1 /* 2 Defines profile/logging in PETSc. 3 */ 4 #ifndef PETSCLOG_H 5 #define PETSCLOG_H 6 7 #include <petscsys.h> 8 #include <petsctime.h> 9 #include <petscbt.h> 10 #include <petsclogtypes.h> 11 12 /* SUBMANSEC = Profiling */ 13 14 /* General logging of information; different from event logging */ 15 PETSC_EXTERN PetscErrorCode PetscInfo_Private(const char[], PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(3, 4); 16 #if defined(PETSC_USE_INFO) 17 #define PetscInfo(A, ...) PetscInfo_Private(PETSC_FUNCTION_NAME, ((PetscObject)A), __VA_ARGS__) 18 #else 19 #define PetscInfo(A, ...) PETSC_SUCCESS 20 #endif 21 22 #define PetscInfo1(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 23 #define PetscInfo2(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 24 #define PetscInfo3(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 25 #define PetscInfo4(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 26 #define PetscInfo5(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 27 #define PetscInfo6(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 28 #define PetscInfo7(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 29 #define PetscInfo8(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 30 #define PetscInfo9(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__) 31 32 /*E 33 PetscInfoCommFlag - Describes the method by which to filter information displayed by `PetscInfo()` by communicator size 34 35 Values: 36 + `PETSC_INFO_COMM_ALL` - Default uninitialized value. `PetscInfo()` will not filter based on 37 communicator size (i.e. will print for all communicators) 38 . `PETSC_INFO_COMM_NO_SELF` - `PetscInfo()` will NOT print for communicators with size = 1 (i.e. *_COMM_SELF) 39 - `PETSC_INFO_COMM_ONLY_SELF` - `PetscInfo()` will ONLY print for communicators with size = 1 40 41 Level: intermediate 42 43 Note: 44 Used as an input for `PetscInfoSetFilterCommSelf()` 45 46 .seealso: `PetscInfo()`, `PetscInfoSetFromOptions()`, `PetscInfoSetFilterCommSelf()` 47 E*/ 48 typedef enum { 49 PETSC_INFO_COMM_ALL = -1, 50 PETSC_INFO_COMM_NO_SELF = 0, 51 PETSC_INFO_COMM_ONLY_SELF = 1 52 } PetscInfoCommFlag; 53 54 PETSC_EXTERN const char *const PetscInfoCommFlags[]; 55 PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId); 56 PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId); 57 PETSC_EXTERN PetscErrorCode PetscInfoEnabled(PetscClassId, PetscBool *); 58 PETSC_EXTERN PetscErrorCode PetscInfoAllow(PetscBool); 59 PETSC_EXTERN PetscErrorCode PetscInfoSetFile(const char[], const char[]); 60 PETSC_EXTERN PetscErrorCode PetscInfoGetFile(char **, FILE **); 61 PETSC_EXTERN PetscErrorCode PetscInfoSetClasses(PetscBool, PetscInt, const char *const *); 62 PETSC_EXTERN PetscErrorCode PetscInfoGetClass(const char *, PetscBool *); 63 PETSC_EXTERN PetscErrorCode PetscInfoGetInfo(PetscBool *, PetscBool *, PetscBool *, PetscBool *, PetscInfoCommFlag *); 64 PETSC_EXTERN PetscErrorCode PetscInfoProcessClass(const char[], PetscInt, const PetscClassId[]); 65 PETSC_EXTERN PetscErrorCode PetscInfoSetFilterCommSelf(PetscInfoCommFlag); 66 PETSC_EXTERN PetscErrorCode PetscInfoSetFromOptions(PetscOptions); 67 PETSC_EXTERN PetscErrorCode PetscInfoDestroy(void); 68 PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */ 69 70 #define PETSC_EVENT 1311311 71 PETSC_EXTERN PetscLogEvent PETSC_LARGEST_EVENT; 72 73 PETSC_EXTERN PetscErrorCode PetscIntStackCreate(PetscIntStack *); 74 PETSC_EXTERN PetscErrorCode PetscIntStackDestroy(PetscIntStack); 75 PETSC_EXTERN PetscErrorCode PetscIntStackPush(PetscIntStack, int); 76 PETSC_EXTERN PetscErrorCode PetscIntStackPop(PetscIntStack, int *); 77 PETSC_EXTERN PetscErrorCode PetscIntStackTop(PetscIntStack, int *); 78 PETSC_EXTERN PetscErrorCode PetscIntStackEmpty(PetscIntStack, PetscBool *); 79 80 PETSC_EXTERN PetscErrorCode PetscLogStateCreate(PetscLogState *); 81 PETSC_EXTERN PetscErrorCode PetscLogStateDestroy(PetscLogState *); 82 PETSC_EXTERN PetscErrorCode PetscLogStateGetRegistry(PetscLogState, PetscLogRegistry *); 83 84 PETSC_EXTERN PetscErrorCode PetscLogStateClassRegister(PetscLogState, const char[], PetscClassId, PetscLogStage *); 85 PETSC_EXTERN PetscErrorCode PetscLogStateClassSetActive(PetscLogState, PetscLogStage, PetscClassId, PetscBool); 86 PETSC_EXTERN PetscErrorCode PetscLogStateClassSetActiveAll(PetscLogState, PetscClassId, PetscBool); 87 88 PETSC_EXTERN PetscErrorCode PetscLogStateStageRegister(PetscLogState, const char[], PetscLogStage *); 89 PETSC_EXTERN PetscErrorCode PetscLogStateStagePush(PetscLogState, PetscLogStage); 90 PETSC_EXTERN PetscErrorCode PetscLogStateStagePop(PetscLogState); 91 PETSC_EXTERN PetscErrorCode PetscLogStateStageSetActive(PetscLogState, PetscLogStage, PetscBool); 92 PETSC_EXTERN PetscErrorCode PetscLogStateStageGetActive(PetscLogState, PetscLogStage, PetscBool *); 93 PETSC_EXTERN PetscErrorCode PetscLogStateGetCurrentStage(PetscLogState, PetscLogStage *); 94 95 PETSC_EXTERN PetscErrorCode PetscLogStateEventRegister(PetscLogState, const char[], PetscClassId, PetscLogEvent *); 96 PETSC_EXTERN PetscErrorCode PetscLogStateEventSetCollective(PetscLogState, PetscLogEvent, PetscBool); 97 PETSC_EXTERN PetscErrorCode PetscLogStateEventSetActive(PetscLogState, PetscLogStage, PetscLogEvent, PetscBool); 98 PETSC_EXTERN PetscErrorCode PetscLogStateEventSetActiveAll(PetscLogState, PetscLogEvent, PetscBool); 99 PETSC_EXTERN PetscErrorCode PetscLogStateEventGetActive(PetscLogState, PetscLogStage, PetscLogEvent, PetscBool *); 100 101 PETSC_EXTERN PetscErrorCode PetscLogStateGetEventFromName(PetscLogState, const char[], PetscLogEvent *); 102 PETSC_EXTERN PetscErrorCode PetscLogStateGetStageFromName(PetscLogState, const char[], PetscLogStage *); 103 PETSC_EXTERN PetscErrorCode PetscLogStateGetClassFromName(PetscLogState, const char[], PetscLogClass *); 104 PETSC_EXTERN PetscErrorCode PetscLogStateGetClassFromClassId(PetscLogState, PetscClassId, PetscLogClass *); 105 PETSC_EXTERN PetscErrorCode PetscLogStateGetNumEvents(PetscLogState, PetscInt *); 106 PETSC_EXTERN PetscErrorCode PetscLogStateGetNumStages(PetscLogState, PetscInt *); 107 PETSC_EXTERN PetscErrorCode PetscLogStateGetNumClasses(PetscLogState, PetscInt *); 108 PETSC_EXTERN PetscErrorCode PetscLogStateEventGetInfo(PetscLogState, PetscLogEvent, PetscLogEventInfo *); 109 PETSC_EXTERN PetscErrorCode PetscLogStateStageGetInfo(PetscLogState, PetscLogStage, PetscLogStageInfo *); 110 PETSC_EXTERN PetscErrorCode PetscLogStateClassGetInfo(PetscLogState, PetscLogClass, PetscLogClassInfo *); 111 112 PETSC_EXTERN PetscClassId PETSCLOGHANDLER_CLASSID; 113 114 PETSC_EXTERN PetscFunctionList PetscLogHandlerList; 115 116 PETSC_EXTERN PetscErrorCode PetscLogHandlerRegister(const char[], PetscErrorCode (*)(PetscLogHandler)); 117 PETSC_EXTERN PetscErrorCode PetscLogHandlerCreate(MPI_Comm, PetscLogHandler *); 118 PETSC_EXTERN PetscErrorCode PetscLogHandlerSetType(PetscLogHandler, PetscLogHandlerType); 119 PETSC_EXTERN PetscErrorCode PetscLogHandlerGetType(PetscLogHandler, PetscLogHandlerType *); 120 PETSC_EXTERN PetscErrorCode PetscLogHandlerDestroy(PetscLogHandler *); 121 PETSC_EXTERN PetscErrorCode PetscLogHandlerSetState(PetscLogHandler, PetscLogState); 122 PETSC_EXTERN PetscErrorCode PetscLogHandlerGetState(PetscLogHandler, PetscLogState *); 123 PETSC_EXTERN PetscErrorCode PetscLogHandlerEventBegin(PetscLogHandler, PetscLogEvent, PetscObject, PetscObject, PetscObject, PetscObject); 124 PETSC_EXTERN PetscErrorCode PetscLogHandlerEventEnd(PetscLogHandler, PetscLogEvent, PetscObject, PetscObject, PetscObject, PetscObject); 125 PETSC_EXTERN PetscErrorCode PetscLogHandlerEventSync(PetscLogHandler, PetscLogEvent, MPI_Comm); 126 PETSC_EXTERN PetscErrorCode PetscLogHandlerObjectCreate(PetscLogHandler, PetscObject); 127 PETSC_EXTERN PetscErrorCode PetscLogHandlerObjectDestroy(PetscLogHandler, PetscObject); 128 PETSC_EXTERN PetscErrorCode PetscLogHandlerStagePush(PetscLogHandler, PetscLogStage); 129 PETSC_EXTERN PetscErrorCode PetscLogHandlerStagePop(PetscLogHandler, PetscLogStage); 130 PETSC_EXTERN PetscErrorCode PetscLogHandlerView(PetscLogHandler, PetscViewer); 131 132 PETSC_EXTERN PetscErrorCode PetscLogHandlerCreateTrace(MPI_Comm, FILE *, PetscLogHandler *); 133 134 /* All events are inactive if an invalid stage is set, like if there have been more stage pops than stage pushes */ 135 #define PetscLogStateStageEventIsActive(state, stage, event) ((stage >= 0) && PetscBTLookup((state)->active, (stage)) && PetscBTLookup((state)->active, (stage) + (event + 1) * (state)->bt_num_stages)) 136 137 /* Handle multithreading */ 138 #if defined(PETSC_HAVE_THREADSAFETY) 139 #if defined(__cplusplus) 140 #define PETSC_TLS thread_local 141 #else 142 #define PETSC_TLS _Thread_local 143 #endif 144 #define PETSC_EXTERN_TLS extern PETSC_TLS PETSC_VISIBILITY_PUBLIC 145 PETSC_EXTERN PetscErrorCode PetscAddLogDouble(PetscLogDouble *, PetscLogDouble *, PetscLogDouble); 146 PETSC_EXTERN PetscErrorCode PetscAddLogDoubleCnt(PetscLogDouble *, PetscLogDouble *, PetscLogDouble *, PetscLogDouble *, PetscLogDouble); 147 #else 148 #define PETSC_EXTERN_TLS PETSC_EXTERN 149 #define PETSC_TLS 150 #define PetscAddLogDouble(a, b, c) ((PetscErrorCode)((*(a) += (c), PETSC_SUCCESS) || ((*(b) += (c)), PETSC_SUCCESS))) 151 #define PetscAddLogDoubleCnt(a, b, c, d, e) ((PetscErrorCode)(PetscAddLogDouble(a, c, 1) || PetscAddLogDouble(b, d, e))) 152 #endif 153 154 /* 155 PetscClassRegInfo, PetscClassPerfInfo - Each class has two data structures associated with it. The first has 156 static information about it, the second collects statistics on how many objects of the class are created, 157 how much memory they use, etc. 158 159 PetscClassRegLog, PetscClassPerfLog - arrays of the PetscClassRegInfo and PetscClassPerfInfo for all classes. 160 */ 161 typedef struct { 162 char *name; /* The class name */ 163 PetscClassId classid; /* The integer identifying this class */ 164 } PetscClassRegInfo; 165 166 typedef struct { 167 PetscClassId id; /* The integer identifying this class */ 168 int creations; /* The number of objects of this class created */ 169 int destructions; /* The number of objects of this class destroyed */ 170 PetscLogDouble mem; /* The total memory allocated by objects of this class; this is completely wrong and should possibly be removed */ 171 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects; this is completely wrong and should possibly be removed */ 172 } PetscClassPerfInfo; 173 174 typedef struct _n_PetscClassRegLog *PetscClassRegLog; 175 struct _n_PetscClassRegLog { 176 int numClasses; /* The number of classes registered */ 177 int maxClasses; /* The maximum number of classes */ 178 PetscClassRegInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 179 }; 180 181 typedef struct _n_PetscClassPerfLog *PetscClassPerfLog; 182 struct _n_PetscClassPerfLog { 183 int numClasses; /* The number of logging classes */ 184 int maxClasses; /* The maximum number of classes */ 185 PetscClassPerfInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 186 }; 187 188 /* 189 PetscEventRegInfo, PetscEventPerfInfo - Each event has two data structures associated with it. The first has 190 static information about it, the second collects statistics on how many times the event is used, how 191 much time it takes, etc. 192 193 PetscEventRegLog, PetscEventPerfLog - an array of all PetscEventRegInfo and PetscEventPerfInfo for all events. There is one 194 of these for each stage. 195 196 */ 197 typedef struct { 198 char *name; /* The name of this event */ 199 PetscClassId classid; /* The class the event is associated with */ 200 PetscBool collective; /* Flag this event as collective */ 201 #if defined(PETSC_HAVE_TAU_PERFSTUBS) 202 void *timer; /* Associated external tool timer for this event */ 203 #endif 204 #if defined(PETSC_HAVE_MPE) 205 int mpe_id_begin; /* MPE IDs that define the event */ 206 int mpe_id_end; 207 #endif 208 } PetscEventRegInfo; 209 210 typedef struct _n_PetscEventRegLog *PetscEventRegLog; 211 struct _n_PetscEventRegLog { 212 int numEvents; /* The number of registered events */ 213 int maxEvents; /* The maximum number of events */ 214 PetscEventRegInfo *eventInfo; /* The registration information for each event */ 215 }; 216 217 typedef struct _n_PetscEventPerfLog *PetscEventPerfLog; 218 struct _n_PetscEventPerfLog { 219 int numEvents; /* The number of logging events */ 220 int maxEvents; /* The maximum number of events */ 221 PetscEventPerfInfo *eventInfo; /* The performance information for each event */ 222 }; 223 224 /* 225 PetscStageInfo - Contains all the information about a particular stage. 226 227 PetscStageLog - An array of PetscStageInfo for each registered stage. There is a single one of these in the code. 228 */ 229 typedef struct _PetscStageInfo { 230 char *name; /* The stage name */ 231 PetscBool used; /* The stage was pushed on this processor */ 232 PetscEventPerfInfo perfInfo; /* The stage performance information */ 233 PetscEventPerfLog eventLog; /* The event information for this stage */ 234 PetscClassPerfLog classLog; /* The class information for this stage */ 235 #if defined(PETSC_HAVE_TAU_PERFSTUBS) 236 void *timer; /* Associated external tool timer for this stage */ 237 #endif 238 } PetscStageInfo; 239 240 typedef struct _n_PetscStageLog *PetscStageLog; 241 struct _n_PetscStageLog { 242 int numStages; /* The number of registered stages */ 243 int maxStages; /* The maximum number of stages */ 244 PetscIntStack stack; /* The stack for active stages */ 245 int curStage; /* The current stage (only used in macros so we don't call PetscIntStackTop) */ 246 PetscStageInfo *stageInfo; /* The information for each stage */ 247 PetscEventRegLog eventLog; /* The registered events */ 248 PetscClassRegLog classLog; /* The registered classes */ 249 }; 250 251 PETSC_DEPRECATED_FUNCTION(3, 18, 0, "PetscLogObjectParent()", ) static inline PetscErrorCode PetscLogObjectParent(PetscObject o, PetscObject p) 252 { 253 (void)o; 254 (void)p; 255 return PETSC_SUCCESS; 256 } 257 258 PETSC_DEPRECATED_FUNCTION(3, 18, 0, "PetscLogObjectMemory()", ) static inline PetscErrorCode PetscLogObjectMemory(PetscObject o, PetscLogDouble m) 259 { 260 (void)o; 261 (void)m; 262 return PETSC_SUCCESS; 263 } 264 265 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 266 PETSC_EXTERN PetscStageLog petsc_stageLog; 267 PETSC_EXTERN PetscErrorCode PetscLogGetStageLog(PetscStageLog *); 268 PETSC_EXTERN PetscErrorCode PetscStageLogGetCurrent(PetscStageLog, int *); 269 PETSC_EXTERN PetscErrorCode PetscStageLogGetEventPerfLog(PetscStageLog, int, PetscEventPerfLog *); 270 271 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 272 273 #if defined(PETSC_HAVE_MPE) 274 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 275 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 276 #endif 277 278 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject); 279 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject); 280 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 281 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 282 283 #define PetscLogObjectParents(p, n, d) PetscMacroReturnStandard(for (int _i = 0; _i < (n); ++_i) PetscCall(PetscLogObjectParent((PetscObject)(p), (PetscObject)(d)[_i]));) 284 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : PETSC_SUCCESS) 285 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : PETSC_SUCCESS) 286 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2, 3); 287 288 /* Initialization functions */ 289 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 290 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 291 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 292 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 293 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 294 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 295 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble, PetscLogDouble *); 296 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 297 298 /* Output functions */ 299 PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer); 300 PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void); 301 PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]); 302 303 /* Status checking functions */ 304 PETSC_EXTERN PetscErrorCode PetscLogIsActive(PetscBool *); 305 306 /* Stage functions */ 307 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[], PetscLogStage *); 308 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 309 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 310 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage, PetscBool); 311 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage, PetscBool *); 312 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage, PetscBool); 313 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage, PetscBool *); 314 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[], PetscLogStage *); 315 316 /* Event functions */ 317 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[], PetscClassId, PetscLogEvent *); 318 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent, PetscBool); 319 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 320 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 321 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 322 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 323 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePush(PetscLogEvent); 324 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePop(PetscLogEvent); 325 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent, PetscBool); 326 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 327 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 328 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[], PetscLogEvent *); 329 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int, PetscLogEvent, PetscEventPerfInfo *); 330 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 331 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 332 PETSC_EXTERN PetscErrorCode PetscLogPushCurrentEvent_Internal(PetscLogEvent); 333 PETSC_EXTERN PetscErrorCode PetscLogPopCurrentEvent_Internal(void); 334 335 PETSC_EXTERN PetscBool PetscLogMemory; 336 337 PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */ 338 PETSC_EXTERN PetscErrorCode PetscLogEventSynchronize(PetscLogEvent, MPI_Comm); 339 340 #define PetscLogEventSync(e, comm) \ 341 ((PetscErrorCode)(((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? PetscLogEventSynchronize((e), (comm)) : PETSC_SUCCESS))) 342 343 #define PetscLogEventBegin(e, o1, o2, o3, o4) \ 344 ((PetscErrorCode)((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? (PetscErrorCode)(((*PetscLogPLB)((e), 0, (PetscObject)(o1), (PetscObject)(o2), (PetscObject)(o3), (PetscObject)(o4))) || PetscLogPushCurrentEvent_Internal(e)) : PETSC_SUCCESS)) 345 346 #define PetscLogEventEnd(e, o1, o2, o3, o4) \ 347 ((PetscErrorCode)((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? (PetscErrorCode)(((*PetscLogPLE)((e), 0, (PetscObject)(o1), (PetscObject)(o2), (PetscObject)(o3), (PetscObject)(o4))) || PetscLogPopCurrentEvent_Internal()) : PETSC_SUCCESS)) 348 349 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent, PetscLogDouble *); 350 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 351 352 /* Global flop counter */ 353 PETSC_EXTERN PetscLogDouble petsc_TotalFlops; 354 PETSC_EXTERN PetscLogDouble petsc_irecv_ct; 355 PETSC_EXTERN PetscLogDouble petsc_isend_ct; 356 PETSC_EXTERN PetscLogDouble petsc_recv_ct; 357 PETSC_EXTERN PetscLogDouble petsc_send_ct; 358 PETSC_EXTERN PetscLogDouble petsc_irecv_len; 359 PETSC_EXTERN PetscLogDouble petsc_isend_len; 360 PETSC_EXTERN PetscLogDouble petsc_recv_len; 361 PETSC_EXTERN PetscLogDouble petsc_send_len; 362 PETSC_EXTERN PetscLogDouble petsc_allreduce_ct; 363 PETSC_EXTERN PetscLogDouble petsc_gather_ct; 364 PETSC_EXTERN PetscLogDouble petsc_scatter_ct; 365 PETSC_EXTERN PetscLogDouble petsc_wait_ct; 366 PETSC_EXTERN PetscLogDouble petsc_wait_any_ct; 367 PETSC_EXTERN PetscLogDouble petsc_wait_all_ct; 368 PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct; 369 370 /* Thread local storage */ 371 PETSC_EXTERN_TLS PetscLogDouble petsc_TotalFlops_th; 372 PETSC_EXTERN_TLS PetscLogDouble petsc_irecv_ct_th; 373 PETSC_EXTERN_TLS PetscLogDouble petsc_isend_ct_th; 374 PETSC_EXTERN_TLS PetscLogDouble petsc_recv_ct_th; 375 PETSC_EXTERN_TLS PetscLogDouble petsc_send_ct_th; 376 PETSC_EXTERN_TLS PetscLogDouble petsc_irecv_len_th; 377 PETSC_EXTERN_TLS PetscLogDouble petsc_isend_len_th; 378 PETSC_EXTERN_TLS PetscLogDouble petsc_recv_len_th; 379 PETSC_EXTERN_TLS PetscLogDouble petsc_send_len_th; 380 PETSC_EXTERN_TLS PetscLogDouble petsc_allreduce_ct_th; 381 PETSC_EXTERN_TLS PetscLogDouble petsc_gather_ct_th; 382 PETSC_EXTERN_TLS PetscLogDouble petsc_scatter_ct_th; 383 PETSC_EXTERN_TLS PetscLogDouble petsc_wait_ct_th; 384 PETSC_EXTERN_TLS PetscLogDouble petsc_wait_any_ct_th; 385 PETSC_EXTERN_TLS PetscLogDouble petsc_wait_all_ct_th; 386 PETSC_EXTERN_TLS PetscLogDouble petsc_sum_of_waits_ct_th; 387 388 /* 389 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 390 391 For the complex numbers version, note that 392 1 complex addition = 2 flops 393 1 complex multiplication = 6 flops, 394 where we define 1 flop as that for a double precision scalar. We roughly approximate 395 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 396 to the assumption that we're counting mostly additions and multiplications -- and 397 roughly the same number of each. More accurate counting could be done by distinguishing 398 among the various arithmetic operations. 399 */ 400 401 #if defined(PETSC_USE_COMPLEX) 402 #define PETSC_FLOPS_PER_OP 4.0 403 #else 404 #define PETSC_FLOPS_PER_OP 1.0 405 #endif 406 407 /*@C 408 PetscLogFlops - Log how many flops are performed in a calculation 409 410 Input Parameter: 411 . flops - the number of flops 412 413 Level: intermediate 414 415 Note: 416 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 417 not an integer. Use `PetscLogFlops`(4.0*n) not `PetscLogFlops`(4*n) 418 419 .seealso: [](ch_profiling), `PetscLogView()`, `PetscLogGpuFlops()` 420 @*/ 421 static inline PetscErrorCode PetscLogFlops(PetscLogDouble n) 422 { 423 PetscAssert(n >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Cannot log negative flops"); 424 return PetscAddLogDouble(&petsc_TotalFlops, &petsc_TotalFlops_th, PETSC_FLOPS_PER_OP * n); 425 } 426 427 /* 428 These are used internally in the PETSc routines to keep a count of MPI messages and 429 their sizes. 430 431 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 432 uses macros to defined the MPI operations. 433 434 It does not work correctly from HP-UX because it processes the 435 macros in a way that sometimes it double counts, hence 436 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 437 438 It does not work with Windows because winmpich lacks MPI_Type_size() 439 */ 440 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) 441 /* 442 Logging of MPI activities 443 */ 444 static inline PetscErrorCode PetscMPITypeSize(PetscInt count, MPI_Datatype type, PetscLogDouble *length, PetscLogDouble *length_th) 445 { 446 PetscMPIInt typesize; 447 448 if (type == MPI_DATATYPE_NULL) return PETSC_SUCCESS; 449 PetscCallMPI(MPI_Type_size(type, &typesize)); 450 return PetscAddLogDouble(length, length_th, (PetscLogDouble)(count * typesize)); 451 } 452 453 static inline PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm, const PetscMPIInt *counts, MPI_Datatype type, PetscLogDouble *length, PetscLogDouble *length_th) 454 { 455 PetscMPIInt typesize, size, p; 456 PetscLogDouble l; 457 458 if (type == MPI_DATATYPE_NULL) return PETSC_SUCCESS; 459 PetscCallMPI(MPI_Comm_size(comm, &size)); 460 PetscCallMPI(MPI_Type_size(type, &typesize)); 461 for (p = 0, l = 0.0; p < size; ++p) l += (PetscLogDouble)(counts[p] * typesize); 462 return PetscAddLogDouble(length, length_th, l); 463 } 464 465 /* 466 Returns 1 if the communicator is parallel else zero 467 */ 468 static inline int PetscMPIParallelComm(MPI_Comm comm) 469 { 470 PetscMPIInt size; 471 MPI_Comm_size(comm, &size); 472 return size > 1; 473 } 474 475 #define MPI_Irecv(buf, count, datatype, source, tag, comm, request) \ 476 (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || MPI_Irecv((buf), (count), (datatype), (source), (tag), (comm), (request))) 477 478 #define MPI_Irecv_c(buf, count, datatype, source, tag, comm, request) \ 479 (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || MPI_Irecv_c((buf), (count), (datatype), (source), (tag), (comm), (request))) 480 481 #define MPI_Isend(buf, count, datatype, dest, tag, comm, request) \ 482 (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || MPI_Isend((buf), (count), (datatype), (dest), (tag), (comm), (request))) 483 484 #define MPI_Isend_c(buf, count, datatype, dest, tag, comm, request) \ 485 (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || MPI_Isend_c((buf), (count), (datatype), (dest), (tag), (comm), (request))) 486 487 #define MPI_Startall_irecv(count, datatype, number, requests) \ 488 (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, number) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || ((number) && MPI_Startall((number), (requests)))) 489 490 #define MPI_Startall_isend(count, datatype, number, requests) \ 491 (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, number) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || ((number) && MPI_Startall((number), (requests)))) 492 493 #define MPI_Start_isend(count, datatype, requests) (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_isend_len), (&petsc_isend_len_th)) || MPI_Start((requests))) 494 495 #define MPI_Recv(buf, count, datatype, source, tag, comm, status) \ 496 (PetscAddLogDouble(&petsc_recv_ct, &petsc_recv_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPI_Recv((buf), (count), (datatype), (source), (tag), (comm), (status))) 497 498 #define MPI_Recv_c(buf, count, datatype, source, tag, comm, status) \ 499 (PetscAddLogDouble(&petsc_recv_ct, &petsc_recv_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len), &(petsc_recv_len_th)) || MPI_Recv_c((buf), (count), (datatype), (source), (tag), (comm), (status))) 500 501 #define MPI_Send(buf, count, datatype, dest, tag, comm) \ 502 (PetscAddLogDouble(&petsc_send_ct, &petsc_send_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Send((buf), (count), (datatype), (dest), (tag), (comm))) 503 504 #define MPI_Send_c(buf, count, datatype, dest, tag, comm) \ 505 (PetscAddLogDouble(&petsc_send_ct, &petsc_send_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Send_c((buf), (count), (datatype), (dest), (tag), (comm))) 506 507 #define MPI_Wait(request, status) (PetscAddLogDouble(&petsc_wait_ct, &petsc_wait_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, 1) || MPI_Wait((request), (status))) 508 509 #define MPI_Waitany(a, b, c, d) (PetscAddLogDouble(&petsc_wait_any_ct, &petsc_wait_any_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, 1) || MPI_Waitany((a), (b), (c), (d))) 510 511 #define MPI_Waitall(count, array_of_requests, array_of_statuses) \ 512 (PetscAddLogDouble(&petsc_wait_all_ct, &petsc_wait_all_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, count) || MPI_Waitall((count), (array_of_requests), (array_of_statuses))) 513 514 #define MPI_Allreduce(sendbuf, recvbuf, count, datatype, op, comm) (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPI_Allreduce((sendbuf), (recvbuf), (count), (datatype), (op), (comm))) 515 516 #define MPI_Bcast(buffer, count, datatype, root, comm) (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPI_Bcast((buffer), (count), (datatype), (root), (comm))) 517 518 #define MPI_Reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm) \ 519 (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPI_Reduce_scatter_block((sendbuf), (recvbuf), (recvcount), (datatype), (op), (comm))) 520 521 #define MPI_Alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm) \ 522 (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Alltoall((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm))) 523 524 #define MPI_Alltoallv(sendbuf, sendcnts, sdispls, sendtype, recvbuf, recvcnts, rdispls, recvtype, comm) \ 525 (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSizeComm((comm), (sendcnts), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Alltoallv((sendbuf), (sendcnts), (sdispls), (sendtype), (recvbuf), (recvcnts), (rdispls), (recvtype), (comm))) 526 527 #define MPI_Allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm) \ 528 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Allgather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm))) 529 530 #define MPI_Allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, comm) \ 531 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Allgatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (comm))) 532 533 #define MPI_Gather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm) \ 534 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Gather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm))) 535 536 #define MPI_Gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, root, comm) \ 537 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Gatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (root), (comm))) 538 539 #define MPI_Scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm) \ 540 (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), &(petsc_recv_len_th)) || MPI_Scatter((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm))) 541 542 #define MPI_Scatterv(sendbuf, sendcount, displs, sendtype, recvbuf, recvcount, recvtype, root, comm) \ 543 (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), &(petsc_recv_len_th)) || MPI_Scatterv((sendbuf), (sendcount), (displs), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm))) 544 545 #define MPI_Ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request) \ 546 (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Ialltoall((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm), (request))) 547 548 #define MPI_Ialltoallv(sendbuf, sendcnts, sdispls, sendtype, recvbuf, recvcnts, rdispls, recvtype, comm, request) \ 549 (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSizeComm((comm), (sendcnts), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Ialltoallv((sendbuf), (sendcnts), (sdispls), (sendtype), (recvbuf), (recvcnts), (rdispls), (recvtype), (comm), (request))) 550 551 #define MPI_Iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request) \ 552 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Iallgather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm), (request))) 553 554 #define MPI_Iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, comm, request) \ 555 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Iallgatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (comm), (request))) 556 557 #define MPI_Igather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \ 558 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Igather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request))) 559 560 #define MPI_Igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, root, comm, request) \ 561 (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Igatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (root), (comm), (request))) 562 563 #define MPI_Iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \ 564 (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPI_Iscatter((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request))) 565 566 #define MPI_Iscatterv(sendbuf, sendcount, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \ 567 (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPI_Iscatterv((sendbuf), (sendcount), (displs), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request))) 568 569 #define MPIX_Send_enqueue(buf, count, datatype, dest, tag, comm) \ 570 (PetscAddLogDouble(&petsc_send_ct, &petsc_send_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_send_len), (&petsc_send_len_th)) || MPIX_Send_enqueue((buf), (count), (datatype), (dest), (tag), (comm))) 571 572 #define MPIX_Recv_enqueue(buf, count, datatype, source, tag, comm, status) \ 573 (PetscAddLogDouble(&petsc_recv_ct, &petsc_recv_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPIX_Recv_enqueue((buf), (count), (datatype), (source), (tag), (comm), (status))) 574 575 #define MPIX_Isend_enqueue(buf, count, datatype, dest, tag, comm, request) \ 576 (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || MPIX_Isend_enqueue((buf), (count), (datatype), (dest), (tag), (comm), (request))) 577 578 #define MPIX_Irecv_enqueue(buf, count, datatype, source, tag, comm, request) \ 579 (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || MPIX_Irecv_enqueue((buf), (count), (datatype), (source), (tag), (comm), (request))) 580 581 #define MPIX_Allreduce_enqueue(sendbuf, recvbuf, count, datatype, op, comm) \ 582 (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPIX_Allreduce_enqueue((sendbuf), (recvbuf), (count), (datatype), (op), (comm))) 583 584 #define MPIX_Wait_enqueue(request, status) (PetscAddLogDouble(&petsc_wait_ct, &petsc_wait_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, 1) || MPIX_Wait_enqueue((request), (status))) 585 586 #define MPIX_Waitall_enqueue(count, array_of_requests, array_of_statuses) \ 587 (PetscAddLogDouble(&petsc_wait_all_ct, &petsc_wait_all_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, count) || MPIX_Waitall_enqueue((count), (array_of_requests), (array_of_statuses))) 588 #else 589 590 #define MPI_Startall_irecv(count, datatype, number, requests) ((number) && MPI_Startall((number), (requests))) 591 592 #define MPI_Startall_isend(count, datatype, number, requests) ((number) && MPI_Startall((number), (requests))) 593 594 #define MPI_Start_isend(count, datatype, requests) (MPI_Start((requests))) 595 596 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 597 598 #else /* ---Logging is turned off --------------------------------------------*/ 599 600 #define PetscLogMemory PETSC_FALSE 601 602 #define PetscLogFlops(n) ((void)(n), PETSC_SUCCESS) 603 #define PetscGetFlops(a) (*(a) = 0.0, PETSC_SUCCESS) 604 605 #define PetscLogStageRegister(a, b) PETSC_SUCCESS 606 #define PetscLogStagePush(a) PETSC_SUCCESS 607 #define PetscLogStagePop() PETSC_SUCCESS 608 #define PetscLogStageSetActive(a, b) PETSC_SUCCESS 609 #define PetscLogStageGetActive(a, b) PETSC_SUCCESS 610 #define PetscLogStageGetVisible(a, b) PETSC_SUCCESS 611 #define PetscLogStageSetVisible(a, b) PETSC_SUCCESS 612 #define PetscLogStageGetId(a, b) (*(b) = 0, PETSC_SUCCESS) 613 614 #define PetscLogEventRegister(a, b, c) PETSC_SUCCESS 615 #define PetscLogEventSetCollective(a, b) PETSC_SUCCESS 616 #define PetscLogEventIncludeClass(a) PETSC_SUCCESS 617 #define PetscLogEventExcludeClass(a) PETSC_SUCCESS 618 #define PetscLogEventActivate(a) PETSC_SUCCESS 619 #define PetscLogEventDeactivate(a) PETSC_SUCCESS 620 #define PetscLogEventDeactivatePush(a) PETSC_SUCCESS 621 #define PetscLogEventDeactivatePop(a) PETSC_SUCCESS 622 #define PetscLogEventActivateClass(a) PETSC_SUCCESS 623 #define PetscLogEventDeactivateClass(a) PETSC_SUCCESS 624 #define PetscLogEventSetActiveAll(a, b) PETSC_SUCCESS 625 #define PetscLogEventGetId(a, b) (*(b) = 0, PETSC_SUCCESS) 626 #define PetscLogEventGetPerfInfo(a, b, c) PETSC_SUCCESS 627 #define PetscLogEventSetDof(a, b, c) PETSC_SUCCESS 628 #define PetscLogEventSetError(a, b, c) PETSC_SUCCESS 629 630 #define PetscLogPLB PETSC_SUCCESS 631 #define PetscLogPLE PETSC_SUCCESS 632 #define PetscLogPHC PETSC_SUCCESS 633 #define PetscLogPHD PETSC_SUCCESS 634 635 #define PetscLogObjectParents(p, n, c) PETSC_SUCCESS 636 #define PetscLogObjectCreate(h) PETSC_SUCCESS 637 #define PetscLogObjectDestroy(h) PETSC_SUCCESS 638 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2, 3); 639 640 #define PetscLogDefaultBegin() PETSC_SUCCESS 641 #define PetscLogAllBegin() PETSC_SUCCESS 642 #define PetscLogNestedBegin() PETSC_SUCCESS 643 #define PetscLogTraceBegin(file) PETSC_SUCCESS 644 #define PetscLogActions(a) PETSC_SUCCESS 645 #define PetscLogObjects(a) PETSC_SUCCESS 646 #define PetscLogSetThreshold(a, b) PETSC_SUCCESS 647 #define PetscLogSet(lb, le) PETSC_SUCCESS 648 #define PetscLogIsActive(flag) (*(flag) = PETSC_FALSE, PETSC_SUCCESS) 649 650 #define PetscLogView(viewer) PETSC_SUCCESS 651 #define PetscLogViewFromOptions() PETSC_SUCCESS 652 #define PetscLogDump(c) PETSC_SUCCESS 653 654 #define PetscLogEventSync(e, comm) PETSC_SUCCESS 655 #define PetscLogEventBegin(e, o1, o2, o3, o4) PETSC_SUCCESS 656 #define PetscLogEventEnd(e, o1, o2, o3, o4) PETSC_SUCCESS 657 658 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 659 #define MPI_Startall_irecv(count, datatype, number, requests) ((number) && MPI_Startall(number, requests)) 660 #define MPI_Startall_isend(count, datatype, number, requests) ((number) && MPI_Startall(number, requests)) 661 #define MPI_Start_isend(count, datatype, requests) MPI_Start(requests) 662 663 #endif /* PETSC_USE_LOG */ 664 665 #define PetscPreLoadBegin(flag, name) \ 666 do { \ 667 PetscBool PetscPreLoading = flag; \ 668 int PetscPreLoadMax, PetscPreLoadIt; \ 669 PetscLogStage _stageNum; \ 670 PetscCall(PetscOptionsGetBool(NULL, NULL, "-preload", &PetscPreLoading, NULL)); \ 671 PetscPreLoadMax = (int)(PetscPreLoading); \ 672 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed; \ 673 for (PetscPreLoadIt = 0; PetscPreLoadIt <= PetscPreLoadMax; PetscPreLoadIt++) { \ 674 PetscPreLoadingOn = PetscPreLoading; \ 675 PetscCall(PetscBarrier(NULL)); \ 676 if (PetscPreLoadIt > 0) PetscCall(PetscLogStageGetId(name, &_stageNum)); \ 677 else PetscCall(PetscLogStageRegister(name, &_stageNum)); \ 678 PetscCall(PetscLogStageSetActive(_stageNum, (PetscBool)(!PetscPreLoadMax || PetscPreLoadIt))); \ 679 PetscCall(PetscLogStagePush(_stageNum)) 680 681 #define PetscPreLoadEnd() \ 682 PetscCall(PetscLogStagePop()); \ 683 PetscPreLoading = PETSC_FALSE; \ 684 } \ 685 } \ 686 while (0) 687 688 #define PetscPreLoadStage(name) \ 689 do { \ 690 PetscCall(PetscLogStagePop()); \ 691 if (PetscPreLoadIt > 0) PetscCall(PetscLogStageGetId(name, &_stageNum)); \ 692 else PetscCall(PetscLogStageRegister(name, &_stageNum)); \ 693 PetscCall(PetscLogStageSetActive(_stageNum, (PetscBool)(!PetscPreLoadMax || PetscPreLoadIt))); \ 694 PetscCall(PetscLogStagePush(_stageNum)); \ 695 } while (0) 696 697 /* some vars for logging */ 698 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 699 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 700 701 #if defined(PETSC_USE_LOG) && defined(PETSC_HAVE_DEVICE) 702 703 /* Global GPU counters */ 704 PETSC_EXTERN PetscLogDouble petsc_ctog_ct; 705 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct; 706 PETSC_EXTERN PetscLogDouble petsc_ctog_sz; 707 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz; 708 PETSC_EXTERN PetscLogDouble petsc_ctog_ct_scalar; 709 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct_scalar; 710 PETSC_EXTERN PetscLogDouble petsc_ctog_sz_scalar; 711 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz_scalar; 712 PETSC_EXTERN PetscLogDouble petsc_gflops; 713 PETSC_EXTERN PetscLogDouble petsc_gtime; 714 715 /* Thread local storage */ 716 PETSC_EXTERN_TLS PetscLogDouble petsc_ctog_ct_th; 717 PETSC_EXTERN_TLS PetscLogDouble petsc_gtoc_ct_th; 718 PETSC_EXTERN_TLS PetscLogDouble petsc_ctog_sz_th; 719 PETSC_EXTERN_TLS PetscLogDouble petsc_gtoc_sz_th; 720 PETSC_EXTERN_TLS PetscLogDouble petsc_ctog_ct_scalar_th; 721 PETSC_EXTERN_TLS PetscLogDouble petsc_gtoc_ct_scalar_th; 722 PETSC_EXTERN_TLS PetscLogDouble petsc_ctog_sz_scalar_th; 723 PETSC_EXTERN_TLS PetscLogDouble petsc_gtoc_sz_scalar_th; 724 PETSC_EXTERN_TLS PetscLogDouble petsc_gflops_th; 725 PETSC_EXTERN_TLS PetscLogDouble petsc_gtime_th; 726 727 PETSC_EXTERN PetscErrorCode PetscLogGpuTime(void); 728 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeBegin(void); 729 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeEnd(void); 730 731 /*@C 732 PetscLogGpuFlops - Log how many flops are performed in a calculation on the device 733 734 Input Parameter: 735 . flops - the number of flops 736 737 Level: intermediate 738 739 Notes: 740 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 741 not an integer. Use `PetscLogFlops`(4.0*n) not `PetscLogFlops`(4*n) 742 743 The values are also added to the total flop count for the MPI rank that is set with `PetscLogFlops()`; hence the number of flops 744 just on the CPU would be the value from set from `PetscLogFlops()` minus the value set from `PetscLogGpuFlops()` 745 746 .seealso: [](ch_profiling), `PetscLogView()`, `PetscLogFlops()`, `PetscLogGpuTimeBegin()`, `PetscLogGpuTimeEnd()` 747 @*/ 748 static inline PetscErrorCode PetscLogGpuFlops(PetscLogDouble n) 749 { 750 PetscAssert(n >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Cannot log negative flops"); 751 PetscCall(PetscAddLogDouble(&petsc_TotalFlops, &petsc_TotalFlops_th, PETSC_FLOPS_PER_OP * n)); 752 PetscCall(PetscAddLogDouble(&petsc_gflops, &petsc_gflops_th, PETSC_FLOPS_PER_OP * n)); 753 return PETSC_SUCCESS; 754 } 755 756 static inline PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t) 757 { 758 return PetscAddLogDouble(&petsc_gtime, &petsc_gtime_th, t); 759 } 760 761 static inline PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size) 762 { 763 return PetscAddLogDoubleCnt(&petsc_ctog_ct, &petsc_ctog_sz, &petsc_ctog_ct_th, &petsc_ctog_sz_th, size); 764 } 765 766 static inline PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size) 767 { 768 return PetscAddLogDoubleCnt(&petsc_gtoc_ct, &petsc_gtoc_sz, &petsc_gtoc_ct_th, &petsc_gtoc_sz_th, size); 769 } 770 771 static inline PetscErrorCode PetscLogCpuToGpuScalar(PetscLogDouble size) 772 { 773 return PetscAddLogDoubleCnt(&petsc_ctog_ct_scalar, &petsc_ctog_sz_scalar, &petsc_ctog_ct_scalar_th, &petsc_ctog_sz_scalar_th, size); 774 } 775 776 static inline PetscErrorCode PetscLogGpuToCpuScalar(PetscLogDouble size) 777 { 778 return PetscAddLogDoubleCnt(&petsc_gtoc_ct_scalar, &petsc_gtoc_sz_scalar, &petsc_gtoc_ct_scalar_th, &petsc_gtoc_sz_scalar_th, size); 779 } 780 #else 781 782 #define PetscLogCpuToGpu(a) PETSC_SUCCESS 783 #define PetscLogGpuToCpu(a) PETSC_SUCCESS 784 #define PetscLogCpuToGpuScalar(a) PETSC_SUCCESS 785 #define PetscLogGpuToCpuScalar(a) PETSC_SUCCESS 786 #define PetscLogGpuFlops(a) PETSC_SUCCESS 787 #define PetscLogGpuTime() PETSC_SUCCESS 788 #define PetscLogGpuTimeAdd(a) PETSC_SUCCESS 789 #define PetscLogGpuTimeBegin() PETSC_SUCCESS 790 #define PetscLogGpuTimeEnd() PETSC_SUCCESS 791 792 #endif /* PETSC_USE_LOG && PETSC_HAVE_DEVICE */ 793 794 /* remove TLS defines */ 795 #undef PETSC_EXTERN_TLS 796 #undef PETSC_TLS 797 798 #endif 799