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