|
libflame
revision_anchor
|
Go to the source code of this file.
Functions | |
| FLA_Error | FLASH_Apply_CAQ_UT_inc (dim_t p, FLA_Side side, FLA_Trans trans, FLA_Direct direct, FLA_Store storev, FLA_Obj A, FLA_Obj ATW, FLA_Obj R, FLA_Obj RTW, FLA_Obj W, FLA_Obj B) |
| FLA_Error | FLA_Apply_CAQ_UT_inc_apply_panels (dim_t nb_part, FLA_Obj A, FLA_Obj ATW, FLA_Obj W, FLA_Obj B) |
| FLA_Error | FLASH_Apply_CAQ_UT_inc_create_workspace (dim_t p, FLA_Obj TW, FLA_Obj B, FLA_Obj *W) |
| FLA_Error | FLA_Apply_CAQ_UT_inc_internal (FLA_Side side, FLA_Trans trans, FLA_Direct direct, FLA_Store storev, FLA_Obj R, FLA_Obj TW, FLA_Obj W1, FLA_Obj B, fla_apcaqutinc_t *cntl) |
| FLA_Error | FLA_Apply_CAQ_UT_inc_lhfc (FLA_Obj R, FLA_Obj TW, FLA_Obj W1, FLA_Obj B, fla_apcaqutinc_t *cntl) |
| FLA_Error FLA_Apply_CAQ_UT_inc_apply_panels | ( | dim_t | nb_part, |
| FLA_Obj | A, | ||
| FLA_Obj | ATW, | ||
| FLA_Obj | W, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLASH_Apply_Q_UT_inc().
Referenced by FLASH_Apply_CAQ_UT_inc().
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj TWT, TW0,
TWB, TW1,
TW2;
FLA_Obj WT, W0,
WB, W1,
W2;
FLA_Obj BT, B0,
BB, B1,
B2;
dim_t b;
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_2x1( TW, &TWT,
&TWB, 0, FLA_TOP );
FLA_Part_2x1( W, &WT,
&WB, 0, FLA_TOP );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
while ( FLA_Obj_length( AB ) > 0 ){
b = min( nb_part, FLA_Obj_length( AB ) );
FLA_Repart_2x1_to_3x1( AT, &A0,
/* ** */ /* ** */
&A1,
AB, &A2, b, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( TWT, &TW0,
/* ** */ /* ** */
&TW1,
TWB, &TW2, b, FLA_BOTTOM );
// NOTE: we use a blocksize of 1 for W since it has exactly nb_part
// rows (where each row is a row panels of b_alg x b_flash blocks).
FLA_Repart_2x1_to_3x1( WT, &W0,
/* ** */ /* ** */
&W1,
WB, &W2, 1, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
// Apply incremental Q's associated with each block A1 to the
// corresponding block of right-hand side B1.
FLASH_Apply_Q_UT_inc( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
A1, TW1, W1, B1 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &TWT, TW0,
TW1,
/* ** */ /* ** */
&TWB, TW2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &WT, W0,
W1,
/* ** */ /* ** */
&WB, W2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Apply_CAQ_UT_inc_internal | ( | FLA_Side | side, |
| FLA_Trans | trans, | ||
| FLA_Direct | direct, | ||
| FLA_Store | storev, | ||
| FLA_Obj | R, | ||
| FLA_Obj | TW, | ||
| FLA_Obj | W1, | ||
| FLA_Obj | B, | ||
| fla_apcaqutinc_t * | cntl | ||
| ) |
References FLA_Apply_CAQ_UT_inc_internal_check(), FLA_Apply_CAQ_UT_inc_lhfc(), and FLA_Check_error_level().
Referenced by FLASH_Apply_CAQ_UT_inc().
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Apply_CAQ_UT_inc_internal_check( side, trans, direct, storev, R, TW, W, B, cntl );
if ( side == FLA_LEFT )
{
if ( trans == FLA_NO_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
}
else if ( trans == FLA_TRANSPOSE || trans == FLA_CONJ_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
r_val = FLA_Apply_CAQ_UT_inc_lhfc( R, TW, W, B, cntl );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
}
}
else if ( side == FLA_RIGHT )
{
if ( trans == FLA_NO_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
}
else if ( trans == FLA_TRANSPOSE || trans == FLA_CONJ_TRANSPOSE )
{
if ( direct == FLA_FORWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
else if ( direct == FLA_BACKWARD )
{
if ( storev == FLA_COLUMNWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
else if ( storev == FLA_ROWWISE )
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
}
}
return r_val;
}
| FLA_Error FLA_Apply_CAQ_UT_inc_lhfc | ( | FLA_Obj | R, |
| FLA_Obj | TW, | ||
| FLA_Obj | W1, | ||
| FLA_Obj | B, | ||
| fla_apcaqutinc_t * | cntl | ||
| ) |
References FLA_Apply_CAQ_UT_inc_lhfc_blk_var1().
Referenced by FLA_Apply_CAQ_UT_inc_internal().
{
return FLA_Apply_CAQ_UT_inc_lhfc_blk_var1( R, TW, W1, B, cntl );
}
| FLA_Error FLASH_Apply_CAQ_UT_inc | ( | dim_t | p, |
| FLA_Side | side, | ||
| FLA_Trans | trans, | ||
| FLA_Direct | direct, | ||
| FLA_Store | storev, | ||
| FLA_Obj | A, | ||
| FLA_Obj | ATW, | ||
| FLA_Obj | R, | ||
| FLA_Obj | RTW, | ||
| FLA_Obj | W, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Apply_CAQ_UT_inc_apply_panels(), FLA_Apply_CAQ_UT_inc_check(), FLA_Apply_CAQ_UT_inc_internal(), FLA_CAQR_UT_inc_compute_blocks_per_part(), FLA_Check_error_level(), FLA_Part_2x1(), FLASH_Queue_begin(), and FLASH_Queue_end().
Referenced by FLASH_CAQR_UT_inc_solve().
{
FLA_Error r_val;
dim_t nb_part;
FLA_Obj WT, WB;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Apply_CAQ_UT_inc_check( side, trans, direct, storev, A, ATW, R, RTW, W, B );
// Compute the partition length from the number of partitions.
nb_part = FLA_CAQR_UT_inc_compute_blocks_per_part( p, R );
// Begin a parallel region.
FLASH_Queue_begin();
// Apply the individual Q's from the incremental QR factorizations.
FLA_Apply_CAQ_UT_inc_apply_panels( nb_part, A, ATW, W, B );
FLA_Part_2x1( W, &WT,
&WB, 1, FLA_TOP );
// Apply the Q from the factorization of the upper triangular R's.
r_val = FLA_Apply_CAQ_UT_inc_internal( side, trans, direct, storev,
R, RTW, WT, B, flash_apcaqutinc_cntl );
// End the parallel region.
FLASH_Queue_end();
return r_val;
}
References FLA_Abort(), FLA_Obj_datatype(), FLA_Obj_width(), FLA_Print_message(), FLASH_Obj_create_ext(), FLASH_Obj_depth(), FLASH_Obj_scalar_length_tl(), and FLASH_Obj_scalar_width_tl().
Referenced by FLASH_CAQR_UT_inc_solve().
{
FLA_Datatype datatype;
dim_t depth;
dim_t b_alg;
dim_t b_flash;
dim_t m, n;
// Query the depth.
depth = FLASH_Obj_depth( TW );
// *** The current Apply_CAQ_UT_inc algorithm implemented assumes that
// the matrix has a hierarchical depth of 1.
if ( depth != 1 )
{
FLA_Print_message( "FLASH_Apply_CAQ_UT_inc() currently only supports matrices of depth 1",
__FILE__, __LINE__ );
FLA_Abort();
}
// Query the datatype of matrix TW.
datatype = FLA_Obj_datatype( TW );
// Inspect the length of a the top-left element of TW to get the
// algorithmic blocksize we'll use throughout the Apply_CAQ_UT_inc
// algorithm.
b_alg = FLASH_Obj_scalar_length_tl( TW );
// The width of the top-left element gives us the storage blocksize.
b_flash = FLASH_Obj_scalar_width_tl( TW );
// The element length of W need to be p: one panel for each
// factorized subproblem.
m = p;
// Query the element (not scalar) width of the right-hand side
// matrix B. This is done so we can create W with full blocks for the
// right "edge cases" of B.
n = FLA_Obj_width( B );
// Create hierarchical matrix W.
FLASH_Obj_create_ext( datatype, m * b_alg, n * b_flash,
depth, &b_alg, &b_flash,
W );
return FLA_SUCCESS;
}
1.7.6.1