1 /* 2 Defines profile/logging in PETSc. 3 */ 4 #if !defined(PETSCLOG_H) 5 #define PETSCLOG_H 6 7 #include <petscsys.h> 8 #include <petsctime.h> 9 10 /* SUBMANSEC = Sys */ 11 12 /* General logging of information; different from event logging */ 13 PETSC_EXTERN PetscErrorCode PetscInfo_Private(const char[], PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(3, 4); 14 #if defined(PETSC_USE_INFO) 15 #define PetscInfo(A, ...) PetscInfo_Private(PETSC_FUNCTION_NAME, ((PetscObject)A), __VA_ARGS__) 16 #else 17 #define PetscInfo(A, ...) 0 18 #endif 19 20 #define PetscInfo1(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 21 #define PetscInfo2(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 22 #define PetscInfo3(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 23 #define PetscInfo4(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 24 #define PetscInfo5(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 25 #define PetscInfo6(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 26 #define PetscInfo7(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 27 #define PetscInfo8(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 28 #define PetscInfo9(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 29 30 /*E 31 PetscInfoCommFlag - Describes the method by which to filter `PetscInfo()` by communicator size 32 33 Used as an input for PetscInfoSetFilterCommSelf() 34 35 $ `PETSC_INFO_COMM_ALL` - Default uninitialized value. PetscInfo() will not filter based on communicator size (i.e. will 36 print for all communicators) 37 $ `PETSC_INFO_COMM_NO_SELF` - PetscInfo() will NOT print for communicators with size = 1 (i.e. *_COMM_SELF) 38 $ `PETSC_INFO_COMM_ONLY_SELF` - PetscInfo will ONLY print for communicators with size = 1 39 40 Level: intermediate 41 42 .seealso: `PetscInfo()`, `PetscInfoSetFromOptions()`, `PetscInfoSetFilterCommSelf()` 43 E*/ 44 typedef enum { 45 PETSC_INFO_COMM_ALL = -1, 46 PETSC_INFO_COMM_NO_SELF = 0, 47 PETSC_INFO_COMM_ONLY_SELF = 1 48 } PetscInfoCommFlag; 49 50 PETSC_EXTERN const char *const PetscInfoCommFlags[]; 51 PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId); 52 PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId); 53 PETSC_EXTERN PetscErrorCode PetscInfoEnabled(PetscClassId, PetscBool *); 54 PETSC_EXTERN PetscErrorCode PetscInfoAllow(PetscBool); 55 PETSC_EXTERN PetscErrorCode PetscInfoSetFile(const char[], const char[]); 56 PETSC_EXTERN PetscErrorCode PetscInfoGetFile(char **, FILE **); 57 PETSC_EXTERN PetscErrorCode PetscInfoSetClasses(PetscBool, PetscInt, const char *const *); 58 PETSC_EXTERN PetscErrorCode PetscInfoGetClass(const char *, PetscBool *); 59 PETSC_EXTERN PetscErrorCode PetscInfoGetInfo(PetscBool *, PetscBool *, PetscBool *, PetscBool *, PetscInfoCommFlag *); 60 PETSC_EXTERN PetscErrorCode PetscInfoProcessClass(const char[], PetscInt, PetscClassId[]); 61 PETSC_EXTERN PetscErrorCode PetscInfoSetFilterCommSelf(PetscInfoCommFlag); 62 PETSC_EXTERN PetscErrorCode PetscInfoSetFromOptions(PetscOptions); 63 PETSC_EXTERN PetscErrorCode PetscInfoDestroy(void); 64 PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */ 65 66 /*MC 67 PetscLogEvent - id used to identify PETSc or user events which timed portions (blocks of executable) 68 code. 69 70 Level: intermediate 71 72 .seealso: `PetscLogEventRegister()`, `PetscLogEventBegin()`, `PetscLogEventEnd()`, `PetscLogStage` 73 M*/ 74 typedef int PetscLogEvent; 75 76 /*MC 77 PetscLogStage - id used to identify user stages (phases, sections) of runs - for logging 78 79 Level: intermediate 80 81 .seealso: `PetscLogStageRegister()`, `PetscLogStagePush()`, `PetscLogStagePop()`, `PetscLogEvent` 82 M*/ 83 typedef int PetscLogStage; 84 85 #define PETSC_EVENT 1311311 86 PETSC_EXTERN PetscLogEvent PETSC_LARGEST_EVENT; 87 88 /* Global flop counter */ 89 PETSC_EXTERN PetscLogDouble petsc_TotalFlops; 90 PETSC_EXTERN PetscLogDouble petsc_tmp_flops; 91 92 /* We must make the following structures available to access the event 93 activation flags in the PetscLogEventBegin/End() macros. These are not part of the PETSc public 94 API and are not intended to be used by other parts of PETSc or by users. 95 96 The code that manipulates these structures is in src/sys/logging/utils. 97 */ 98 typedef struct _n_PetscIntStack *PetscIntStack; 99 100 /* -----------------------------------------------------------------------------------------------------*/ 101 /* 102 PetscClassRegInfo, PetscClassPerfInfo - Each class has two data structures associated with it. The first has 103 static information about it, the second collects statistics on how many objects of the class are created, 104 how much memory they use, etc. 105 106 PetscClassRegLog, PetscClassPerfLog - arrays of the PetscClassRegInfo and PetscClassPerfInfo for all classes. 107 */ 108 typedef struct { 109 char *name; /* The class name */ 110 PetscClassId classid; /* The integer identifying this class */ 111 } PetscClassRegInfo; 112 113 typedef struct { 114 PetscClassId id; /* The integer identifying this class */ 115 int creations; /* The number of objects of this class created */ 116 int destructions; /* The number of objects of this class destroyed */ 117 PetscLogDouble mem; /* The total memory allocated by objects of this class; this is completely wrong and should possibly be removed */ 118 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects; this is completely wrong and should possibly be removed */ 119 } PetscClassPerfInfo; 120 121 typedef struct _n_PetscClassRegLog *PetscClassRegLog; 122 struct _n_PetscClassRegLog { 123 int numClasses; /* The number of classes registered */ 124 int maxClasses; /* The maximum number of classes */ 125 PetscClassRegInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 126 }; 127 128 typedef struct _n_PetscClassPerfLog *PetscClassPerfLog; 129 struct _n_PetscClassPerfLog { 130 int numClasses; /* The number of logging classes */ 131 int maxClasses; /* The maximum number of classes */ 132 PetscClassPerfInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 133 }; 134 /* -----------------------------------------------------------------------------------------------------*/ 135 /* 136 PetscEventRegInfo, PetscEventPerfInfo - Each event has two data structures associated with it. The first has 137 static information about it, the second collects statistics on how many times the event is used, how 138 much time it takes, etc. 139 140 PetscEventRegLog, PetscEventPerfLog - an array of all PetscEventRegInfo and PetscEventPerfInfo for all events. There is one 141 of these for each stage. 142 143 */ 144 typedef struct { 145 char *name; /* The name of this event */ 146 PetscClassId classid; /* The class the event is associated with */ 147 PetscBool collective; /* Flag this event as collective */ 148 #if defined(PETSC_HAVE_MPE) 149 int mpe_id_begin; /* MPE IDs that define the event */ 150 int mpe_id_end; 151 #endif 152 } PetscEventRegInfo; 153 154 typedef struct { 155 int id; /* The integer identifying this event */ 156 PetscBool active; /* The flag to activate logging */ 157 PetscBool visible; /* The flag to print info in summary */ 158 int depth; /* The nesting depth of the event call */ 159 int count; /* The number of times this event was executed */ 160 PetscLogDouble flops, flops2, flopsTmp; /* The flops and flops^2 used in this event */ 161 PetscLogDouble time, time2, timeTmp; /* The time and time^2 taken for this event */ 162 PetscLogDouble syncTime; /* The synchronization barrier time */ 163 PetscLogDouble dof[8]; /* The number of degrees of freedom associated with this event */ 164 PetscLogDouble errors[8]; /* The errors (user-defined) associated with this event */ 165 PetscLogDouble numMessages; /* The number of messages in this event */ 166 PetscLogDouble messageLength; /* The total message lengths in this event */ 167 PetscLogDouble numReductions; /* The number of reductions in this event */ 168 PetscLogDouble memIncrease; /* How much the resident memory has increased in this event */ 169 PetscLogDouble mallocIncrease; /* How much the maximum malloced space has increased in this event */ 170 PetscLogDouble mallocSpace; /* How much the space was malloced and kept during this event */ 171 PetscLogDouble mallocIncreaseEvent; /* Maximum of the high water mark with in event minus memory available at the end of the event */ 172 #if defined(PETSC_HAVE_DEVICE) 173 PetscLogDouble CpuToGpuCount; /* The total number of CPU to GPU copies */ 174 PetscLogDouble GpuToCpuCount; /* The total number of GPU to CPU copies */ 175 PetscLogDouble CpuToGpuSize; /* The total size of CPU to GPU copies */ 176 PetscLogDouble GpuToCpuSize; /* The total size of GPU to CPU copies */ 177 PetscLogDouble GpuFlops; /* The flops done on a GPU in this event */ 178 PetscLogDouble GpuTime; /* The time spent on a GPU in this event */ 179 #endif 180 } PetscEventPerfInfo; 181 182 typedef struct _n_PetscEventRegLog *PetscEventRegLog; 183 struct _n_PetscEventRegLog { 184 int numEvents; /* The number of registered events */ 185 int maxEvents; /* The maximum number of events */ 186 PetscEventRegInfo *eventInfo; /* The registration information for each event */ 187 }; 188 189 typedef struct _n_PetscEventPerfLog *PetscEventPerfLog; 190 struct _n_PetscEventPerfLog { 191 int numEvents; /* The number of logging events */ 192 int maxEvents; /* The maximum number of events */ 193 PetscEventPerfInfo *eventInfo; /* The performance information for each event */ 194 }; 195 /* ------------------------------------------------------------------------------------------------------------*/ 196 /* 197 PetscStageInfo - Contains all the information about a particular stage. 198 199 PetscStageLog - An array of PetscStageInfo for each registered stage. There is a single one of these in the code. 200 */ 201 typedef struct _PetscStageInfo { 202 char *name; /* The stage name */ 203 PetscBool used; /* The stage was pushed on this processor */ 204 PetscEventPerfInfo perfInfo; /* The stage performance information */ 205 PetscEventPerfLog eventLog; /* The event information for this stage */ 206 PetscClassPerfLog classLog; /* The class information for this stage */ 207 } PetscStageInfo; 208 209 typedef struct _n_PetscStageLog *PetscStageLog; 210 struct _n_PetscStageLog { 211 int numStages; /* The number of registered stages */ 212 int maxStages; /* The maximum number of stages */ 213 PetscIntStack stack; /* The stack for active stages */ 214 int curStage; /* The current stage (only used in macros so we don't call PetscIntStackTop) */ 215 PetscStageInfo *stageInfo; /* The information for each stage */ 216 PetscEventRegLog eventLog; /* The registered events */ 217 PetscClassRegLog classLog; /* The registered classes */ 218 }; 219 /* -----------------------------------------------------------------------------------------------------*/ 220 221 PETSC_EXTERN PetscErrorCode PetscLogObjectParent(PetscObject, PetscObject); 222 PETSC_EXTERN PetscErrorCode PetscLogObjectMemory(PetscObject, PetscLogDouble); 223 224 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 225 PETSC_EXTERN PetscStageLog petsc_stageLog; 226 PETSC_EXTERN PetscErrorCode PetscLogGetStageLog(PetscStageLog *); 227 PETSC_EXTERN PetscErrorCode PetscStageLogGetCurrent(PetscStageLog, int *); 228 PETSC_EXTERN PetscErrorCode PetscStageLogGetEventPerfLog(PetscStageLog, int, PetscEventPerfLog *); 229 230 /* 231 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 232 233 For the complex numbers version, note that 234 1 complex addition = 2 flops 235 1 complex multiplication = 6 flops, 236 where we define 1 flop as that for a double precision scalar. We roughly approximate 237 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 238 to the assumption that we're counting mostly additions and multiplications -- and 239 roughly the same number of each. More accurate counting could be done by distinguishing 240 among the various arithmetic operations. 241 */ 242 243 #if defined(PETSC_USE_COMPLEX) 244 #define PETSC_FLOPS_PER_OP 4.0 245 #else 246 #define PETSC_FLOPS_PER_OP 1.0 247 #endif 248 249 /*@C 250 PetscLogFlops - Log how many flops are performed in a calculation 251 252 Input Parameter: 253 . flops - the number of flops 254 255 Notes: 256 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 257 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 258 259 Level: intermediate 260 261 .seealso: `PetscLogView()`, `PetscLogGpuFlops()` 262 @*/ 263 264 static inline PetscErrorCode PetscLogFlops(PetscLogDouble n) { 265 PetscFunctionBegin; 266 #if defined(PETSC_USE_DEBUG) 267 PetscCheck(n >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Cannot log negative flops"); 268 #endif 269 petsc_TotalFlops += PETSC_FLOPS_PER_OP * n; 270 PetscFunctionReturn(0); 271 } 272 273 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 274 275 #if defined(PETSC_HAVE_MPE) 276 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 277 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 278 #endif 279 280 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject); 281 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject); 282 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 283 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 284 285 #define PetscLogObjectParents(p, n, d) PetscMacroReturnStandard(for (int _i = 0; _i < (n); ++_i) PetscCall(PetscLogObjectParent((PetscObject)(p), (PetscObject)(d)[_i]));) 286 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : 0) 287 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : 0) 288 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2, 3); 289 290 /* Initialization functions */ 291 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 292 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 293 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 294 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 295 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 296 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 297 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble, PetscLogDouble *); 298 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 299 300 /* Output functions */ 301 PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer); 302 PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void); 303 PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]); 304 305 /* Status checking functions */ 306 PETSC_EXTERN PetscErrorCode PetscLogIsActive(PetscBool *); 307 308 /* Stage functions */ 309 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[], PetscLogStage *); 310 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 311 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 312 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage, PetscBool); 313 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage, PetscBool *); 314 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage, PetscBool); 315 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage, PetscBool *); 316 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[], PetscLogStage *); 317 318 /* Event functions */ 319 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[], PetscClassId, PetscLogEvent *); 320 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent, PetscBool); 321 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 322 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 323 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 324 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 325 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePush(PetscLogEvent); 326 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePop(PetscLogEvent); 327 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent, PetscBool); 328 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 329 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 330 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[], PetscLogEvent *); 331 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int, PetscLogEvent, PetscEventPerfInfo *); 332 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 333 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 334 PETSC_EXTERN PetscErrorCode PetscLogPushCurrentEvent_Internal(PetscLogEvent); 335 PETSC_EXTERN PetscErrorCode PetscLogPopCurrentEvent_Internal(void); 336 337 /* Global counters */ 338 PETSC_EXTERN PetscLogDouble petsc_irecv_ct; 339 PETSC_EXTERN PetscLogDouble petsc_isend_ct; 340 PETSC_EXTERN PetscLogDouble petsc_recv_ct; 341 PETSC_EXTERN PetscLogDouble petsc_send_ct; 342 PETSC_EXTERN PetscLogDouble petsc_irecv_len; 343 PETSC_EXTERN PetscLogDouble petsc_isend_len; 344 PETSC_EXTERN PetscLogDouble petsc_recv_len; 345 PETSC_EXTERN PetscLogDouble petsc_send_len; 346 PETSC_EXTERN PetscLogDouble petsc_allreduce_ct; 347 PETSC_EXTERN PetscLogDouble petsc_gather_ct; 348 PETSC_EXTERN PetscLogDouble petsc_scatter_ct; 349 PETSC_EXTERN PetscLogDouble petsc_wait_ct; 350 PETSC_EXTERN PetscLogDouble petsc_wait_any_ct; 351 PETSC_EXTERN PetscLogDouble petsc_wait_all_ct; 352 PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct; 353 354 PETSC_EXTERN PetscBool PetscLogMemory; 355 356 PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */ 357 PETSC_EXTERN PetscErrorCode PetscLogEventSynchronize(PetscLogEvent, MPI_Comm); 358 359 #define PetscLogEventSync(e, comm) \ 360 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? PetscLogEventSynchronize((e), (comm)) : 0)) 361 362 #define PetscLogEventBegin(e, o1, o2, o3, o4) \ 363 ((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? (((*PetscLogPLB)((e), 0, (PetscObject)(o1), (PetscObject)(o2), (PetscObject)(o3), (PetscObject)(o4))) || PetscLogPushCurrentEvent_Internal(e)) : 0) 364 365 #define PetscLogEventEnd(e, o1, o2, o3, o4) \ 366 ((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? (((*PetscLogPLE)((e), 0, (PetscObject)(o1), (PetscObject)(o2), (PetscObject)(o3), (PetscObject)(o4))) || PetscLogPopCurrentEvent_Internal()) : 0) 367 368 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent, PetscLogDouble *); 369 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 370 371 /* 372 These are used internally in the PETSc routines to keep a count of MPI messages and 373 their sizes. 374 375 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 376 uses macros to defined the MPI operations. 377 378 It does not work correctly from HP-UX because it processes the 379 macros in a way that sometimes it double counts, hence 380 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 381 382 It does not work with Windows because winmpich lacks MPI_Type_size() 383 */ 384 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined(PETSC_HAVE_MPI_MISSING_TYPESIZE) 385 /* 386 Logging of MPI activities 387 */ 388 static inline PetscErrorCode PetscMPITypeSize(PetscInt count, MPI_Datatype type, PetscLogDouble *length) { 389 PetscMPIInt typesize; 390 391 if (type == MPI_DATATYPE_NULL) return 0; 392 PetscCallMPI(MPI_Type_size(type, &typesize)); 393 *length += (PetscLogDouble)(count * typesize); 394 return 0; 395 } 396 397 static inline PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm, const PetscMPIInt *counts, MPI_Datatype type, PetscLogDouble *length) { 398 PetscMPIInt typesize, size, p; 399 400 if (type == MPI_DATATYPE_NULL) return 0; 401 PetscCallMPI(MPI_Comm_size(comm, &size)); 402 PetscCallMPI(MPI_Type_size(type, &typesize)); 403 for (p = 0; p < size; ++p) *length += (PetscLogDouble)(counts[p] * typesize); 404 return 0; 405 } 406 407 static inline PetscErrorCode PetscMPITypeSizeCount(PetscInt n, const PetscMPIInt *counts, MPI_Datatype type, PetscLogDouble *length) { 408 PetscMPIInt typesize, p; 409 410 if (type == MPI_DATATYPE_NULL) return 0; 411 PetscCallMPI(MPI_Type_size(type, &typesize)); 412 for (p = 0; p < n; ++p) *length += (PetscLogDouble)(counts[p] * typesize); 413 return 0; 414 } 415 416 /* 417 Returns 1 if the communicator is parallel else zero 418 */ 419 static inline int PetscMPIParallelComm(MPI_Comm comm) { 420 PetscMPIInt size; 421 MPI_Comm_size(comm, &size); 422 return size > 1; 423 } 424 425 #define MPI_Irecv(buf, count, datatype, source, tag, comm, request) ((petsc_irecv_ct++, 0) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len)) || MPI_Irecv((buf), (count), (datatype), (source), (tag), (comm), (request))) 426 427 #define MPI_Irecv_c(buf, count, datatype, source, tag, comm, request) ((petsc_irecv_ct++, 0) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len)) || MPI_Irecv_c((buf), (count), (datatype), (source), (tag), (comm), (request))) 428 429 #define MPI_Isend(buf, count, datatype, dest, tag, comm, request) ((petsc_isend_ct++, 0) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len)) || MPI_Isend((buf), (count), (datatype), (dest), (tag), (comm), (request))) 430 431 #define MPI_Isend_c(buf, count, datatype, dest, tag, comm, request) ((petsc_isend_ct++, 0) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len)) || MPI_Isend_c((buf), (count), (datatype), (dest), (tag), (comm), (request))) 432 433 #define MPI_Startall_irecv(count, datatype, number, requests) ((petsc_irecv_ct += (PetscLogDouble)(number), 0) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len)) || ((number) && MPI_Startall((number), (requests)))) 434 435 #define MPI_Startall_isend(count, datatype, number, requests) ((petsc_isend_ct += (PetscLogDouble)(number), 0) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len)) || ((number) && MPI_Startall((number), (requests)))) 436 437 #define MPI_Start_isend(count, datatype, requests) ((petsc_isend_ct++, 0) || PetscMPITypeSize((count), (datatype), (&petsc_isend_len)) || MPI_Start((requests))) 438 439 #define MPI_Recv(buf, count, datatype, source, tag, comm, status) ((petsc_recv_ct++, 0) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len)) || MPI_Recv((buf), (count), (datatype), (source), (tag), (comm), (status))) 440 441 #define MPI_Recv_c(buf, count, datatype, source, tag, comm, status) ((petsc_recv_ct++, 0) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len)) || MPI_Recv_c((buf), (count), (datatype), (source), (tag), (comm), (status))) 442 443 #define MPI_Send(buf, count, datatype, dest, tag, comm) ((petsc_send_ct++, 0) || PetscMPITypeSize((count), (datatype), (&petsc_send_len)) || MPI_Send((buf), (count), (datatype), (dest), (tag), (comm))) 444 445 #define MPI_Send_c(buf, count, datatype, dest, tag, comm) ((petsc_send_ct++, 0) || PetscMPITypeSize((count), (datatype), (&petsc_send_len)) || MPI_Send_c((buf), (count), (datatype), (dest), (tag), (comm))) 446 447 #define MPI_Wait(request, status) ((petsc_wait_ct++, petsc_sum_of_waits_ct++, 0) || MPI_Wait((request), (status))) 448 449 #define MPI_Waitany(a, b, c, d) ((petsc_wait_any_ct++, petsc_sum_of_waits_ct++, 0) || MPI_Waitany((a), (b), (c), (d))) 450 451 #define MPI_Waitall(count, array_of_requests, array_of_statuses) ((petsc_wait_all_ct++, petsc_sum_of_waits_ct += (PetscLogDouble)(count), 0) || MPI_Waitall((count), (array_of_requests), (array_of_statuses))) 452 453 #define MPI_Allreduce(sendbuf, recvbuf, count, datatype, op, comm) (petsc_allreduce_ct += PetscMPIParallelComm((comm)), MPI_Allreduce((sendbuf), (recvbuf), (count), (datatype), (op), (comm))) 454 455 #define MPI_Bcast(buffer, count, datatype, root, comm) ((petsc_allreduce_ct += PetscMPIParallelComm((comm)), 0) || MPI_Bcast((buffer), (count), (datatype), (root), (comm))) 456 457 #define MPI_Reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm) ((petsc_allreduce_ct += PetscMPIParallelComm((comm)), 0) || MPI_Reduce_scatter_block((sendbuf), (recvbuf), (recvcount), (datatype), (op), (comm))) 458 459 #define MPI_Alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm) \ 460 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)), 0) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len)) || MPI_Alltoall((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm))) 461 462 #define MPI_Alltoallv(sendbuf, sendcnts, sdispls, sendtype, recvbuf, recvcnts, rdispls, recvtype, comm) \ 463 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)), 0) || PetscMPITypeSizeComm((comm), (sendcnts), (sendtype), (&petsc_send_len)) || MPI_Alltoallv((sendbuf), (sendcnts), (sdispls), (sendtype), (recvbuf), (recvcnts), (rdispls), (recvtype), (comm))) 464 465 #define MPI_Allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm) ((petsc_gather_ct += PetscMPIParallelComm((comm)), 0) || MPI_Allgather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm))) 466 467 #define MPI_Allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, comm) \ 468 ((petsc_gather_ct += PetscMPIParallelComm((comm)), 0) || MPI_Allgatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (comm))) 469 470 #define MPI_Gather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm) \ 471 ((petsc_gather_ct++, 0) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len)) || MPI_Gather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm))) 472 473 #define MPI_Gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, root, comm) \ 474 ((petsc_gather_ct++, 0) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len)) || MPI_Gatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (root), (comm))) 475 476 #define MPI_Scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm) \ 477 ((petsc_scatter_ct++, 0) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len)) || MPI_Scatter((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm))) 478 479 #define MPI_Scatterv(sendbuf, sendcount, displs, sendtype, recvbuf, recvcount, recvtype, root, comm) \ 480 ((petsc_scatter_ct++, 0) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len)) || MPI_Scatterv((sendbuf), (sendcount), (displs), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm))) 481 482 #define MPI_Ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request) \ 483 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)), 0) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len)) || MPI_Ialltoall((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm), (request))) 484 485 #define MPI_Ialltoallv(sendbuf, sendcnts, sdispls, sendtype, recvbuf, recvcnts, rdispls, recvtype, comm, request) \ 486 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)), 0) || PetscMPITypeSizeComm((comm), (sendcnts), (sendtype), (&petsc_send_len)) || MPI_Ialltoallv((sendbuf), (sendcnts), (sdispls), (sendtype), (recvbuf), (recvcnts), (rdispls), (recvtype), (comm), (request))) 487 488 #define MPI_Iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request) \ 489 ((petsc_gather_ct += PetscMPIParallelComm((comm)), 0) || MPI_Iallgather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm), (request))) 490 491 #define MPI_Iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, comm, request) \ 492 ((petsc_gather_ct += PetscMPIParallelComm((comm)), 0) || MPI_Iallgatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (comm), (request))) 493 494 #define MPI_Igather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \ 495 ((petsc_gather_ct++, 0) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len)) || MPI_Igather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request))) 496 497 #define MPI_Igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, root, comm, request) \ 498 ((petsc_gather_ct++, 0) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len)) || MPI_Igatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (root), (comm), (request))) 499 500 #define MPI_Iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \ 501 ((petsc_scatter_ct++, 0) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len)) || MPI_Iscatter((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request))) 502 503 #define MPI_Iscatterv(sendbuf, sendcount, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \ 504 ((petsc_scatter_ct++, 0) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len)) || MPI_Iscatterv((sendbuf), (sendcount), (displs), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request))) 505 506 #else 507 508 #define MPI_Startall_irecv(count, datatype, number, requests) ((number) && MPI_Startall((number), (requests))) 509 510 #define MPI_Startall_isend(count, datatype, number, requests) ((number) && MPI_Startall((number), (requests))) 511 512 #define MPI_Start_isend(count, datatype, requests) (MPI_Start((requests))) 513 514 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 515 516 #else /* ---Logging is turned off --------------------------------------------*/ 517 518 #define PetscLogMemory PETSC_FALSE 519 520 #define PetscLogFlops(n) 0 521 #define PetscGetFlops(a) (*(a) = 0.0, 0) 522 523 #define PetscLogStageRegister(a, b) 0 524 #define PetscLogStagePush(a) 0 525 #define PetscLogStagePop() 0 526 #define PetscLogStageSetActive(a, b) 0 527 #define PetscLogStageGetActive(a, b) 0 528 #define PetscLogStageGetVisible(a, b) 0 529 #define PetscLogStageSetVisible(a, b) 0 530 #define PetscLogStageGetId(a, b) (*(b) = 0, 0) 531 532 #define PetscLogEventRegister(a, b, c) 0 533 #define PetscLogEventSetCollective(a, b) 0 534 #define PetscLogEventIncludeClass(a) 0 535 #define PetscLogEventExcludeClass(a) 0 536 #define PetscLogEventActivate(a) 0 537 #define PetscLogEventDeactivate(a) 0 538 #define PetscLogEventDeactivatePush(a) 0 539 #define PetscLogEventDeactivatePop(a) 0 540 #define PetscLogEventActivateClass(a) 0 541 #define PetscLogEventDeactivateClass(a) 0 542 #define PetscLogEventSetActiveAll(a, b) 0 543 #define PetscLogEventGetId(a, b) (*(b) = 0, 0) 544 #define PetscLogEventGetPerfInfo(a, b, c) 0 545 #define PetscLogEventSetDof(a, b, c) 0 546 #define PetscLogEventSetError(a, b, c) 0 547 548 #define PetscLogPLB 0 549 #define PetscLogPLE 0 550 #define PetscLogPHC 0 551 #define PetscLogPHD 0 552 553 #define PetscLogObjectParents(p, n, c) 0 554 #define PetscLogObjectCreate(h) 0 555 #define PetscLogObjectDestroy(h) 0 556 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2, 3); 557 558 #define PetscLogDefaultBegin() 0 559 #define PetscLogAllBegin() 0 560 #define PetscLogNestedBegin() 0 561 #define PetscLogTraceBegin(file) 0 562 #define PetscLogActions(a) 0 563 #define PetscLogObjects(a) 0 564 #define PetscLogSetThreshold(a, b) 0 565 #define PetscLogSet(lb, le) 0 566 #define PetscLogIsActive(flag) (*(flag) = PETSC_FALSE, 0) 567 568 #define PetscLogView(viewer) 0 569 #define PetscLogViewFromOptions() 0 570 #define PetscLogDump(c) 0 571 572 #define PetscLogEventSync(e, comm) 0 573 #define PetscLogEventBegin(e, o1, o2, o3, o4) 0 574 #define PetscLogEventEnd(e, o1, o2, o3, o4) 0 575 576 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 577 #define MPI_Startall_irecv(count, datatype, number, requests) ((number) && MPI_Startall(number, requests)) 578 #define MPI_Startall_isend(count, datatype, number, requests) ((number) && MPI_Startall(number, requests)) 579 #define MPI_Start_isend(count, datatype, requests) MPI_Start(requests) 580 581 #endif /* PETSC_USE_LOG */ 582 583 #if defined(PETSC_USE_LOG) && defined(PETSC_HAVE_DEVICE) 584 585 /* Global GPU counters */ 586 PETSC_EXTERN PetscLogDouble petsc_ctog_ct; 587 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct; 588 PETSC_EXTERN PetscLogDouble petsc_ctog_sz; 589 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz; 590 PETSC_EXTERN PetscLogDouble petsc_ctog_ct_scalar; 591 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct_scalar; 592 PETSC_EXTERN PetscLogDouble petsc_ctog_sz_scalar; 593 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz_scalar; 594 PETSC_EXTERN PetscLogDouble petsc_gflops; 595 PETSC_EXTERN PetscLogDouble petsc_gtime; 596 597 static inline PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size) { 598 PetscFunctionBegin; 599 petsc_ctog_ct += 1; 600 petsc_ctog_sz += size; 601 PetscFunctionReturn(0); 602 } 603 604 static inline PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size) { 605 PetscFunctionBegin; 606 petsc_gtoc_ct += 1; 607 petsc_gtoc_sz += size; 608 PetscFunctionReturn(0); 609 } 610 611 static inline PetscErrorCode PetscLogCpuToGpuScalar(PetscLogDouble size) { 612 PetscFunctionBegin; 613 petsc_ctog_ct_scalar += 1; 614 petsc_ctog_sz_scalar += size; 615 PetscFunctionReturn(0); 616 } 617 618 static inline PetscErrorCode PetscLogGpuToCpuScalar(PetscLogDouble size) { 619 PetscFunctionBegin; 620 petsc_gtoc_ct_scalar += 1; 621 petsc_gtoc_sz_scalar += size; 622 PetscFunctionReturn(0); 623 } 624 625 /*@C 626 PetscLogGpuFlops - Log how many flops are performed in a calculation on the device 627 628 Input Parameter: 629 . flops - the number of flops 630 631 Notes: 632 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 633 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 634 635 The values are also added to the total flop count for the MPI rank that is set with `PetscLogFlops()`; hence the number of flops 636 just on the CPU would be the value from set from `PetscLogFlops()` minus the value set from `PetscLogGpuFlops()` 637 638 Level: intermediate 639 640 .seealso: `PetscLogView()`, `PetscLogFlops()`, `PetscLogGpuTimeBegin()`, `PetscLogGpuTimeEnd()` 641 @*/ 642 static inline PetscErrorCode PetscLogGpuFlops(PetscLogDouble n) { 643 PetscFunctionBegin; 644 PetscCheck(n >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Cannot log negative flops"); 645 petsc_TotalFlops += PETSC_FLOPS_PER_OP * n; 646 petsc_gflops += PETSC_FLOPS_PER_OP * n; 647 PetscFunctionReturn(0); 648 } 649 650 static inline PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t) { 651 PetscFunctionBegin; 652 petsc_gtime += t; 653 PetscFunctionReturn(0); 654 } 655 656 PETSC_EXTERN PetscErrorCode PetscLogGpuTime(void); 657 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeBegin(void); 658 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeEnd(void); 659 660 #else 661 662 #define PetscLogCpuToGpu(a) 0 663 #define PetscLogGpuToCpu(a) 0 664 #define PetscLogCpuToGpuScalar(a) 0 665 #define PetscLogGpuToCpuScalar(a) 0 666 #define PetscLogGpuFlops(a) 0 667 #define PetscLogGpuTime() 0 668 #define PetscLogGpuTimeAdd(a) 0 669 #define PetscLogGpuTimeBegin() 0 670 #define PetscLogGpuTimeEnd() 0 671 672 #endif /* PETSC_USE_LOG && PETSC_HAVE_DEVICE */ 673 674 #define PetscPreLoadBegin(flag, name) \ 675 do { \ 676 PetscBool PetscPreLoading = flag; \ 677 int PetscPreLoadMax, PetscPreLoadIt; \ 678 PetscLogStage _stageNum; \ 679 PetscCall(PetscOptionsGetBool(NULL, NULL, "-preload", &PetscPreLoading, NULL)); \ 680 PetscPreLoadMax = (int)(PetscPreLoading); \ 681 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed; \ 682 for (PetscPreLoadIt = 0; PetscPreLoadIt <= PetscPreLoadMax; PetscPreLoadIt++) { \ 683 PetscPreLoadingOn = PetscPreLoading; \ 684 PetscCall(PetscBarrier(NULL)); \ 685 if (PetscPreLoadIt > 0) PetscCall(PetscLogStageGetId(name, &_stageNum)); \ 686 else PetscCall(PetscLogStageRegister(name, &_stageNum)); \ 687 PetscCall(PetscLogStageSetActive(_stageNum, (PetscBool)(!PetscPreLoadMax || PetscPreLoadIt))); \ 688 PetscCall(PetscLogStagePush(_stageNum)); 689 690 #define PetscPreLoadEnd() \ 691 PetscCall(PetscLogStagePop()); \ 692 PetscPreLoading = PETSC_FALSE; \ 693 } \ 694 } \ 695 while (0) 696 697 #define PetscPreLoadStage(name) \ 698 do { \ 699 PetscCall(PetscLogStagePop()); \ 700 if (PetscPreLoadIt > 0) PetscCall(PetscLogStageGetId(name, &_stageNum)); \ 701 else PetscCall(PetscLogStageRegister(name, &_stageNum)); \ 702 PetscCall(PetscLogStageSetActive(_stageNum, (PetscBool)(!PetscPreLoadMax || PetscPreLoadIt))); \ 703 PetscCall(PetscLogStagePush(_stageNum)); \ 704 } while (0) 705 706 /* some vars for logging */ 707 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 708 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 709 710 #endif 711