cpucblk_cadd.c File Reference

#include "common/common.h"
#include "blend/solver.h"
#include "kernels_trace.h"
#include "pastix_ccores.h"
#include "pastix_clrcores.h"

Go to the source code of this file.

Functions
static pastix_fixdbl_t	cpucblk_cadd_frlr (pastix_complex32_t alpha, const SolverCblk *cblkA, SolverCblk *cblkB, const pastix_complex32_t *A, pastix_lrblock_t *lrB, pastix_complex32_t *work, pastix_int_t lwork, const pastix_lr_t *lowrank)
	Add a column blok in full rank format to a column blok in low rank format.
static pastix_fixdbl_t	cpucblk_cadd_lrlr (pastix_complex32_t alpha, const SolverCblk *cblkA, SolverCblk *cblkB, const pastix_lrblock_t *lrA, pastix_lrblock_t *lrB, pastix_complex32_t *work, pastix_int_t lwork, const pastix_lr_t *lowrank)
	Add two column bloks in low rank format.
static pastix_fixdbl_t	cpucblk_cadd_frfr (pastix_complex32_t alpha, const SolverCblk *cblkA, SolverCblk *cblkB, const pastix_complex32_t *A, pastix_complex32_t *B)
	Add two column bloks in full rank format.
pastix_fixdbl_t	cpucblk_cadd (pastix_complex32_t alpha, const SolverCblk *cblkA, SolverCblk *cblkB, const void *A, void *B, pastix_complex32_t *work, pastix_int_t lwork, const pastix_lr_t *lowrank)
	Add two column bloks in full rank format.

Detailed Description

Precision dependent routines to add different cblks.

Copyright

2015-2024 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.

Version

6.4.0

Author

Pierre Ramet

Mathieu Faverge

Tony Delarue

Alycia Lisito

Nolan Bredel

Date

2024-07-05

Generated arithmetic file

from /builds/2mk6rsew/0/solverstack/pastix/kernels/cpucblk_zadd.c, normal z -> c, Tue Feb 25 14:34:59 2025

Definition in file cpucblk_cadd.c.

Function Documentation

cpucblk_cadd_frlr()

static pastix_fixdbl_t cpucblk_cadd_frlr ( pastix_complex32_t  alpha, const SolverCblk *  cblkA, SolverCblk *  cblkB, const pastix_complex32_t *  A, pastix_lrblock_t *  lrB, pastix_complex32_t *  work, pastix_int_t  lwork, const pastix_lr_t *  lowrank  )

inlinestatic

Add a column blok in full rank format to a column blok in low rank format.

The second cblk is overwritten by the sum of the two column blocks. B <- alpha * A + B

Parameters

[in]	alpha	The scalar alpha
[in]	cblkA	The column block of the A matrix.
[in,out]	cblkB	The column block of the B matrix On exit, cblkB coefficient arrays are overwritten by the result of alpha * A + B.
[in,out]	A	The pointer to the coeftab of the cblk.lcoeftab matrix storing the coefficients of the panel when the Lower part is computed, cblk.ucoeftab otherwise. Must be of size cblk.stride -by- cblk.width
[in]	lrB	Pointer to the low-rank representation of the column block B. Must be followed by the low-rank representation of the following blocks.
[in]	work	Temporary memory buffer.
[in]	lwork	Temporary workspace dimension.
[in]	lowrank	The structure with low-rank parameters.

Returns

The number of flops of the operation.

Definition at line 73 of file cpucblk_cadd.c.

References core_clrmm_s::A, core_clrmm_s::alpha, core_clrmm_s::B, core_clrmm_s::beta, blok_rownbr(), core_clrmm_s::C, cblk_colnbr(), solver_cblk_s::cblktype, core_clrmm_s::Cm, core_clrmm_s::Cn, solver_blok_s::coefind, core_clradd(), solver_cblk_s::fblokptr, solver_cblk_s::fcolnum, solver_blok_s::frownum, is_block_inside_fblock(), core_clrmm_s::K, solver_cblk_s::lock, core_clrmm_s::lock, core_clrmm_s::lowrank, core_clrmm_s::lwork, core_clrmm_s::lwused, core_clrmm_s::M, core_clrmm_s::N, core_clrmm_s::offx, core_clrmm_s::offy, PastixNoTrans, pastix_lrblock_s::rk, pastix_lrblock_s::rkmax, core_clrmm_s::transA, core_clrmm_s::transB, pastix_lrblock_s::u, pastix_lrblock_s::v, and core_clrmm_s::work.

Referenced by cpucblk_cadd().

cpucblk_cadd_lrlr()

static pastix_fixdbl_t cpucblk_cadd_lrlr ( pastix_complex32_t  alpha, const SolverCblk *  cblkA, SolverCblk *  cblkB, const pastix_lrblock_t *  lrA, pastix_lrblock_t *  lrB, pastix_complex32_t *  work, pastix_int_t  lwork, const pastix_lr_t *  lowrank  )

inlinestatic

Add two column bloks in low rank format.

The second cblk is overwritten by the sum of the two column blocks. B <- alpha * A + B

Parameters

[in]	alpha	The scalar alpha
[in]	cblkA	The column block of the A matrix.
[in,out]	cblkB	The column block of the B matrix On exit, cblkB coefficient arrays are overwritten by the result of alpha * A + B.
[in]	lrA	Pointer to the low-rank representation of the column block A. Must be followed by the low-rank representation of the following blocks.
[in]	lrB	Pointer to the low-rank representation of the column block B. Must be followed by the low-rank representation of the following blocks.
[in]	work	Temporary memory buffer.
[in]	lwork	Temporary workspace dimension.
[in]	lowrank	The structure with low-rank parameters.

Returns

The number of flops of the operation.

Definition at line 185 of file cpucblk_cadd.c.

References core_clrmm_s::A, core_clrmm_s::alpha, core_clrmm_s::B, core_clrmm_s::beta, blok_rownbr(), core_clrmm_s::C, cblk_colnbr(), solver_cblk_s::cblktype, core_clrmm_s::Cm, core_clrmm_s::Cn, core_clradd(), solver_cblk_s::fblokptr, solver_cblk_s::fcolnum, solver_blok_s::frownum, is_block_inside_fblock(), core_clrmm_s::K, solver_cblk_s::lock, core_clrmm_s::lock, core_clrmm_s::lowrank, core_clrmm_s::lwork, core_clrmm_s::lwused, core_clrmm_s::M, core_clrmm_s::N, core_clrmm_s::offx, core_clrmm_s::offy, PASTIX_LRM3_ORTHOU, PastixNoTrans, core_clrmm_s::transA, core_clrmm_s::transB, and core_clrmm_s::work.

Referenced by cpucblk_cadd().

cpucblk_cadd_frfr()

static pastix_fixdbl_t cpucblk_cadd_frfr ( pastix_complex32_t  alpha, const SolverCblk *  cblkA, SolverCblk *  cblkB, const pastix_complex32_t *  A, pastix_complex32_t *  B  )

inlinestatic

Add two column bloks in full rank format.

The second cblk is overwritten by the sum of the two column blocks. B <- alpha * A + B

Parameters

[in]	alpha	The scalar alpha
[in]	cblkA	The column block of the A matrix.
[in,out]	cblkB	The column block of the B matrix On exit, cblkB coefficient arrays are overwritten by the result of alpha * A + B.
[in,out]	A	The pointer to the coeftab of the cblk.lcoeftab matrix storing the coefficients of the panel when the Lower part is computed, cblk.ucoeftab otherwise. Must be of size cblk.stride -by- cblk.width
[in,out]	B	The pointer to the coeftab of the cblk.lcoeftab matrix storing the coefficients of the panel, if Symmetric/Hermitian cases or if upper part is computed; cblk.ucoeftab otherwise. Must be of size cblk.stride -by- cblk.width

Returns

The number of flops of the operation.

Definition at line 279 of file cpucblk_cadd.c.

References blok_rownbr(), cblk_colnbr(), solver_cblk_s::cblktype, solver_blok_s::coefind, core_cgeadd(), solver_cblk_s::fblokptr, solver_cblk_s::fcolnum, solver_blok_s::frownum, is_block_inside_fblock(), pastix_int_t, PastixNoTrans, and solver_cblk_s::stride.

Referenced by cpucblk_cadd().