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[],...); 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 #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 32 /*E 33 PetscInfoCommFlag - Describes the method by which to filter PetscInfo() by communicator size 34 35 Used as an input for PetscInfoSetFilterCommSelf() 36 37 $ PETSC_INFO_COMM_ALL - Default uninitialized value. PetscInfo() will not filter based on communicator size (i.e. will 38 print for all communicators) 39 $ PETSC_INFO_COMM_NO_SELF - PetscInfo() will NOT print for communicators with size = 1 (i.e. *_COMM_SELF) 40 $ PETSC_INFO_COMM_ONLY_SELF - PetscInfo will ONLY print for communicators with size = 1 41 42 Level: intermediate 43 44 .seealso: PetscInfo(), PetscInfoSetFromOptions(), PetscInfoSetFilterCommSelf() 45 E*/ 46 typedef enum { 47 PETSC_INFO_COMM_ALL = -1, 48 PETSC_INFO_COMM_NO_SELF = 0, 49 PETSC_INFO_COMM_ONLY_SELF = 1 50 } PetscInfoCommFlag; 51 52 PETSC_EXTERN const char * const PetscInfoCommFlags[]; 53 PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId); 54 PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId); 55 PETSC_EXTERN PetscErrorCode PetscInfoEnabled(PetscClassId, PetscBool *); 56 PETSC_EXTERN PetscErrorCode PetscInfoAllow(PetscBool); 57 PETSC_EXTERN PetscErrorCode PetscInfoSetFile(const char[],const char[]); 58 PETSC_EXTERN PetscErrorCode PetscInfoGetFile(char **,FILE **); 59 PETSC_EXTERN PetscErrorCode PetscInfoSetClasses(PetscBool,PetscInt,const char *const *); 60 PETSC_EXTERN PetscErrorCode PetscInfoGetClass(const char *, PetscBool *); 61 PETSC_EXTERN PetscErrorCode PetscInfoGetInfo(PetscBool *,PetscBool *,PetscBool *,PetscBool *,PetscInfoCommFlag *); 62 PETSC_EXTERN PetscErrorCode PetscInfoProcessClass(const char[],PetscInt,PetscClassId[]); 63 PETSC_EXTERN PetscErrorCode PetscInfoSetFilterCommSelf(PetscInfoCommFlag); 64 PETSC_EXTERN PetscErrorCode PetscInfoSetFromOptions(PetscOptions); 65 PETSC_EXTERN PetscErrorCode PetscInfoDestroy(void); 66 PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */ 67 68 /*MC 69 PetscLogEvent - id used to identify PETSc or user events which timed portions (blocks of executable) 70 code. 71 72 Level: intermediate 73 74 .seealso: PetscLogEventRegister(), PetscLogEventBegin(), PetscLogEventEnd(), PetscLogStage 75 M*/ 76 typedef int PetscLogEvent; 77 78 /*MC 79 PetscLogStage - id used to identify user stages (phases, sections) of runs - for logging 80 81 Level: intermediate 82 83 .seealso: PetscLogStageRegister(), PetscLogStagePush(), PetscLogStagePop(), PetscLogEvent 84 M*/ 85 typedef int PetscLogStage; 86 87 #define PETSC_EVENT 1311311 88 PETSC_EXTERN PetscLogEvent PETSC_LARGEST_EVENT; 89 90 /* Global flop counter */ 91 PETSC_EXTERN PetscLogDouble petsc_TotalFlops; 92 PETSC_EXTERN PetscLogDouble petsc_tmp_flops; 93 94 /* Global GPU counters */ 95 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 96 PETSC_EXTERN PetscLogDouble petsc_ctog_ct; 97 PETSC_EXTERN PetscLogDouble petsc_gtoc_ct; 98 PETSC_EXTERN PetscLogDouble petsc_ctog_sz; 99 PETSC_EXTERN PetscLogDouble petsc_gtoc_sz; 100 PETSC_EXTERN PetscLogDouble petsc_gflops; 101 PETSC_EXTERN PetscLogDouble petsc_gtime; 102 #endif 103 104 /* We must make the following structures available to access the event 105 activation flags in the PetscLogEventBegin/End() macros. These are not part of the PETSc public 106 API and are not intended to be used by other parts of PETSc or by users. 107 108 The code that manipulates these structures is in src/sys/logging/utils. 109 */ 110 typedef struct _n_PetscIntStack *PetscIntStack; 111 112 /* -----------------------------------------------------------------------------------------------------*/ 113 /* 114 PetscClassRegInfo, PetscClassPerfInfo - Each class has two data structures associated with it. The first has 115 static information about it, the second collects statistics on how many objects of the class are created, 116 how much memory they use, etc. 117 118 PetscClassRegLog, PetscClassPerfLog - arrays of the PetscClassRegInfo and PetscClassPerfInfo for all classes. 119 */ 120 typedef struct { 121 char *name; /* The class name */ 122 PetscClassId classid; /* The integer identifying this class */ 123 } PetscClassRegInfo; 124 125 typedef struct { 126 PetscClassId id; /* The integer identifying this class */ 127 int creations; /* The number of objects of this class created */ 128 int destructions; /* The number of objects of this class destroyed */ 129 PetscLogDouble mem; /* The total memory allocated by objects of this class */ 130 PetscLogDouble descMem; /* The total memory allocated by descendents of these objects */ 131 } PetscClassPerfInfo; 132 133 typedef struct _n_PetscClassRegLog *PetscClassRegLog; 134 struct _n_PetscClassRegLog { 135 int numClasses; /* The number of classes registered */ 136 int maxClasses; /* The maximum number of classes */ 137 PetscClassRegInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 138 }; 139 140 typedef struct _n_PetscClassPerfLog *PetscClassPerfLog; 141 struct _n_PetscClassPerfLog { 142 int numClasses; /* The number of logging classes */ 143 int maxClasses; /* The maximum number of classes */ 144 PetscClassPerfInfo *classInfo; /* The structure for class information (classids are monotonicly increasing) */ 145 }; 146 /* -----------------------------------------------------------------------------------------------------*/ 147 /* 148 PetscEventRegInfo, PetscEventPerfInfo - Each event has two data structures associated with it. The first has 149 static information about it, the second collects statistics on how many times the event is used, how 150 much time it takes, etc. 151 152 PetscEventRegLog, PetscEventPerfLog - an array of all PetscEventRegInfo and PetscEventPerfInfo for all events. There is one 153 of these for each stage. 154 155 */ 156 typedef struct { 157 char *name; /* The name of this event */ 158 PetscClassId classid; /* The class the event is associated with */ 159 PetscBool collective; /* Flag this event as collective */ 160 #if defined (PETSC_HAVE_MPE) 161 int mpe_id_begin; /* MPE IDs that define the event */ 162 int mpe_id_end; 163 #endif 164 } PetscEventRegInfo; 165 166 typedef struct { 167 int id; /* The integer identifying this event */ 168 PetscBool active; /* The flag to activate logging */ 169 PetscBool visible; /* The flag to print info in summary */ 170 int depth; /* The nesting depth of the event call */ 171 int count; /* The number of times this event was executed */ 172 PetscLogDouble flops, flops2, flopsTmp; /* The flops and flops^2 used in this event */ 173 PetscLogDouble time, time2, timeTmp; /* The time and time^2 taken for this event */ 174 PetscLogDouble syncTime; /* The synchronization barrier time */ 175 PetscLogDouble dof[8]; /* The number of degrees of freedom associated with this event */ 176 PetscLogDouble errors[8]; /* The errors (user-defined) associated with this event */ 177 PetscLogDouble numMessages; /* The number of messages in this event */ 178 PetscLogDouble messageLength; /* The total message lengths in this event */ 179 PetscLogDouble numReductions; /* The number of reductions in this event */ 180 PetscLogDouble memIncrease; /* How much the resident memory has increased in this event */ 181 PetscLogDouble mallocIncrease;/* How much the maximum malloced space has increased in this event */ 182 PetscLogDouble mallocSpace; /* How much the space was malloced and kept during this event */ 183 PetscLogDouble mallocIncreaseEvent; /* Maximum of the high water mark with in event minus memory available at the end of the event */ 184 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 185 PetscLogDouble CpuToGpuCount; /* The total number of CPU to GPU copies */ 186 PetscLogDouble GpuToCpuCount; /* The total number of GPU to CPU copies */ 187 PetscLogDouble CpuToGpuSize; /* The total size of CPU to GPU copies */ 188 PetscLogDouble GpuToCpuSize; /* The total size of GPU to CPU copies */ 189 PetscLogDouble GpuFlops; /* The flops done on a GPU in this event */ 190 PetscLogDouble GpuTime; /* The time spent on a GPU in this event */ 191 #endif 192 } PetscEventPerfInfo; 193 194 typedef struct _n_PetscEventRegLog *PetscEventRegLog; 195 struct _n_PetscEventRegLog { 196 int numEvents; /* The number of registered events */ 197 int maxEvents; /* The maximum number of events */ 198 PetscEventRegInfo *eventInfo; /* The registration information for each event */ 199 }; 200 201 typedef struct _n_PetscEventPerfLog *PetscEventPerfLog; 202 struct _n_PetscEventPerfLog { 203 int numEvents; /* The number of logging events */ 204 int maxEvents; /* The maximum number of events */ 205 PetscEventPerfInfo *eventInfo; /* The performance information for each event */ 206 }; 207 /* ------------------------------------------------------------------------------------------------------------*/ 208 /* 209 PetscStageInfo - Contains all the information about a particular stage. 210 211 PetscStageLog - An array of PetscStageInfo for each registered stage. There is a single one of these in the code. 212 */ 213 typedef struct _PetscStageInfo { 214 char *name; /* The stage name */ 215 PetscBool used; /* The stage was pushed on this processor */ 216 PetscEventPerfInfo perfInfo; /* The stage performance information */ 217 PetscEventPerfLog eventLog; /* The event information for this stage */ 218 PetscClassPerfLog classLog; /* The class information for this stage */ 219 } PetscStageInfo; 220 221 typedef struct _n_PetscStageLog *PetscStageLog; 222 struct _n_PetscStageLog { 223 int numStages; /* The number of registered stages */ 224 int maxStages; /* The maximum number of stages */ 225 PetscIntStack stack; /* The stack for active stages */ 226 int curStage; /* The current stage (only used in macros so we don't call PetscIntStackTop) */ 227 PetscStageInfo *stageInfo; /* The information for each stage */ 228 PetscEventRegLog eventLog; /* The registered events */ 229 PetscClassRegLog classLog; /* The registered classes */ 230 }; 231 /* -----------------------------------------------------------------------------------------------------*/ 232 233 PETSC_EXTERN PetscErrorCode PetscLogObjectParent(PetscObject,PetscObject); 234 PETSC_EXTERN PetscErrorCode PetscLogObjectMemory(PetscObject,PetscLogDouble); 235 236 #if defined(PETSC_USE_LOG) /* --- Logging is turned on --------------------------------*/ 237 PETSC_EXTERN PetscStageLog petsc_stageLog; 238 PETSC_EXTERN PetscErrorCode PetscLogGetStageLog(PetscStageLog*); 239 PETSC_EXTERN PetscErrorCode PetscStageLogGetCurrent(PetscStageLog,int*); 240 PETSC_EXTERN PetscErrorCode PetscStageLogGetEventPerfLog(PetscStageLog,int,PetscEventPerfLog*); 241 242 /* 243 Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately. 244 245 For the complex numbers version, note that 246 1 complex addition = 2 flops 247 1 complex multiplication = 6 flops, 248 where we define 1 flop as that for a double precision scalar. We roughly approximate 249 flop counting for complex numbers by multiplying the total flops by 4; this corresponds 250 to the assumption that we're counting mostly additions and multiplications -- and 251 roughly the same number of each. More accurate counting could be done by distinguishing 252 among the various arithmetic operations. 253 */ 254 255 #if defined(PETSC_USE_COMPLEX) 256 #define PETSC_FLOPS_PER_OP 4.0 257 #else 258 #define PETSC_FLOPS_PER_OP 1.0 259 #endif 260 261 /*@ 262 PetscLogFlops - Log how many flops are performed in a calculation 263 264 Input Paramters: 265 flops - the number of flops 266 267 Notes: 268 To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double, 269 not an integer. Use PetscLogFlops(4.0*n) not PetscLogFlops(4*n) 270 271 Level: intermediate 272 273 .seealso: PetscLogView() 274 @*/ 275 276 PETSC_STATIC_INLINE PetscErrorCode PetscLogFlops(PetscLogDouble n) 277 { 278 PetscFunctionBegin; 279 #if defined(PETSC_USE_DEBUG) 280 if (n < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 281 #endif 282 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 283 PetscFunctionReturn(0); 284 } 285 286 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 287 PETSC_STATIC_INLINE PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size) 288 { 289 PetscFunctionBegin; 290 petsc_ctog_ct += 1; 291 petsc_ctog_sz += size; 292 PetscFunctionReturn(0); 293 } 294 295 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size) 296 { 297 PetscFunctionBegin; 298 petsc_gtoc_ct += 1; 299 petsc_gtoc_sz += size; 300 PetscFunctionReturn(0); 301 } 302 303 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuFlops(PetscLogDouble n) 304 { 305 PetscFunctionBegin; 306 #if defined(PETSC_USE_DEBUG) 307 if (n < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Cannot log negative flops"); 308 #endif 309 petsc_TotalFlops += PETSC_FLOPS_PER_OP*n; 310 petsc_gflops += PETSC_FLOPS_PER_OP*n; 311 PetscFunctionReturn(0); 312 } 313 314 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeBegin() 315 { 316 PetscErrorCode ierr; 317 PetscFunctionBegin; 318 ierr = PetscTimeSubtract(&petsc_gtime);CHKERRQ(ierr); 319 PetscFunctionReturn(0); 320 } 321 322 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeEnd() 323 { 324 PetscErrorCode ierr; 325 PetscFunctionBegin; 326 ierr = PetscTimeAdd(&petsc_gtime);CHKERRQ(ierr); 327 PetscFunctionReturn(0); 328 } 329 330 PETSC_STATIC_INLINE PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t) 331 { 332 PetscFunctionBegin; 333 petsc_gtime += t; 334 PetscFunctionReturn(0); 335 } 336 #endif 337 338 PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *); 339 340 #if defined (PETSC_HAVE_MPE) 341 PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void); 342 PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]); 343 #endif 344 345 PETSC_EXTERN PetscErrorCode (*PetscLogPLB)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 346 PETSC_EXTERN PetscErrorCode (*PetscLogPLE)(PetscLogEvent,int,PetscObject,PetscObject,PetscObject,PetscObject); 347 PETSC_EXTERN PetscErrorCode (*PetscLogPHC)(PetscObject); 348 PETSC_EXTERN PetscErrorCode (*PetscLogPHD)(PetscObject); 349 350 #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) 351 #define PetscLogObjectCreate(h) ((PetscLogPHC) ? (*PetscLogPHC)((PetscObject)(h)) : 0) 352 #define PetscLogObjectDestroy(h) ((PetscLogPHD) ? (*PetscLogPHD)((PetscObject)(h)) : 0) 353 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...); 354 355 /* Initialization functions */ 356 PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void); 357 PETSC_EXTERN PetscErrorCode PetscLogAllBegin(void); 358 PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void); 359 PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *); 360 PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool); 361 PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool); 362 PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble,PetscLogDouble*); 363 PETSC_EXTERN PetscErrorCode PetscLogSet(PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject), 364 PetscErrorCode (*)(int, int, PetscObject, PetscObject, PetscObject, PetscObject)); 365 366 /* Output functions */ 367 PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer); 368 PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void); 369 PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]); 370 371 /* Stage functions */ 372 PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[],PetscLogStage*); 373 PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage); 374 PETSC_EXTERN PetscErrorCode PetscLogStagePop(void); 375 PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage,PetscBool); 376 PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage,PetscBool*); 377 PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage,PetscBool); 378 PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage,PetscBool*); 379 PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[],PetscLogStage*); 380 381 /* Event functions */ 382 PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[],PetscClassId,PetscLogEvent*); 383 PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent,PetscBool); 384 PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId); 385 PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId); 386 PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent); 387 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent); 388 PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent,PetscBool); 389 PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId); 390 PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId); 391 PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[],PetscLogEvent*); 392 PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(int,PetscLogEvent,PetscEventPerfInfo*); 393 PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble); 394 PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble); 395 396 /* Global counters */ 397 PETSC_EXTERN PetscLogDouble petsc_irecv_ct; 398 PETSC_EXTERN PetscLogDouble petsc_isend_ct; 399 PETSC_EXTERN PetscLogDouble petsc_recv_ct; 400 PETSC_EXTERN PetscLogDouble petsc_send_ct; 401 PETSC_EXTERN PetscLogDouble petsc_irecv_len; 402 PETSC_EXTERN PetscLogDouble petsc_isend_len; 403 PETSC_EXTERN PetscLogDouble petsc_recv_len; 404 PETSC_EXTERN PetscLogDouble petsc_send_len; 405 PETSC_EXTERN PetscLogDouble petsc_allreduce_ct; 406 PETSC_EXTERN PetscLogDouble petsc_gather_ct; 407 PETSC_EXTERN PetscLogDouble petsc_scatter_ct; 408 PETSC_EXTERN PetscLogDouble petsc_wait_ct; 409 PETSC_EXTERN PetscLogDouble petsc_wait_any_ct; 410 PETSC_EXTERN PetscLogDouble petsc_wait_all_ct; 411 PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct; 412 413 PETSC_EXTERN PetscBool PetscLogMemory; 414 415 PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */ 416 PETSC_EXTERN PetscErrorCode PetscLogEventSynchronize(PetscLogEvent, MPI_Comm); 417 418 #define PetscLogEventSync(e,comm) \ 419 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 420 PetscLogEventSynchronize((e),(comm)) : 0 )) 421 422 #define PetscLogEventBegin(e,o1,o2,o3,o4) \ 423 (((PetscLogPLB && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 424 (*PetscLogPLB)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0 )) 425 426 #define PetscLogEventEnd(e,o1,o2,o3,o4) \ 427 (((PetscLogPLE && petsc_stageLog->stageInfo[petsc_stageLog->curStage].perfInfo.active && petsc_stageLog->stageInfo[petsc_stageLog->curStage].eventLog->eventInfo[e].active) ? \ 428 (*PetscLogPLE)((e),0,(PetscObject)(o1),(PetscObject)(o2),(PetscObject)(o3),(PetscObject)(o4)) : 0 )) 429 430 PETSC_EXTERN PetscErrorCode PetscLogEventGetFlops(PetscLogEvent,PetscLogDouble*); 431 PETSC_EXTERN PetscErrorCode PetscLogEventZeroFlops(PetscLogEvent); 432 433 /* 434 These are used internally in the PETSc routines to keep a count of MPI messages and 435 their sizes. 436 437 This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file 438 uses macros to defined the MPI operations. 439 440 It does not work correctly from HP-UX because it processes the 441 macros in a way that sometimes it double counts, hence 442 PETSC_HAVE_BROKEN_RECURSIVE_MACRO 443 444 It does not work with Windows because winmpich lacks MPI_Type_size() 445 */ 446 #if !defined(MPIUNI_H) && !defined(PETSC_HAVE_BROKEN_RECURSIVE_MACRO) && !defined (PETSC_HAVE_MPI_MISSING_TYPESIZE) 447 /* 448 Logging of MPI activities 449 */ 450 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSize(PetscInt count,MPI_Datatype type,PetscLogDouble *length) 451 { 452 PetscMPIInt typesize; 453 PetscErrorCode ierr; 454 if (type == MPI_DATATYPE_NULL) return 0; 455 ierr = MPI_Type_size(type,&typesize);CHKERRQ(ierr); 456 *length += (PetscLogDouble) (count*typesize); 457 return 0; 458 } 459 460 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 461 { 462 PetscMPIInt typesize,size,p; 463 PetscErrorCode ierr; 464 465 if (type == MPI_DATATYPE_NULL) return 0; 466 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 467 ierr = MPI_Type_size(type,&typesize);CHKERRQ(ierr); 468 for (p=0; p<size; ++p) { 469 *length += (PetscLogDouble) (counts[p]*typesize); 470 } 471 return 0; 472 } 473 474 PETSC_STATIC_INLINE PetscErrorCode PetscMPITypeSizeCount(PetscInt n,const PetscMPIInt *counts,MPI_Datatype type,PetscLogDouble *length) 475 { 476 PetscMPIInt typesize,p; 477 PetscErrorCode ierr; 478 479 if (type == MPI_DATATYPE_NULL) return 0; 480 ierr = MPI_Type_size(type,&typesize);CHKERRQ(ierr); 481 for (p=0; p<n; ++p) { 482 *length += (PetscLogDouble) (counts[p]*typesize); 483 } 484 return 0; 485 } 486 487 /* 488 Returns 1 if the communicator is parallel else zero 489 */ 490 PETSC_STATIC_INLINE int PetscMPIParallelComm(MPI_Comm comm) 491 { 492 PetscMPIInt size; MPI_Comm_size(comm,&size); return size > 1; 493 } 494 495 #define MPI_Irecv(buf,count,datatype,source,tag,comm,request) \ 496 ((petsc_irecv_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || MPI_Irecv((buf),(count),(datatype),(source),(tag),(comm),(request))) 497 498 #define MPI_Isend(buf,count,datatype,dest,tag,comm,request) \ 499 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || MPI_Isend((buf),(count),(datatype),(dest),(tag),(comm),(request))) 500 501 #define MPI_Startall_irecv(count,datatype,number,requests) \ 502 ((petsc_irecv_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_irecv_len)) || ((number) && MPI_Startall((number),(requests)))) 503 504 #define MPI_Startall_isend(count,datatype,number,requests) \ 505 ((petsc_isend_ct += (PetscLogDouble)(number),0) || PetscMPITypeSize((count),(datatype),&(petsc_isend_len)) || ((number) && MPI_Startall((number),(requests)))) 506 507 #define MPI_Start_isend(count,datatype,requests) \ 508 ((petsc_isend_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_isend_len)) || MPI_Start((requests))) 509 510 #define MPI_Recv(buf,count,datatype,source,tag,comm,status) \ 511 ((petsc_recv_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_recv_len)) || MPI_Recv((buf),(count),(datatype),(source),(tag),(comm),(status))) 512 513 #define MPI_Send(buf,count,datatype,dest,tag,comm) \ 514 ((petsc_send_ct++,0) || PetscMPITypeSize((count),(datatype),(&petsc_send_len)) || MPI_Send((buf),(count),(datatype),(dest),(tag),(comm))) 515 516 #define MPI_Wait(request,status) \ 517 ((petsc_wait_ct++,petsc_sum_of_waits_ct++,0) || MPI_Wait((request),(status))) 518 519 #define MPI_Waitany(a,b,c,d) \ 520 ((petsc_wait_any_ct++,petsc_sum_of_waits_ct++,0) || MPI_Waitany((a),(b),(c),(d))) 521 522 #define MPI_Waitall(count,array_of_requests,array_of_statuses) \ 523 ((petsc_wait_all_ct++,petsc_sum_of_waits_ct += (PetscLogDouble) (count),0) || MPI_Waitall((count),(array_of_requests),(array_of_statuses))) 524 525 #define MPI_Allreduce(sendbuf,recvbuf,count,datatype,op,comm) \ 526 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Allreduce((sendbuf),(recvbuf),(count),(datatype),(op),(comm))) 527 528 #define MPI_Bcast(buffer,count,datatype,root,comm) \ 529 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Bcast((buffer),(count),(datatype),(root),(comm))) 530 531 #define MPI_Reduce_scatter_block(sendbuf,recvbuf,recvcount,datatype,op,comm) \ 532 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || MPI_Reduce_scatter_block((sendbuf),(recvbuf),(recvcount),(datatype),(op),(comm))) 533 534 #define MPI_Alltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 535 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Alltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 536 537 #define MPI_Alltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 538 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSizeComm((comm),(sendcnts),(sendtype),(&petsc_send_len)) || MPI_Alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 539 540 #define MPI_Allgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm) \ 541 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm))) 542 543 #define MPI_Allgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm) \ 544 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Allgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm))) 545 546 #define MPI_Gather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 547 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 548 549 #define MPI_Gatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm) \ 550 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Gatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm))) 551 552 #define MPI_Scatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 553 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 554 555 #define MPI_Scatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm) \ 556 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Scatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm))) 557 558 #define MPI_Ialltoall(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 559 ((petsc_allreduce_ct += PetscMPIParallelComm((comm)),0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Ialltoall((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 560 561 #define MPI_Ialltoallv(sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 562 ((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))) 563 564 #define MPI_Iallgather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,comm,request) \ 565 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(comm),(request))) 566 567 #define MPI_Iallgatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,comm,request) \ 568 ((petsc_gather_ct += PetscMPIParallelComm((comm)),0) || MPI_Iallgatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(comm),(request))) 569 570 #define MPI_Igather(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 571 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igather((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 572 573 #define MPI_Igatherv(sendbuf,sendcount,sendtype,recvbuf,recvcount,displs,recvtype,root,comm,request) \ 574 ((petsc_gather_ct++,0) || PetscMPITypeSize((sendcount),(sendtype),(&petsc_send_len)) || MPI_Igatherv((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(displs),(recvtype),(root),(comm),(request))) 575 576 #define MPI_Iscatter(sendbuf,sendcount,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 577 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatter((sendbuf),(sendcount),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 578 579 #define MPI_Iscatterv(sendbuf,sendcount,displs,sendtype,recvbuf,recvcount,recvtype,root,comm,request) \ 580 ((petsc_scatter_ct++,0) || PetscMPITypeSize((recvcount),(recvtype),(&petsc_recv_len)) || MPI_Iscatterv((sendbuf),(sendcount),(displs),(sendtype),(recvbuf),(recvcount),(recvtype),(root),(comm),(request))) 581 582 /* We treat MPI_Ineighbor_alltoallv as a set of isend/irecv instead of a traditional MPI collective. 583 OpenMPI-3.0 ran into error with outdegree = indegree = 0, so we use ((outdegree) || (indegree)) as a workaround. 584 */ 585 #define MPI_Start_ineighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 586 ((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)))) 587 588 #define MPI_Start_neighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 589 ((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)))) 590 591 #else 592 593 #define MPI_Startall_irecv(count,datatype,number,requests) \ 594 ((number) && MPI_Startall((number),(requests))) 595 596 #define MPI_Startall_isend(count,datatype,number,requests) \ 597 ((number) && MPI_Startall((number),(requests))) 598 599 #define MPI_Start_isend(count,datatype,requests) \ 600 (MPI_Start((requests))) 601 602 #define MPI_Start_ineighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 603 (((outdegree) || (indegree)) && MPI_Ineighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm),(request))) 604 605 #define MPI_Start_neighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 606 (((outdegree) || (indegree)) && MPI_Neighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 607 #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */ 608 609 #else /* ---Logging is turned off --------------------------------------------*/ 610 611 #define PetscLogMemory PETSC_FALSE 612 613 #define PetscLogFlops(n) 0 614 #define PetscGetFlops(a) (*(a) = 0.0,0) 615 616 #define PetscLogStageRegister(a,b) 0 617 #define PetscLogStagePush(a) 0 618 #define PetscLogStagePop() 0 619 #define PetscLogStageSetActive(a,b) 0 620 #define PetscLogStageGetActive(a,b) 0 621 #define PetscLogStageGetVisible(a,b) 0 622 #define PetscLogStageSetVisible(a,b) 0 623 #define PetscLogStageGetId(a,b) (*(b)=0,0) 624 625 #define PetscLogEventRegister(a,b,c) 0 626 #define PetscLogEventSetCollective(a,b) 0 627 #define PetscLogEventIncludeClass(a) 0 628 #define PetscLogEventExcludeClass(a) 0 629 #define PetscLogEventActivate(a) 0 630 #define PetscLogEventDeactivate(a) 0 631 #define PetscLogEventActivateClass(a) 0 632 #define PetscLogEventDeactivateClass(a) 0 633 #define PetscLogEventSetActiveAll(a,b) 0 634 #define PetscLogEventGetId(a,b) (*(b)=0,0) 635 #define PetscLogEventGetPerfInfo(a,b,c) 0 636 #define PetscLogEventSetDof(a,b,c) 0 637 #define PetscLogEventSetError(a,b,c) 0 638 639 #define PetscLogPLB 0 640 #define PetscLogPLE 0 641 #define PetscLogPHC 0 642 #define PetscLogPHD 0 643 644 #define PetscLogObjectParents(p,n,c) 0 645 #define PetscLogObjectCreate(h) 0 646 #define PetscLogObjectDestroy(h) 0 647 PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject,const char[],...); 648 649 #define PetscLogDefaultBegin() 0 650 #define PetscLogAllBegin() 0 651 #define PetscLogNestedBegin() 0 652 #define PetscLogTraceBegin(file) 0 653 #define PetscLogActions(a) 0 654 #define PetscLogObjects(a) 0 655 #define PetscLogSetThreshold(a,b) 0 656 #define PetscLogSet(lb,le) 0 657 658 #define PetscLogView(viewer) 0 659 #define PetscLogViewFromOptions() 0 660 #define PetscLogDump(c) 0 661 662 #define PetscLogEventSync(e,comm) 0 663 #define PetscLogEventBegin(e,o1,o2,o3,o4) 0 664 #define PetscLogEventEnd(e,o1,o2,o3,o4) 0 665 666 #define PetscLogCpuToGpu(a) 0 667 #define PetscLogGpuToCpu(a) 0 668 #define PetscLogGpuFlops(a) 0 669 #define PetscLogGpuTimeBegin() 0 670 #define PetscLogGpuTimeEnd() 0 671 #define PetscLogGpuTimeAdd(a) 0 672 673 /* If PETSC_USE_LOG is NOT defined, these still need to be! */ 674 #define MPI_Startall_irecv(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 675 #define MPI_Startall_isend(count,datatype,number,requests) ((number) && MPI_Startall(number,requests)) 676 #define MPI_Start_isend(count,datatype,requests) MPI_Start(requests) 677 #define MPI_Start_ineighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm,request) \ 678 (((outdegree) || (indegree)) && MPI_Ineighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm),(request))) 679 #define MPI_Start_neighbor_alltoallv(outdegree,indegree,sendbuf,sendcnts,sdispls,sendtype,recvbuf,recvcnts,rdispls,recvtype,comm) \ 680 (((outdegree) || (indegree)) && MPI_Neighbor_alltoallv((sendbuf),(sendcnts),(sdispls),(sendtype),(recvbuf),(recvcnts),(rdispls),(recvtype),(comm))) 681 682 #endif /* PETSC_USE_LOG */ 683 684 #define PetscPreLoadBegin(flag,name) \ 685 do {\ 686 PetscBool PetscPreLoading = flag;\ 687 int PetscPreLoadMax,PetscPreLoadIt;\ 688 PetscLogStage _stageNum;\ 689 PetscErrorCode _3_ierr; \ 690 _3_ierr = PetscOptionsGetBool(NULL,NULL,"-preload",&PetscPreLoading,NULL);CHKERRQ(_3_ierr); \ 691 PetscPreLoadMax = (int)(PetscPreLoading);\ 692 PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed;\ 693 for (PetscPreLoadIt=0; PetscPreLoadIt<=PetscPreLoadMax; PetscPreLoadIt++) {\ 694 PetscPreLoadingOn = PetscPreLoading;\ 695 _3_ierr = PetscBarrier(NULL);CHKERRQ(_3_ierr);\ 696 if (PetscPreLoadIt>0) {\ 697 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr);\ 698 } else {\ 699 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 700 }\ 701 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt));\ 702 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); 703 704 #define PetscPreLoadEnd() \ 705 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr);\ 706 PetscPreLoading = PETSC_FALSE;\ 707 }\ 708 } while (0) 709 710 #define PetscPreLoadStage(name) do { \ 711 _3_ierr = PetscLogStagePop();CHKERRQ(_3_ierr); \ 712 if (PetscPreLoadIt>0) { \ 713 _3_ierr = PetscLogStageGetId(name,&_stageNum);CHKERRQ(_3_ierr); \ 714 } else { \ 715 _3_ierr = PetscLogStageRegister(name,&_stageNum);CHKERRQ(_3_ierr); \ 716 } \ 717 _3_ierr = PetscLogStageSetActive(_stageNum,(PetscBool)(!PetscPreLoadMax || PetscPreLoadIt)); \ 718 _3_ierr = PetscLogStagePush(_stageNum);CHKERRQ(_3_ierr); \ 719 } while (0) 720 721 /* some vars for logging */ 722 PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */ 723 PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */ 724 725 #endif 726