Actual source code: plog.c
petsc-3.6.1 2015-07-22
2: /*
3: PETSc code to log object creation and destruction and PETSc events.
5: This provides the public API used by the rest of PETSc and by users.
7: These routines use a private API that is not used elsewhere in PETSc and is not
8: accessible to users. The private API is defined in logimpl.h and the utils directory.
10: */
11: #include <petsc/private/logimpl.h> /*I "petscsys.h" I*/
12: #include <petsctime.h>
13: #include <petscviewer.h>
15: PetscErrorCode PetscLogObjectParent(PetscObject p,PetscObject c)
16: {
17: if (!c || !p) return 0;
18: c->parent = p;
19: c->parentid = p->id;
20: return 0;
21: }
23: PetscErrorCode PetscLogObjectMemory(PetscObject p,PetscLogDouble m)
24: {
25: if (!p) return 0;
26: p->mem += m;
27: return 0;
28: }
30: PetscLogEvent PETSC_LARGEST_EVENT = PETSC_EVENT;
32: #if defined(PETSC_USE_LOG)
33: #include <petscmachineinfo.h>
34: #include <petscconfiginfo.h>
36: /* used in the MPI_XXX() count macros in petsclog.h */
38: /* Action and object logging variables */
39: Action *petsc_actions = NULL;
40: Object *petsc_objects = NULL;
41: PetscBool petsc_logActions = PETSC_FALSE;
42: PetscBool petsc_logObjects = PETSC_FALSE;
43: int petsc_numActions = 0, petsc_maxActions = 100;
44: int petsc_numObjects = 0, petsc_maxObjects = 100;
45: int petsc_numObjectsDestroyed = 0;
47: /* Global counters */
48: PetscLogDouble petsc_BaseTime = 0.0;
49: PetscLogDouble petsc_TotalFlops = 0.0; /* The number of flops */
50: PetscLogDouble petsc_tmp_flops = 0.0; /* The incremental number of flops */
51: PetscLogDouble petsc_send_ct = 0.0; /* The number of sends */
52: PetscLogDouble petsc_recv_ct = 0.0; /* The number of receives */
53: PetscLogDouble petsc_send_len = 0.0; /* The total length of all sent messages */
54: PetscLogDouble petsc_recv_len = 0.0; /* The total length of all received messages */
55: PetscLogDouble petsc_isend_ct = 0.0; /* The number of immediate sends */
56: PetscLogDouble petsc_irecv_ct = 0.0; /* The number of immediate receives */
57: PetscLogDouble petsc_isend_len = 0.0; /* The total length of all immediate send messages */
58: PetscLogDouble petsc_irecv_len = 0.0; /* The total length of all immediate receive messages */
59: PetscLogDouble petsc_wait_ct = 0.0; /* The number of waits */
60: PetscLogDouble petsc_wait_any_ct = 0.0; /* The number of anywaits */
61: PetscLogDouble petsc_wait_all_ct = 0.0; /* The number of waitalls */
62: PetscLogDouble petsc_sum_of_waits_ct = 0.0; /* The total number of waits */
63: PetscLogDouble petsc_allreduce_ct = 0.0; /* The number of reductions */
64: PetscLogDouble petsc_gather_ct = 0.0; /* The number of gathers and gathervs */
65: PetscLogDouble petsc_scatter_ct = 0.0; /* The number of scatters and scattervs */
67: /* Logging functions */
68: PetscErrorCode (*PetscLogPHC)(PetscObject) = NULL;
69: PetscErrorCode (*PetscLogPHD)(PetscObject) = NULL;
70: PetscErrorCode (*PetscLogPLB)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject) = NULL;
71: PetscErrorCode (*PetscLogPLE)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject) = NULL;
73: /* Tracing event logging variables */
74: FILE *petsc_tracefile = NULL;
75: int petsc_tracelevel = 0;
76: const char *petsc_traceblanks = " ";
77: char petsc_tracespace[128] = " ";
78: PetscLogDouble petsc_tracetime = 0.0;
79: static PetscBool PetscLogBegin_PrivateCalled = PETSC_FALSE;
81: /*---------------------------------------------- General Functions --------------------------------------------------*/
84: /*@C
85: PetscLogDestroy - Destroys the object and event logging data and resets the global counters.
87: Not Collective
89: Notes:
90: This routine should not usually be used by programmers. Instead employ
91: PetscLogStagePush() and PetscLogStagePop().
93: Level: developer
95: .keywords: log, destroy
96: .seealso: PetscLogDump(), PetscLogAllBegin(), PetscLogView(), PetscLogStagePush(), PlogStagePop()
97: @*/
98: PetscErrorCode PetscLogDestroy(void)
99: {
100: PetscStageLog stageLog;
104: PetscFree(petsc_actions);
105: PetscFree(petsc_objects);
106: PetscLogSet(NULL, NULL);
108: /* Resetting phase */
109: PetscLogGetStageLog(&stageLog);
110: PetscStageLogDestroy(stageLog);
112: petsc_TotalFlops = 0.0;
113: petsc_numActions = 0;
114: petsc_numObjects = 0;
115: petsc_numObjectsDestroyed = 0;
116: petsc_maxActions = 100;
117: petsc_maxObjects = 100;
118: petsc_actions = NULL;
119: petsc_objects = NULL;
120: petsc_logActions = PETSC_FALSE;
121: petsc_logObjects = PETSC_FALSE;
122: petsc_BaseTime = 0.0;
123: petsc_TotalFlops = 0.0;
124: petsc_tmp_flops = 0.0;
125: petsc_send_ct = 0.0;
126: petsc_recv_ct = 0.0;
127: petsc_send_len = 0.0;
128: petsc_recv_len = 0.0;
129: petsc_isend_ct = 0.0;
130: petsc_irecv_ct = 0.0;
131: petsc_isend_len = 0.0;
132: petsc_irecv_len = 0.0;
133: petsc_wait_ct = 0.0;
134: petsc_wait_any_ct = 0.0;
135: petsc_wait_all_ct = 0.0;
136: petsc_sum_of_waits_ct = 0.0;
137: petsc_allreduce_ct = 0.0;
138: petsc_gather_ct = 0.0;
139: petsc_scatter_ct = 0.0;
140: PETSC_LARGEST_EVENT = PETSC_EVENT;
141: PetscLogPHC = NULL;
142: PetscLogPHD = NULL;
143: petsc_tracefile = NULL;
144: petsc_tracelevel = 0;
145: petsc_traceblanks = " ";
146: petsc_tracespace[0] = ' '; petsc_tracespace[1] = 0;
147: petsc_tracetime = 0.0;
148: PETSC_LARGEST_CLASSID = PETSC_SMALLEST_CLASSID;
149: PETSC_OBJECT_CLASSID = 0;
150: petsc_stageLog = 0;
151: PetscLogBegin_PrivateCalled = PETSC_FALSE;
152: return(0);
153: }
157: /*@C
158: PetscLogSet - Sets the logging functions called at the beginning and ending of every event.
160: Not Collective
162: Input Parameters:
163: + b - The function called at beginning of event
164: - e - The function called at end of event
166: Level: developer
168: .seealso: PetscLogDump(), PetscLogBegin(), PetscLogAllBegin(), PetscLogTraceBegin()
169: @*/
170: PetscErrorCode PetscLogSet(PetscErrorCode (*b)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject),
171: PetscErrorCode (*e)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject))
172: {
174: PetscLogPLB = b;
175: PetscLogPLE = e;
176: return(0);
177: }
179: #if defined(PETSC_HAVE_PAPI)
180: #include <papi.h>
181: int PAPIEventSet = PAPI_NULL;
182: #endif
184: /*------------------------------------------- Initialization Functions ----------------------------------------------*/
187: PetscErrorCode PetscLogBegin_Private(void)
188: {
189: int stage;
190: PetscBool opt;
194: if (PetscLogBegin_PrivateCalled) return(0);
195: PetscLogBegin_PrivateCalled = PETSC_TRUE;
197: PetscOptionsHasName(NULL, "-log_exclude_actions", &opt);
198: if (opt) petsc_logActions = PETSC_FALSE;
199: PetscOptionsHasName(NULL, "-log_exclude_objects", &opt);
200: if (opt) petsc_logObjects = PETSC_FALSE;
201: if (petsc_logActions) {
202: PetscMalloc1(petsc_maxActions, &petsc_actions);
203: }
204: if (petsc_logObjects) {
205: PetscMalloc1(petsc_maxObjects, &petsc_objects);
206: }
207: PetscLogPHC = PetscLogObjCreateDefault;
208: PetscLogPHD = PetscLogObjDestroyDefault;
209: /* Setup default logging structures */
210: PetscStageLogCreate(&petsc_stageLog);
211: PetscStageLogRegister(petsc_stageLog, "Main Stage", &stage);
212: #if defined(PETSC_HAVE_PAPI)
213: PAPI_library_init(PAPI_VER_CURRENT);
214: if (ierr != PAPI_VER_CURRENT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"Cannot initialize PAPI");
215: PAPI_query_event(PAPI_FP_INS);
216: PAPI_create_eventset(&PAPIEventSet);
217: PAPI_add_event(PAPIEventSet,PAPI_FP_INS);
218: PAPI_start(PAPIEventSet);
219: #endif
221: /* All processors sync here for more consistent logging */
222: MPI_Barrier(PETSC_COMM_WORLD);
223: PetscTime(&petsc_BaseTime);
224: PetscLogStagePush(stage);
225: return(0);
226: }
230: /*@C
231: PetscLogBegin - Turns on logging of objects and events. This logs flop
232: rates and object creation and should not slow programs down too much.
233: This routine may be called more than once.
235: Logically Collective over PETSC_COMM_WORLD
237: Options Database Keys:
238: + -log_summary - Prints summary of flop and timing information to the
239: screen (for code compiled with PETSC_USE_LOG)
240: - -log - Prints detailed log information (for code compiled with PETSC_USE_LOG)
242: Usage:
243: .vb
244: PetscInitialize(...);
245: PetscLogBegin();
246: ... code ...
247: PetscLogView(viewer); or PetscLogDump();
248: PetscFinalize();
249: .ve
251: Notes:
252: PetscLogView(viewer) or PetscLogDump() actually cause the printing of
253: the logging information.
255: Level: advanced
257: .keywords: log, begin
258: .seealso: PetscLogDump(), PetscLogAllBegin(), PetscLogView(), PetscLogTraceBegin()
259: @*/
260: PetscErrorCode PetscLogBegin(void)
261: {
265: PetscLogSet(PetscLogEventBeginDefault, PetscLogEventEndDefault);
266: PetscLogBegin_Private();
267: return(0);
268: }
272: /*@C
273: PetscLogAllBegin - Turns on extensive logging of objects and events. Logs
274: all events. This creates large log files and slows the program down.
276: Logically Collective on PETSC_COMM_WORLD
278: Options Database Keys:
279: . -log_all - Prints extensive log information (for code compiled with PETSC_USE_LOG)
281: Usage:
282: .vb
283: PetscInitialize(...);
284: PetscLogAllBegin();
285: ... code ...
286: PetscLogDump(filename);
287: PetscFinalize();
288: .ve
290: Notes:
291: A related routine is PetscLogBegin() (with the options key -log), which is
292: intended for production runs since it logs only flop rates and object
293: creation (and shouldn't significantly slow the programs).
295: Level: advanced
297: .keywords: log, all, begin
298: .seealso: PetscLogDump(), PetscLogBegin(), PetscLogTraceBegin()
299: @*/
300: PetscErrorCode PetscLogAllBegin(void)
301: {
305: PetscLogSet(PetscLogEventBeginComplete, PetscLogEventEndComplete);
306: PetscLogBegin_Private();
307: return(0);
308: }
312: /*@
313: PetscLogTraceBegin - Activates trace logging. Every time a PETSc event
314: begins or ends, the event name is printed.
316: Logically Collective on PETSC_COMM_WORLD
318: Input Parameter:
319: . file - The file to print trace in (e.g. stdout)
321: Options Database Key:
322: . -log_trace [filename] - Activates PetscLogTraceBegin()
324: Notes:
325: PetscLogTraceBegin() prints the processor number, the execution time (sec),
326: then "Event begin:" or "Event end:" followed by the event name.
328: PetscLogTraceBegin() allows tracing of all PETSc calls, which is useful
329: to determine where a program is hanging without running in the
330: debugger. Can be used in conjunction with the -info option.
332: Level: intermediate
334: .seealso: PetscLogDump(), PetscLogAllBegin(), PetscLogView(), PetscLogBegin()
335: @*/
336: PetscErrorCode PetscLogTraceBegin(FILE *file)
337: {
341: petsc_tracefile = file;
343: PetscLogSet(PetscLogEventBeginTrace, PetscLogEventEndTrace);
344: PetscLogBegin_Private();
345: return(0);
346: }
350: /*@
351: PetscLogActions - Determines whether actions are logged for the graphical viewer.
353: Not Collective
355: Input Parameter:
356: . flag - PETSC_TRUE if actions are to be logged
358: Level: intermediate
360: Note: Logging of actions continues to consume more memory as the program
361: runs. Long running programs should consider turning this feature off.
363: Options Database Keys:
364: . -log_exclude_actions - Turns off actions logging
366: .keywords: log, stage, register
367: .seealso: PetscLogStagePush(), PetscLogStagePop()
368: @*/
369: PetscErrorCode PetscLogActions(PetscBool flag)
370: {
372: petsc_logActions = flag;
373: return(0);
374: }
378: /*@
379: PetscLogObjects - Determines whether objects are logged for the graphical viewer.
381: Not Collective
383: Input Parameter:
384: . flag - PETSC_TRUE if objects are to be logged
386: Level: intermediate
388: Note: Logging of objects continues to consume more memory as the program
389: runs. Long running programs should consider turning this feature off.
391: Options Database Keys:
392: . -log_exclude_objects - Turns off objects logging
394: .keywords: log, stage, register
395: .seealso: PetscLogStagePush(), PetscLogStagePop()
396: @*/
397: PetscErrorCode PetscLogObjects(PetscBool flag)
398: {
400: petsc_logObjects = flag;
401: return(0);
402: }
404: /*------------------------------------------------ Stage Functions --------------------------------------------------*/
407: /*@C
408: PetscLogStageRegister - Attaches a charactor string name to a logging stage.
410: Not Collective
412: Input Parameter:
413: . sname - The name to associate with that stage
415: Output Parameter:
416: . stage - The stage number
418: Level: intermediate
420: .keywords: log, stage, register
421: .seealso: PetscLogStagePush(), PetscLogStagePop()
422: @*/
423: PetscErrorCode PetscLogStageRegister(const char sname[],PetscLogStage *stage)
424: {
425: PetscStageLog stageLog;
426: PetscLogEvent event;
430: PetscLogGetStageLog(&stageLog);
431: PetscStageLogRegister(stageLog, sname, stage);
432: /* Copy events already changed in the main stage, this sucks */
433: EventPerfLogEnsureSize(stageLog->stageInfo[*stage].eventLog, stageLog->eventLog->numEvents);
434: for (event = 0; event < stageLog->eventLog->numEvents; event++) {
435: EventPerfInfoCopy(&stageLog->stageInfo[0].eventLog->eventInfo[event],&stageLog->stageInfo[*stage].eventLog->eventInfo[event]);
436: }
437: ClassPerfLogEnsureSize(stageLog->stageInfo[*stage].classLog, stageLog->classLog->numClasses);
438: return(0);
439: }
443: /*@C
444: PetscLogStagePush - This function pushes a stage on the stack.
446: Not Collective
448: Input Parameter:
449: . stage - The stage on which to log
451: Usage:
452: If the option -log_sumary is used to run the program containing the
453: following code, then 2 sets of summary data will be printed during
454: PetscFinalize().
455: .vb
456: PetscInitialize(int *argc,char ***args,0,0);
457: [stage 0 of code]
458: PetscLogStagePush(1);
459: [stage 1 of code]
460: PetscLogStagePop();
461: PetscBarrier(...);
462: [more stage 0 of code]
463: PetscFinalize();
464: .ve
466: Notes:
467: Use PetscLogStageRegister() to register a stage.
469: Level: intermediate
471: .keywords: log, push, stage
472: .seealso: PetscLogStagePop(), PetscLogStageRegister(), PetscBarrier()
473: @*/
474: PetscErrorCode PetscLogStagePush(PetscLogStage stage)
475: {
476: PetscStageLog stageLog;
480: PetscLogGetStageLog(&stageLog);
481: PetscStageLogPush(stageLog, stage);
482: return(0);
483: }
487: /*@C
488: PetscLogStagePop - This function pops a stage from the stack.
490: Not Collective
492: Usage:
493: If the option -log_sumary is used to run the program containing the
494: following code, then 2 sets of summary data will be printed during
495: PetscFinalize().
496: .vb
497: PetscInitialize(int *argc,char ***args,0,0);
498: [stage 0 of code]
499: PetscLogStagePush(1);
500: [stage 1 of code]
501: PetscLogStagePop();
502: PetscBarrier(...);
503: [more stage 0 of code]
504: PetscFinalize();
505: .ve
507: Notes:
508: Use PetscLogStageRegister() to register a stage.
510: Level: intermediate
512: .keywords: log, pop, stage
513: .seealso: PetscLogStagePush(), PetscLogStageRegister(), PetscBarrier()
514: @*/
515: PetscErrorCode PetscLogStagePop(void)
516: {
517: PetscStageLog stageLog;
521: PetscLogGetStageLog(&stageLog);
522: PetscStageLogPop(stageLog);
523: return(0);
524: }
528: /*@
529: PetscLogStageSetActive - Determines stage activity for PetscLogEventBegin() and PetscLogEventEnd().
531: Not Collective
533: Input Parameters:
534: + stage - The stage
535: - isActive - The activity flag, PETSC_TRUE for logging, else PETSC_FALSE (defaults to PETSC_TRUE)
537: Level: intermediate
539: .seealso: PetscLogStagePush(), PetscLogStagePop(), PetscLogEventBegin(), PetscLogEventEnd(), PetscPreLoadBegin(), PetscPreLoadEnd(), PetscPreLoadStage()
540: @*/
541: PetscErrorCode PetscLogStageSetActive(PetscLogStage stage, PetscBool isActive)
542: {
543: PetscStageLog stageLog;
547: PetscLogGetStageLog(&stageLog);
548: PetscStageLogSetActive(stageLog, stage, isActive);
549: return(0);
550: }
554: /*@
555: PetscLogStageGetActive - Returns stage activity for PetscLogEventBegin() and PetscLogEventEnd().
557: Not Collective
559: Input Parameter:
560: . stage - The stage
562: Output Parameter:
563: . isActive - The activity flag, PETSC_TRUE for logging, else PETSC_FALSE (defaults to PETSC_TRUE)
565: Level: intermediate
567: .seealso: PetscLogStagePush(), PetscLogStagePop(), PetscLogEventBegin(), PetscLogEventEnd(), PetscPreLoadBegin(), PetscPreLoadEnd(), PetscPreLoadStage()
568: @*/
569: PetscErrorCode PetscLogStageGetActive(PetscLogStage stage, PetscBool *isActive)
570: {
571: PetscStageLog stageLog;
575: PetscLogGetStageLog(&stageLog);
576: PetscStageLogGetActive(stageLog, stage, isActive);
577: return(0);
578: }
582: /*@
583: PetscLogStageSetVisible - Determines stage visibility in PetscLogView()
585: Not Collective
587: Input Parameters:
588: + stage - The stage
589: - isVisible - The visibility flag, PETSC_TRUE to print, else PETSC_FALSE (defaults to PETSC_TRUE)
591: Level: intermediate
593: .seealso: PetscLogStagePush(), PetscLogStagePop(), PetscLogView()
594: @*/
595: PetscErrorCode PetscLogStageSetVisible(PetscLogStage stage, PetscBool isVisible)
596: {
597: PetscStageLog stageLog;
601: PetscLogGetStageLog(&stageLog);
602: PetscStageLogSetVisible(stageLog, stage, isVisible);
603: return(0);
604: }
608: /*@
609: PetscLogStageGetVisible - Returns stage visibility in PetscLogView()
611: Not Collective
613: Input Parameter:
614: . stage - The stage
616: Output Parameter:
617: . isVisible - The visibility flag, PETSC_TRUE to print, else PETSC_FALSE (defaults to PETSC_TRUE)
619: Level: intermediate
621: .seealso: PetscLogStagePush(), PetscLogStagePop(), PetscLogView()
622: @*/
623: PetscErrorCode PetscLogStageGetVisible(PetscLogStage stage, PetscBool *isVisible)
624: {
625: PetscStageLog stageLog;
629: PetscLogGetStageLog(&stageLog);
630: PetscStageLogGetVisible(stageLog, stage, isVisible);
631: return(0);
632: }
636: /*@C
637: PetscLogStageGetId - Returns the stage id when given the stage name.
639: Not Collective
641: Input Parameter:
642: . name - The stage name
644: Output Parameter:
645: . stage - The stage, , or -1 if no stage with that name exists
647: Level: intermediate
649: .seealso: PetscLogStagePush(), PetscLogStagePop(), PetscPreLoadBegin(), PetscPreLoadEnd(), PetscPreLoadStage()
650: @*/
651: PetscErrorCode PetscLogStageGetId(const char name[], PetscLogStage *stage)
652: {
653: PetscStageLog stageLog;
657: PetscLogGetStageLog(&stageLog);
658: PetscStageLogGetStage(stageLog, name, stage);
659: return(0);
660: }
662: /*------------------------------------------------ Event Functions --------------------------------------------------*/
665: /*@C
666: PetscLogEventRegister - Registers an event name for logging operations in an application code.
668: Not Collective
670: Input Parameter:
671: + name - The name associated with the event
672: - classid - The classid associated to the class for this event, obtain either with
673: PetscClassIdRegister() or use a predefined one such as KSP_CLASSID, SNES_CLASSID, the predefined ones
674: are only available in C code
676: Output Parameter:
677: . event - The event id for use with PetscLogEventBegin() and PetscLogEventEnd().
679: Example of Usage:
680: .vb
681: PetscLogEvent USER_EVENT;
682: PetscClassId classid;
683: PetscLogDouble user_event_flops;
684: PetscClassIdRegister("class name",&classid);
685: PetscLogEventRegister("User event name",classid,&USER_EVENT);
686: PetscLogEventBegin(USER_EVENT,0,0,0,0);
687: [code segment to monitor]
688: PetscLogFlops(user_event_flops);
689: PetscLogEventEnd(USER_EVENT,0,0,0,0);
690: .ve
692: Notes:
693: PETSc automatically logs library events if the code has been
694: compiled with -DPETSC_USE_LOG (which is the default) and -log,
695: -log_summary, or -log_all are specified. PetscLogEventRegister() is
696: intended for logging user events to supplement this PETSc
697: information.
699: PETSc can gather data for use with the utilities Jumpshot
700: (part of the MPICH distribution). If PETSc has been compiled
701: with flag -DPETSC_HAVE_MPE (MPE is an additional utility within
702: MPICH), the user can employ another command line option, -log_mpe,
703: to create a logfile, "mpe.log", which can be visualized
704: Jumpshot.
706: The classid is associated with each event so that classes of events
707: can be disabled simultaneously, such as all matrix events. The user
708: can either use an existing classid, such as MAT_CLASSID, or create
709: their own as shown in the example.
711: If an existing event with the same name exists, its event handle is
712: returned instead of creating a new event.
714: Level: intermediate
716: .keywords: log, event, register
717: .seealso: PetscLogEventBegin(), PetscLogEventEnd(), PetscLogFlops(),
718: PetscLogEventMPEActivate(), PetscLogEventMPEDeactivate(),
719: PetscLogEventActivate(), PetscLogEventDeactivate(), PetscClassIdRegister()
720: @*/
721: PetscErrorCode PetscLogEventRegister(const char name[],PetscClassId classid,PetscLogEvent *event)
722: {
723: PetscStageLog stageLog;
724: int stage;
728: *event = PETSC_DECIDE;
729: PetscLogGetStageLog(&stageLog);
730: EventRegLogGetEvent(stageLog->eventLog, name, event);
731: if (*event > 0) return(0);
732: EventRegLogRegister(stageLog->eventLog, name, classid, event);
733: for (stage = 0; stage < stageLog->numStages; stage++) {
734: EventPerfLogEnsureSize(stageLog->stageInfo[stage].eventLog, stageLog->eventLog->numEvents);
735: ClassPerfLogEnsureSize(stageLog->stageInfo[stage].classLog, stageLog->classLog->numClasses);
736: }
737: return(0);
738: }
742: /*@
743: PetscLogEventActivate - Indicates that a particular event should be logged.
745: Not Collective
747: Input Parameter:
748: . event - The event id
750: Usage:
751: .vb
752: PetscLogEventDeactivate(VEC_SetValues);
753: [code where you do not want to log VecSetValues()]
754: PetscLogEventActivate(VEC_SetValues);
755: [code where you do want to log VecSetValues()]
756: .ve
758: Note:
759: The event may be either a pre-defined PETSc event (found in include/petsclog.h)
760: or an event number obtained with PetscLogEventRegister().
762: Level: advanced
764: .keywords: log, event, activate
765: .seealso: PetscLogEventMPEDeactivate(),PetscLogEventMPEActivate(),PlogEventDeactivate()
766: @*/
767: PetscErrorCode PetscLogEventActivate(PetscLogEvent event)
768: {
769: PetscStageLog stageLog;
770: int stage;
774: PetscLogGetStageLog(&stageLog);
775: PetscStageLogGetCurrent(stageLog, &stage);
776: EventPerfLogActivate(stageLog->stageInfo[stage].eventLog, event);
777: return(0);
778: }
782: /*@
783: PetscLogEventDeactivate - Indicates that a particular event should not be logged.
785: Not Collective
787: Input Parameter:
788: . event - The event id
790: Usage:
791: .vb
792: PetscLogEventDeactivate(VEC_SetValues);
793: [code where you do not want to log VecSetValues()]
794: PetscLogEventActivate(VEC_SetValues);
795: [code where you do want to log VecSetValues()]
796: .ve
798: Note:
799: The event may be either a pre-defined PETSc event (found in
800: include/petsclog.h) or an event number obtained with PetscLogEventRegister()).
802: Level: advanced
804: .keywords: log, event, deactivate
805: .seealso: PetscLogEventMPEDeactivate(),PetscLogEventMPEActivate(),PlogEventActivate()
806: @*/
807: PetscErrorCode PetscLogEventDeactivate(PetscLogEvent event)
808: {
809: PetscStageLog stageLog;
810: int stage;
814: PetscLogGetStageLog(&stageLog);
815: PetscStageLogGetCurrent(stageLog, &stage);
816: EventPerfLogDeactivate(stageLog->stageInfo[stage].eventLog, event);
817: return(0);
818: }
822: /*@
823: PetscLogEventSetActiveAll - Sets the event activity in every stage.
825: Not Collective
827: Input Parameters:
828: + event - The event id
829: - isActive - The activity flag determining whether the event is logged
831: Level: advanced
833: .keywords: log, event, activate
834: .seealso: PetscLogEventMPEDeactivate(),PetscLogEventMPEActivate(),PlogEventActivate(),PlogEventDeactivate()
835: @*/
836: PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent event, PetscBool isActive)
837: {
838: PetscStageLog stageLog;
839: int stage;
843: PetscLogGetStageLog(&stageLog);
844: for (stage = 0; stage < stageLog->numStages; stage++) {
845: if (isActive) {
846: EventPerfLogActivate(stageLog->stageInfo[stage].eventLog, event);
847: } else {
848: EventPerfLogDeactivate(stageLog->stageInfo[stage].eventLog, event);
849: }
850: }
851: return(0);
852: }
856: /*@
857: PetscLogEventActivateClass - Activates event logging for a PETSc object class.
859: Not Collective
861: Input Parameter:
862: . classid - The event class, for example MAT_CLASSID, SNES_CLASSID, etc.
864: Level: developer
866: .keywords: log, event, activate, class
867: .seealso: PetscInfoActivate(),PetscInfo(),PetscInfoAllow(),PetscLogEventDeactivateClass(), PetscLogEventActivate(),PetscLogEventDeactivate()
868: @*/
869: PetscErrorCode PetscLogEventActivateClass(PetscClassId classid)
870: {
871: PetscStageLog stageLog;
872: int stage;
876: PetscLogGetStageLog(&stageLog);
877: PetscStageLogGetCurrent(stageLog, &stage);
878: EventPerfLogActivateClass(stageLog->stageInfo[stage].eventLog, stageLog->eventLog, classid);
879: return(0);
880: }
884: /*@
885: PetscLogEventDeactivateClass - Deactivates event logging for a PETSc object class.
887: Not Collective
889: Input Parameter:
890: . classid - The event class, for example MAT_CLASSID, SNES_CLASSID, etc.
892: Level: developer
894: .keywords: log, event, deactivate, class
895: .seealso: PetscInfoActivate(),PetscInfo(),PetscInfoAllow(),PetscLogEventActivateClass(), PetscLogEventActivate(),PetscLogEventDeactivate()
896: @*/
897: PetscErrorCode PetscLogEventDeactivateClass(PetscClassId classid)
898: {
899: PetscStageLog stageLog;
900: int stage;
904: PetscLogGetStageLog(&stageLog);
905: PetscStageLogGetCurrent(stageLog, &stage);
906: EventPerfLogDeactivateClass(stageLog->stageInfo[stage].eventLog, stageLog->eventLog, classid);
907: return(0);
908: }
910: /*MC
911: PetscLogEventBegin - Logs the beginning of a user event.
913: Synopsis:
914: #include <petsclog.h>
915: PetscErrorCode PetscLogEventBegin(int e,PetscObject o1,PetscObject o2,PetscObject o3,PetscObject o4)
917: Not Collective
919: Input Parameters:
920: + e - integer associated with the event obtained from PetscLogEventRegister()
921: - o1,o2,o3,o4 - objects associated with the event, or 0
924: Fortran Synopsis:
925: void PetscLogEventBegin(int e,PetscErrorCode ierr)
927: Usage:
928: .vb
929: PetscLogEvent USER_EVENT;
930: PetscLogDouble user_event_flops;
931: PetscLogEventRegister("User event",0,&USER_EVENT);
932: PetscLogEventBegin(USER_EVENT,0,0,0,0);
933: [code segment to monitor]
934: PetscLogFlops(user_event_flops);
935: PetscLogEventEnd(USER_EVENT,0,0,0,0);
936: .ve
938: Notes:
939: You need to register each integer event with the command
940: PetscLogEventRegister(). The source code must be compiled with
941: -DPETSC_USE_LOG, which is the default.
943: PETSc automatically logs library events if the code has been
944: compiled with -DPETSC_USE_LOG, and -log, -log_summary, or -log_all are
945: specified. PetscLogEventBegin() is intended for logging user events
946: to supplement this PETSc information.
948: Level: intermediate
950: .seealso: PetscLogEventRegister(), PetscLogEventEnd(), PetscLogFlops()
952: .keywords: log, event, begin
953: M*/
955: /*MC
956: PetscLogEventEnd - Log the end of a user event.
958: Synopsis:
959: #include <petsclog.h>
960: PetscErrorCode PetscLogEventEnd(int e,PetscObject o1,PetscObject o2,PetscObject o3,PetscObject o4)
962: Not Collective
964: Input Parameters:
965: + e - integer associated with the event obtained with PetscLogEventRegister()
966: - o1,o2,o3,o4 - objects associated with the event, or 0
969: Fortran Synopsis:
970: void PetscLogEventEnd(int e,PetscErrorCode ierr)
972: Usage:
973: .vb
974: PetscLogEvent USER_EVENT;
975: PetscLogDouble user_event_flops;
976: PetscLogEventRegister("User event",0,&USER_EVENT,);
977: PetscLogEventBegin(USER_EVENT,0,0,0,0);
978: [code segment to monitor]
979: PetscLogFlops(user_event_flops);
980: PetscLogEventEnd(USER_EVENT,0,0,0,0);
981: .ve
983: Notes:
984: You should also register each additional integer event with the command
985: PetscLogEventRegister(). Source code must be compiled with
986: -DPETSC_USE_LOG, which is the default.
988: PETSc automatically logs library events if the code has been
989: compiled with -DPETSC_USE_LOG, and -log, -log_summary, or -log_all are
990: specified. PetscLogEventEnd() is intended for logging user events
991: to supplement this PETSc information.
993: Level: intermediate
995: .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogFlops()
997: .keywords: log, event, end
998: M*/
1000: /*MC
1001: PetscLogEventBarrierBegin - Logs the time in a barrier before an event.
1003: Synopsis:
1004: #include <petsclog.h>
1005: PetscErrorCode PetscLogEventBarrierBegin(int e,PetscObject o1,PetscObject o2,PetscObject o3,PetscObject o4,MPI_Comm comm)
1007: Not Collective
1009: Input Parameters:
1010: . e - integer associated with the event obtained from PetscLogEventRegister()
1011: . o1,o2,o3,o4 - objects associated with the event, or 0
1012: . comm - communicator the barrier takes place over
1015: Usage:
1016: .vb
1017: PetscLogEventBarrierBegin(VEC_NormBarrier,0,0,0,0,comm);
1018: MPI_Allreduce()
1019: PetscLogEventBarrierEnd(VEC_NormBarrier,0,0,0,0,comm);
1020: .ve
1022: Notes:
1023: This is for logging the amount of time spent in a barrier for an event
1024: that requires synchronization.
1026: Additional Notes:
1027: Synchronization events always come in pairs; for example, VEC_NormBarrier and
1028: VEC_NormComm = VEC_NormBarrier + 1
1030: Level: advanced
1032: .seealso: PetscLogEventRegister(), PetscLogEventEnd(), PetscLogFlops(), PetscLogEventBegin(),
1033: PetscLogEventBarrierEnd()
1035: .keywords: log, event, begin, barrier
1036: M*/
1038: /*MC
1039: PetscLogEventBarrierEnd - Logs the time in a barrier before an event.
1041: Synopsis:
1042: #include <petsclog.h>
1043: PetscErrorCode PetscLogEventBarrierEnd(int e,PetscObject o1,PetscObject o2,PetscObject o3,PetscObject o4,MPI_Comm comm)
1045: Logically Collective on MPI_Comm
1047: Input Parameters:
1048: . e - integer associated with the event obtained from PetscLogEventRegister()
1049: . o1,o2,o3,o4 - objects associated with the event, or 0
1050: . comm - communicator the barrier takes place over
1053: Usage:
1054: .vb
1055: PetscLogEventBarrierBegin(VEC_NormBarrier,0,0,0,0,comm);
1056: MPI_Allreduce()
1057: PetscLogEventBarrierEnd(VEC_NormBarrier,0,0,0,0,comm);
1058: .ve
1060: Notes:
1061: This is for logging the amount of time spent in a barrier for an event
1062: that requires synchronization.
1064: Additional Notes:
1065: Synchronization events always come in pairs; for example, VEC_NormBarrier and
1066: VEC_NormComm = VEC_NormBarrier + 1
1068: Level: advanced
1070: .seealso: PetscLogEventRegister(), PetscLogEventEnd(), PetscLogFlops(), PetscLogEventBegin(),
1071: PetscLogEventBarrierBegin()
1073: .keywords: log, event, begin, barrier
1074: M*/
1078: /*@C
1079: PetscLogEventGetId - Returns the event id when given the event name.
1081: Not Collective
1083: Input Parameter:
1084: . name - The event name
1086: Output Parameter:
1087: . event - The event, or -1 if no event with that name exists
1089: Level: intermediate
1091: .seealso: PetscLogEventBegin(), PetscLogEventEnd(), PetscLogStageGetId()
1092: @*/
1093: PetscErrorCode PetscLogEventGetId(const char name[], PetscLogEvent *event)
1094: {
1095: PetscStageLog stageLog;
1099: PetscLogGetStageLog(&stageLog);
1100: EventRegLogGetEvent(stageLog->eventLog, name, event);
1101: return(0);
1102: }
1105: /*------------------------------------------------ Output Functions -------------------------------------------------*/
1108: /*@C
1109: PetscLogDump - Dumps logs of objects to a file. This file is intended to
1110: be read by bin/petscview. This program no longer exists.
1112: Collective on PETSC_COMM_WORLD
1114: Input Parameter:
1115: . name - an optional file name
1117: Options Database Keys:
1118: + -log - Prints basic log information (for code compiled with PETSC_USE_LOG)
1119: - -log_all - Prints extensive log information (for code compiled with PETSC_USE_LOG)
1121: Usage:
1122: .vb
1123: PetscInitialize(...);
1124: PetscLogBegin(); or PetscLogAllBegin();
1125: ... code ...
1126: PetscLogDump(filename);
1127: PetscFinalize();
1128: .ve
1130: Notes:
1131: The default file name is
1132: $ Log.<rank>
1133: where <rank> is the processor number. If no name is specified,
1134: this file will be used.
1136: Level: advanced
1138: .keywords: log, dump
1139: .seealso: PetscLogBegin(), PetscLogAllBegin(), PetscLogView()
1140: @*/
1141: PetscErrorCode PetscLogDump(const char sname[])
1142: {
1143: PetscStageLog stageLog;
1144: PetscEventPerfInfo *eventInfo;
1145: FILE *fd;
1146: char file[PETSC_MAX_PATH_LEN], fname[PETSC_MAX_PATH_LEN];
1147: PetscLogDouble flops, _TotalTime;
1148: PetscMPIInt rank;
1149: int action, object, curStage;
1150: PetscLogEvent event;
1151: PetscErrorCode ierr;
1154: /* Calculate the total elapsed time */
1155: PetscTime(&_TotalTime);
1156: _TotalTime -= petsc_BaseTime;
1157: /* Open log file */
1158: MPI_Comm_rank(PETSC_COMM_WORLD, &rank);
1159: if (sname) sprintf(file, "%s.%d", sname, rank);
1160: else sprintf(file, "Log.%d", rank);
1161: PetscFixFilename(file, fname);
1162: PetscFOpen(PETSC_COMM_WORLD, fname, "w", &fd);
1163: if ((!rank) && (!fd)) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_FILE_OPEN, "Cannot open file: %s", fname);
1164: /* Output totals */
1165: PetscFPrintf(PETSC_COMM_WORLD, fd, "Total Flops %14e %16.8e\n", petsc_TotalFlops, _TotalTime);
1166: PetscFPrintf(PETSC_COMM_WORLD, fd, "Clock Resolution %g\n", 0.0);
1167: /* Output actions */
1168: if (petsc_logActions) {
1169: PetscFPrintf(PETSC_COMM_WORLD, fd, "Actions accomplished %d\n", petsc_numActions);
1170: for (action = 0; action < petsc_numActions; action++) {
1171: PetscFPrintf(PETSC_COMM_WORLD, fd, "%g %d %d %d %d %d %d %g %g %g\n",
1172: petsc_actions[action].time, petsc_actions[action].action, (int)petsc_actions[action].event, (int)petsc_actions[action].classid, petsc_actions[action].id1,
1173: petsc_actions[action].id2, petsc_actions[action].id3, petsc_actions[action].flops, petsc_actions[action].mem, petsc_actions[action].maxmem);
1174: }
1175: }
1176: /* Output objects */
1177: if (petsc_logObjects) {
1178: PetscFPrintf(PETSC_COMM_WORLD, fd, "Objects created %d destroyed %d\n", petsc_numObjects, petsc_numObjectsDestroyed);
1179: for (object = 0; object < petsc_numObjects; object++) {
1180: PetscFPrintf(PETSC_COMM_WORLD, fd, "Parent ID: %d Memory: %d\n", petsc_objects[object].parent, (int) petsc_objects[object].mem);
1181: if (!petsc_objects[object].name[0]) {
1182: PetscFPrintf(PETSC_COMM_WORLD, fd,"No Name\n");
1183: } else {
1184: PetscFPrintf(PETSC_COMM_WORLD, fd, "Name: %s\n", petsc_objects[object].name);
1185: }
1186: if (petsc_objects[object].info[0] != 0) {
1187: PetscFPrintf(PETSC_COMM_WORLD, fd, "No Info\n");
1188: } else {
1189: PetscFPrintf(PETSC_COMM_WORLD, fd, "Info: %s\n", petsc_objects[object].info);
1190: }
1191: }
1192: }
1193: /* Output events */
1194: PetscFPrintf(PETSC_COMM_WORLD, fd, "Event log:\n");
1195: PetscLogGetStageLog(&stageLog);
1196: PetscIntStackTop(stageLog->stack, &curStage);
1197: eventInfo = stageLog->stageInfo[curStage].eventLog->eventInfo;
1198: for (event = 0; event < stageLog->stageInfo[curStage].eventLog->numEvents; event++) {
1199: if (eventInfo[event].time != 0.0) flops = eventInfo[event].flops/eventInfo[event].time;
1200: else flops = 0.0;
1201: PetscFPrintf(PETSC_COMM_WORLD, fd, "%d %16d %16g %16g %16g\n", event, eventInfo[event].count,
1202: eventInfo[event].flops, eventInfo[event].time, flops);
1203: }
1204: PetscFClose(PETSC_COMM_WORLD, fd);
1205: return(0);
1206: }
1210: /*
1211: PetscLogView_Detailed - Each process prints the times for its own events
1213: */
1214: PetscErrorCode PetscLogView_Detailed(PetscViewer viewer)
1215: {
1216: MPI_Comm comm = PetscObjectComm((PetscObject) viewer);
1217: PetscEventPerfInfo *eventInfo = NULL;
1218: PetscLogDouble locTotalTime, numRed, maxMem;
1219: PetscStageLog stageLog;
1220: int numStages,numEvents,stage,event;
1221: PetscMPIInt rank,size;
1222: PetscErrorCode ierr;
1225: MPI_Comm_size(comm, &size);
1226: MPI_Comm_rank(comm, &rank);
1227: /* Must preserve reduction count before we go on */
1228: numRed = petsc_allreduce_ct + petsc_gather_ct + petsc_scatter_ct;
1229: /* Get the total elapsed time */
1230: PetscTime(&locTotalTime); locTotalTime -= petsc_BaseTime;
1231: PetscViewerASCIIPrintf(viewer,"numProcs = %d\n",size);
1232: PetscViewerASCIIPrintf(viewer,"LocalTimes = {}\n");
1233: PetscViewerASCIIPrintf(viewer,"LocalFlops = {}\n");
1234: PetscViewerASCIIPrintf(viewer,"LocalMessageLens = {}\n");
1235: PetscViewerASCIIPrintf(viewer,"LocalMessages = {}\n");
1236: PetscViewerASCIIPrintf(viewer,"LocalReductions = {}\n");
1237: PetscViewerASCIIPrintf(viewer,"LocalObjects = {}\n");
1238: PetscViewerASCIIPrintf(viewer,"LocalMemory = {}\n");
1239: PetscLogGetStageLog(&stageLog);
1240: MPI_Allreduce(&stageLog->numStages, &numStages, 1, MPI_INT, MPI_MAX, comm);
1241: PetscViewerASCIIPrintf(viewer,"Stages = {}\n");
1242: for (stage=0; stage<numStages; stage++) {
1243: PetscViewerASCIIPrintf(viewer,"Stages[\"%s\"] = {}\n",stageLog->stageInfo[stage].name);
1244: PetscViewerASCIIPrintf(viewer,"Stages[\"%s\"][\"summary\"] = {}\n",stageLog->stageInfo[stage].name);
1245: MPI_Allreduce(&stageLog->stageInfo[stage].eventLog->numEvents, &numEvents, 1, MPI_INT, MPI_MAX, comm);
1246: for (event = 0; event < numEvents; event++) {
1247: PetscViewerASCIIPrintf(viewer,"Stages[\"%s\"][\"%s\"] = {}\n",stageLog->stageInfo[stage].name,stageLog->eventLog->eventInfo[event].name);
1248: }
1249: }
1250: PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
1251: PetscViewerASCIISynchronizedPrintf(viewer,"LocalTimes[%d] = %g\n",rank,locTotalTime);
1252: PetscViewerASCIISynchronizedPrintf(viewer,"LocalFlops[%d] = %g\n",rank,petsc_TotalFlops);
1253: PetscViewerASCIISynchronizedPrintf(viewer,"LocalMessageLens[%d] = %g\n",rank,(petsc_irecv_len + petsc_isend_len + petsc_recv_len + petsc_send_len));
1254: PetscViewerASCIISynchronizedPrintf(viewer,"LocalMessages[%d] = %g\n",rank,(petsc_irecv_ct + petsc_isend_ct + petsc_recv_ct + petsc_send_ct));
1255: PetscViewerASCIISynchronizedPrintf(viewer,"LocalReductions[%d] = %g\n",rank,numRed);
1256: PetscViewerASCIISynchronizedPrintf(viewer,"LocalObjects[%d] = %g\n",rank,petsc_numObjects);
1257: PetscMallocGetMaximumUsage(&maxMem);
1258: PetscViewerASCIISynchronizedPrintf(viewer,"LocalMemory[%d] = %g\n",rank,maxMem);
1259: PetscViewerFlush(viewer);
1260: for (stage=0; stage<numStages; stage++) {
1261: PetscViewerASCIISynchronizedPrintf(viewer,"Stages[\"%s\"][\"summary\"][%d] = {\"time\" : %g, \"numMessages\" : %g, \"messageLength\" : %g, \"numReductions\" : %g, \"flops\" : %g}\n",
1262: stageLog->stageInfo[stage].name,rank,
1263: stageLog->stageInfo[stage].perfInfo.time,stageLog->stageInfo[stage].perfInfo.numMessages,stageLog->stageInfo[stage].perfInfo.messageLength,
1264: stageLog->stageInfo[stage].perfInfo.numReductions,stageLog->stageInfo[stage].perfInfo.flops);
1265: MPI_Allreduce(&stageLog->stageInfo[stage].eventLog->numEvents, &numEvents, 1, MPI_INT, MPI_MAX, comm);
1266: for (event = 0; event < numEvents; event++) {
1267: eventInfo = stageLog->stageInfo[stage].eventLog->eventInfo;
1268: PetscViewerASCIISynchronizedPrintf(viewer,"Stages[\"%s\"][\"%s\"][%d] = {\"count\" : %D, \"time\" : %g, \"numMessages\" : %g, \"messageLength\" : %g, \"numReductions\" : %g, \"flops\" : %g}\n",stageLog->stageInfo[stage].name,stageLog->eventLog->eventInfo[event].name,rank,
1269: eventInfo[event].count, eventInfo[event].time,eventInfo[event].numMessages, eventInfo[event].messageLength,
1270: eventInfo[event].numReductions,eventInfo[event].flops);
1271: }
1272: }
1273: PetscViewerFlush(viewer);
1274: PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
1275: return(0);
1276: }
1280: PetscErrorCode PetscLogView_Default(PetscViewer viewer)
1281: {
1282: FILE *fd;
1283: PetscLogDouble zero = 0.0;
1284: PetscStageLog stageLog;
1285: PetscStageInfo *stageInfo = NULL;
1286: PetscEventPerfInfo *eventInfo = NULL;
1287: PetscClassPerfInfo *classInfo;
1288: char arch[128],hostname[128],username[128],pname[PETSC_MAX_PATH_LEN],date[128];
1289: const char *name;
1290: PetscLogDouble locTotalTime, TotalTime, TotalFlops;
1291: PetscLogDouble numMessages, messageLength, avgMessLen, numReductions;
1292: PetscLogDouble stageTime, flops, flopr, mem, mess, messLen, red;
1293: PetscLogDouble fracTime, fracFlops, fracMessages, fracLength, fracReductions, fracMess, fracMessLen, fracRed;
1294: PetscLogDouble fracStageTime, fracStageFlops, fracStageMess, fracStageMessLen, fracStageRed;
1295: PetscLogDouble min, max, tot, ratio, avg, x, y;
1296: PetscLogDouble minf, maxf, totf, ratf, mint, maxt, tott, ratt, ratCt, totm, totml, totr;
1297: PetscMPIInt minCt, maxCt;
1298: PetscMPIInt size, rank;
1299: PetscBool *localStageUsed, *stageUsed;
1300: PetscBool *localStageVisible, *stageVisible;
1301: int numStages, localNumEvents, numEvents;
1302: int stage, oclass;
1303: PetscLogEvent event;
1304: PetscErrorCode ierr;
1305: char version[256];
1306: MPI_Comm comm;
1309: PetscObjectGetComm((PetscObject)viewer,&comm);
1310: PetscViewerASCIIGetPointer(viewer,&fd);
1311: MPI_Comm_size(comm, &size);
1312: MPI_Comm_rank(comm, &rank);
1313: /* Get the total elapsed time */
1314: PetscTime(&locTotalTime); locTotalTime -= petsc_BaseTime;
1316: PetscFPrintf(comm, fd, "************************************************************************************************************************\n");
1317: PetscFPrintf(comm, fd, "*** WIDEN YOUR WINDOW TO 120 CHARACTERS. Use 'enscript -r -fCourier9' to print this document ***\n");
1318: PetscFPrintf(comm, fd, "************************************************************************************************************************\n");
1319: PetscFPrintf(comm, fd, "\n---------------------------------------------- PETSc Performance Summary: ----------------------------------------------\n\n");
1320: PetscGetArchType(arch,sizeof(arch));
1321: PetscGetHostName(hostname,sizeof(hostname));
1322: PetscGetUserName(username,sizeof(username));
1323: PetscGetProgramName(pname,sizeof(pname));
1324: PetscGetDate(date,sizeof(date));
1325: PetscGetVersion(version,sizeof(version));
1326: if (size == 1) {
1327: PetscFPrintf(comm,fd,"%s on a %s named %s with %d processor, by %s %s\n", pname, arch, hostname, size, username, date);
1328: } else {
1329: PetscFPrintf(comm,fd,"%s on a %s named %s with %d processors, by %s %s\n", pname, arch, hostname, size, username, date);
1330: }
1332: PetscFPrintf(comm, fd, "Using %s\n", version);
1334: /* Must preserve reduction count before we go on */
1335: red = petsc_allreduce_ct + petsc_gather_ct + petsc_scatter_ct;
1337: /* Calculate summary information */
1338: PetscFPrintf(comm, fd, "\n Max Max/Min Avg Total \n");
1339: /* Time */
1340: MPI_Allreduce(&locTotalTime, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1341: MPI_Allreduce(&locTotalTime, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1342: MPI_Allreduce(&locTotalTime, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1343: avg = (tot)/((PetscLogDouble) size);
1344: if (min != 0.0) ratio = max/min;
1345: else ratio = 0.0;
1346: PetscFPrintf(comm, fd, "Time (sec): %5.3e %10.5f %5.3e\n", max, ratio, avg);
1347: TotalTime = tot;
1348: /* Objects */
1349: avg = (PetscLogDouble) petsc_numObjects;
1350: MPI_Allreduce(&avg, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1351: MPI_Allreduce(&avg, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1352: MPI_Allreduce(&avg, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1353: avg = (tot)/((PetscLogDouble) size);
1354: if (min != 0.0) ratio = max/min;
1355: else ratio = 0.0;
1356: PetscFPrintf(comm, fd, "Objects: %5.3e %10.5f %5.3e\n", max, ratio, avg);
1357: /* Flops */
1358: MPI_Allreduce(&petsc_TotalFlops, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1359: MPI_Allreduce(&petsc_TotalFlops, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1360: MPI_Allreduce(&petsc_TotalFlops, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1361: avg = (tot)/((PetscLogDouble) size);
1362: if (min != 0.0) ratio = max/min;
1363: else ratio = 0.0;
1364: PetscFPrintf(comm, fd, "Flops: %5.3e %10.5f %5.3e %5.3e\n", max, ratio, avg, tot);
1365: TotalFlops = tot;
1366: /* Flops/sec -- Must talk to Barry here */
1367: if (locTotalTime != 0.0) flops = petsc_TotalFlops/locTotalTime;
1368: else flops = 0.0;
1369: MPI_Allreduce(&flops, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1370: MPI_Allreduce(&flops, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1371: MPI_Allreduce(&flops, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1372: avg = (tot)/((PetscLogDouble) size);
1373: if (min != 0.0) ratio = max/min;
1374: else ratio = 0.0;
1375: PetscFPrintf(comm, fd, "Flops/sec: %5.3e %10.5f %5.3e %5.3e\n", max, ratio, avg, tot);
1376: /* Memory */
1377: PetscMallocGetMaximumUsage(&mem);
1378: if (mem > 0.0) {
1379: MPI_Allreduce(&mem, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1380: MPI_Allreduce(&mem, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1381: MPI_Allreduce(&mem, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1382: avg = (tot)/((PetscLogDouble) size);
1383: if (min != 0.0) ratio = max/min;
1384: else ratio = 0.0;
1385: PetscFPrintf(comm, fd, "Memory: %5.3e %10.5f %5.3e\n", max, ratio, tot);
1386: }
1387: /* Messages */
1388: mess = 0.5*(petsc_irecv_ct + petsc_isend_ct + petsc_recv_ct + petsc_send_ct);
1389: MPI_Allreduce(&mess, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1390: MPI_Allreduce(&mess, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1391: MPI_Allreduce(&mess, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1392: avg = (tot)/((PetscLogDouble) size);
1393: if (min != 0.0) ratio = max/min;
1394: else ratio = 0.0;
1395: PetscFPrintf(comm, fd, "MPI Messages: %5.3e %10.5f %5.3e %5.3e\n", max, ratio, avg, tot);
1396: numMessages = tot;
1397: /* Message Lengths */
1398: mess = 0.5*(petsc_irecv_len + petsc_isend_len + petsc_recv_len + petsc_send_len);
1399: MPI_Allreduce(&mess, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1400: MPI_Allreduce(&mess, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1401: MPI_Allreduce(&mess, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1402: if (numMessages != 0) avg = (tot)/(numMessages);
1403: else avg = 0.0;
1404: if (min != 0.0) ratio = max/min;
1405: else ratio = 0.0;
1406: PetscFPrintf(comm, fd, "MPI Message Lengths: %5.3e %10.5f %5.3e %5.3e\n", max, ratio, avg, tot);
1407: messageLength = tot;
1408: /* Reductions */
1409: MPI_Allreduce(&red, &min, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1410: MPI_Allreduce(&red, &max, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1411: MPI_Allreduce(&red, &tot, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1412: if (min != 0.0) ratio = max/min;
1413: else ratio = 0.0;
1414: PetscFPrintf(comm, fd, "MPI Reductions: %5.3e %10.5f\n", max, ratio);
1415: numReductions = red; /* wrong because uses count from process zero */
1416: PetscFPrintf(comm, fd, "\nFlop counting convention: 1 flop = 1 real number operation of type (multiply/divide/add/subtract)\n");
1417: PetscFPrintf(comm, fd, " e.g., VecAXPY() for real vectors of length N --> 2N flops\n");
1418: PetscFPrintf(comm, fd, " and VecAXPY() for complex vectors of length N --> 8N flops\n");
1420: /* Get total number of stages --
1421: Currently, a single processor can register more stages than another, but stages must all be registered in order.
1422: We can removed this requirement if necessary by having a global stage numbering and indirection on the stage ID.
1423: This seems best accomplished by assoicating a communicator with each stage.
1424: */
1425: PetscLogGetStageLog(&stageLog);
1426: MPI_Allreduce(&stageLog->numStages, &numStages, 1, MPI_INT, MPI_MAX, comm);
1427: PetscMalloc1(numStages, &localStageUsed);
1428: PetscMalloc1(numStages, &stageUsed);
1429: PetscMalloc1(numStages, &localStageVisible);
1430: PetscMalloc1(numStages, &stageVisible);
1431: if (numStages > 0) {
1432: stageInfo = stageLog->stageInfo;
1433: for (stage = 0; stage < numStages; stage++) {
1434: if (stage < stageLog->numStages) {
1435: localStageUsed[stage] = stageInfo[stage].used;
1436: localStageVisible[stage] = stageInfo[stage].perfInfo.visible;
1437: } else {
1438: localStageUsed[stage] = PETSC_FALSE;
1439: localStageVisible[stage] = PETSC_TRUE;
1440: }
1441: }
1442: MPI_Allreduce(localStageUsed, stageUsed, numStages, MPIU_BOOL, MPI_LOR, comm);
1443: MPI_Allreduce(localStageVisible, stageVisible, numStages, MPIU_BOOL, MPI_LAND, comm);
1444: for (stage = 0; stage < numStages; stage++) {
1445: if (stageUsed[stage]) {
1446: PetscFPrintf(comm, fd, "\nSummary of Stages: ----- Time ------ ----- Flops ----- --- Messages --- -- Message Lengths -- -- Reductions --\n");
1447: PetscFPrintf(comm, fd, " Avg %%Total Avg %%Total counts %%Total Avg %%Total counts %%Total \n");
1448: break;
1449: }
1450: }
1451: for (stage = 0; stage < numStages; stage++) {
1452: if (!stageUsed[stage]) continue;
1453: if (localStageUsed[stage]) {
1454: MPI_Allreduce(&stageInfo[stage].perfInfo.time, &stageTime, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1455: MPI_Allreduce(&stageInfo[stage].perfInfo.flops, &flops, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1456: MPI_Allreduce(&stageInfo[stage].perfInfo.numMessages, &mess, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1457: MPI_Allreduce(&stageInfo[stage].perfInfo.messageLength, &messLen, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1458: MPI_Allreduce(&stageInfo[stage].perfInfo.numReductions, &red, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1459: name = stageInfo[stage].name;
1460: } else {
1461: MPI_Allreduce(&zero, &stageTime, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1462: MPI_Allreduce(&zero, &flops, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1463: MPI_Allreduce(&zero, &mess, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1464: MPI_Allreduce(&zero, &messLen, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1465: MPI_Allreduce(&zero, &red, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1466: name = "";
1467: }
1468: mess *= 0.5; messLen *= 0.5; red /= size;
1469: if (TotalTime != 0.0) fracTime = stageTime/TotalTime; else fracTime = 0.0;
1470: if (TotalFlops != 0.0) fracFlops = flops/TotalFlops; else fracFlops = 0.0;
1471: /* Talk to Barry if (stageTime != 0.0) flops = (size*flops)/stageTime; else flops = 0.0; */
1472: if (numMessages != 0.0) fracMessages = mess/numMessages; else fracMessages = 0.0;
1473: if (numMessages != 0.0) avgMessLen = messLen/numMessages; else avgMessLen = 0.0;
1474: if (messageLength != 0.0) fracLength = messLen/messageLength; else fracLength = 0.0;
1475: if (numReductions != 0.0) fracReductions = red/numReductions; else fracReductions = 0.0;
1476: PetscFPrintf(comm, fd, "%2d: %15s: %6.4e %5.1f%% %6.4e %5.1f%% %5.3e %5.1f%% %5.3e %5.1f%% %5.3e %5.1f%% \n",
1477: stage, name, stageTime/size, 100.0*fracTime, flops, 100.0*fracFlops,
1478: mess, 100.0*fracMessages, avgMessLen, 100.0*fracLength, red, 100.0*fracReductions);
1479: }
1480: }
1482: PetscFPrintf(comm, fd,"\n------------------------------------------------------------------------------------------------------------------------\n");
1483: PetscFPrintf(comm, fd, "See the 'Profiling' chapter of the users' manual for details on interpreting output.\n");
1484: PetscFPrintf(comm, fd, "Phase summary info:\n");
1485: PetscFPrintf(comm, fd, " Count: number of times phase was executed\n");
1486: PetscFPrintf(comm, fd, " Time and Flops: Max - maximum over all processors\n");
1487: PetscFPrintf(comm, fd, " Ratio - ratio of maximum to minimum over all processors\n");
1488: PetscFPrintf(comm, fd, " Mess: number of messages sent\n");
1489: PetscFPrintf(comm, fd, " Avg. len: average message length (bytes)\n");
1490: PetscFPrintf(comm, fd, " Reduct: number of global reductions\n");
1491: PetscFPrintf(comm, fd, " Global: entire computation\n");
1492: PetscFPrintf(comm, fd, " Stage: stages of a computation. Set stages with PetscLogStagePush() and PetscLogStagePop().\n");
1493: PetscFPrintf(comm, fd, " %%T - percent time in this phase %%F - percent flops in this phase\n");
1494: PetscFPrintf(comm, fd, " %%M - percent messages in this phase %%L - percent message lengths in this phase\n");
1495: PetscFPrintf(comm, fd, " %%R - percent reductions in this phase\n");
1496: PetscFPrintf(comm, fd, " Total Mflop/s: 10e-6 * (sum of flops over all processors)/(max time over all processors)\n");
1497: PetscFPrintf(comm, fd, "------------------------------------------------------------------------------------------------------------------------\n");
1499: #if defined(PETSC_USE_DEBUG)
1500: PetscFPrintf(comm, fd, "\n\n");
1501: PetscFPrintf(comm, fd, " ##########################################################\n");
1502: PetscFPrintf(comm, fd, " # #\n");
1503: PetscFPrintf(comm, fd, " # WARNING!!! #\n");
1504: PetscFPrintf(comm, fd, " # #\n");
1505: PetscFPrintf(comm, fd, " # This code was compiled with a debugging option, #\n");
1506: PetscFPrintf(comm, fd, " # To get timing results run ./configure #\n");
1507: PetscFPrintf(comm, fd, " # using --with-debugging=no, the performance will #\n");
1508: PetscFPrintf(comm, fd, " # be generally two or three times faster. #\n");
1509: PetscFPrintf(comm, fd, " # #\n");
1510: PetscFPrintf(comm, fd, " ##########################################################\n\n\n");
1511: #endif
1512: #if defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_FORTRAN_KERNELS)
1513: PetscFPrintf(comm, fd, "\n\n");
1514: PetscFPrintf(comm, fd, " ##########################################################\n");
1515: PetscFPrintf(comm, fd, " # #\n");
1516: PetscFPrintf(comm, fd, " # WARNING!!! #\n");
1517: PetscFPrintf(comm, fd, " # #\n");
1518: PetscFPrintf(comm, fd, " # The code for various complex numbers numerical #\n");
1519: PetscFPrintf(comm, fd, " # kernels uses C++, which generally is not well #\n");
1520: PetscFPrintf(comm, fd, " # optimized. For performance that is about 4-5 times #\n");
1521: PetscFPrintf(comm, fd, " # faster, specify --with-fortran-kernels=1 #\n");
1522: PetscFPrintf(comm, fd, " # when running ./configure.py. #\n");
1523: PetscFPrintf(comm, fd, " # #\n");
1524: PetscFPrintf(comm, fd, " ##########################################################\n\n\n");
1525: #endif
1527: /* Report events */
1528: PetscFPrintf(comm, fd,"Event Count Time (sec) Flops --- Global --- --- Stage --- Total\n");
1529: PetscFPrintf(comm, fd," Max Ratio Max Ratio Max Ratio Mess Avg len Reduct %%T %%F %%M %%L %%R %%T %%F %%M %%L %%R Mflop/s\n");
1530: PetscFPrintf(comm,fd,"------------------------------------------------------------------------------------------------------------------------\n");
1532: /* Problem: The stage name will not show up unless the stage executed on proc 1 */
1533: for (stage = 0; stage < numStages; stage++) {
1534: if (!stageVisible[stage]) continue;
1535: if (localStageUsed[stage]) {
1536: PetscFPrintf(comm, fd, "\n--- Event Stage %d: %s\n\n", stage, stageInfo[stage].name);
1537: MPI_Allreduce(&stageInfo[stage].perfInfo.time, &stageTime, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1538: MPI_Allreduce(&stageInfo[stage].perfInfo.flops, &flops, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1539: MPI_Allreduce(&stageInfo[stage].perfInfo.numMessages, &mess, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1540: MPI_Allreduce(&stageInfo[stage].perfInfo.messageLength, &messLen, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1541: MPI_Allreduce(&stageInfo[stage].perfInfo.numReductions, &red, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1542: } else {
1543: PetscFPrintf(comm, fd, "\n--- Event Stage %d: Unknown\n\n", stage);
1544: MPI_Allreduce(&zero, &stageTime, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1545: MPI_Allreduce(&zero, &flops, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1546: MPI_Allreduce(&zero, &mess, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1547: MPI_Allreduce(&zero, &messLen, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1548: MPI_Allreduce(&zero, &red, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1549: }
1550: mess *= 0.5; messLen *= 0.5; red /= size;
1552: /* Get total number of events in this stage --
1553: Currently, a single processor can register more events than another, but events must all be registered in order,
1554: just like stages. We can removed this requirement if necessary by having a global event numbering and indirection
1555: on the event ID. This seems best accomplished by assoicating a communicator with each stage.
1557: Problem: If the event did not happen on proc 1, its name will not be available.
1558: Problem: Event visibility is not implemented
1559: */
1560: if (localStageUsed[stage]) {
1561: eventInfo = stageLog->stageInfo[stage].eventLog->eventInfo;
1562: localNumEvents = stageLog->stageInfo[stage].eventLog->numEvents;
1563: } else localNumEvents = 0;
1564: MPI_Allreduce(&localNumEvents, &numEvents, 1, MPI_INT, MPI_MAX, comm);
1565: for (event = 0; event < numEvents; event++) {
1566: if (localStageUsed[stage] && (event < stageLog->stageInfo[stage].eventLog->numEvents) && (eventInfo[event].depth == 0)) {
1567: if ((eventInfo[event].count > 0) && (eventInfo[event].time > 0.0)) flopr = eventInfo[event].flops;
1568: else flopr = 0.0;
1570: MPI_Allreduce(&flopr, &minf, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1571: MPI_Allreduce(&flopr, &maxf, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1572: MPI_Allreduce(&eventInfo[event].flops, &totf, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1573: MPI_Allreduce(&eventInfo[event].time, &mint, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1574: MPI_Allreduce(&eventInfo[event].time, &maxt, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1575: MPI_Allreduce(&eventInfo[event].time, &tott, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1576: MPI_Allreduce(&eventInfo[event].numMessages, &totm, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1577: MPI_Allreduce(&eventInfo[event].messageLength, &totml, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1578: MPI_Allreduce(&eventInfo[event].numReductions, &totr, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1579: MPI_Allreduce(&eventInfo[event].count, &minCt, 1, MPI_INT, MPI_MIN, comm);
1580: MPI_Allreduce(&eventInfo[event].count, &maxCt, 1, MPI_INT, MPI_MAX, comm);
1581: name = stageLog->eventLog->eventInfo[event].name;
1582: } else {
1583: flopr = 0.0;
1584: MPI_Allreduce(&flopr, &minf, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1585: MPI_Allreduce(&flopr, &maxf, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1586: MPI_Allreduce(&zero, &totf, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1587: MPI_Allreduce(&zero, &mint, 1, MPIU_PETSCLOGDOUBLE, MPI_MIN, comm);
1588: MPI_Allreduce(&zero, &maxt, 1, MPIU_PETSCLOGDOUBLE, MPI_MAX, comm);
1589: MPI_Allreduce(&zero, &tott, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1590: MPI_Allreduce(&zero, &totm, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1591: MPI_Allreduce(&zero, &totml, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1592: MPI_Allreduce(&zero, &totr, 1, MPIU_PETSCLOGDOUBLE, MPI_SUM, comm);
1593: MPI_Allreduce(&ierr, &minCt, 1, MPI_INT, MPI_MIN, comm);
1594: MPI_Allreduce(&ierr, &maxCt, 1, MPI_INT, MPI_MAX, comm);
1595: name = "";
1596: }
1597: if (mint < 0.0) {
1598: PetscFPrintf(comm, fd, "WARNING!!! Minimum time %g over all processors for %s is negative! This happens\n on some machines whose times cannot handle too rapid calls.!\n artificially changing minimum to zero.\n",mint,name);
1599: mint = 0;
1600: }
1601: if (minf < 0.0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Minimum flops %g over all processors for %s is negative! Not possible!",minf,name);
1602: totm *= 0.5; totml *= 0.5; totr /= size;
1604: if (maxCt != 0) {
1605: if (minCt != 0) ratCt = ((PetscLogDouble) maxCt)/minCt; else ratCt = 0.0;
1606: if (mint != 0.0) ratt = maxt/mint; else ratt = 0.0;
1607: if (minf != 0.0) ratf = maxf/minf; else ratf = 0.0;
1608: if (TotalTime != 0.0) fracTime = tott/TotalTime; else fracTime = 0.0;
1609: if (TotalFlops != 0.0) fracFlops = totf/TotalFlops; else fracFlops = 0.0;
1610: if (stageTime != 0.0) fracStageTime = tott/stageTime; else fracStageTime = 0.0;
1611: if (flops != 0.0) fracStageFlops = totf/flops; else fracStageFlops = 0.0;
1612: if (numMessages != 0.0) fracMess = totm/numMessages; else fracMess = 0.0;
1613: if (messageLength != 0.0) fracMessLen = totml/messageLength; else fracMessLen = 0.0;
1614: if (numReductions != 0.0) fracRed = totr/numReductions; else fracRed = 0.0;
1615: if (mess != 0.0) fracStageMess = totm/mess; else fracStageMess = 0.0;
1616: if (messLen != 0.0) fracStageMessLen = totml/messLen; else fracStageMessLen = 0.0;
1617: if (red != 0.0) fracStageRed = totr/red; else fracStageRed = 0.0;
1618: if (totm != 0.0) totml /= totm; else totml = 0.0;
1619: if (maxt != 0.0) flopr = totf/maxt; else flopr = 0.0;
1620: if (fracStageTime > 1.00) PetscFPrintf(comm, fd,"Warning -- total time of even greater than time of entire stage -- something is wrong with the timer\n");
1621: PetscFPrintf(comm, fd,
1622: "%-16s %7d%4.1f %5.4e%4.1f %3.2e%4.1f %2.1e %2.1e %2.1e%3.0f%3.0f%3.0f%3.0f%3.0f %3.0f%3.0f%3.0f%3.0f%3.0f %5.0f\n",
1623: name, maxCt, ratCt, maxt, ratt, maxf, ratf, totm, totml, totr,
1624: 100.0*fracTime, 100.0*fracFlops, 100.0*fracMess, 100.0*fracMessLen, 100.0*fracRed,
1625: 100.0*fracStageTime, 100.0*fracStageFlops, 100.0*fracStageMess, 100.0*fracStageMessLen, 100.0*fracStageRed,
1626: PetscAbsReal(flopr/1.0e6));
1627: }
1628: }
1629: }
1631: /* Memory usage and object creation */
1632: PetscFPrintf(comm, fd, "------------------------------------------------------------------------------------------------------------------------\n");
1633: PetscFPrintf(comm, fd, "\n");
1634: PetscFPrintf(comm, fd, "Memory usage is given in bytes:\n\n");
1636: /* Right now, only stages on the first processor are reported here, meaning only objects associated with
1637: the global communicator, or MPI_COMM_SELF for proc 1. We really should report global stats and then
1638: stats for stages local to processor sets.
1639: */
1640: /* We should figure out the longest object name here (now 20 characters) */
1641: PetscFPrintf(comm, fd, "Object Type Creations Destructions Memory Descendants' Mem.\n");
1642: PetscFPrintf(comm, fd, "Reports information only for process 0.\n");
1643: for (stage = 0; stage < numStages; stage++) {
1644: if (localStageUsed[stage]) {
1645: classInfo = stageLog->stageInfo[stage].classLog->classInfo;
1646: PetscFPrintf(comm, fd, "\n--- Event Stage %d: %s\n\n", stage, stageInfo[stage].name);
1647: for (oclass = 0; oclass < stageLog->stageInfo[stage].classLog->numClasses; oclass++) {
1648: if ((classInfo[oclass].creations > 0) || (classInfo[oclass].destructions > 0)) {
1649: PetscFPrintf(comm, fd, "%20s %5d %5d %11.0f %g\n", stageLog->classLog->classInfo[oclass].name,
1650: classInfo[oclass].creations, classInfo[oclass].destructions, classInfo[oclass].mem,
1651: classInfo[oclass].descMem);
1652: }
1653: }
1654: } else {
1655: PetscFPrintf(comm, fd, "\n--- Event Stage %d: Unknown\n\n", stage);
1656: }
1657: }
1659: PetscFree(localStageUsed);
1660: PetscFree(stageUsed);
1661: PetscFree(localStageVisible);
1662: PetscFree(stageVisible);
1664: /* Information unrelated to this particular run */
1665: PetscFPrintf(comm, fd, "========================================================================================================================\n");
1666: PetscTime(&y);
1667: PetscTime(&x);
1668: PetscTime(&y); PetscTime(&y); PetscTime(&y); PetscTime(&y); PetscTime(&y);
1669: PetscTime(&y); PetscTime(&y); PetscTime(&y); PetscTime(&y); PetscTime(&y);
1670: PetscFPrintf(comm,fd,"Average time to get PetscTime(): %g\n", (y-x)/10.0);
1671: /* MPI information */
1672: if (size > 1) {
1673: MPI_Status status;
1674: PetscMPIInt tag;
1675: MPI_Comm newcomm;
1677: MPI_Barrier(comm);
1678: PetscTime(&x);
1679: MPI_Barrier(comm);
1680: MPI_Barrier(comm);
1681: MPI_Barrier(comm);
1682: MPI_Barrier(comm);
1683: MPI_Barrier(comm);
1684: PetscTime(&y);
1685: PetscFPrintf(comm, fd, "Average time for MPI_Barrier(): %g\n", (y-x)/5.0);
1686: PetscCommDuplicate(comm,&newcomm, &tag);
1687: MPI_Barrier(comm);
1688: if (rank) {
1689: MPI_Recv(0, 0, MPI_INT, rank-1, tag, newcomm, &status);
1690: MPI_Send(0, 0, MPI_INT, (rank+1)%size, tag, newcomm);
1691: } else {
1692: PetscTime(&x);
1693: MPI_Send(0, 0, MPI_INT, 1, tag, newcomm);
1694: MPI_Recv(0, 0, MPI_INT, size-1, tag, newcomm, &status);
1695: PetscTime(&y);
1696: PetscFPrintf(comm,fd,"Average time for zero size MPI_Send(): %g\n", (y-x)/size);
1697: }
1698: PetscCommDestroy(&newcomm);
1699: }
1700: PetscOptionsView(viewer);
1702: /* Machine and compile information */
1703: #if defined(PETSC_USE_FORTRAN_KERNELS)
1704: PetscFPrintf(comm, fd, "Compiled with FORTRAN kernels\n");
1705: #else
1706: PetscFPrintf(comm, fd, "Compiled without FORTRAN kernels\n");
1707: #endif
1708: #if defined(PETSC_USE_REAL_SINGLE)
1709: PetscFPrintf(comm, fd, "Compiled with single precision PetscScalar and PetscReal\n");
1710: #elif defined(PETSC_USE_LONGDOUBLE)
1711: PetscFPrintf(comm, fd, "Compiled with long double precision PetscScalar and PetscReal\n");
1712: #endif
1714: #if defined(PETSC_USE_REAL_MAT_SINGLE)
1715: PetscFPrintf(comm, fd, "Compiled with single precision matrices\n");
1716: #else
1717: PetscFPrintf(comm, fd, "Compiled with full precision matrices (default)\n");
1718: #endif
1719: PetscFPrintf(comm, fd, "sizeof(short) %d sizeof(int) %d sizeof(long) %d sizeof(void*) %d sizeof(PetscScalar) %d sizeof(PetscInt) %d\n",
1720: (int) sizeof(short), (int) sizeof(int), (int) sizeof(long), (int) sizeof(void*),(int) sizeof(PetscScalar),(int) sizeof(PetscInt));
1722: PetscFPrintf(comm, fd, "Configure options: %s",petscconfigureoptions);
1723: PetscFPrintf(comm, fd, "%s", petscmachineinfo);
1724: PetscFPrintf(comm, fd, "%s", petsccompilerinfo);
1725: PetscFPrintf(comm, fd, "%s", petsccompilerflagsinfo);
1726: PetscFPrintf(comm, fd, "%s", petsclinkerinfo);
1728: /* Cleanup */
1729: PetscFPrintf(comm, fd, "\n");
1730: return(0);
1731: }
1735: /*@C
1736: PetscLogView - Prints a summary of the logging.
1738: Collective over MPI_Comm
1740: Input Parameter:
1741: . viewer - an ASCII viewer
1743: Options Database Keys:
1744: . -log_view [viewertype[:filename[:format]]] - Prints summary of log information (for code compiled with PETSC_USE_LOG)
1746: Usage:
1747: .vb
1748: PetscInitialize(...);
1749: PetscLogBegin();
1750: ... code ...
1751: PetscLogView(PetscViewer);
1752: PetscFinalize(...);
1753: .ve
1755: Notes:
1756: By default the summary is printed to stdout.
1758: Level: beginner
1760: .keywords: log, dump, print
1761: .seealso: PetscLogBegin(), PetscLogDump()
1762: @*/
1763: PetscErrorCode PetscLogView(PetscViewer viewer)
1764: {
1765: PetscErrorCode ierr;
1766: PetscBool isascii;
1767: PetscViewerFormat format;
1768: int stage, lastStage;
1769: PetscStageLog stageLog;
1772: if (!PetscLogBegin_PrivateCalled) SETERRQ(PetscObjectComm((PetscObject)viewer), PETSC_ERR_ORDER, "No call to PetscLogBegin() before PetscLogView()");
1773: /* Pop off any stages the user forgot to remove */
1774: lastStage = 0;
1775: PetscLogGetStageLog(&stageLog);
1776: PetscStageLogGetCurrent(stageLog, &stage);
1777: while (stage >= 0) {
1778: lastStage = stage;
1779: PetscStageLogPop(stageLog);
1780: PetscStageLogGetCurrent(stageLog, &stage);
1781: }
1782: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);
1783: if (!isascii) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Currently can only view logging to ASCII");
1784: PetscViewerGetFormat(viewer,&format);
1785: if (format == PETSC_VIEWER_DEFAULT || format == PETSC_VIEWER_ASCII_INFO) {
1786: PetscLogView_Default(viewer);
1787: } else if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1788: PetscLogView_Detailed(viewer);
1789: }
1790: PetscStageLogPush(stageLog, lastStage);
1791: return(0);
1792: }
1796: /*@C
1797: PetscLogViewFromOptions - Processes command line options to determine if/how a PetscLog is to be viewed.
1799: Collective on PETSC_COMM_WORLD
1801: Not normally called by user
1803: Level: intermediate
1805: @*/
1806: PetscErrorCode PetscLogViewFromOptions(void)
1807: {
1808: PetscErrorCode ierr;
1809: PetscViewer viewer;
1810: PetscBool flg;
1811: PetscViewerFormat format;
1814: PetscOptionsGetViewer(PETSC_COMM_WORLD,NULL,"-log_view",&viewer,&format,&flg);
1815: if (flg) {
1816: PetscViewerPushFormat(viewer,format);
1817: PetscLogView(viewer);
1818: PetscViewerPopFormat(viewer);
1819: PetscViewerDestroy(&viewer);
1820: }
1821: return(0);
1822: }
1826: /*----------------------------------------------- Counter Functions -------------------------------------------------*/
1829: /*@C
1830: PetscGetFlops - Returns the number of flops used on this processor
1831: since the program began.
1833: Not Collective
1835: Output Parameter:
1836: flops - number of floating point operations
1838: Notes:
1839: A global counter logs all PETSc flop counts. The user can use
1840: PetscLogFlops() to increment this counter to include flops for the
1841: application code.
1843: PETSc automatically logs library events if the code has been
1844: compiled with -DPETSC_USE_LOG (which is the default), and -log,
1845: -log_summary, or -log_all are specified. PetscLogFlops() is
1846: intended for logging user flops to supplement this PETSc
1847: information.
1849: Level: intermediate
1851: .keywords: log, flops, floating point operations
1853: .seealso: PetscTime(), PetscLogFlops()
1854: @*/
1855: PetscErrorCode PetscGetFlops(PetscLogDouble *flops)
1856: {
1858: *flops = petsc_TotalFlops;
1859: return(0);
1860: }
1864: PetscErrorCode PetscLogObjectState(PetscObject obj, const char format[], ...)
1865: {
1867: size_t fullLength;
1868: va_list Argp;
1871: if (!petsc_logObjects) return(0);
1872: va_start(Argp, format);
1873: PetscVSNPrintf(petsc_objects[obj->id].info, 64,format,&fullLength, Argp);
1874: va_end(Argp);
1875: return(0);
1876: }
1879: /*MC
1880: PetscLogFlops - Adds floating point operations to the global counter.
1882: Synopsis:
1883: #include <petsclog.h>
1884: PetscErrorCode PetscLogFlops(PetscLogDouble f)
1886: Not Collective
1888: Input Parameter:
1889: . f - flop counter
1892: Usage:
1893: .vb
1894: PetscLogEvent USER_EVENT;
1895: PetscLogEventRegister("User event",0,&USER_EVENT);
1896: PetscLogEventBegin(USER_EVENT,0,0,0,0);
1897: [code segment to monitor]
1898: PetscLogFlops(user_flops)
1899: PetscLogEventEnd(USER_EVENT,0,0,0,0);
1900: .ve
1902: Notes:
1903: A global counter logs all PETSc flop counts. The user can use
1904: PetscLogFlops() to increment this counter to include flops for the
1905: application code.
1907: PETSc automatically logs library events if the code has been
1908: compiled with -DPETSC_USE_LOG (which is the default), and -log,
1909: -log_summary, or -log_all are specified. PetscLogFlops() is
1910: intended for logging user flops to supplement this PETSc
1911: information.
1913: Level: intermediate
1915: .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscGetFlops()
1917: .keywords: log, flops, floating point operations
1918: M*/
1920: /*MC
1921: PetscPreLoadBegin - Begin a segment of code that may be preloaded (run twice)
1922: to get accurate timings
1924: Synopsis:
1925: #include <petsclog.h>
1926: void PetscPreLoadBegin(PetscBool flag,char *name);
1928: Not Collective
1930: Input Parameter:
1931: + flag - PETSC_TRUE to run twice, PETSC_FALSE to run once, may be overridden
1932: with command line option -preload true or -preload false
1933: - name - name of first stage (lines of code timed separately with -log_summary) to
1934: be preloaded
1936: Usage:
1937: .vb
1938: PetscPreLoadBegin(PETSC_TRUE,"first stage);
1939: lines of code
1940: PetscPreLoadStage("second stage");
1941: lines of code
1942: PetscPreLoadEnd();
1943: .ve
1945: Notes: Only works in C/C++, not Fortran
1947: Flags available within the macro.
1948: + PetscPreLoadingUsed - true if we are or have done preloading
1949: . PetscPreLoadingOn - true if it is CURRENTLY doing preload
1950: . PetscPreLoadIt - 0 for the first computation (with preloading turned off it is only 0) 1 for the second
1951: - PetscPreLoadMax - number of times it will do the computation, only one when preloading is turned on
1952: The first two variables are available throughout the program, the second two only between the PetscPreLoadBegin()
1953: and PetscPreLoadEnd()
1955: Level: intermediate
1957: .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscPreLoadEnd(), PetscPreLoadStage()
1959: Concepts: preloading
1960: Concepts: timing^accurate
1961: Concepts: paging^eliminating effects of
1964: M*/
1966: /*MC
1967: PetscPreLoadEnd - End a segment of code that may be preloaded (run twice)
1968: to get accurate timings
1970: Synopsis:
1971: #include <petsclog.h>
1972: void PetscPreLoadEnd(void);
1974: Not Collective
1976: Usage:
1977: .vb
1978: PetscPreLoadBegin(PETSC_TRUE,"first stage);
1979: lines of code
1980: PetscPreLoadStage("second stage");
1981: lines of code
1982: PetscPreLoadEnd();
1983: .ve
1985: Notes: only works in C/C++ not fortran
1987: Level: intermediate
1989: .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscPreLoadBegin(), PetscPreLoadStage()
1991: M*/
1993: /*MC
1994: PetscPreLoadStage - Start a new segment of code to be timed separately.
1995: to get accurate timings
1997: Synopsis:
1998: #include <petsclog.h>
1999: void PetscPreLoadStage(char *name);
2001: Not Collective
2003: Usage:
2004: .vb
2005: PetscPreLoadBegin(PETSC_TRUE,"first stage);
2006: lines of code
2007: PetscPreLoadStage("second stage");
2008: lines of code
2009: PetscPreLoadEnd();
2010: .ve
2012: Notes: only works in C/C++ not fortran
2014: Level: intermediate
2016: .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscPreLoadBegin(), PetscPreLoadEnd()
2018: M*/
2021: #else /* end of -DPETSC_USE_LOG section */
2025: PetscErrorCode PetscLogObjectState(PetscObject obj, const char format[], ...)
2026: {
2028: return(0);
2029: }
2031: #endif /* PETSC_USE_LOG*/
2034: PetscClassId PETSC_LARGEST_CLASSID = PETSC_SMALLEST_CLASSID;
2035: PetscClassId PETSC_OBJECT_CLASSID = 0;
2039: /*@C
2040: PetscClassIdRegister - Registers a new class name for objects and logging operations in an application code.
2042: Not Collective
2044: Input Parameter:
2045: . name - The class name
2047: Output Parameter:
2048: . oclass - The class id or classid
2050: Level: developer
2052: .keywords: log, class, register
2054: @*/
2055: PetscErrorCode PetscClassIdRegister(const char name[],PetscClassId *oclass)
2056: {
2057: #if defined(PETSC_USE_LOG)
2058: PetscStageLog stageLog;
2059: PetscInt stage;
2061: #endif
2064: *oclass = ++PETSC_LARGEST_CLASSID;
2065: #if defined(PETSC_USE_LOG)
2066: PetscLogGetStageLog(&stageLog);
2067: PetscClassRegLogRegister(stageLog->classLog, name, *oclass);
2068: for (stage = 0; stage < stageLog->numStages; stage++) {
2069: ClassPerfLogEnsureSize(stageLog->stageInfo[stage].classLog, stageLog->classLog->numClasses);
2070: }
2071: #endif
2072: return(0);
2073: }
2075: #if defined(PETSC_USE_LOG) && defined(PETSC_HAVE_MPE)
2076: #include <mpe.h>
2078: PetscBool PetscBeganMPE = PETSC_FALSE;
2080: PETSC_INTERN PetscErrorCode PetscLogEventBeginMPE(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject);
2081: PETSC_INTERN PetscErrorCode PetscLogEventEndMPE(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject);
2085: /*@C
2086: PetscLogMPEBegin - Turns on MPE logging of events. This creates large log files
2087: and slows the program down.
2089: Collective over PETSC_COMM_WORLD
2091: Options Database Keys:
2092: . -log_mpe - Prints extensive log information (for code compiled with PETSC_USE_LOG)
2094: Notes:
2095: A related routine is PetscLogBegin() (with the options key -log_summary), which is
2096: intended for production runs since it logs only flop rates and object
2097: creation (and should not significantly slow the programs).
2099: Level: advanced
2101: Concepts: logging^MPE
2102: Concepts: logging^message passing
2104: .seealso: PetscLogDump(), PetscLogBegin(), PetscLogAllBegin(), PetscLogEventActivate(),
2105: PetscLogEventDeactivate()
2106: @*/
2107: PetscErrorCode PetscLogMPEBegin(void)
2108: {
2112: /* Do MPE initialization */
2113: if (!MPE_Initialized_logging()) { /* This function exists in mpich 1.1.2 and higher */
2114: PetscInfo(0,"Initializing MPE.\n");
2115: MPE_Init_log();
2117: PetscBeganMPE = PETSC_TRUE;
2118: } else {
2119: PetscInfo(0,"MPE already initialized. Not attempting to reinitialize.\n");
2120: }
2121: PetscLogSet(PetscLogEventBeginMPE, PetscLogEventEndMPE);
2122: return(0);
2123: }
2127: /*@C
2128: PetscLogMPEDump - Dumps the MPE logging info to file for later use with Jumpshot.
2130: Collective over PETSC_COMM_WORLD
2132: Level: advanced
2134: .seealso: PetscLogDump(), PetscLogAllBegin(), PetscLogMPEBegin()
2135: @*/
2136: PetscErrorCode PetscLogMPEDump(const char sname[])
2137: {
2138: char name[PETSC_MAX_PATH_LEN];
2142: if (PetscBeganMPE) {
2143: PetscInfo(0,"Finalizing MPE.\n");
2144: if (sname) {
2145: PetscStrcpy(name,sname);
2146: } else {
2147: PetscGetProgramName(name,PETSC_MAX_PATH_LEN);
2148: }
2149: MPE_Finish_log(name);
2150: } else {
2151: PetscInfo(0,"Not finalizing MPE (not started by PETSc).\n");
2152: }
2153: return(0);
2154: }
2156: #define PETSC_RGB_COLORS_MAX 39
2157: static const char *PetscLogMPERGBColors[PETSC_RGB_COLORS_MAX] = {
2158: "OliveDrab: ",
2159: "BlueViolet: ",
2160: "CadetBlue: ",
2161: "CornflowerBlue: ",
2162: "DarkGoldenrod: ",
2163: "DarkGreen: ",
2164: "DarkKhaki: ",
2165: "DarkOliveGreen: ",
2166: "DarkOrange: ",
2167: "DarkOrchid: ",
2168: "DarkSeaGreen: ",
2169: "DarkSlateGray: ",
2170: "DarkTurquoise: ",
2171: "DeepPink: ",
2172: "DarkKhaki: ",
2173: "DimGray: ",
2174: "DodgerBlue: ",
2175: "GreenYellow: ",
2176: "HotPink: ",
2177: "IndianRed: ",
2178: "LavenderBlush: ",
2179: "LawnGreen: ",
2180: "LemonChiffon: ",
2181: "LightCoral: ",
2182: "LightCyan: ",
2183: "LightPink: ",
2184: "LightSalmon: ",
2185: "LightSlateGray: ",
2186: "LightYellow: ",
2187: "LimeGreen: ",
2188: "MediumPurple: ",
2189: "MediumSeaGreen: ",
2190: "MediumSlateBlue:",
2191: "MidnightBlue: ",
2192: "MintCream: ",
2193: "MistyRose: ",
2194: "NavajoWhite: ",
2195: "NavyBlue: ",
2196: "OliveDrab: "
2197: };
2201: /*@C
2202: PetscLogMPEGetRGBColor - This routine returns a rgb color useable with PetscLogEventRegister()
2204: Not collective. Maybe it should be?
2206: Output Parameter
2207: . str - character string representing the color
2209: Level: developer
2211: .keywords: log, mpe , color
2212: .seealso: PetscLogEventRegister
2213: @*/
2214: PetscErrorCode PetscLogMPEGetRGBColor(const char *str[])
2215: {
2216: static int idx = 0;
2219: *str = PetscLogMPERGBColors[idx];
2220: idx = (idx + 1)% PETSC_RGB_COLORS_MAX;
2221: return(0);
2222: }
2224: #endif /* PETSC_USE_LOG && PETSC_HAVE_MPE */