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[],void*,const char[],...); 12 #if defined(PETSC_USE_INFO) 13 #define PetscInfo(A,S) PetscInfo_Private(PETSC_FUNCTION_NAME,A,S) 14 #define PetscInfo1(A,S,a1) PetscInfo_Private(PETSC_FUNCTION_NAME,A,S,a1) 15 #define PetscInfo2(A,S,a1,a2) PetscInfo_Private(PETSC_FUNCTION_NAME,A,S,a1,a2) 16 #define PetscInfo3(A,S,a1,a2,a3) PetscInfo_Private(PETSC_FUNCTION_NAME,A,S,a1,a2,a3) 17 #define PetscInfo4(A,S,a1,a2,a3,a4) PetscInfo_Private(PETSC_FUNCTION_NAME,A,S,a1,a2,a3,a4) 18 #define PetscInfo5(A,S,a1,a2,a3,a4,a5) PetscInfo_Private(PETSC_FUNCTION_NAME,A,S,a1,a2,a3,a4,a5) 19 #define PetscInfo6(A,S,a1,a2,a3,a4,a5,a6) PetscInfo_Private(PETSC_FUNCTION_NAME,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,A,S,a1,a2,a3,a4,a5,a6,a7) 21 #else 22 #define PetscInfo(A,S) 0 23 #define PetscInfo1(A,S,a1) 0 24 #define PetscInfo2(A,S,a1,a2) 0 25 #define PetscInfo3(A,S,a1,a2,a3) 0 26 #define PetscInfo4(A,S,a1,a2,a3,a4) 0 27 #define PetscInfo5(A,S,a1,a2,a3,a4,a5) 0 28 #define PetscInfo6(A,S,a1,a2,a3,a4,a5,a6) 0 29 #define PetscInfo7(A,S,a1,a2,a3,a4,a5,a6,a7) 0 30 #endif 31 PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId); 32 PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId); 33 PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */ 34 35 /*MC 36 PetscLogEvent - id used to identify PETSc or user events which timed portions (blocks of executable) 37 code. 38 39 Level: intermediate 40 41 .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscLogStage 42 M*/ 43 typedef int PetscLogEvent; 44 45 /*MC 46 PetscLogStage - id used to identify user stages (phases, sections) of runs - for logging 47 48 Level: intermediate 49 50 .seealso: PetscLogStageRegister(), PetscLogStagePush(), PetscLogStagePop(), PetscLogEvent 51 M*/ 52 typedef int PetscLogStage; 53 54 #define PETSC_EVENT 1311311 55 PETSC_EXTERN PetscLogEvent PETSC_LARGEST_EVENT; 56 57 /* Global flop counter */ 58 PETSC_EXTERN PetscLogDouble petsc_TotalFlops; 59 PETSC_EXTERN PetscLogDouble petsc_tmp_flops; 60 61 /* Global GPU counters */ 62 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 63 PETSC_EXTERN PetscLogDouble petsc_ctog_ct; 64 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct; 65 PETSC_EXTERN PetscLogDouble petsc_ctog_sz; 66 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz; 67 PETSC_EXTERN PetscLogDouble petsc_gflops; 68 PETSC_EXTERN PetscLogDouble petsc_gtime; 69 #endif 70 71 /* We must make the following structures available to access the event 72 activation flags in the PetscLogEventBegin/End() macros. These are not part of the PETSc public 73 API and are not intended to be used by other parts of PETSc or by users. 74 75 The code that manipulates these structures is in src/sys/logging/utils. 76 */ 77 typedef struct _n_PetscIntStack *PetscIntStack; 78 79 /* -----------------------------------------------------------------------------------------------------*/ 80 /* 81 PetscClassRegInfo, PetscClassPerfInfo - Each class has two data structures associated with it. The first has 82 static information about it, the second collects statistics on how many objects of the class are created, 83 how much memory they use, etc. 84 85 PetscClassRegLog, PetscClassPerfLog - arrays of the PetscClassRegInfo and PetscClassPerfInfo for all classes. 86 */ 87 typedef struct { 88 char *name; /* The class name */ 89 PetscClassId classid; /* The integer identifying this class */ 90 } PetscClassRegInfo; 91 92 typedef struct { 93 PetscClassId id; /* The integer identifying this class */ 94 int creations; /* The number of objects of this class created */ 95 int destructions; /* The number of objects of this class destroyed */ 96 PetscLogDouble mem; /* The total memory allocated by objects of this class */ 97 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 98 } PetscClassPerfInfo; 99 100 typedef struct _n_PetscClassRegLog *PetscClassRegLog; 101 struct _n_PetscClassRegLog { 102 int numClasses; /* The number of classes registered */ 103 int maxClasses; /* The maximum number of classes */ 104 PetscClassRegInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 105 }; 106 107 typedef struct _n_PetscClassPerfLog *PetscClassPerfLog; 108 struct _n_PetscClassPerfLog { 109 int numClasses; /* The number of logging classes */ 110 int maxClasses; /* The maximum number of classes */ 111 PetscClassPerfInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 112 }; 113 /* -----------------------------------------------------------------------------------------------------*/ 114 /* 115 PetscEventRegInfo, PetscEventPerfInfo - Each event has two data structures associated with it. The first has 116 static information about it, the second collects statistics on how many times the event is used, how 117 much time it takes, etc. 118 119 PetscEventRegLog, PetscEventPerfLog - an array of all PetscEventRegInfo and PetscEventPerfInfo for all events. There is one 120 of these for each stage. 121 122 */ 123 typedef struct { 124 char *name; /* The name of this event */ 125 PetscClassId classid; /* The class the event is associated with */ 126 PetscBool collective; /* Flag this event as collective */ 127 #if defined (PETSC_HAVE_MPE) 128 int mpe_id_begin; /* MPE IDs that define the event */ 129 int mpe_id_end; 130 #endif 131 } PetscEventRegInfo; 132 133 typedef struct { 134 int id; /* The integer identifying this event */ 135 PetscBool active; /* The flag to activate logging */ 136 PetscBool visible; /* The flag to print info in summary */ 137 int depth; /* The nesting depth of the event call */ 138 int count; /* The number of times this event was executed */ 139 PetscLogDouble flops, flops2, flopsTmp; /* The flops and flops^2 used in this event */ 140 PetscLogDouble time, time2, timeTmp; /* The time and time^2 taken for this event */ 141 PetscLogDouble syncTime; /* The synchronization barrier time */ 142 PetscLogDouble dof[8]; /* The number of degrees of freedom associated with this event */ 143 PetscLogDouble errors[8]; /* The errors (user-defined) associated with this event */ 144 PetscLogDouble numMessages; /* The number of messages in this event */ 145 PetscLogDouble messageLength; /* The total message lengths in this event */ 146 PetscLogDouble numReductions; /* The number of reductions in this event */ 147 PetscLogDouble memIncrease; /* How much the resident memory has increased in this event */ 148 PetscLogDouble mallocIncrease;/* How much the maximum malloced space has increased in this event */ 149 PetscLogDouble mallocSpace; /* How much the space was malloced and kept during this event */ 150 PetscLogDouble mallocIncreaseEvent; /* Maximum of the high water mark with in event minus memory available at the end of the event */ 151 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 152 PetscLogDouble CpuToGpuCount; /* The total number of CPU to GPU copies */ 153 PetscLogDouble GpuToCpuCount; /* The total number of GPU to CPU copies */ 154 PetscLogDouble CpuToGpuSize; /* The total size of CPU to GPU copies */ 155 PetscLogDouble GpuToCpuSize; /* The total size of GPU to CPU copies */ 156 PetscLogDouble GpuFlops; /* The flops done on a GPU in this event */ 157 PetscLogDouble GpuTime; /* The time spent on a GPU in this event */ 158 #endif 159 } PetscEventPerfInfo; 160 161 typedef struct _n_PetscEventRegLog *PetscEventRegLog; 162 struct _n_PetscEventRegLog { 163 int numEvents; /* The number of registered events */ 164 int maxEvents; /* The maximum number of events */ 165 PetscEventRegInfo *eventInfo; /* The registration information for each event */ 166 }; 167 168 typedef struct _n_PetscEventPerfLog *PetscEventPerfLog; 169 struct _n_PetscEventPerfLog { 170 int numEvents; /* The number of logging events */ 171 int maxEvents; /* The maximum number of events */ 172 PetscEventPerfInfo *eventInfo; /* The performance information for each event */ 173 }; 174 /* ------------------------------------------------------------------------------------------------------------*/ 175 /* 176 PetscStageInfo - Contains all the information about a particular stage. 177 178 PetscStageLog - An array of PetscStageInfo for each registered stage. There is a single one of these in the code. 179 */ 180 typedef struct _PetscStageInfo { 181 char *name; /* The stage name */ 182 PetscBool used; /* The stage was pushed on this processor */ 183 PetscEventPerfInfo perfInfo; /* The stage performance information */ 184 PetscEventPerfLog eventLog; /* The event information for this stage */ 185 PetscClassPerfLog classLog; /* The class information for this stage */ 186 } PetscStageInfo; 187 188 typedef struct _n_PetscStageLog *PetscStageLog; 189 struct _n_PetscStageLog { 190 int numStages; /* The number of registered stages */ 191 int maxStages; /* The maximum number of stages */ 192 PetscIntStack stack; /* The stack for active stages */ 193 int curStage; /* The current stage (only used in macros so we don't call PetscIntStackTop) */ 194 PetscStageInfo *stageInfo; /* The information for each stage */ 195 PetscEventRegLog eventLog; /* The registered events */ 196 PetscClassRegLog classLog; /* The registered classes */ 197 }; 198 /* -----------------------------------------------------------------------------------------------------*/ 199 200 PETSC_EXTERN PetscErrorCode PetscLogObjectParent(PetscObject,PetscObject); 201 PETSC_EXTERN PetscErrorCode PetscLogObjectMemory(PetscObject,PetscLogDouble); 202 203 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 204 PETSC_EXTERN PetscStageLog petsc_stageLog; 205 PETSC_EXTERN PetscErrorCode PetscLogGetStageLog(PetscStageLog*); 206 PETSC_EXTERN PetscErrorCode PetscStageLogGetCurrent(PetscStageLog,int*); 207 PETSC_EXTERN PetscErrorCode PetscStageLogGetEventPerfLog(PetscStageLog,int,PetscEventPerfLog*); 208 209 /* 210 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 211 212 For the complex numbers version, note that 213 1 complex addition = 2 flops 214 1 complex multiplication = 6 flops, 215 where we define 1 flop as that for a double precision scalar. We roughly approximate 216 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 217 to the assumption that we're counting mostly additions and multiplications -- and 218 roughly the same number of each. More accurate counting could be done by distinguishing 219 among the various arithmetic operations. 220 */ 221 222 #if defined(PETSC_USE_COMPLEX) 223 #define PETSC_FLOPS_PER_OP 4.0 224 #else 225 #define PETSC_FLOPS_PER_OP 1.0 226 #endif 227 228 PETSC_STATIC_INLINE PetscErrorCode PetscLogFlops(PetscLogDouble n) 229 { 230 PetscFunctionBegin; 231 #if defined(PETSC_USE_DEBUG) 232 if (n < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 233 #endif 234 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 235 PetscFunctionReturn(0); 236 } 237 238 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 239 PETSC_STATIC_INLINE PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size) 240 { 241 PetscFunctionBegin; 242 petsc_ctog_ct += 1; 243 petsc_ctog_sz += size; 244 PetscFunctionReturn(0); 245 } 246 247 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size) 248 { 249 PetscFunctionBegin; 250 petsc_gtoc_ct += 1; 251 petsc_gtoc_sz += size; 252 PetscFunctionReturn(0); 253 } 254 255 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuFlops(PetscLogDouble n) 256 { 257 PetscFunctionBegin; 258 #if defined(PETSC_USE_DEBUG) 259 if (n < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 260 #endif 261 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 262 petsc_gflops += PETSC_FLOPS_PER_OP*n; 263 PetscFunctionReturn(0); 264 } 265 266 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeBegin() 267 { 268 PetscErrorCode ierr; 269 PetscFunctionBegin; 270 ierr = PetscTimeSubtract(&petsc_gtime);CHKERRQ(ierr); 271 PetscFunctionReturn(0); 272 } 273 274 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeEnd() 275 { 276 PetscErrorCode ierr; 277 PetscFunctionBegin; 278 ierr = PetscTimeAdd(&petsc_gtime);CHKERRQ(ierr); 279 PetscFunctionReturn(0); 280 } 281 282 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t) 283 { 284 PetscFunctionBegin; 285 petsc_gtime += t; 286 PetscFunctionReturn(0); 287 } 288 #endif 289 290 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 291 292 #if defined (PETSC_HAVE_MPE) 293 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 294 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 295 #endif 296 297 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 298 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 299 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 300 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 301 302 #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) 303 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : 0) 304 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : 0) 305 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...); 306 307 /* Initialization functions */ 308 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 309 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 310 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 311 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 312 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 313 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 314 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble,PetscLogDouble*); 315 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 316 PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 317 318 /* Output functions */ 319 PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer); 320 PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void); 321 PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]); 322 323 /* Stage functions */ 324 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[],PetscLogStage*); 325 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 326 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 327 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage,PetscBool); 328 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage,PetscBool*); 329 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage,PetscBool); 330 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage,PetscBool*); 331 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[],PetscLogStage*); 332 333 /* Event functions */ 334 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[],PetscClassId,PetscLogEvent*); 335 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent,PetscBool); 336 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 337 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 338 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 339 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 340 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent,PetscBool); 341 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 342 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 343 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[],PetscLogEvent*); 344 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int,PetscLogEvent,PetscEventPerfInfo*); 345 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 346 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 347 348 /* Global counters */ 349 PETSC_EXTERN PetscLogDouble petsc_irecv_ct; 350 PETSC_EXTERN PetscLogDouble petsc_isend_ct; 351 PETSC_EXTERN PetscLogDouble petsc_recv_ct; 352 PETSC_EXTERN PetscLogDouble petsc_send_ct; 353 PETSC_EXTERN PetscLogDouble petsc_irecv_len; 354 PETSC_EXTERN PetscLogDouble petsc_isend_len; 355 PETSC_EXTERN PetscLogDouble petsc_recv_len; 356 PETSC_EXTERN PetscLogDouble petsc_send_len; 357 PETSC_EXTERN PetscLogDouble petsc_allreduce_ct; 358 PETSC_EXTERN PetscLogDouble petsc_gather_ct; 359 PETSC_EXTERN PetscLogDouble petsc_scatter_ct; 360 PETSC_EXTERN PetscLogDouble petsc_wait_ct; 361 PETSC_EXTERN PetscLogDouble petsc_wait_any_ct; 362 PETSC_EXTERN PetscLogDouble petsc_wait_all_ct; 363 PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct; 364 365 PETSC_EXTERN PetscBool PetscLogMemory; 366 367 PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */ 368 PETSC_EXTERN PetscErrorCode PetscLogEventSynchronize(PetscLogEvent, MPI_Comm); 369 370 #define PetscLogEventSync(e,comm) \ 371 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 372 PetscLogEventSynchronize((e),(comm)) : 0 )) 373 374 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 375 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 376 (*PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0 )) 377 378 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 379 (((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 380 (*PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0 )) 381 382 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent,PetscLogDouble*); 383 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 384 385 /* 386 These are used internally in the PETSc routines to keep a count of MPI messages and 387 their sizes. 388 389 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 390 uses macros to defined the MPI operations. 391 392 It does not work correctly from HP-UX because it processes the 393 macros in a way that sometimes it double counts, hence 394 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 395 396 It does not work with Windows because winmpich lacks MPI_Type_size() 397 */ 398 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 399 /* 400 Logging of MPI activities 401 */ 402 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSize(PetscInt count,MPI_Datatype type,PetscLogDouble *length) 403 { 404 PetscMPIInt typesize; 405 PetscErrorCode ierr; 406 if (type == MPI_DATATYPE_NULL) return 0; 407 ierr = MPI_Type_size(type,&typesize);CHKERRQ(ierr); 408 *length += (PetscLogDouble) (count*typesize); 409 return 0; 410 } 411 412 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 413 { 414 PetscMPIInt typesize,size,p; 415 PetscErrorCode ierr; 416 417 if (type == MPI_DATATYPE_NULL) return 0; 418 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 419 ierr = MPI_Type_size(type,&typesize);CHKERRQ(ierr); 420 for (p=0; p<size; ++p) { 421 *length += (PetscLogDouble) (counts[p]*typesize); 422 } 423 return 0; 424 } 425 426 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSizeCount(PetscInt n,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 427 { 428 PetscMPIInt typesize,p; 429 PetscErrorCode ierr; 430 431 if (type == MPI_DATATYPE_NULL) return 0; 432 ierr = MPI_Type_size(type,&typesize);CHKERRQ(ierr); 433 for (p=0; p<n; ++p) { 434 *length += (PetscLogDouble) (counts[p]*typesize); 435 } 436 return 0; 437 } 438 439 /* 440 Returns 1 if the communicator is parallel else zero 441 */ 442 PETSC_STATIC_INLINE int PetscMPIParallelComm(MPI_Comm comm) 443 { 444 PetscMPIInt size; MPI_Comm_size(comm,&size); return size > 1; 445 } 446 447 #define MPI_Irecv(buf,count,datatype,source,tag,comm,request) \ 448 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv((buf),(count),(datatype),(source),(tag),(comm),(request))) 449 450 #define MPI_Isend(buf,count,datatype,dest,tag,comm,request) \ 451 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend((buf),(count),(datatype),(dest),(tag),(comm),(request))) 452 453 #define MPI_Startall_irecv(count,datatype,number,requests) \ 454 ((petsc_irecv_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || ((number) && MPI_Startall((number),(requests)))) 455 456 #define MPI_Startall_isend(count,datatype,number,requests) \ 457 ((petsc_isend_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || ((number) && MPI_Startall((number),(requests)))) 458 459 #define MPI_Start_isend(count,datatype,requests) \ 460 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_isend_len)) || MPI_Start((requests))) 461 462 #define MPI_Recv(buf,count,datatype,source,tag,comm,status) \ 463 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv((buf),(count),(datatype),(source),(tag),(comm),(status))) 464 465 #define MPI_Send(buf,count,datatype,dest,tag,comm) \ 466 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send((buf),(count),(datatype),(dest),(tag),(comm))) 467 468 #define MPI_Wait(request,status) \ 469 ((petsc_wait_ct++,petsc_sum_of_waits_ct++,0) || MPI_Wait((request),(status))) 470 471 #define MPI_Waitany(a,b,c,d) \ 472 ((petsc_wait_any_ct++,petsc_sum_of_waits_ct++,0) || MPI_Waitany((a),(b),(c),(d))) 473 474 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 475 ((petsc_wait_all_ct++,petsc_sum_of_waits_ct += (PetscLogDouble) (count),0) || MPI_Waitall((count),(array_of_requests),(array_of_statuses))) 476 477 #define MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm) \ 478 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Allreduce((sendbuf),(recvbuf),(count),(datatype),(op),(comm))) 479 480 #define MPI_Bcast(buffer,count,datatype,root,comm) \ 481 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Bcast((buffer),(count),(datatype),(root),(comm))) 482 483 #define MPI_Reduce_scatter_block(sendbuf,recvbuf,recvcount,datatype,op,comm) \ 484 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Reduce_scatter_block((sendbuf),(recvbuf),(recvcount),(datatype),(op),(comm))) 485 486 #define MPI_Alltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 487 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Alltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 488 489 #define MPI_Alltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 490 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSizeComm((comm),(sendcnts),(sendtype),(&petsc_send_len)) || MPI_Alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 491 492 #define MPI_Allgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 493 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 494 495 #define MPI_Allgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm) \ 496 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm))) 497 498 #define MPI_Gather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 499 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 500 501 #define MPI_Gatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm) \ 502 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm))) 503 504 #define MPI_Scatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 505 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 506 507 #define MPI_Scatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 508 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 509 510 #define MPI_Ialltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 511 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Ialltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 512 513 #define MPI_Ialltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 514 ((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))) 515 516 #define MPI_Iallgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 517 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 518 519 #define MPI_Iallgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm,request) \ 520 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm),(request))) 521 522 #define MPI_Igather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 523 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 524 525 #define MPI_Igatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm,request) \ 526 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm),(request))) 527 528 #define MPI_Iscatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 529 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 530 531 #define MPI_Iscatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 532 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 533 534 /* We treat MPI_Ineighbor_alltoallv as a set of isend/irecv instead of a traditional MPI collective. 535 OpenMPI-3.0 ran into error with outdegree = indegree = 0, so we use ((outdegree) || (indegree)) as a workaround. 536 */ 537 #define MPI_Start_ineighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 538 ((petsc_isend_ct += (PetscLogDouble)(outdegree),0) || (petsc_irecv_ct += (PetscLogDouble)(indegree),0) || PetscMPITypeSizeCount((outdegree),(sendcnts),(sendtype),(&petsc_isend_len)) || PetscMPITypeSizeCount((indegree),(recvcnts),(recvtype),(&petsc_irecv_len)) || (((outdegree) || (indegree)) && MPI_Ineighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm),(request)))) 539 540 #define MPI_Start_neighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 541 ((petsc_isend_ct += (PetscLogDouble)(outdegree),0) || (petsc_irecv_ct += (PetscLogDouble)(indegree),0) || PetscMPITypeSizeCount((outdegree),(sendcnts),(sendtype),(&petsc_isend_len)) || PetscMPITypeSizeCount((indegree),(recvcnts),(recvtype),(&petsc_irecv_len)) || (((outdegree) || (indegree)) && MPI_Neighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm)))) 542 543 #else 544 545 #define MPI_Startall_irecv(count,datatype,number,requests) \ 546 ((number) && MPI_Startall((number),(requests))) 547 548 #define MPI_Startall_isend(count,datatype,number,requests) \ 549 ((number) && MPI_Startall((number),(requests))) 550 551 #define MPI_Start_isend(count,datatype,requests) \ 552 (MPI_Start((requests))) 553 554 #define MPI_Start_ineighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 555 (((outdegree) || (indegree)) && MPI_Ineighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm),(request))) 556 557 #define MPI_Start_neighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 558 (((outdegree) || (indegree)) && MPI_Neighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 559 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 560 561 #else /* ---Logging is turned off --------------------------------------------*/ 562 563 #define PetscLogMemory PETSC_FALSE 564 565 #define PetscLogFlops(n) 0 566 #define PetscGetFlops(a) (*(a) = 0.0,0) 567 568 #define PetscLogStageRegister(a,b) 0 569 #define PetscLogStagePush(a) 0 570 #define PetscLogStagePop() 0 571 #define PetscLogStageSetActive(a,b) 0 572 #define PetscLogStageGetActive(a,b) 0 573 #define PetscLogStageGetVisible(a,b) 0 574 #define PetscLogStageSetVisible(a,b) 0 575 #define PetscLogStageGetId(a,b) (*(b)=0,0) 576 577 #define PetscLogEventRegister(a,b,c) 0 578 #define PetscLogEventSetCollective(a,b) 0 579 #define PetscLogEventIncludeClass(a) 0 580 #define PetscLogEventExcludeClass(a) 0 581 #define PetscLogEventActivate(a) 0 582 #define PetscLogEventDeactivate(a) 0 583 #define PetscLogEventActivateClass(a) 0 584 #define PetscLogEventDeactivateClass(a) 0 585 #define PetscLogEventSetActiveAll(a,b) 0 586 #define PetscLogEventGetId(a,b) (*(b)=0,0) 587 #define PetscLogEventGetPerfInfo(a,b,c) 0 588 #define PetscLogEventSetDof(a,b,c) 0 589 #define PetscLogEventSetError(a,b,c) 0 590 591 #define PetscLogPLB 0 592 #define PetscLogPLE 0 593 #define PetscLogPHC 0 594 #define PetscLogPHD 0 595 596 #define PetscLogObjectParents(p,n,c) 0 597 #define PetscLogObjectCreate(h) 0 598 #define PetscLogObjectDestroy(h) 0 599 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject,const char[],...); 600 601 #define PetscLogDefaultBegin() 0 602 #define PetscLogAllBegin() 0 603 #define PetscLogNestedBegin() 0 604 #define PetscLogTraceBegin(file) 0 605 #define PetscLogActions(a) 0 606 #define PetscLogObjects(a) 0 607 #define PetscLogSetThreshold(a,b) 0 608 #define PetscLogSet(lb,le) 0 609 610 #define PetscLogView(viewer) 0 611 #define PetscLogViewFromOptions() 0 612 #define PetscLogDump(c) 0 613 614 #define PetscLogEventSync(e,comm) 0 615 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0 616 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0 617 618 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 619 #define MPI_Startall_irecv(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 620 #define MPI_Startall_isend(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 621 #define MPI_Start_isend(count,datatype,requests) MPI_Start(requests) 622 #define MPI_Start_ineighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 623 (((outdegree) || (indegree)) && MPI_Ineighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm),(request))) 624 #define MPI_Start_neighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 625 (((outdegree) || (indegree)) && MPI_Neighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 626 627 #endif /* PETSC_USE_LOG */ 628 629 #define PetscPreLoadBegin(flag,name) \ 630 do {\ 631 PetscBool PetscPreLoading = flag;\ 632 int PetscPreLoadMax,PetscPreLoadIt;\ 633 PetscLogStage _stageNum;\ 634 PetscErrorCode _3_ierr; \ 635 _3_ierr = PetscOptionsGetBool(NULL,NULL,"-preload",&PetscPreLoading,NULL);CHKERRQ(_3_ierr); \ 636 PetscPreLoadMax = (int)(PetscPreLoading);\ 637 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 638 for (PetscPreLoadIt=0; PetscPreLoadIt<=PetscPreLoadMax; PetscPreLoadIt++) {\ 639 PetscPreLoadingOn = PetscPreLoading;\ 640 _3_ierr = PetscBarrier(NULL);CHKERRQ(_3_ierr);\ 641 if (PetscPreLoadIt>0) {\ 642 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 643 } else {\ 644 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 645 }\ 646 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt));\ 647 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 648 649 #define PetscPreLoadEnd() \ 650 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 651 PetscPreLoading = PETSC_FALSE;\ 652 }\ 653 } while (0) 654 655 #define PetscPreLoadStage(name) do { \ 656 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr); \ 657 if (PetscPreLoadIt>0) { \ 658 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr); \ 659 } else { \ 660 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 661 } \ 662 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt)); \ 663 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); \ 664 } while (0) 665 666 /* some vars for logging */ 667 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 668 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 669 670 #endif 671