT, p0,
pB, p1,
p2;
FLA_Obj LT, L0,
LB, L1,
L2;
FLA_Part_2x1( D, &DT,
&DB, 0, FLA_TOP );
FLA_Part_2x1( E, &ET,
&EB, 0, FLA_TOP );
FLA_Part_2x1( p, &pT,
&pB, 0, FLA_TOP );
FLA_Part_2x1( L, <,
&LB, 0, FLA_TOP );
while ( FLA_Obj_length( DT ) < FLA_Obj_length( D ) )
{
FLA_Repart_2x1_to_3x1( DT, &D0,
/* ** */ /* ** */
&D1,
DB, &D2, 1, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( ET, &E0,
/* ** */ /* ** */
&E1,
EB, &E2, 1, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( pT, &p0,
/* ** */ /* ** */
&p1,
pB, &p2, 1, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( LT, &L0,
/* ** */ /* ** */
&L1,
LB, &L2, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
if ( FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_SA_LU( *FLASH_OBJ_PTR_AT( B ),
*FLASH_OBJ_PTR_AT( D1 ),
*FLASH_OBJ_PTR_AT( p1 ),
*FLASH_OBJ_PTR_AT( L1 ),
nb_alg,
FLA_Cntl_sub_lu( cntl ) );
}
else
{
// Execute leaf
FLA_SA_LU_task( *FLASH_OBJ_PTR_AT( B ),
*FLASH_OBJ_PTR_AT( D1 ),
*FLASH_OBJ_PTR_AT( p1 ),
*FLASH_OBJ_PTR_AT( L1 ),
nb_alg,
FLA_Cntl_sub_lu( cntl ) );
}
FLASH_SA_FS( L1,
D1, p1, C,
E1, nb_alg, FLA_Cntl_sub_gemm1( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &DT, D0,
D1,
/* ** */ /* ** */
&DB, D2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &ET, E0,
E1,
/* ** */ /* ** */
&EB, E2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &pT, p0,
p1,
/* ** */ /* ** */
&pB, p2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( <, L0,
L1,
/* ** */ /* ** */
&LB, L2, FLA_TOP );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Apply_Q_UT_internal( FLA_Side side, FLA_Trans trans, FLA_Direct direct, FLA_Store storev, FLA_Obj A, FLA_Obj T, FLA_Obj W, FLA_Obj B, fla_apqut_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Apply_Q_UT_internal_check( side, trans, direct, storev, A, T, W, B, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Apply_Q_UT_internal( side,
trans,
direct,
storev,
*FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( T ),
*FLASH_OBJ_PTR_AT( W ),
*FLASH_OBJ_PTR_AT( B ),
flash_apqut_cntl );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Apply_Q_UT( side, trans, direct, storev, A, T, W, B, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf.
cntl = fla_apqut_cntl_leaf;
}
if ( side == FLA_LEFT )
{
if ( trans == FLA_NO_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_lnfc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_lnfr( A, T, W, B, cntl );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_lnbc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_lnbr( A, T, W, B, cntl );
}
}
else if ( trans == FLA_TRANSPOSE || trans == FLA_CONJ_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_lhfc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_lhfr( A, T, W, B, cntl );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_lhbc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_lhbr( A, T, W, B, cntl );
}
}
}
else if ( side == FLA_RIGHT )
{
if ( trans == FLA_NO_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_rnfc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_rnfr( A, T, W, B, cntl );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_rnbc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_rnbr( A, T, W, B, cntl );
}
}
else if ( trans == FLA_TRANSPOSE || trans == FLA_CONJ_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_rhfc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_rhfr( A, T, W, B, cntl );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_Q_UT_rhbc( A, T, W, B, cntl );
else if ( storev == FLA_ROWWISE )
r_val = FLA_Apply_Q_UT_rhbr( A, T, W, B, cntl );
}
}
}
}
return r_val;
}
FLA_Error FLA_Copy_internal( FLA_Obj A, FLA_Obj B, fla_copy_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Copy_internal_check( A, B, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Copy_internal( *FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( B ),
flash_copy_cntl );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Copy( A, B, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = flash_copy_cntl_blas;
}
// Parameter combinations
if ( FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
r_val = FLA_Copy_task( A, B, cntl );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT1 )
{
r_val = FLA_Copy_blk_var1( A, B, cntl );
}
#ifdef FLA_ENABLE_NON_CRITICAL_CODE
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT2 )
{
r_val = FLA_Copy_blk_var2( A, B, cntl );
}
#endif
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT3 )
{
r_val = FLA_Copy_blk_var3( A, B, cntl );
}
#ifdef FLA_ENABLE_NON_CRITICAL_CODE
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT4 )
{
r_val = FLA_Copy_blk_var4( A, B, cntl );
}
#endif
else
{
r_val = FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
}
return r_val;
}
FLA_Error FLA_Trsm_internal( FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Trsm_internal_check( side, uplo, transa, diag, alpha, A, B, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Trsm_internal( side,
uplo,
transa,
diag,
alpha,
*FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( B ),
flash_trsm_cntl_mm );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Trsm( side, uplo, transa, diag, alpha, A, B, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = flash_trsm_cntl_blas;
}
// Parameter combinations
if ( side == FLA_LEFT )
{
if ( uplo == FLA_LOWER_TRIANGULAR )
{
if ( transa == FLA_NO_TRANSPOSE )
r_val = FLA_Trsm_lln( diag, alpha, A, B, cntl );
else if ( transa == FLA_TRANSPOSE )
r_val = FLA_Trsm_llt( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Trsm_llc( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_TRANSPOSE )
r_val = FLA_Trsm_llh( diag, alpha, A, B, cntl );
}
else if ( uplo == FLA_UPPER_TRIANGULAR )
{
if ( transa == FLA_NO_TRANSPOSE )
r_val = FLA_Trsm_lun( diag, alpha, A, B, cntl );
else if ( transa == FLA_TRANSPOSE )
r_val = FLA_Trsm_lut( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Trsm_luc( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_TRANSPOSE )
r_val = FLA_Trsm_luh( diag, alpha, A, B, cntl );
}
}
else if ( side == FLA_RIGHT )
{
if ( uplo == FLA_LOWER_TRIANGULAR )
{
if ( transa == FLA_NO_TRANSPOSE )
r_val = FLA_Trsm_rln( diag, alpha, A, B, cntl );
else if ( transa == FLA_TRANSPOSE )
r_val = FLA_Trsm_rlt( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Trsm_rlc( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_TRANSPOSE )
r_val = FLA_Trsm_rlh( diag, alpha, A, B, cntl );
}
else if ( uplo == FLA_UPPER_TRIANGULAR )
{
if ( transa == FLA_NO_TRANSPOSE )
r_val = FLA_Trsm_run( diag, alpha, A, B, cntl );
else if ( transa == FLA_TRANSPOSE )
r_val = FLA_Trsm_rut( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Trsm_ruc( diag, alpha, A, B, cntl );
else if ( transa == FLA_CONJ_TRANSPOSE )
r_val = FLA_Trsm_ruh( diag, alpha, A, B, cntl );
}
}
}
return r_val;
}
