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