1 /* $Id: petsclog.h,v 1.155 2001/09/06 14:51:20 bsmith Exp $ */ 2 3 /* 4 Defines profile/logging in PETSc. 5 */ 6 7 #if !defined(__PetscLog_H) 8 #define __PetscLog_H 9 #include "petsc.h" 10 11 /* 12 Each PETSc object class has it's own cookie (internal integer in the 13 data structure used for error checking). These are all defined by an offset 14 from the lowest one, PETSC_COOKIE. 15 */ 16 #define PETSC_COOKIE 1211211 17 extern int PETSC_LARGEST_COOKIE; 18 #define PETSC_EVENT 1311311 19 extern int PETSC_LARGEST_EVENT; 20 21 /* Events for the Petsc standard library */ 22 enum {PETSC_Barrier, PETSC_MAX_EVENTS}; 23 extern int PetscEvents[PETSC_MAX_EVENTS]; 24 25 /* Global flop counter */ 26 extern PetscLogDouble _TotalFlops; 27 28 /* General logging of information; different from event logging */ 29 EXTERN int PetscLogInfo(void*,const char[],...); 30 EXTERN int PetscLogInfoDeactivateClass(int); 31 EXTERN int PetscLogInfoActivateClass(int); 32 extern PetscTruth PetscLogPrintInfo; /* if true, indicates PetscLogInfo() is turned on */ 33 34 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 35 36 /* 37 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 38 39 For the complex numbers version, note that 40 1 complex addition = 2 flops 41 1 complex multiplication = 6 flops, 42 where we define 1 flop as that for a double precision scalar. We roughly approximate 43 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 44 to the assumption that we're counting mostly additions and multiplications -- and 45 roughly the same number of each. More accurate counting could be done by distinguishing 46 among the various arithmetic operations. 47 */ 48 49 #if defined(PETSC_USE_COMPLEX) 50 #define PetscLogFlops(n) (_TotalFlops += (4*n),0) 51 #else 52 #define PetscLogFlops(n) (_TotalFlops += (n),0) 53 #endif 54 55 #if defined (PETSC_HAVE_MPE) 56 #include "mpe.h" 57 #define MPEBEGIN 1000 58 EXTERN int PetscLogMPEBegin(void); 59 EXTERN int PetscLogMPEDump(const char[]); 60 extern PetscTruth UseMPE; 61 extern int PetscLogEventMPEFlags[]; 62 EXTERN int PetscLogEventMPEActivate(int); 63 EXTERN int PetscLogEventMPEDeactivate(int); 64 #else 65 #define PetscLogEventMPEActivate(a) 0 66 #define PetscLogEventMPEDeactivate(a) 0 67 #endif 68 69 EXTERN int (*_PetscLogPLB)(int,int,PetscObject,PetscObject,PetscObject,PetscObject); 70 EXTERN int (*_PetscLogPLE)(int,int,PetscObject,PetscObject,PetscObject,PetscObject); 71 EXTERN int (*_PetscLogPHC)(PetscObject); 72 EXTERN int (*_PetscLogPHD)(PetscObject); 73 74 #define PetscLogObjectParent(p,c) if (c) {PetscValidHeader((PetscObject)(c)); \ 75 PetscValidHeader((PetscObject)(p));\ 76 ((PetscObject)(c))->parent = (PetscObject)(p);\ 77 ((PetscObject)(c))->parentid = ((PetscObject)p)->id;} 78 #define PetscLogObjectParents(p,n,d) {int _i; for (_i=0; _i<n; _i++) \ 79 PetscLogObjectParent(p,(d)[_i]);} 80 #define PetscLogObjectCreate(h) {if (_PetscLogPHC) (*_PetscLogPHC)((PetscObject)h);} 81 #define PetscLogObjectDestroy(h) {if (_PetscLogPHD) (*_PetscLogPHD)((PetscObject)h);} 82 #define PetscLogObjectMemory(p,m) {PetscValidHeader((PetscObject)p);\ 83 ((PetscObject)(p))->mem += (m);} 84 /* Initialization functions */ 85 EXTERN int PetscLogBegin(void); 86 EXTERN int PetscLogAllBegin(void); 87 EXTERN int PetscLogTraceBegin(FILE *); 88 /* General functions */ 89 EXTERN int PetscLogDestroy(void); 90 EXTERN int PetscLogSet(int (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 91 int (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 92 EXTERN int PetscLogObjectState(PetscObject, const char[], ...); 93 /* Output functions */ 94 EXTERN int PetscLogPrintSummary(MPI_Comm, const char[]); 95 EXTERN int PetscLogDump(const char[]); 96 /* Counter functions */ 97 EXTERN int PetscGetFlops(PetscLogDouble *); 98 /* Stage functions */ 99 EXTERN int PetscLogStageRegister(int *, const char[]); 100 EXTERN int PetscLogStagePush(int); 101 EXTERN int PetscLogStagePop(void); 102 EXTERN int PetscLogStageSetVisible(int, PetscTruth); 103 EXTERN int PetscLogStageGetVisible(int, PetscTruth *); 104 EXTERN int PetscLogStageGetId(const char [], int *); 105 /* Event functions */ 106 EXTERN int PetscLogEventRegister(int *, const char[], const char[], int); 107 EXTERN int PetscLogEventActivate(int); 108 EXTERN int PetscLogEventDeactivate(int); 109 EXTERN int PetscLogEventActivateClass(int); 110 EXTERN int PetscLogEventDeactivateClass(int); 111 /* Class functions */ 112 EXTERN int PetscLogClassRegister(int *, const char []); 113 114 /* Default log */ 115 typedef struct _StageLog *StageLog; 116 extern StageLog _stageLog; 117 118 /* Global counters */ 119 extern PetscLogDouble irecv_ct, isend_ct, recv_ct, send_ct; 120 extern PetscLogDouble irecv_len, isend_len, recv_len, send_len; 121 extern PetscLogDouble allreduce_ct; 122 extern PetscLogDouble wait_ct, wait_any_ct, wait_all_ct, sum_of_waits_ct; 123 extern int PETSC_DUMMY, PETSC_DUMMY_SIZE; 124 125 /* We must make these structures available if we are to access the event 126 activation flags in the PetscLogEventBegin/End() macros. If we forced a 127 function call each time, we could leave these structures in plog.h 128 */ 129 /* The structure for logging performance */ 130 typedef struct _PerfInfo { 131 char *name; /* The name of this section */ 132 char *color; /* The color of this section */ 133 int id; /* The integer identifying this event */ 134 int cookie; /* The class id for this section */ 135 int depth; /* The nesting depth of the event call */ 136 int count; /* The number of times this section was executed */ 137 PetscLogDouble flops; /* The flops used in this section */ 138 PetscLogDouble time; /* The time taken for this section */ 139 PetscLogDouble numMessages; /* The number of messages in this section */ 140 PetscLogDouble messageLength; /* The total message lengths in this section */ 141 PetscLogDouble numReductions; /* The number of reductions in this section */ 142 } PerfInfo; 143 144 /* The structure for logging events */ 145 typedef struct _EventLog *EventLog; 146 struct _EventLog { 147 /* Size information */ 148 int numEvents; /* The number of registered events */ 149 int maxEvents; /* The maximum number of events */ 150 /* Event specific information */ 151 int *eventActive; /* The flag for active events */ 152 PerfInfo *eventInfo; /* The performance information for each event */ 153 }; 154 155 /* The structure for logging class information */ 156 typedef struct _ClassInfo { 157 char *name; /* The class name */ 158 int cookie; /* The integer identifying this class */ 159 int creations; /* The number of objects of this class created */ 160 int destructions; /* The number of objects of this class destroyed */ 161 PetscLogDouble mem; /* The total memory allocated by objects of this class */ 162 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 163 } ClassInfo; 164 165 typedef struct _ClassLog *ClassLog; 166 struct _ClassLog { 167 int numClasses; /* The number of classes registered */ 168 int maxClasses; /* The maximum number of classes */ 169 ClassInfo *classInfo; /* The structure for classs information (cookies are monotonicly increasing) */ 170 }; 171 172 /* A simple stack (should replace) */ 173 typedef struct _IntStack *IntStack; 174 175 /* The structure for logging in stages */ 176 struct _StageLog { 177 /* Size information */ 178 int numStages; /* The number of registered stages */ 179 int maxStages; /* The maximum number of stages */ 180 /* Runtime information */ 181 int *stageUsed; /* The flags for stages which were executed */ 182 PetscTruth *stageVisible; /* The flags for stages which are visible to PetscLogPrintSummary() */ 183 IntStack stack; /* The stack for active stages */ 184 int curStage; /* The current stage (only used in macros so we don't call StackTop) */ 185 /* Stage specific information */ 186 PerfInfo *stageInfo; /* The performance information for each stage */ 187 EventLog *eventLog; /* The event log for each stage */ 188 ClassLog *classLog; /* The class information for each stage */ 189 }; 190 191 EXTERN int EventLogGetEvent(EventLog, int, int *); 192 193 #if defined(PETSC_HAVE_MPE) 194 #define PetscLogEventBarrierBegin(e,o1,o2,o3,o4,cm) \ 195 0; { int _1_ierr, _1_eventNum; \ 196 _1_ierr = EventLogGetEvent(_stageLog->eventLog[_stageLog->curStage], (e), &_1_eventNum); \ 197 if (_PetscLogPLB && _stageLog->eventLog[_stageLog->curStage]->eventActive[_1_eventNum]) { \ 198 _1_ierr = PetscLogEventBegin((e),o1,o2,o3,o4);CHKERRQ(_1_ierr); \ 199 if (UseMPE && PetscLogEventMPEFlags[(e)]) \ 200 MPE_Log_event(MPEBEGIN+2*(e),0,""); \ 201 _1_ierr = MPI_Barrier(cm);CHKERRQ(_1_ierr); \ 202 _1_ierr = PetscLogEventEnd((e),o1,o2,o3,o4);CHKERRQ(_1_ierr); \ 203 if (UseMPE && PetscLogEventMPEFlags[(e)]) \ 204 MPE_Log_event(MPEBEGIN+2*((e)+1),0,""); \ 205 } \ 206 _1_ierr = PetscLogEventBegin(e+1,o1,o2,o3,o4);CHKERRQ(_1_ierr); \ 207 if (UseMPE && PetscLogEventMPEFlags[(e)+1]) \ 208 MPE_Log_event(MPEBEGIN+2*((e)+1),0,""); \ 209 } 210 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 211 0; { int _3_ierr, _3_eventNum; \ 212 _3_ierr = EventLogGetEvent(_stageLog->eventLog[_stageLog->curStage], (e), &_3_eventNum); \ 213 if (_PetscLogPLB && _stageLog->eventLog[_stageLog->curStage]->eventActive[_3_eventNum]) {\ 214 (*_PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4));}\ 215 if (UseMPE && PetscLogEventMPEFlags[(e)])\ 216 MPE_Log_event(MPEBEGIN+2*(e),0,"");\ 217 } 218 #else 219 #define PetscLogEventBarrierBegin(e,o1,o2,o3,o4,cm) \ 220 0; { int _2_ierr, _2_eventNum;\ 221 _2_ierr = EventLogGetEvent(_stageLog->eventLog[_stageLog->curStage], (e), &_2_eventNum); \ 222 if (_PetscLogPLB && _stageLog->eventLog[_stageLog->curStage]->eventActive[_2_eventNum]) { \ 223 _2_ierr = PetscLogEventBegin((e),o1,o2,o3,o4);CHKERRQ(_2_ierr); \ 224 _2_ierr = MPI_Barrier(cm);CHKERRQ(_2_ierr); \ 225 _2_ierr = PetscLogEventEnd((e),o1,o2,o3,o4);CHKERRQ(_2_ierr); \ 226 } \ 227 _2_ierr = PetscLogEventBegin((e)+1,o1,o2,o3,o4);CHKERRQ(_2_ierr); \ 228 } 229 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 230 0; { int _4_ierr, _4_eventNum; \ 231 _4_ierr = EventLogGetEvent(_stageLog->eventLog[_stageLog->curStage], (e), &_4_eventNum); \ 232 if (_PetscLogPLB && _stageLog->eventLog[_stageLog->curStage]->eventActive[_4_eventNum]) {\ 233 (*_PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4));}\ 234 } 235 #endif 236 237 #if defined(PETSC_HAVE_MPE) 238 #define PetscLogEventBarrierEnd(e,o1,o2,o3,o4,cm) PetscLogEventEnd(e+1,o1,o2,o3,o4) 239 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 240 0; { int _5_ierr, _5_eventNum; \ 241 _5_ierr = EventLogGetEvent(_stageLog->eventLog[_stageLog->curStage], (e), &_5_eventNum); \ 242 if (_PetscLogPLE && _stageLog->eventLog[_stageLog->curStage]->eventActive[_5_eventNum]) {\ 243 (*_PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4));}\ 244 if (UseMPE && PetscLogEventMPEFlags[(e)])\ 245 MPE_Log_event(MPEBEGIN+2*(e)+1,0,"");\ 246 } 247 #else 248 #define PetscLogEventBarrierEnd(e,o1,o2,o3,o4,cm) PetscLogEventEnd(e+1,o1,o2,o3,o4) 249 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 250 0; { int _6_ierr, _6_eventNum; \ 251 _6_ierr = EventLogGetEvent(_stageLog->eventLog[_stageLog->curStage], (e), &_6_eventNum); \ 252 if (_PetscLogPLE && _stageLog->eventLog[_stageLog->curStage]->eventActive[_6_eventNum]) {\ 253 (*_PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4));}\ 254 } 255 #endif 256 257 /* 258 This does not work for MPI-Uni because our src/mpiuni/mpi.h file 259 uses macros to defined the MPI operations. 260 261 It does not work correctly from HP-UX because it processes the 262 macros in a way that sometimes it double counts, hence 263 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 264 265 It does not work with Windows NT because winmpich lacks MPI_Type_size() 266 */ 267 #if !defined(HAVE_MPI_UNI) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 268 /* 269 Logging of MPI activities 270 */ 271 272 #define TypeSize(buff,count,type) \ 273 (\ 274 MPI_Type_size(type,&PETSC_DUMMY_SIZE),buff += ((PetscLogDouble) ((count)*PETSC_DUMMY_SIZE)) \ 275 ) 276 277 #define MPI_Irecv(buf,count, datatype,source,tag,comm,request) \ 278 (\ 279 PETSC_DUMMY = MPI_Irecv(buf,count, datatype,source,tag,comm,request), \ 280 irecv_ct++,TypeSize(irecv_len,count,datatype),PETSC_DUMMY \ 281 ) 282 283 #define MPI_Isend(buf,count, datatype,dest,tag,comm,request) \ 284 (\ 285 PETSC_DUMMY = MPI_Isend(buf,count, datatype,dest,tag,comm,request), \ 286 isend_ct++, TypeSize(isend_len,count,datatype),PETSC_DUMMY \ 287 ) 288 289 #define MPI_Startall_irecv(count,number,requests) \ 290 (\ 291 PETSC_DUMMY = MPI_Startall(number,requests), \ 292 irecv_ct += (PetscLogDouble)(number),irecv_len += ((PetscLogDouble) ((count)*sizeof(PetscScalar))),PETSC_DUMMY \ 293 ) 294 295 #define MPI_Startall_isend(count,number,requests) \ 296 (\ 297 PETSC_DUMMY = MPI_Startall(number,requests), \ 298 isend_ct += (PetscLogDouble)(number),isend_len += ((PetscLogDouble) ((count)*sizeof(PetscScalar))),PETSC_DUMMY \ 299 ) 300 301 #define MPI_Start_isend(count, requests)\ 302 (\ 303 PETSC_DUMMY = MPI_Start(requests),\ 304 isend_ct++,isend_len += ((PetscLogDouble) ((count)*sizeof(PetscScalar))),PETSC_DUMMY\ 305 ) 306 307 #define MPI_Recv(buf,count, datatype,source,tag,comm,status) \ 308 (\ 309 PETSC_DUMMY = MPI_Recv(buf,count, datatype,source,tag,comm,status), \ 310 recv_ct++,TypeSize(recv_len,count,datatype),PETSC_DUMMY \ 311 ) 312 313 #define MPI_Send(buf,count, datatype,dest,tag,comm) \ 314 (\ 315 PETSC_DUMMY = MPI_Send(buf,count, datatype,dest,tag,comm), \ 316 send_ct++, TypeSize(send_len,count,datatype),PETSC_DUMMY \ 317 ) 318 319 #define MPI_Wait(request,status) \ 320 (\ 321 wait_ct++,sum_of_waits_ct++, \ 322 MPI_Wait(request,status) \ 323 ) 324 325 #define MPI_Waitany(a,b,c,d) \ 326 (\ 327 wait_any_ct++,sum_of_waits_ct++,\ 328 MPI_Waitany(a,b,c,d) \ 329 ) 330 331 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 332 (\ 333 wait_all_ct++,sum_of_waits_ct += (PetscLogDouble) (count), \ 334 MPI_Waitall(count,array_of_requests,array_of_statuses) \ 335 ) 336 337 #define MPI_Allreduce(sendbuf, recvbuf,count,datatype,op,comm) \ 338 (\ 339 allreduce_ct++,MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm)\ 340 ) 341 342 #else 343 344 #define MPI_Startall_irecv(count,number,requests) \ 345 (\ 346 MPI_Startall(number,requests) \ 347 ) 348 349 #define MPI_Startall_isend(count,number,requests) \ 350 (\ 351 MPI_Startall(number,requests) \ 352 ) 353 354 #define MPI_Start_isend(count, requests) \ 355 (\ 356 MPI_Start(requests) \ 357 ) 358 359 #endif /* !USING_MPIUNI && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 360 361 #else /* ---Logging is turned off --------------------------------------------*/ 362 363 #define PetscLogFlops(n) 0 364 365 /* 366 With logging turned off, then MPE has to be turned off 367 */ 368 #define MPEBEGIN 1000 369 #define PetscLogMPEBegin() 0 370 #define PetscLogMPEDump(a) 0 371 #define PetscLogEventMPEActivate(a) 0 372 #define PetscLogEventMPEDeactivate(a) 0 373 374 #define PetscLogEventActivate(a) 0 375 #define PetscLogEventDeactivate(a) 0 376 377 #define PetscLogEventActivateClass(a) 0 378 #define PetscLogEventDeactivateClass(a) 0 379 380 #define _PetscLogPLB 0 381 #define _PetscLogPLE 0 382 #define _PetscLogPHC 0 383 #define _PetscLogPHD 0 384 #define PetscGetFlops(a) (*(a) = 0.0,0) 385 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0 386 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0 387 #define PetscLogEventBarrierBegin(e,o1,o2,o3,o4,cm) 0 388 #define PetscLogEventBarrierEnd(e,o1,o2,o3,o4,cm) 0 389 #define PetscLogObjectParent(p,c) 390 #define PetscLogObjectParents(p,n,c) 391 #define PetscLogObjectCreate(h) 392 #define PetscLogObjectDestroy(h) 393 #define PetscLogObjectMemory(p,m) 394 #define PetscLogDestroy() 0 395 #define PetscLogStagePush(a) 0 396 #define PetscLogStagePop() 0 397 #define PetscLogStageRegister(a,b) 0 398 #define PetscLogStagePrint(a,flg) 0 399 #define PetscLogPrintSummary(comm,file) 0 400 #define PetscLogBegin() 0 401 #define PetscLogTraceBegin(file) 0 402 #define PetscLogSet(lb,le) 0 403 #define PetscLogAllBegin() 0 404 #define PetscLogDump(c) 0 405 #define PetscLogEventRegister(a,b,c) 0 406 EXTERN int PetscLogObjectState(PetscObject,const char[],...); 407 408 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 409 #define MPI_Startall_irecv(count,number,requests) MPI_Startall(number,requests) 410 411 #define MPI_Startall_isend(count,number,requests) MPI_Startall(number,requests) 412 413 #define MPI_Start_isend(count,requests) MPI_Start(requests) 414 415 #endif /* PETSC_USE_LOG */ 416 417 extern PetscTruth PetscPreLoadingUsed; /* true if we are or have done preloading */ 418 extern PetscTruth PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 419 420 #define PreLoadBegin(flag,name) {PetscTruth PreLoading = flag; \ 421 int PreLoadMax,PreLoadIt,_stageNum,_3_ierr;\ 422 _3_ierr = PetscOptionsGetLogical(PETSC_NULL,"-preload",&PreLoading,PETSC_NULL);CHKERRQ(_3_ierr);\ 423 PreLoadMax = (int)(PreLoading);PetscPreLoadingUsed = PreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 424 for (PreLoadIt=0; PreLoadIt<=PreLoadMax; PreLoadIt++) {\ 425 PetscPreLoadingOn = PreLoading;\ 426 _3_ierr = PetscBarrier(PETSC_NULL);CHKERRQ(_3_ierr);\ 427 if (PreLoadIt>0) {\ 428 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 429 } else {\ 430 _3_ierr = PetscLogStageRegister(&_stageNum,name);CHKERRQ(_3_ierr);\ 431 _3_ierr = PetscLogStageSetVisible(_stageNum,(PetscTruth)(!PreLoadMax || PreLoadIt));\ 432 }\ 433 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 434 435 #define PreLoadEnd() _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);PreLoading = PETSC_FALSE;}} 436 437 #define PreLoadStage(name) _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 438 if (PreLoadIt>0) {\ 439 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 440 } else {\ 441 _3_ierr = PetscLogStageRegister(&_stageNum,name);CHKERRQ(_3_ierr);\ 442 _3_ierr = PetscLogStageSetVisible(_stageNum,(PetscTruth)(!PreLoadMax || PreLoadIt));\ 443 }\ 444 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 445 #endif 446