Actual source code: petsclog.h
1: /*
2: Defines profile/logging in PETSc.
3: */
4: #pragma once
6: #include <petscsys.h>
7: #include <petsctime.h>
8: #include <petscbt.h>
9: #include <petsclogtypes.h>
11: /* MANSEC = Sys */
12: /* SUBMANSEC = Log */
14: /* General logging of information; different from event logging */
15: PETSC_EXTERN PetscErrorCode PetscInfo_Private(const char[], PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(3, 4);
16: #if PetscDefined(USE_INFO)
17: #define PetscInfo(A, ...) PetscInfo_Private(PETSC_FUNCTION_NAME, ((PetscObject)(A)), __VA_ARGS__)
18: #else
19: #define PetscInfo(A, ...) PETSC_SUCCESS
20: #endif
22: #define PetscInfo1(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
23: #define PetscInfo2(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
24: #define PetscInfo3(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
25: #define PetscInfo4(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
26: #define PetscInfo5(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
27: #define PetscInfo6(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
28: #define PetscInfo7(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
29: #define PetscInfo8(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
30: #define PetscInfo9(...) PETSC_DEPRECATED_MACRO(3, 17, 0, "PetscInfo()", ) PetscInfo(__VA_ARGS__)
32: /*E
33: PetscInfoCommFlag - Describes the method by which to filter information displayed by `PetscInfo()` by communicator size
35: Values:
36: + `PETSC_INFO_COMM_ALL` - Default uninitialized value. `PetscInfo()` will not filter based on
37: communicator size (i.e. will print for all communicators)
38: . `PETSC_INFO_COMM_NO_SELF` - `PetscInfo()` will NOT print for communicators with size = 1 (i.e. *_COMM_SELF)
39: - `PETSC_INFO_COMM_ONLY_SELF` - `PetscInfo()` will ONLY print for communicators with size = 1
41: Level: intermediate
43: Note:
44: Used as an input for `PetscInfoSetFilterCommSelf()`
46: .seealso: `PetscInfo()`, `PetscInfoSetFromOptions()`, `PetscInfoSetFilterCommSelf()`
47: E*/
48: typedef enum {
49: PETSC_INFO_COMM_ALL = -1,
50: PETSC_INFO_COMM_NO_SELF = 0,
51: PETSC_INFO_COMM_ONLY_SELF = 1
52: } PetscInfoCommFlag;
54: PETSC_EXTERN const char *const PetscInfoCommFlags[];
55: PETSC_EXTERN PetscErrorCode PetscInfoDeactivateClass(PetscClassId);
56: PETSC_EXTERN PetscErrorCode PetscInfoActivateClass(PetscClassId);
57: PETSC_EXTERN PetscErrorCode PetscInfoEnabled(PetscClassId, PetscBool *);
58: PETSC_EXTERN PetscErrorCode PetscInfoAllow(PetscBool);
59: PETSC_EXTERN PetscErrorCode PetscInfoSetFile(const char[], const char[]);
60: PETSC_EXTERN PetscErrorCode PetscInfoGetFile(char **, FILE **);
61: PETSC_EXTERN PetscErrorCode PetscInfoSetClasses(PetscBool, PetscInt, const char *const *);
62: PETSC_EXTERN PetscErrorCode PetscInfoGetClass(const char *, PetscBool *);
63: PETSC_EXTERN PetscErrorCode PetscInfoGetInfo(PetscBool *, PetscBool *, PetscBool *, PetscBool *, PetscInfoCommFlag *);
64: PETSC_EXTERN PetscErrorCode PetscInfoProcessClass(const char[], PetscInt, const PetscClassId[]);
65: PETSC_EXTERN PetscErrorCode PetscInfoSetFilterCommSelf(PetscInfoCommFlag);
66: PETSC_EXTERN PetscErrorCode PetscInfoSetFromOptions(PetscOptions);
67: PETSC_EXTERN PetscErrorCode PetscInfoDestroy(void);
68: PETSC_EXTERN PetscBool PetscLogPrintInfo; /* if true, indicates PetscInfo() is turned on */
70: PETSC_EXTERN PetscErrorCode PetscIntStackCreate(PetscIntStack *);
71: PETSC_EXTERN PetscErrorCode PetscIntStackDestroy(PetscIntStack);
72: PETSC_EXTERN PetscErrorCode PetscIntStackPush(PetscIntStack, int);
73: PETSC_EXTERN PetscErrorCode PetscIntStackPop(PetscIntStack, int *);
74: PETSC_EXTERN PetscErrorCode PetscIntStackTop(PetscIntStack, int *);
75: PETSC_EXTERN PetscErrorCode PetscIntStackEmpty(PetscIntStack, PetscBool *);
77: PETSC_EXTERN PetscErrorCode PetscLogStateCreate(PetscLogState *);
78: PETSC_EXTERN PetscErrorCode PetscLogStateDestroy(PetscLogState *);
80: PETSC_EXTERN PetscErrorCode PetscLogStateClassRegister(PetscLogState, const char[], PetscClassId, PetscLogStage *);
81: PETSC_EXTERN PetscErrorCode PetscLogStateClassSetActive(PetscLogState, PetscLogStage, PetscClassId, PetscBool);
82: PETSC_EXTERN PetscErrorCode PetscLogStateClassSetActiveAll(PetscLogState, PetscClassId, PetscBool);
84: PETSC_EXTERN PetscErrorCode PetscLogStateStageRegister(PetscLogState, const char[], PetscLogStage *);
85: PETSC_EXTERN PetscErrorCode PetscLogStateStagePush(PetscLogState, PetscLogStage);
86: PETSC_EXTERN PetscErrorCode PetscLogStateStagePop(PetscLogState);
87: PETSC_EXTERN PetscErrorCode PetscLogStateStageSetActive(PetscLogState, PetscLogStage, PetscBool);
88: PETSC_EXTERN PetscErrorCode PetscLogStateStageGetActive(PetscLogState, PetscLogStage, PetscBool *);
89: PETSC_EXTERN PetscErrorCode PetscLogStateGetCurrentStage(PetscLogState, PetscLogStage *);
91: PETSC_EXTERN PetscErrorCode PetscLogStateEventRegister(PetscLogState, const char[], PetscClassId, PetscLogEvent *);
92: PETSC_EXTERN PetscErrorCode PetscLogStateEventSetCollective(PetscLogState, PetscLogEvent, PetscBool);
93: PETSC_EXTERN PetscErrorCode PetscLogStateEventSetActive(PetscLogState, PetscLogStage, PetscLogEvent, PetscBool);
94: PETSC_EXTERN PetscErrorCode PetscLogStateEventSetActiveAll(PetscLogState, PetscLogEvent, PetscBool);
95: PETSC_EXTERN PetscErrorCode PetscLogStateEventGetActive(PetscLogState, PetscLogStage, PetscLogEvent, PetscBool *);
97: PETSC_EXTERN PetscErrorCode PetscLogStateGetEventFromName(PetscLogState, const char[], PetscLogEvent *);
98: PETSC_EXTERN PetscErrorCode PetscLogStateGetStageFromName(PetscLogState, const char[], PetscLogStage *);
99: PETSC_EXTERN PetscErrorCode PetscLogStateGetClassFromName(PetscLogState, const char[], PetscLogClass *);
100: PETSC_EXTERN PetscErrorCode PetscLogStateGetClassFromClassId(PetscLogState, PetscClassId, PetscLogClass *);
101: PETSC_EXTERN PetscErrorCode PetscLogStateGetNumEvents(PetscLogState, PetscInt *);
102: PETSC_EXTERN PetscErrorCode PetscLogStateGetNumStages(PetscLogState, PetscInt *);
103: PETSC_EXTERN PetscErrorCode PetscLogStateGetNumClasses(PetscLogState, PetscInt *);
104: PETSC_EXTERN PetscErrorCode PetscLogStateEventGetInfo(PetscLogState, PetscLogEvent, PetscLogEventInfo *);
105: PETSC_EXTERN PetscErrorCode PetscLogStateStageGetInfo(PetscLogState, PetscLogStage, PetscLogStageInfo *);
106: PETSC_EXTERN PetscErrorCode PetscLogStateClassGetInfo(PetscLogState, PetscLogClass, PetscLogClassInfo *);
108: PETSC_EXTERN PetscClassId PETSCLOGHANDLER_CLASSID;
110: PETSC_EXTERN PetscFunctionList PetscLogHandlerList;
112: PETSC_EXTERN PetscErrorCode PetscLogHandlerRegister(const char[], PetscErrorCode (*)(PetscLogHandler));
113: PETSC_EXTERN PetscErrorCode PetscLogHandlerCreate(MPI_Comm, PetscLogHandler *);
114: PETSC_EXTERN PetscErrorCode PetscLogHandlerSetType(PetscLogHandler, PetscLogHandlerType);
115: PETSC_EXTERN PetscErrorCode PetscLogHandlerGetType(PetscLogHandler, PetscLogHandlerType *);
116: PETSC_EXTERN PetscErrorCode PetscLogHandlerDestroy(PetscLogHandler *);
117: PETSC_EXTERN PetscErrorCode PetscLogHandlerSetState(PetscLogHandler, PetscLogState);
118: PETSC_EXTERN PetscErrorCode PetscLogHandlerGetState(PetscLogHandler, PetscLogState *);
119: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventBegin(PetscLogHandler, PetscLogEvent, PetscObject, PetscObject, PetscObject, PetscObject);
120: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventEnd(PetscLogHandler, PetscLogEvent, PetscObject, PetscObject, PetscObject, PetscObject);
121: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventSync(PetscLogHandler, PetscLogEvent, MPI_Comm);
122: PETSC_EXTERN PetscErrorCode PetscLogHandlerObjectCreate(PetscLogHandler, PetscObject);
123: PETSC_EXTERN PetscErrorCode PetscLogHandlerObjectDestroy(PetscLogHandler, PetscObject);
124: PETSC_EXTERN PetscErrorCode PetscLogHandlerStagePush(PetscLogHandler, PetscLogStage);
125: PETSC_EXTERN PetscErrorCode PetscLogHandlerStagePop(PetscLogHandler, PetscLogStage);
126: PETSC_EXTERN PetscErrorCode PetscLogHandlerView(PetscLogHandler, PetscViewer);
128: PETSC_EXTERN PetscErrorCode PetscLogHandlerGetEventPerfInfo(PetscLogHandler, PetscLogStage, PetscLogEvent, PetscEventPerfInfo **);
129: PETSC_EXTERN PetscErrorCode PetscLogHandlerGetStagePerfInfo(PetscLogHandler, PetscLogStage, PetscEventPerfInfo **);
130: PETSC_EXTERN PetscErrorCode PetscLogHandlerSetLogActions(PetscLogHandler, PetscBool);
131: PETSC_EXTERN PetscErrorCode PetscLogHandlerSetLogObjects(PetscLogHandler, PetscBool);
132: PETSC_EXTERN PetscErrorCode PetscLogHandlerLogObjectState(PetscLogHandler, PetscObject, const char[], ...);
133: PETSC_EXTERN PetscErrorCode PetscLogHandlerGetNumObjects(PetscLogHandler, PetscInt *);
134: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventDeactivatePush(PetscLogHandler, PetscLogStage, PetscLogEvent);
135: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventDeactivatePop(PetscLogHandler, PetscLogStage, PetscLogEvent);
136: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventsPause(PetscLogHandler);
137: PETSC_EXTERN PetscErrorCode PetscLogHandlerEventsResume(PetscLogHandler);
138: PETSC_EXTERN PetscErrorCode PetscLogHandlerDump(PetscLogHandler, const char[]);
139: PETSC_EXTERN PetscErrorCode PetscLogHandlerStageSetVisible(PetscLogHandler, PetscLogStage, PetscBool);
140: PETSC_EXTERN PetscErrorCode PetscLogHandlerStageGetVisible(PetscLogHandler, PetscLogStage, PetscBool *);
142: PETSC_EXTERN PetscErrorCode PetscLogHandlerCreateTrace(MPI_Comm, FILE *, PetscLogHandler *);
143: PETSC_EXTERN PetscErrorCode PetscLogHandlerCreateLegacy(MPI_Comm, PetscErrorCode (*)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject), PetscErrorCode (*)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject), PetscErrorCode (*)(PetscObject), PetscErrorCode (*)(PetscObject), PetscLogHandler *);
145: /*MC
146: PetscLogStateStageEventIsActive - Returns whether a specific `PetscLogEvent` is active in a specific stage of a `PetscLogState`
148: Synopsis:
149: #include <petsclog.h>
150: PetscBool PetscLogStateStageEventIsActive(PetscLogState state, PetscLogStage stage, PetscLogEvent event)
152: Not Collective; No Fortran Support
154: Input Parameters:
155: + state - the `PetscLogState`
156: . stage - the stage index
157: - event - the event index
159: Level: developer
161: Note:
162: Returns `PETSC_FALSE` when `stage` is negative, which indicates that there have been more stage pops than pushes.
164: .seealso: `PetscLogState`, `PetscLogStage`, `PetscLogEvent`, `PetscLogStateEventCurrentlyActive()`, `PetscLogStateStageGetActive()`
165: M*/
166: /* All events are inactive if an invalid stage is set, like if there have been more stage pops than stage pushes */
167: #define PetscLogStateStageEventIsActive(state, stage, event) (((stage) >= 0) && PetscBTLookup((state)->active, (stage)) && PetscBTLookup((state)->active, (stage) + ((event) + 1) * (state)->bt_num_stages))
169: /*MC
170: PetscLogStateEventCurrentlyActive - Returns whether a specific `PetscLogEvent` is active in the currently active stage of a `PetscLogState`
172: Synopsis:
173: #include <petsclog.h>
174: PetscBool PetscLogStateEventCurrentlyActive(PetscLogState state, PetscLogEvent event)
176: Not Collective; No Fortran Support
178: Input Parameters:
179: + state - the `PetscLogState` (may be `NULL`, in which case `PETSC_FALSE` is returned)
180: - event - the event index
182: Level: developer
184: .seealso: `PetscLogState`, `PetscLogEvent`, `PetscLogStateStageEventIsActive()`
185: M*/
186: #define PetscLogStateEventCurrentlyActive(state, event) ((state) && PetscLogStateStageEventIsActive(state, (state)->current_stage, event))
188: /* PetscLogHandler with critical methods exposed for speed */
189: typedef struct _n_PetscLogHandlerHot {
190: PetscLogHandler handler;
191: PetscErrorCode (*eventBegin)(PetscLogHandler, PetscLogEvent, PetscObject, PetscObject, PetscObject, PetscObject);
192: PetscErrorCode (*eventEnd)(PetscLogHandler, PetscLogEvent, PetscObject, PetscObject, PetscObject, PetscObject);
193: PetscErrorCode (*eventSync)(PetscLogHandler, PetscLogEvent, MPI_Comm);
194: PetscErrorCode (*objectCreate)(PetscLogHandler, PetscObject);
195: PetscErrorCode (*objectDestroy)(PetscLogHandler, PetscObject);
196: } PetscLogHandlerHot;
198: #if PetscDefined(HAVE_THREADSAFETY) && PetscDefined(USE_LOG)
199: PETSC_EXTERN PetscErrorCode PetscAddLogDouble(PetscLogDouble *, PetscLogDouble *, PetscLogDouble);
200: PETSC_EXTERN PetscErrorCode PetscAddLogDoubleCnt(PetscLogDouble *, PetscLogDouble *, PetscLogDouble *, PetscLogDouble *, PetscLogDouble);
201: #else
202: #define PetscAddLogDouble(a, b, c) ((PetscErrorCode)((*(a) += (c), PETSC_SUCCESS) || ((*(b) += (c)), PETSC_SUCCESS)))
203: #define PetscAddLogDoubleCnt(a, b, c, d, e) ((PetscErrorCode)(PetscAddLogDouble(a, c, 1) || PetscAddLogDouble(b, d, e)))
204: #endif
206: PETSC_DEPRECATED_FUNCTION(3, 18, 0, "PetscLogObjectParent()", ) static inline PetscErrorCode PetscLogObjectParent(PetscObject o, PetscObject p)
207: {
208: (void)o;
209: (void)p;
210: return PETSC_SUCCESS;
211: }
212: /*MC
213: PetscLogObjectParents - Record a parent/child relationship between a `PetscObject` and an array of `PetscObject`s for performance logging
215: Synopsis:
216: #include <petsclog.h>
217: PetscErrorCode PetscLogObjectParents(PetscObject p, int n, PetscObject d[])
219: Not Collective; No Fortran Support
221: Input Parameters:
222: + p - the parent `PetscObject`
223: . n - the number of child objects
224: - d - array of child `PetscObject`s
226: Level: developer
228: Note:
229: This routine is a no-op unless PETSc is configured with logging enabled.
231: .seealso: `PetscLogObjectParent()`, `PetscLogObjectCreate()`, `PetscLogObjectDestroy()`
232: M*/
233: #define PetscLogObjectParents(p, n, d) PetscMacroReturnStandard(for (int _i = 0; _i < (n); ++_i) PetscCall(PetscLogObjectParent((PetscObject)(p), (PetscObject)(d)[_i]));)
235: PETSC_DEPRECATED_FUNCTION(3, 18, 0, "PetscLogObjectMemory()", ) static inline PetscErrorCode PetscLogObjectMemory(PetscObject o, PetscLogDouble m)
236: {
237: (void)o;
238: (void)m;
239: return PETSC_SUCCESS;
240: }
242: /* Global flop counter */
243: PETSC_EXTERN PetscLogDouble petsc_TotalFlops;
244: PETSC_EXTERN PetscLogDouble petsc_irecv_ct;
245: PETSC_EXTERN PetscLogDouble petsc_isend_ct;
246: PETSC_EXTERN PetscLogDouble petsc_recv_ct;
247: PETSC_EXTERN PetscLogDouble petsc_send_ct;
248: PETSC_EXTERN PetscLogDouble petsc_irecv_len;
249: PETSC_EXTERN PetscLogDouble petsc_isend_len;
250: PETSC_EXTERN PetscLogDouble petsc_recv_len;
251: PETSC_EXTERN PetscLogDouble petsc_send_len;
252: PETSC_EXTERN PetscLogDouble petsc_allreduce_ct;
253: PETSC_EXTERN PetscLogDouble petsc_gather_ct;
254: PETSC_EXTERN PetscLogDouble petsc_scatter_ct;
255: PETSC_EXTERN PetscLogDouble petsc_wait_ct;
256: PETSC_EXTERN PetscLogDouble petsc_wait_any_ct;
257: PETSC_EXTERN PetscLogDouble petsc_wait_all_ct;
258: PETSC_EXTERN PetscLogDouble petsc_sum_of_waits_ct;
260: /* Thread local storage */
261: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_TotalFlops_th;
262: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_irecv_ct_th;
263: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_isend_ct_th;
264: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_recv_ct_th;
265: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_send_ct_th;
266: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_irecv_len_th;
267: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_isend_len_th;
268: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_recv_len_th;
269: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_send_len_th;
270: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_allreduce_ct_th;
271: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gather_ct_th;
272: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_scatter_ct_th;
273: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_wait_ct_th;
274: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_wait_any_ct_th;
275: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_wait_all_ct_th;
276: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_sum_of_waits_ct_th;
278: /* Global GPU counters */
279: PETSC_EXTERN PetscLogDouble petsc_ctog_ct;
280: PETSC_EXTERN PetscLogDouble petsc_gtoc_ct;
281: PETSC_EXTERN PetscLogDouble petsc_ctog_sz;
282: PETSC_EXTERN PetscLogDouble petsc_gtoc_sz;
283: PETSC_EXTERN PetscLogDouble petsc_ctog_ct_scalar;
284: PETSC_EXTERN PetscLogDouble petsc_gtoc_ct_scalar;
285: PETSC_EXTERN PetscLogDouble petsc_ctog_sz_scalar;
286: PETSC_EXTERN PetscLogDouble petsc_gtoc_sz_scalar;
287: PETSC_EXTERN PetscLogDouble petsc_gflops;
288: PETSC_EXTERN PetscLogDouble petsc_gtime;
289: PETSC_EXTERN PetscLogDouble petsc_genergy;
290: PETSC_EXTERN PetscLogDouble petsc_genergy_meter;
292: /* Thread local storage */
293: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_ctog_ct_th;
294: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gtoc_ct_th;
295: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_ctog_sz_th;
296: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gtoc_sz_th;
297: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_ctog_ct_scalar_th;
298: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gtoc_ct_scalar_th;
299: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_ctog_sz_scalar_th;
300: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gtoc_sz_scalar_th;
301: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gflops_th;
302: PETSC_EXTERN PETSC_TLS PetscLogDouble petsc_gtime_th;
304: PETSC_EXTERN PetscBool PetscLogMemory;
305: PETSC_EXTERN PetscBool PetscLogSyncOn; /* true if logging synchronization is enabled */
307: PETSC_EXTERN PetscLogState petsc_log_state;
309: #define PETSC_LOG_HANDLER_MAX 4
310: PETSC_EXTERN PetscLogHandlerHot PetscLogHandlers[PETSC_LOG_HANDLER_MAX];
312: #if PetscDefined(USE_LOG) /* --- Logging is turned on --------------------------------*/
313: PETSC_EXTERN PetscErrorCode PetscGetFlops(PetscLogDouble *);
315: PETSC_EXTERN PetscErrorCode PetscLogObjectState(PetscObject, const char[], ...) PETSC_ATTRIBUTE_FORMAT(2, 3);
317: /* Initialization functions */
318: PETSC_EXTERN PetscErrorCode PetscLogDefaultBegin(void);
319: PETSC_EXTERN PetscErrorCode PetscLogNestedBegin(void);
320: PETSC_EXTERN PetscErrorCode PetscLogTraceBegin(FILE *);
321: PETSC_EXTERN PetscErrorCode PetscLogMPEBegin(void);
322: PETSC_EXTERN PetscErrorCode PetscLogPerfstubsBegin(void);
323: PETSC_EXTERN PetscErrorCode PetscLogLegacyCallbacksBegin(PetscErrorCode (*)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject), PetscErrorCode (*)(PetscLogEvent, int, PetscObject, PetscObject, PetscObject, PetscObject), PetscErrorCode (*)(PetscObject), PetscErrorCode (*)(PetscObject));
324: PETSC_EXTERN PetscErrorCode PetscLogActions(PetscBool);
325: PETSC_EXTERN PetscErrorCode PetscLogObjects(PetscBool);
326: PETSC_EXTERN PetscErrorCode PetscLogSetThreshold(PetscLogDouble, PetscLogDouble *);
328: /* Output functions */
329: PETSC_EXTERN PetscErrorCode PetscLogView(PetscViewer);
330: PETSC_EXTERN PetscErrorCode PetscLogViewFromOptions(void);
331: PETSC_EXTERN PetscErrorCode PetscLogDump(const char[]);
332: PETSC_EXTERN PetscErrorCode PetscLogMPEDump(const char[]);
334: PETSC_EXTERN PetscErrorCode PetscLogGetState(PetscLogState *);
335: PETSC_EXTERN PetscErrorCode PetscLogGetDefaultHandler(PetscLogHandler *);
336: PETSC_EXTERN PetscErrorCode PetscLogHandlerStart(PetscLogHandler);
337: PETSC_EXTERN PetscErrorCode PetscLogHandlerStop(PetscLogHandler);
339: /* Status checking functions */
340: PETSC_EXTERN PetscErrorCode PetscLogIsActive(PetscBool *);
342: /* Stage functions */
343: PETSC_EXTERN PetscErrorCode PetscLogStageRegister(const char[], PetscLogStage *);
344: PETSC_EXTERN PetscErrorCode PetscLogStagePush(PetscLogStage);
345: PETSC_EXTERN PetscErrorCode PetscLogStagePop(void);
346: PETSC_EXTERN PetscErrorCode PetscLogStageSetActive(PetscLogStage, PetscBool);
347: PETSC_EXTERN PetscErrorCode PetscLogStageGetActive(PetscLogStage, PetscBool *);
348: PETSC_EXTERN PetscErrorCode PetscLogStageSetVisible(PetscLogStage, PetscBool);
349: PETSC_EXTERN PetscErrorCode PetscLogStageGetVisible(PetscLogStage, PetscBool *);
350: PETSC_EXTERN PetscErrorCode PetscLogStageGetId(const char[], PetscLogStage *);
351: PETSC_EXTERN PetscErrorCode PetscLogStageGetName(PetscLogEvent, const char *[]);
352: PETSC_EXTERN PetscErrorCode PetscLogStageGetPerfInfo(PetscLogStage, PetscEventPerfInfo *);
354: /* Event functions */
355: PETSC_EXTERN PetscErrorCode PetscLogEventRegister(const char[], PetscClassId, PetscLogEvent *);
356: PETSC_EXTERN PetscErrorCode PetscLogEventSetCollective(PetscLogEvent, PetscBool);
357: PETSC_EXTERN PetscErrorCode PetscLogEventIncludeClass(PetscClassId);
358: PETSC_EXTERN PetscErrorCode PetscLogEventExcludeClass(PetscClassId);
359: PETSC_EXTERN PetscErrorCode PetscLogEventActivate(PetscLogEvent);
360: PETSC_EXTERN PetscErrorCode PetscLogEventDeactivate(PetscLogEvent);
361: PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePush(PetscLogEvent);
362: PETSC_EXTERN PetscErrorCode PetscLogEventDeactivatePop(PetscLogEvent);
363: PETSC_EXTERN PetscErrorCode PetscLogEventSetActiveAll(PetscLogEvent, PetscBool);
364: PETSC_EXTERN PetscErrorCode PetscLogEventActivateClass(PetscClassId);
365: PETSC_EXTERN PetscErrorCode PetscLogEventDeactivateClass(PetscClassId);
366: PETSC_EXTERN PetscErrorCode PetscLogEventGetId(const char[], PetscLogEvent *);
367: PETSC_EXTERN PetscErrorCode PetscLogEventGetName(PetscLogEvent, const char *[]);
368: PETSC_EXTERN PetscErrorCode PetscLogEventGetPerfInfo(PetscLogStage, PetscLogEvent, PetscEventPerfInfo *);
369: PETSC_EXTERN PetscErrorCode PetscLogEventSetDof(PetscLogEvent, PetscInt, PetscLogDouble);
370: PETSC_EXTERN PetscErrorCode PetscLogEventSetError(PetscLogEvent, PetscInt, PetscLogDouble);
371: PETSC_EXTERN PetscErrorCode PetscLogEventsPause(void);
372: PETSC_EXTERN PetscErrorCode PetscLogEventsResume(void);
374: /* Class functions */
375: PETSC_EXTERN PetscErrorCode PetscLogClassGetClassId(const char[], PetscClassId *);
376: PETSC_EXTERN PetscErrorCode PetscLogClassIdGetName(PetscClassId, const char *[]);
378: /*@
379: PetscLogEventSync - Synchronize an `MPI_Comm` so that the wall-clock time spent waiting at the implicit barrier is not attributed to a subsequent event
381: Logically Collective on `comm`; No Fortran Support
383: Input Parameters:
384: + e - the `PetscLogEvent` to associate with the synchronization
385: - comm - the `MPI_Comm` whose ranks are synchronized
387: Level: developer
389: Notes:
390: Forwards the call to every active `PetscLogHandler` that implements an `eventSync` method.
392: Has no effect when logging is not active or when the event is not active in the current stage.
394: .seealso: `PetscLogEvent`, `PetscLogEventBegin()`, `PetscLogEventEnd()`, `PetscLogHandler`
395: @*/
396: static inline PetscErrorCode PetscLogEventSync(PetscLogEvent e, MPI_Comm comm)
397: {
398: if (PetscLogStateEventCurrentlyActive(petsc_log_state, e)) {
399: for (int i = 0; i < PETSC_LOG_HANDLER_MAX; i++) {
400: PetscLogHandlerHot *h = &PetscLogHandlers[i];
401: if (h->eventSync) {
402: PetscErrorCode err = (*h->eventSync)(h->handler, e, comm);
403: if (err != PETSC_SUCCESS) return err;
404: }
405: }
406: }
407: return PETSC_SUCCESS;
408: }
410: static inline PetscErrorCode PetscLogEventBegin_Internal(PetscLogEvent e, PetscObject o1, PetscObject o2, PetscObject o3, PetscObject o4)
411: {
412: if (PetscLogStateEventCurrentlyActive(petsc_log_state, e)) {
413: for (int i = 0; i < PETSC_LOG_HANDLER_MAX; i++) {
414: PetscLogHandlerHot *h = &PetscLogHandlers[i];
415: if (h->eventBegin) {
416: PetscErrorCode err = (*h->eventBegin)(h->handler, e, o1, o2, o3, o4);
417: if (err != PETSC_SUCCESS) return err;
418: }
419: }
420: }
421: return PETSC_SUCCESS;
422: }
423: /*MC
424: PetscLogEventBegin - Log the start of an instance of a `PetscLogEvent`
426: Synopsis:
427: #include <petsclog.h>
428: PetscErrorCode PetscLogEventBegin(PetscLogEvent e, PetscObject o1, PetscObject o2, PetscObject o3, PetscObject o4)
430: Not Collective; No Fortran Support
432: Input Parameters:
433: + e - the `PetscLogEvent`
434: . o1 - first object involved in the event (may be `NULL`)
435: . o2 - second object involved in the event (may be `NULL`)
436: . o3 - third object involved in the event (may be `NULL`)
437: - o4 - fourth object involved in the event (may be `NULL`)
439: Level: intermediate
441: Note:
442: Forwards the call to every active `PetscLogHandler` that implements an `eventBegin` method. Use
443: `PetscLogEventEnd()` to log the end of the same event.
445: .seealso: `PetscLogEvent`, `PetscLogEventRegister()`, `PetscLogEventEnd()`, `PetscLogEventSync()`
446: M*/
447: #define PetscLogEventBegin(e, o1, o2, o3, o4) PetscLogEventBegin_Internal(e, (PetscObject)(o1), (PetscObject)(o2), (PetscObject)(o3), (PetscObject)(o4))
449: static inline PetscErrorCode PetscLogEventEnd_Internal(PetscLogEvent e, PetscObject o1, PetscObject o2, PetscObject o3, PetscObject o4)
450: {
451: if (PetscLogStateEventCurrentlyActive(petsc_log_state, e)) {
452: for (int i = 0; i < PETSC_LOG_HANDLER_MAX; i++) {
453: PetscLogHandlerHot *h = &PetscLogHandlers[i];
454: if (h->eventEnd) {
455: PetscErrorCode err = (*h->eventEnd)(h->handler, e, o1, o2, o3, o4);
456: if (err != PETSC_SUCCESS) return err;
457: }
458: }
459: }
460: return PETSC_SUCCESS;
461: }
462: /*MC
463: PetscLogEventEnd - Log the end of an instance of a `PetscLogEvent`
465: Synopsis:
466: #include <petsclog.h>
467: PetscErrorCode PetscLogEventEnd(PetscLogEvent e, PetscObject o1, PetscObject o2, PetscObject o3, PetscObject o4)
469: Not Collective; No Fortran Support
471: Input Parameters:
472: + e - the `PetscLogEvent`
473: . o1 - first object involved in the event (may be `NULL`)
474: . o2 - second object involved in the event (may be `NULL`)
475: . o3 - third object involved in the event (may be `NULL`)
476: - o4 - fourth object involved in the event (may be `NULL`)
478: Level: intermediate
480: Note:
481: Must be paired with a prior `PetscLogEventBegin()` for the same event.
483: .seealso: `PetscLogEvent`, `PetscLogEventRegister()`, `PetscLogEventBegin()`, `PetscLogEventSync()`
484: M*/
485: #define PetscLogEventEnd(e, o1, o2, o3, o4) PetscLogEventEnd_Internal(e, (PetscObject)(o1), (PetscObject)(o2), (PetscObject)(o3), (PetscObject)(o4))
487: /* Object functions */
488: /*@
489: PetscLogObjectCreate - Notify the active `PetscLogHandler`s that a new `PetscObject` has been created
491: Not Collective; No Fortran Support
493: Input Parameter:
494: . o - the newly created `PetscObject`
496: Level: developer
498: Note:
499: Called internally by PETSc object constructors; users normally do not need to call this directly.
501: .seealso: `PetscLogObjectDestroy()`, `PetscLogHandler`, `PetscObject`
502: @*/
503: static inline PetscErrorCode PetscLogObjectCreate(PetscObject o)
504: {
505: if (petsc_log_state) {
506: for (int i = 0; i < PETSC_LOG_HANDLER_MAX; i++) {
507: PetscLogHandlerHot *h = &PetscLogHandlers[i];
508: if (h->objectCreate) {
509: PetscErrorCode err = (*h->objectCreate)(h->handler, o);
510: if (err != PETSC_SUCCESS) return err;
511: }
512: }
513: }
514: return PETSC_SUCCESS;
515: }
517: /*@
518: PetscLogObjectDestroy - Notify the active `PetscLogHandler`s that a `PetscObject` is being destroyed
520: Not Collective; No Fortran Support
522: Input Parameter:
523: . o - the `PetscObject` that is being destroyed
525: Level: developer
527: Note:
528: Called internally by PETSc object destructors; users normally do not need to call this directly.
530: .seealso: `PetscLogObjectCreate()`, `PetscLogHandler`, `PetscObject`
531: @*/
532: static inline PetscErrorCode PetscLogObjectDestroy(PetscObject o)
533: {
534: if (petsc_log_state) {
535: for (int i = 0; i < PETSC_LOG_HANDLER_MAX; i++) {
536: PetscLogHandlerHot *h = &PetscLogHandlers[i];
537: if (h->objectDestroy) {
538: PetscErrorCode err = (*h->objectDestroy)(h->handler, o);
539: if (err != PETSC_SUCCESS) return err;
540: }
541: }
542: }
543: return PETSC_SUCCESS;
544: }
546: /*
547: Flop counting: We count each arithmetic operation (e.g., addition, multiplication) separately.
549: For the complex numbers version, note that
550: 1 complex addition = 2 flops
551: 1 complex multiplication = 6 flops,
552: where we define 1 flop as that for a double precision scalar. We roughly approximate
553: flop counting for complex numbers by multiplying the total flops by 4; this corresponds
554: to the assumption that we're counting mostly additions and multiplications -- and
555: roughly the same number of each. More accurate counting could be done by distinguishing
556: among the various arithmetic operations.
557: */
559: #if PetscDefined(USE_COMPLEX)
560: #define PETSC_FLOPS_PER_OP 4.0
561: #else
562: #define PETSC_FLOPS_PER_OP 1.0
563: #endif
565: /*@
566: PetscLogFlops - Log how many flops are performed in a calculation
568: Not Collective
570: Input Parameter:
571: . n - the number of flops
573: Level: intermediate
575: Note:
576: To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double,
577: not an integer. Use `PetscLogFlops`(4.0*n) not `PetscLogFlops`(4*n)
579: .seealso: [](ch_profiling), `PetscLogView()`, `PetscLogGpuFlops()`
580: @*/
581: static inline PetscErrorCode PetscLogFlops(PetscLogDouble n)
582: {
583: PetscAssert(n >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Cannot log negative flops");
584: return PetscAddLogDouble(&petsc_TotalFlops, &petsc_TotalFlops_th, PETSC_FLOPS_PER_OP * n);
585: }
587: /*
588: These are used internally in the PETSc routines to keep a count of MPI messages and
589: their sizes.
591: This does not work for MPI-Uni because our include/petsc/mpiuni/mpi.h file
592: uses macros to defined the MPI operations.
594: It does not work correctly from HP-UX because it processes the
595: macros in a way that sometimes it double counts, hence
596: PETSC_HAVE_BROKEN_RECURSIVE_MACRO
598: It does not work with Windows because winmpich lacks MPI_Type_size()
599: */
600: #if !defined(MPIUNI_H) && !PetscDefined(HAVE_BROKEN_RECURSIVE_MACRO)
601: /*
602: Logging of MPI activities
603: */
604: /*@
605: PetscMPITypeSize - Atomically add a `PetscLogDouble` value to both a global counter and its per-thread counterpart based on the
606: length of an `MPI_Datatype`
608: Not Collective; No Fortran Support
610: Input Parameters:
611: + count - the length of the array containing a specific `MPI_Datatype`
612: . type - the `MPI_Datatype`
613: . length - pointer to the global counter to update
614: - length_th - pointer to the per-thread counter to update
616: Level: developer
618: Note:
619: When PETSc is built without thread safety this is a fast routine that performs the same update without locking.
621: Developer Note:
622: This should have a better name related to its functionality
624: .seealso: `PetscAddLogDouble()`, `PetscAddLogDoubleCnt()`, `PetscLogFlops()`, `PetscLogDouble`, `PetscMPITypeSizeComm()`
625: @*/
626: static inline PetscErrorCode PetscMPITypeSize(PetscCount count, MPI_Datatype type, PetscLogDouble *length, PetscLogDouble *length_th)
627: {
628: PetscMPIInt typesize;
630: if (type == MPI_DATATYPE_NULL) return PETSC_SUCCESS;
631: PetscCallMPI(MPI_Type_size(type, &typesize));
632: return PetscAddLogDouble(length, length_th, (PetscLogDouble)(count * typesize));
633: }
635: /*@
636: PetscMPITypeSizeComm - Atomically add a `PetscLogDouble` value to both a global counter and its per-thread counterpart based on the
637: length of an `MPI_Datatype`
639: Not Collective; No Fortran Support
641: Input Parameters:
642: + comm - the `MPI_Comm` that shared the communication of the `MPI_Datatype`
643: . counts - the length of the array containing a specific `MPI_Datatype` being sent to each MPI process
644: . type - the `MPI_Datatype`
645: . length - pointer to the global counter to update
646: - length_th - pointer to the per-thread counter to update
648: Level: developer
650: Note:
651: When PETSc is built without thread safety this is a fast routine that performs the same update without locking.
653: Developer Note:
654: This should have a better name related to its functionality
656: .seealso: `PetscAddLogDouble()`, `PetscAddLogDoubleCnt()`, `PetscLogFlops()`, `PetscLogDouble`, `PetscMPITypeSize()`
657: @*/
658: static inline PetscErrorCode PetscMPITypeSizeComm(MPI_Comm comm, const PetscMPIInt *counts, MPI_Datatype type, PetscLogDouble *length, PetscLogDouble *length_th)
659: {
660: PetscMPIInt typesize, size, p;
661: PetscLogDouble l;
663: if (type == MPI_DATATYPE_NULL) return PETSC_SUCCESS;
664: PetscCallMPI(MPI_Comm_size(comm, &size));
665: PetscCallMPI(MPI_Type_size(type, &typesize));
666: for (p = 0, l = 0.0; p < size; ++p) l += (PetscLogDouble)(counts[p] * typesize);
667: return PetscAddLogDouble(length, length_th, l);
668: }
670: /*
671: Returns 1 if the communicator is parallel else zero
672: */
673: static inline int PetscMPIParallelComm(MPI_Comm comm)
674: {
675: PetscMPIInt size;
676: MPI_Comm_size(comm, &size);
677: return size > 1;
678: }
680: #define MPI_Irecv(buf, count, datatype, source, tag, comm, request) \
681: (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || MPI_Irecv((buf), (count), (datatype), (source), (tag), (comm), (request)))
683: #define MPI_Irecv_c(buf, count, datatype, source, tag, comm, request) \
684: (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || MPI_Irecv_c((buf), (count), (datatype), (source), (tag), (comm), (request)))
686: #define MPI_Isend(buf, count, datatype, dest, tag, comm, request) \
687: (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || MPI_Isend((buf), (count), (datatype), (dest), (tag), (comm), (request)))
689: #define MPI_Isend_c(buf, count, datatype, dest, tag, comm, request) \
690: (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || MPI_Isend_c((buf), (count), (datatype), (dest), (tag), (comm), (request)))
692: #define MPI_Startall_irecv(count, datatype, number, requests) \
693: (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, number) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || ((number) && MPI_Startall((number), (requests))))
695: #define MPI_Startall_isend(count, datatype, number, requests) \
696: (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, number) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || ((number) && MPI_Startall((number), (requests))))
698: #define MPI_Start_isend(count, datatype, requests) (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_isend_len), (&petsc_isend_len_th)) || MPI_Start(requests))
700: #define MPI_Recv(buf, count, datatype, source, tag, comm, status) \
701: (PetscAddLogDouble(&petsc_recv_ct, &petsc_recv_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPI_Recv((buf), (count), (datatype), (source), (tag), (comm), (status)))
703: #define MPI_Recv_c(buf, count, datatype, source, tag, comm, status) \
704: (PetscAddLogDouble(&petsc_recv_ct, &petsc_recv_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len), &(petsc_recv_len_th)) || MPI_Recv_c((buf), (count), (datatype), (source), (tag), (comm), (status)))
706: #define MPI_Send(buf, count, datatype, dest, tag, comm) \
707: (PetscAddLogDouble(&petsc_send_ct, &petsc_send_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Send((buf), (count), (datatype), (dest), (tag), (comm)))
709: #define MPI_Send_c(buf, count, datatype, dest, tag, comm) \
710: (PetscAddLogDouble(&petsc_send_ct, &petsc_send_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Send_c((buf), (count), (datatype), (dest), (tag), (comm)))
712: #define MPI_Wait(request, status) (PetscAddLogDouble(&petsc_wait_ct, &petsc_wait_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, 1) || MPI_Wait((request), (status)))
714: #define MPI_Waitany(a, b, c, d) (PetscAddLogDouble(&petsc_wait_any_ct, &petsc_wait_any_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, 1) || MPI_Waitany((a), (b), (c), (d)))
716: #define MPI_Waitall(count, array_of_requests, array_of_statuses) \
717: (PetscAddLogDouble(&petsc_wait_all_ct, &petsc_wait_all_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, count) || MPI_Waitall((count), (array_of_requests), (array_of_statuses)))
719: #define MPI_Allreduce(sendbuf, recvbuf, count, datatype, op, comm) (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPI_Allreduce((sendbuf), (recvbuf), (count), (datatype), (op), (comm)))
721: #define MPI_Bcast(buffer, count, datatype, root, comm) (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPI_Bcast((buffer), (count), (datatype), (root), (comm)))
723: #define MPI_Reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm) \
724: (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPI_Reduce_scatter_block((sendbuf), (recvbuf), (recvcount), (datatype), (op), (comm)))
726: #define MPI_Alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm) \
727: (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Alltoall((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm)))
729: #define MPI_Alltoallv(sendbuf, sendcnts, sdispls, sendtype, recvbuf, recvcnts, rdispls, recvtype, comm) \
730: (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSizeComm((comm), (sendcnts), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Alltoallv((sendbuf), (sendcnts), (sdispls), (sendtype), (recvbuf), (recvcnts), (rdispls), (recvtype), (comm)))
732: #define MPI_Allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm) \
733: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Allgather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm)))
735: #define MPI_Allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, comm) \
736: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Allgatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (comm)))
738: #define MPI_Gather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm) \
739: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Gather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm)))
741: #define MPI_Gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, root, comm) \
742: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Gatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (root), (comm)))
744: #define MPI_Scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm) \
745: (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), &(petsc_recv_len_th)) || MPI_Scatter((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm)))
747: #define MPI_Scatterv(sendbuf, sendcount, displs, sendtype, recvbuf, recvcount, recvtype, root, comm) \
748: (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), &(petsc_recv_len_th)) || MPI_Scatterv((sendbuf), (sendcount), (displs), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm)))
750: #define MPI_Ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request) \
751: (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Ialltoall((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm), (request)))
753: #define MPI_Ialltoallv(sendbuf, sendcnts, sdispls, sendtype, recvbuf, recvcnts, rdispls, recvtype, comm, request) \
754: (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || PetscMPITypeSizeComm((comm), (sendcnts), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Ialltoallv((sendbuf), (sendcnts), (sdispls), (sendtype), (recvbuf), (recvcnts), (rdispls), (recvtype), (comm), (request)))
756: #define MPI_Iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request) \
757: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Iallgather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (comm), (request)))
759: #define MPI_Iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, comm, request) \
760: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, PetscMPIParallelComm(comm)) || MPI_Iallgatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (comm), (request)))
762: #define MPI_Igather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \
763: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Igather((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request)))
765: #define MPI_Igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcount, displs, recvtype, root, comm, request) \
766: (PetscAddLogDouble(&petsc_gather_ct, &petsc_gather_ct_th, 1) || PetscMPITypeSize((sendcount), (sendtype), (&petsc_send_len), (&petsc_send_len_th)) || MPI_Igatherv((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (displs), (recvtype), (root), (comm), (request)))
768: #define MPI_Iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \
769: (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPI_Iscatter((sendbuf), (sendcount), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request)))
771: #define MPI_Iscatterv(sendbuf, sendcount, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, request) \
772: (PetscAddLogDouble(&petsc_scatter_ct, &petsc_scatter_ct_th, 1) || PetscMPITypeSize((recvcount), (recvtype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPI_Iscatterv((sendbuf), (sendcount), (displs), (sendtype), (recvbuf), (recvcount), (recvtype), (root), (comm), (request)))
774: #define MPIX_Send_enqueue(buf, count, datatype, dest, tag, comm) \
775: (PetscAddLogDouble(&petsc_send_ct, &petsc_send_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_send_len), (&petsc_send_len_th)) || MPIX_Send_enqueue((buf), (count), (datatype), (dest), (tag), (comm)))
777: #define MPIX_Recv_enqueue(buf, count, datatype, source, tag, comm, status) \
778: (PetscAddLogDouble(&petsc_recv_ct, &petsc_recv_ct_th, 1) || PetscMPITypeSize((count), (datatype), (&petsc_recv_len), (&petsc_recv_len_th)) || MPIX_Recv_enqueue((buf), (count), (datatype), (source), (tag), (comm), (status)))
780: #define MPIX_Isend_enqueue(buf, count, datatype, dest, tag, comm, request) \
781: (PetscAddLogDouble(&petsc_isend_ct, &petsc_isend_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_isend_len), &(petsc_isend_len_th)) || MPIX_Isend_enqueue((buf), (count), (datatype), (dest), (tag), (comm), (request)))
783: #define MPIX_Irecv_enqueue(buf, count, datatype, source, tag, comm, request) \
784: (PetscAddLogDouble(&petsc_irecv_ct, &petsc_irecv_ct_th, 1) || PetscMPITypeSize((count), (datatype), &(petsc_irecv_len), &(petsc_irecv_len_th)) || MPIX_Irecv_enqueue((buf), (count), (datatype), (source), (tag), (comm), (request)))
786: #define MPIX_Allreduce_enqueue(sendbuf, recvbuf, count, datatype, op, comm) \
787: (PetscAddLogDouble(&petsc_allreduce_ct, &petsc_allreduce_ct_th, PetscMPIParallelComm(comm)) || MPIX_Allreduce_enqueue((sendbuf), (recvbuf), (count), (datatype), (op), (comm)))
789: #define MPIX_Wait_enqueue(request, status) (PetscAddLogDouble(&petsc_wait_ct, &petsc_wait_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, 1) || MPIX_Wait_enqueue((request), (status)))
791: #define MPIX_Waitall_enqueue(count, array_of_requests, array_of_statuses) \
792: (PetscAddLogDouble(&petsc_wait_all_ct, &petsc_wait_all_ct_th, 1) || PetscAddLogDouble(&petsc_sum_of_waits_ct, &petsc_sum_of_waits_ct_th, count) || MPIX_Waitall_enqueue((count), (array_of_requests), (array_of_statuses)))
793: #else
795: #define MPI_Startall_irecv(count, datatype, number, requests) ((number) && MPI_Startall((number), (requests)))
797: #define MPI_Startall_isend(count, datatype, number, requests) ((number) && MPI_Startall((number), (requests)))
799: #define MPI_Start_isend(count, datatype, requests) (MPI_Start(requests))
801: #endif /* !MPIUNI_H && ! PETSC_HAVE_BROKEN_RECURSIVE_MACRO */
803: #else /* ---Logging is turned off --------------------------------------------*/
805: #define PetscLogGetState(a) (*(a) = NULL, PETSC_SUCCESS)
806: #define PetscLogGetDefaultHandler(a) (*(a) = NULL, PETSC_SUCCESS)
807: #define PetscLogHandlerStart(a) ((void)(a), PETSC_SUCCESS)
808: #define PetscLogHandlerStop(a) ((void)(a), PETSC_SUCCESS)
810: #define PetscLogFlops(n) ((void)(n), PETSC_SUCCESS)
811: #define PetscGetFlops(a) (*(a) = 0.0, PETSC_SUCCESS)
813: #define PetscLogStageRegister(a, b) ((void)(a), *(b) = -1, PETSC_SUCCESS)
814: #define PetscLogStagePush(a) ((void)(a), PETSC_SUCCESS)
815: #define PetscLogStagePop() PETSC_SUCCESS
816: #define PetscLogStageSetActive(a, b) ((void)(a), (void)(b), PETSC_SUCCESS)
817: #define PetscLogStageGetActive(a, b) ((void)(a), *(b) = PETSC_FALSE, PETSC_SUCCESS)
818: #define PetscLogStageGetVisible(a, b) ((void)(a), *(b) = PETSC_FALSE, PETSC_SUCCESS)
819: #define PetscLogStageSetVisible(a, b) ((void)(a), (void)(b), PETSC_SUCCESS)
820: #define PetscLogStageGetId(a, b) ((void)(a), *(b) = -1, PETSC_SUCCESS)
821: #define PetscLogStageGetName(a, b) ((void)(a), *(b) = NULL, PETSC_SUCCESS)
822: #define PetscLogStageGetPerfInfo(a, b) ((void)(a), *(b) = (const PetscEventPerfInfo){0}, PETSC_SUCCESS)
824: #define PetscLogEventRegister(a, b, c) ((void)(a), (void)(b), *(c) = -1, PETSC_SUCCESS)
825: #define PetscLogEventSetCollective(a, b) ((void)(a), (void)(b), PETSC_SUCCESS)
826: #define PetscLogEventIncludeClass(a) ((void)(a), PETSC_SUCCESS)
827: #define PetscLogEventExcludeClass(a) ((void)(a), PETSC_SUCCESS)
828: #define PetscLogEventActivate(a) ((void)(a), PETSC_SUCCESS)
829: #define PetscLogEventDeactivate(a) ((void)(a), PETSC_SUCCESS)
830: #define PetscLogEventDeactivatePush(a) ((void)(a), PETSC_SUCCESS)
831: #define PetscLogEventDeactivatePop(a) ((void)(a), PETSC_SUCCESS)
832: #define PetscLogEventActivateClass(a) ((void)(a), PETSC_SUCCESS)
833: #define PetscLogEventDeactivateClass(a) ((void)(a), PETSC_SUCCESS)
834: #define PetscLogEventSetActiveAll(a, b) ((void)(a), PETSC_SUCCESS)
835: #define PetscLogEventGetId(a, b) ((void)(a), *(b) = -1, PETSC_SUCCESS)
836: #define PetscLogEventGetName(a, b) ((void)(a), *(b) = NULL, PETSC_SUCCESS)
837: #define PetscLogEventGetPerfInfo(a, b, c) ((void)(a), (void)(b), *(c) = (const PetscEventPerfInfo){0}, PETSC_SUCCESS)
839: #define PetscLogEventSetDof(a, b, c) ((void)(a), (void)(b), (void)(c), PETSC_SUCCESS)
840: #define PetscLogEventSetError(a, b, c) ((void)(a), (void)(b), (void)(c), PETSC_SUCCESS)
841: #define PetscLogEventsPause() PETSC_SUCCESS
842: #define PetscLogEventsResume() PETSC_SUCCESS
844: #define PetscLogClassGetClassId(a, b) (*(b) = -1, PETSC_SUCCESS)
845: #define PetscLogClassIdGetName(a, b) (*(b) = NULL, PETSC_SUCCESS)
847: #define PetscLogObjectCreate(h) ((void)(h), PETSC_SUCCESS)
848: #define PetscLogObjectDestroy(h) ((void)(h), PETSC_SUCCESS)
849: #define PetscLogObjectState(h, c, ...) ((void)(h), (void)(c), PETSC_SUCCESS)
851: #define PetscLogDefaultBegin() PETSC_SUCCESS
852: #define PetscLogNestedBegin() PETSC_SUCCESS
853: #define PetscLogTraceBegin(file) ((void)(file), PETSC_SUCCESS)
854: #define PetscLogMPEBegin() PETSC_SUCCESS
855: #define PetscLogPerfstubsBegin() PETSC_SUCCESS
856: #define PetscLogLegacyCallbacksBegin(a, b, c, d) ((void)(a), (void)(b), (void)(c), (void)(d), PETSC_SUCCESS)
857: #define PetscLogActions(a) ((void)(a), PETSC_SUCCESS)
858: #define PetscLogObjects(a) ((void)(a), PETSC_SUCCESS)
859: #define PetscLogSetThreshold(a, b) ((void)(a), (void)(b), PETSC_SUCCESS)
861: #define PetscLogIsActive(flag) (*(flag) = PETSC_FALSE, PETSC_SUCCESS)
863: #define PetscLogView(viewer) ((void)(viewer), PETSC_SUCCESS)
864: #define PetscLogViewFromOptions() PETSC_SUCCESS
865: #define PetscLogDump(c) ((void)(c), PETSC_SUCCESS)
866: #define PetscLogMPEDump(c) ((void)(c), PETSC_SUCCESS)
868: #define PetscLogEventSync(e, comm) ((void)(e), (void)(comm), PETSC_SUCCESS)
869: #define PetscLogEventBegin(e, o1, o2, o3, o4) ((void)(e), (void)(o1), (void)(o2), (void)(o3), PETSC_SUCCESS)
870: #define PetscLogEventEnd(e, o1, o2, o3, o4) ((void)(e), (void)(o1), (void)(o2), (void)(o3), PETSC_SUCCESS)
872: /* If PETSC_USE_LOG is NOT defined, these still need to be! */
873: #define MPI_Startall_irecv(count, datatype, number, requests) ((number) && MPI_Startall(number, requests))
874: #define MPI_Startall_isend(count, datatype, number, requests) ((number) && MPI_Startall(number, requests))
875: #define MPI_Start_isend(count, datatype, requests) MPI_Start(requests)
877: #endif /* PETSC_USE_LOG */
879: /*MC
880: PetscPreLoadBegin - Begin a block of code that is timed twice so that startup costs (such as JIT or first-touch allocation) are not attributed to the measured run
882: Synopsis:
883: #include <petsclog.h>
884: PetscPreLoadBegin(PetscBool flag, const char name[])
886: Not Collective; No Fortran Support
888: Input Parameters:
889: + flag - whether preloading is desired (may be overridden by the command-line option `-preload`)
890: - name - name to use for the `PetscLogStage` created for the measured run
892: Level: intermediate
894: Note:
895: Use this macro in the form
896: .vb
897: PetscPreLoadBegin(PETSC_TRUE, "Compute");
898: // ... code to be timed
899: PetscPreLoadEnd();
900: .ve
901: It expands to a loop that runs the enclosed code twice when preloading is enabled and once otherwise.
902: Pair with `PetscPreLoadEnd()`. Use `PetscPreLoadStage()` to advance to a new named stage between phases.
904: .seealso: `PetscPreLoadEnd`, `PetscPreLoadStage`, `PetscLogStageRegister()`, `PetscLogStagePush()`
905: M*/
906: #define PetscPreLoadBegin(flag, name) \
907: do { \
908: PetscBool PetscPreLoading = flag; \
909: int PetscPreLoadMax, PetscPreLoadIt; \
910: PetscLogStage _stageNum; \
911: PetscCall(PetscOptionsGetBool(NULL, NULL, "-preload", &PetscPreLoading, NULL)); \
912: PetscPreLoadMax = (int)(PetscPreLoading); \
913: PetscPreLoadingUsed = PetscPreLoading ? PETSC_TRUE : PetscPreLoadingUsed; \
914: PetscCall(PetscLogStageGetId(name, &_stageNum)); \
915: for (PetscPreLoadIt = (_stageNum == -1) ? 0 : PetscPreLoadMax; PetscPreLoadIt <= PetscPreLoadMax; PetscPreLoadIt++) { \
916: PetscPreLoadingOn = (PetscBool)(PetscPreLoadIt < PetscPreLoadMax); \
917: PetscCall(PetscBarrier(NULL)); \
918: if (_stageNum == -1) PetscCall(PetscLogStageRegister(name, &_stageNum)); \
919: PetscCall(PetscLogStageSetActive(_stageNum, (PetscBool)(PetscPreLoadIt == PetscPreLoadMax))); \
920: PetscCall(PetscLogStagePush(_stageNum))
922: /*MC
923: PetscPreLoadEnd - Close a preload block started with `PetscPreLoadBegin()`
925: Synopsis:
926: #include <petsclog.h>
927: PetscPreLoadEnd()
929: Not Collective; No Fortran Support
931: Level: intermediate
933: .seealso: `PetscPreLoadBegin`, `PetscPreLoadStage`
934: M*/
935: #define PetscPreLoadEnd() \
936: PetscCall(PetscLogStagePop()); \
937: } \
938: } \
939: while (0)
941: /*MC
942: PetscPreLoadStage - Advance to a new named stage inside a `PetscPreLoadBegin()` / `PetscPreLoadEnd()` block
944: Synopsis:
945: #include <petsclog.h>
946: PetscPreLoadStage(const char name[])
948: Not Collective; No Fortran Support
950: Input Parameter:
951: . name - name for the new `PetscLogStage`
953: Level: intermediate
955: Note:
956: Pops the previous stage, registers (or reuses) the new stage, and pushes it. Only valid between
957: `PetscPreLoadBegin()` and `PetscPreLoadEnd()`.
959: .seealso: `PetscPreLoadBegin`, `PetscPreLoadEnd`, `PetscLogStagePush()`, `PetscLogStagePop()`
960: M*/
961: #define PetscPreLoadStage(name) \
962: do { \
963: PetscCall(PetscLogStagePop()); \
964: PetscCall(PetscLogStageGetId(name, &_stageNum)); \
965: if (_stageNum == -1) PetscCall(PetscLogStageRegister(name, &_stageNum)); \
966: PetscCall(PetscLogStageSetActive(_stageNum, (PetscBool)(!PetscPreLoadMax || PetscPreLoadIt))); \
967: PetscCall(PetscLogStagePush(_stageNum)); \
968: } while (0)
970: /* some vars for logging */
971: PETSC_EXTERN PetscBool PetscPreLoadingUsed; /* true if we are or have done preloading */
972: PETSC_EXTERN PetscBool PetscPreLoadingOn; /* true if we are currently in a preloading calculation */
974: #if PetscDefined(USE_LOG) && PetscDefined(HAVE_DEVICE)
976: PETSC_EXTERN PetscErrorCode PetscLogGpuTime(void);
977: PETSC_EXTERN PetscErrorCode PetscLogGpuTimeBegin(void);
978: PETSC_EXTERN PetscErrorCode PetscLogGpuTimeEnd(void);
979: PETSC_EXTERN PetscErrorCode PetscLogGpuEnergy(void);
980: PETSC_EXTERN PetscErrorCode PetscLogGpuEnergyMeter(void);
981: PETSC_EXTERN PetscErrorCode PetscLogGpuEnergyMeterBegin(void);
982: PETSC_EXTERN PetscErrorCode PetscLogGpuEnergyMeterEnd(void);
984: /*@
985: PetscLogGpuFlops - Log how many flops are performed in a calculation on the device
987: Not Collective
989: Input Parameter:
990: . n - the number of flops
992: Level: intermediate
994: Notes:
995: To limit the chance of integer overflow when multiplying by a constant, represent the constant as a double,
996: not an integer. Use `PetscLogFlops`(4.0*n) not `PetscLogFlops`(4*n)
998: The values are also added to the total flop count for the MPI rank that is set with `PetscLogFlops()`; hence the number of flops
999: just on the CPU would be the value from set from `PetscLogFlops()` minus the value set from `PetscLogGpuFlops()`
1001: .seealso: [](ch_profiling), `PetscLogView()`, `PetscLogFlops()`, `PetscLogGpuTimeBegin()`, `PetscLogGpuTimeEnd()`
1002: @*/
1003: static inline PetscErrorCode PetscLogGpuFlops(PetscLogDouble n)
1004: {
1005: PetscAssert(n >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Cannot log negative flops");
1006: PetscCall(PetscAddLogDouble(&petsc_TotalFlops, &petsc_TotalFlops_th, PETSC_FLOPS_PER_OP * n));
1007: PetscCall(PetscAddLogDouble(&petsc_gflops, &petsc_gflops_th, PETSC_FLOPS_PER_OP * n));
1008: return PETSC_SUCCESS;
1009: }
1011: /*@
1012: PetscLogGpuTimeAdd - Add elapsed GPU computation time to PETSc's GPU time counter
1014: Not Collective; No Fortran Support
1016: Input Parameter:
1017: . t - the elapsed time, in seconds, to add to the GPU time counter
1019: Level: developer
1021: Note:
1022: Used internally by GPU backends to report the time spent in a kernel or library call.
1024: .seealso: `PetscLogGpuTime()`, `PetscLogGpuTimeBegin()`, `PetscLogGpuTimeEnd()`, `PetscLogGpuFlops()`
1025: @*/
1026: static inline PetscErrorCode PetscLogGpuTimeAdd(PetscLogDouble t)
1027: {
1028: return PetscAddLogDouble(&petsc_gtime, &petsc_gtime_th, t);
1029: }
1031: /*@
1032: PetscLogCpuToGpu - Log a CPU-to-GPU memory transfer for performance reporting
1034: Not Collective; No Fortran Support
1036: Input Parameter:
1037: . size - number of bytes transferred from CPU to GPU
1039: Level: developer
1041: Note:
1042: Updates both the transfer-count and transfer-size counters used by `PetscLogView()`.
1044: .seealso: `PetscLogGpuToCpu()`, `PetscLogCpuToGpuScalar()`, `PetscLogGpuFlops()`
1045: @*/
1046: static inline PetscErrorCode PetscLogCpuToGpu(PetscLogDouble size)
1047: {
1048: return PetscAddLogDoubleCnt(&petsc_ctog_ct, &petsc_ctog_sz, &petsc_ctog_ct_th, &petsc_ctog_sz_th, size);
1049: }
1051: /*@
1052: PetscLogGpuToCpu - Log a GPU-to-CPU memory transfer for performance reporting
1054: Not Collective; No Fortran Support
1056: Input Parameter:
1057: . size - number of bytes transferred from GPU to CPU
1059: Level: developer
1061: .seealso: `PetscLogCpuToGpu()`, `PetscLogGpuToCpuScalar()`, `PetscLogGpuFlops()`
1062: @*/
1063: static inline PetscErrorCode PetscLogGpuToCpu(PetscLogDouble size)
1064: {
1065: return PetscAddLogDoubleCnt(&petsc_gtoc_ct, &petsc_gtoc_sz, &petsc_gtoc_ct_th, &petsc_gtoc_sz_th, size);
1066: }
1068: /*@
1069: PetscLogCpuToGpuScalar - Log a CPU-to-GPU memory transfer of `PetscScalar` data for performance reporting
1071: Not Collective; No Fortran Support
1073: Input Parameter:
1074: . size - number of bytes of scalar data transferred from CPU to GPU
1076: Level: developer
1078: Note:
1079: Tracks scalar-only transfers separately from generic byte counters tracked by `PetscLogCpuToGpu()`.
1081: .seealso: `PetscLogCpuToGpu()`, `PetscLogGpuToCpuScalar()`
1082: @*/
1083: static inline PetscErrorCode PetscLogCpuToGpuScalar(PetscLogDouble size)
1084: {
1085: return PetscAddLogDoubleCnt(&petsc_ctog_ct_scalar, &petsc_ctog_sz_scalar, &petsc_ctog_ct_scalar_th, &petsc_ctog_sz_scalar_th, size);
1086: }
1088: /*@
1089: PetscLogGpuToCpuScalar - Log a GPU-to-CPU memory transfer of `PetscScalar` data for performance reporting
1091: Not Collective; No Fortran Support
1093: Input Parameter:
1094: . size - number of bytes of scalar data transferred from GPU to CPU
1096: Level: developer
1098: .seealso: `PetscLogGpuToCpu()`, `PetscLogCpuToGpuScalar()`
1099: @*/
1100: static inline PetscErrorCode PetscLogGpuToCpuScalar(PetscLogDouble size)
1101: {
1102: return PetscAddLogDoubleCnt(&petsc_gtoc_ct_scalar, &petsc_gtoc_sz_scalar, &petsc_gtoc_ct_scalar_th, &petsc_gtoc_sz_scalar_th, size);
1103: }
1104: #else
1106: #define PetscLogCpuToGpu(a) ((void)(a), PETSC_SUCCESS)
1107: #define PetscLogGpuToCpu(a) ((void)(a), PETSC_SUCCESS)
1108: #define PetscLogCpuToGpuScalar(a) ((void)(a), PETSC_SUCCESS)
1109: #define PetscLogGpuToCpuScalar(a) ((void)(a), PETSC_SUCCESS)
1110: #define PetscLogGpuFlops(a) ((void)(a), PETSC_SUCCESS)
1111: #define PetscLogGpuTimeAdd(a) ((void)(a), PETSC_SUCCESS)
1112: #define PetscLogGpuTime() PETSC_SUCCESS
1113: #define PetscLogGpuTimeBegin() PETSC_SUCCESS
1114: #define PetscLogGpuTimeEnd() PETSC_SUCCESS
1116: #define PetscLogGpuEnergy() PETSC_SUCCESS
1117: #define PetscLogGpuEnergyMeter() PETSC_SUCCESS
1118: #define PetscLogGpuEnergyMeterBegin() PETSC_SUCCESS
1119: #define PetscLogGpuEnergyMeterEnd() PETSC_SUCCESS
1121: #endif /* PETSC_USE_LOG && PETSC_HAVE_DEVICE */
1123: #include <petsclogdeprecated.h>