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 */ 118 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 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 { 266 PetscFunctionBegin; 267 #if defined(PETSC_USE_DEBUG) 268 PetscCheck(n >= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 269 #endif 270 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 271 PetscFunctionReturn(0); 272 } 273 274 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 275 276 #if defined (PETSC_HAVE_MPE) 277 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 278 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 279 #endif 280 281 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 282 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 283 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 284 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 285 286 #define PetscLogObjectParents(p,n,d) PetscMacroReturnStandard(for (int _i=0; _i<(n); ++_i) PetscCall(PetscLogObjectParent((PetscObject)(p),(PetscObject)(d)[_i]));) 287 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : 0) 288 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : 0) 289 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2,3); 290 291 /* Initialization functions */ 292 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 293 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 294 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 295 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 296 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 297 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 298 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble,PetscLogDouble*); 299 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 300 PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 301 302 /* Output functions */ 303 PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer); 304 PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void); 305 PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]); 306 307 /* Status checking functions */ 308 PETSC_EXTERN PetscErrorCode PetscLogIsActive(PetscBool*); 309 310 /* Stage functions */ 311 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[],PetscLogStage*); 312 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 313 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 314 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage,PetscBool); 315 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage,PetscBool*); 316 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage,PetscBool); 317 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage,PetscBool*); 318 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[],PetscLogStage*); 319 320 /* Event functions */ 321 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[],PetscClassId,PetscLogEvent*); 322 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent,PetscBool); 323 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 324 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 325 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 326 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 327 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePush(PetscLogEvent); 328 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePop(PetscLogEvent); 329 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent,PetscBool); 330 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 331 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 332 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[],PetscLogEvent*); 333 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int,PetscLogEvent,PetscEventPerfInfo*); 334 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 335 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 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) ? \ 361 PetscLogEventSynchronize((e),(comm)) : 0)) 362 363 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 364 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 365 (*PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0)) 366 367 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 368 (((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 369 (*PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0)) 370 371 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent,PetscLogDouble*); 372 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 373 374 /* 375 These are used internally in the PETSc routines to keep a count of MPI messages and 376 their sizes. 377 378 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 379 uses macros to defined the MPI operations. 380 381 It does not work correctly from HP-UX because it processes the 382 macros in a way that sometimes it double counts, hence 383 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 384 385 It does not work with Windows because winmpich lacks MPI_Type_size() 386 */ 387 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 388 /* 389 Logging of MPI activities 390 */ 391 static inline PetscErrorCode PetscMPITypeSize(PetscInt count,MPI_Datatype type,PetscLogDouble *length) 392 { 393 PetscMPIInt typesize; 394 395 if (type == MPI_DATATYPE_NULL) return 0; 396 PetscCallMPI(MPI_Type_size(type,&typesize)); 397 *length += (PetscLogDouble) (count*typesize); 398 return 0; 399 } 400 401 static inline PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 402 { 403 PetscMPIInt typesize,size,p; 404 405 if (type == MPI_DATATYPE_NULL) return 0; 406 PetscCallMPI(MPI_Comm_size(comm,&size)); 407 PetscCallMPI(MPI_Type_size(type,&typesize)); 408 for (p=0; p<size; ++p) *length += (PetscLogDouble)(counts[p]*typesize); 409 return 0; 410 } 411 412 static inline PetscErrorCode PetscMPITypeSizeCount(PetscInt n,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 413 { 414 PetscMPIInt typesize,p; 415 416 if (type == MPI_DATATYPE_NULL) return 0; 417 PetscCallMPI(MPI_Type_size(type,&typesize)); 418 for (p=0; p<n; ++p) *length += (PetscLogDouble)(counts[p]*typesize); 419 return 0; 420 } 421 422 /* 423 Returns 1 if the communicator is parallel else zero 424 */ 425 static inline int PetscMPIParallelComm(MPI_Comm comm) 426 { 427 PetscMPIInt size; MPI_Comm_size(comm,&size); return size > 1; 428 } 429 430 #define MPI_Irecv(buf,count,datatype,source,tag,comm,request) \ 431 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv((buf),(count),(datatype),(source),(tag),(comm),(request))) 432 433 #define MPI_Irecv_c(buf,count,datatype,source,tag,comm,request) \ 434 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv_c((buf),(count),(datatype),(source),(tag),(comm),(request))) 435 436 #define MPI_Isend(buf,count,datatype,dest,tag,comm,request) \ 437 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend((buf),(count),(datatype),(dest),(tag),(comm),(request))) 438 439 #define MPI_Isend_c(buf,count,datatype,dest,tag,comm,request) \ 440 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend_c((buf),(count),(datatype),(dest),(tag),(comm),(request))) 441 442 #define MPI_Startall_irecv(count,datatype,number,requests) \ 443 ((petsc_irecv_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || ((number) && MPI_Startall((number),(requests)))) 444 445 #define MPI_Startall_isend(count,datatype,number,requests) \ 446 ((petsc_isend_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || ((number) && MPI_Startall((number),(requests)))) 447 448 #define MPI_Start_isend(count,datatype,requests) \ 449 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_isend_len)) || MPI_Start((requests))) 450 451 #define MPI_Recv(buf,count,datatype,source,tag,comm,status) \ 452 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv((buf),(count),(datatype),(source),(tag),(comm),(status))) 453 454 #define MPI_Recv_c(buf,count,datatype,source,tag,comm,status) \ 455 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv_c((buf),(count),(datatype),(source),(tag),(comm),(status))) 456 457 #define MPI_Send(buf,count,datatype,dest,tag,comm) \ 458 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send((buf),(count),(datatype),(dest),(tag),(comm))) 459 460 #define MPI_Send_c(buf,count,datatype,dest,tag,comm) \ 461 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send_c((buf),(count),(datatype),(dest),(tag),(comm))) 462 463 #define MPI_Wait(request,status) \ 464 ((petsc_wait_ct++,petsc_sum_of_waits_ct++,0) || MPI_Wait((request),(status))) 465 466 #define MPI_Waitany(a,b,c,d) \ 467 ((petsc_wait_any_ct++,petsc_sum_of_waits_ct++,0) || MPI_Waitany((a),(b),(c),(d))) 468 469 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 470 ((petsc_wait_all_ct++,petsc_sum_of_waits_ct += (PetscLogDouble) (count),0) || MPI_Waitall((count),(array_of_requests),(array_of_statuses))) 471 472 #define MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm) \ 473 (petsc_allreduce_ct += PetscMPIParallelComm((comm)),MPI_Allreduce((sendbuf),(recvbuf),(count),(datatype),(op),(comm))) 474 475 #define MPI_Bcast(buffer,count,datatype,root,comm) \ 476 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Bcast((buffer),(count),(datatype),(root),(comm))) 477 478 #define MPI_Reduce_scatter_block(sendbuf,recvbuf,recvcount,datatype,op,comm) \ 479 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Reduce_scatter_block((sendbuf),(recvbuf),(recvcount),(datatype),(op),(comm))) 480 481 #define MPI_Alltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 482 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Alltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 483 484 #define MPI_Alltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 485 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSizeComm((comm),(sendcnts),(sendtype),(&petsc_send_len)) || MPI_Alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 486 487 #define MPI_Allgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 488 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 489 490 #define MPI_Allgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm) \ 491 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm))) 492 493 #define MPI_Gather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 494 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 495 496 #define MPI_Gatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm) \ 497 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm))) 498 499 #define MPI_Scatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 500 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 501 502 #define MPI_Scatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 503 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 504 505 #define MPI_Ialltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 506 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Ialltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 507 508 #define MPI_Ialltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 509 ((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))) 510 511 #define MPI_Iallgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 512 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 513 514 #define MPI_Iallgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm,request) \ 515 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm),(request))) 516 517 #define MPI_Igather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 518 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 519 520 #define MPI_Igatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm,request) \ 521 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm),(request))) 522 523 #define MPI_Iscatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 524 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 525 526 #define MPI_Iscatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 527 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 528 529 #else 530 531 #define MPI_Startall_irecv(count,datatype,number,requests) \ 532 ((number) && MPI_Startall((number),(requests))) 533 534 #define MPI_Startall_isend(count,datatype,number,requests) \ 535 ((number) && MPI_Startall((number),(requests))) 536 537 #define MPI_Start_isend(count,datatype,requests) \ 538 (MPI_Start((requests))) 539 540 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 541 542 #else /* ---Logging is turned off --------------------------------------------*/ 543 544 #define PetscLogMemory PETSC_FALSE 545 546 #define PetscLogFlops(n) 0 547 #define PetscGetFlops(a) (*(a) = 0.0,0) 548 549 #define PetscLogStageRegister(a,b) 0 550 #define PetscLogStagePush(a) 0 551 #define PetscLogStagePop() 0 552 #define PetscLogStageSetActive(a,b) 0 553 #define PetscLogStageGetActive(a,b) 0 554 #define PetscLogStageGetVisible(a,b) 0 555 #define PetscLogStageSetVisible(a,b) 0 556 #define PetscLogStageGetId(a,b) (*(b)=0,0) 557 558 #define PetscLogEventRegister(a,b,c) 0 559 #define PetscLogEventSetCollective(a,b) 0 560 #define PetscLogEventIncludeClass(a) 0 561 #define PetscLogEventExcludeClass(a) 0 562 #define PetscLogEventActivate(a) 0 563 #define PetscLogEventDeactivate(a) 0 564 #define PetscLogEventDeactivatePush(a) 0 565 #define PetscLogEventDeactivatePop(a) 0 566 #define PetscLogEventActivateClass(a) 0 567 #define PetscLogEventDeactivateClass(a) 0 568 #define PetscLogEventSetActiveAll(a,b) 0 569 #define PetscLogEventGetId(a,b) (*(b)=0,0) 570 #define PetscLogEventGetPerfInfo(a,b,c) 0 571 #define PetscLogEventSetDof(a,b,c) 0 572 #define PetscLogEventSetError(a,b,c) 0 573 574 #define PetscLogPLB 0 575 #define PetscLogPLE 0 576 #define PetscLogPHC 0 577 #define PetscLogPHD 0 578 579 #define PetscLogObjectParents(p,n,c) 0 580 #define PetscLogObjectCreate(h) 0 581 #define PetscLogObjectDestroy(h) 0 582 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject,const char[],...) PETSC_ATTRIBUTE_FORMAT(2,3); 583 584 #define PetscLogDefaultBegin() 0 585 #define PetscLogAllBegin() 0 586 #define PetscLogNestedBegin() 0 587 #define PetscLogTraceBegin(file) 0 588 #define PetscLogActions(a) 0 589 #define PetscLogObjects(a) 0 590 #define PetscLogSetThreshold(a,b) 0 591 #define PetscLogSet(lb,le) 0 592 #define PetscLogIsActive(flag) (*(flag) = PETSC_FALSE,0) 593 594 #define PetscLogView(viewer) 0 595 #define PetscLogViewFromOptions() 0 596 #define PetscLogDump(c) 0 597 598 #define PetscLogEventSync(e,comm) 0 599 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0 600 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0 601 602 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 603 #define MPI_Startall_irecv(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 604 #define MPI_Startall_isend(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 605 #define MPI_Start_isend(count,datatype,requests) MPI_Start(requests) 606 607 #endif /* PETSC_USE_LOG */ 608 609 #if defined (PETSC_USE_LOG) && defined(PETSC_HAVE_DEVICE) 610 611 /* Global GPU counters */ 612 PETSC_EXTERN PetscLogDouble petsc_ctog_ct; 613 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct; 614 PETSC_EXTERN PetscLogDouble petsc_ctog_sz; 615 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz; 616 PETSC_EXTERN PetscLogDouble petsc_ctog_ct_scalar; 617 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct_scalar; 618 PETSC_EXTERN PetscLogDouble petsc_ctog_sz_scalar; 619 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz_scalar; 620 PETSC_EXTERN PetscLogDouble petsc_gflops; 621 PETSC_EXTERN PetscLogDouble petsc_gtime; 622 623 static inline PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size) 624 { 625 PetscFunctionBegin; 626 petsc_ctog_ct += 1; 627 petsc_ctog_sz += size; 628 PetscFunctionReturn(0); 629 } 630 631 static inline PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size) 632 { 633 PetscFunctionBegin; 634 petsc_gtoc_ct += 1; 635 petsc_gtoc_sz += size; 636 PetscFunctionReturn(0); 637 } 638 639 static inline PetscErrorCode PetscLogCpuToGpuScalar(PetscLogDouble size) 640 { 641 PetscFunctionBegin; 642 petsc_ctog_ct_scalar += 1; 643 petsc_ctog_sz_scalar += size; 644 PetscFunctionReturn(0); 645 } 646 647 static inline PetscErrorCode PetscLogGpuToCpuScalar(PetscLogDouble size) 648 { 649 PetscFunctionBegin; 650 petsc_gtoc_ct_scalar += 1; 651 petsc_gtoc_sz_scalar += size; 652 PetscFunctionReturn(0); 653 } 654 655 /*@C 656 PetscLogGpuFlops - Log how many flops are performed in a calculation on the device 657 658 Input Parameter: 659 . flops - the number of flops 660 661 Notes: 662 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 663 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 664 665 Level: intermediate 666 667 .seealso: `PetscLogView()`, `PetscLogFlops()`, `PetscLogGpuTimeBegin()`, `PetscLogGpuTimeEnd()` 668 @*/ 669 static inline PetscErrorCode PetscLogGpuFlops(PetscLogDouble n) 670 { 671 PetscFunctionBegin; 672 PetscCheck(n >= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 673 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 674 petsc_gflops += PETSC_FLOPS_PER_OP*n; 675 PetscFunctionReturn(0); 676 } 677 678 static inline PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t) 679 { 680 PetscFunctionBegin; 681 petsc_gtime += t; 682 PetscFunctionReturn(0); 683 } 684 685 PETSC_EXTERN PetscErrorCode PetscLogGpuTime(void); 686 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeBegin(void); 687 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeEnd(void); 688 689 #else 690 691 #define PetscLogCpuToGpu(a) 0 692 #define PetscLogGpuToCpu(a) 0 693 #define PetscLogCpuToGpuScalar(a) 0 694 #define PetscLogGpuToCpuScalar(a) 0 695 #define PetscLogGpuFlops(a) 0 696 #define PetscLogGpuTimeAdd(a) 0 697 #define PetscLogGpuTimeBegin() 0 698 #define PetscLogGpuTimeEnd() 0 699 700 #endif /* PETSC_USE_LOG && PETSC_HAVE_DEVICE */ 701 702 #define PetscPreLoadBegin(flag,name) \ 703 do {\ 704 PetscBool PetscPreLoading = flag;\ 705 int PetscPreLoadMax,PetscPreLoadIt;\ 706 PetscLogStage _stageNum;\ 707 PetscCall(PetscOptionsGetBool(NULL,NULL,"-preload",&PetscPreLoading,NULL)); \ 708 PetscPreLoadMax = (int)(PetscPreLoading);\ 709 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 710 for (PetscPreLoadIt=0; PetscPreLoadIt<=PetscPreLoadMax; PetscPreLoadIt++) {\ 711 PetscPreLoadingOn = PetscPreLoading;\ 712 PetscCall(PetscBarrier(NULL));\ 713 if (PetscPreLoadIt>0) PetscCall(PetscLogStageGetId(name,&_stageNum));\ 714 else PetscCall(PetscLogStageRegister(name,&_stageNum));\ 715 PetscCall(PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt)));\ 716 PetscCall(PetscLogStagePush(_stageNum)); 717 718 #define PetscPreLoadEnd() \ 719 PetscCall(PetscLogStagePop());\ 720 PetscPreLoading = PETSC_FALSE;\ 721 }\ 722 } while (0) 723 724 #define PetscPreLoadStage(name) do { \ 725 PetscCall(PetscLogStagePop()); \ 726 if (PetscPreLoadIt>0) PetscCall(PetscLogStageGetId(name,&_stageNum)); \ 727 else PetscCall(PetscLogStageRegister(name,&_stageNum)); \ 728 PetscCall(PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt))); \ 729 PetscCall(PetscLogStagePush(_stageNum)); \ 730 } while (0) 731 732 /* some vars for logging */ 733 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 734 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 735 736 #endif 737