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,S) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S) 14 #define PetscInfo1(A,S,a1) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1) 15 #define PetscInfo2(A,S,a1,a2) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2) 16 #define PetscInfo3(A,S,a1,a2,a3) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3) 17 #define PetscInfo4(A,S,a1,a2,a3,a4) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3,a4) 18 #define PetscInfo5(A,S,a1,a2,a3,a4,a5) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3,a4,a5) 19 #define PetscInfo6(A,S,a1,a2,a3,a4,a5,a6) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3,a4,a5,a6) 20 #define PetscInfo7(A,S,a1,a2,a3,a4,a5,a6,a7) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3,a4,a5,a6,a7) 21 #define PetscInfo8(A,S,a1,a2,a3,a4,a5,a6,a7,a8) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3,a4,a5,a6,a7,a8) 22 #define PetscInfo9(A,S,a1,a2,a3,a4,a5,a6,a7,a8,a9) PetscInfo_Private(PETSC_FUNCTION_NAME,((PetscObject)A),S,a1,a2,a3,a4,a5,a6,a7,a8,a9) 23 #else 24 #define PetscInfo(A,S) 0 25 #define PetscInfo1(A,S,a1) 0 26 #define PetscInfo2(A,S,a1,a2) 0 27 #define PetscInfo3(A,S,a1,a2,a3) 0 28 #define PetscInfo4(A,S,a1,a2,a3,a4) 0 29 #define PetscInfo5(A,S,a1,a2,a3,a4,a5) 0 30 #define PetscInfo6(A,S,a1,a2,a3,a4,a5,a6) 0 31 #define PetscInfo7(A,S,a1,a2,a3,a4,a5,a6,a7) 0 32 #define PetscInfo8(A,S,a1,a2,a3,a4,a5,a6,a7,a8) 0 33 #define PetscInfo9(A,S,a1,a2,a3,a4,a5,a6,a7,a8,a9) 0 34 #endif 35 36 /*E 37 PetscInfoCommFlag - Describes the method by which to filter PetscInfo() by communicator size 38 39 Used as an input for PetscInfoSetFilterCommSelf() 40 41 $ PETSC_INFO_COMM_ALL - Default uninitialized value. PetscInfo() will not filter based on communicator size (i.e. will 42 print for all communicators) 43 $ PETSC_INFO_COMM_NO_SELF - PetscInfo() will NOT print for communicators with size = 1 (i.e. *_COMM_SELF) 44 $ PETSC_INFO_COMM_ONLY_SELF - PetscInfo will ONLY print for communicators with size = 1 45 46 Level: intermediate 47 48 .seealso: PetscInfo(), PetscInfoSetFromOptions(), PetscInfoSetFilterCommSelf() 49 E*/ 50 typedef enum { 51 PETSC_INFO_COMM_ALL = -1, 52 PETSC_INFO_COMM_NO_SELF = 0, 53 PETSC_INFO_COMM_ONLY_SELF = 1 54 } PetscInfoCommFlag; 55 56 PETSC_EXTERN const char * const PetscInfoCommFlags[]; 57 PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId); 58 PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId); 59 PETSC_EXTERN PetscErrorCode PetscInfoEnabled(PetscClassId, PetscBool *); 60 PETSC_EXTERN PetscErrorCode PetscInfoAllow(PetscBool); 61 PETSC_EXTERN PetscErrorCode PetscInfoSetFile(const char[],const char[]); 62 PETSC_EXTERN PetscErrorCode PetscInfoGetFile(char **,FILE **); 63 PETSC_EXTERN PetscErrorCode PetscInfoSetClasses(PetscBool,PetscInt,const char *const *); 64 PETSC_EXTERN PetscErrorCode PetscInfoGetClass(const char *, PetscBool *); 65 PETSC_EXTERN PetscErrorCode PetscInfoGetInfo(PetscBool *,PetscBool *,PetscBool *,PetscBool *,PetscInfoCommFlag *); 66 PETSC_EXTERN PetscErrorCode PetscInfoProcessClass(const char[],PetscInt,PetscClassId[]); 67 PETSC_EXTERN PetscErrorCode PetscInfoSetFilterCommSelf(PetscInfoCommFlag); 68 PETSC_EXTERN PetscErrorCode PetscInfoSetFromOptions(PetscOptions); 69 PETSC_EXTERN PetscErrorCode PetscInfoDestroy(void); 70 PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */ 71 72 /*MC 73 PetscLogEvent - id used to identify PETSc or user events which timed portions (blocks of executable) 74 code. 75 76 Level: intermediate 77 78 .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscLogStage 79 M*/ 80 typedef int PetscLogEvent; 81 82 /*MC 83 PetscLogStage - id used to identify user stages (phases, sections) of runs - for logging 84 85 Level: intermediate 86 87 .seealso: PetscLogStageRegister(), PetscLogStagePush(), PetscLogStagePop(), PetscLogEvent 88 M*/ 89 typedef int PetscLogStage; 90 91 #define PETSC_EVENT 1311311 92 PETSC_EXTERN PetscLogEvent PETSC_LARGEST_EVENT; 93 94 /* Global flop counter */ 95 PETSC_EXTERN PetscLogDouble petsc_TotalFlops; 96 PETSC_EXTERN PetscLogDouble petsc_tmp_flops; 97 98 /* We must make the following structures available to access the event 99 activation flags in the PetscLogEventBegin/End() macros. These are not part of the PETSc public 100 API and are not intended to be used by other parts of PETSc or by users. 101 102 The code that manipulates these structures is in src/sys/logging/utils. 103 */ 104 typedef struct _n_PetscIntStack *PetscIntStack; 105 106 /* -----------------------------------------------------------------------------------------------------*/ 107 /* 108 PetscClassRegInfo, PetscClassPerfInfo - Each class has two data structures associated with it. The first has 109 static information about it, the second collects statistics on how many objects of the class are created, 110 how much memory they use, etc. 111 112 PetscClassRegLog, PetscClassPerfLog - arrays of the PetscClassRegInfo and PetscClassPerfInfo for all classes. 113 */ 114 typedef struct { 115 char *name; /* The class name */ 116 PetscClassId classid; /* The integer identifying this class */ 117 } PetscClassRegInfo; 118 119 typedef struct { 120 PetscClassId id; /* The integer identifying this class */ 121 int creations; /* The number of objects of this class created */ 122 int destructions; /* The number of objects of this class destroyed */ 123 PetscLogDouble mem; /* The total memory allocated by objects of this class */ 124 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 125 } PetscClassPerfInfo; 126 127 typedef struct _n_PetscClassRegLog *PetscClassRegLog; 128 struct _n_PetscClassRegLog { 129 int numClasses; /* The number of classes registered */ 130 int maxClasses; /* The maximum number of classes */ 131 PetscClassRegInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 132 }; 133 134 typedef struct _n_PetscClassPerfLog *PetscClassPerfLog; 135 struct _n_PetscClassPerfLog { 136 int numClasses; /* The number of logging classes */ 137 int maxClasses; /* The maximum number of classes */ 138 PetscClassPerfInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 139 }; 140 /* -----------------------------------------------------------------------------------------------------*/ 141 /* 142 PetscEventRegInfo, PetscEventPerfInfo - Each event has two data structures associated with it. The first has 143 static information about it, the second collects statistics on how many times the event is used, how 144 much time it takes, etc. 145 146 PetscEventRegLog, PetscEventPerfLog - an array of all PetscEventRegInfo and PetscEventPerfInfo for all events. There is one 147 of these for each stage. 148 149 */ 150 typedef struct { 151 char *name; /* The name of this event */ 152 PetscClassId classid; /* The class the event is associated with */ 153 PetscBool collective; /* Flag this event as collective */ 154 #if defined (PETSC_HAVE_MPE) 155 int mpe_id_begin; /* MPE IDs that define the event */ 156 int mpe_id_end; 157 #endif 158 } PetscEventRegInfo; 159 160 typedef struct { 161 int id; /* The integer identifying this event */ 162 PetscBool active; /* The flag to activate logging */ 163 PetscBool visible; /* The flag to print info in summary */ 164 int depth; /* The nesting depth of the event call */ 165 int count; /* The number of times this event was executed */ 166 PetscLogDouble flops, flops2, flopsTmp; /* The flops and flops^2 used in this event */ 167 PetscLogDouble time, time2, timeTmp; /* The time and time^2 taken for this event */ 168 PetscLogDouble syncTime; /* The synchronization barrier time */ 169 PetscLogDouble dof[8]; /* The number of degrees of freedom associated with this event */ 170 PetscLogDouble errors[8]; /* The errors (user-defined) associated with this event */ 171 PetscLogDouble numMessages; /* The number of messages in this event */ 172 PetscLogDouble messageLength; /* The total message lengths in this event */ 173 PetscLogDouble numReductions; /* The number of reductions in this event */ 174 PetscLogDouble memIncrease; /* How much the resident memory has increased in this event */ 175 PetscLogDouble mallocIncrease;/* How much the maximum malloced space has increased in this event */ 176 PetscLogDouble mallocSpace; /* How much the space was malloced and kept during this event */ 177 PetscLogDouble mallocIncreaseEvent; /* Maximum of the high water mark with in event minus memory available at the end of the event */ 178 #if defined(PETSC_HAVE_DEVICE) 179 PetscLogDouble CpuToGpuCount; /* The total number of CPU to GPU copies */ 180 PetscLogDouble GpuToCpuCount; /* The total number of GPU to CPU copies */ 181 PetscLogDouble CpuToGpuSize; /* The total size of CPU to GPU copies */ 182 PetscLogDouble GpuToCpuSize; /* The total size of GPU to CPU copies */ 183 PetscLogDouble GpuFlops; /* The flops done on a GPU in this event */ 184 PetscLogDouble GpuTime; /* The time spent on a GPU in this event */ 185 #endif 186 } PetscEventPerfInfo; 187 188 typedef struct _n_PetscEventRegLog *PetscEventRegLog; 189 struct _n_PetscEventRegLog { 190 int numEvents; /* The number of registered events */ 191 int maxEvents; /* The maximum number of events */ 192 PetscEventRegInfo *eventInfo; /* The registration information for each event */ 193 }; 194 195 typedef struct _n_PetscEventPerfLog *PetscEventPerfLog; 196 struct _n_PetscEventPerfLog { 197 int numEvents; /* The number of logging events */ 198 int maxEvents; /* The maximum number of events */ 199 PetscEventPerfInfo *eventInfo; /* The performance information for each event */ 200 }; 201 /* ------------------------------------------------------------------------------------------------------------*/ 202 /* 203 PetscStageInfo - Contains all the information about a particular stage. 204 205 PetscStageLog - An array of PetscStageInfo for each registered stage. There is a single one of these in the code. 206 */ 207 typedef struct _PetscStageInfo { 208 char *name; /* The stage name */ 209 PetscBool used; /* The stage was pushed on this processor */ 210 PetscEventPerfInfo perfInfo; /* The stage performance information */ 211 PetscEventPerfLog eventLog; /* The event information for this stage */ 212 PetscClassPerfLog classLog; /* The class information for this stage */ 213 } PetscStageInfo; 214 215 typedef struct _n_PetscStageLog *PetscStageLog; 216 struct _n_PetscStageLog { 217 int numStages; /* The number of registered stages */ 218 int maxStages; /* The maximum number of stages */ 219 PetscIntStack stack; /* The stack for active stages */ 220 int curStage; /* The current stage (only used in macros so we don't call PetscIntStackTop) */ 221 PetscStageInfo *stageInfo; /* The information for each stage */ 222 PetscEventRegLog eventLog; /* The registered events */ 223 PetscClassRegLog classLog; /* The registered classes */ 224 }; 225 /* -----------------------------------------------------------------------------------------------------*/ 226 227 PETSC_EXTERN PetscErrorCode PetscLogObjectParent(PetscObject,PetscObject); 228 PETSC_EXTERN PetscErrorCode PetscLogObjectMemory(PetscObject,PetscLogDouble); 229 230 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 231 PETSC_EXTERN PetscStageLog petsc_stageLog; 232 PETSC_EXTERN PetscErrorCode PetscLogGetStageLog(PetscStageLog*); 233 PETSC_EXTERN PetscErrorCode PetscStageLogGetCurrent(PetscStageLog,int*); 234 PETSC_EXTERN PetscErrorCode PetscStageLogGetEventPerfLog(PetscStageLog,int,PetscEventPerfLog*); 235 236 /* 237 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 238 239 For the complex numbers version, note that 240 1 complex addition = 2 flops 241 1 complex multiplication = 6 flops, 242 where we define 1 flop as that for a double precision scalar. We roughly approximate 243 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 244 to the assumption that we're counting mostly additions and multiplications -- and 245 roughly the same number of each. More accurate counting could be done by distinguishing 246 among the various arithmetic operations. 247 */ 248 249 #if defined(PETSC_USE_COMPLEX) 250 #define PETSC_FLOPS_PER_OP 4.0 251 #else 252 #define PETSC_FLOPS_PER_OP 1.0 253 #endif 254 255 /*@C 256 PetscLogFlops - Log how many flops are performed in a calculation 257 258 Input Parameter: 259 . flops - the number of flops 260 261 Notes: 262 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 263 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 264 265 Level: intermediate 266 267 .seealso: PetscLogView(), PetscLogGpuFlops() 268 @*/ 269 270 PETSC_STATIC_INLINE PetscErrorCode PetscLogFlops(PetscLogDouble n) 271 { 272 PetscFunctionBegin; 273 #if defined(PETSC_USE_DEBUG) 274 if (n < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 275 #endif 276 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 277 PetscFunctionReturn(0); 278 } 279 280 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 281 282 #if defined (PETSC_HAVE_MPE) 283 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 284 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 285 #endif 286 287 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 288 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 289 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 290 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 291 292 #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) 293 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : 0) 294 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : 0) 295 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2,3); 296 297 /* Initialization functions */ 298 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 299 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 300 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 301 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 302 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 303 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 304 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble,PetscLogDouble*); 305 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 306 PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 307 308 /* Output functions */ 309 PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer); 310 PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void); 311 PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]); 312 313 /* Stage functions */ 314 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[],PetscLogStage*); 315 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 316 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 317 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage,PetscBool); 318 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage,PetscBool*); 319 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage,PetscBool); 320 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage,PetscBool*); 321 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[],PetscLogStage*); 322 323 /* Event functions */ 324 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[],PetscClassId,PetscLogEvent*); 325 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent,PetscBool); 326 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 327 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 328 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 329 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 330 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePush(PetscLogEvent); 331 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePop(PetscLogEvent); 332 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent,PetscBool); 333 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 334 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 335 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[],PetscLogEvent*); 336 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int,PetscLogEvent,PetscEventPerfInfo*); 337 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 338 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 339 340 /* Global counters */ 341 PETSC_EXTERN PetscLogDouble petsc_irecv_ct; 342 PETSC_EXTERN PetscLogDouble petsc_isend_ct; 343 PETSC_EXTERN PetscLogDouble petsc_recv_ct; 344 PETSC_EXTERN PetscLogDouble petsc_send_ct; 345 PETSC_EXTERN PetscLogDouble petsc_irecv_len; 346 PETSC_EXTERN PetscLogDouble petsc_isend_len; 347 PETSC_EXTERN PetscLogDouble petsc_recv_len; 348 PETSC_EXTERN PetscLogDouble petsc_send_len; 349 PETSC_EXTERN PetscLogDouble petsc_allreduce_ct; 350 PETSC_EXTERN PetscLogDouble petsc_gather_ct; 351 PETSC_EXTERN PetscLogDouble petsc_scatter_ct; 352 PETSC_EXTERN PetscLogDouble petsc_wait_ct; 353 PETSC_EXTERN PetscLogDouble petsc_wait_any_ct; 354 PETSC_EXTERN PetscLogDouble petsc_wait_all_ct; 355 PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct; 356 357 PETSC_EXTERN PetscBool PetscLogMemory; 358 359 PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */ 360 PETSC_EXTERN PetscErrorCode PetscLogEventSynchronize(PetscLogEvent, MPI_Comm); 361 362 #define PetscLogEventSync(e,comm) \ 363 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 364 PetscLogEventSynchronize((e),(comm)) : 0)) 365 366 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 367 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 368 (*PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0)) 369 370 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 371 (((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 372 (*PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0)) 373 374 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent,PetscLogDouble*); 375 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 376 377 /* 378 These are used internally in the PETSc routines to keep a count of MPI messages and 379 their sizes. 380 381 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 382 uses macros to defined the MPI operations. 383 384 It does not work correctly from HP-UX because it processes the 385 macros in a way that sometimes it double counts, hence 386 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 387 388 It does not work with Windows because winmpich lacks MPI_Type_size() 389 */ 390 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 391 /* 392 Logging of MPI activities 393 */ 394 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSize(PetscInt count,MPI_Datatype type,PetscLogDouble *length) 395 { 396 PetscMPIInt typesize; 397 PetscErrorCode ierr; 398 if (type == MPI_DATATYPE_NULL) return 0; 399 ierr = MPI_Type_size(type,&typesize);CHKERRMPI(ierr); 400 *length += (PetscLogDouble) (count*typesize); 401 return 0; 402 } 403 404 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 405 { 406 PetscMPIInt typesize,size,p; 407 PetscErrorCode ierr; 408 409 if (type == MPI_DATATYPE_NULL) return 0; 410 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 411 ierr = MPI_Type_size(type,&typesize);CHKERRMPI(ierr); 412 for (p=0; p<size; ++p) { 413 *length += (PetscLogDouble) (counts[p]*typesize); 414 } 415 return 0; 416 } 417 418 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSizeCount(PetscInt n,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 419 { 420 PetscMPIInt typesize,p; 421 PetscErrorCode ierr; 422 423 if (type == MPI_DATATYPE_NULL) return 0; 424 ierr = MPI_Type_size(type,&typesize);CHKERRMPI(ierr); 425 for (p=0; p<n; ++p) { 426 *length += (PetscLogDouble) (counts[p]*typesize); 427 } 428 return 0; 429 } 430 431 /* 432 Returns 1 if the communicator is parallel else zero 433 */ 434 PETSC_STATIC_INLINE int PetscMPIParallelComm(MPI_Comm comm) 435 { 436 PetscMPIInt size; MPI_Comm_size(comm,&size); return size > 1; 437 } 438 439 #define MPI_Irecv(buf,count,datatype,source,tag,comm,request) \ 440 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv((buf),(count),(datatype),(source),(tag),(comm),(request))) 441 442 #define MPI_Irecv_c(buf,count,datatype,source,tag,comm,request) \ 443 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv_c((buf),(count),(datatype),(source),(tag),(comm),(request))) 444 445 #define MPI_Isend(buf,count,datatype,dest,tag,comm,request) \ 446 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend((buf),(count),(datatype),(dest),(tag),(comm),(request))) 447 448 #define MPI_Isend_c(buf,count,datatype,dest,tag,comm,request) \ 449 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend_c((buf),(count),(datatype),(dest),(tag),(comm),(request))) 450 451 #define MPI_Startall_irecv(count,datatype,number,requests) \ 452 ((petsc_irecv_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || ((number) && MPI_Startall((number),(requests)))) 453 454 #define MPI_Startall_isend(count,datatype,number,requests) \ 455 ((petsc_isend_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || ((number) && MPI_Startall((number),(requests)))) 456 457 #define MPI_Start_isend(count,datatype,requests) \ 458 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_isend_len)) || MPI_Start((requests))) 459 460 #define MPI_Recv(buf,count,datatype,source,tag,comm,status) \ 461 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv((buf),(count),(datatype),(source),(tag),(comm),(status))) 462 463 #define MPI_Recv_c(buf,count,datatype,source,tag,comm,status) \ 464 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv_c((buf),(count),(datatype),(source),(tag),(comm),(status))) 465 466 #define MPI_Send(buf,count,datatype,dest,tag,comm) \ 467 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send((buf),(count),(datatype),(dest),(tag),(comm))) 468 469 #define MPI_Send_c(buf,count,datatype,dest,tag,comm) \ 470 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send_c((buf),(count),(datatype),(dest),(tag),(comm))) 471 472 #define MPI_Wait(request,status) \ 473 ((petsc_wait_ct++,petsc_sum_of_waits_ct++,0) || MPI_Wait((request),(status))) 474 475 #define MPI_Waitany(a,b,c,d) \ 476 ((petsc_wait_any_ct++,petsc_sum_of_waits_ct++,0) || MPI_Waitany((a),(b),(c),(d))) 477 478 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 479 ((petsc_wait_all_ct++,petsc_sum_of_waits_ct += (PetscLogDouble) (count),0) || MPI_Waitall((count),(array_of_requests),(array_of_statuses))) 480 481 #define MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm) \ 482 (petsc_allreduce_ct += PetscMPIParallelComm((comm)),MPI_Allreduce((sendbuf),(recvbuf),(count),(datatype),(op),(comm))) 483 484 #define MPI_Bcast(buffer,count,datatype,root,comm) \ 485 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Bcast((buffer),(count),(datatype),(root),(comm))) 486 487 #define MPI_Reduce_scatter_block(sendbuf,recvbuf,recvcount,datatype,op,comm) \ 488 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Reduce_scatter_block((sendbuf),(recvbuf),(recvcount),(datatype),(op),(comm))) 489 490 #define MPI_Alltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 491 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Alltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 492 493 #define MPI_Alltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 494 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSizeComm((comm),(sendcnts),(sendtype),(&petsc_send_len)) || MPI_Alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 495 496 #define MPI_Allgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 497 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 498 499 #define MPI_Allgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm) \ 500 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm))) 501 502 #define MPI_Gather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 503 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 504 505 #define MPI_Gatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm) \ 506 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm))) 507 508 #define MPI_Scatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 509 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 510 511 #define MPI_Scatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 512 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 513 514 #define MPI_Ialltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 515 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Ialltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 516 517 #define MPI_Ialltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 518 ((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))) 519 520 #define MPI_Iallgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 521 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 522 523 #define MPI_Iallgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm,request) \ 524 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm),(request))) 525 526 #define MPI_Igather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 527 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 528 529 #define MPI_Igatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm,request) \ 530 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm),(request))) 531 532 #define MPI_Iscatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 533 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 534 535 #define MPI_Iscatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 536 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 537 538 #else 539 540 #define MPI_Startall_irecv(count,datatype,number,requests) \ 541 ((number) && MPI_Startall((number),(requests))) 542 543 #define MPI_Startall_isend(count,datatype,number,requests) \ 544 ((number) && MPI_Startall((number),(requests))) 545 546 #define MPI_Start_isend(count,datatype,requests) \ 547 (MPI_Start((requests))) 548 549 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 550 551 #else /* ---Logging is turned off --------------------------------------------*/ 552 553 #define PetscLogMemory PETSC_FALSE 554 555 #define PetscLogFlops(n) 0 556 #define PetscGetFlops(a) (*(a) = 0.0,0) 557 558 #define PetscLogStageRegister(a,b) 0 559 #define PetscLogStagePush(a) 0 560 #define PetscLogStagePop() 0 561 #define PetscLogStageSetActive(a,b) 0 562 #define PetscLogStageGetActive(a,b) 0 563 #define PetscLogStageGetVisible(a,b) 0 564 #define PetscLogStageSetVisible(a,b) 0 565 #define PetscLogStageGetId(a,b) (*(b)=0,0) 566 567 #define PetscLogEventRegister(a,b,c) 0 568 #define PetscLogEventSetCollective(a,b) 0 569 #define PetscLogEventIncludeClass(a) 0 570 #define PetscLogEventExcludeClass(a) 0 571 #define PetscLogEventActivate(a) 0 572 #define PetscLogEventDeactivate(a) 0 573 #define PetscLogEventDeactivatePush(a) 0 574 #define PetscLogEventDeactivatePop(a) 0 575 #define PetscLogEventActivateClass(a) 0 576 #define PetscLogEventDeactivateClass(a) 0 577 #define PetscLogEventSetActiveAll(a,b) 0 578 #define PetscLogEventGetId(a,b) (*(b)=0,0) 579 #define PetscLogEventGetPerfInfo(a,b,c) 0 580 #define PetscLogEventSetDof(a,b,c) 0 581 #define PetscLogEventSetError(a,b,c) 0 582 583 #define PetscLogPLB 0 584 #define PetscLogPLE 0 585 #define PetscLogPHC 0 586 #define PetscLogPHD 0 587 588 #define PetscLogObjectParents(p,n,c) 0 589 #define PetscLogObjectCreate(h) 0 590 #define PetscLogObjectDestroy(h) 0 591 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject,const char[],...) PETSC_ATTRIBUTE_FORMAT(2,3); 592 593 #define PetscLogDefaultBegin() 0 594 #define PetscLogAllBegin() 0 595 #define PetscLogNestedBegin() 0 596 #define PetscLogTraceBegin(file) 0 597 #define PetscLogActions(a) 0 598 #define PetscLogObjects(a) 0 599 #define PetscLogSetThreshold(a,b) 0 600 #define PetscLogSet(lb,le) 0 601 602 #define PetscLogView(viewer) 0 603 #define PetscLogViewFromOptions() 0 604 #define PetscLogDump(c) 0 605 606 #define PetscLogEventSync(e,comm) 0 607 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0 608 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0 609 610 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 611 #define MPI_Startall_irecv(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 612 #define MPI_Startall_isend(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 613 #define MPI_Start_isend(count,datatype,requests) MPI_Start(requests) 614 615 #endif /* PETSC_USE_LOG */ 616 617 #if defined (PETSC_USE_LOG) && defined(PETSC_HAVE_DEVICE) 618 619 /* Global GPU counters */ 620 PETSC_EXTERN PetscLogDouble petsc_ctog_ct; 621 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct; 622 PETSC_EXTERN PetscLogDouble petsc_ctog_sz; 623 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz; 624 PETSC_EXTERN PetscLogDouble petsc_ctog_ct_scalar; 625 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct_scalar; 626 PETSC_EXTERN PetscLogDouble petsc_ctog_sz_scalar; 627 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz_scalar; 628 PETSC_EXTERN PetscLogDouble petsc_gflops; 629 PETSC_EXTERN PetscLogDouble petsc_gtime; 630 631 PETSC_STATIC_INLINE PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size) 632 { 633 PetscFunctionBegin; 634 petsc_ctog_ct += 1; 635 petsc_ctog_sz += size; 636 PetscFunctionReturn(0); 637 } 638 639 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size) 640 { 641 PetscFunctionBegin; 642 petsc_gtoc_ct += 1; 643 petsc_gtoc_sz += size; 644 PetscFunctionReturn(0); 645 } 646 647 PETSC_STATIC_INLINE PetscErrorCode PetscLogCpuToGpuScalar(PetscLogDouble size) 648 { 649 PetscFunctionBegin; 650 petsc_ctog_ct_scalar += 1; 651 petsc_ctog_sz_scalar += size; 652 PetscFunctionReturn(0); 653 } 654 655 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuToCpuScalar(PetscLogDouble size) 656 { 657 PetscFunctionBegin; 658 petsc_gtoc_ct_scalar += 1; 659 petsc_gtoc_sz_scalar += size; 660 PetscFunctionReturn(0); 661 } 662 663 /*@C 664 PetscLogGpuFlops - Log how many flops are performed in a calculation on the device 665 666 Input Parameter: 667 . flops - the number of flops 668 669 Notes: 670 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 671 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 672 673 Level: intermediate 674 675 .seealso: PetscLogView(), PetscLogFlops(), PetscLogGpuTimeBegin(), PetscLogGpuTimeEnd() 676 @*/ 677 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuFlops(PetscLogDouble n) 678 { 679 PetscFunctionBegin; 680 if (PetscUnlikelyDebug(n < 0)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 681 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 682 petsc_gflops += PETSC_FLOPS_PER_OP*n; 683 PetscFunctionReturn(0); 684 } 685 686 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t) 687 { 688 PetscFunctionBegin; 689 petsc_gtime += t; 690 PetscFunctionReturn(0); 691 } 692 693 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeBegin(void); 694 PETSC_EXTERN PetscErrorCode PetscLogGpuTimeEnd(void); 695 696 #else 697 698 #define PetscLogCpuToGpu(a) 0 699 #define PetscLogGpuToCpu(a) 0 700 #define PetscLogCpuToGpuScalar(a) 0 701 #define PetscLogGpuToCpuScalar(a) 0 702 #define PetscLogGpuFlops(a) 0 703 #define PetscLogGpuTimeAdd(a) 0 704 #define PetscLogGpuTimeBegin() 0 705 #define PetscLogGpuTimeEnd() 0 706 707 #endif /* PETSC_USE_LOG && PETSC_HAVE_DEVICE */ 708 709 #define PetscPreLoadBegin(flag,name) \ 710 do {\ 711 PetscBool PetscPreLoading = flag;\ 712 int PetscPreLoadMax,PetscPreLoadIt;\ 713 PetscLogStage _stageNum;\ 714 PetscErrorCode _3_ierr; \ 715 _3_ierr = PetscOptionsGetBool(NULL,NULL,"-preload",&PetscPreLoading,NULL);CHKERRQ(_3_ierr); \ 716 PetscPreLoadMax = (int)(PetscPreLoading);\ 717 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 718 for (PetscPreLoadIt=0; PetscPreLoadIt<=PetscPreLoadMax; PetscPreLoadIt++) {\ 719 PetscPreLoadingOn = PetscPreLoading;\ 720 _3_ierr = PetscBarrier(NULL);CHKERRQ(_3_ierr);\ 721 if (PetscPreLoadIt>0) {\ 722 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 723 } else {\ 724 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 725 }\ 726 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt));\ 727 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 728 729 #define PetscPreLoadEnd() \ 730 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 731 PetscPreLoading = PETSC_FALSE;\ 732 }\ 733 } while (0) 734 735 #define PetscPreLoadStage(name) do { \ 736 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr); \ 737 if (PetscPreLoadIt>0) { \ 738 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr); \ 739 } else { \ 740 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 741 } \ 742 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt)); \ 743 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); \ 744 } while (0) 745 746 /* some vars for logging */ 747 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 748 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 749 750 #endif 751