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 extern int PETSC_Barrier; 23 24 /* Global flop counter */ 25 extern PetscLogDouble _TotalFlops; 26 27 /* General logging of information; different from event logging */ 28 EXTERN int PetscLogInfo(void*,const char[],...) PETSC_PRINTF_FORMAT_CHECK(2,3); 29 EXTERN int PetscLogInfoDeactivateClass(int); 30 EXTERN int PetscLogInfoActivateClass(int); 31 extern PetscTruth PetscLogPrintInfo; /* if true, indicates PetscLogInfo() is turned on */ 32 33 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 34 35 /* 36 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 37 38 For the complex numbers version, note that 39 1 complex addition = 2 flops 40 1 complex multiplication = 6 flops, 41 where we define 1 flop as that for a double precision scalar. We roughly approximate 42 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 43 to the assumption that we're counting mostly additions and multiplications -- and 44 roughly the same number of each. More accurate counting could be done by distinguishing 45 among the various arithmetic operations. 46 */ 47 48 #if defined(PETSC_USE_COMPLEX) 49 #define PetscLogFlops(n) (_TotalFlops += (4*n),0) 50 #else 51 #define PetscLogFlops(n) (_TotalFlops += (n),0) 52 #endif 53 54 #if defined (PETSC_HAVE_MPE) 55 #include "mpe.h" 56 EXTERN int PetscLogMPEBegin(void); 57 EXTERN int PetscLogMPEDump(const char[]); 58 extern PetscTruth UseMPE; 59 #define PETSC_LOG_EVENT_MPE_BEGIN(e) \ 60 if(UseMPE && _stageLog->stageInfo[_stageLog->curStage].eventLog->eventInfo[e].active) \ 61 MPE_Log_event(_stageLog->eventLog->eventInfo[e].mpe_id_begin,0,""); 62 63 #define PETSC_LOG_EVENT_MPE_END(e) \ 64 if(UseMPE && _stageLog->stageInfo[_stageLog->curStage].eventLog->eventInfo[e].active) \ 65 MPE_Log_event(_stageLog->eventLog->eventInfo[e].mpe_id_end,0,""); 66 67 #else 68 #define PETSC_LOG_EVENT_MPE_BEGIN(e) 69 #define PETSC_LOG_EVENT_MPE_END(e) 70 #endif 71 72 EXTERN int (*_PetscLogPLB)(int,int,PetscObject,PetscObject,PetscObject,PetscObject); 73 EXTERN int (*_PetscLogPLE)(int,int,PetscObject,PetscObject,PetscObject,PetscObject); 74 EXTERN int (*_PetscLogPHC)(PetscObject); 75 EXTERN int (*_PetscLogPHD)(PetscObject); 76 77 #define PetscLogObjectParent(p,c) \ 78 if (c) {\ 79 PetscValidHeader((PetscObject)(c));\ 80 PetscValidHeader((PetscObject)(p));\ 81 ((PetscObject)(c))->parent = (PetscObject)(p);\ 82 ((PetscObject)(c))->parentid = ((PetscObject)p)->id;\ 83 } 84 #define PetscLogObjectParents(p,n,d) {int _i; for (_i=0; _i<n; _i++) PetscLogObjectParent(p,(d)[_i]);} 85 #define PetscLogObjectCreate(h) {if (_PetscLogPHC) (*_PetscLogPHC)((PetscObject)h);} 86 #define PetscLogObjectDestroy(h) {if (_PetscLogPHD) (*_PetscLogPHD)((PetscObject)h);} 87 #define PetscLogObjectMemory(p,m) {PetscValidHeader((PetscObject)p);((PetscObject)(p))->mem += (m);} 88 /* Initialization functions */ 89 EXTERN int PetscLogBegin(void); 90 EXTERN int PetscLogAllBegin(void); 91 EXTERN int PetscLogTraceBegin(FILE *); 92 EXTERN int PetscLogActions(PetscTruth); 93 EXTERN int PetscLogObjects(PetscTruth); 94 /* General functions */ 95 EXTERN int PetscLogGetRGBColor(char **); 96 EXTERN int PetscLogDestroy(void); 97 EXTERN int PetscLogSet(int (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 98 int (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 99 EXTERN int PetscLogObjectState(PetscObject, const char[], ...) PETSC_PRINTF_FORMAT_CHECK(2,3); 100 /* Output functions */ 101 EXTERN int PetscLogPrintSummary(MPI_Comm, const char[]); 102 EXTERN int PetscLogDump(const char[]); 103 /* Counter functions */ 104 EXTERN int PetscGetFlops(PetscLogDouble *); 105 /* Stage functions */ 106 EXTERN int PetscLogStageRegister(int *, const char[]); 107 EXTERN int PetscLogStagePush(int); 108 EXTERN int PetscLogStagePop(void); 109 EXTERN int PetscLogStageSetActive(int, PetscTruth); 110 EXTERN int PetscLogStageGetActive(int, PetscTruth *); 111 EXTERN int PetscLogStageSetVisible(int, PetscTruth); 112 EXTERN int PetscLogStageGetVisible(int, PetscTruth *); 113 EXTERN int PetscLogStageGetId(const char [], int *); 114 /* Event functions */ 115 EXTERN int PetscLogEventRegister(int *, const char[], int); 116 EXTERN int PetscLogEventActivate(int); 117 EXTERN int PetscLogEventDeactivate(int); 118 EXTERN int PetscLogEventActivateClass(int); 119 EXTERN int PetscLogEventDeactivateClass(int); 120 /* Class functions */ 121 EXTERN int PetscLogClassRegister(int *, const char []); 122 123 /* Global counters */ 124 extern PetscLogDouble irecv_ct, isend_ct, recv_ct, send_ct; 125 extern PetscLogDouble irecv_len, isend_len, recv_len, send_len; 126 extern PetscLogDouble allreduce_ct; 127 extern PetscLogDouble wait_ct, wait_any_ct, wait_all_ct, sum_of_waits_ct; 128 extern int PETSC_DUMMY, PETSC_DUMMY_SIZE; 129 130 /* We must make these structures available if we are to access the event 131 activation flags in the PetscLogEventBegin/End() macros. If we forced a 132 function call each time, we could leave these structures in plog.h 133 */ 134 /* Default log */ 135 typedef struct _StageLog *StageLog; 136 extern StageLog _stageLog; 137 138 /* A simple stack (should replace) */ 139 typedef struct _IntStack *IntStack; 140 141 /* The structures for logging performance */ 142 typedef struct _EventPerfInfo { 143 int id; /* The integer identifying this section */ 144 PetscTruth active; /* The flag to activate logging */ 145 PetscTruth visible; /* The flag to print info in summary */ 146 int depth; /* The nesting depth of the event call */ 147 int count; /* The number of times this section was executed */ 148 PetscLogDouble flops; /* The flops used in this section */ 149 PetscLogDouble time; /* The time taken for this section */ 150 PetscLogDouble numMessages; /* The number of messages in this section */ 151 PetscLogDouble messageLength; /* The total message lengths in this section */ 152 PetscLogDouble numReductions; /* The number of reductions in this section */ 153 } EventPerfInfo; 154 155 typedef struct _ClassPerfInfo { 156 int id; /* The integer identifying this class */ 157 int creations; /* The number of objects of this class created */ 158 int destructions; /* The number of objects of this class destroyed */ 159 PetscLogDouble mem; /* The total memory allocated by objects of this class */ 160 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 161 } ClassPerfInfo; 162 163 /* The structures for logging registration */ 164 typedef struct _ClassRegInfo { 165 char *name; /* The class name */ 166 int cookie; /* The integer identifying this class */ 167 } ClassRegInfo; 168 169 typedef struct _EventRegInfo { 170 char *name; /* The name of this event */ 171 int cookie; /* The class id for this event (should maybe give class ID instead) */ 172 #if defined (PETSC_HAVE_MPE) 173 int mpe_id_begin; /* MPE IDs that define the event */ 174 int mpe_id_end; 175 #endif 176 } EventRegInfo; 177 178 /* The structure for logging events */ 179 typedef int PetscEvent; 180 181 typedef struct _EventRegLog *EventRegLog; 182 struct _EventRegLog { 183 int numEvents; /* The number of registered events */ 184 int maxEvents; /* The maximum number of events */ 185 EventRegInfo *eventInfo; /* The registration information for each event */ 186 }; 187 188 typedef struct _EventPerfLog *EventPerfLog; 189 struct _EventPerfLog { 190 int numEvents; /* The number of logging events */ 191 int maxEvents; /* The maximum number of events */ 192 EventPerfInfo *eventInfo; /* The performance information for each event */ 193 }; 194 195 /* The structure for logging class information */ 196 typedef struct _ClassRegLog *ClassRegLog; 197 struct _ClassRegLog { 198 int numClasses; /* The number of classes registered */ 199 int maxClasses; /* The maximum number of classes */ 200 ClassRegInfo *classInfo; /* The structure for class information (cookies are monotonicly increasing) */ 201 }; 202 203 typedef struct _ClassPerfLog *ClassPerfLog; 204 struct _ClassPerfLog { 205 int numClasses; /* The number of logging classes */ 206 int maxClasses; /* The maximum number of classes */ 207 ClassPerfInfo *classInfo; /* The structure for class information (cookies are monotonicly increasing) */ 208 }; 209 210 /* The structures for logging in stages */ 211 typedef struct _StageInfo { 212 char *name; /* The stage name */ 213 PetscTruth used; /* The stage was pushed on this processor */ 214 EventPerfInfo perfInfo; /* The stage performance information */ 215 EventPerfLog eventLog; /* The event information for this stage */ 216 ClassPerfLog classLog; /* The class information for this stage */ 217 } StageInfo; 218 219 struct _StageLog { 220 /* Size information */ 221 int numStages; /* The number of registered stages */ 222 int maxStages; /* The maximum number of stages */ 223 /* Runtime information */ 224 IntStack stack; /* The stack for active stages */ 225 int curStage; /* The current stage (only used in macros so we don't call StackTop) */ 226 /* Stage specific information */ 227 StageInfo *stageInfo; /* The information for each stage */ 228 EventRegLog eventLog; /* The registered events */ 229 ClassRegLog classLog; /* The registered classes */ 230 }; 231 232 #define PetscLogEventBarrierBegin(e,o1,o2,o3,o4,cm) 0; \ 233 {\ 234 int _2_ierr;\ 235 if (_PetscLogPLB && \ 236 _stageLog->stageInfo[_stageLog->curStage].perfInfo.active && \ 237 _stageLog->stageInfo[_stageLog->curStage].eventLog->eventInfo[e].active) {\ 238 _2_ierr = PetscLogEventBegin((e),o1,o2,o3,o4);CHKERRQ(_2_ierr);\ 239 _2_ierr = MPI_Barrier(cm);CHKERRQ(_2_ierr);\ 240 _2_ierr = PetscLogEventEnd((e),o1,o2,o3,o4);CHKERRQ(_2_ierr);\ 241 }\ 242 _2_ierr = PetscLogEventBegin((e)+1,o1,o2,o3,o4);CHKERRQ(_2_ierr);\ 243 } 244 245 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0; \ 246 {\ 247 if (_PetscLogPLB && \ 248 _stageLog->stageInfo[_stageLog->curStage].perfInfo.active && \ 249 _stageLog->stageInfo[_stageLog->curStage].eventLog->eventInfo[e].active) {\ 250 (*_PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4));\ 251 }\ 252 PETSC_LOG_EVENT_MPE_BEGIN(e); \ 253 } 254 255 #define PetscLogEventBarrierEnd(e,o1,o2,o3,o4,cm) PetscLogEventEnd(e+1,o1,o2,o3,o4) 256 257 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0; \ 258 {\ 259 if (_PetscLogPLE && \ 260 _stageLog->stageInfo[_stageLog->curStage].perfInfo.active && \ 261 _stageLog->stageInfo[_stageLog->curStage].eventLog->eventInfo[e].active) {\ 262 (*_PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4));\ 263 }\ 264 PETSC_LOG_EVENT_MPE_END(e); \ 265 } 266 267 /* 268 This does not work for MPI-Uni because our src/mpiuni/mpi.h file 269 uses macros to defined the MPI operations. 270 271 It does not work correctly from HP-UX because it processes the 272 macros in a way that sometimes it double counts, hence 273 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 274 275 It does not work with Windows NT because winmpich lacks MPI_Type_size() 276 */ 277 #if !defined(_petsc_mpi_uni) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 278 /* 279 Logging of MPI activities 280 */ 281 282 #define TypeSize(buff,count,type) \ 283 (\ 284 MPI_Type_size(type,&PETSC_DUMMY_SIZE),buff += ((PetscLogDouble) ((count)*PETSC_DUMMY_SIZE))\ 285 ) 286 287 #define MPI_Irecv(buf,count, datatype,source,tag,comm,request) \ 288 (\ 289 PETSC_DUMMY = MPI_Irecv(buf,count, datatype,source,tag,comm,request),\ 290 irecv_ct++,TypeSize(irecv_len,count,datatype),PETSC_DUMMY\ 291 ) 292 293 #define MPI_Isend(buf,count, datatype,dest,tag,comm,request) \ 294 (\ 295 PETSC_DUMMY = MPI_Isend(buf,count, datatype,dest,tag,comm,request),\ 296 isend_ct++, TypeSize(isend_len,count,datatype),PETSC_DUMMY\ 297 ) 298 299 #define MPI_Startall_irecv(count,number,requests) \ 300 (\ 301 PETSC_DUMMY = MPI_Startall(number,requests),\ 302 irecv_ct += (PetscLogDouble)(number),irecv_len += ((PetscLogDouble) ((count)*sizeof(PetscScalar))),PETSC_DUMMY\ 303 ) 304 305 #define MPI_Startall_isend(count,number,requests) \ 306 (\ 307 PETSC_DUMMY = MPI_Startall(number,requests),\ 308 isend_ct += (PetscLogDouble)(number),isend_len += ((PetscLogDouble) ((count)*sizeof(PetscScalar))),PETSC_DUMMY\ 309 ) 310 311 #define MPI_Start_isend(count, requests) \ 312 (\ 313 PETSC_DUMMY = MPI_Start(requests),\ 314 isend_ct++,isend_len += ((PetscLogDouble) ((count)*sizeof(PetscScalar))),PETSC_DUMMY\ 315 ) 316 317 #define MPI_Recv(buf,count, datatype,source,tag,comm,status) \ 318 (\ 319 PETSC_DUMMY = MPI_Recv(buf,count, datatype,source,tag,comm,status),\ 320 recv_ct++,TypeSize(recv_len,count,datatype),PETSC_DUMMY\ 321 ) 322 323 #define MPI_Send(buf,count, datatype,dest,tag,comm) \ 324 (\ 325 PETSC_DUMMY = MPI_Send(buf,count, datatype,dest,tag,comm),\ 326 send_ct++, TypeSize(send_len,count,datatype),PETSC_DUMMY\ 327 ) 328 329 #define MPI_Wait(request,status) \ 330 (\ 331 wait_ct++,sum_of_waits_ct++,\ 332 MPI_Wait(request,status)\ 333 ) 334 335 #define MPI_Waitany(a,b,c,d) \ 336 (\ 337 wait_any_ct++,sum_of_waits_ct++,\ 338 MPI_Waitany(a,b,c,d)\ 339 ) 340 341 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 342 (\ 343 wait_all_ct++,sum_of_waits_ct += (PetscLogDouble) (count),\ 344 MPI_Waitall(count,array_of_requests,array_of_statuses)\ 345 ) 346 347 #define MPI_Allreduce(sendbuf, recvbuf,count,datatype,op,comm) \ 348 (\ 349 allreduce_ct++,MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm)\ 350 ) 351 352 #else 353 354 #define MPI_Startall_irecv(count,number,requests) \ 355 (\ 356 MPI_Startall(number,requests)\ 357 ) 358 359 #define MPI_Startall_isend(count,number,requests) \ 360 (\ 361 MPI_Startall(number,requests)\ 362 ) 363 364 #define MPI_Start_isend(count, requests) \ 365 (\ 366 MPI_Start(requests)\ 367 ) 368 369 #endif /* !_petsc_mpi_uni && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 370 371 #else /* ---Logging is turned off --------------------------------------------*/ 372 373 #define PetscLogFlops(n) 0 374 375 /* 376 With logging turned off, then MPE has to be turned off 377 */ 378 #define PetscLogMPEBegin() 0 379 #define PetscLogMPEDump(a) 0 380 381 #define PetscLogEventActivate(a) 0 382 #define PetscLogEventDeactivate(a) 0 383 384 #define PetscLogEventActivateClass(a) 0 385 #define PetscLogEventDeactivateClass(a) 0 386 387 #define _PetscLogPLB 0 388 #define _PetscLogPLE 0 389 #define _PetscLogPHC 0 390 #define _PetscLogPHD 0 391 #define PetscGetFlops(a) (*(a) = 0.0,0) 392 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0 393 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0 394 #define PetscLogEventBarrierBegin(e,o1,o2,o3,o4,cm) 0 395 #define PetscLogEventBarrierEnd(e,o1,o2,o3,o4,cm) 0 396 #define PetscLogObjectParent(p,c) 397 #define PetscLogObjectParents(p,n,c) 398 #define PetscLogObjectCreate(h) 399 #define PetscLogObjectDestroy(h) 400 #define PetscLogObjectMemory(p,m) 401 #define PetscLogDestroy() 0 402 #define PetscLogStagePush(a) 0 403 #define PetscLogStagePop() 0 404 #define PetscLogStageRegister(a,b) 0 405 #define PetscLogStagePrint(a,flg) 0 406 #define PetscLogPrintSummary(comm,file) 0 407 #define PetscLogBegin() 0 408 #define PetscLogTraceBegin(file) 0 409 #define PetscLogSet(lb,le) 0 410 #define PetscLogAllBegin() 0 411 #define PetscLogDump(c) 0 412 #define PetscLogEventRegister(a,b,c) 0 413 EXTERN int PetscLogObjectState(PetscObject,const char[],...) PETSC_PRINTF_FORMAT_CHECK(2,3); 414 415 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 416 #define MPI_Startall_irecv(count,number,requests) MPI_Startall(number,requests) 417 #define MPI_Startall_isend(count,number,requests) MPI_Startall(number,requests) 418 #define MPI_Start_isend(count,requests) MPI_Start(requests) 419 420 #endif /* PETSC_USE_LOG */ 421 422 extern PetscTruth PetscPreLoadingUsed; /* true if we are or have done preloading */ 423 extern PetscTruth PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 424 425 #define PreLoadBegin(flag,name) \ 426 {\ 427 PetscTruth PreLoading = flag;\ 428 int PreLoadMax,PreLoadIt,_stageNum,_3_ierr;\ 429 _3_ierr = PetscOptionsGetLogical(PETSC_NULL,"-preload",&PreLoading,PETSC_NULL);CHKERRQ(_3_ierr);\ 430 PreLoadMax = (int)(PreLoading);\ 431 PetscPreLoadingUsed = PreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 432 for (PreLoadIt=0; PreLoadIt<=PreLoadMax; PreLoadIt++) {\ 433 PetscPreLoadingOn = PreLoading;\ 434 _3_ierr = PetscBarrier(PETSC_NULL);CHKERRQ(_3_ierr);\ 435 if (PreLoadIt>0) {\ 436 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 437 } else {\ 438 _3_ierr = PetscLogStageRegister(&_stageNum,name);CHKERRQ(_3_ierr);\ 439 }\ 440 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscTruth)(!PreLoadMax || PreLoadIt));\ 441 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 442 443 #define PreLoadEnd() \ 444 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 445 PreLoading = PETSC_FALSE;\ 446 }\ 447 } 448 449 #define PreLoadStage(name) \ 450 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 451 if (PreLoadIt>0) {\ 452 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 453 } else {\ 454 _3_ierr = PetscLogStageRegister(&_stageNum,name);CHKERRQ(_3_ierr);\ 455 }\ 456 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscTruth)(!PreLoadMax || PreLoadIt));\ 457 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 458 #endif 459