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;
}
FLA_Error FLASH_Axpy_hierarchy( int direction, FLA_Obj alpha, FLA_Obj F, FLA_Obj* H )
{
// Once we get down to a submatrix whose elements are scalars, we are down
// to our base case.
if ( FLA_Obj_elemtype( *H ) == FLA_SCALAR )
{
// Depending on which top-level function invoked us, we either axpy
// the source data in the flat matrix to the leaf-level submatrix of
// the hierarchical matrix, or axpy the data in the hierarchical
// submatrix to the flat matrix.
if ( direction == FLA_FLAT_TO_HIER )
{
#ifdef FLA_ENABLE_SCC
if ( FLA_is_owner() )
#endif
FLA_Axpy_external( alpha, F, *H );
}
else if ( direction == FLA_HIER_TO_FLAT )
{
#ifdef FLA_ENABLE_SCC
if ( FLA_is_owner() )
#endif
FLA_Axpy_external( alpha, *H, F );
}
}
else
{
FLA_Obj HL, HR, H0, H1, H2;
FLA_Obj FL, FR, F0, F1, F2;
FLA_Obj H1T, H01,
H1B, H11,
H21;
FLA_Obj F1T, F01,
F1B, F11,
F21;
dim_t b_m;
dim_t b_n;
FLA_Part_1x2( *H, &HL, &HR, 0, FLA_LEFT );
FLA_Part_1x2( F, &FL, &FR, 0, FLA_LEFT );
while ( FLA_Obj_width( HL ) < FLA_Obj_width( *H ) )
{
FLA_Repart_1x2_to_1x3( HL, /**/ HR, &H0, /**/ &H1, &H2,
1, FLA_RIGHT );
// Get the scalar width of H1 and use that to determine the
// width of F1.
b_n = FLASH_Obj_scalar_width( H1 );
FLA_Repart_1x2_to_1x3( FL, /**/ FR, &F0, /**/ &F1, &F2,
b_n, FLA_RIGHT );
// -------------------------------------------------------------
FLA_Part_2x1( H1, &H1T,
&H1B, 0, FLA_TOP );
FLA_Part_2x1( F1, &F1T,
&F1B, 0, FLA_TOP );
while ( FLA_Obj_length( H1T ) < FLA_Obj_length( H1 ) )
{
FLA_Repart_2x1_to_3x1( H1T, &H01,
/* ** */ /* *** */
&H11,
H1B, &H21, 1, FLA_BOTTOM );
// Get the scalar length of H11 and use that to determine the
// length of F11.
b_m = FLASH_Obj_scalar_length( H11 );
FLA_Repart_2x1_to_3x1( F1T, &F01,
/* ** */ /* *** */
&F11,
F1B, &F21, b_m, FLA_BOTTOM );
// -------------------------------------------------------------
// Recursively axpy between F11 and H11.
FLASH_Axpy_hierarchy( direction, alpha, F11,
FLASH_OBJ_PTR_AT( H11 ) );
// -------------------------------------------------------------
FLA_Cont_with_3x1_to_2x1( &H1T, H01,
H11,
/* ** */ /* *** */
&H1B, H21, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &F1T, F01,
F11,
/* ** */ /* *** */
&F1B, F21, FLA_TOP );
}
// -------------------------------------------------------------
FLA_Cont_with_1x3_to_1x2( &HL, /**/ &HR, H0, H1, /**/ H2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &FL, /**/ &FR, F0, F1, /**/ F2,
FLA_LEFT );
}
}
return FLA_SUCCESS;
}
