OpenRadioss 2025.1.11
OpenRadioss project
Loading...
Searching...
No Matches
sfac_omp_m.F
Go to the documentation of this file.
1C
2C This file is part of MUMPS 5.5.1, released
3C on Tue Jul 12 13:17:24 UTC 2022
4C
5C
6C Copyright 1991-2022 CERFACS, CNRS, ENS Lyon, INP Toulouse, Inria,
7C Mumps Technologies, University of Bordeaux.
8C
9C This version of MUMPS is provided to you free of charge. It is
10C released under the CeCILL-C license
11C (see doc/CeCILL-C_V1-en.txt, doc/CeCILL-C_V1-fr.txt, and
12C https://cecill.info/licences/Licence_CeCILL-C_V1-en.html)
13C
14 MODULE smumps_fac_omp_m
15 INTEGER(8), PARAMETER :: UnderL0 = -20_8
16 INTEGER(8), PARAMETER :: copynotstarted = -19_8
17 INTEGER(8), PARAMETER :: waitmem = -18_8
18 INTEGER(8), PARAMETER :: copyfactorsfinished = -17_8
19 INTEGER(8), PARAMETER :: allocatevidercbencours = -16_8
20 INTEGER(8), PARAMETER :: finished = -15_8
21 CONTAINS
22 SUBROUTINE smumps_fac_l0_omp(N,LIW, NSTK_STEPS, ND,
23 & FILS,STEP, FRERE, DAD, ISTEP_TO_INIV2, TAB_POS_IN_PERE,
24 & PTRIST, PTRAST, PIMASTER, PAMASTER, PTRARW, PTRAIW, ITLOC,
25 & RHS_MUMPS, RINFO, NBROOT, NBRTOT, NBROOT_UNDER_L0, UU, ICNTL,
26 & PTLUST_S, PTRFAC, INFO, KEEP, KEEP8, PROCNODE_STEPS,SLAVEF,
27 & COMM_NODES, MYID_NODES, BUFR, LBUFR, LBUFR_BYTES,
28 & INTARR, DBLARR, root, PERM, NELT, FRTPTR, FRTELT, LPTRAR,
29 & COMM_LOAD, ASS_IRECV, SEUIL, SEUIL_LDLT_NIV2,
30 & MEM_DISTRIB, NE, DKEEP,PIVNUL_LIST,LPN_LIST,
31 & LPOOL_B_L0_OMP, IPOOL_B_L0_OMP,
32 & L_VIRT_L0_OMP, VIRT_L0_OMP, VIRT_L0_OMP_MAPPING,
33 & L_PHYS_L0_OMP, PHYS_L0_OMP, PERM_L0_OMP, PTR_LEAFS_L0_OMP,
34 & L0_OMP_MAPPING, LL0_OMP_MAPPING, THREAD_LA,
35 & MUMPS_TPS_ARR, SMUMPS_TPS_ARR,
36 & NSTEPSW, OPASSW, OPELIW, NELVAW, COMP,
37 & MAXFRW, NMAXNPIVW, NPVW, NOFFNEGW,
38 & NB22T1W, NBTINYW, DET_EXPW, DET_MANTW, DET_SIGNW,
39 & LRGROUPS, L0_OMP_FACTORS, LL0_OMP_FACTORS,
40 & I4_L0_OMP, NBSTATS_I4, NBCOLS_I4,
41 & I8_L0_OMP, NBSTATS_I8, NBCOLS_I8 )
42 USE smumps_load
43!$ USE OMP_LIB
44 USE mumps_tps_m, ONLY : mumps_tps_t
45 USE smumps_tps_m, ONLY : smumps_tps_t
46 USE smumps_lr_stats
47 USE smumps_struc_def, ONLY : smumps_root_struc,
48 & smumps_l0ompfac_t
49 USE smumps_dynamic_memory_m, ONLY :
50 & smumps_dm_fac_alloc_allowed
51 IMPLICIT NONE
52 include 'mpif.h'
53 include 'mumps_tags.h'
54 include 'mumps_headers.h'
55 TYPE (SMUMPS_ROOT_STRUC) :: root
56 INTEGER N,LIW, LPTRAR,
57 & NSTEPSW, INFO(80)
58 INTEGER SLAVEF, COMM_NODES, MYID_NODES
59 INTEGER, DIMENSION(0: SLAVEF - 1) :: MEM_DISTRIB
60 INTEGER KEEP(500), ICNTL(60)
61 INTEGER(8) KEEP8(150)
62 INTEGER(8), INTENT(IN) :: THREAD_LA
63 INTEGER PROCNODE_STEPS(KEEP(28))
64 INTEGER ITLOC(N+KEEP(253))
65 REAL :: RHS_MUMPS(KEEP(255))
66 INTEGER NSTK_STEPS(KEEP(28))
67 INTEGER(8), INTENT(IN) :: PTRARW(LPTRAR), PTRAIW(LPTRAR)
68 INTEGER ND(KEEP(28))
69 INTEGER FILS(N),PTRIST(KEEP(28))
70 INTEGER STEP(N), FRERE(KEEP(28)), DAD(KEEP(28))
71 INTEGER PIMASTER(KEEP(28))
72 INTEGER PTLUST_S(KEEP(28)), PERM(N)
73 INTEGER ISTEP_TO_INIV2(KEEP(71)),
74 & tab_pos_in_pere(slavef+2,max(1,keep(56)))
75 INTEGER NE(KEEP(28))
76 REAL RINFO(40)
77 INTEGER(8) :: PAMASTER(KEEP(28)), PTRAST(KEEP(28))
78 INTEGER(8) :: PTRFAC(KEEP(28))
79 INTEGER NBROOT
80 INTEGER NBRTOT
81 INTEGER, intent(out) :: NBROOT_UNDER_L0
82 INTEGER COMM_LOAD, ASS_IRECV
83 REAL UU, SEUIL, SEUIL_LDLT_NIV2
84 INTEGER NELT
85 INTEGER FRTPTR( N+1 ), FRTELT( NELT )
86 INTEGER LBUFR, LBUFR_BYTES
87 INTEGER BUFR( LBUFR )
88 REAL DBLARR( KEEP8(26) )
89 INTEGER INTARR( KEEP8(27) )
90 INTEGER LPN_LIST
91 INTEGER PIVNUL_LIST(LPN_LIST)
92 REAL DKEEP(230)
93 DOUBLE PRECISION :: OPASSW, OPELIW
94 INTEGER LRGROUPS(N)
95 INTEGER, INTENT ( IN ) :: LPOOL_B_L0_OMP
96 INTEGER, INTENT ( IN ) :: IPOOL_B_L0_OMP
97 & ( lpool_b_l0_omp )
98 INTEGER, INTENT ( IN ) :: L_PHYS_L0_OMP
99 INTEGER, INTENT ( IN ) :: PHYS_L0_OMP( L_PHYS_L0_OMP )
100 INTEGER, INTENT ( IN ) :: L_VIRT_L0_OMP
101 INTEGER, INTENT ( IN ) :: VIRT_L0_OMP( L_VIRT_L0_OMP )
102 INTEGER, INTENT ( IN ) :: VIRT_L0_OMP_MAPPING( L_VIRT_L0_OMP )
103 INTEGER, INTENT ( IN ) :: PERM_L0_OMP( L_PHYS_L0_OMP )
104 INTEGER, INTENT ( IN ) :: PTR_LEAFS_L0_OMP( L_PHYS_L0_OMP + 1)
105 INTEGER, INTENT ( IN ) :: LL0_OMP_MAPPING
106 INTEGER, INTENT ( OUT ) :: L0_OMP_MAPPING( LL0_OMP_MAPPING )
107 TYPE (MUMPS_TPS_T), DIMENSION(:) :: MUMPS_TPS_ARR
108 TYPE (SMUMPS_TPS_T), DIMENSION(:) :: SMUMPS_TPS_ARR
109 INTEGER, INTENT ( IN ) :: LL0_OMP_FACTORS
110 TYPE (SMUMPS_L0OMPFAC_T), INTENT(INOUT) ::
111 & l0_omp_factors(ll0_omp_factors)
112 INTEGER, INTENT (IN) :: NBSTATS_I4, NBSTATS_I8
113 INTEGER, INTENT (IN) :: NBCOLS_I4, NBCOLS_I8
114 INTEGER, INTENT (IN) :: I4_L0_OMP(NBSTATS_I4, NBCOLS_I4)
115 INTEGER(8), INTENT (IN) :: I8_L0_OMP(NBSTATS_I8, NBCOLS_I8)
116 EXTERNAL MUMPS_TYPENODE, MUMPS_PROCNODE
117 INTEGER MUMPS_TYPENODE, MUMPS_PROCNODE
118 LOGICAL MUMPS_INSSARBR,MUMPS_ROOTSSARBR
119 EXTERNAL mumps_inssarbr,mumps_rootssarbr
120 LOGICAL SMUMPS_POOL_EMPTY
121 EXTERNAL smumps_pool_empty, smumps_extract_pool
122 INTEGER numroc
123 EXTERNAL numroc
124 INTEGER :: MYTHREAD_ID, ITH
125 INTEGER :: THREAD_ID_P
126 DOUBLE PRECISION, PARAMETER :: DZERO = 0.0d0, done = 1.0d0
127 INTEGER INODE, LEAF
128 INTEGER TYPEF
129 INTEGER NBFIN
130 INTEGER TYPE
131 LOGICAL AVOID_DELAYED
132 INTEGER NBROOT_PROCESSED
133 INTEGER MAXFRW, NPVW, NMAXNPIVW, NOFFNEGW, NELVAW, COMP
134 INTEGER :: NB22T1W, NBTINYW, DET_EXPW, DET_SIGNW
135 REAL :: DET_MANTW
136 DOUBLE PRECISION FLOP_ESTIM_ACC
137#if defined(PANEL)
138 INTEGER(8) :: LAFAC
139 INTEGER LIWFAC, STRAT, TYPEFile, NextPiv2beWritten
140 INTEGER(8) :: ITMP8
141#endif
142 INTEGER :: LPOOL_P
143 INTEGER, DIMENSION(:), ALLOCATABLE :: IPOOL_P
144 INTEGER(8) :: TO_ALLOCATE
145 INTEGER, DIMENSION(:), ALLOCATABLE :: ID
146 INTEGER(8), DIMENSION(:), ALLOCATABLE :: VAL
147 INTEGER(8), ALLOCATABLE, DIMENSION(:) :: STATE, SIZE_COPIED
148 INTEGER :: NbWaitMem, NbFinished, NbOnGoingCopies, NbUnderL0
149 INTEGER(8) :: KEEP8_77_SAVE
150 DOUBLE PRECISION :: GTIME
151 INTEGER(8) :: MEMDISPO_UNDERL0, MEMDISPO_PERTHREAD
152 INTEGER :: BLR_STRAT
153 INTEGER :: VIRTUAL_TASK, PHYSICAL_TASK
154 INTEGER :: IFATH
155 INTEGER :: I, INFO_P(2), allocok
156 INTEGER(8) :: I8
157!$ INTEGER(8) :: CHUNK8
158!$ LOGICAL :: OMP_FLAG
159!$ INTEGER :: NOMP_SAVE
160 CALL mumps_load_disable()
161 gtime = mpi_wtime()
162 l0_omp_mapping = 0
163 nbroot_processed = 0
164 nstepsw = 0
165 opassw = dzero
166 opeliw = dzero
167 nelvaw = 0
168 comp = 0
169 maxfrw = 0
170 nmaxnpivw = 0
171 noffnegw = 0
172 flop_estim_acc = dzero
173 npvw = 0
174 nb22t1w = 0
175 nbtinyw = 0
176 det_expw = 0
177 det_mantw = cmplx(1.0e0,0.0e0, kind=kind(1.0e0))
178 det_signw = 1
179 DO ith = 1, keep(400)
180 NULLIFY(mumps_tps_arr(ith)%IW)
181 NULLIFY(mumps_tps_arr(ith)%ITLOC)
182 NULLIFY(smumps_tps_arr(ith)%A)
183 CALL smumps_set_maxs_maxis_thread(
184 & mumps_tps_arr(ith)%LA,
185 & mumps_tps_arr(ith)%LIW, blr_strat,
186 & keep,
187 & i4_l0_omp(1,ith), nbstats_i4,
188 & i8_l0_omp(1,ith), nbstats_i8)
189 ENDDO
190 IF (keep8(4) .NE. 0_8) THEN
191 CALL smumps_ma_eff_mem_dispo (
192 & mumps_tps_arr, keep(400),keep8, keep,
193 & n, blr_strat, lpool_b_l0_omp,
194 & i8_l0_omp, nbstats_i8,
195 & memdispo_underl0
196 & )
197 memdispo_underl0 = max(memdispo_underl0/2_8,0_8)
198 keep8(77) = keep8(77) + memdispo_underl0
199 memdispo_perthread = 0_8
200 IF (memdispo_underl0.GT.0) THEN
201 memdispo_perthread = memdispo_underl0/(int(keep(400),8))
202 ENDIF
203 DO ith = 1, keep(400)
204 mumps_tps_arr(ith)%LA = mumps_tps_arr(ith)%LA +
205 & memdispo_perthread
206 ENDDO
207 ENDIF
208 DO ith = 1, keep(400)
209 mumps_tps_arr(ith)%LRLU = mumps_tps_arr(ith)%LA
210 mumps_tps_arr(ith)%LRLUS = mumps_tps_arr(ith)%LA
211 mumps_tps_arr(ith)%LRLUSM = mumps_tps_arr(ith)%LA
212 mumps_tps_arr(ith)%IPTRLU = mumps_tps_arr(ith)%LA
213 mumps_tps_arr(ith)%POSFAC = 1_8
214 mumps_tps_arr(ith)%IWPOS = 1
215 mumps_tps_arr(ith)%IWPOSCB = mumps_tps_arr(ith)%LIW
216 ENDDO
217 IF (keep(406) .EQ. 2 ) THEN
218 ALLOCATE(state(keep(400)), size_copied(keep(400)), stat=allocok)
219 IF (allocok .GT. 0 ) THEN
220 WRITE(*,*) "Problem allocating STATE/SIZE_COPIED", keep(400)
221 CALL mumps_abort()
222 ENDIF
223 CALL smumps_perform_copies_init( state, keep8_77_save,
224 & nbwaitmem, nbfinished, nbongoingcopies, nbunderl0,
225 & keep, keep8 )
226 ENDIF
227!$OMP PARALLEL
228!$OMP& SHARED ( IPOOL_B_L0_OMP, LPOOL_B_L0_OMP )
229!$OMP& PRIVATE ( VIRTUAL_TASK, PHYSICAL_TASK,
230!$omp& ipool_p, lpool_p, leaf, inode, ifath, info_p, i, i8,
231!$OMP& TO_ALLOCATE, THREAD_ID_P,
232!$OMP& TYPE, TYPEF, AVOID_DELAYED, NOMP_SAVE, allocok )
233!$OMP& REDUCTION ( + : NPVW, OPASSW, OPELIW, NOFFNEGW, NELVAW, COMP,
234!$OMP& NB22T1W, NBTINYW, DET_EXPW,
235!$OMP& FLOP_ESTIM_ACC, NBROOT_PROCESSED, NSTEPSW )
236!$OMP& REDUCTION ( * : DET_MANTW, DET_SIGNW )
237!$OMP& REDUCTION ( max : MAXFRW, NMAXNPIVW )
238 thread_id_p = 1
239!$ THREAD_ID_P = OMP_GET_THREAD_NUM () + 1
240!$OMP BARRIER
241!$ NOMP_SAVE = omp_get_max_threads()
242!$ CALL omp_set_num_threads(1)
243 lpool_p = lpool_b_l0_omp
244 leaf = 1
245 info_p = 0
246 avoid_delayed = .false.
247 virtual_task = 0
248!$ IF ( omp_get_num_threads() .NE. KEEP(400) ) THEN
249!$ INFO_P(1)=-58
250!$ INFO_P(2)=-100-omp_get_num_threads()
251!$ GOTO 700
252!$ ENDIF
253 CALL smumps_dm_fac_alloc_allowed( mumps_tps_arr(thread_id_p)%LA,
254 & keep8,
255 & info_p(1), info_p(2) )
256 IF (info_p(1) .LT. 0) GOTO 700
257 ALLOCATE ( smumps_tps_arr(thread_id_p)%A(
258 & max(1_8,mumps_tps_arr(thread_id_p)%LA) ),
259 & stat=allocok)
260 IF (allocok.GT.0) THEN
261 info_p(1) = -13
262 CALL mumps_seti8toi4( mumps_tps_arr(thread_id_p)%LA,
263 & info_p(2))
264 GOTO 700
265 ELSE
266 CALL mumps_dm_fac_upd_dyn_memcnts(
267 & mumps_tps_arr(thread_id_p)%LA,
268 & keep(405).EQ.1,
269 & keep8, info_p(1), info_p(2),
270 & .false., .false. )
271 IF (info_p(1) .LT. 0) THEN
272 GOTO 700
273 ENDIF
274 ENDIF
275 to_allocate =
276 & ((int(mumps_tps_arr(thread_id_p)%LIW,8) * int(keep(34),8 )) /
277 & int(keep(35),8 ))+
278 & ((int(lpool_p,8) * int(keep(34),8 )) / int(keep(35),8 ))+
279 & ((int(n+keep(253),8) * int(keep(34),8 )) / int(keep(35),8 ))
280 CALL smumps_dm_fac_alloc_allowed( to_allocate,
281 & keep8, info_p(1), info_p(2) )
282 IF ( info_p(1) .LT. 0 ) GOTO 700
283 ALLOCATE ( mumps_tps_arr(thread_id_p)%IW(
284 & mumps_tps_arr(thread_id_p)%LIW ),
285 & ipool_p( lpool_p ),
286 & mumps_tps_arr(thread_id_p)%ITLOC ( n + keep(253) ),
287 & stat = allocok )
288 IF ( allocok .GT. 0 ) THEN
289 info_p(1) = -13
290 info_p(2) = mumps_tps_arr(thread_id_p)%LIW +
291 & lpool_p + n+keep(253)
292 GOTO 700
293 ELSE
294 CALL mumps_dm_fac_upd_dyn_memcnts( to_allocate,
295 & keep(405).EQ.1, keep8,
296 & info_p(1), info_p(2), .true., .false. )
297 IF (info_p(1) .LT. 0) THEN
298 GOTO 700
299 ENDIF
300 ENDIF
301 CALL smumps_alloc_cb( .false., 0_8,
302 & .false., .false., myid_nodes, n, keep, keep8, dkeep,
303 & mumps_tps_arr(thread_id_p)%IW(1),
304 & mumps_tps_arr(thread_id_p)%LIW,
305 & smumps_tps_arr(thread_id_p)%A(1),
306 & mumps_tps_arr(thread_id_p)%LA,
307 & mumps_tps_arr(thread_id_p)%LRLU,
308 & mumps_tps_arr(thread_id_p)%IPTRLU,
309 & mumps_tps_arr(thread_id_p)%IWPOS,
310 & mumps_tps_arr(thread_id_p)%IWPOSCB,
311 & slavef, procnode_steps, dad,
312 & ptrist, ptrast, step, pimaster,
313 & pamaster, keep(ixsz), 0_8, -444, -444, .true.,
314 & comp, mumps_tps_arr(thread_id_p)%LRLUS,
315 & mumps_tps_arr(thread_id_p)%LRLUSM,
316 & info_p(1), info_p(2)
317 & )
318 CALL smumps_init_pool_last3( ipool_p(1), lpool_p,
319 & leaf )
320 mumps_tps_arr(thread_id_p)%ITLOC = 0
321600 CONTINUE
322 virtual_task = virtual_task + 1
323 IF ( virtual_task .LT. l_virt_l0_omp ) THEN
324 IF ( virt_l0_omp_mapping( virtual_task ) .EQ. thread_id_p ) THEN
325 DO physical_task =
326 & virt_l0_omp( virtual_task ),
327 & virt_l0_omp( virtual_task + 1 ) - 1
328 DO i = ptr_leafs_l0_omp( perm_l0_omp( physical_task )+1 ) + 1,
329 & ptr_leafs_l0_omp( perm_l0_omp( physical_task ) )
330 IF (ipool_b_l0_omp(i) .GT. 0) THEN
331 CALL smumps_insert_pool_n( n, ipool_p(1),
332 & lpool_p,
333 & procnode_steps, slavef, keep(199), keep(28), 3, 0, 1, step,
334 & ipool_b_l0_omp(i) )
335 END IF
336 END DO
337 DO WHILE (
338 & .NOT. smumps_pool_empty( ipool_p(1), lpool_p )
339 & .AND. info_p(1) .GE. 0 )
340 CALL smumps_extract_pool( n, ipool_p(1), lpool_p,
341 & procnode_steps, slavef, step, inode, keep, keep8, myid_nodes,
342 & nd, .false. )
343 10 CONTINUE
344 l0_omp_mapping( step( inode ) ) = thread_id_p
345 ifath = dad( step( inode ) )
346 TYPE = 1
347 IF ( ifath .NE. 0 ) THEN
348 typef = 1
349 ELSE
350 typef = -9999
351 ENDIF
352 CALL smumps_process_front_niv1( comm_load, ass_irecv, n,
353 & inode, TYPE, typef, mumps_tps_arr(thread_id_p)%LA, MUMPS_TPS_ARR
354 & (thread_id_p)%IW(1), mumps_tps_arr(thread_id_p)%LIW,
355 & smumps_tps_arr(thread_id_p)%A(1), maxfrw, noffnegw, npvw,
356 & nb22t1w, nbtinyw, det_expw, det_mantw, det_signw, info_p, uu,
357 & seuil, seuil_ldlt_niv2, opeliw, nelvaw, nmaxnpivw, nstepsw,
358 & ptrist, ptlust_s, ptrfac, ptrast, step, pimaster, pamaster,
359 & ne, mumps_tps_arr(thread_id_p)%POSFAC,
360 & mumps_tps_arr(thread_id_p)%LRLU,
361 & mumps_tps_arr(thread_id_p)%LRLUS, mumps_tps_arr(thread_id_p)%
362 % LRLUSM, mumps_tps_arr(thread_id_p)%IPTRLU, icntl, keep, keep8,
363 & dkeep, pivnul_list, lpn_list, comp, mumps_tps_arr(thread_id_p)%
364 & iwpos, mumps_tps_arr(thread_id_p)%IWPOSCB, procnode_steps,
365 & slavef, ifath, comm_nodes, myid_nodes, ipool_p, lpool_p, leaf,
366 & perm, nstk_steps, bufr, lbufr, lbufr_bytes,
367 & nbfin, root, opassw, mumps_tps_arr(thread_id_p)%ITLOC(1),
368 & rhs_mumps, fils, ptrarw, ptraiw, intarr, dblarr, nd, frere, dad,
369 & lptrar, nelt, frtptr, frtelt, istep_to_iniv2, tab_pos_in_pere,
370 & lrgroups, flop_estim_acc )
371 IF (info_p(1) .LT. 0) THEN
372 GOTO 700
373 ENDIF
374 IF ( ifath .NE. 0 ) THEN
375 IF ( phys_l0_omp( perm_l0_omp( physical_task ) )
376 & .NE. inode ) THEN
377 nstk_steps( step( ifath ) ) =
378 & nstk_steps( step ( ifath ) ) - 1
379 IF ( nstk_steps( step( ifath ) ) .EQ. 0 ) THEN
380 inode = ifath
381 GOTO 10
382 ENDIF
383 ELSE
384!$OMP ATOMIC UPDATE
385 nstk_steps( step( ifath ) ) =
386 & nstk_steps( step( ifath ) ) - 1
387!$OMP END ATOMIC
388 END IF
389 ELSE
390 nbroot_processed = nbroot_processed + 1
391 END IF
392 END DO
393 END DO
394 ENDIF
395 GOTO 600
396 ENDIF
397 700 CONTINUE
398 IF (associated(mumps_tps_arr(thread_id_p)%ITLOC)) THEN
399 DEALLOCATE(mumps_tps_arr(thread_id_p)%ITLOC)
400 NULLIFY(mumps_tps_arr(thread_id_p)%ITLOC)
401 CALL mumps_dm_fac_upd_dyn_memcnts(
402 & -(int(n+keep(253),8) * int(keep(34),8 )) / int(keep(35),8),
403 & keep(405).EQ.1, keep8,
404 & info_p(1), info_p(2), .true., .false. )
405 ENDIF
406 IF (allocated(ipool_p)) THEN
407 DEALLOCATE(ipool_p);
408 CALL mumps_dm_fac_upd_dyn_memcnts(
409 & -(int(lpool_p,8) * int(keep(34),8 )) / int(keep(35),8),
410 & keep(405).EQ.1, keep8,
411 & info_p(1), info_p(2), .true., .false. )
412 ENDIF
413 IF ( keep(406) .EQ. 2) THEN
414 CALL smumps_perform_copies( thread_id_p,
415 & mumps_tps_arr, smumps_tps_arr,
416 & l0_omp_factors, ll0_omp_factors,
417 & state, size_copied,
418 & nbwaitmem, nbfinished, nbongoingcopies, nbunderl0,
419 & myid_nodes, n, slavef,
420 & step, ptrast, pamaster, procnode_steps, dad,
421 & keep, keep8, info_p
422 & )
423 ELSE
424 IF ((keep(407) .EQ. 1) .OR. (keep(406) .EQ.1) ) THEN
425 IF (info_p(1) .GE. 0) THEN
426 CALL smumps_dm_cbstatic2dynamic_i
427 & (2,
428 & 0_8,
429 & .false.,
430 & myid_nodes, n, slavef, keep, keep8,
431 & mumps_tps_arr(thread_id_p)%IW(1),
432 & mumps_tps_arr(thread_id_p)%LIW,
433 & mumps_tps_arr(thread_id_p)%IWPOSCB,
434 & mumps_tps_arr(thread_id_p)%IWPOS,
435 & smumps_tps_arr(thread_id_p)%A(1),
436 & mumps_tps_arr(thread_id_p)%LA,
437 & mumps_tps_arr(thread_id_p)%LRLU,
438 & mumps_tps_arr(thread_id_p)%IPTRLU,
439 & mumps_tps_arr(thread_id_p)%LRLUS,
440 & step, ptrast, pamaster, procnode_steps, dad,
441 & info_p(1), info_p(2) )
442 ENDIF
443 ENDIF
444 IF (keep(406) .EQ.1) THEN
445 IF (info_p(1) .GE.0 )THEN
446 to_allocate = max(mumps_tps_arr(thread_id_p)%POSFAC-1,1_8)
447 CALL smumps_dm_fac_alloc_allowed( to_allocate,
448 & keep8, info_p(1), info_p(2) )
449 ENDIF
450 IF (info_p(1) .GE.0 )THEN
451 ALLOCATE(l0_omp_factors(thread_id_p)%A(to_allocate),
452 & stat=allocok)
453 IF (allocok .GT. 0) THEN
454 info_p(1) = -13
455 CALL mumps_seti8toi4(to_allocate, info_p(2))
456 l0_omp_factors(thread_id_p)%LA = 0_8
457 ELSE
458 l0_omp_factors(thread_id_p)%LA =
459 & mumps_tps_arr(thread_id_p)%POSFAC-1_8
460 CALL mumps_dm_fac_upd_dyn_memcnts(
461 & l0_omp_factors(thread_id_p)%LA, keep(405).EQ.1, keep8,
462 & info_p(1), info_p(2), .true., .false. )
463 ENDIF
464 ENDIF
465 IF (info_p(1) .GE.0 ) THEN
466 DO i8 = 1_8, mumps_tps_arr(thread_id_p)%POSFAC-1_8
467 l0_omp_factors(thread_id_p)%A(i8) =
468 & smumps_tps_arr(thread_id_p)%A(i8)
469 ENDDO
470 ENDIF
471 IF ( associated(smumps_tps_arr(thread_id_p)%A)) THEN
472 DEALLOCATE(smumps_tps_arr(thread_id_p)%A)
473 NULLIFY(smumps_tps_arr(thread_id_p)%A)
474 CALL mumps_dm_fac_upd_dyn_memcnts(
475 & -mumps_tps_arr(thread_id_p)%LA,
476 & keep(405).EQ.1,
477 & keep8,
478 & info_p(1), info_p(2),
479 & .false., .false. )
480 IF (info_p(1) .GE. 0) THEN
481!$OMP ATOMIC UPDATE
482 keep8(69) = keep8(69) - l0_omp_factors(thread_id_p)%LA
483!$OMP END ATOMIC
484 ENDIF
485 ENDIF
486 ENDIF
487 ENDIF
488 IF (info_p(1) .LT.0) THEN
489!$OMP CRITICAL(critical_info)
490 info(1) = info_p(1)
491 info(2) = info_p(2)
492!$OMP END CRITICAL(critical_info)
493 ELSE IF (info_p(1) .GE. 0) THEN
494!$OMP CRITICAL(critical_info)
495 IF (info(1) .EQ. 0) THEN
496 info(1) = info_p(1)
497 info(2) = info_p(2)
498 ENDIF
499!$OMP END CRITICAL(critical_info)
500 ENDIF
501#if defined(WORKAROUNDINTELILP64OPENMPLIMITATION)
502!$ CALL omp_set_num_threads(int(nomp_save,4))
503#else
504!$ CALL omp_set_num_threads(NOMP_SAVE)
505#endif
506!$OMP END PARALLEL
507 IF (info(1) .LT. 0) THEN
508 IF (icntl(1) .GT. 0 .AND. icntl(4) .GE.1 ) THEN
509 WRITE(icntl(1),'(A,I6,I16,A,I5,A)')
510 & "** ERROR DURING L0_OMP: INFO(1:2)=",
511 & info(1), info(2), " (MPI worker ", myid_nodes,")"
512 ENDIF
513 ENDIF
514 IF ( keep(406) .EQ. 0 ) THEN
515 ALLOCATE(id(keep(400)), val(keep(400)),
516 & stat = allocok)
517 IF ( allocok .GT. 0 ) THEN
518 info(1) = -13
519 info(2) = keep(400)
520 GOTO 800
521 ENDIF
522 DO mythread_id = 1, keep(400)
523 val(mythread_id) = mumps_tps_arr( mythread_id )%POSFAC-1_8
524 id(mythread_id) = mythread_id
525 ENDDO
526 CALL mumps_sort_int8(keep(400), val, id)
527 DO ith=1, keep(400)
528 mythread_id = id(ith)
529 IF ((keep(407).NE.1) .AND. (keep(406).EQ.0)) THEN
530 IF (info(1) .GE. 0) THEN
531 CALL smumps_dm_cbstatic2dynamic_i
532 & (2,
533 & 0_8,
534 & .false.,
535 & myid_nodes, n, slavef, keep, keep8,
536 & mumps_tps_arr(mythread_id)%IW(1),
537 & mumps_tps_arr(mythread_id)%LIW,
538 & mumps_tps_arr(mythread_id)%IWPOSCB,
539 & mumps_tps_arr(mythread_id)%IWPOS,
540 & smumps_tps_arr(mythread_id)%A(1),
541 & mumps_tps_arr(mythread_id)%LA,
542 & mumps_tps_arr(mythread_id)%LRLU,
543 & mumps_tps_arr(mythread_id)%IPTRLU,
544 & mumps_tps_arr(mythread_id)%LRLUS,
545 & step, ptrast, pamaster, procnode_steps, dad,
546 & info(1), info(2) )
547 ENDIF
548 ENDIF
549 IF (keep(406).EQ.0) THEN
550 IF (info(1) .GE. 0 )THEN
551 to_allocate = max(mumps_tps_arr(mythread_id)%POSFAC-1,1_8)
552 CALL smumps_dm_fac_alloc_allowed( to_allocate,
553 & keep8, info(1), info(2) )
554 ENDIF
555 IF (info(1) .GE.0 ) THEN
556 ALLOCATE(l0_omp_factors(mythread_id)%A(to_allocate),
557 & stat=allocok)
558 IF (allocok .GT. 0) THEN
559 info(1) = -13
560 CALL mumps_seti8toi4(to_allocate, info(2))
561 l0_omp_factors(mythread_id)%LA = 0_8
562 ELSE
563 l0_omp_factors(mythread_id)%LA =
564 & mumps_tps_arr(mythread_id)%POSFAC-1_8
565 CALL mumps_dm_fac_upd_dyn_memcnts(
566 & l0_omp_factors(mythread_id)%LA,
567 & keep(405).EQ.1, keep8,
568 & info(1), info(2), .true., .false. )
569 ENDIF
570 ENDIF
571 IF (info(1) .GE. 0) THEN
572!$ CHUNK8 = max( int(KEEP(361),8),
573!$ & (MUMPS_TPS_ARR(MYTHREAD_ID)%POSFAC+KEEP(400)-2_8) /
574!$ & keep(400) )
575!$ OMP_FLAG = ( (MUMPS_TPS_ARR(MYTHREAD_ID)%POSFAC-1_8 >
576!$ & int(KEEP(361),8))
577!$ & .AND. (KEEP(400).GT.1)
578!$ & )
579!$OMP PARALLEL DO PRIVATE(I8) SCHEDULE(STATIC, CHUNK8)
580!$OMP& IF (OMP_FLAG)
581 DO i8 = 1_8, mumps_tps_arr(mythread_id)%POSFAC-1_8
582 l0_omp_factors(mythread_id)%A(i8) =
583 & smumps_tps_arr(mythread_id)%A(i8)
584 ENDDO
585!$OMP END PARALLEL DO
586 ENDIF
587 IF ( associated(smumps_tps_arr(mythread_id)%A)) THEN
588 DEALLOCATE(smumps_tps_arr(mythread_id)%A)
589 NULLIFY(smumps_tps_arr(mythread_id)%A)
590 CALL mumps_dm_fac_upd_dyn_memcnts(
591 & -mumps_tps_arr(mythread_id)%LA,
592 & keep(405).EQ.1, keep8,
593 & info(1), info(2),
594 & .false., .false. )
595 IF (info(1).GE.0) THEN
596 keep8(69) = keep8(69) - l0_omp_factors(mythread_id)%LA
597 ENDIF
598 ENDIF
599 ENDIF
600 ENDDO
601 IF (ALLOCATED(id)) DEALLOCATE(id)
602 IF (ALLOCATED(val)) DEALLOCATE(val)
603 ENDIF
604 800 CONTINUE
605 keep8(64) = 0_8
606 DO i = 1, keep(400)
607 keep8(64) = keep8(64) + mumps_tps_arr(i)%POSFAC - 1_8
608 ENDDO
609 keep8(62) = 0_8
610 DO i = 1, keep(400)
611 keep8(62) = keep8(62) + mumps_tps_arr(i)%LRLUSM
612 ENDDO
613 nbroot_under_l0 = nbroot_processed
614 dkeep(95) = real(mpi_wtime() - gtime)
615 time_update = time_update/dble(keep(400))
616 time_update_lrlr1 = time_update_lrlr1/dble(keep(400))
617 time_update_lrlr2 = time_update_lrlr2/dble(keep(400))
618 time_update_lrlr3 = time_update_lrlr3/dble(keep(400))
619 time_update_frlr = time_update_frlr/dble(keep(400))
620 time_update_frfr = time_update_frfr/dble(keep(400))
621 time_midblk_compress = time_midblk_compress/dble(keep(400))
622 time_compress = time_compress/dble(keep(400))
623 time_frswap_compress = time_frswap_compress/dble(keep(400))
624 time_cb_compress = time_cb_compress/dble(keep(400))
625 time_panel = time_panel/dble(keep(400))
626 time_fac_i = time_fac_i/dble(keep(400))
627 time_fac_mq = time_fac_mq/dble(keep(400))
628 time_fac_sq = time_fac_sq/dble(keep(400))
629 time_frfronts = time_frfronts/dble(keep(400))
630 time_lrtrsm = time_lrtrsm/dble(keep(400))
631 time_frtrsm = time_frtrsm/dble(keep(400))
632 time_lr_module = time_lr_module/dble(keep(400))
633 time_decomp = time_decomp/dble(keep(400))
634 time_diagcopy = time_diagcopy/dble(keep(400))
635 time_decomp_ucfs = time_decomp_ucfs/dble(keep(400))
636 time_decomp_asm1 = time_decomp_asm1/dble(keep(400))
637 time_decomp_locasm2 = time_decomp_locasm2/dble(keep(400))
638 time_decomp_maplig1 = time_decomp_maplig1/dble(keep(400))
639 time_decomp_asms2s = time_decomp_asms2s/dble(keep(400))
640 time_decomp_asms2m = time_decomp_asms2m/dble(keep(400))
641 CALL mumps_load_enable()
642 CALL smumps_load_update(0,.false., flop_estim_acc,keep,keep8)
643 RETURN
644 END SUBROUTINE smumps_fac_l0_omp
645 SUBROUTINE smumps_set_maxs_maxis_thread(MAXS_BASE_RELAXED8TH,
646 & MAXIS_BASE_RELAXEDTH, BLR_STRAT,
647 & KEEP,
648 & I4_L0_OMPTH, NBSTATS_I4,
649 & I8_L0_OMPTH, NBSTATS_I8)
650 IMPLICIT NONE
651 INTEGER, INTENT(IN) :: KEEP(500), NBSTATS_I4, NBSTATS_I8
652 INTEGER, INTENT(IN) :: I4_L0_OMPTH(NBSTATS_I4)
653 INTEGER(8), INTENT(IN) :: I8_L0_OMPTH(NBSTATS_I8)
654 INTEGER(8), INTENT(OUT) :: MAXS_BASE_RELAXED8TH
655 INTEGER, INTENT(OUT) :: MAXIS_BASE_RELAXEDTH
656 INTEGER, INTENT(OUT) :: BLR_STRAT
657 INTEGER :: PERLU
658 INTEGER(8) :: MAXS_BASE8TH
659 INTEGER(8) :: MAXIS_BASE_RELAXEDTH8
660 perlu = keep(12)
661 CALL smumps_set_blrstrat_and_maxs ( maxs_base8th,
662 & maxs_base_relaxed8th, blr_strat, keep(1),
663 & i8_l0_ompth(2), i8_l0_ompth(3), i8_l0_ompth(5),
664 & i8_l0_ompth(6), i8_l0_ompth(7), i8_l0_ompth(8) )
665 IF ( keep(201) .EQ. 0 ) THEN
666 maxis_base_relaxedth8 = int(i4_l0_ompth(2),8)
667 ELSE
668 maxis_base_relaxedth8 = int(i4_l0_ompth(4),8)
669 ENDIF
670 maxis_base_relaxedth8 = max( 1_8,
671 & maxis_base_relaxedth8 + 3 * max(perlu,10) *
672 & ( maxis_base_relaxedth8 / 100 + 1 )
673 & )
674 maxis_base_relaxedth8 = min(maxis_base_relaxedth8,
675 & int( huge( maxis_base_relaxedth ) ,8)
676 & )
677 maxis_base_relaxedth = int( maxis_base_relaxedth8 )
678 RETURN
679 END SUBROUTINE smumps_set_maxs_maxis_thread
680 SUBROUTINE smumps_ma_eff_mem_dispo(
681 & MUMPS_TPS_ARR, NBTHREADS, KEEP8, KEEP,
682 & N, BLR_STRAT, LPOOL_P,
683 & I8_L0_OMP, NBSTATS_I8,
684 & MEMDISPO_UNDERL0)
685 USE mumps_tps_m, ONLY : mumps_tps_t
686 IMPLICIT NONE
687 INTEGER, INTENT(IN) :: N, BLR_STRAT, KEEP(500)
688 INTEGER, INTENT(IN) :: NBSTATS_I8, NBTHREADS, LPOOL_P
689 INTEGER(8), INTENT(IN) :: KEEP8(150)
690 INTEGER(8), INTENT(IN) :: I8_L0_OMP(NBSTATS_I8,NBTHREADS)
691 INTEGER(8), INTENT(OUT) :: MEMDISPO_UNDERL0
692 TYPE (MUMPS_TPS_T), INTENT(IN) :: MUMPS_TPS_ARR(:)
693 INTEGER :: PERLU, ITH, ITHMIN, ITHMIN_if_LRLU, OOC_STRAT
694 INTEGER(8) :: TO_ALLOCATE, BLR_RELATED, COPY_RELATED
695 INTEGER(8) :: MIN_NRLADU_underL0, MIN_NRLADU_if_LR_LU_underL0
696 perlu = keep(12)
697 ooc_strat = keep(201)
698 to_allocate = 0_8
699 DO ith = 1, nbthreads
700 to_allocate = to_allocate +
701 & ((int(mumps_tps_arr(ith)%LIW,8) * int(keep(34),8 )) /
702 & int(keep(35),8 ))
703 & + mumps_tps_arr(ith)%LA
704 ENDDO
705 to_allocate = to_allocate + int(nbthreads,8)* (
706 & ((int(lpool_p,8) * int(keep(34),8 )) / int(keep(35),8 )) +
707 & ((int(n+keep(253),8) * int(keep(34),8 )) / int(keep(35),8 ))
708 & )
709 blr_related = 0_8
710 DO ith = 1, nbthreads
711 IF (blr_strat.EQ.1) THEN
712 blr_related = blr_related +
713 & i8_l0_omp(10,ith) +
714 & int(perlu,8) * ( i8_l0_omp(10,ith) / 100_8 + 1_8)
715 ELSE IF (blr_strat.EQ.2) THEN
716 blr_related = blr_related +
717 & i8_l0_omp(13,ith) +
718 & int(perlu,8) * ( i8_l0_omp(13,ith) / 100_8 + 1_8)
719 ELSE IF (blr_strat.EQ.3) THEN
720 blr_related = blr_related +
721 & i8_l0_omp(8,ith) +
722 & int(perlu,8) * ( i8_l0_omp(8,ith) / 100_8 + 1_8)
723 ENDIF
724 ENDDO
725 copy_related = 0_8
726 ithmin = 1
727 ithmin_if_lrlu = 1
728 min_nrladu_if_lr_lu_underl0 = i8_l0_omp(4,1)
729 min_nrladu_underl0 = i8_l0_omp(1,1)
730 DO ith = 1, nbthreads
731 IF (i8_l0_omp(1,ith).LT.min_nrladu_underl0)
732 & THEN
733 min_nrladu_underl0 = i8_l0_omp(1,ith)
734 ithmin = ith
735 ENDIF
736 IF (i8_l0_omp(4,ith).LT.min_nrladu_if_lr_lu_underl0)
737 & THEN
738 min_nrladu_if_lr_lu_underl0 = i8_l0_omp(4,ith)
739 ithmin_if_lrlu = ith
740 ENDIF
741 ENDDO
742 IF (blr_strat.EQ.0) THEN
743 IF (ooc_strat .LE. 0 .AND. ooc_strat .NE. -1 ) THEN
744 copy_related = copy_related +
745 & i8_l0_omp(1,ithmin) +
746 & i8_l0_omp(23, ithmin)
747 ELSE
748 copy_related = copy_related +
749 & i8_l0_omp(23, ithmin)
750 ENDIF
751 ELSE
752 IF (ooc_strat .LE. 0 .AND. ooc_strat .NE. -1 ) THEN
753 copy_related = copy_related +
754 & i8_l0_omp(4,ithmin_if_lrlu) +
755 & i8_l0_omp(23,ithmin_if_lrlu )
756 ELSE
757 copy_related = copy_related +
758 & i8_l0_omp(23, ithmin_if_lrlu)
759 ENDIF
760 ENDIF
761 copy_related = copy_related +
762 & int(perlu,8)*(copy_related / 100_8 + 1_8 )
763 to_allocate = to_allocate + copy_related + blr_related
764 memdispo_underl0 = keep8(75) - to_allocate
765 RETURN
766 END SUBROUTINE smumps_ma_eff_mem_dispo
767 SUBROUTINE smumps_l0omp_copy_iw( IW, LIW, IWPOS,
768 & MUMPS_TPS_ARR, KEEP,
769 & PTLUST, ICNTL, INFO )
770 USE mumps_tps_m, ONLY : mumps_tps_t
771 IMPLICIT NONE
772 INTEGER :: KEEP(500)
773 INTEGER, INTENT( IN ) :: LIW
774 INTEGER, INTENT( INOUT ) :: IW(:)
775 INTEGER, INTENT( INOUT ) :: IWPOS
776 INTEGER, INTENT( INOUT ) :: PTLUST(KEEP(28))
777 INTEGER, INTENT( IN ) :: ICNTL(60)
778 INTEGER, INTENT( INOUT ) :: INFO(80)
779 TYPE (MUMPS_TPS_T), TARGET :: MUMPS_TPS_ARR(:)
780 INTEGER :: ITHREAD, JTHREAD
781 INTEGER :: REQUESTED_SIZE
782 INTEGER :: IWPOS_TO_COPY
783 INTEGER :: LOC_IPOS
784 INTEGER :: LOC_SIZE, LOC_ISTEP
785 TYPE (MUMPS_TPS_T), POINTER :: MUMPS_TPS
786 include 'mumps_headers.h'
787 requested_size = 0
788 DO ithread = 1, size(mumps_tps_arr)
789 mumps_tps => mumps_tps_arr(ithread)
790 requested_size = requested_size + mumps_tps%IWPOS - 1
791 ENDDO
792 IF ( liw - iwpos + 1 .LT. requested_size ) THEN
793 WRITE(*,*) " LIW too small in SMUMPS_L0OMP_COPY_IW !!", liw,
794 & requested_size
795 info(1) = -8
796 info(2) = requested_size-liw+iwpos-1
797 IF ( icntl(1) .GT. 0 .AND. icntl(4).GE.1) THEN
798 WRITE(icntl(1),*) " ** ERROR IN SMUMPS_L0OMP_COPY_IW: ",
799 & "LIW TOO SMALL TO COPY LOCAL FACTOR INFORMATION",
800 & info(2)
801 ENDIF
802 GOTO 500
803 ENDIF
804 DO ithread = 1, size(mumps_tps_arr)
805 mumps_tps => mumps_tps_arr(ithread)
806 iwpos_to_copy = iwpos
807 DO jthread=1, ithread - 1
808 iwpos_to_copy = iwpos_to_copy+mumps_tps_arr(jthread)%IWPOS-1
809 ENDDO
810 iw(iwpos_to_copy: iwpos_to_copy+mumps_tps%IWPOS - 2) =
811 & mumps_tps%IW(1:mumps_tps%IWPOS-1)
812 loc_ipos = 1
813 DO WHILE ( loc_ipos .NE. mumps_tps%IWPOS )
814 loc_size = mumps_tps%IW(loc_ipos+xxi)
815 loc_istep = mumps_tps%IW(loc_ipos+keep(ixsz)+4)
816 ptlust(loc_istep) = iwpos_to_copy+loc_ipos-1
817 loc_ipos = loc_ipos + loc_size
818 ENDDO
819 ENDDO
820 iwpos = iwpos + requested_size
821 500 CONTINUE
822 RETURN
823 END SUBROUTINE smumps_l0omp_copy_iw
824 SUBROUTINE smumps_perform_copies_init( STATE, KEEP8_77_SAVE,
825 & NbWaitMem, NbFinished, NbOnGoingCopies, NbUnderL0,
826 & KEEP, KEEP8 )
827 INTEGER, INTENT(IN) :: KEEP(500)
828 INTEGER(8), INTENT(IN) :: KEEP8(150)
829 INTEGER, INTENT(OUT) :: NbWaitMem,
830 & NbFinished,
831 & NbOnGoingCopies,
832 & nbunderl0
833 INTEGER(8), INTENT(OUT) :: STATE(KEEP(400)), KEEP8_77_SAVE
834 INTEGER :: ITH
835 nbwaitmem = 0
836 nbfinished = 0
837 nbongoingcopies = 0
838 nbunderl0 = keep(400)
839 DO ith=1, keep(400)
840 state(ith) = underl0
841 ENDDO
842 keep8_77_save = keep8(77)
843 RETURN
844 END SUBROUTINE smumps_perform_copies_init
845 SUBROUTINE smumps_perform_copies( THREAD_ID_P,
846 & MUMPS_TPS_ARR, SMUMPS_TPS_ARR,
847 & L0_OMP_FACTORS, LL0_OMP_FACTORS,
848 & STATE, SIZE_COPIED,
849 & NbWaitMem, NbFinished, NbOngoingCopies, NbUnderL0,
850 & MYID_NODES, N, SLAVEF,
851 & STEP, PTRAST, PAMASTER, PROCNODE_STEPS, DAD,
852 & KEEP, KEEP8, INFO_P
853 & )
854 USE mumps_tps_m, ONLY : mumps_tps_t
855 USE smumps_tps_m, ONLY : smumps_tps_t
856 USE smumps_struc_def, ONLY : smumps_l0ompfac_t
857 INTEGER :: KEEP(500)
858 INTEGER(8) :: KEEP8(150)
859 INTEGER, INTENT(IN) :: THREAD_ID_P
860 INTEGER, INTENT(INOUT) :: INFO_P(2)
861 INTEGER, INTENT(IN) :: MYID_NODES, N, SLAVEF
862 INTEGER, INTENT(IN) :: STEP(N), DAD(KEEP(28))
863 INTEGER(8), INTENT(IN) :: PAMASTER(KEEP(28)), PTRAST(KEEP(28))
864 INTEGER, INTENT(IN) :: PROCNODE_STEPS(KEEP(28))
865 INTEGER, INTENT(INOUT) :: NbWaitMem,
866 & nbfinished,
867 & nbongoingcopies,
868 & nbunderl0
869 INTEGER(8), INTENT(INOUT) :: STATE( KEEP(400) )
870 INTEGER(8), INTENT(INOUT) :: SIZE_COPIED(KEEP(400) )
871 TYPE (MUMPS_TPS_T), DIMENSION(:) :: MUMPS_TPS_ARR
872 TYPE (SMUMPS_TPS_T), DIMENSION(:) :: SMUMPS_TPS_ARR
873 INTEGER, INTENT ( IN ) :: LL0_OMP_FACTORS
874 TYPE (SMUMPS_L0OMPFAC_T), INTENT(INOUT) ::
875 & l0_omp_factors(ll0_omp_factors)
876 INTEGER :: LOCAL_ACTION
877 INTEGER, PARAMETER :: NOTHING = 0
878 INTEGER, PARAMETER :: FREE_WORK_MYID = 1
879 INTEGER, PARAMETER :: COPY_FACTORS = 2
880 INTEGER, PARAMETER :: AllocateViderCB = 3
881 INTEGER, PARAMETER :: DORMIR = 4
882 INTEGER(8) :: COPY_START, CHUNK8, I8, TO_ALLOCATE
883 INTEGER :: ITH, K
884 INTEGER :: allocok
885 INTEGER(8) :: PeakAuthorized_P
886 INTEGER(8) :: MemNeeded_P, MemNeededForCB_P, MemDispo_P,
887 & CBCopiedToDynamic_P, LRLUS_SAVE_P
888 INTEGER(8) :: KEEP8_71, KEEP8_73
889!$OMP CRITICAL(L0_COPIES)
890 STATE(THREAD_ID_P) = copynotstarted
891 IF ( info_p(1) .LT. 0 ) THEN
892 nbfinished = nbfinished + 1
893 state(thread_id_p) = finished
894 ENDIF
895 DO ith = 1, keep(400)
896 IF ( state(ith) .EQ. waitmem ) THEN
897 state(ith)=copynotstarted
898 ENDIF
899 ENDDO
900 nbwaitmem = 0
901!$OMP END CRITICAL(L0_COPIES)
902 DO WHILE ( nbfinished .NE. keep(400) )
903 local_action = dormir
904!$OMP CRITICAL(L0_COPIES)
905 IF ( nbfinished.EQ. keep(400)) THEN
906 local_action = nothing
907 ELSE IF ( (nbfinished+nbwaitmem) .EQ. keep(400) ) THEN
908!$OMP ATOMIC READ
909 keep8_73 = keep8(73)
910!$OMP END ATOMIC
911!$OMP ATOMIC READ
912 keep8_71 = keep8(71)
913!$OMP END ATOMIC
914 memdispo_p = keep8(77) - (keep8_73 -keep8_71)
915 memdispo_p = min(memdispo_p, keep8(75)-keep8_73)
916 memneeded_p = huge(memneeded_p)
917 DO ith = 1, keep(400)
918 IF (state(ith).EQ.waitmem) THEN
919 memneeded_p = min( memneeded_p,
920 & mumps_tps_arr(ith)%LA -
921 & mumps_tps_arr(ith)%LRLUS )
922 ENDIF
923 ENDDO
924 IF ((keep8(75)-keep8_73).LT.memneeded_p) THEN
925 info_p(1) = -19
926 CALL mumps_set_ierror (
927 & memneeded_p-(keep8(75)-keep8_73), info_p(2))
928 DO ith = 1, keep(400)
929 state(ith) = finished
930 ENDDO
931 nbfinished = keep(400)
932 ELSE
933 keep8(77) = memneeded_p + (keep8_73 -keep8_71)
934 DO ith = 1, keep(400)
935 IF ( state(ith) .EQ. waitmem ) THEN
936 state(ith)=copynotstarted
937 ENDIF
938 ENDDO
939 nbwaitmem = 0
940 ENDIF
941 local_action = nothing
942 ELSE
943 SELECT CASE (state(thread_id_p))
944 CASE ( copyfactorsfinished )
945 local_action = free_work_myid
946 CASE ( copynotstarted )
947!$OMP ATOMIC READ
948 keep8_73 = keep8(73)
949!$OMP END ATOMIC
950!$OMP ATOMIC READ
951 keep8_71 = keep8(71)
952!$OMP END ATOMIC
953 peakauthorized_p = keep8(77)
954 memdispo_p = peakauthorized_p - (keep8_73 -keep8_71)
955 memdispo_p = min(memdispo_p, keep8(75)-keep8_73)
956 memneeded_p = mumps_tps_arr(thread_id_p)%LA -
957 & mumps_tps_arr(thread_id_p)%LRLUS
958 memneededforcb_p = memneeded_p -
959 & ( mumps_tps_arr(thread_id_p)%POSFAC - 1_8 )
960 IF ( memdispo_p .GE. memneeded_p ) THEN
961!$OMP ATOMIC CAPTURE
962 keep8(73) = keep8(73) + memneeded_p
963 keep8_73 = keep8(73)
964!$OMP END ATOMIC
965!$OMP ATOMIC READ
966 keep8_71 = keep8(71)
967!$OMP END ATOMIC
968 memdispo_p = peakauthorized_p - (keep8_73 -keep8_71)
969 memdispo_p = min(memdispo_p, keep8(75)-keep8_73)
970 IF ( memdispo_p .LT. 0 ) THEN
971!$OMP ATOMIC UPDATE
972 keep8(73) = keep8(73) - memneeded_p
973!$OMP END ATOMIC
974 IF ( state(thread_id_p) .NE. waitmem ) THEN
975 state( thread_id_p ) = waitmem
976 nbwaitmem = nbwaitmem + 1
977 ENDIF
978 ELSE
979!$OMP ATOMIC UPDATE
980 keep8(74) = max(keep8(74), keep8_73 )
981!$OMP END ATOMIC
982 IF ( state( thread_id_p ) .EQ. waitmem ) THEN
983 nbwaitmem = nbwaitmem - 1
984 ENDIF
985 state( thread_id_p ) = allocatevidercbencours
986 local_action = allocatevidercb
987 nbongoingcopies = nbongoingcopies + 1
988 ENDIF
989 ELSE
990 IF ( state(thread_id_p) .NE. waitmem ) THEN
991 state( thread_id_p ) = waitmem
992 nbwaitmem = nbwaitmem + 1
993 ENDIF
994 ENDIF
995 CASE DEFAULT
996 ith = -1
997 DO k = thread_id_p, thread_id_p + keep(400) - 1
998 IF ( k > keep(400) ) THEN
999 ith = k - keep(400)
1000 ELSE
1001 ith = k
1002 ENDIF
1003 IF ( state(ith) .GE. 0 .AND.
1004 & state(ith) .LT. mumps_tps_arr(ith)%POSFAC - 1_8 ) THEN
1005 EXIT
1006 ELSE
1007 ith = -1
1008 ENDIF
1009 ENDDO
1010 IF ( ith .GT. 0 ) THEN
1011 local_action = copy_factors
1012 copy_start = state(ith) + 1
1013 chunk8 = max(
1014 &
1015 & int(keep(361),8),
1016 &
1017 & (mumps_tps_arr(ith)%POSFAC+keep(400)-2_8) /
1018 & (int(keep(400)*2,8))
1019 &
1020 & )
1021 IF (keep(72) .EQ. 1) THEN
1022 chunk8 = 4_8
1023 ENDIF
1024 chunk8 = min( chunk8,
1025 & mumps_tps_arr(ith)%POSFAC - 1_8 - copy_start + 1_8
1026 & )
1027 state(ith) = state(ith) + chunk8
1028 ENDIF
1029 END SELECT
1030 ENDIF
1031!$OMP END CRITICAL(L0_COPIES)
1032 SELECT CASE ( local_action )
1033 CASE ( free_work_myid )
1034 IF ( associated(smumps_tps_arr(thread_id_p)%A) ) THEN
1035 DEALLOCATE(smumps_tps_arr(thread_id_p)%A)
1036 NULLIFY(smumps_tps_arr(thread_id_p)%A)
1037 CALL mumps_dm_fac_upd_dyn_memcnts(
1038 & -mumps_tps_arr(thread_id_p)%LA,
1039 & keep(405).EQ.1,
1040 & keep8,
1041 & info_p(1), info_p(2),
1042 & .false., .false. )
1043 IF (info_p(1) .GE. 0) THEN
1044!$OMP ATOMIC UPDATE
1045 keep8(69) = keep8(69) - l0_omp_factors(thread_id_p)%LA
1046!$OMP END ATOMIC
1047!$OMP CRITICAL(L0_COPIES)
1048 DO ith = 1, keep(400)
1049 IF ( state(ith) .EQ. waitmem ) THEN
1050 state(ith)=copynotstarted
1051 ENDIF
1052 ENDDO
1053 nbwaitmem = 0
1054 nbfinished = nbfinished + 1
1055 state( thread_id_p ) = finished
1056 nbongoingcopies = nbongoingcopies -1
1057!$OMP END CRITICAL(L0_COPIES)
1058 ENDIF
1059 ENDIF
1060 CASE ( allocatevidercb )
1061 to_allocate = max(mumps_tps_arr(thread_id_p)%POSFAC-1_8,1_8)
1062 ALLOCATE( l0_omp_factors(thread_id_p)%A(to_allocate),
1063 & stat=allocok )
1064 IF ( allocok .GT. 0 ) THEN
1065 info_p(1) = -13
1066 CALL mumps_seti8toi4(to_allocate, info_p(2))
1067 l0_omp_factors(thread_id_p)%LA = 0_8
1068!$omp critical(l0_copies)
1069!$OMP ATOMIC UPDATE
1070 keep8(73) = keep8(73) - memneeded_p
1071!$OMP END ATOMIC
1072!$OMP END CRITICAL(L0_COPIES)
1073 ELSE
1074 l0_omp_factors(thread_id_p)%LA =
1075 & mumps_tps_arr(thread_id_p)%POSFAC-1_8
1076!$OMP ATOMIC UPDATE
1077 keep8(69) = keep8(69) + l0_omp_factors(thread_id_p)%LA
1078!$OMP END ATOMIC
1079!$OMP CRITICAL(L0_COPIES)
1080 IF ( mumps_tps_arr(thread_id_p)%POSFAC-1_8 == 0_8 ) THEN
1081 state(thread_id_p) = copyfactorsfinished
1082 ELSE
1083 state( thread_id_p ) = 0
1084 size_copied( thread_id_p ) = 0
1085 ENDIF
1086!$OMP END CRITICAL(L0_COPIES)
1087 lrlus_save_p = mumps_tps_arr(thread_id_p)%LRLUS
1088 CALL smumps_dm_cbstatic2dynamic_i
1089 & (3,
1090 & 0_8,
1091 & .false.,
1092 & myid_nodes, n, slavef, keep, keep8,
1093 & mumps_tps_arr(thread_id_p)%IW(1),
1094 & mumps_tps_arr(thread_id_p)%LIW,
1095 & mumps_tps_arr(thread_id_p)%IWPOSCB,
1096 & mumps_tps_arr(thread_id_p)%IWPOS,
1097 & smumps_tps_arr(thread_id_p)%A(1),
1098 & mumps_tps_arr(thread_id_p)%LA,
1099 & mumps_tps_arr(thread_id_p)%LRLU,
1100 & mumps_tps_arr(thread_id_p)%IPTRLU,
1101 & mumps_tps_arr(thread_id_p)%LRLUS,
1102 & step, ptrast, pamaster, procnode_steps, dad,
1103 & info_p(1), info_p(2) )
1104 cbcopiedtodynamic_p =
1105 & mumps_tps_arr(thread_id_p)%LRLUS - lrlus_save_p
1106 IF (info_p(1) .LT. 0 ) THEN
1107!$OMP ATOMIC UPDATE
1108 keep8(73) = keep8(73) -
1109 & ( memneededforcb_p - cbcopiedtodynamic_p )
1110!$OMP END ATOMIC
1111 ELSE
1112 ENDIF
1113 ENDIF
1114 CASE ( copy_factors )
1115 DO i8 = copy_start, copy_start + chunk8 - 1
1116 l0_omp_factors(ith)%A(i8) = smumps_tps_arr(ith)%A(i8)
1117 ENDDO
1118!$OMP CRITICAL(L0_COPIES)
1119 size_copied(ith) = size_copied(ith) + chunk8
1120 IF ( size_copied(ith) .EQ. l0_omp_factors(ith)%LA ) THEN
1121 state(ith) = copyfactorsfinished
1122 ENDIF
1123!$OMP END CRITICAL(L0_COPIES)
1124 CASE ( nothing )
1125 CASE ( dormir )
1126 CALL mumps_usleep(1000)
1127 CASE DEFAULT
1128 WRITE(*,*) " Internal error in SMUMPS_PERFORM_COPIES",
1129 & local_action
1130 END SELECT
1131 ENDDO
1132 RETURN
1133 END SUBROUTINE smumps_perform_copies
1134 END MODULE smumps_fac_omp_m
1135 RECURSIVE SUBROUTINE smumps_process_front_niv1( COMM_LOAD,
1136 & ASS_IRECV, N, INODE, TYPE, TYPEF, LA, IW, LIW, A,
1137 & MAXFRW, NOFFNEGW, NPVW, NB22T1W, NBTINYW,
1138 & DET_EXPW, DET_MANTW, DET_SIGNW,
1139 & INFO_P, UU, SEUIL, SEUIL_LDLT_NIV2,
1140 & OPELIW, NELVAW, NMAXNPIVW, NSTEPSW, PTRIST, PTLUST_S,
1141 & PTRFAC, PTRAST, STEP, PIMASTER, PAMASTER, NE, POSFAC,
1142 & LRLU, LRLUS, LRLUSM, IPTRLU, ICNTL, KEEP, KEEP8, DKEEP,
1143 & PIVNUL_LIST, LPN_LIST, COMP, IWPOS, IWPOSCB, PROCNODE_STEPS,
1144 & SLAVEF, IFATH, COMM_NODES, MYID_NODES, IPOOL_P,
1145 & LPOOL_P, LEAF, PERM, NSTK_STEPS, BUFR, LBUFR,
1146 & LBUFR_BYTES, NBFIN, root, OPASSW, ITLOC, RHS_MUMPS, FILS,
1147 & PTRARW, PTRAIW, INTARR, DBLARR, ND, FRERE, DAD, LPTRAR, NELT,
1148 & FRTPTR, FRTELT, ISTEP_TO_INIV2, TAB_POS_IN_PERE, LRGROUPS,
1149 & FLOP_ESTIM_ACC )
1150 USE smumps_fac_asm_master_m
1151 USE smumps_fac_asm_master_elt_m
1152 USE smumps_fac1_lu_m
1153 USE smumps_fac1_ldlt_m
1154 USE smumps_struc_def, ONLY : smumps_root_struc
1155 IMPLICIT NONE
1156 include 'mumps_headers.h'
1157 TYPE (smumps_root_struc) :: root
1158 INTEGER comm_load, ass_irecv
1159 INTEGER comm_nodes, myid_nodes, TYPE, TYPEF
1160 INTEGER n, liw, inode,info_p(2)
1161 INTEGER icntl(60), keep(500)
1162 REAL dkeep(230)
1163 REAL UU, seuil, seuil_ldlt_niv2
1164 INTEGER(8) keep8(150)
1165 INTEGER(8) :: la, posfac, lrlu, lrlus, lrlusm, iptrlu
1166 INTEGER iwposcb, iwpos,
1167 & ifath, slavef, nelvaw, nmaxnpivw, nstepsw
1168 INTEGER iw(liw),procnode_steps(keep(28))
1169 INTEGER(8) :: ptrast (keep(28))
1170 INTEGER(8) :: ptrfac (keep(28))
1171 INTEGER(8) :: pamaster(keep(28))
1172 INTEGER ptrist(keep(28)), ptlust_s(keep(28))
1173 INTEGER step(n), pimaster(keep(28)), ne(keep(28))
1174 REAL a(la)
1175 INTEGER :: maxfrw, noffnegw, npvw, nbtinyw
1176 INTEGER, intent(in) :: lrgroups(n)
1177 DOUBLE PRECISION opassw, opeliw
1178 REAL dblarr(keep8(26))
1179 INTEGER intarr(keep8(27))
1180 INTEGER itloc( n + keep(253) ), fils( n ),
1181 & nd( keep(28) ), frere( keep(28) )
1182 INTEGER dad( keep(28) )
1183 REAL :: rhs_mumps(keep(255))
1184 INTEGER istep_to_iniv2(KEEP(71)),
1185 & tab_pos_in_pere(slavef+2,max(1,keep(56)))
1186 INTEGER nelt, lptrar
1187 INTEGER frtptr( n+1 ), frtelt( nelt )
1188 INTEGER(8), INTENT(IN) :: ptraiw( lptrar ), ptrarw( lptrar )
1189 INTEGER leaf, comp
1190 INTEGER :: nb22t1w, det_expw, det_signw
1191 REAL :: det_mantw
1192 INTEGER perm( n )
1193 INTEGER nstk_steps( keep(28) )
1194 INTEGER lbufr, lbufr_bytes
1195 INTEGER bufr( lbufr )
1196 INTEGER nbfin
1197 INTEGER :: lpn_list
1198 INTEGER :: pivnul_list(lpn_list)
1199 DOUBLE PRECISION flop_estim_acc
1200 INTEGER, INTENT(IN) :: lpool_p
1201 INTEGER, INTENT(IN) :: ipool_p(lpool_p)
1202 INTEGER :: ioldps, jobass, etatass
1203 INTEGER(8) :: poselt
1204 LOGICAL :: avoid_delayed, son_level2
1205 jobass = 0
1206 etatass = 0
1207 IF ( keep(55) .EQ. 0 ) THEN
1208 jobass = 0
1209 CALL smumps_fac_asm_niv1( comm_load, ass_irecv, n, inode,
1210 & iw, liw, a, la,
1211 & info_p, nd, fils, frere, dad, maxfrw, root, opassw,
1212 & opeliw, ptrist, ptlust_s, ptrfac, ptrast, step, pimaster,
1213 & pamaster, ptrarw, ptraiw, itloc, rhs_mumps, nstepsw,
1214 & son_level2,comp, lrlu, iptrlu,
1215 & iwpos, iwposcb, posfac,
1216 & lrlus, lrlusm, icntl, keep, keep8, dkeep,
1217 & intarr, keep8(27), dblarr, keep8(26),
1218 & nstk_steps, procnode_steps, slavef, comm_nodes,
1219 & myid_nodes, bufr, lbufr, lbufr_bytes, nbfin, ipool_p,
1220 & lpool_p, leaf, perm, istep_to_iniv2, tab_pos_in_pere,
1221 & jobass, etatass
1222 & , lrgroups
1223 & )
1224 ELSE
1225 CALL smumps_fac_asm_niv1_elt(comm_load,ass_irecv,nelt,frtptr,
1226 & frtelt, n, inode, iw, liw, a,
1227 & la, info_p, nd, fils, frere, dad, maxfrw,
1228 & root, opassw, opeliw, ptrist, ptlust_s, ptrfac, ptrast, step,
1229 & pimaster, pamaster, ptrarw, ptraiw, itloc,
1230 & rhs_mumps, nstepsw, son_level2, comp, lrlu,
1231 & iptrlu, iwpos, iwposcb,
1232 & posfac, lrlus, lrlusm, icntl, keep, keep8, dkeep,
1233 & intarr, keep8(27), dblarr, keep8(26),
1234 & nstk_steps, procnode_steps, slavef,
1235 & comm_nodes, myid_nodes, bufr, lbufr, lbufr_bytes, nbfin,
1236 & ipool_p, lpool_p, leaf, perm, istep_to_iniv2,
1237 & tab_pos_in_pere
1238 & , lrgroups
1239 & )
1240 ENDIF
1241 IF (info_p(1) .LT. 0) THEN
1242 RETURN
1243 ENDIF
1244 avoid_delayed = ( ( ifath .EQ. keep(20)
1245 & .OR.
1246 & ifath .EQ. keep(38) )
1247 & .AND.
1248 & ( keep(60) .NE. 0 ) )
1249 poselt = ptrast(step(inode))
1250 ioldps = ptlust_s(step(inode))
1251 IF ( keep(50) .EQ. 0 ) THEN
1252 CALL smumps_fac1_lu( n, inode,
1253 & iw, liw,
1254 & a, la, ioldps,
1255 & poselt,
1256 & info_p(1), info_p(2), uu, noffnegw, npvw, nbtinyw,
1257 & det_expw, det_mantw, det_signw,
1258 & keep, keep8,
1259 & step, procnode_steps, myid_nodes, slavef,
1260 & seuil, avoid_delayed, etatass,
1261 & dkeep(1), pivnul_list(1), lpn_list, iwpos
1262 & , lrgroups
1263 & , perm
1264 & )
1265 ELSE
1266 iw( ioldps + 4 + keep(ixsz) ) = 1
1267 CALL smumps_fac1_ldlt( n, inode,
1268 & iw, liw, a,
1269 & la,
1270 & ioldps, poselt,
1271 & info_p(1), info_p(2), uu, noffnegw, npvw,
1272 & nb22t1w, nbtinyw, det_expw, det_mantw, det_signw,
1273 & keep, keep8, myid_nodes, seuil,
1274 & avoid_delayed,
1275 & etatass,
1276 & dkeep(1), pivnul_list(1), lpn_list, iwpos
1277 & , lrgroups
1278 & , perm
1279 & )
1280 iw(ioldps + 4 + keep(ixsz)) = step(inode)
1281 ENDIF
1282 IF (info_p(1) .LT. 0) THEN
1283 RETURN
1284 ENDIF
1285 CALL smumps_fac_stack( comm_load, ass_irecv, n, inode, TYPE,
1286 &typef, la, iw, liw, a,
1287 &info_p(1), info_p(2), opeliw, nelvaw, nmaxnpivw, ptrist, ptlust_s,
1288 &ptrfac, ptrast, step, pimaster, pamaster, ne, posfac,
1289 &lrlu, lrlus, lrlusm, iptrlu, icntl, keep,
1290 &keep8, dkeep,
1291 &comp,iwpos, iwposcb, procnode_steps,
1292 &slavef, IFATH, comm_nodes, myid_nodes, ipool_p,
1293 &lpool_p, leaf, nstk_steps, perm, bufr, lbufr,
1294 &lbufr_bytes, nbfin, root, opassw, itloc, rhs_mumps,
1295 &fils, dad, ptrarw, ptraiw, intarr, dblarr, nd, frere,
1296 &lptrar, nelt, frtptr, frtelt, istep_to_iniv2, tab_pos_in_pere,
1297 & lrgroups,
1298 & flop_estim_acc
1299 &)
1300 RETURN
1301 END SUBROUTINE smumps_process_front_niv1
cmplx
float cmplx[2]
Definition pblas.h:136
mumps_abort
#define mumps_abort
Definition VE_Metis.h:25
min
#define min(a, b)
Definition macros.h:20
max
#define max(a, b)
Definition macros.h:21
mpi_wtime
double precision function mpi_wtime()
Definition mpi.f:561
mumps_tps_m
Definition omp_tps_common_m.F:14
smumps_dynamic_memory_m
Definition sfac_mem_dynamic.F:14
smumps_dynamic_memory_m::smumps_dm_fac_alloc_allowed
subroutine smumps_dm_fac_alloc_allowed(mem_count_to_allocate, keep8, iflag, ierror)
Definition sfac_mem_dynamic.F:123
smumps_fac1_ldlt_m
Definition sfac_front_LDLT_type1.F:14
smumps_fac1_ldlt_m::smumps_fac1_ldlt
subroutine smumps_fac1_ldlt(n, inode, iw, liw, a, la, ioldps, poselt, iflag, ierror, uu, nnegw, npvw, nb22t1w, nbtinyw, det_expw, det_mantw, det_signw, keep, keep8, myid, seuil, avoid_delayed, etatass, dkeep, pivnul_list, lpn_list, iwpos, lrgroups, perm)
Definition sfac_front_LDLT_type1.F:26
smumps_fac1_lu_m
Definition sfac_front_LU_type1.F:14
smumps_fac1_lu_m::smumps_fac1_lu
subroutine smumps_fac1_lu(n, inode, iw, liw, a, la, ioldps, poselt, iflag, ierror, uu, noffw, npvw, nbtinyw, det_expw, det_mantw, det_signw, keep, keep8, step, procnode_steps, myid, slavef, seuil, avoid_delayed, etatass, dkeep, pivnul_list, lpn_list, iwpos, lrgroups, perm)
Definition sfac_front_LU_type1.F:30
smumps_fac_asm_master_elt_m
Definition sfac_asm_master_ELT_m.F:14
smumps_fac_asm_master_elt_m::smumps_fac_asm_niv1_elt
subroutine smumps_fac_asm_niv1_elt(comm_load, ass_irecv, nelt, frt_ptr, frt_elt, n, inode, iw, liw, a, la, info, nd, fils, frere, dad, maxfrw, root, opassw, opeliw, ptrist, ptlust, ptrfac, ptrast, step, pimaster, pamaster, ptrarw, ptraiw, itloc, rhs_mumps, nsteps, son_level2, comp, lrlu, iptrlu, iwpos, iwposcb, posfac, lrlus, lrlusm, icntl, keep, keep8, dkeep, intarr, lintarr, dblarr, ldblarr nstk_s, procnode_steps, slavef, comm, myid, bufr, lbufr, lbufr_bytes, nbfin, ipool, lpool, leaf, perm, istep_to_iniv2, tab_pos_in_pere, lrgroups, mumps_tps_arr, smumps_tps_arr, l0_omp_mapping)
Definition sfac_asm_master_ELT_m.F:33
smumps_fac_asm_master_m
Definition sfac_asm_master_m.F:14
smumps_fac_asm_master_m::smumps_fac_asm_niv1
subroutine smumps_fac_asm_niv1(comm_load, ass_irecv, n, inode, iw, liw, a, la, info, nd, fils, frere, dad, maxfrw, root, opassw, opeliw, ptrist, ptlust, ptrfac, ptrast, step, pimaster, pamaster, ptrarw, ptraiw, itloc, rhs_mumps, nsteps, son_level2, comp, lrlu, iptrlu, iwpos, iwposcb, posfac, lrlus, lrlusm, icntl, keep, keep8, dkeep, intarr, lintarr, dblarr, ldblarr nstk_s, procnode_steps, slavef, comm, myid, bufr, lbufr, lbufr_bytes, nbfin, ipool, lpool, leaf, perm, istep_to_iniv2, tab_pos_in_pere, jobass, etatass, lrgroups, mumps_tps_arr, smumps_tps_arr, l0_omp_mapping)
Definition sfac_asm_master_m.F:32
smumps_fac_omp_m
Definition sfac_omp_m.F:14
smumps_fac_omp_m::smumps_fac_l0_omp
subroutine smumps_fac_l0_omp(n, liw, nstk_steps, nd, fils, step, frere, dad, istep_to_iniv2, tab_pos_in_pere, ptrist, ptrast, pimaster, pamaster, ptrarw, ptraiw, itloc, rhs_mumps, rinfo, nbroot, nbrtot, nbroot_under_l0, uu, icntl, ptlust_s, ptrfac, info, keep, keep8, procnode_steps, slavef, comm_nodes, myid_nodes, bufr, lbufr, lbufr_bytes, intarr, dblarr, root, perm, nelt, frtptr, frtelt, lptrar, comm_load, ass_irecv, seuil, seuil_ldlt_niv2, mem_distrib, ne, dkeep, pivnul_list, lpn_list, lpool_b_l0_omp, ipool_b_l0_omp, l_virt_l0_omp, virt_l0_omp, virt_l0_omp_mapping, l_phys_l0_omp, phys_l0_omp, perm_l0_omp, ptr_leafs_l0_omp, l0_omp_mapping, ll0_omp_mapping, thread_la, mumps_tps_arr, smumps_tps_arr, nstepsw, opassw, opeliw, nelvaw, comp, maxfrw, nmaxnpivw, npvw, noffnegw, nb22t1w, nbtinyw, det_expw, det_mantw, det_signw, lrgroups, l0_omp_factors, ll0_omp_factors, i4_l0_omp, nbstats_i4, nbcols_i4, i8_l0_omp, nbstats_i8, nbcols_i8)
Definition sfac_omp_m.F:42
smumps_fac_omp_m::smumps_ma_eff_mem_dispo
subroutine smumps_ma_eff_mem_dispo(mumps_tps_arr, nbthreads, keep8, keep, n, blr_strat, lpool_p, i8_l0_omp, nbstats_i8, memdispo_underl0)
Definition sfac_omp_m.F:685
smumps_fac_omp_m::finished
integer(8), parameter finished
Definition sfac_omp_m.F:20
smumps_fac_omp_m::smumps_set_maxs_maxis_thread
subroutine smumps_set_maxs_maxis_thread(maxs_base_relaxed8th, maxis_base_relaxedth, blr_strat, keep, i4_l0_ompth, nbstats_i4, i8_l0_ompth, nbstats_i8)
Definition sfac_omp_m.F:650
smumps_fac_omp_m::underl0
integer(8), parameter underl0
Definition sfac_omp_m.F:15
smumps_fac_omp_m::smumps_perform_copies
subroutine smumps_perform_copies(thread_id_p, mumps_tps_arr, smumps_tps_arr, l0_omp_factors, ll0_omp_factors, state, size_copied, nbwaitmem, nbfinished, nbongoingcopies, nbunderl0, myid_nodes, n, slavef, step, ptrast, pamaster, procnode_steps, dad, keep, keep8, info_p)
Definition sfac_omp_m.F:854
smumps_fac_omp_m::smumps_perform_copies_init
subroutine smumps_perform_copies_init(state, keep8_77_save, nbwaitmem, nbfinished, nbongoingcopies, nbunderl0, keep, keep8)
Definition sfac_omp_m.F:827
smumps_fac_omp_m::allocatevidercbencours
integer(8), parameter allocatevidercbencours
Definition sfac_omp_m.F:19
smumps_fac_omp_m::smumps_l0omp_copy_iw
subroutine smumps_l0omp_copy_iw(iw, liw, iwpos, mumps_tps_arr, keep, ptlust, icntl, info)
Definition sfac_omp_m.F:770
smumps_fac_omp_m::copynotstarted
integer(8), parameter copynotstarted
Definition sfac_omp_m.F:16
smumps_fac_omp_m::copyfactorsfinished
integer(8), parameter copyfactorsfinished
Definition sfac_omp_m.F:18
smumps_fac_omp_m::waitmem
integer(8), parameter waitmem
Definition sfac_omp_m.F:17
smumps_load
Definition smumps_load.F:14
smumps_load::mumps_load_disable
subroutine, public mumps_load_disable()
Definition smumps_load.F:135
smumps_load::smumps_load_update
subroutine, public smumps_load_update(check_flops, process_bande, inc_load, keep, keep8)
Definition smumps_load.F:824
smumps_load::mumps_load_enable
subroutine, public mumps_load_enable()
Definition smumps_load.F:130
smumps_lr_stats
Definition slr_stats.F:14
smumps_lr_stats::time_frtrsm
double precision time_frtrsm
Definition slr_stats.F:62
smumps_lr_stats::time_update_lrlr3
double precision time_update_lrlr3
Definition slr_stats.F:52
smumps_lr_stats::time_lrtrsm
double precision time_lrtrsm
Definition slr_stats.F:61
smumps_lr_stats::time_decomp_asm1
double precision time_decomp_asm1
Definition slr_stats.F:71
smumps_lr_stats::time_update_frlr
double precision time_update_frlr
Definition slr_stats.F:53
smumps_lr_stats::time_decomp_asms2s
double precision time_decomp_asms2s
Definition slr_stats.F:74
smumps_lr_stats::time_diagcopy
double precision time_diagcopy
Definition slr_stats.F:68
smumps_lr_stats::time_decomp_locasm2
double precision time_decomp_locasm2
Definition slr_stats.F:72
smumps_lr_stats::time_fac_mq
double precision time_fac_mq
Definition slr_stats.F:65
smumps_lr_stats::time_decomp
double precision time_decomp
Definition slr_stats.F:69
smumps_lr_stats::time_update
double precision time_update
Definition slr_stats.F:49
smumps_lr_stats::time_panel
double precision time_panel
Definition slr_stats.F:63
smumps_lr_stats::time_update_frfr
double precision time_update_frfr
Definition slr_stats.F:54
smumps_lr_stats::time_decomp_maplig1
double precision time_decomp_maplig1
Definition slr_stats.F:73
smumps_lr_stats::time_fac_i
double precision time_fac_i
Definition slr_stats.F:64
smumps_lr_stats::time_frswap_compress
double precision time_frswap_compress
Definition slr_stats.F:57
smumps_lr_stats::time_update_lrlr1
double precision time_update_lrlr1
Definition slr_stats.F:50
smumps_lr_stats::time_fac_sq
double precision time_fac_sq
Definition slr_stats.F:66
smumps_lr_stats::time_decomp_asms2m
double precision time_decomp_asms2m
Definition slr_stats.F:75
smumps_lr_stats::time_decomp_ucfs
double precision time_decomp_ucfs
Definition slr_stats.F:70
smumps_lr_stats::time_lr_module
double precision time_lr_module
Definition slr_stats.F:59
smumps_lr_stats::time_cb_compress
double precision time_cb_compress
Definition slr_stats.F:58
smumps_lr_stats::time_frfronts
double precision time_frfronts
Definition slr_stats.F:67
smumps_lr_stats::time_update_lrlr2
double precision time_update_lrlr2
Definition slr_stats.F:51
smumps_lr_stats::time_compress
double precision time_compress
Definition slr_stats.F:55
smumps_lr_stats::time_midblk_compress
double precision time_midblk_compress
Definition slr_stats.F:56
smumps_struc_def
Definition smumps_struc_def.F:14
smumps_tps_m
Definition somp_tps_m.F:14
comp
int comp(int a, int b)
Definition set_surface_lines.cpp:102
smumps_alloc_cb
subroutine smumps_alloc_cb(inplace, min_space_in_place, ssarbr, process_bande, myid, n, keep, keep8, dkeep, iw, liw, a, la, lrlu, iptrlu, iwpos, iwposcb, slavef, procnode_steps, dad, ptrist, ptrast, step, pimaster, pamaster, lreq, lreqcb, node_arg, state_arg, set_header, comp, lrlus, lrlusm, iflag, ierror)
Definition sfac_mem_alloc_cb.F:23
smumps_dm_cbstatic2dynamic_i
subroutine smumps_dm_cbstatic2dynamic_i(strategy, sizer_needed, skip_top_stack, myid, n, slavef, keep, keep8, iw, liw, iwposcb, iwpos, a, la, lrlu, iptrlu, lrlus, step, ptrast, pamaster, procnode_steps, dad, iflag, ierror)
Definition sfac_mem_dynamic.F:483
smumps_fac_stack
subroutine smumps_fac_stack(comm_load, ass_irecv, n, inode, type, typef, la, iw, liw, a, iflag, ierror, opeliw, nelvaw, nmaxnpiv, ptrist, ptlust_s, ptrfac, ptrast, step, pimaster, pamaster, ne, posfac, lrlu, lrlus, lrlusm, iptrlu, icntl, keep, keep8, dkeep, comp, iwpos, iwposcb, procnode_steps, slavef, fpere, comm, myid, ipool, lpool, leaf, nstk_s, perm, bufr, lbufr, lbufr_bytes, nbfin, root, opassw, itloc, rhs_mumps, fils, dad, ptrarw, ptraiw, intarr, dblarr, nd, frere, lptrar, nelt, frtptr, frtelt, istep_to_iniv2, tab_pos_in_pere, lrgroups, flop_estim_acc)
Definition sfac_mem_stack.F:32
smumps_process_front_niv1
recursive subroutine smumps_process_front_niv1(comm_load, ass_irecv, n, inode, type, typef, la, iw, liw, a, maxfrw, noffnegw, npvw, nb22t1w, nbtinyw, det_expw, det_mantw, det_signw, info_p, uu, seuil, seuil_ldlt_niv2, opeliw, nelvaw, nmaxnpivw, nstepsw, ptrist, ptlust_s, ptrfac, ptrast, step, pimaster, pamaster, ne, posfac, lrlu, lrlus, lrlusm, iptrlu, icntl, keep, keep8, dkeep, pivnul_list, lpn_list, comp, iwpos, iwposcb, procnode_steps, slavef, ifath, comm_nodes, myid_nodes, ipool_p, lpool_p, leaf, perm, nstk_steps, bufr, lbufr, lbufr_bytes, nbfin, root, opassw, itloc, rhs_mumps, fils, ptrarw, ptraiw, intarr, dblarr, nd, frere, dad, lptrar, nelt, frtptr, frtelt, istep_to_iniv2, tab_pos_in_pere, lrgroups, flop_estim_acc)
Definition sfac_omp_m.F:1150
smumps_init_pool_last3
subroutine smumps_init_pool_last3(ipool, lpool, leaf)
Definition sfac_sol_pool.F:15
smumps_extract_pool
subroutine smumps_extract_pool(n, pool, lpool, procnode, slavef, step, inode, keep, keep8, myid, nd, force_extract_top_sbtr)
Definition sfac_sol_pool.F:171
smumps_insert_pool_n
subroutine smumps_insert_pool_n(n, pool, lpool, procnode, slavef, keep199, k28, k76, k80, k47, step, inode)
Definition sfac_sol_pool.F:27
smumps_set_blrstrat_and_maxs
subroutine smumps_set_blrstrat_and_maxs(maxs_base8, maxs_base_relaxed8, blr_strat, keep, nrlnec, nrlnec_active, nrlnec_if_lr_lu, nrlnec_if_lr_lucb, nrlnecooc_if_lr_lucb, nrlnec_if_lr_cb)
Definition stools.F:1191
mumps_tps_m::mumps_tps_t
Definition omp_tps_common_m.F:15
smumps_tps_m::smumps_tps_t
Definition somp_tps_m.F:15
mumps_seti8toi4
subroutine mumps_seti8toi4(i8, i)
Definition tools_common.F:843
mumps_sort_int8
subroutine mumps_sort_int8(n, val, id)
Definition tools_common.F:757
mumps_inssarbr
logical function mumps_inssarbr(procinfo_inode, k199)
Definition tools_common.F:201
mumps_set_ierror
subroutine mumps_set_ierror(size8, ierror)
Definition tools_common.F:865
mumps_dm_fac_upd_dyn_memcnts
subroutine mumps_dm_fac_upd_dyn_memcnts(mem_count_allocated, atomic_updates, keep8, iflag, ierror, k69upd, k71upd)
Definition tools_common.F:1476