1 /* 2 Defines profile/logging in PETSc. 3 */ 4 5 #if !defined(PETSCLOG_H) 6 #define PETSCLOG_H 7 #include <petscsys.h> 8 #include <petsctime.h> 9 10 /* General logging of information; different from event logging */ 11 PETSC_EXTERN PetscErrorCode PetscInfo_Private(const char[],PetscObject,const char[],...) PETSC_ATTRIBUTE_FORMAT(3,4); 12 #if defined(PETSC_USE_INFO) 13 #define PetscInfo(A,...) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),__VA_ARGS__) 14 #else 15 #define PetscInfo(A,...) 0 16 #endif 17 18 #define PetscInfo1(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 19 #define PetscInfo2(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 20 #define PetscInfo3(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS_)_ 21 #define PetscInfo4(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 22 #define PetscInfo5(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 23 #define PetscInfo6(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 24 #define PetscInfo7(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 25 #define PetscInfo8(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 26 #define PetscInfo9(...) PETSC_DEPRECATED_MACRO("GCC warning \"Use PetscInfo() (since version 3.17)\"") PetscInfo(__VA_ARGS__) 27 28 /*E 29 PetscInfoCommFlag - Describes the method by which to filter PetscInfo() by communicator size 30 31 Used as an input for PetscInfoSetFilterCommSelf() 32 33 $ PETSC_INFO_COMM_ALL - Default uninitialized value. PetscInfo() will not filter based on communicator size (i.e. will 34 print for all communicators) 35 $ PETSC_INFO_COMM_NO_SELF - PetscInfo() will NOT print for communicators with size = 1 (i.e. *_COMM_SELF) 36 $ PETSC_INFO_COMM_ONLY_SELF - PetscInfo will ONLY print for communicators with size = 1 37 38 Level: intermediate 39 40 .seealso: PetscInfo(), PetscInfoSetFromOptions(), PetscInfoSetFilterCommSelf() 41 E*/ 42 typedef enum { 43 PETSC_INFO_COMM_ALL = -1, 44 PETSC_INFO_COMM_NO_SELF = 0, 45 PETSC_INFO_COMM_ONLY_SELF = 1 46 } PetscInfoCommFlag; 47 48 PETSC_EXTERN const char * const PetscInfoCommFlags[]; 49 PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId); 50 PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId); 51 PETSC_EXTERN PetscErrorCode PetscInfoEnabled(PetscClassId, PetscBool *); 52 PETSC_EXTERN PetscErrorCode PetscInfoAllow(PetscBool); 53 PETSC_EXTERN PetscErrorCode PetscInfoSetFile(const char[],const char[]); 54 PETSC_EXTERN PetscErrorCode PetscInfoGetFile(char **,FILE **); 55 PETSC_EXTERN PetscErrorCode PetscInfoSetClasses(PetscBool,PetscInt,const char *const *); 56 PETSC_EXTERN PetscErrorCode PetscInfoGetClass(const char *, PetscBool *); 57 PETSC_EXTERN PetscErrorCode PetscInfoGetInfo(PetscBool *,PetscBool *,PetscBool *,PetscBool *,PetscInfoCommFlag *); 58 PETSC_EXTERN PetscErrorCode PetscInfoProcessClass(const char[],PetscInt,PetscClassId[]); 59 PETSC_EXTERN PetscErrorCode PetscInfoSetFilterCommSelf(PetscInfoCommFlag); 60 PETSC_EXTERN PetscErrorCode PetscInfoSetFromOptions(PetscOptions); 61 PETSC_EXTERN PetscErrorCode PetscInfoDestroy(void); 62 PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */ 63 64 /*MC 65 PetscLogEvent - id used to identify PETSc or user events which timed portions (blocks of executable) 66 code. 67 68 Level: intermediate 69 70 .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscLogStage 71 M*/ 72 typedef int PetscLogEvent; 73 74 /*MC 75 PetscLogStage - id used to identify user stages (phases, sections) of runs - for logging 76 77 Level: intermediate 78 79 .seealso: PetscLogStageRegister(), PetscLogStagePush(), PetscLogStagePop(), PetscLogEvent 80 M*/ 81 typedef int PetscLogStage; 82 83 #define PETSC_EVENT 1311311 84 PETSC_EXTERN PetscLogEvent PETSC_LARGEST_EVENT; 85 86 /* Global flop counter */ 87 PETSC_EXTERN PetscLogDouble petsc_TotalFlops; 88 PETSC_EXTERN PetscLogDouble petsc_tmp_flops; 89 90 /* We must make the following structures available to access the event 91 activation flags in the PetscLogEventBegin/End() macros. These are not part of the PETSc public 92 API and are not intended to be used by other parts of PETSc or by users. 93 94 The code that manipulates these structures is in src/sys/logging/utils. 95 */ 96 typedef struct _n_PetscIntStack *PetscIntStack; 97 98 /* -----------------------------------------------------------------------------------------------------*/ 99 /* 100 PetscClassRegInfo, PetscClassPerfInfo - Each class has two data structures associated with it. The first has 101 static information about it, the second collects statistics on how many objects of the class are created, 102 how much memory they use, etc. 103 104 PetscClassRegLog, PetscClassPerfLog - arrays of the PetscClassRegInfo and PetscClassPerfInfo for all classes. 105 */ 106 typedef struct { 107 char *name; /* The class name */ 108 PetscClassId classid; /* The integer identifying this class */ 109 } PetscClassRegInfo; 110 111 typedef struct { 112 PetscClassId id; /* The integer identifying this class */ 113 int creations; /* The number of objects of this class created */ 114 int destructions; /* The number of objects of this class destroyed */ 115 PetscLogDouble mem; /* The total memory allocated by objects of this class */ 116 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 117 } PetscClassPerfInfo; 118 119 typedef struct _n_PetscClassRegLog *PetscClassRegLog; 120 struct _n_PetscClassRegLog { 121 int numClasses; /* The number of classes registered */ 122 int maxClasses; /* The maximum number of classes */ 123 PetscClassRegInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 124 }; 125 126 typedef struct _n_PetscClassPerfLog *PetscClassPerfLog; 127 struct _n_PetscClassPerfLog { 128 int numClasses; /* The number of logging classes */ 129 int maxClasses; /* The maximum number of classes */ 130 PetscClassPerfInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 131 }; 132 /* -----------------------------------------------------------------------------------------------------*/ 133 /* 134 PetscEventRegInfo, PetscEventPerfInfo - Each event has two data structures associated with it. The first has 135 static information about it, the second collects statistics on how many times the event is used, how 136 much time it takes, etc. 137 138 PetscEventRegLog, PetscEventPerfLog - an array of all PetscEventRegInfo and PetscEventPerfInfo for all events. There is one 139 of these for each stage. 140 141 */ 142 typedef struct { 143 char *name; /* The name of this event */ 144 PetscClassId classid; /* The class the event is associated with */ 145 PetscBool collective; /* Flag this event as collective */ 146 #if defined (PETSC_HAVE_MPE) 147 int mpe_id_begin; /* MPE IDs that define the event */ 148 int mpe_id_end; 149 #endif 150 } PetscEventRegInfo; 151 152 typedef struct { 153 int id; /* The integer identifying this event */ 154 PetscBool active; /* The flag to activate logging */ 155 PetscBool visible; /* The flag to print info in summary */ 156 int depth; /* The nesting depth of the event call */ 157 int count; /* The number of times this event was executed */ 158 PetscLogDouble flops, flops2, flopsTmp; /* The flops and flops^2 used in this event */ 159 PetscLogDouble time, time2, timeTmp; /* The time and time^2 taken for this event */ 160 PetscLogDouble syncTime; /* The synchronization barrier time */ 161 PetscLogDouble dof[8]; /* The number of degrees of freedom associated with this event */ 162 PetscLogDouble errors[8]; /* The errors (user-defined) associated with this event */ 163 PetscLogDouble numMessages; /* The number of messages in this event */ 164 PetscLogDouble messageLength; /* The total message lengths in this event */ 165 PetscLogDouble numReductions; /* The number of reductions in this event */ 166 PetscLogDouble memIncrease; /* How much the resident memory has increased in this event */ 167 PetscLogDouble mallocIncrease;/* How much the maximum malloced space has increased in this event */ 168 PetscLogDouble mallocSpace; /* How much the space was malloced and kept during this event */ 169 PetscLogDouble mallocIncreaseEvent; /* Maximum of the high water mark with in event minus memory available at the end of the event */ 170 #if defined(PETSC_HAVE_DEVICE) 171 PetscLogDouble CpuToGpuCount; /* The total number of CPU to GPU copies */ 172 PetscLogDouble GpuToCpuCount; /* The total number of GPU to CPU copies */ 173 PetscLogDouble CpuToGpuSize; /* The total size of CPU to GPU copies */ 174 PetscLogDouble GpuToCpuSize; /* The total size of GPU to CPU copies */ 175 PetscLogDouble GpuFlops; /* The flops done on a GPU in this event */ 176 PetscLogDouble GpuTime; /* The time spent on a GPU in this event */ 177 #endif 178 } PetscEventPerfInfo; 179 180 typedef struct _n_PetscEventRegLog *PetscEventRegLog; 181 struct _n_PetscEventRegLog { 182 int numEvents; /* The number of registered events */ 183 int maxEvents; /* The maximum number of events */ 184 PetscEventRegInfo *eventInfo; /* The registration information for each event */ 185 }; 186 187 typedef struct _n_PetscEventPerfLog *PetscEventPerfLog; 188 struct _n_PetscEventPerfLog { 189 int numEvents; /* The number of logging events */ 190 int maxEvents; /* The maximum number of events */ 191 PetscEventPerfInfo *eventInfo; /* The performance information for each event */ 192 }; 193 /* ------------------------------------------------------------------------------------------------------------*/ 194 /* 195 PetscStageInfo - Contains all the information about a particular stage. 196 197 PetscStageLog - An array of PetscStageInfo for each registered stage. There is a single one of these in the code. 198 */ 199 typedef struct _PetscStageInfo { 200 char *name; /* The stage name */ 201 PetscBool used; /* The stage was pushed on this processor */ 202 PetscEventPerfInfo perfInfo; /* The stage performance information */ 203 PetscEventPerfLog eventLog; /* The event information for this stage */ 204 PetscClassPerfLog classLog; /* The class information for this stage */ 205 } PetscStageInfo; 206 207 typedef struct _n_PetscStageLog *PetscStageLog; 208 struct _n_PetscStageLog { 209 int numStages; /* The number of registered stages */ 210 int maxStages; /* The maximum number of stages */ 211 PetscIntStack stack; /* The stack for active stages */ 212 int curStage; /* The current stage (only used in macros so we don't call PetscIntStackTop) */ 213 PetscStageInfo *stageInfo; /* The information for each stage */ 214 PetscEventRegLog eventLog; /* The registered events */ 215 PetscClassRegLog classLog; /* The registered classes */ 216 }; 217 /* -----------------------------------------------------------------------------------------------------*/ 218 219 PETSC_EXTERN PetscErrorCode PetscLogObjectParent(PetscObject,PetscObject); 220 PETSC_EXTERN PetscErrorCode PetscLogObjectMemory(PetscObject,PetscLogDouble); 221 222 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 223 PETSC_EXTERN PetscStageLog petsc_stageLog; 224 PETSC_EXTERN PetscErrorCode PetscLogGetStageLog(PetscStageLog*); 225 PETSC_EXTERN PetscErrorCode PetscStageLogGetCurrent(PetscStageLog,int*); 226 PETSC_EXTERN PetscErrorCode PetscStageLogGetEventPerfLog(PetscStageLog,int,PetscEventPerfLog*); 227 228 /* 229 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 230 231 For the complex numbers version, note that 232 1 complex addition = 2 flops 233 1 complex multiplication = 6 flops, 234 where we define 1 flop as that for a double precision scalar. We roughly approximate 235 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 236 to the assumption that we're counting mostly additions and multiplications -- and 237 roughly the same number of each. More accurate counting could be done by distinguishing 238 among the various arithmetic operations. 239 */ 240 241 #if defined(PETSC_USE_COMPLEX) 242 #define PETSC_FLOPS_PER_OP 4.0 243 #else 244 #define PETSC_FLOPS_PER_OP 1.0 245 #endif 246 247 /*@C 248 PetscLogFlops - Log how many flops are performed in a calculation 249 250 Input Parameter: 251 . flops - the number of flops 252 253 Notes: 254 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 255 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 256 257 Level: intermediate 258 259 .seealso: PetscLogView(), PetscLogGpuFlops() 260 @*/ 261 262 static inline PetscErrorCode PetscLogFlops(PetscLogDouble n) 263 { 264 PetscFunctionBegin; 265 #if defined(PETSC_USE_DEBUG) 266 PetscCheck(n >= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 267 #endif 268 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 269 PetscFunctionReturn(0); 270 } 271 272 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 273 274 #if defined (PETSC_HAVE_MPE) 275 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 276 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 277 #endif 278 279 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 280 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 281 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 282 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 283 284 #define PetscLogObjectParents(p,n,d) 0;do{int _i; for (_i=0; _i<(n); _i++) {ierr = PetscLogObjectParent((PetscObject)(p),(PetscObject)(d)[_i]);CHKERRQ(ierr);}}while (0) 285 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : 0) 286 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : 0) 287 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2,3); 288 289 /* Initialization functions */ 290 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 291 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 292 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 293 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 294 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 295 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 296 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble,PetscLogDouble*); 297 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 298 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 /* Stage functions */ 306 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[],PetscLogStage*); 307 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 308 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 309 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage,PetscBool); 310 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage,PetscBool*); 311 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage,PetscBool); 312 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage,PetscBool*); 313 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[],PetscLogStage*); 314 315 /* Event functions */ 316 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[],PetscClassId,PetscLogEvent*); 317 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent,PetscBool); 318 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 319 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 320 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 321 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 322 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePush(PetscLogEvent); 323 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePop(PetscLogEvent); 324 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent,PetscBool); 325 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 326 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 327 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[],PetscLogEvent*); 328 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int,PetscLogEvent,PetscEventPerfInfo*); 329 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 330 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 331 332 /* Global counters */ 333 PETSC_EXTERN PetscLogDouble petsc_irecv_ct; 334 PETSC_EXTERN PetscLogDouble petsc_isend_ct; 335 PETSC_EXTERN PetscLogDouble petsc_recv_ct; 336 PETSC_EXTERN PetscLogDouble petsc_send_ct; 337 PETSC_EXTERN PetscLogDouble petsc_irecv_len; 338 PETSC_EXTERN PetscLogDouble petsc_isend_len; 339 PETSC_EXTERN PetscLogDouble petsc_recv_len; 340 PETSC_EXTERN PetscLogDouble petsc_send_len; 341 PETSC_EXTERN PetscLogDouble petsc_allreduce_ct; 342 PETSC_EXTERN PetscLogDouble petsc_gather_ct; 343 PETSC_EXTERN PetscLogDouble petsc_scatter_ct; 344 PETSC_EXTERN PetscLogDouble petsc_wait_ct; 345 PETSC_EXTERN PetscLogDouble petsc_wait_any_ct; 346 PETSC_EXTERN PetscLogDouble petsc_wait_all_ct; 347 PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct; 348 349 PETSC_EXTERN PetscBool PetscLogMemory; 350 351 PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */ 352 PETSC_EXTERN PetscErrorCode PetscLogEventSynchronize(PetscLogEvent, MPI_Comm); 353 354 #define PetscLogEventSync(e,comm) \ 355 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 356 PetscLogEventSynchronize((e),(comm)) : 0)) 357 358 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 359 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 360 (*PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0)) 361 362 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 363 (((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 364 (*PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0)) 365 366 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent,PetscLogDouble*); 367 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 368 369 /* 370 These are used internally in the PETSc routines to keep a count of MPI messages and 371 their sizes. 372 373 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 374 uses macros to defined the MPI operations. 375 376 It does not work correctly from HP-UX because it processes the 377 macros in a way that sometimes it double counts, hence 378 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 379 380 It does not work with Windows because winmpich lacks MPI_Type_size() 381 */ 382 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 383 /* 384 Logging of MPI activities 385 */ 386 static inline PetscErrorCode PetscMPITypeSize(PetscInt count,MPI_Datatype type,PetscLogDouble *length) 387 { 388 PetscMPIInt typesize; 389 PetscErrorCode ierr; 390 if (type == MPI_DATATYPE_NULL) return 0; 391 ierr = MPI_Type_size(type,&typesize);CHKERRMPI(ierr); 392 *length += (PetscLogDouble) (count*typesize); 393 return 0; 394 } 395 396 static inline PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 397 { 398 PetscMPIInt typesize,size,p; 399 PetscErrorCode ierr; 400 401 if (type == MPI_DATATYPE_NULL) return 0; 402 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 403 ierr = MPI_Type_size(type,&typesize);CHKERRMPI(ierr); 404 for (p=0; p<size; ++p) { 405 *length += (PetscLogDouble) (counts[p]*typesize); 406 } 407 return 0; 408 } 409 410 static inline PetscErrorCode PetscMPITypeSizeCount(PetscInt n,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 411 { 412 PetscMPIInt typesize,p; 413 PetscErrorCode ierr; 414 415 if (type == MPI_DATATYPE_NULL) return 0; 416 ierr = MPI_Type_size(type,&typesize);CHKERRMPI(ierr); 417 for (p=0; p<n; ++p) { 418 *length += (PetscLogDouble) (counts[p]*typesize); 419 } 420 return 0; 421 } 422 423 /* 424 Returns 1 if the communicator is parallel else zero 425 */ 426 static inline int PetscMPIParallelComm(MPI_Comm comm) 427 { 428 PetscMPIInt size; MPI_Comm_size(comm,&size); return size > 1; 429 } 430 431 #define MPI_Irecv(buf,count,datatype,source,tag,comm,request) \ 432 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv((buf),(count),(datatype),(source),(tag),(comm),(request))) 433 434 #define MPI_Irecv_c(buf,count,datatype,source,tag,comm,request) \ 435 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv_c((buf),(count),(datatype),(source),(tag),(comm),(request))) 436 437 #define MPI_Isend(buf,count,datatype,dest,tag,comm,request) \ 438 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend((buf),(count),(datatype),(dest),(tag),(comm),(request))) 439 440 #define MPI_Isend_c(buf,count,datatype,dest,tag,comm,request) \ 441 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend_c((buf),(count),(datatype),(dest),(tag),(comm),(request))) 442 443 #define MPI_Startall_irecv(count,datatype,number,requests) \ 444 ((petsc_irecv_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || ((number) && MPI_Startall((number),(requests)))) 445 446 #define MPI_Startall_isend(count,datatype,number,requests) \ 447 ((petsc_isend_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || ((number) && MPI_Startall((number),(requests)))) 448 449 #define MPI_Start_isend(count,datatype,requests) \ 450 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_isend_len)) || MPI_Start((requests))) 451 452 #define MPI_Recv(buf,count,datatype,source,tag,comm,status) \ 453 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv((buf),(count),(datatype),(source),(tag),(comm),(status))) 454 455 #define MPI_Recv_c(buf,count,datatype,source,tag,comm,status) \ 456 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv_c((buf),(count),(datatype),(source),(tag),(comm),(status))) 457 458 #define MPI_Send(buf,count,datatype,dest,tag,comm) \ 459 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send((buf),(count),(datatype),(dest),(tag),(comm))) 460 461 #define MPI_Send_c(buf,count,datatype,dest,tag,comm) \ 462 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send_c((buf),(count),(datatype),(dest),(tag),(comm))) 463 464 #define MPI_Wait(request,status) \ 465 ((petsc_wait_ct++,petsc_sum_of_waits_ct++,0) || MPI_Wait((request),(status))) 466 467 #define MPI_Waitany(a,b,c,d) \ 468 ((petsc_wait_any_ct++,petsc_sum_of_waits_ct++,0) || MPI_Waitany((a),(b),(c),(d))) 469 470 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 471 ((petsc_wait_all_ct++,petsc_sum_of_waits_ct += (PetscLogDouble) (count),0) || MPI_Waitall((count),(array_of_requests),(array_of_statuses))) 472 473 #define MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm) \ 474 (petsc_allreduce_ct += PetscMPIParallelComm((comm)),MPI_Allreduce((sendbuf),(recvbuf),(count),(datatype),(op),(comm))) 475 476 #define MPI_Bcast(buffer,count,datatype,root,comm) \ 477 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Bcast((buffer),(count),(datatype),(root),(comm))) 478 479 #define MPI_Reduce_scatter_block(sendbuf,recvbuf,recvcount,datatype,op,comm) \ 480 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Reduce_scatter_block((sendbuf),(recvbuf),(recvcount),(datatype),(op),(comm))) 481 482 #define MPI_Alltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 483 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Alltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 484 485 #define MPI_Alltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 486 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSizeComm((comm),(sendcnts),(sendtype),(&petsc_send_len)) || MPI_Alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 487 488 #define MPI_Allgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 489 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 490 491 #define MPI_Allgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm) \ 492 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm))) 493 494 #define MPI_Gather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 495 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 496 497 #define MPI_Gatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm) \ 498 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm))) 499 500 #define MPI_Scatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 501 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 502 503 #define MPI_Scatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 504 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 505 506 #define MPI_Ialltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 507 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Ialltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 508 509 #define MPI_Ialltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 510 ((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))) 511 512 #define MPI_Iallgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 513 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 514 515 #define MPI_Iallgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm,request) \ 516 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm),(request))) 517 518 #define MPI_Igather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 519 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 520 521 #define MPI_Igatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm,request) \ 522 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm),(request))) 523 524 #define MPI_Iscatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 525 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 526 527 #define MPI_Iscatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 528 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 529 530 #else 531 532 #define MPI_Startall_irecv(count,datatype,number,requests) \ 533 ((number) && MPI_Startall((number),(requests))) 534 535 #define MPI_Startall_isend(count,datatype,number,requests) \ 536 ((number) && MPI_Startall((number),(requests))) 537 538 #define MPI_Start_isend(count,datatype,requests) \ 539 (MPI_Start((requests))) 540 541 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 542 543 #else /* ---Logging is turned off --------------------------------------------*/ 544 545 #define PetscLogMemory PETSC_FALSE 546 547 #define PetscLogFlops(n) 0 548 #define PetscGetFlops(a) (*(a) = 0.0,0) 549 550 #define PetscLogStageRegister(a,b) 0 551 #define PetscLogStagePush(a) 0 552 #define PetscLogStagePop() 0 553 #define PetscLogStageSetActive(a,b) 0 554 #define PetscLogStageGetActive(a,b) 0 555 #define PetscLogStageGetVisible(a,b) 0 556 #define PetscLogStageSetVisible(a,b) 0 557 #define PetscLogStageGetId(a,b) (*(b)=0,0) 558 559 #define PetscLogEventRegister(a,b,c) 0 560 #define PetscLogEventSetCollective(a,b) 0 561 #define PetscLogEventIncludeClass(a) 0 562 #define PetscLogEventExcludeClass(a) 0 563 #define PetscLogEventActivate(a) 0 564 #define PetscLogEventDeactivate(a) 0 565 #define PetscLogEventDeactivatePush(a) 0 566 #define PetscLogEventDeactivatePop(a) 0 567 #define PetscLogEventActivateClass(a) 0 568 #define PetscLogEventDeactivateClass(a) 0 569 #define PetscLogEventSetActiveAll(a,b) 0 570 #define PetscLogEventGetId(a,b) (*(b)=0,0) 571 #define PetscLogEventGetPerfInfo(a,b,c) 0 572 #define PetscLogEventSetDof(a,b,c) 0 573 #define PetscLogEventSetError(a,b,c) 0 574 575 #define PetscLogPLB 0 576 #define PetscLogPLE 0 577 #define PetscLogPHC 0 578 #define PetscLogPHD 0 579 580 #define PetscLogObjectParents(p,n,c) 0 581 #define PetscLogObjectCreate(h) 0 582 #define PetscLogObjectDestroy(h) 0 583 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject,const char[],...) PETSC_ATTRIBUTE_FORMAT(2,3); 584 585 #define PetscLogDefaultBegin() 0 586 #define PetscLogAllBegin() 0 587 #define PetscLogNestedBegin() 0 588 #define PetscLogTraceBegin(file) 0 589 #define PetscLogActions(a) 0 590 #define PetscLogObjects(a) 0 591 #define PetscLogSetThreshold(a,b) 0 592 #define PetscLogSet(lb,le) 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 PetscLogGpuTimeBegin(void); 686 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeEnd(void); 687 688 #else 689 690 #define PetscLogCpuToGpu(a) 0 691 #define PetscLogGpuToCpu(a) 0 692 #define PetscLogCpuToGpuScalar(a) 0 693 #define PetscLogGpuToCpuScalar(a) 0 694 #define PetscLogGpuFlops(a) 0 695 #define PetscLogGpuTimeAdd(a) 0 696 #define PetscLogGpuTimeBegin() 0 697 #define PetscLogGpuTimeEnd() 0 698 699 #endif /* PETSC_USE_LOG && PETSC_HAVE_DEVICE */ 700 701 #define PetscPreLoadBegin(flag,name) \ 702 do {\ 703 PetscBool PetscPreLoading = flag;\ 704 int PetscPreLoadMax,PetscPreLoadIt;\ 705 PetscLogStage _stageNum;\ 706 PetscErrorCode _3_ierr; \ 707 _3_ierr = PetscOptionsGetBool(NULL,NULL,"-preload",&PetscPreLoading,NULL);CHKERRQ(_3_ierr); \ 708 PetscPreLoadMax = (int)(PetscPreLoading);\ 709 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 710 for (PetscPreLoadIt=0; PetscPreLoadIt<=PetscPreLoadMax; PetscPreLoadIt++) {\ 711 PetscPreLoadingOn = PetscPreLoading;\ 712 _3_ierr = PetscBarrier(NULL);CHKERRQ(_3_ierr);\ 713 if (PetscPreLoadIt>0) {\ 714 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 715 } else {\ 716 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 717 }\ 718 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt));\ 719 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 720 721 #define PetscPreLoadEnd() \ 722 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 723 PetscPreLoading = PETSC_FALSE;\ 724 }\ 725 } while (0) 726 727 #define PetscPreLoadStage(name) do { \ 728 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr); \ 729 if (PetscPreLoadIt>0) { \ 730 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr); \ 731 } else { \ 732 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 733 } \ 734 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt)); \ 735 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); \ 736 } while (0) 737 738 /* some vars for logging */ 739 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 740 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 741 742 #endif 743