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