SuperLU Distributed 8.2.1
Distributed memory sparse direct solver
superlu_zdefs.h
Go to the documentation of this file.
1
27#ifndef __SUPERLU_ZDEFS /* allow multiple inclusions */
28#define __SUPERLU_ZDEFS
29
30/*
31 * File name: superlu_zdefs.h
32 * Purpose: Distributed SuperLU data types and function prototypes
33 * History:
34 */
35
36#include "superlu_defs.h"
37#include "dcomplex.h"
38
39/*
40 *-- The structure used to store matrix A of the linear system and
41 * several vectors describing the transformations done to matrix A.
42 *
43 * A (SuperMatrix*)
44 * Matrix A in A*X=B, of dimension (A->nrow, A->ncol).
45 * The number of linear equations is A->nrow. The type of A can be:
46 * Stype = SLU_NC; Dtype = SLU_D; Mtype = SLU_GE.
47 *
48 * DiagScale (DiagScale_t)
49 * Specifies the form of equilibration that was done.
50 * = NOEQUIL: No equilibration.
51 * = ROW: Row equilibration, i.e., A was premultiplied by diag(R).
52 * = COL: Column equilibration, i.e., A was postmultiplied by diag(C).
53 * = BOTH: Both row and column equilibration, i.e., A was replaced
54 * by diag(R)*A*diag(C).
55 *
56 * R double*, dimension (A->nrow)
57 * The row scale factors for A.
58 * If DiagScale = ROW or BOTH, A is multiplied on the left by diag(R).
59 * If DiagScale = NOEQUIL or COL, R is not defined.
60 *
61 * C double*, dimension (A->ncol)
62 * The column scale factors for A.
63 * If DiagScale = COL or BOTH, A is multiplied on the right by diag(C).
64 * If DiagScale = NOEQUIL or ROW, C is not defined.
65 *
66 * perm_r (int*) dimension (A->nrow)
67 * Row permutation vector which defines the permutation matrix Pr,
68 * perm_r[i] = j means row i of A is in position j in Pr*A.
69 *
70 * perm_c (int*) dimension (A->ncol)
71 * Column permutation vector, which defines the
72 * permutation matrix Pc; perm_c[i] = j means column i of A is
73 * in position j in A*Pc.
74 *
75 */
76typedef struct {
77 DiagScale_t DiagScale;
78 double *R;
79 double *C;
80 int_t *perm_r;
81 int_t *perm_c;
82} zScalePermstruct_t;
83
84#if 0 // Sherry: move to superlu_defs.h
85/*-- Auxiliary data type used in PxGSTRS/PxGSTRS1. */
86typedef struct {
87 int_t lbnum; /* Row block number (local). */
88 int_t indpos; /* Starting position in Uindex[]. */
89} Ucb_indptr_t;
90#endif
91
92/*
93 * On each processor, the blocks in L are stored in compressed block
94 * column format, the blocks in U are stored in compressed block row format.
95 */
96#define MAX_LOOKAHEADS 50
97typedef struct {
98 int_t **Lrowind_bc_ptr; /* size ceil(NSUPERS/Pc);
99 free'd in distribution routinies */
100 int_t *Lrowind_bc_dat; /* size sum of sizes of Lrowind_bc_ptr[lk]) */
101 long int *Lrowind_bc_offset; /* size ceil(NSUPERS/Pc) */
102 long int Lrowind_bc_cnt;
103
104 doublecomplex **Lnzval_bc_ptr; /* size ceil(NSUPERS/Pc);
105 free'd in distribution routinies */
106 doublecomplex *Lnzval_bc_dat; /* size sum of sizes of Lnzval_bc_ptr[lk]) */
107 long int *Lnzval_bc_offset; /* size ceil(NSUPERS/Pc) */
108 long int Lnzval_bc_cnt;
109
110 doublecomplex **Linv_bc_ptr; /* size ceil(NSUPERS/Pc);
111 free'd in distribution routinies */
112 doublecomplex *Linv_bc_dat; /* size sum of sizes of Linv_bc_ptr[lk]) */
113 long int *Linv_bc_offset; /* size ceil(NSUPERS/Pc) */
114 long int Linv_bc_cnt;
115
116 int_t **Lindval_loc_bc_ptr; /* size ceil(NSUPERS/Pc);
117 pointers to locations in Lrowind_bc_ptr and Lnzval_bc_ptr;
118 free'd in distribution routinies */
119
120 int_t *Lindval_loc_bc_dat; /* size: sum of sizes of Lindval_loc_bc_ptr[lk]) */
121 long int *Lindval_loc_bc_offset; /* size ceil(NSUPERS/Pc) */
122 long int Lindval_loc_bc_cnt;
123
124 /* for new U format -> */
125 int_t **Ucolind_bc_ptr; /* size ceil(NSUPERS/Pc) */
126 int_t *Ucolind_bc_dat; /* size: sum of sizes of Ucolind_bc_ptr[lk]) */
127 int64_t *Ucolind_bc_offset; /* size ceil(NSUPERS/Pc) */
128 int64_t Ucolind_bc_cnt;
129
130 doublecomplex **Unzval_bc_ptr; /* size ceil(NSUPERS/Pc) */
131 doublecomplex *Unzval_bc_dat; /* size: sum of sizes of Unzval_bc_ptr[lk]) */
132 int64_t *Unzval_bc_offset; /* size ceil(NSUPERS/Pc) */
133 int64_t Unzval_bc_cnt;
134
135 int_t **Uindval_loc_bc_ptr; /* size ceil(NSUPERS/Pc) pointers to locations in Ucolind_bc_ptr and Unzval_bc_ptr */
136 int_t *Uindval_loc_bc_dat; /* size: sum of sizes of Uindval_loc_bc_ptr[lk]) */
137 int64_t *Uindval_loc_bc_offset; /* size ceil(NSUPERS/Pc) */
138 int64_t Uindval_loc_bc_cnt;
139 /* end for new U format <- */
140
141 int_t *Unnz; /* number of nonzeros per block column in U*/
142 int_t **Lrowind_bc_2_lsum; /* size ceil(NSUPERS/Pc) map indices of Lrowind_bc_ptr to indices of lsum */
143 doublecomplex **Uinv_bc_ptr; /* size ceil(NSUPERS/Pc) */
144 doublecomplex *Uinv_bc_dat; /* size sum of sizes of Linv_bc_ptr[lk]) */
145 long int *Uinv_bc_offset; /* size ceil(NSUPERS/Pc) */
146 long int Uinv_bc_cnt;
147
148 int_t **Ufstnz_br_ptr; /* size ceil(NSUPERS/Pr) */
149 int_t *Ufstnz_br_dat; /* size sum of sizes of Ufstnz_br_ptr[lk]) */
150 long int *Ufstnz_br_offset; /* size ceil(NSUPERS/Pr) */
151 long int Ufstnz_br_cnt;
152
153 doublecomplex **Unzval_br_ptr; /* size ceil(NSUPERS/Pr) */
154 doublecomplex *Unzval_br_dat; /* size sum of sizes of Unzval_br_ptr[lk]) */
155 long int *Unzval_br_offset; /* size ceil(NSUPERS/Pr) */
156 long int Unzval_br_cnt;
157
158 /*-- Data structures used for broadcast and reduction trees. --*/
159 C_Tree *LBtree_ptr; /* size ceil(NSUPERS/Pc) */
160 C_Tree *LRtree_ptr; /* size ceil(NSUPERS/Pr) */
161 C_Tree *UBtree_ptr; /* size ceil(NSUPERS/Pc) */
162 C_Tree *URtree_ptr; /* size ceil(NSUPERS/Pr) */
163#if 0
164 int_t *Lsub_buf; /* Buffer for the remote subscripts of L */
165 doublecomplex *Lval_buf; /* Buffer for the remote nonzeros of L */
166 int_t *Usub_buf; /* Buffer for the remote subscripts of U */
167 doublecomplex *Uval_buf; /* Buffer for the remote nonzeros of U */
168#endif
169 int_t *Lsub_buf_2[MAX_LOOKAHEADS]; /* Buffers for the remote subscripts of L*/
170 doublecomplex *Lval_buf_2[MAX_LOOKAHEADS]; /* Buffers for the remote nonzeros of L */
171 int_t *Usub_buf_2[MAX_LOOKAHEADS]; /* Buffer for the remote subscripts of U */
172 doublecomplex *Uval_buf_2[MAX_LOOKAHEADS]; /* Buffer for the remote nonzeros of U */
173 doublecomplex *ujrow; /* used in panel factorization. */
174 int_t bufmax[NBUFFERS]; /* Maximum buffer size across all MPI ranks:
175 * 0 : maximum size of Lsub_buf[]
176 * 1 : maximum size of Lval_buf[]
177 * 2 : maximum size of Usub_buf[]
178 * 3 : maximum size of Uval_buf[]
179 * 4 : maximum size of tempv[LDA]
180 */
181
182 /*-- Record communication schedule for factorization. --*/
183 int *ToRecv; /* Recv from no one (0), left (1), and up (2).*/
184 int *ToSendD; /* Whether need to send down block row. */
185 int **ToSendR; /* List of processes to send right block col. */
186
187 /*-- Record communication schedule for forward/back solves. --*/
188 /* 1/15/22 Sherry: changed int_t to int type */
189 int *fmod; /* Modification count for L-solve */
190 int **fsendx_plist; /* Column process list to send down Xk */
191 int *frecv; /* Modifications to be recv'd in proc row */
192 int nfrecvx; /* Number of Xk I will receive in L-solve */
193 int nfsendx; /* Number of Xk I will send in L-solve */
194 int *bmod; /* Modification count for U-solve */
195 int **bsendx_plist; /* Column process list to send down Xk */
196 int *brecv; /* Modifications to be recv'd in proc row */
197 int nbrecvx; /* Number of Xk I will receive in U-solve */
198 int nbsendx; /* Number of Xk I will send in U-solve */
199 int *mod_bit; /* Flag contribution from each row blocks */
200
201 /*-- Auxiliary arrays used for forward/back solves. --*/
202 int_t *ilsum; /* Starting position of each supernode in lsum
203 (local) */
204 int_t ldalsum; /* LDA of lsum (local) */
205 int_t SolveMsgSent; /* Number of actual messages sent in LU-solve */
206 int_t SolveMsgVol; /* Volume of messages sent in the solve phase */
207
208
209 /*********************/
210 /* The following variables are used in the hybrid solver */
211
212 /*-- Counts to be used in U^{-T} triangular solve. -- */
213 int_t UT_SOLVE;
214 int_t L_SOLVE;
215 int_t FRECV;
216 int_t ut_ldalsum; /* LDA of lsum (local) */
217 int_t *ut_ilsum; /* ilsum in column-wise */
218 int_t *utmod; /* Modification count for Ut-solve. */
219 int_t **ut_sendx_plist; /* Row process list to send down Xk */
220 int_t *utrecv; /* Modifications to be recev'd in proc column. */
221 int_t n_utsendx; /* Number of Xk I will receive */
222 int_t n_utrecvx; /* Number of Xk I will send */
223 int_t n_utrecvmod;
224 int_t nroot;
225 int_t *ut_modbit;
226 int_t *Urbs;
227 Ucb_indptr_t **Ucb_indptr;/* Vertical linked list pointing to Uindex[] */
228 Ucb_indptr_t *Ucb_inddat;
229 long int *Ucb_indoffset;
230 long int Ucb_indcnt;
231
232 int_t **Ucb_valptr; /* Vertical linked list pointing to Unzval[] */
233 int_t *Ucb_valdat;
234 long int *Ucb_valoffset;
235 long int Ucb_valcnt;
236
237 /* some additional counters for L solve */
238 int_t n;
239 int_t nleaf;
240 int_t nfrecvmod;
241 int_t inv; /* whether the diagonal block is inverted*/
242
243#ifdef GPU_ACC
244 /* The following variables are used in GPU trisolve */
245
246 int_t *d_Lrowind_bc_dat;
247 long int *d_Lrowind_bc_offset;
248 doublecomplex *d_Lnzval_bc_dat;
249 long int *d_Lnzval_bc_offset;
250 int_t *d_Ucolind_bc_dat;
251 int64_t *d_Ucolind_bc_offset;
252 doublecomplex *d_Unzval_bc_dat;
253 long int *d_Unzval_bc_offset;
254
255 doublecomplex *d_Linv_bc_dat ;
256 doublecomplex *d_Uinv_bc_dat ;
257 long int *d_Linv_bc_offset ;
258 long int *d_Uinv_bc_offset ;
259 int_t *d_Lindval_loc_bc_dat ;
260 int64_t *d_Lindval_loc_bc_offset ;
261 int_t *d_Uindval_loc_bc_dat ;
262 int64_t *d_Uindval_loc_bc_offset ;
263
264 // long int *d_Lindval_loc_bc_offset ;
265 // int_t *d_Urbs;
266 // int_t *d_Ufstnz_br_dat;
267 // long int *d_Ufstnz_br_offset;
268 // doublecomplex *d_Unzval_br_dat;
269 // long int *d_Unzval_br_offset;
270 // int_t *d_Ucb_valdat;
271 // long int *d_Ucb_valoffset;
272 // Ucb_indptr_t *d_Ucb_inddat;
273 // long int *d_Ucb_indoffset;
274
275 int_t *d_ilsum ;
276 int_t *d_xsup ;
277 C_Tree *d_LBtree_ptr ;
278 C_Tree *d_LRtree_ptr ;
279 C_Tree *d_UBtree_ptr ;
280 C_Tree *d_URtree_ptr ;
281#endif
282
283} zLocalLU_t;
284
285
286typedef struct {
287 int_t *etree;
288 Glu_persist_t *Glu_persist;
289 zLocalLU_t *Llu;
290 char dt;
291} zLUstruct_t;
292
293
294/*-- Data structure for communication during matrix-vector multiplication. */
295typedef struct {
296 int_t *extern_start;
297 int_t *ind_tosend; /* X indeices to be sent to other processes */
298 int_t *ind_torecv; /* X indeices to be received from other processes */
299 int_t *ptr_ind_tosend;/* Printers to ind_tosend[] (Size procs)
300 (also point to val_torecv) */
301 int_t *ptr_ind_torecv;/* Printers to ind_torecv[] (Size procs)
302 (also point to val_tosend) */
303 int *SendCounts; /* Numbers of X indices to be sent
304 (also numbers of X values to be received) */
305 int *RecvCounts; /* Numbers of X indices to be received
306 (also numbers of X values to be sent) */
307 void *val_tosend; /* X values to be sent to other processes */
308 void *val_torecv; /* X values to be received from other processes */
309 int_t TotalIndSend; /* Total number of indices to be sent
310 (also total number of values to be received) */
311 int_t TotalValSend; /* Total number of values to be sent.
312 (also total number of indices to be received) */
313} pzgsmv_comm_t;
314
315/*-- Data structure holding the information for the solution phase --*/
316typedef struct {
317 int_t *row_to_proc;
318 int_t *inv_perm_c;
319 int_t num_diag_procs, *diag_procs, *diag_len;
320 pzgsmv_comm_t *gsmv_comm; /* communication metadata for SpMV,
321 required by IterRefine. */
322 pxgstrs_comm_t *gstrs_comm; /* communication metadata for SpTRSV. */
323 int_t *A_colind_gsmv; /* After pzgsmv_init(), the global column
324 indices of A are translated into the relative
325 positions in the gathered x-vector.
326 This is re-used in repeated calls to pzgsmv() */
327 int_t *xrow_to_proc; /* used by PDSLin */
328 NRformat_loc3d* A3d; /* Point to 3D {A, B} gathered on 2D layer 0.
329 This needs to be peresistent between
330 3D factorization and solve. */
331} zSOLVEstruct_t;
332
333
334
335/*==== For 3D code ====*/
336
337// new structures for pdgstrf_4_8
338
339#if 0 // Sherry: moved to superlu_defs.h
340typedef struct
341{
342 int_t nub;
343 int_t klst;
344 int_t ldu;
345 int_t* usub;
346 doublecomplex* uval;
347} uPanelInfo_t;
348
349typedef struct
350{
351 int_t *lsub;
352 doublecomplex *lusup;
353 int_t luptr0;
354 int_t nlb; //number of l blocks
355 int_t nsupr;
356} lPanelInfo_t;
357
358/* HyP_t is the data structure to assist HALO offload of Schur-complement. */
359typedef struct
360{
361 Remain_info_t *lookAhead_info, *Remain_info;
362 Ublock_info_t *Ublock_info, *Ublock_info_Phi;
363
364 int_t first_l_block_acc , first_u_block_acc;
365 int_t last_offload ;
366 int_t *Lblock_dirty_bit, * Ublock_dirty_bit;
367 doublecomplex *lookAhead_L_buff, *Remain_L_buff;
368 int_t lookAheadBlk; /* number of blocks in look-ahead window */
369 int_t RemainBlk ; /* number of blocks outside look-ahead window */
370 int_t num_look_aheads, nsupers;
371 int_t ldu, ldu_Phi;
372 int_t num_u_blks, num_u_blks_Phi;
373
374 int_t jj_cpu;
375 doublecomplex *bigU_Phi;
376 doublecomplex *bigU_host;
377 int_t Lnbrow;
378 int_t Rnbrow;
379
380 int_t buffer_size;
381 int_t bigu_size;
382 int offloadCondition;
383 int superlu_acc_offload;
384 int nGPUStreams;
385} HyP_t;
386
387#endif // Above are moved to superlu_defs.h
388
389
390typedef struct
391{
392 int_t * Lsub_buf ;
393 doublecomplex * Lval_buf ;
394 int_t * Usub_buf ;
395 doublecomplex * Uval_buf ;
396} zLUValSubBuf_t;
397
398int_t scuStatUpdate(
399 int_t knsupc,
400 HyP_t* HyP,
401 SCT_t* SCT,
402 SuperLUStat_t *stat
403 );
404
405typedef struct
406{
407 gEtreeInfo_t gEtreeInfo;
408 int_t* iperm_c_supno;
409 int_t* myNodeCount;
410 int_t* myTreeIdxs;
411 int_t* myZeroTrIdxs;
412 int_t** treePerm;
413 sForest_t** sForests;
414 int_t* supernode2treeMap;
415 zLUValSubBuf_t *LUvsb;
416} ztrf3Dpartition_t;
417
418typedef struct
419{
420 doublecomplex *bigU;
421 doublecomplex *bigV;
422} zscuBufs_t;
423
424typedef struct
425{
426 doublecomplex* BlockLFactor;
427 doublecomplex* BlockUFactor;
428} zdiagFactBufs_t;
429
430/*=====================*/
431
432/***********************************************************************
433 * Function prototypes
434 ***********************************************************************/
435
436#ifdef __cplusplus
437extern "C" {
438#endif
439
440
441/* Supernodal LU factor related */
442extern void
443zCreate_CompCol_Matrix_dist(SuperMatrix *, int_t, int_t, int_t, doublecomplex *,
444 int_t *, int_t *, Stype_t, Dtype_t, Mtype_t);
445extern void
446zCreate_CompRowLoc_Matrix_dist(SuperMatrix *, int_t, int_t, int_t, int_t,
447 int_t, doublecomplex *, int_t *, int_t *,
448 Stype_t, Dtype_t, Mtype_t);
449extern void
450zCompRow_to_CompCol_dist(int_t, int_t, int_t, doublecomplex *, int_t *, int_t *,
451 doublecomplex **, int_t **, int_t **);
452extern int
453pzCompRow_loc_to_CompCol_global(int_t, SuperMatrix *, gridinfo_t *,
454 SuperMatrix *);
455extern void
456zCopy_CompCol_Matrix_dist(SuperMatrix *, SuperMatrix *);
457extern void
458zCreate_Dense_Matrix_dist(SuperMatrix *, int_t, int_t, doublecomplex *, int_t,
459 Stype_t, Dtype_t, Mtype_t);
460extern void
461zCreate_SuperNode_Matrix_dist(SuperMatrix *, int_t, int_t, int_t, doublecomplex *,
462 int_t *, int_t *, int_t *, int_t *, int_t *,
463 Stype_t, Dtype_t, Mtype_t);
464extern void
465zCopy_Dense_Matrix_dist(int_t, int_t, doublecomplex *, int_t,
466 doublecomplex *, int_t);
467
468extern void zallocateA_dist (int_t, int_t, doublecomplex **, int_t **, int_t **);
469extern void zGenXtrue_dist (int_t, int_t, doublecomplex *, int_t);
470extern void zFillRHS_dist (char *, int_t, doublecomplex *, int_t,
471 SuperMatrix *, doublecomplex *, int_t);
472extern int zcreate_matrix(SuperMatrix *, int, doublecomplex **, int *,
473 doublecomplex **, int *, FILE *, gridinfo_t *);
474extern int zcreate_matrix_rb(SuperMatrix *, int, doublecomplex **, int *,
475 doublecomplex **, int *, FILE *, gridinfo_t *);
476extern int zcreate_matrix_dat(SuperMatrix *, int, doublecomplex **, int *,
477 doublecomplex **, int *, FILE *, gridinfo_t *);
478extern int zcreate_matrix_postfix(SuperMatrix *, int, doublecomplex **, int *,
479 doublecomplex **, int *, FILE *, char *, gridinfo_t *);
480
481extern void zScalePermstructInit(const int_t, const int_t,
482 zScalePermstruct_t *);
483extern void zScalePermstructFree(zScalePermstruct_t *);
484
485/* Driver related */
486extern void zgsequ_dist (SuperMatrix *, double *, double *, double *,
487 double *, double *, int_t *);
488extern double zlangs_dist (char *, SuperMatrix *);
489extern void zlaqgs_dist (SuperMatrix *, double *, double *, double,
490 double, double, char *);
491extern void pzgsequ (SuperMatrix *, double *, double *, double *,
492 double *, double *, int_t *, gridinfo_t *);
493extern double pzlangs (char *, SuperMatrix *, gridinfo_t *);
494extern void pzlaqgs (SuperMatrix *, double *, double *, double,
495 double, double, char *);
496extern int pzPermute_Dense_Matrix(int_t, int_t, int_t [], int_t[],
497 doublecomplex [], int, doublecomplex [], int, int,
498 gridinfo_t *);
499
500extern int sp_ztrsv_dist (char *, char *, char *, SuperMatrix *,
501 SuperMatrix *, doublecomplex *, int *);
502extern int sp_zgemv_dist (char *, doublecomplex, SuperMatrix *, doublecomplex *,
503 int, doublecomplex, doublecomplex *, int);
504extern int sp_zgemm_dist (char *, int, doublecomplex, SuperMatrix *,
505 doublecomplex *, int, doublecomplex, doublecomplex *, int);
506
507extern float zdistribute(superlu_dist_options_t *,
508 int_t, SuperMatrix *, Glu_freeable_t *,
509 zLUstruct_t *, gridinfo_t *);
510extern void pzgssvx_ABglobal(superlu_dist_options_t *, SuperMatrix *,
511 zScalePermstruct_t *, doublecomplex *,
512 int, int, gridinfo_t *, zLUstruct_t *, double *,
513 SuperLUStat_t *, int *);
514extern float pzdistribute(superlu_dist_options_t *, int_t, SuperMatrix *,
515 zScalePermstruct_t *, Glu_freeable_t *,
516 zLUstruct_t *, gridinfo_t *);
517extern void pzgssvx(superlu_dist_options_t *, SuperMatrix *,
518 zScalePermstruct_t *, doublecomplex *,
519 int, int, gridinfo_t *, zLUstruct_t *,
520 zSOLVEstruct_t *, double *, SuperLUStat_t *, int *);
521extern void pzCompute_Diag_Inv(int_t, zLUstruct_t *,gridinfo_t *, SuperLUStat_t *, int *);
522extern int zSolveInit(superlu_dist_options_t *, SuperMatrix *, int_t [], int_t [],
523 int_t, zLUstruct_t *, gridinfo_t *, zSOLVEstruct_t *);
524extern void zSolveFinalize(superlu_dist_options_t *, zSOLVEstruct_t *);
525extern void zDestroy_A3d_gathered_on_2d(zSOLVEstruct_t *, gridinfo3d_t *);
526extern int_t pzgstrs_init(int_t, int_t, int_t, int_t,
527 int_t [], int_t [], gridinfo_t *grid,
528 Glu_persist_t *, zSOLVEstruct_t *);
529extern void pxgstrs_finalize(pxgstrs_comm_t *);
530extern int zldperm_dist(int, int, int_t, int_t [], int_t [],
531 doublecomplex [], int_t *, double [], double []);
532extern int zstatic_schedule(superlu_dist_options_t *, int, int,
533 zLUstruct_t *, gridinfo_t *, SuperLUStat_t *,
534 int_t *, int_t *, int *);
535extern void zLUstructInit(const int_t, zLUstruct_t *);
536extern void zLUstructFree(zLUstruct_t *);
537extern void zDestroy_LU(int_t, gridinfo_t *, zLUstruct_t *);
538extern void zDestroy_Tree(int_t, gridinfo_t *, zLUstruct_t *);
539extern void zscatter_l (int ib, int ljb, int nsupc, int_t iukp, int_t* xsup,
540 int klst, int nbrow, int_t lptr, int temp_nbrow,
541 int_t* usub, int_t* lsub, doublecomplex *tempv,
542 int* indirect_thread, int* indirect2,
543 int_t ** Lrowind_bc_ptr, doublecomplex **Lnzval_bc_ptr,
544 gridinfo_t * grid);
545extern void zscatter_u (int ib, int jb, int nsupc, int_t iukp, int_t * xsup,
546 int klst, int nbrow, int_t lptr, int temp_nbrow,
547 int_t* lsub, int_t* usub, doublecomplex* tempv,
548 int_t ** Ufstnz_br_ptr, doublecomplex **Unzval_br_ptr,
549 gridinfo_t * grid);
550extern int_t pzgstrf(superlu_dist_options_t *, int, int, double anorm,
551 zLUstruct_t*, gridinfo_t*, SuperLUStat_t*, int*);
552
553/* #define GPU_PROF
554#define IPM_PROF */
555
556/* Solve related */
557extern void pzgstrs_Bglobal(superlu_dist_options_t *,
558 int_t, zLUstruct_t *, gridinfo_t *,
559 doublecomplex *, int_t, int, SuperLUStat_t *, int *);
560extern void pzgstrs(superlu_dist_options_t *, int_t,
561 zLUstruct_t *, zScalePermstruct_t *, gridinfo_t *,
562 doublecomplex *, int_t, int_t, int_t, int, zSOLVEstruct_t *,
563 SuperLUStat_t *, int *);
564extern void pzgstrf2_trsm(superlu_dist_options_t * options, int_t k0, int_t k,
565 double thresh, Glu_persist_t *, gridinfo_t *,
566 zLocalLU_t *, MPI_Request *, int tag_ub,
567 SuperLUStat_t *, int *info);
568extern void pzgstrs2_omp(int_t k0, int_t k, Glu_persist_t *, gridinfo_t *,
569 zLocalLU_t *, Ublock_info_t *, SuperLUStat_t *);
570extern int_t pzReDistribute_B_to_X(doublecomplex *B, int_t m_loc, int nrhs, int_t ldb,
571 int_t fst_row, int_t *ilsum, doublecomplex *x,
572 zScalePermstruct_t *, Glu_persist_t *,
573 gridinfo_t *, zSOLVEstruct_t *);
574extern void zlsum_fmod(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *,
575 int, int, int_t , int *fmod, int_t, int_t, int_t,
576 int_t *, gridinfo_t *, zLocalLU_t *,
577 MPI_Request [], SuperLUStat_t *);
578extern void zlsum_bmod(doublecomplex *, doublecomplex *, doublecomplex *,
579 int, int_t, int *bmod, int_t *, Ucb_indptr_t **,
580 int_t **, int_t *, gridinfo_t *, zLocalLU_t *,
581 MPI_Request [], SuperLUStat_t *);
582
583extern void zlsum_fmod_inv(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *,
584 int, int_t , int *fmod,
585 int_t *, gridinfo_t *, zLocalLU_t *,
586 SuperLUStat_t **, int_t *, int_t *, int_t, int_t, int_t, int_t, int, int);
587extern void zlsum_fmod_inv_master(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *,
588 int, int, int_t , int *fmod, int_t,
589 int_t *, gridinfo_t *, zLocalLU_t *,
590 SuperLUStat_t **, int_t, int_t, int_t, int_t, int, int);
591extern void zlsum_bmod_inv(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *,
592 int, int_t, int *bmod, int_t *, Ucb_indptr_t **,
593 int_t **, int_t *, gridinfo_t *, zLocalLU_t *,
594 SuperLUStat_t **, int_t *, int_t *, int_t, int_t, int, int);
595extern void zlsum_bmod_inv_master(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *,
596 int, int_t, int *bmod, int_t *, Ucb_indptr_t **,
597 int_t **, int_t *, gridinfo_t *, zLocalLU_t *,
598 SuperLUStat_t **, int_t, int_t, int, int);
599
600extern void zComputeLevelsets(int , int_t , gridinfo_t *,
601 Glu_persist_t *, zLocalLU_t *, int_t *);
602
603#ifdef GPU_ACC
604extern void pzconvertU(superlu_dist_options_t *, gridinfo_t *, zLUstruct_t *, SuperLUStat_t *, int);
605extern void zlsum_fmod_inv_gpu_wrap(int_t, int_t, int_t, int_t, doublecomplex *, doublecomplex *, int, int, int_t , int *fmod, C_Tree *, C_Tree *, int_t *, int_t *, int64_t *, doublecomplex *, int64_t *, doublecomplex *, int64_t *, int_t *, int64_t *, int_t *, gridinfo_t *, doublecomplex * , doublecomplex * , int_t );
606extern void zlsum_bmod_inv_gpu_wrap(superlu_dist_options_t *, int_t, int_t, int_t, int_t, doublecomplex *, doublecomplex *, int, int, int_t , int *bmod, C_Tree *, C_Tree *, int_t *, int_t *, int64_t *, doublecomplex *, int64_t *, doublecomplex *, int64_t *, int_t *, int64_t *,int_t *,gridinfo_t *);
607#endif
608
609extern void pzgsrfs(superlu_dist_options_t *, int_t,
610 SuperMatrix *, double, zLUstruct_t *,
611 zScalePermstruct_t *, gridinfo_t *,
612 doublecomplex [], int_t, doublecomplex [], int_t, int,
613 zSOLVEstruct_t *, double *, SuperLUStat_t *, int *);
614extern void pzgsrfs_ABXglobal(superlu_dist_options_t *, int_t,
615 SuperMatrix *, double, zLUstruct_t *,
616 gridinfo_t *, doublecomplex *, int_t, doublecomplex *, int_t,
617 int, double *, SuperLUStat_t *, int *);
618extern int pzgsmv_AXglobal_setup(SuperMatrix *, Glu_persist_t *,
619 gridinfo_t *, int_t *, int_t *[],
620 doublecomplex *[], int_t *[], int_t []);
621extern int pzgsmv_AXglobal(int_t, int_t [], doublecomplex [], int_t [],
622 doublecomplex [], doublecomplex []);
623extern int pzgsmv_AXglobal_abs(int_t, int_t [], doublecomplex [], int_t [],
624 doublecomplex [], double []);
625extern void pzgsmv_init(SuperMatrix *, int_t *, gridinfo_t *,
626 pzgsmv_comm_t *);
627extern void pzgsmv(int_t, SuperMatrix *, gridinfo_t *, pzgsmv_comm_t *,
628 doublecomplex x[], doublecomplex ax[]);
629extern void pzgsmv_finalize(pzgsmv_comm_t *);
630
631/* Memory-related */
632extern doublecomplex *doublecomplexMalloc_dist(int_t);
633extern doublecomplex *doublecomplexCalloc_dist(int_t);
634extern double *doubleMalloc_dist(int_t);
635extern double *doubleCalloc_dist(int_t);
636extern void *duser_malloc_dist (int_t, int_t);
637extern void duser_free_dist (int_t, int_t);
638extern int_t zQuerySpace_dist(int_t, zLUstruct_t *, gridinfo_t *,
639 SuperLUStat_t *, superlu_dist_mem_usage_t *);
640
641/* Auxiliary routines */
642
643extern void zClone_CompRowLoc_Matrix_dist(SuperMatrix *, SuperMatrix *);
644extern void zCopy_CompRowLoc_Matrix_dist(SuperMatrix *, SuperMatrix *);
645extern void zZero_CompRowLoc_Matrix_dist(SuperMatrix *);
646extern void zScaleAddId_CompRowLoc_Matrix_dist(SuperMatrix *, doublecomplex);
647extern void zScaleAdd_CompRowLoc_Matrix_dist(SuperMatrix *, SuperMatrix *, doublecomplex);
648extern void zZeroLblocks(int, int, gridinfo_t *, zLUstruct_t *);
649extern void zZeroUblocks(int iam, int n, gridinfo_t *, zLUstruct_t *);
650extern void zfill_dist (doublecomplex *, int_t, doublecomplex);
651extern void zinf_norm_error_dist (int_t, int_t, doublecomplex*, int_t,
652 doublecomplex*, int_t, gridinfo_t*);
653extern void pzinf_norm_error(int, int_t, int_t, doublecomplex [], int_t,
654 doublecomplex [], int_t , MPI_Comm);
655extern void zreadhb_dist (int, FILE *, int_t *, int_t *, int_t *,
656 doublecomplex **, int_t **, int_t **);
657extern void zreadtriple_dist(FILE *, int_t *, int_t *, int_t *,
658 doublecomplex **, int_t **, int_t **);
659extern void zreadtriple_noheader(FILE *, int_t *, int_t *, int_t *,
660 doublecomplex **, int_t **, int_t **);
661extern void zreadrb_dist(int, FILE *, int_t *, int_t *, int_t *,
662 doublecomplex **, int_t **, int_t **);
663extern void zreadMM_dist(FILE *, int_t *, int_t *, int_t *,
664 doublecomplex **, int_t **, int_t **);
665extern int zread_binary(FILE *, int_t *, int_t *, int_t *,
666 doublecomplex **, int_t **, int_t **);
667
668/* Distribute the data for numerical factorization */
669extern float zdist_psymbtonum(superlu_dist_options_t *, int_t, SuperMatrix *,
670 zScalePermstruct_t *, Pslu_freeable_t *,
671 zLUstruct_t *, gridinfo_t *);
672extern void pzGetDiagU(int_t, zLUstruct_t *, gridinfo_t *, doublecomplex *);
673
674extern int z_c2cpp_GetHWPM(SuperMatrix *, gridinfo_t *, zScalePermstruct_t *);
675
676/* Routines for debugging */
677extern void zPrintLblocks(int, int_t, gridinfo_t *, Glu_persist_t *,
678 zLocalLU_t *);
679extern void zPrintUblocks(int, int_t, gridinfo_t *, Glu_persist_t *,
680 zLocalLU_t *);
681extern void zPrint_CompCol_Matrix_dist(SuperMatrix *);
682extern void zPrint_Dense_Matrix_dist(SuperMatrix *);
683extern int zPrint_CompRowLoc_Matrix_dist(SuperMatrix *);
684extern int file_zPrint_CompRowLoc_Matrix_dist(FILE *fp, SuperMatrix *A);
685extern void PrintDoublecomplex(char *, int_t, doublecomplex *);
686extern int file_PrintDoublecomplex(FILE *fp, char *, int_t, doublecomplex *);
687
688extern void zGenCOOLblocks(int, int_t, gridinfo_t*,
689 Glu_persist_t*, zLocalLU_t *, int_t** , int_t** , doublecomplex ** , int_t* , int_t* );
690extern void zGenCSCLblocks(int, int_t, gridinfo_t*,
691 Glu_persist_t*, zLocalLU_t *, doublecomplex **, int_t **, int_t **, int_t*, int_t*);
692extern void zGenCSRLblocks(int, int_t, gridinfo_t*,
693 Glu_persist_t*, zLocalLU_t *, doublecomplex **, int_t **, int_t **, int_t*, int_t*);
694
695
696/* BLAS */
697
698#ifdef USE_VENDOR_BLAS
699extern void zgemm_(const char*, const char*, const int*, const int*, const int*,
700 const doublecomplex*, const doublecomplex*, const int*, const doublecomplex*,
701 const int*, const doublecomplex*, doublecomplex*, const int*, int, int);
702extern void ztrsv_(char*, char*, char*, int*, doublecomplex*, int*,
703 doublecomplex*, int*, int, int, int);
704extern void ztrsm_(const char*, const char*, const char*, const char*,
705 const int*, const int*, const doublecomplex*, const doublecomplex*, const int*,
706 doublecomplex*, const int*, int, int, int, int);
707extern void zgemv_(const char *, const int *, const int *, const doublecomplex *,
708 const doublecomplex *a, const int *, const doublecomplex *, const int *,
709 const doublecomplex *, doublecomplex *, const int *, int);
710
711#else
712extern int zgemm_(const char*, const char*, const int*, const int*, const int*,
713 const doublecomplex*, const doublecomplex*, const int*, const doublecomplex*,
714 const int*, const doublecomplex*, doublecomplex*, const int*);
715extern int ztrsv_(char*, char*, char*, int*, doublecomplex*, int*,
716 doublecomplex*, int*);
717extern int ztrsm_(const char*, const char*, const char*, const char*,
718 const int*, const int*, const doublecomplex*, const doublecomplex*, const int*,
719 doublecomplex*, const int*);
720extern void zgemv_(const char *, const int *, const int *, const doublecomplex *,
721 const doublecomplex *a, const int *, const doublecomplex *, const int *,
722 const doublecomplex *, doublecomplex *, const int *);
723#endif
724
725extern void zgeru_(const int*, const int*, const doublecomplex*,
726 const doublecomplex*, const int*, const doublecomplex*, const int*,
727 doublecomplex*, const int*);
728
729extern int zscal_(const int *n, const doublecomplex *alpha, doublecomplex *dx, const int *incx);
730extern int zaxpy_(const int *n, const doublecomplex *alpha, const doublecomplex *x,
731 const int *incx, doublecomplex *y, const int *incy);
732
733/* SuperLU BLAS interface: zsuperlu_blas.c */
734extern int superlu_zgemm(const char *transa, const char *transb,
735 int m, int n, int k, doublecomplex alpha, doublecomplex *a,
736 int lda, doublecomplex *b, int ldb, doublecomplex beta, doublecomplex *c, int ldc);
737extern int superlu_ztrsm(const char *sideRL, const char *uplo,
738 const char *transa, const char *diag, const int m, const int n,
739 const doublecomplex alpha, const doublecomplex *a,
740 const int lda, doublecomplex *b, const int ldb);
741extern int superlu_zger(const int m, const int n, const doublecomplex alpha,
742 const doublecomplex *x, const int incx, const doublecomplex *y,
743 const int incy, doublecomplex *a, const int lda);
744extern int superlu_zscal(const int n, const doublecomplex alpha, doublecomplex *x, const int incx);
745extern int superlu_zaxpy(const int n, const doublecomplex alpha,
746 const doublecomplex *x, const int incx, doublecomplex *y, const int incy);
747extern int superlu_zgemv(const char *trans, const int m,
748 const int n, const doublecomplex alpha, const doublecomplex *a,
749 const int lda, const doublecomplex *x, const int incx,
750 const doublecomplex beta, doublecomplex *y, const int incy);
751extern int superlu_ztrsv(char *uplo, char *trans, char *diag,
752 int n, doublecomplex *a, int lda, doublecomplex *x, int incx);
753
754#ifdef SLU_HAVE_LAPACK
755extern void ztrtri_(char*, char*, int*, doublecomplex*, int*, int*);
756#endif
757
758/*==== For 3D code ====*/
759extern int zcreate_matrix3d(SuperMatrix *A, int nrhs, doublecomplex **rhs,
760 int *ldb, doublecomplex **x, int *ldx,
761 FILE *fp, gridinfo3d_t *grid3d);
762extern int zcreate_matrix_postfix3d(SuperMatrix *A, int nrhs, doublecomplex **rhs,
763 int *ldb, doublecomplex **x, int *ldx,
764 FILE *fp, char * postfix, gridinfo3d_t *grid3d);
765
766/* Matrix distributed in NRformat_loc in 3D process grid. It converts
767 it to a NRformat_loc distributed in 2D grid in grid-0 */
768extern void zGatherNRformat_loc3d(fact_t Fact, NRformat_loc *A, doublecomplex *B,
769 int ldb, int nrhs, gridinfo3d_t *grid3d,
770 NRformat_loc3d **);
771extern int zScatter_B3d(NRformat_loc3d *A3d, gridinfo3d_t *grid3d);
772
773extern void pzgssvx3d (superlu_dist_options_t *, SuperMatrix *,
774 zScalePermstruct_t *, doublecomplex B[], int ldb, int nrhs,
775 gridinfo3d_t *, zLUstruct_t *, zSOLVEstruct_t *,
776 double *berr, SuperLUStat_t *, int *info);
777extern int_t pzgstrf3d(superlu_dist_options_t *, int m, int n, double anorm,
778 ztrf3Dpartition_t*, SCT_t *, zLUstruct_t *,
779 gridinfo3d_t *, SuperLUStat_t *, int *);
780extern void zInit_HyP(HyP_t* HyP, zLocalLU_t *Llu, int_t mcb, int_t mrb );
781extern void Free_HyP(HyP_t* HyP);
782extern int updateDirtyBit(int_t k0, HyP_t* HyP, gridinfo_t* grid);
783
784 /* from scatter.h */
785extern void
786zblock_gemm_scatter( int_t lb, int_t j, Ublock_info_t *Ublock_info,
787 Remain_info_t *Remain_info, doublecomplex *L_mat, int ldl,
788 doublecomplex *U_mat, int ldu, doublecomplex *bigV,
789 // int_t jj0,
790 int_t knsupc, int_t klst,
791 int_t *lsub, int_t *usub, int_t ldt,
792 int_t thread_id,
793 int *indirect, int *indirect2,
794 int_t **Lrowind_bc_ptr, doublecomplex **Lnzval_bc_ptr,
795 int_t **Ufstnz_br_ptr, doublecomplex **Unzval_br_ptr,
796 int_t *xsup, gridinfo_t *, SuperLUStat_t *
797#ifdef SCATTER_PROFILE
798 , double *Host_TheadScatterMOP, double *Host_TheadScatterTimer
799#endif
800 );
801
802#ifdef _OPENMP
803/*this version uses a lock to prevent multiple thread updating the same block*/
804extern void
805zblock_gemm_scatter_lock( int_t lb, int_t j, omp_lock_t* lock,
806 Ublock_info_t *Ublock_info, Remain_info_t *Remain_info,
807 doublecomplex *L_mat, int_t ldl, doublecomplex *U_mat, int_t ldu,
808 doublecomplex *bigV,
809 // int_t jj0,
810 int_t knsupc, int_t klst,
811 int_t *lsub, int_t *usub, int_t ldt,
812 int_t thread_id,
813 int *indirect, int *indirect2,
814 int_t **Lrowind_bc_ptr, doublecomplex **Lnzval_bc_ptr,
815 int_t **Ufstnz_br_ptr, doublecomplex **Unzval_br_ptr,
816 int_t *xsup, gridinfo_t *
817#ifdef SCATTER_PROFILE
818 , double *Host_TheadScatterMOP, double *Host_TheadScatterTimer
819#endif
820 );
821#endif
822
823extern int_t
824zblock_gemm_scatterTopLeft( int_t lb, int_t j, doublecomplex* bigV,
825 int_t knsupc, int_t klst, int_t* lsub,
826 int_t * usub, int_t ldt,
827 int* indirect, int* indirect2,
828 HyP_t* HyP, zLUstruct_t *, gridinfo_t*,
829 SCT_t*SCT, SuperLUStat_t *
830 );
831extern int_t
832zblock_gemm_scatterTopRight( int_t lb, int_t j, doublecomplex* bigV,
833 int_t knsupc, int_t klst, int_t* lsub,
834 int_t * usub, int_t ldt,
835 int* indirect, int* indirect2,
836 HyP_t* HyP, zLUstruct_t *, gridinfo_t*,
837 SCT_t*SCT, SuperLUStat_t * );
838extern int_t
839zblock_gemm_scatterBottomLeft( int_t lb, int_t j, doublecomplex* bigV,
840 int_t knsupc, int_t klst, int_t* lsub,
841 int_t * usub, int_t ldt,
842 int* indirect, int* indirect2,
843 HyP_t* HyP, zLUstruct_t *, gridinfo_t*,
844 SCT_t*SCT, SuperLUStat_t * );
845extern int_t
846zblock_gemm_scatterBottomRight( int_t lb, int_t j, doublecomplex* bigV,
847 int_t knsupc, int_t klst, int_t* lsub,
848 int_t * usub, int_t ldt,
849 int* indirect, int* indirect2,
850 HyP_t* HyP, zLUstruct_t *, gridinfo_t*,
851 SCT_t*SCT, SuperLUStat_t * );
852
853 /* from gather.h */
854extern void zgather_u(int_t num_u_blks,
855 Ublock_info_t *Ublock_info, int_t * usub,
856 doublecomplex *uval, doublecomplex *bigU, int_t ldu,
857 int_t *xsup, int_t klst /* for SuperSize */
858 );
859
860extern void zgather_l( int_t num_LBlk, int_t knsupc,
861 Remain_info_t *L_info,
862 doublecomplex * lval, int_t LD_lval,
863 doublecomplex * L_buff );
864
865extern void zRgather_L(int_t k, int_t *lsub, doublecomplex *lusup, gEtreeInfo_t*,
866 Glu_persist_t *, gridinfo_t *, HyP_t *,
867 int_t *myIperm, int_t *iperm_c_supno );
868extern void zRgather_U(int_t k, int_t jj0, int_t *usub, doublecomplex *uval,
869 doublecomplex *bigU, gEtreeInfo_t*, Glu_persist_t *,
870 gridinfo_t *, HyP_t *, int_t *myIperm,
871 int_t *iperm_c_supno, int_t *perm_u);
872
873 /* from xtrf3Dpartition.h */
874extern ztrf3Dpartition_t* zinitTrf3Dpartition(int_t nsupers,
875 superlu_dist_options_t *options,
876 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d);
877extern void zDestroy_trf3Dpartition(ztrf3Dpartition_t *trf3Dpartition, gridinfo3d_t *grid3d);
878
879extern void z3D_printMemUse(ztrf3Dpartition_t* trf3Dpartition,
880 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d);
881
882//extern int* getLastDep(gridinfo_t *grid, SuperLUStat_t *stat,
883// superlu_dist_options_t *options, zLocalLU_t *Llu,
884// int_t* xsup, int_t num_look_aheads, int_t nsupers,
885// int_t * iperm_c_supno);
886
887extern void zinit3DLUstructForest( int_t* myTreeIdxs, int_t* myZeroTrIdxs,
888 sForest_t** sForests, zLUstruct_t* LUstruct,
889 gridinfo3d_t* grid3d);
890
891extern int_t zgatherAllFactoredLUFr(int_t* myZeroTrIdxs, sForest_t* sForests,
892 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d,
893 SCT_t* SCT );
894
895 /* The following are from pdgstrf2.h */
896extern int_t zLpanelUpdate(int_t off0, int_t nsupc, doublecomplex* ublk_ptr,
897 int_t ld_ujrow, doublecomplex* lusup, int_t nsupr, SCT_t*);
898extern void Local_Zgstrf2(superlu_dist_options_t *options, int_t k,
899 double thresh, doublecomplex *BlockUFactor, Glu_persist_t *,
900 gridinfo_t *, zLocalLU_t *,
901 SuperLUStat_t *, int *info, SCT_t*);
902extern int_t zTrs2_GatherU(int_t iukp, int_t rukp, int_t klst,
903 int_t nsupc, int_t ldu, int_t *usub,
904 doublecomplex* uval, doublecomplex *tempv);
905extern int_t zTrs2_ScatterU(int_t iukp, int_t rukp, int_t klst,
906 int_t nsupc, int_t ldu, int_t *usub,
907 doublecomplex* uval, doublecomplex *tempv);
908extern int_t zTrs2_GatherTrsmScatter(int_t klst, int_t iukp, int_t rukp,
909 int_t *usub, doublecomplex* uval, doublecomplex *tempv,
910 int_t knsupc, int nsupr, doublecomplex* lusup,
911 Glu_persist_t *Glu_persist) ;
912extern void pzgstrs2
913#ifdef _CRAY
914(
915 int_t m, int_t k0, int_t k, Glu_persist_t *Glu_persist, gridinfo_t *grid,
916 zLocalLU_t *Llu, SuperLUStat_t *stat, _fcd ftcs1, _fcd ftcs2, _fcd ftcs3
917);
918#else
919(
920 int_t m, int_t k0, int_t k, Glu_persist_t *Glu_persist, gridinfo_t *grid,
921 zLocalLU_t *Llu, SuperLUStat_t *stat
922);
923#endif
924
925extern void pzgstrf2(superlu_dist_options_t *, int_t nsupers, int_t k0,
926 int_t k, double thresh, Glu_persist_t *, gridinfo_t *,
927 zLocalLU_t *, MPI_Request *, int, SuperLUStat_t *, int *);
928
929 /* from p3dcomm.h */
930extern int_t zAllocLlu_3d(int_t nsupers, zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
931extern int_t zp3dScatter(int_t n, zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
932extern int_t zscatter3dLPanels(int_t nsupers,
933 zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
934extern int_t zscatter3dUPanels(int_t nsupers,
935 zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
936extern int_t zcollect3dLpanels(int_t layer, int_t nsupers, zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
937extern int_t zcollect3dUpanels(int_t layer, int_t nsupers, zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
938extern int_t zp3dCollect(int_t layer, int_t n, zLUstruct_t * LUstruct, gridinfo3d_t* grid3d);
939/*zero out LU non zero entries*/
940extern int_t zzeroSetLU(int_t nnodes, int_t* nodeList , zLUstruct_t *, gridinfo3d_t*);
941extern int zAllocGlu_3d(int_t n, int_t nsupers, zLUstruct_t *);
942extern int zDeAllocLlu_3d(int_t n, zLUstruct_t *, gridinfo3d_t*);
943extern int zDeAllocGlu_3d(zLUstruct_t *);
944
945/* Reduces L and U panels of nodes in the List nodeList (size=nnnodes)
946receiver[L(nodelist)] =sender[L(nodelist)] +receiver[L(nodelist)]
947receiver[U(nodelist)] =sender[U(nodelist)] +receiver[U(nodelist)]
948*/
949int_t zreduceAncestors3d(int_t sender, int_t receiver,
950 int_t nnodes, int_t* nodeList,
951 doublecomplex* Lval_buf, doublecomplex* Uval_buf,
952 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d, SCT_t* SCT);
953/*reduces all nodelists required in a level*/
954extern int zreduceAllAncestors3d(int_t ilvl, int_t* myNodeCount,
955 int_t** treePerm,
956 zLUValSubBuf_t* LUvsb,
957 zLUstruct_t* LUstruct,
958 gridinfo3d_t* grid3d,
959 SCT_t* SCT );
960/*
961 Copies factored L and U panels from sender grid to receiver grid
962 receiver[L(nodelist)] <-- sender[L(nodelist)];
963 receiver[U(nodelist)] <-- sender[U(nodelist)];
964*/
965int_t zgatherFactoredLU(int_t sender, int_t receiver,
966 int_t nnodes, int_t *nodeList, zLUValSubBuf_t* LUvsb,
967 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d,SCT_t* SCT );
968
969/*Gathers all the L and U factors to grid 0 for solve stage
970 By repeatidly calling above function*/
971int_t zgatherAllFactoredLU(ztrf3Dpartition_t* trf3Dpartition, zLUstruct_t* LUstruct,
972 gridinfo3d_t* grid3d, SCT_t* SCT );
973
974/*Distributes data in each layer and initilizes ancestors
975 as zero in required nodes*/
976int_t zinit3DLUstruct( int_t* myTreeIdxs, int_t* myZeroTrIdxs,
977 int_t* nodeCount, int_t** nodeList,
978 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d);
979
980int_t zzSendLPanel(int_t k, int_t receiver,
981 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d, SCT_t* SCT);
982int_t zzRecvLPanel(int_t k, int_t sender, doublecomplex alpha,
983 doublecomplex beta, doublecomplex* Lval_buf,
984 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d, SCT_t* SCT);
985int_t zzSendUPanel(int_t k, int_t receiver,
986 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d, SCT_t* SCT);
987int_t zzRecvUPanel(int_t k, int_t sender, doublecomplex alpha,
988 doublecomplex beta, doublecomplex* Uval_buf,
989 zLUstruct_t* LUstruct, gridinfo3d_t* grid3d, SCT_t* SCT);
990
991 /* from communication_aux.h */
992extern int_t zIBcast_LPanel (int_t k, int_t k0, int_t* lsub, doublecomplex* lusup,
993 gridinfo_t *, int* msgcnt, MPI_Request *,
994 int **ToSendR, int_t *xsup, int );
995extern int_t zBcast_LPanel(int_t k, int_t k0, int_t* lsub, doublecomplex* lusup,
996 gridinfo_t *, int* msgcnt, int **ToSendR,
997 int_t *xsup , SCT_t*, int);
998extern int_t zIBcast_UPanel(int_t k, int_t k0, int_t* usub, doublecomplex* uval,
999 gridinfo_t *, int* msgcnt, MPI_Request *,
1000 int *ToSendD, int );
1001extern int_t zBcast_UPanel(int_t k, int_t k0, int_t* usub, doublecomplex* uval,
1002 gridinfo_t *, int* msgcnt, int *ToSendD, SCT_t*, int);
1003extern int_t zIrecv_LPanel (int_t k, int_t k0, int_t* Lsub_buf,
1004 doublecomplex* Lval_buf, gridinfo_t *,
1005 MPI_Request *, zLocalLU_t *, int);
1006extern int_t zIrecv_UPanel(int_t k, int_t k0, int_t* Usub_buf, doublecomplex*,
1007 zLocalLU_t *, gridinfo_t*, MPI_Request *, int);
1008extern int_t zWait_URecv(MPI_Request *, int* msgcnt, SCT_t *);
1009extern int_t zWait_LRecv(MPI_Request*, int* msgcnt, int* msgcntsU,
1010 gridinfo_t *, SCT_t*);
1011extern int_t zISend_UDiagBlock(int_t k0, doublecomplex *ublk_ptr, int_t size,
1012 MPI_Request *, gridinfo_t *, int);
1013extern int_t zRecv_UDiagBlock(int_t k0, doublecomplex *ublk_ptr, int_t size,
1014 int_t src, gridinfo_t *, SCT_t*, int);
1015extern int_t zPackLBlock(int_t k, doublecomplex* Dest, Glu_persist_t *,
1016 gridinfo_t *, zLocalLU_t *);
1017extern int_t zISend_LDiagBlock(int_t k0, doublecomplex *lblk_ptr, int_t size,
1018 MPI_Request *, gridinfo_t *, int);
1019extern int_t zIRecv_UDiagBlock(int_t k0, doublecomplex *ublk_ptr, int_t size,
1020 int_t src, MPI_Request *, gridinfo_t *,
1021 SCT_t*, int);
1022extern int_t zIRecv_LDiagBlock(int_t k0, doublecomplex *L_blk_ptr, int_t size,
1023 int_t src, MPI_Request *, gridinfo_t*, SCT_t*, int);
1024extern int_t zUDiagBlockRecvWait( int_t k, int_t* IrecvPlcd_D, int_t* factored_L,
1025 MPI_Request *, gridinfo_t *, zLUstruct_t *, SCT_t *);
1026extern int_t LDiagBlockRecvWait( int_t k, int_t* factored_U, MPI_Request *, gridinfo_t *);
1027
1028#if (MPI_VERSION>2)
1029extern int_t zIBcast_UDiagBlock(int_t k, doublecomplex *ublk_ptr, int_t size,
1030 MPI_Request *, gridinfo_t *);
1031extern int_t zIBcast_LDiagBlock(int_t k, doublecomplex *lblk_ptr, int_t size,
1032 MPI_Request *, gridinfo_t *);
1033#endif
1034
1035 /* from trfCommWrapper.h */
1036extern int_t zDiagFactIBCast(int_t k, int_t k0,
1037 doublecomplex *BlockUFactor, doublecomplex *BlockLFactor,
1038 int_t* IrecvPlcd_D, MPI_Request *, MPI_Request *,
1039 MPI_Request *, MPI_Request *, gridinfo_t *,
1040 superlu_dist_options_t *, double thresh,
1041 zLUstruct_t *LUstruct, SuperLUStat_t *, int *info,
1042 SCT_t *, int tag_ub);
1043extern int_t zUPanelTrSolve( int_t k, doublecomplex* BlockLFactor, doublecomplex* bigV,
1044 int_t ldt, Ublock_info_t*, gridinfo_t *,
1045 zLUstruct_t *, SuperLUStat_t *, SCT_t *);
1046extern int_t zLPanelUpdate(int_t k, int_t* IrecvPlcd_D, int_t* factored_L,
1047 MPI_Request *, doublecomplex* BlockUFactor, gridinfo_t *,
1048 zLUstruct_t *, SCT_t *);
1049extern int_t zUPanelUpdate(int_t k, int_t* factored_U, MPI_Request *,
1050 doublecomplex* BlockLFactor, doublecomplex* bigV,
1051 int_t ldt, Ublock_info_t*, gridinfo_t *,
1052 zLUstruct_t *, SuperLUStat_t *, SCT_t *);
1053extern int_t zIBcastRecvLPanel(int_t k, int_t k0, int* msgcnt,
1054 MPI_Request *, MPI_Request *,
1055 int_t* Lsub_buf, doublecomplex* Lval_buf,
1056 int * factored, gridinfo_t *, zLUstruct_t *,
1057 SCT_t *, int tag_ub);
1058extern int_t zIBcastRecvUPanel(int_t k, int_t k0, int* msgcnt, MPI_Request *,
1059 MPI_Request *, int_t* Usub_buf, doublecomplex* Uval_buf,
1060 gridinfo_t *, zLUstruct_t *, SCT_t *, int tag_ub);
1061extern int_t zWaitL(int_t k, int* msgcnt, int* msgcntU, MPI_Request *,
1062 MPI_Request *, gridinfo_t *, zLUstruct_t *, SCT_t *);
1063extern int_t zWaitU(int_t k, int* msgcnt, MPI_Request *, MPI_Request *,
1064 gridinfo_t *, zLUstruct_t *, SCT_t *);
1065extern int_t zLPanelTrSolve(int_t k, int_t* factored_L, doublecomplex* BlockUFactor,
1066 gridinfo_t *, zLUstruct_t *);
1067
1068 /* from trfAux.h */
1069extern int getNsupers(int, Glu_persist_t *);
1070extern int_t initPackLUInfo(int_t nsupers, packLUInfo_t* packLUInfo);
1071extern int freePackLUInfo(packLUInfo_t* packLUInfo);
1072extern int_t zSchurComplementSetup(int_t k, int *msgcnt, Ublock_info_t*,
1073 Remain_info_t*, uPanelInfo_t *,
1074 lPanelInfo_t *, int_t*, int_t *, int_t *,
1075 doublecomplex *bigU, int_t* Lsub_buf,
1076 doublecomplex* Lval_buf, int_t* Usub_buf,
1077 doublecomplex* Uval_buf, gridinfo_t *, zLUstruct_t *);
1078extern int_t zSchurComplementSetupGPU(int_t k, msgs_t* msgs, packLUInfo_t*,
1079 int_t*, int_t*, int_t*, gEtreeInfo_t*,
1080 factNodelists_t*, zscuBufs_t*,
1081 zLUValSubBuf_t* LUvsb, gridinfo_t *,
1082 zLUstruct_t *, HyP_t*);
1083extern doublecomplex* zgetBigV(int_t, int_t);
1084extern doublecomplex* zgetBigU(superlu_dist_options_t *,
1085 int_t, gridinfo_t *, zLUstruct_t *);
1086// permutation from superLU default
1087
1088 /* from treeFactorization.h */
1089extern int_t zLluBufInit(zLUValSubBuf_t*, zLUstruct_t *);
1090extern int_t zinitScuBufs(superlu_dist_options_t *,
1091 int_t ldt, int_t num_threads, int_t nsupers,
1092 zscuBufs_t*, zLUstruct_t*, gridinfo_t *);
1093extern int zfreeScuBufs(zscuBufs_t* scuBufs);
1094
1095#if 0 // NOT CALLED
1096// the generic tree factoring code
1097extern int_t treeFactor(
1098 int_t nnnodes, // number of nodes in the tree
1099 int_t *perm_c_supno, // list of nodes in the order of factorization
1100 commRequests_t *comReqs, // lists of communication requests
1101 zscuBufs_t *scuBufs, // contains buffers for schur complement update
1102 packLUInfo_t*packLUInfo,
1103 msgs_t*msgs,
1104 zLUValSubBuf_t* LUvsb,
1105 zdiagFactBufs_t *dFBuf,
1106 factStat_t *factStat,
1107 factNodelists_t *fNlists,
1108 superlu_dist_options_t *options,
1109 int_t * gIperm_c_supno,
1110 int_t ldt,
1111 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d, SuperLUStat_t *stat,
1112 double thresh, SCT_t *SCT,
1113 int *info
1114);
1115#endif
1116
1117extern int_t zsparseTreeFactor(
1118 int_t nnodes, // number of nodes in the tree
1119 int_t *perm_c_supno, // list of nodes in the order of factorization
1120 treeTopoInfo_t* treeTopoInfo,
1121 commRequests_t *comReqs, // lists of communication requests
1122 zscuBufs_t *scuBufs, // contains buffers for schur complement update
1123 packLUInfo_t*packLUInfo,
1124 msgs_t*msgs,
1125 zLUValSubBuf_t* LUvsb,
1126 zdiagFactBufs_t *dFBuf,
1127 factStat_t *factStat,
1128 factNodelists_t *fNlists,
1129 superlu_dist_options_t *options,
1130 int_t * gIperm_c_supno,
1131 int_t ldt,
1132 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d, SuperLUStat_t *stat,
1133 double thresh, SCT_t *SCT,
1134 int *info
1135);
1136
1137extern int_t zdenseTreeFactor(
1138 int_t nnnodes, // number of nodes in the tree
1139 int_t *perm_c_supno, // list of nodes in the order of factorization
1140 commRequests_t *comReqs, // lists of communication requests
1141 zscuBufs_t *scuBufs, // contains buffers for schur complement update
1142 packLUInfo_t*packLUInfo,
1143 msgs_t*msgs,
1144 zLUValSubBuf_t* LUvsb,
1145 zdiagFactBufs_t *dFBuf,
1146 factStat_t *factStat,
1147 factNodelists_t *fNlists,
1148 superlu_dist_options_t *options,
1149 int_t * gIperm_c_supno,
1150 int_t ldt,
1151 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d, SuperLUStat_t *stat,
1152 double thresh, SCT_t *SCT, int tag_ub,
1153 int *info
1154);
1155
1156extern int_t zsparseTreeFactor_ASYNC(
1157 sForest_t* sforest,
1158 commRequests_t **comReqss, // lists of communication requests // size maxEtree level
1159 zscuBufs_t *scuBufs, // contains buffers for schur complement update
1160 packLUInfo_t*packLUInfo,
1161 msgs_t**msgss, // size=num Look ahead
1162 zLUValSubBuf_t** LUvsbs, // size=num Look ahead
1163 zdiagFactBufs_t **dFBufs, // size maxEtree level
1164 factStat_t *factStat,
1165 factNodelists_t *fNlists,
1166 gEtreeInfo_t* gEtreeInfo, // global etree info
1167 superlu_dist_options_t *options,
1168 int_t * gIperm_c_supno,
1169 int_t ldt,
1170 HyP_t* HyP,
1171 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d, SuperLUStat_t *stat,
1172 double thresh, SCT_t *SCT, int tag_ub,
1173 int *info
1174);
1175extern zLUValSubBuf_t** zLluBufInitArr(int_t numLA, zLUstruct_t *LUstruct);
1176extern int zLluBufFreeArr(int_t numLA, zLUValSubBuf_t **LUvsbs);
1177extern zdiagFactBufs_t** zinitDiagFactBufsArr(int_t mxLeafNode, int_t ldt, gridinfo_t* grid);
1178extern int zfreeDiagFactBufsArr(int_t mxLeafNode, zdiagFactBufs_t** dFBufs);
1179extern int_t zinitDiagFactBufs(int_t ldt, zdiagFactBufs_t* dFBuf);
1180extern int_t checkRecvUDiag(int_t k, commRequests_t *comReqs,
1181 gridinfo_t *grid, SCT_t *SCT);
1182extern int_t checkRecvLDiag(int_t k, commRequests_t *comReqs, gridinfo_t *, SCT_t *);
1183
1184#if 0 // NOT CALLED
1185/* from ancFactorization.h (not called) */
1186extern int_t ancestorFactor(
1187 int_t ilvl, // level of factorization
1188 sForest_t* sforest,
1189 commRequests_t **comReqss, // lists of communication requests // size maxEtree level
1190 zscuBufs_t *scuBufs, // contains buffers for schur complement update
1191 packLUInfo_t*packLUInfo,
1192 msgs_t**msgss, // size=num Look ahead
1193 zLUValSubBuf_t** LUvsbs, // size=num Look ahead
1194 zdiagFactBufs_t **dFBufs, // size maxEtree level
1195 factStat_t *factStat,
1196 factNodelists_t *fNlists,
1197 gEtreeInfo_t* gEtreeInfo, // global etree info
1198 superlu_dist_options_t *options,
1199 int_t * gIperm_c_supno,
1200 int_t ldt,
1201 HyP_t* HyP,
1202 zLUstruct_t *LUstruct, gridinfo3d_t * grid3d, SuperLUStat_t *stat,
1203 double thresh, SCT_t *SCT, int tag_ub, int *info
1204);
1205#endif
1206
1207/*== end 3D prototypes ===================*/
1208
1209
1210#ifdef __cplusplus
1211 }
1212#endif
1213
1214#endif /* __SUPERLU_dDEFS */
1215
j
int j
Definition: dutil_dist.c:248
ztrtri_
#define ztrtri_
Definition: superlu_FCnames.h:163
NBUFFERS
#define NBUFFERS
Definition: superlu_defs.h:194
int_t
int int_t
Definition: superlu_defs.h:114
DiagScale_t
DiagScale_t
Definition: superlu_enum_consts.h:35
fact_t
fact_t
Definition: superlu_enum_consts.h:30
zcreate_matrix_postfix3d
int zcreate_matrix_postfix3d(SuperMatrix *A, int nrhs, doublecomplex **rhs, int *ldb, doublecomplex **x, int *ldx, FILE *fp, char *postfix, gridinfo3d_t *grid3d)
Definition: zcreate_matrix3d.c:71
pzgstrs_init
int_t pzgstrs_init(int_t, int_t, int_t, int_t, int_t[], int_t[], gridinfo_t *grid, Glu_persist_t *, zSOLVEstruct_t *)
Destroy distributed L & U matrices. *‍/ void zDestroy_LU(int_t n, gridinfo_t *grid,...
Definition: pzutil.c:724
initPackLUInfo
int_t initPackLUInfo(int_t nsupers, packLUInfo_t *packLUInfo)
Definition: treeFactorization.c:367
zCreate_CompRowLoc_Matrix_dist
void zCreate_CompRowLoc_Matrix_dist(SuperMatrix *, int_t, int_t, int_t, int_t, int_t, doublecomplex *, int_t *, int_t *, Stype_t, Dtype_t, Mtype_t)
zgatherFactoredLU
int_t zgatherFactoredLU(int_t sender, int_t receiver, int_t nnodes, int_t *nodeList, zLUValSubBuf_t *LUvsb, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zlsum_bmod_inv_master
void zlsum_bmod_inv_master(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *, int, int_t, int *bmod, int_t *, Ucb_indptr_t **, int_t **, int_t *, gridinfo_t *, zLocalLU_t *, SuperLUStat_t **, int_t, int_t, int, int)
Definition: pzgstrs_lsum.c:1883
zinit3DLUstructForest
void zinit3DLUstructForest(int_t *myTreeIdxs, int_t *myZeroTrIdxs, sForest_t **sForests, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zGenCSRLblocks
void zGenCSRLblocks(int, int_t, gridinfo_t *, Glu_persist_t *, zLocalLU_t *, doublecomplex **, int_t **, int_t **, int_t *, int_t *)
pzgstrs2
void pzgstrs2(int_t m, int_t k0, int_t k, Glu_persist_t *Glu_persist, gridinfo_t *grid, zLocalLU_t *Llu, SuperLUStat_t *stat)
zUDiagBlockRecvWait
int_t zUDiagBlockRecvWait(int_t k, int_t *IrecvPlcd_D, int_t *factored_L, MPI_Request *, gridinfo_t *, zLUstruct_t *, SCT_t *)
zFillRHS_dist
void zFillRHS_dist(char *, int_t, doublecomplex *, int_t, SuperMatrix *, doublecomplex *, int_t)
Let rhs[i] = sum of i-th row of A, so the solution vector is all 1's.
Definition: zutil_dist.c:506
zPrint_CompCol_Matrix_dist
void zPrint_CompCol_Matrix_dist(SuperMatrix *)
updateDirtyBit
int updateDirtyBit(int_t k0, HyP_t *HyP, gridinfo_t *grid)
Definition: sec_structs.c:618
zreduceAllAncestors3d
int zreduceAllAncestors3d(int_t ilvl, int_t *myNodeCount, int_t **treePerm, zLUValSubBuf_t *LUvsb, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zUPanelUpdate
int_t zUPanelUpdate(int_t k, int_t *factored_U, MPI_Request *, doublecomplex *BlockLFactor, doublecomplex *bigV, int_t ldt, Ublock_info_t *, gridinfo_t *, zLUstruct_t *, SuperLUStat_t *, SCT_t *)
zLPanelUpdate
int_t zLPanelUpdate(int_t k, int_t *IrecvPlcd_D, int_t *factored_L, MPI_Request *, doublecomplex *BlockUFactor, gridinfo_t *, zLUstruct_t *, SCT_t *)
pzgsmv_init
void pzgsmv_init(SuperMatrix *, int_t *, gridinfo_t *, pzgsmv_comm_t *)
Definition: pzgsmv.c:26
zDeAllocLlu_3d
int zDeAllocLlu_3d(int_t n, zLUstruct_t *, gridinfo3d_t *)
Definition: zutil_dist.c:451
zIRecv_UDiagBlock
int_t zIRecv_UDiagBlock(int_t k0, doublecomplex *ublk_ptr, int_t size, int_t src, MPI_Request *, gridinfo_t *, SCT_t *, int)
zgetBigV
doublecomplex * zgetBigV(int_t, int_t)
sp_ztrsv_dist
int sp_ztrsv_dist(char *, char *, char *, SuperMatrix *, SuperMatrix *, doublecomplex *, int *)
Definition: zsp_blas2_dist.c:94
ztrsm_
int ztrsm_(const char *, const char *, const char *, const char *, const int *, const int *, const doublecomplex *, const doublecomplex *, const int *, doublecomplex *, const int *)
zlaqgs_dist
void zlaqgs_dist(SuperMatrix *, double *, double *, double, double, double, char *)
Definition: zlaqgs_dist.c:82
pzgssvx_ABglobal
void pzgssvx_ABglobal(superlu_dist_options_t *, SuperMatrix *, zScalePermstruct_t *, doublecomplex *, int, int, gridinfo_t *, zLUstruct_t *, double *, SuperLUStat_t *, int *)
pzgsequ
void pzgsequ(SuperMatrix *, double *, double *, double *, double *, double *, int_t *, gridinfo_t *)
Definition: pzgsequ.c:85
ancestorFactor
int_t ancestorFactor(int_t ilvl, sForest_t *sforest, commRequests_t **comReqss, zscuBufs_t *scuBufs, packLUInfo_t *packLUInfo, msgs_t **msgss, zLUValSubBuf_t **LUvsbs, zdiagFactBufs_t **dFBufs, factStat_t *factStat, factNodelists_t *fNlists, gEtreeInfo_t *gEtreeInfo, superlu_dist_options_t *options, int_t *gIperm_c_supno, int_t ldt, HyP_t *HyP, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SuperLUStat_t *stat, double thresh, SCT_t *SCT, int tag_ub, int *info)
pzgsrfs
void pzgsrfs(int_t, SuperMatrix *, double, zLUstruct_t *, zScalePermstruct_t *, gridinfo_t *, doublecomplex[], int_t, doublecomplex[], int_t, int, zSOLVEstruct_t *, double *, SuperLUStat_t *, int *)
pzgsmv
void pzgsmv(int_t, SuperMatrix *, gridinfo_t *, pzgsmv_comm_t *, doublecomplex x[], doublecomplex ax[])
Definition: pzgsmv.c:234
scuStatUpdate
int_t scuStatUpdate(int_t knsupc, HyP_t *HyP, SCT_t *SCT, SuperLUStat_t *stat)
Definition: sec_structs.c:635
Local_Zgstrf2
void Local_Zgstrf2(superlu_dist_options_t *options, int_t k, double thresh, doublecomplex *BlockUFactor, Glu_persist_t *, gridinfo_t *, zLocalLU_t *, SuperLUStat_t *, int *info, SCT_t *)
sp_zgemv_dist
int sp_zgemv_dist(char *, doublecomplex, SuperMatrix *, doublecomplex *, int, doublecomplex, doublecomplex *, int)
SpGEMV.
Definition: zsp_blas2_dist.c:395
zreadtriple_dist
void zreadtriple_dist(FILE *, int_t *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
Definition: zreadtriple.c:34
zcreate_matrix3d
int zcreate_matrix3d(SuperMatrix *A, int nrhs, doublecomplex **rhs, int *ldb, doublecomplex **x, int *ldx, FILE *fp, gridinfo3d_t *grid3d)
Definition: zcreate_matrix3d_Jake.c:67
zreadMM_dist
void zreadMM_dist(FILE *, int_t *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
Definition: zreadMM.c:37
zinitTrf3Dpartition
trf3Dpartition_t * zinitTrf3Dpartition(int_t nsupers, superlu_dist_options_t *options, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zIBcast_UPanel
int_t zIBcast_UPanel(int_t k, int_t k0, int_t *usub, doublecomplex *uval, gridinfo_t *, int *msgcnt, MPI_Request *, int *ToSendD, int)
pzgstrf2
void pzgstrf2(superlu_dist_options_t *, int_t nsupers, int_t k0, int_t k, double thresh, Glu_persist_t *, gridinfo_t *, zLocalLU_t *, MPI_Request *, int, SuperLUStat_t *, int *)
pzlangs
double pzlangs(char *, SuperMatrix *, gridinfo_t *)
Definition: pzlangs.c:64
getNsupers
int getNsupers(int, Glu_persist_t *)
Definition: trfAux.c:42
pzinf_norm_error
void pzinf_norm_error(int, int_t, int_t, doublecomplex[], int_t, doublecomplex[], int_t, MPI_Comm)
Check the inf-norm of the error vector.
Definition: pzutil.c:988
pzgssvx
void pzgssvx(superlu_dist_options_t *, SuperMatrix *, zScalePermstruct_t *, doublecomplex *, int, int, gridinfo_t *, zLUstruct_t *, zSOLVEstruct_t *, double *, SuperLUStat_t *, int *)
zCreate_CompCol_Matrix_dist
void zCreate_CompCol_Matrix_dist(SuperMatrix *, int_t, int_t, int_t, doublecomplex *, int_t *, int_t *, Stype_t, Dtype_t, Mtype_t)
pzgstrs2_omp
void pzgstrs2_omp(int_t k0, int_t k, Glu_persist_t *, gridinfo_t *, zLocalLU_t *, Ublock_info_t *, SuperLUStat_t *)
Definition: pzgstrf2.c:762
zAllocLlu_3d
int_t zAllocLlu_3d(int_t nsupers, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
superlu_zgemv
int superlu_zgemv(const char *trans, const int m, const int n, const doublecomplex alpha, const doublecomplex *a, const int lda, const doublecomplex *x, const int incx, const doublecomplex beta, doublecomplex *y, const int incy)
zInit_HyP
void zInit_HyP(HyP_t *HyP, zLocalLU_t *Llu, int_t mcb, int_t mrb)
pzCompute_Diag_Inv
void pzCompute_Diag_Inv(int_t, zLUstruct_t *, gridinfo_t *, SuperLUStat_t *, int *)
Definition: pzgstrs.c:650
zinitDiagFactBufs
int_t zinitDiagFactBufs(int_t ldt, zdiagFactBufs_t *dFBuf)
zDestroy_LU
void zDestroy_LU(int_t, gridinfo_t *, zLUstruct_t *)
Destroy distributed L & U matrices.
Definition: pzutil.c:483
Free_HyP
void Free_HyP(HyP_t *HyP)
Definition: sec_structs.c:594
zp3dScatter
int_t zp3dScatter(int_t n, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zPackLBlock
int_t zPackLBlock(int_t k, doublecomplex *Dest, Glu_persist_t *, gridinfo_t *, zLocalLU_t *)
zdist_psymbtonum
float zdist_psymbtonum(fact_t, int_t, SuperMatrix *, zScalePermstruct_t *, Pslu_freeable_t *, zLUstruct_t *, gridinfo_t *)
Definition: pzsymbfact_distdata.c:1187
zPrint_CompRowLoc_Matrix_dist
int zPrint_CompRowLoc_Matrix_dist(SuperMatrix *)
zcreate_matrix_postfix
int zcreate_matrix_postfix(SuperMatrix *, int, doublecomplex **, int *, doublecomplex **, int *, FILE *, char *, gridinfo_t *)
Definition: zcreate_matrix.c:75
zWait_LRecv
int_t zWait_LRecv(MPI_Request *, int *msgcnt, int *msgcntsU, gridinfo_t *, SCT_t *)
zQuerySpace_dist
int_t zQuerySpace_dist(int_t, zLUstruct_t *, gridinfo_t *, SuperLUStat_t *, superlu_dist_mem_usage_t *)
Definition: zmemory_dist.c:72
zlsum_fmod
void zlsum_fmod(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *, int, int, int_t, int *fmod, int_t, int_t, int_t, int_t *, gridinfo_t *, zLocalLU_t *, MPI_Request[], SuperLUStat_t *)
Definition: pzgstrs_lsum.c:61
zinitDiagFactBufsArr
zdiagFactBufs_t ** zinitDiagFactBufsArr(int_t mxLeafNode, int_t ldt, gridinfo_t *grid)
checkRecvUDiag
int_t checkRecvUDiag(int_t k, commRequests_t *comReqs, gridinfo_t *grid, SCT_t *SCT)
Definition: treeFactorization.c:401
zinitScuBufs
int_t zinitScuBufs(int_t ldt, int_t num_threads, int_t nsupers, zscuBufs_t *, zLUstruct_t *, gridinfo_t *)
zzRecvLPanel
int_t zzRecvLPanel(int_t k, int_t sender, doublecomplex alpha, doublecomplex beta, doublecomplex *Lval_buf, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zscatter3dLPanels
int_t zscatter3dLPanels(int_t nsupers, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zdistribute
float zdistribute(fact_t, int_t, SuperMatrix *, Glu_freeable_t *, zLUstruct_t *, gridinfo_t *)
Definition: zdistribute.c:62
doublecomplexMalloc_dist
doublecomplex * doublecomplexMalloc_dist(int_t)
Definition: zmemory_dist.c:154
zgatherAllFactoredLUFr
int_t zgatherAllFactoredLUFr(int_t *myZeroTrIdxs, sForest_t *sForests, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zzSendUPanel
int_t zzSendUPanel(int_t k, int_t receiver, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zblock_gemm_scatterTopRight
int_t zblock_gemm_scatterTopRight(int_t lb, int_t j, doublecomplex *bigV, int_t knsupc, int_t klst, int_t *lsub, int_t *usub, int_t ldt, int *indirect, int *indirect2, HyP_t *HyP, zLUstruct_t *, gridinfo_t *, SCT_t *SCT, SuperLUStat_t *)
zISend_LDiagBlock
int_t zISend_LDiagBlock(int_t k0, doublecomplex *lblk_ptr, int_t size, MPI_Request *, gridinfo_t *, int)
pzgstrs
void pzgstrs(int_t, zLUstruct_t *, zScalePermstruct_t *, gridinfo_t *, doublecomplex *, int_t, int_t, int_t, int, zSOLVEstruct_t *, SuperLUStat_t *, int *)
Definition: pzgstrs.c:842
zgather_l
void zgather_l(int_t num_LBlk, int_t knsupc, Remain_info_t *L_info, doublecomplex *lval, int_t LD_lval, doublecomplex *L_buff)
pzPermute_Dense_Matrix
int pzPermute_Dense_Matrix(int_t, int_t, int_t[], int_t[], doublecomplex[], int, doublecomplex[], int, int, gridinfo_t *)
Permute the distributed dense matrix: B <= perm(X). perm[i] = j means the i-th row of X is in the j-t...
Definition: pzutil.c:290
zDestroy_Tree
void zDestroy_Tree(int_t, gridinfo_t *, zLUstruct_t *)
Destroy broadcast and reduction trees used in triangular solve.
Definition: pzutil.c:433
zfreeScuBufs
int zfreeScuBufs(zscuBufs_t *scuBufs)
zgetBigU
doublecomplex * zgetBigU(int_t, gridinfo_t *, zLUstruct_t *)
zinf_norm_error_dist
void zinf_norm_error_dist(int_t, int_t, doublecomplex *, int_t, doublecomplex *, int_t, gridinfo_t *)
Check the inf-norm of the error vector.
Definition: zutil_dist.c:529
pzgstrs_Bglobal
void pzgstrs_Bglobal(int_t, zLUstruct_t *, gridinfo_t *, doublecomplex *, int_t, int, SuperLUStat_t *, int *)
Definition: pzgstrs_Bglobal.c:103
freePackLUInfo
int freePackLUInfo(packLUInfo_t *packLUInfo)
Definition: treeFactorization.c:376
zZero_CompRowLoc_Matrix_dist
void zZero_CompRowLoc_Matrix_dist(SuperMatrix *)
Sets all entries of a matrix to zero, A_{i,j}=0, for i,j=1,..,n.
Definition: zutil_dist.c:338
PrintDoublecomplex
void PrintDoublecomplex(char *, int_t, doublecomplex *)
Definition: zutil_dist.c:552
zreadrb_dist
void zreadrb_dist(int, FILE *, int_t *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
Definition: zreadrb.c:283
zTrs2_GatherTrsmScatter
int_t zTrs2_GatherTrsmScatter(int_t klst, int_t iukp, int_t rukp, int_t *usub, doublecomplex *uval, doublecomplex *tempv, int_t knsupc, int nsupr, doublecomplex *lusup, Glu_persist_t *Glu_persist)
Definition: pzgstrf2.c:716
zScatter_B3d
int zScatter_B3d(NRformat_loc3d *A3d, gridinfo3d_t *grid3d)
zdenseTreeFactor
int_t zdenseTreeFactor(int_t nnnodes, int_t *perm_c_supno, commRequests_t *comReqs, zscuBufs_t *scuBufs, packLUInfo_t *packLUInfo, msgs_t *msgs, zLUValSubBuf_t *LUvsb, zdiagFactBufs_t *dFBuf, factStat_t *factStat, factNodelists_t *fNlists, superlu_dist_options_t *options, int_t *gIperm_c_supno, int_t ldt, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SuperLUStat_t *stat, double thresh, SCT_t *SCT, int tag_ub, int *info)
zscatter_u
void zscatter_u(int ib, int jb, int nsupc, int_t iukp, int_t *xsup, int klst, int nbrow, int_t lptr, int temp_nbrow, int_t *lsub, int_t *usub, doublecomplex *tempv, int_t **Ufstnz_br_ptr, doublecomplex **Unzval_br_ptr, gridinfo_t *grid)
zaxpy_
int zaxpy_(const int *n, const doublecomplex *alpha, const doublecomplex *x, const int *incx, doublecomplex *y, const int *incy)
zComputeLevelsets
void zComputeLevelsets(int, int_t, gridinfo_t *, Glu_persist_t *, zLocalLU_t *, int_t *)
zblock_gemm_scatterTopLeft
int_t zblock_gemm_scatterTopLeft(int_t lb, int_t j, doublecomplex *bigV, int_t knsupc, int_t klst, int_t *lsub, int_t *usub, int_t ldt, int *indirect, int *indirect2, HyP_t *HyP, zLUstruct_t *, gridinfo_t *, SCT_t *SCT, SuperLUStat_t *)
zTrs2_ScatterU
int_t zTrs2_ScatterU(int_t iukp, int_t rukp, int_t klst, int_t nsupc, int_t ldu, int_t *usub, doublecomplex *uval, doublecomplex *tempv)
Definition: pzgstrf2.c:694
zIBcast_LPanel
int_t zIBcast_LPanel(int_t k, int_t k0, int_t *lsub, doublecomplex *lusup, gridinfo_t *, int *msgcnt, MPI_Request *, int **ToSendR, int_t *xsup, int)
zBcast_LPanel
int_t zBcast_LPanel(int_t k, int_t k0, int_t *lsub, doublecomplex *lusup, gridinfo_t *, int *msgcnt, int **ToSendR, int_t *xsup, SCT_t *, int)
zCompRow_to_CompCol_dist
void zCompRow_to_CompCol_dist(int_t, int_t, int_t, doublecomplex *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
zgemm_
int zgemm_(const char *, const char *, const int *, const int *, const int *, const doublecomplex *, const doublecomplex *, const int *, const doublecomplex *, const int *, const doublecomplex *, doublecomplex *, const int *)
treeFactor
int_t treeFactor(int_t nnnodes, int_t *perm_c_supno, commRequests_t *comReqs, zscuBufs_t *scuBufs, packLUInfo_t *packLUInfo, msgs_t *msgs, zLUValSubBuf_t *LUvsb, zdiagFactBufs_t *dFBuf, factStat_t *factStat, factNodelists_t *fNlists, superlu_dist_options_t *options, int_t *gIperm_c_supno, int_t ldt, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SuperLUStat_t *stat, double thresh, SCT_t *SCT, int *info)
z_c2cpp_GetHWPM
int z_c2cpp_GetHWPM(SuperMatrix *, gridinfo_t *, zScalePermstruct_t *)
Definition: z_c2cpp_GetHWPM.cpp:59
zldperm_dist
int zldperm_dist(int, int, int_t, int_t[], int_t[], doublecomplex[], int_t *, double[], double[])
Definition: zldperm_dist.c:95
pzgssvx3d
void pzgssvx3d(superlu_dist_options_t *, SuperMatrix *, zScalePermstruct_t *, doublecomplex B[], int ldb, int nrhs, gridinfo3d_t *, zLUstruct_t *, zSOLVEstruct_t *, double *berr, SuperLUStat_t *, int *info)
Definition: pzgssvx3d.c:500
zIBcastRecvLPanel
int_t zIBcastRecvLPanel(int_t k, int_t k0, int *msgcnt, MPI_Request *, MPI_Request *, int_t *Lsub_buf, doublecomplex *Lval_buf, int_t *factored, gridinfo_t *, zLUstruct_t *, SCT_t *, int tag_ub)
zBcast_UPanel
int_t zBcast_UPanel(int_t k, int_t k0, int_t *usub, doublecomplex *uval, gridinfo_t *, int *msgcnt, int *ToSendD, SCT_t *, int)
superlu_ztrsv
int superlu_ztrsv(char *uplo, char *trans, char *diag, int n, doublecomplex *a, int lda, doublecomplex *x, int incx)
zcollect3dUpanels
int_t zcollect3dUpanels(int_t layer, int_t nsupers, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zGenCOOLblocks
void zGenCOOLblocks(int, int_t, gridinfo_t *, Glu_persist_t *, zLocalLU_t *, int_t **, int_t **, doublecomplex **, int_t *, int_t *)
zcollect3dLpanels
int_t zcollect3dLpanels(int_t layer, int_t nsupers, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zgatherAllFactoredLU
int_t zgatherAllFactoredLU(trf3Dpartition_t *trf3Dpartition, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
superlu_zgemm
int superlu_zgemm(const char *transa, const char *transb, int m, int n, int k, doublecomplex alpha, doublecomplex *a, int lda, doublecomplex *b, int ldb, doublecomplex beta, doublecomplex *c, int ldc)
zCreate_Dense_Matrix_dist
void zCreate_Dense_Matrix_dist(SuperMatrix *, int_t, int_t, doublecomplex *, int_t, Stype_t, Dtype_t, Mtype_t)
zRgather_U
void zRgather_U(int_t k, int_t jj0, int_t *usub, doublecomplex *uval, doublecomplex *bigU, gEtreeInfo_t *, Glu_persist_t *, gridinfo_t *, HyP_t *, int_t *myIperm, int_t *iperm_c_supno, int_t *perm_u)
duser_malloc_dist
void * duser_malloc_dist(int_t, int_t)
Definition: dmemory_dist.c:30
LDiagBlockRecvWait
int_t LDiagBlockRecvWait(int_t k, int_t *factored_U, MPI_Request *, gridinfo_t *)
Definition: communication_aux.c:218
zsparseTreeFactor_ASYNC
int_t zsparseTreeFactor_ASYNC(sForest_t *sforest, commRequests_t **comReqss, zscuBufs_t *scuBufs, packLUInfo_t *packLUInfo, msgs_t **msgss, zLUValSubBuf_t **LUvsbs, zdiagFactBufs_t **dFBufs, factStat_t *factStat, factNodelists_t *fNlists, gEtreeInfo_t *gEtreeInfo, superlu_dist_options_t *options, int_t *gIperm_c_supno, int_t ldt, HyP_t *HyP, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SuperLUStat_t *stat, double thresh, SCT_t *SCT, int tag_ub, int *info)
zlsum_bmod_inv
void zlsum_bmod_inv(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *, int, int_t, int *bmod, int_t *, Ucb_indptr_t **, int_t **, int_t *, gridinfo_t *, zLocalLU_t *, SuperLUStat_t **, int_t *, int_t *, int_t, int_t, int, int)
Definition: pzgstrs_lsum.c:1395
doublecomplexCalloc_dist
doublecomplex * doublecomplexCalloc_dist(int_t)
Definition: zmemory_dist.c:161
zGenXtrue_dist
void zGenXtrue_dist(int_t, int_t, doublecomplex *, int_t)
Definition: zutil_dist.c:487
zfill_dist
void zfill_dist(doublecomplex *, int_t, doublecomplex)
Fills a doublecomplex precision array with a given value.
Definition: zutil_dist.c:519
pzgsmv_AXglobal_abs
int pzgsmv_AXglobal_abs(int_t, int_t[], doublecomplex[], int_t[], doublecomplex[], double[])
Definition: pzgsmv_AXglobal.c:288
zstatic_schedule
int zstatic_schedule(superlu_dist_options_t *, int, int, zLUstruct_t *, gridinfo_t *, SuperLUStat_t *, int_t *, int_t *, int *)
Definition: zstatic_schedule.c:45
superlu_zscal
int superlu_zscal(const int n, const doublecomplex alpha, doublecomplex *x, const int incx)
zLUstructFree
void zLUstructFree(zLUstruct_t *)
Deallocate LUstruct.
Definition: pzutil.c:415
zCopy_CompRowLoc_Matrix_dist
void zCopy_CompRowLoc_Matrix_dist(SuperMatrix *, SuperMatrix *)
Definition: zutil_dist.c:323
zscatter3dUPanels
int_t zscatter3dUPanels(int_t nsupers, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zcreate_matrix
int zcreate_matrix(SuperMatrix *, int, doublecomplex **, int *, doublecomplex **, int *, FILE *, gridinfo_t *)
Definition: zcreate_matrix.c:347
sp_zgemm_dist
int sp_zgemm_dist(char *, int, doublecomplex, SuperMatrix *, doublecomplex *, int, doublecomplex, doublecomplex *, int)
Definition: zsp_blas3_dist.c:124
zallocateA_dist
void zallocateA_dist(int_t, int_t, doublecomplex **, int_t **, int_t **)
Definition: zmemory_dist.c:146
zinit3DLUstruct
int_t zinit3DLUstruct(int_t *myTreeIdxs, int_t *myZeroTrIdxs, int_t *nodeCount, int_t **nodeList, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zSolveInit
int zSolveInit(superlu_dist_options_t *, SuperMatrix *, int_t[], int_t[], int_t, zLUstruct_t *, gridinfo_t *, zSOLVEstruct_t *)
Initialize the data structure for the solution phase.
Definition: pzutil.c:853
zblock_gemm_scatterBottomRight
int_t zblock_gemm_scatterBottomRight(int_t lb, int_t j, doublecomplex *bigV, int_t knsupc, int_t klst, int_t *lsub, int_t *usub, int_t ldt, int *indirect, int *indirect2, HyP_t *HyP, zLUstruct_t *, gridinfo_t *, SCT_t *SCT, SuperLUStat_t *)
zLUstructInit
void zLUstructInit(const int_t, zLUstruct_t *)
Allocate storage in LUstruct.
Definition: pzutil.c:401
zsparseTreeFactor
int_t zsparseTreeFactor(int_t nnodes, int_t *perm_c_supno, treeTopoInfo_t *treeTopoInfo, commRequests_t *comReqs, zscuBufs_t *scuBufs, packLUInfo_t *packLUInfo, msgs_t *msgs, zLUValSubBuf_t *LUvsb, zdiagFactBufs_t *dFBuf, factStat_t *factStat, factNodelists_t *fNlists, superlu_dist_options_t *options, int_t *gIperm_c_supno, int_t ldt, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SuperLUStat_t *stat, double thresh, SCT_t *SCT, int *info)
pzlaqgs
void pzlaqgs(SuperMatrix *, double *, double *, double, double, double, char *)
Definition: pzlaqgs.c:84
pzCompRow_loc_to_CompCol_global
int pzCompRow_loc_to_CompCol_global(int_t, SuperMatrix *, gridinfo_t *, SuperMatrix *)
Gather A from the distributed compressed row format to global A in compressed column format.
Definition: pzutil.c:28
zAllocGlu_3d
int zAllocGlu_3d(int_t n, int_t nsupers, zLUstruct_t *)
Definition: zutil_dist.c:433
zZeroUblocks
void zZeroUblocks(int iam, int n, gridinfo_t *, zLUstruct_t *)
Sets all entries of matrix U to zero.
Definition: zutil_dist.c:794
zWait_URecv
int_t zWait_URecv(MPI_Request *, int *msgcnt, SCT_t *)
zDiagFactIBCast
int_t zDiagFactIBCast(int_t k, int_t k0, doublecomplex *BlockUFactor, doublecomplex *BlockLFactor, int_t *IrecvPlcd_D, MPI_Request *, MPI_Request *, MPI_Request *, MPI_Request *, gridinfo_t *, superlu_dist_options_t *, double thresh, zLUstruct_t *LUstruct, SuperLUStat_t *, int *info, SCT_t *, int tag_ub)
zScalePermstructFree
void zScalePermstructFree(zScalePermstruct_t *)
Deallocate ScalePermstruct.
Definition: zutil_dist.c:410
zIrecv_UPanel
int_t zIrecv_UPanel(int_t k, int_t k0, int_t *Usub_buf, doublecomplex *, zLocalLU_t *, gridinfo_t *, MPI_Request *, int)
zGatherNRformat_loc3d
void zGatherNRformat_loc3d(fact_t Fact, NRformat_loc *A, doublecomplex *B, int ldb, int nrhs, gridinfo3d_t *grid3d, NRformat_loc3d **)
pxgstrs_finalize
void pxgstrs_finalize(pxgstrs_comm_t *)
Definition: util.c:266
zgsequ_dist
void zgsequ_dist(SuperMatrix *, double *, double *, double *, double *, double *, int_t *)
Definition: zgsequ_dist.c:84
zgeru_
void zgeru_(const int *, const int *, const doublecomplex *, const doublecomplex *, const int *, const doublecomplex *, const int *, doublecomplex *, const int *)
zPrintLblocks
void zPrintLblocks(int, int_t, gridinfo_t *, Glu_persist_t *, zLocalLU_t *)
Print the blocks in the factored matrix L.
Definition: zutil_dist.c:573
zTrs2_GatherU
int_t zTrs2_GatherU(int_t iukp, int_t rukp, int_t klst, int_t nsupc, int_t ldu, int_t *usub, doublecomplex *uval, doublecomplex *tempv)
Definition: pzgstrf2.c:669
z3D_printMemUse
void z3D_printMemUse(trf3Dpartition_t *trf3Dpartition, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
Definition: zmemory_dist.c:242
zWaitU
int_t zWaitU(int_t k, int *msgcnt, MPI_Request *, MPI_Request *, gridinfo_t *, zLUstruct_t *, SCT_t *)
doubleMalloc_dist
double * doubleMalloc_dist(int_t)
Definition: dmemory_dist.c:155
zscal_
int zscal_(const int *n, const doublecomplex *alpha, doublecomplex *dx, const int *incx)
zblock_gemm_scatterBottomLeft
int_t zblock_gemm_scatterBottomLeft(int_t lb, int_t j, doublecomplex *bigV, int_t knsupc, int_t klst, int_t *lsub, int_t *usub, int_t ldt, int *indirect, int *indirect2, HyP_t *HyP, zLUstruct_t *, gridinfo_t *, SCT_t *SCT, SuperLUStat_t *)
zscatter_l
void zscatter_l(int ib, int ljb, int nsupc, int_t iukp, int_t *xsup, int klst, int nbrow, int_t lptr, int temp_nbrow, int_t *usub, int_t *lsub, doublecomplex *tempv, int *indirect_thread, int *indirect2, int_t **Lrowind_bc_ptr, doublecomplex **Lnzval_bc_ptr, gridinfo_t *grid)
pzGetDiagU
void pzGetDiagU(int_t, zLUstruct_t *, gridinfo_t *, doublecomplex *)
Definition: pzGetDiagU.c:65
superlu_zaxpy
int superlu_zaxpy(const int n, const doublecomplex alpha, const doublecomplex *x, const int incx, doublecomplex *y, const int incy)
zcreate_matrix_rb
int zcreate_matrix_rb(SuperMatrix *, int, doublecomplex **, int *, doublecomplex **, int *, FILE *, gridinfo_t *)
zCopy_CompCol_Matrix_dist
void zCopy_CompCol_Matrix_dist(SuperMatrix *, SuperMatrix *)
zClone_CompRowLoc_Matrix_dist
void zClone_CompRowLoc_Matrix_dist(SuperMatrix *, SuperMatrix *)
pzgstrf
int_t pzgstrf(superlu_dist_options_t *, int, int, double anorm, zLUstruct_t *, gridinfo_t *, SuperLUStat_t *, int *)
Definition: pzgstrf.c:241
zCreate_SuperNode_Matrix_dist
void zCreate_SuperNode_Matrix_dist(SuperMatrix *, int_t, int_t, int_t, doublecomplex *, int_t *, int_t *, int_t *, int_t *, int_t *, Stype_t, Dtype_t, Mtype_t)
zRecv_UDiagBlock
int_t zRecv_UDiagBlock(int_t k0, doublecomplex *ublk_ptr, int_t size, int_t src, gridinfo_t *, SCT_t *, int)
zIrecv_LPanel
int_t zIrecv_LPanel(int_t k, int_t k0, int_t *Lsub_buf, doublecomplex *Lval_buf, gridinfo_t *, MPI_Request *, zLocalLU_t *, int)
file_zPrint_CompRowLoc_Matrix_dist
int file_zPrint_CompRowLoc_Matrix_dist(FILE *fp, SuperMatrix *A)
zPrint_Dense_Matrix_dist
void zPrint_Dense_Matrix_dist(SuperMatrix *)
zblock_gemm_scatter
void zblock_gemm_scatter(int_t lb, int_t j, Ublock_info_t *Ublock_info, Remain_info_t *Remain_info, doublecomplex *L_mat, int ldl, doublecomplex *U_mat, int ldu, doublecomplex *bigV, int_t knsupc, int_t klst, int_t *lsub, int_t *usub, int_t ldt, int_t thread_id, int *indirect, int *indirect2, int_t **Lrowind_bc_ptr, doublecomplex **Lnzval_bc_ptr, int_t **Ufstnz_br_ptr, doublecomplex **Unzval_br_ptr, int_t *xsup, gridinfo_t *, SuperLUStat_t *)
zPrintUblocks
void zPrintUblocks(int, int_t, gridinfo_t *, Glu_persist_t *, zLocalLU_t *)
Print the blocks in the factored matrix U.
Definition: zutil_dist.c:753
zzSendLPanel
int_t zzSendLPanel(int_t k, int_t receiver, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
pzgstrf3d
int_t pzgstrf3d(superlu_dist_options_t *, int m, int n, double anorm, trf3Dpartition_t *, SCT_t *, zLUstruct_t *, gridinfo3d_t *, SuperLUStat_t *, int *)
Definition: pzgstrf3d.c:120
zISend_UDiagBlock
int_t zISend_UDiagBlock(int_t k0, doublecomplex *ublk_ptr, int_t size, MPI_Request *, gridinfo_t *, int)
pzgsrfs_ABXglobal
void pzgsrfs_ABXglobal(int_t, SuperMatrix *, double, zLUstruct_t *, gridinfo_t *, doublecomplex *, int_t, doublecomplex *, int_t, int, double *, SuperLUStat_t *, int *)
Definition: pzgsrfs_ABXglobal.c:124
zlangs_dist
double zlangs_dist(char *, SuperMatrix *)
Definition: zlangs_dist.c:61
zLPanelTrSolve
int_t zLPanelTrSolve(int_t k, int_t *factored_L, doublecomplex *BlockUFactor, gridinfo_t *, zLUstruct_t *)
zScalePermstructInit
void zScalePermstructInit(const int_t, const int_t, zScalePermstruct_t *)
Allocate storage in ScalePermstruct.
Definition: zutil_dist.c:399
zScaleAddId_CompRowLoc_Matrix_dist
void zScaleAddId_CompRowLoc_Matrix_dist(SuperMatrix *, doublecomplex)
Scale and add I: scales a matrix and adds an identity. A_{i,j} = c * A_{i,j} + \delta_{i,...
Definition: zutil_dist.c:355
zUPanelTrSolve
int_t zUPanelTrSolve(int_t k, doublecomplex *BlockLFactor, doublecomplex *bigV, int_t ldt, Ublock_info_t *, gridinfo_t *, zLUstruct_t *, SuperLUStat_t *, SCT_t *)
zLpanelUpdate
int_t zLpanelUpdate(int_t off0, int_t nsupc, doublecomplex *ublk_ptr, int_t ld_ujrow, doublecomplex *lusup, int_t nsupr, SCT_t *)
pzReDistribute_B_to_X
int_t pzReDistribute_B_to_X(doublecomplex *B, int_t m_loc, int nrhs, int_t ldb, int_t fst_row, int_t *ilsum, doublecomplex *x, zScalePermstruct_t *, Glu_persist_t *, gridinfo_t *, zSOLVEstruct_t *)
Definition: pzgstrs.c:154
zDestroy_A3d_gathered_on_2d
void zDestroy_A3d_gathered_on_2d(zSOLVEstruct_t *, gridinfo3d_t *)
Definition: pzutil.c:961
zlsum_bmod
void zlsum_bmod(doublecomplex *, doublecomplex *, doublecomplex *, int, int_t, int *bmod, int_t *, Ucb_indptr_t **, int_t **, int_t *, gridinfo_t *, zLocalLU_t *, MPI_Request[], SuperLUStat_t *)
Definition: pzgstrs_lsum.c:250
zIBcastRecvUPanel
int_t zIBcastRecvUPanel(int_t k, int_t k0, int *msgcnt, MPI_Request *, MPI_Request *, int_t *Usub_buf, doublecomplex *Uval_buf, gridinfo_t *, zLUstruct_t *, SCT_t *, int tag_ub)
superlu_zger
int superlu_zger(const int m, const int n, const doublecomplex alpha, const doublecomplex *x, const int incx, const doublecomplex *y, const int incy, doublecomplex *a, const int lda)
zreadtriple_noheader
void zreadtriple_noheader(FILE *, int_t *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
Definition: zreadtriple_noheader.c:34
zgemv_
void zgemv_(const char *, const int *, const int *, const doublecomplex *, const doublecomplex *a, const int *, const doublecomplex *, const int *, const doublecomplex *, doublecomplex *, const int *)
zzRecvUPanel
int_t zzRecvUPanel(int_t k, int_t sender, doublecomplex alpha, doublecomplex beta, doublecomplex *Uval_buf, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zCopy_Dense_Matrix_dist
void zCopy_Dense_Matrix_dist(int_t, int_t, doublecomplex *, int_t, doublecomplex *, int_t)
zLluBufInit
int_t zLluBufInit(zLUValSubBuf_t *, zLUstruct_t *)
duser_free_dist
void duser_free_dist(int_t, int_t)
Definition: dmemory_dist.c:49
pzgsmv_AXglobal_setup
int pzgsmv_AXglobal_setup(SuperMatrix *, Glu_persist_t *, gridinfo_t *, int_t *, int_t *[], doublecomplex *[], int_t *[], int_t[])
zDeAllocGlu_3d
int zDeAllocGlu_3d(zLUstruct_t *)
Definition: zutil_dist.c:443
zreadhb_dist
void zreadhb_dist(int, FILE *, int_t *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
Definition: zreadhb.c:105
zcreate_matrix_dat
int zcreate_matrix_dat(SuperMatrix *, int, doublecomplex **, int *, doublecomplex **, int *, FILE *, gridinfo_t *)
zSchurComplementSetupGPU
int_t zSchurComplementSetupGPU(int_t k, msgs_t *msgs, packLUInfo_t *, int_t *, int_t *, int_t *, gEtreeInfo_t *, factNodelists_t *, zscuBufs_t *, zLUValSubBuf_t *LUvsb, gridinfo_t *, zLUstruct_t *, HyP_t *)
checkRecvLDiag
int_t checkRecvLDiag(int_t k, commRequests_t *comReqs, gridinfo_t *, SCT_t *)
Definition: treeFactorization.c:422
file_PrintDoublecomplex
int file_PrintDoublecomplex(FILE *fp, char *, int_t, doublecomplex *)
Definition: zutil_dist.c:561
superlu_ztrsm
int superlu_ztrsm(const char *sideRL, const char *uplo, const char *transa, const char *diag, const int m, const int n, const doublecomplex alpha, const doublecomplex *a, const int lda, doublecomplex *b, const int ldb)
zp3dCollect
int_t zp3dCollect(int_t layer, int_t n, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d)
zSolveFinalize
void zSolveFinalize(superlu_dist_options_t *, zSOLVEstruct_t *)
Release the resources used for the solution phase.
Definition: pzutil.c:941
zGenCSCLblocks
void zGenCSCLblocks(int, int_t, gridinfo_t *, Glu_persist_t *, zLocalLU_t *, doublecomplex **, int_t **, int_t **, int_t *, int_t *)
ztrsv_
int ztrsv_(char *, char *, char *, int *, doublecomplex *, int *, doublecomplex *, int *)
doubleCalloc_dist
double * doubleCalloc_dist(int_t)
Definition: dmemory_dist.c:162
pzgsmv_finalize
void pzgsmv_finalize(pzgsmv_comm_t *)
Definition: pzgsmv.c:373
pzdistribute
float pzdistribute(fact_t, int_t, SuperMatrix *, zScalePermstruct_t *, Glu_freeable_t *, zLUstruct_t *, gridinfo_t *)
Definition: pzdistribute.c:325
zread_binary
int zread_binary(FILE *, int_t *, int_t *, int_t *, doublecomplex **, int_t **, int_t **)
Definition: zbinary_io.c:4
pzgstrf2_trsm
void pzgstrf2_trsm(superlu_dist_options_t *options, int_t k0, int_t k, double thresh, Glu_persist_t *, gridinfo_t *, zLocalLU_t *, MPI_Request *, int tag_ub, SuperLUStat_t *, int *info)
Definition: pzgstrf2.c:142
zreduceAncestors3d
int_t zreduceAncestors3d(int_t sender, int_t receiver, int_t nnodes, int_t *nodeList, doublecomplex *Lval_buf, doublecomplex *Uval_buf, zLUstruct_t *LUstruct, gridinfo3d_t *grid3d, SCT_t *SCT)
zzeroSetLU
int_t zzeroSetLU(int_t nnodes, int_t *nodeList, zLUstruct_t *, gridinfo3d_t *)
zDestroy_trf3Dpartition
void zDestroy_trf3Dpartition(trf3Dpartition_t *trf3Dpartition, gridinfo3d_t *grid3d)
zlsum_fmod_inv
void zlsum_fmod_inv(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *, int, int_t, int *fmod, int_t *, gridinfo_t *, zLocalLU_t *, SuperLUStat_t **, int_t *, int_t *, int_t, int_t, int_t, int_t, int, int)
Definition: pzgstrs_lsum.c:427
zLluBufInitArr
zLUValSubBuf_t ** zLluBufInitArr(int_t numLA, zLUstruct_t *LUstruct)
zZeroLblocks
void zZeroLblocks(int, int, gridinfo_t *, zLUstruct_t *)
Sets all entries of matrix L to zero.
Definition: zutil_dist.c:622
zfreeDiagFactBufsArr
int zfreeDiagFactBufsArr(int_t mxLeafNode, zdiagFactBufs_t **dFBufs)
zgather_u
void zgather_u(int_t num_u_blks, Ublock_info_t *Ublock_info, int_t *usub, doublecomplex *uval, doublecomplex *bigU, int_t ldu, int_t *xsup, int_t klst)
pzgsmv_AXglobal
int pzgsmv_AXglobal(int_t, int_t[], doublecomplex[], int_t[], doublecomplex[], doublecomplex[])
Definition: pzgsmv_AXglobal.c:258
zRgather_L
void zRgather_L(int_t k, int_t *lsub, doublecomplex *lusup, gEtreeInfo_t *, Glu_persist_t *, gridinfo_t *, HyP_t *, int_t *myIperm, int_t *iperm_c_supno)
zIRecv_LDiagBlock
int_t zIRecv_LDiagBlock(int_t k0, doublecomplex *L_blk_ptr, int_t size, int_t src, MPI_Request *, gridinfo_t *, SCT_t *, int)
zSchurComplementSetup
int_t zSchurComplementSetup(int_t k, int *msgcnt, Ublock_info_t *, Remain_info_t *, uPanelInfo_t *, lPanelInfo_t *, int_t *, int_t *, int_t *, doublecomplex *bigU, int_t *Lsub_buf, doublecomplex *Lval_buf, int_t *Usub_buf, doublecomplex *Uval_buf, gridinfo_t *, zLUstruct_t *)
zScaleAdd_CompRowLoc_Matrix_dist
void zScaleAdd_CompRowLoc_Matrix_dist(SuperMatrix *, SuperMatrix *, doublecomplex)
Scale and add: adds a scalar multiple of one matrix to another. A_{i,j} = c * A_{i,...
Definition: zutil_dist.c:381
zlsum_fmod_inv_master
void zlsum_fmod_inv_master(doublecomplex *, doublecomplex *, doublecomplex *, doublecomplex *, int, int, int_t, int *fmod, int_t, int_t *, gridinfo_t *, zLocalLU_t *, SuperLUStat_t **, int_t, int_t, int_t, int_t, int, int)
Definition: pzgstrs_lsum.c:985
zLluBufFreeArr
int zLluBufFreeArr(int_t numLA, zLUValSubBuf_t **LUvsbs)
zWaitL
int_t zWaitL(int_t k, int *msgcnt, int *msgcntU, MPI_Request *, MPI_Request *, gridinfo_t *, zLUstruct_t *, SCT_t *)
Mtype_t
Mtype_t
Definition: supermatrix.h:42
Dtype_t
Dtype_t
Definition: supermatrix.h:35
Stype_t
Stype_t
Definition: supermatrix.h:22
dcomplex.h
Header for dcomplex.c.
superlupara_mod::lsub
integer, parameter, public lsub
Definition: superlupara.f90:35
superlupara_mod::trans
integer, parameter, public trans
Definition: superlupara.f90:35
superlupara_mod::factored
integer, parameter, public factored
Definition: superlupara.f90:35
superlupara_mod::lusup
integer, parameter, public lusup
Definition: superlupara.f90:35
superlupara_mod::usub
integer, parameter, public usub
Definition: superlupara.f90:35
C_Tree
Definition: superlu_defs.h:1142
Glu_freeable_t
Definition: superlu_defs.h:490
Glu_persist_t
Definition: superlu_defs.h:435
HyP_t
Definition: superlu_ddefs.h:329
NRformat_loc3d
Definition: supermatrix.h:194
NRformat_loc
Definition: supermatrix.h:176
Pslu_freeable_t
Definition: psymbfact.h:57
Remain_info_t
Definition: superlu_defs.h:770
SCT_t
Definition: util_dist.h:172
SuperLUStat_t
Definition: util_dist.h:95
SuperMatrix
Definition: supermatrix.h:54
Ublock_info_t
Definition: superlu_defs.h:760
Ucb_indptr_t
Definition: superlu_defs.h:752
commRequests_t
Definition: superlu_defs.h:924
doublecomplex
Definition: dcomplex.h:30
factNodelists_t
Definition: superlu_defs.h:937
factStat_t
Definition: superlu_defs.h:839
gEtreeInfo_t
Definition: superlu_defs.h:890
gridinfo3d_t
Definition: superlu_defs.h:398
gridinfo_t
Definition: superlu_defs.h:388
lPanelInfo_t
Definition: superlu_ddefs.h:317
msgs_t
Definition: superlu_defs.h:947
packLUInfo_t
Definition: superlu_ddefs.h:397
pxgstrs_comm_t
Definition: superlu_defs.h:551
pzgsmv_comm_t
Definition: superlu_zdefs.h:263
pzgsmv_comm_t::val_tosend
void * val_tosend
Definition: superlu_zdefs.h:307
pzgsmv_comm_t::val_torecv
void * val_torecv
Definition: superlu_zdefs.h:308
sForest_t
Definition: superlu_defs.h:901
superlu_dist_mem_usage_t
Definition: superlu_defs.h:744
superlu_dist_options_t
Definition: superlu_defs.h:712
treeTopoInfo_t
Definition: superlu_defs.h:882
uPanelInfo_t
Definition: superlu_ddefs.h:308
zLUValSubBuf_t
Definition: superlu_zdefs.h:357
zLUstruct_t
Definition: superlu_zdefs.h:254
zLocalLU_t
Definition: superlu_zdefs.h:97
zLocalLU_t::Ucolind_bc_offset
int64_t * Ucolind_bc_offset
Definition: superlu_zdefs.h:127
zLocalLU_t::Ucolind_bc_dat
int_t * Ucolind_bc_dat
Definition: superlu_zdefs.h:126
zLocalLU_t::Unzval_bc_ptr
doublecomplex ** Unzval_bc_ptr
Definition: superlu_zdefs.h:130
zLocalLU_t::Uindval_loc_bc_dat
int_t * Uindval_loc_bc_dat
Definition: superlu_zdefs.h:136
zLocalLU_t::Unzval_bc_offset
int64_t * Unzval_bc_offset
Definition: superlu_zdefs.h:132
zLocalLU_t::Uindval_loc_bc_cnt
int64_t Uindval_loc_bc_cnt
Definition: superlu_zdefs.h:138
zLocalLU_t::Uindval_loc_bc_ptr
int_t ** Uindval_loc_bc_ptr
Definition: superlu_zdefs.h:135
zLocalLU_t::Uindval_loc_bc_offset
int64_t * Uindval_loc_bc_offset
Definition: superlu_zdefs.h:137
zLocalLU_t::Unzval_bc_dat
doublecomplex * Unzval_bc_dat
Definition: superlu_zdefs.h:131
zLocalLU_t::Unzval_bc_cnt
int64_t Unzval_bc_cnt
Definition: superlu_zdefs.h:133
zLocalLU_t::Ucolind_bc_ptr
int_t ** Ucolind_bc_ptr
Definition: superlu_zdefs.h:125
zLocalLU_t::Ucolind_bc_cnt
int64_t Ucolind_bc_cnt
Definition: superlu_zdefs.h:128
zSOLVEstruct_t
Definition: superlu_zdefs.h:284
zScalePermstruct_t
Definition: superlu_zdefs.h:76
zdiagFactBufs_t
Definition: superlu_zdefs.h:391
zscuBufs_t
Definition: superlu_zdefs.h:385
ztrf3Dpartition_t
Definition: superlu_zdefs.h:406
ztrf3Dpartition_t::sForests
sForest_t ** sForests
Definition: superlu_zdefs.h:413
ztrf3Dpartition_t::iperm_c_supno
int_t * iperm_c_supno
Definition: superlu_zdefs.h:408
ztrf3Dpartition_t::treePerm
int_t ** treePerm
Definition: superlu_zdefs.h:412
ztrf3Dpartition_t::LUvsb
zLUValSubBuf_t * LUvsb
Definition: superlu_zdefs.h:415
ztrf3Dpartition_t::supernode2treeMap
int_t * supernode2treeMap
Definition: superlu_zdefs.h:414
ztrf3Dpartition_t::myZeroTrIdxs
int_t * myZeroTrIdxs
Definition: superlu_zdefs.h:411
ztrf3Dpartition_t::myTreeIdxs
int_t * myTreeIdxs
Definition: superlu_zdefs.h:410
ztrf3Dpartition_t::myNodeCount
int_t * myNodeCount
Definition: superlu_zdefs.h:409
ztrf3Dpartition_t::gEtreeInfo
gEtreeInfo_t gEtreeInfo
Definition: superlu_zdefs.h:407
superlu_defs.h
Definitions which are precision-neutral.
MAX_LOOKAHEADS
#define MAX_LOOKAHEADS
Definition: superlu_zdefs.h:96