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