Actual source code: iscoloring.c
1: #include <petsc/private/isimpl.h>
2: #include <petscviewer.h>
3: #include <petscsf.h>
5: const char *const ISColoringTypes[] = {"global", "ghosted", "ISColoringType", "IS_COLORING_", NULL};
7: /*@
8: ISColoringReference - Increases the reference count of an `ISColoring` object by one
10: Logically collective
12: Input Parameter:
13: . coloring - the `ISColoring` object
15: Level: developer
17: Note:
18: The reference count is decreased by a matching call to `ISColoringDestroy()`.
20: .seealso: `ISColoring`, `ISColoringCreate()`, `ISColoringDestroy()`
21: @*/
22: PetscErrorCode ISColoringReference(ISColoring coloring)
23: {
24: PetscFunctionBegin;
25: coloring->refct++;
26: PetscFunctionReturn(PETSC_SUCCESS);
27: }
29: /*@
30: ISColoringSetType - indicates if the coloring is for the local representation (including ghost points) or the global representation of a `Mat`
32: Collective
34: Input Parameters:
35: + coloring - the coloring object
36: - type - either `IS_COLORING_LOCAL` or `IS_COLORING_GLOBAL`
38: Level: intermediate
40: Notes:
41: `IS_COLORING_LOCAL` can lead to faster computations since parallel ghost point updates are not needed for each color
43: With `IS_COLORING_LOCAL` the coloring is in the numbering of the local vector, for `IS_COLORING_GLOBAL` it is in the numbering of the global vector
45: .seealso: `MatFDColoringCreate()`, `ISColoring`, `ISColoringType`, `ISColoringCreate()`, `IS_COLORING_LOCAL`, `IS_COLORING_GLOBAL`, `ISColoringGetType()`
46: @*/
47: PetscErrorCode ISColoringSetType(ISColoring coloring, ISColoringType type)
48: {
49: PetscFunctionBegin;
50: coloring->ctype = type;
51: PetscFunctionReturn(PETSC_SUCCESS);
52: }
54: /*@
55: ISColoringGetType - gets if the coloring is for the local representation (including ghost points) or the global representation
57: Collective
59: Input Parameter:
60: . coloring - the coloring object
62: Output Parameter:
63: . type - either `IS_COLORING_LOCAL` or `IS_COLORING_GLOBAL`
65: Level: intermediate
67: .seealso: `MatFDColoringCreate()`, `ISColoring`, `ISColoringType`, `ISColoringCreate()`, `IS_COLORING_LOCAL`, `IS_COLORING_GLOBAL`, `ISColoringSetType()`
68: @*/
69: PetscErrorCode ISColoringGetType(ISColoring coloring, ISColoringType *type)
70: {
71: PetscFunctionBegin;
72: *type = coloring->ctype;
73: PetscFunctionReturn(PETSC_SUCCESS);
74: }
76: /*@
77: ISColoringDestroy - Destroys an `ISColoring` coloring context.
79: Collective
81: Input Parameter:
82: . iscoloring - the coloring context
84: Level: advanced
86: .seealso: `ISColoring`, `ISColoringView()`, `MatColoring`
87: @*/
88: PetscErrorCode ISColoringDestroy(ISColoring *iscoloring)
89: {
90: PetscInt i;
92: PetscFunctionBegin;
93: if (!*iscoloring) PetscFunctionReturn(PETSC_SUCCESS);
94: PetscAssertPointer(*iscoloring, 1);
95: if (--(*iscoloring)->refct > 0) {
96: *iscoloring = NULL;
97: PetscFunctionReturn(PETSC_SUCCESS);
98: }
100: if ((*iscoloring)->is) {
101: for (i = 0; i < (*iscoloring)->n; i++) PetscCall(ISDestroy(&(*iscoloring)->is[i]));
102: PetscCall(PetscFree((*iscoloring)->is));
103: }
104: if ((*iscoloring)->allocated) PetscCall(PetscFree((*iscoloring)->colors));
105: PetscCall(PetscCommDestroy(&(*iscoloring)->comm));
106: PetscCall(PetscFree(*iscoloring));
107: PetscFunctionReturn(PETSC_SUCCESS);
108: }
110: /*@
111: ISColoringViewFromOptions - Processes command line options to determine if/how an `ISColoring` object is to be viewed.
113: Collective
115: Input Parameters:
116: + obj - the `ISColoring` object
117: . bobj - prefix to use for viewing, or `NULL` to use prefix of `mat`
118: - name - option to activate viewing
120: Options Database Key:
121: . -name [viewertype][:...] - option name and values. See `PetscObjectViewFromOptions()` for the possible arguments
123: Level: intermediate
125: Developer Note:
126: This cannot use `PetscObjectViewFromOptions()` because `ISColoring` is not a `PetscObject`
128: .seealso: `ISColoring`, `ISColoringView()`, `PetscObjectViewFromOptions()`
129: @*/
130: PetscErrorCode ISColoringViewFromOptions(ISColoring obj, PetscObject bobj, const char name[])
131: {
132: PetscViewer viewer;
133: PetscBool flg;
134: PetscViewerFormat format;
135: char *prefix;
137: PetscFunctionBegin;
138: prefix = bobj ? bobj->prefix : NULL;
139: PetscCall(PetscOptionsCreateViewer(obj->comm, NULL, prefix, name, &viewer, &format, &flg));
140: if (flg) {
141: PetscCall(PetscViewerPushFormat(viewer, format));
142: PetscCall(ISColoringView(obj, viewer));
143: PetscCall(PetscViewerPopFormat(viewer));
144: PetscCall(PetscViewerDestroy(&viewer));
145: }
146: PetscFunctionReturn(PETSC_SUCCESS);
147: }
149: /*@
150: ISColoringView - Views an `ISColoring` coloring context.
152: Collective
154: Input Parameters:
155: + iscoloring - the coloring context
156: - viewer - the viewer
158: Level: advanced
160: .seealso: `ISColoring()`, `ISColoringViewFromOptions()`, `ISColoringDestroy()`, `ISColoringGetIS()`, `MatColoring`
161: @*/
162: PetscErrorCode ISColoringView(ISColoring iscoloring, PetscViewer viewer)
163: {
164: PetscInt i;
165: PetscBool isascii;
166: IS *is;
168: PetscFunctionBegin;
169: PetscAssertPointer(iscoloring, 1);
170: if (!viewer) PetscCall(PetscViewerASCIIGetStdout(iscoloring->comm, &viewer));
173: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
174: if (isascii) {
175: MPI_Comm comm;
176: PetscMPIInt size, rank;
178: PetscCall(PetscObjectGetComm((PetscObject)viewer, &comm));
179: PetscCallMPI(MPI_Comm_size(comm, &size));
180: PetscCallMPI(MPI_Comm_rank(comm, &rank));
181: PetscCall(PetscViewerASCIIPrintf(viewer, "ISColoring Object: %d MPI processes\n", size));
182: PetscCall(PetscViewerASCIIPrintf(viewer, "ISColoringType: %s\n", ISColoringTypes[iscoloring->ctype]));
183: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
184: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Number of colors %" PetscInt_FMT "\n", rank, iscoloring->n));
185: PetscCall(PetscViewerFlush(viewer));
186: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
187: }
189: PetscCall(ISColoringGetIS(iscoloring, PETSC_USE_POINTER, PETSC_IGNORE, &is));
190: for (i = 0; i < iscoloring->n; i++) PetscCall(ISView(iscoloring->is[i], viewer));
191: PetscCall(ISColoringRestoreIS(iscoloring, PETSC_USE_POINTER, &is));
192: PetscFunctionReturn(PETSC_SUCCESS);
193: }
195: /*@C
196: ISColoringGetColors - Returns an array with the color for each local node
198: Not Collective
200: Input Parameter:
201: . iscoloring - the coloring context
203: Output Parameters:
204: + n - number of nodes
205: . nc - number of colors
206: - colors - color for each node
208: Level: advanced
210: Notes:
211: Do not free the `colors` array.
213: The `colors` array will only be valid for the lifetime of the `ISColoring`
215: .seealso: `ISColoring`, `ISColoringValue`, `ISColoringRestoreIS()`, `ISColoringView()`, `ISColoringGetIS()`
216: @*/
217: PetscErrorCode ISColoringGetColors(ISColoring iscoloring, PetscInt *n, PetscInt *nc, const ISColoringValue **colors)
218: {
219: PetscFunctionBegin;
220: PetscAssertPointer(iscoloring, 1);
222: if (n) *n = iscoloring->N;
223: if (nc) *nc = iscoloring->n;
224: if (colors) *colors = iscoloring->colors;
225: PetscFunctionReturn(PETSC_SUCCESS);
226: }
228: /*@C
229: ISColoringGetIS - Extracts index sets from the coloring context. Each is contains the nodes of one color
231: Collective
233: Input Parameters:
234: + iscoloring - the coloring context
235: - mode - if this value is `PETSC_OWN_POINTER` then the caller owns the pointer and must free the array of `IS` and each `IS` in the array
237: Output Parameters:
238: + nn - number of index sets in the coloring context
239: - isis - array of index sets
241: Level: advanced
243: Note:
244: If mode is `PETSC_USE_POINTER` then `ISColoringRestoreIS()` must be called when the `IS` are no longer needed
246: .seealso: `ISColoring`, `IS`, `ISColoringRestoreIS()`, `ISColoringView()`, `ISColoringGetColoring()`, `ISColoringGetColors()`
247: @*/
248: PetscErrorCode ISColoringGetIS(ISColoring iscoloring, PetscCopyMode mode, PetscInt *nn, IS *isis[])
249: {
250: PetscFunctionBegin;
251: PetscAssertPointer(iscoloring, 1);
253: if (nn) *nn = iscoloring->n;
254: if (isis) {
255: if (!iscoloring->is) {
256: PetscInt *mcolors, **ii, nc = iscoloring->n, i, base, n = iscoloring->N;
257: ISColoringValue *colors = iscoloring->colors;
258: IS *is;
260: if (PetscDefined(USE_DEBUG)) {
261: for (i = 0; i < n; i++) PetscCheck(((PetscInt)colors[i]) < nc, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Coloring is our of range index %" PetscInt_FMT "value %d number colors %" PetscInt_FMT, i, (int)colors[i], nc);
262: }
264: /* generate the lists of nodes for each color */
265: PetscCall(PetscCalloc1(nc, &mcolors));
266: for (i = 0; i < n; i++) mcolors[colors[i]]++;
268: PetscCall(PetscMalloc1(nc, &ii));
269: PetscCall(PetscMalloc1(n, &ii[0]));
270: for (i = 1; i < nc; i++) ii[i] = ii[i - 1] + mcolors[i - 1];
271: PetscCall(PetscArrayzero(mcolors, nc));
273: if (iscoloring->ctype == IS_COLORING_GLOBAL) {
274: PetscCallMPI(MPI_Scan(&iscoloring->N, &base, 1, MPIU_INT, MPI_SUM, iscoloring->comm));
275: base -= iscoloring->N;
276: for (i = 0; i < n; i++) ii[colors[i]][mcolors[colors[i]]++] = i + base; /* global idx */
277: } else if (iscoloring->ctype == IS_COLORING_LOCAL) {
278: for (i = 0; i < n; i++) ii[colors[i]][mcolors[colors[i]]++] = i; /* local idx */
279: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Not provided for this ISColoringType type");
281: PetscCall(PetscMalloc1(nc, &is));
282: for (i = 0; i < nc; i++) PetscCall(ISCreateGeneral(iscoloring->comm, mcolors[i], ii[i], PETSC_COPY_VALUES, is + i));
284: if (mode != PETSC_OWN_POINTER) iscoloring->is = is;
285: *isis = is;
286: PetscCall(PetscFree(ii[0]));
287: PetscCall(PetscFree(ii));
288: PetscCall(PetscFree(mcolors));
289: } else {
290: *isis = iscoloring->is;
291: if (mode == PETSC_OWN_POINTER) iscoloring->is = NULL;
292: }
293: }
294: PetscFunctionReturn(PETSC_SUCCESS);
295: }
297: /*@C
298: ISColoringRestoreIS - Restores the index sets extracted from the coloring context with `ISColoringGetIS()` using `PETSC_USE_POINTER`
300: Collective
302: Input Parameters:
303: + iscoloring - the coloring context
304: . mode - who retains ownership of the is
305: - is - array of index sets
307: Level: advanced
309: .seealso: `ISColoring()`, `IS`, `ISColoringGetIS()`, `ISColoringView()`, `PetscCopyMode`
310: @*/
311: PetscErrorCode ISColoringRestoreIS(ISColoring iscoloring, PetscCopyMode mode, IS *is[])
312: {
313: PetscFunctionBegin;
314: PetscAssertPointer(iscoloring, 1);
316: /* currently nothing is done here */
317: PetscFunctionReturn(PETSC_SUCCESS);
318: }
320: /*@
321: ISColoringCreate - Generates an `ISColoring` context from lists (provided by each MPI process) of colors for each node.
323: Collective
325: Input Parameters:
326: + comm - communicator for the processors creating the coloring
327: . ncolors - max color value
328: . n - number of nodes on this processor
329: . colors - array containing the colors for this MPI rank, color numbers begin at 0, for each local node
330: - mode - see `PetscCopyMode` for meaning of this flag.
332: Output Parameter:
333: . iscoloring - the resulting coloring data structure
335: Options Database Key:
336: . -is_coloring_view - Activates `ISColoringView()`
338: Level: advanced
340: Notes:
341: By default sets coloring type to `IS_COLORING_GLOBAL`
343: .seealso: `ISColoring`, `ISColoringValue`, `MatColoringCreate()`, `ISColoringView()`, `ISColoringDestroy()`, `ISColoringSetType()`
344: @*/
345: PetscErrorCode ISColoringCreate(MPI_Comm comm, PetscInt ncolors, PetscInt n, const ISColoringValue colors[], PetscCopyMode mode, ISColoring *iscoloring)
346: {
347: PetscMPIInt size, rank, tag;
348: PetscInt base, top, i;
349: PetscInt nc, ncwork;
350: MPI_Status status;
352: PetscFunctionBegin;
353: if (ncolors != PETSC_DECIDE && ncolors > IS_COLORING_MAX) {
354: PetscCheck(ncolors <= PETSC_UINT16_MAX, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Max color value exceeds %d limit. This number is unrealistic. Perhaps a bug in code? Current max: %d user requested: %" PetscInt_FMT, PETSC_UINT16_MAX, PETSC_IS_COLORING_MAX, ncolors);
355: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Max color value exceeds limit. Perhaps reconfigure PETSc with --with-is-color-value-type=short? Current max: %d user requested: %" PetscInt_FMT, PETSC_IS_COLORING_MAX, ncolors);
356: }
357: PetscCall(PetscNew(iscoloring));
358: PetscCall(PetscCommDuplicate(comm, &(*iscoloring)->comm, &tag));
359: comm = (*iscoloring)->comm;
361: /* compute the number of the first node on my processor */
362: PetscCallMPI(MPI_Comm_size(comm, &size));
364: /* should use MPI_Scan() */
365: PetscCallMPI(MPI_Comm_rank(comm, &rank));
366: if (rank == 0) {
367: base = 0;
368: top = n;
369: } else {
370: PetscCallMPI(MPI_Recv(&base, 1, MPIU_INT, rank - 1, tag, comm, &status));
371: top = base + n;
372: }
373: if (rank < size - 1) PetscCallMPI(MPI_Send(&top, 1, MPIU_INT, rank + 1, tag, comm));
375: /* compute the total number of colors */
376: ncwork = 0;
377: for (i = 0; i < n; i++) {
378: if (ncwork < colors[i]) ncwork = colors[i];
379: }
380: ncwork++;
381: PetscCallMPI(MPIU_Allreduce(&ncwork, &nc, 1, MPIU_INT, MPI_MAX, comm));
382: PetscCheck(nc <= ncolors, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Number of colors passed in %" PetscInt_FMT " is less than the actual number of colors in array %" PetscInt_FMT, ncolors, nc);
383: (*iscoloring)->n = nc;
384: (*iscoloring)->is = NULL;
385: (*iscoloring)->N = n;
386: (*iscoloring)->refct = 1;
387: (*iscoloring)->ctype = IS_COLORING_GLOBAL;
388: if (mode == PETSC_COPY_VALUES) {
389: PetscCall(PetscMalloc1(n, &(*iscoloring)->colors));
390: PetscCall(PetscArraycpy((*iscoloring)->colors, colors, n));
391: (*iscoloring)->allocated = PETSC_TRUE;
392: } else if (mode == PETSC_OWN_POINTER) {
393: (*iscoloring)->colors = (ISColoringValue *)colors;
394: (*iscoloring)->allocated = PETSC_TRUE;
395: } else {
396: (*iscoloring)->colors = (ISColoringValue *)colors;
397: (*iscoloring)->allocated = PETSC_FALSE;
398: }
399: PetscCall(ISColoringViewFromOptions(*iscoloring, NULL, "-is_coloring_view"));
400: PetscCall(PetscInfo(0, "Number of colors %" PetscInt_FMT "\n", nc));
401: PetscFunctionReturn(PETSC_SUCCESS);
402: }
404: /*@
405: ISBuildTwoSided - Takes an `IS` that describes where each element will be mapped globally over all ranks.
406: Generates an `IS` that contains new numbers from remote or local on the `IS`.
408: Collective
410: Input Parameters:
411: + ito - an `IS` describes to which rank each entry will be mapped. Negative target rank will be ignored
412: - toindx - an `IS` describes what indices should send. `NULL` means sending natural numbering
414: Output Parameter:
415: . rows - contains new numbers from remote or local
417: Level: advanced
419: Developer Note:
420: This manual page is incomprehensible and still needs to be fixed
422: .seealso: [](sec_scatter), `IS`, `MatPartitioningCreate()`, `ISPartitioningToNumbering()`, `ISPartitioningCount()`
423: @*/
424: PetscErrorCode ISBuildTwoSided(IS ito, IS toindx, IS *rows)
425: {
426: const PetscInt *ito_indices, *toindx_indices;
427: PetscInt *send_indices, rstart, *recv_indices, nrecvs, nsends;
428: PetscInt *tosizes, *fromsizes, j, *tosizes_tmp, *tooffsets_tmp, ito_ln;
429: PetscMPIInt *toranks, *fromranks, size, target_rank, *fromperm_newtoold, nto, nfrom;
430: PetscLayout isrmap;
431: MPI_Comm comm;
432: PetscSF sf;
433: PetscSFNode *iremote;
435: PetscFunctionBegin;
436: PetscCall(PetscObjectGetComm((PetscObject)ito, &comm));
437: PetscCallMPI(MPI_Comm_size(comm, &size));
438: PetscCall(ISGetLocalSize(ito, &ito_ln));
439: PetscCall(ISGetLayout(ito, &isrmap));
440: PetscCall(PetscLayoutGetRange(isrmap, &rstart, NULL));
441: PetscCall(ISGetIndices(ito, &ito_indices));
442: PetscCall(PetscCalloc2(size, &tosizes_tmp, size + 1, &tooffsets_tmp));
443: for (PetscInt i = 0; i < ito_ln; i++) {
444: if (ito_indices[i] < 0) continue;
445: else PetscCheck(ito_indices[i] < size, comm, PETSC_ERR_ARG_OUTOFRANGE, "target rank %" PetscInt_FMT " is larger than communicator size %d ", ito_indices[i], size);
446: tosizes_tmp[ito_indices[i]]++;
447: }
448: nto = 0;
449: for (PetscMPIInt i = 0; i < size; i++) {
450: tooffsets_tmp[i + 1] = tooffsets_tmp[i] + tosizes_tmp[i];
451: if (tosizes_tmp[i] > 0) nto++;
452: }
453: PetscCall(PetscCalloc2(nto, &toranks, 2 * nto, &tosizes));
454: nto = 0;
455: for (PetscMPIInt i = 0; i < size; i++) {
456: if (tosizes_tmp[i] > 0) {
457: toranks[nto] = i;
458: tosizes[2 * nto] = tosizes_tmp[i]; /* size */
459: tosizes[2 * nto + 1] = tooffsets_tmp[i]; /* offset */
460: nto++;
461: }
462: }
463: nsends = tooffsets_tmp[size];
464: PetscCall(PetscCalloc1(nsends, &send_indices));
465: if (toindx) PetscCall(ISGetIndices(toindx, &toindx_indices));
466: for (PetscInt i = 0; i < ito_ln; i++) {
467: if (ito_indices[i] < 0) continue;
468: PetscCall(PetscMPIIntCast(ito_indices[i], &target_rank));
469: send_indices[tooffsets_tmp[target_rank]] = toindx ? toindx_indices[i] : (i + rstart);
470: tooffsets_tmp[target_rank]++;
471: }
472: if (toindx) PetscCall(ISRestoreIndices(toindx, &toindx_indices));
473: PetscCall(ISRestoreIndices(ito, &ito_indices));
474: PetscCall(PetscFree2(tosizes_tmp, tooffsets_tmp));
475: PetscCall(PetscCommBuildTwoSided(comm, 2, MPIU_INT, nto, toranks, tosizes, &nfrom, &fromranks, &fromsizes));
476: PetscCall(PetscFree2(toranks, tosizes));
477: PetscCall(PetscMalloc1(nfrom, &fromperm_newtoold));
478: for (PetscMPIInt i = 0; i < nfrom; i++) fromperm_newtoold[i] = i;
479: PetscCall(PetscSortMPIIntWithArray(nfrom, fromranks, fromperm_newtoold));
480: nrecvs = 0;
481: for (PetscMPIInt i = 0; i < nfrom; i++) nrecvs += fromsizes[i * 2];
482: PetscCall(PetscCalloc1(nrecvs, &recv_indices));
483: PetscCall(PetscMalloc1(nrecvs, &iremote));
484: nrecvs = 0;
485: for (PetscMPIInt i = 0; i < nfrom; i++) {
486: for (j = 0; j < fromsizes[2 * fromperm_newtoold[i]]; j++) {
487: iremote[nrecvs].rank = fromranks[i];
488: iremote[nrecvs++].index = fromsizes[2 * fromperm_newtoold[i] + 1] + j;
489: }
490: }
491: PetscCall(PetscSFCreate(comm, &sf));
492: PetscCall(PetscSFSetGraph(sf, nsends, nrecvs, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
493: PetscCall(PetscSFSetType(sf, PETSCSFBASIC));
494: /* how to put a prefix ? */
495: PetscCall(PetscSFSetFromOptions(sf));
496: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, send_indices, recv_indices, MPI_REPLACE));
497: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, send_indices, recv_indices, MPI_REPLACE));
498: PetscCall(PetscSFDestroy(&sf));
499: PetscCall(PetscFree(fromranks));
500: PetscCall(PetscFree(fromsizes));
501: PetscCall(PetscFree(fromperm_newtoold));
502: PetscCall(PetscFree(send_indices));
503: if (rows) {
504: PetscCall(PetscSortInt(nrecvs, recv_indices));
505: PetscCall(ISCreateGeneral(comm, nrecvs, recv_indices, PETSC_OWN_POINTER, rows));
506: } else {
507: PetscCall(PetscFree(recv_indices));
508: }
509: PetscFunctionReturn(PETSC_SUCCESS);
510: }
512: /*@
513: ISPartitioningToNumbering - Takes an `IS' that represents a partitioning (the MPI rank that each local entry belongs to) and on each MPI process
514: generates an `IS` that contains a new global node number in the new ordering for each entry
516: Collective
518: Input Parameter:
519: . part - a partitioning as generated by `MatPartitioningApply()` or `MatPartitioningApplyND()`
521: Output Parameter:
522: . is - on each processor the index set that defines the global numbers
523: (in the new numbering) for all the nodes currently (before the partitioning)
524: on that processor
526: Level: advanced
528: Note:
529: The resulting `IS` tells where each local entry is mapped to in a new global ordering
531: .seealso: [](sec_scatter), `IS`, `MatPartitioningCreate()`, `AOCreateBasic()`, `ISPartitioningCount()`
532: @*/
533: PetscErrorCode ISPartitioningToNumbering(IS part, IS *is)
534: {
535: MPI_Comm comm;
536: IS ndorder;
537: PetscInt n, *starts = NULL, *sums = NULL, *lsizes = NULL, *newi = NULL;
538: const PetscInt *indices = NULL;
539: PetscMPIInt np, npt;
541: PetscFunctionBegin;
543: PetscAssertPointer(is, 2);
544: /* see if the partitioning comes from nested dissection */
545: PetscCall(PetscObjectQuery((PetscObject)part, "_petsc_matpartitioning_ndorder", (PetscObject *)&ndorder));
546: if (ndorder) {
547: PetscCall(PetscObjectReference((PetscObject)ndorder));
548: *is = ndorder;
549: PetscFunctionReturn(PETSC_SUCCESS);
550: }
552: PetscCall(PetscObjectGetComm((PetscObject)part, &comm));
553: /* count the number of partitions, i.e., virtual processors */
554: PetscCall(ISGetLocalSize(part, &n));
555: PetscCall(ISGetIndices(part, &indices));
556: np = 0;
557: for (PetscInt i = 0; i < n; i++) PetscCall(PetscMPIIntCast(PetscMax(np, indices[i]), &np));
558: PetscCallMPI(MPIU_Allreduce(&np, &npt, 1, MPI_INT, MPI_MAX, comm));
559: np = npt + 1; /* so that it looks like a MPI_Comm_size output */
561: /*
562: lsizes - number of elements of each partition on this particular processor
563: sums - total number of "previous" nodes for any particular partition
564: starts - global number of first element in each partition on this processor
565: */
566: PetscCall(PetscMalloc3(np, &lsizes, np, &starts, np, &sums));
567: PetscCall(PetscArrayzero(lsizes, np));
568: for (PetscInt i = 0; i < n; i++) lsizes[indices[i]]++;
569: PetscCallMPI(MPIU_Allreduce(lsizes, sums, np, MPIU_INT, MPI_SUM, comm));
570: PetscCallMPI(MPI_Scan(lsizes, starts, np, MPIU_INT, MPI_SUM, comm));
571: for (PetscMPIInt i = 0; i < np; i++) starts[i] -= lsizes[i];
572: for (PetscMPIInt i = 1; i < np; i++) {
573: sums[i] += sums[i - 1];
574: starts[i] += sums[i - 1];
575: }
577: /*
578: For each local index give it the new global number
579: */
580: PetscCall(PetscMalloc1(n, &newi));
581: for (PetscInt i = 0; i < n; i++) newi[i] = starts[indices[i]]++;
582: PetscCall(PetscFree3(lsizes, starts, sums));
584: PetscCall(ISRestoreIndices(part, &indices));
585: PetscCall(ISCreateGeneral(comm, n, newi, PETSC_OWN_POINTER, is));
586: PetscCall(ISSetPermutation(*is));
587: PetscFunctionReturn(PETSC_SUCCESS);
588: }
590: /*@
591: ISPartitioningCount - Takes a `IS` that represents a partitioning (the MPI rank that each local entry belongs to) and determines the number of
592: resulting elements on each (partition) rank
594: Collective
596: Input Parameters:
597: + part - a partitioning as generated by `MatPartitioningApply()` or `MatPartitioningApplyND()`
598: - len - length of the array count, this is the total number of partitions
600: Output Parameter:
601: . count - array of length size, to contain the number of elements assigned
602: to each partition, where size is the number of partitions generated
603: (see notes below).
605: Level: advanced
607: Notes:
608: By default the number of partitions generated (and thus the length
609: of count) is the size of the communicator associated with `IS`,
610: but it can be set by `MatPartitioningSetNParts()`.
612: The resulting array of lengths can for instance serve as input of `PCBJacobiSetTotalBlocks()`.
614: If the partitioning has been obtained by `MatPartitioningApplyND()`, the returned count does not include the separators.
616: .seealso: [](sec_scatter), `IS`, `MatPartitioningCreate()`, `AOCreateBasic()`, `ISPartitioningToNumbering()`,
617: `MatPartitioningSetNParts()`, `MatPartitioningApply()`, `MatPartitioningApplyND()`
618: @*/
619: PetscErrorCode ISPartitioningCount(IS part, PetscInt len, PetscInt count[])
620: {
621: MPI_Comm comm;
622: PetscInt i, n, *lsizes;
623: const PetscInt *indices;
625: PetscFunctionBegin;
626: PetscCall(PetscObjectGetComm((PetscObject)part, &comm));
627: if (len == PETSC_DEFAULT) {
628: PetscMPIInt size;
630: PetscCallMPI(MPI_Comm_size(comm, &size));
631: len = size;
632: }
634: /* count the number of partitions */
635: PetscCall(ISGetLocalSize(part, &n));
636: PetscCall(ISGetIndices(part, &indices));
637: if (PetscDefined(USE_DEBUG)) {
638: PetscInt np = 0, npt;
639: for (i = 0; i < n; i++) np = PetscMax(np, indices[i]);
640: PetscCallMPI(MPIU_Allreduce(&np, &npt, 1, MPIU_INT, MPI_MAX, comm));
641: np = npt + 1; /* so that it looks like a MPI_Comm_size output */
642: PetscCheck(np <= len, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Length of count array %" PetscInt_FMT " is less than number of partitions %" PetscInt_FMT, len, np);
643: }
645: /*
646: lsizes - number of elements of each partition on this particular processor
647: sums - total number of "previous" nodes for any particular partition
648: starts - global number of first element in each partition on this processor
649: */
650: PetscCall(PetscCalloc1(len, &lsizes));
651: for (i = 0; i < n; i++) {
652: if (indices[i] > -1) lsizes[indices[i]]++;
653: }
654: PetscCall(ISRestoreIndices(part, &indices));
655: PetscCallMPI(MPIU_Allreduce(lsizes, count, len, MPIU_INT, MPI_SUM, comm));
656: PetscCall(PetscFree(lsizes));
657: PetscFunctionReturn(PETSC_SUCCESS);
658: }
660: /*@
661: ISAllGather - Given an index set `IS` on each processor, generates a large
662: index set (same on each processor) by concatenating together each
663: processors index set.
665: Collective
667: Input Parameter:
668: . is - the distributed index set
670: Output Parameter:
671: . isout - the concatenated index set (same on all processors)
673: Level: intermediate
675: Notes:
676: `ISAllGather()` is clearly not scalable for large index sets.
678: The `IS` created on each processor must be created with a common
679: communicator (e.g., `PETSC_COMM_WORLD`). If the index sets were created
680: with `PETSC_COMM_SELF`, this routine will not work as expected, since
681: each process will generate its own new `IS` that consists only of
682: itself.
684: The communicator for this new `IS` is `PETSC_COMM_SELF`
686: .seealso: [](sec_scatter), `IS`, `ISCreateGeneral()`, `ISCreateStride()`, `ISCreateBlock()`
687: @*/
688: PetscErrorCode ISAllGather(IS is, IS *isout)
689: {
690: PetscInt *indices, n, i, N, step, first;
691: const PetscInt *lindices;
692: MPI_Comm comm;
693: PetscMPIInt size, *sizes = NULL, *offsets = NULL, nn;
694: PetscBool stride;
696: PetscFunctionBegin;
698: PetscAssertPointer(isout, 2);
700: PetscCall(PetscObjectGetComm((PetscObject)is, &comm));
701: PetscCallMPI(MPI_Comm_size(comm, &size));
702: PetscCall(ISGetLocalSize(is, &n));
703: PetscCall(PetscObjectTypeCompare((PetscObject)is, ISSTRIDE, &stride));
704: if (size == 1 && stride) { /* should handle parallel ISStride also */
705: PetscCall(ISStrideGetInfo(is, &first, &step));
706: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, first, step, isout));
707: } else {
708: PetscCall(PetscMalloc2(size, &sizes, size, &offsets));
710: PetscCall(PetscMPIIntCast(n, &nn));
711: PetscCallMPI(MPI_Allgather(&nn, 1, MPI_INT, sizes, 1, MPI_INT, comm));
712: offsets[0] = 0;
713: for (i = 1; i < size; i++) {
714: PetscInt s = offsets[i - 1] + sizes[i - 1];
715: PetscCall(PetscMPIIntCast(s, &offsets[i]));
716: }
717: N = offsets[size - 1] + sizes[size - 1];
719: PetscCall(PetscMalloc1(N, &indices));
720: PetscCall(ISGetIndices(is, &lindices));
721: PetscCallMPI(MPI_Allgatherv((void *)lindices, nn, MPIU_INT, indices, sizes, offsets, MPIU_INT, comm));
722: PetscCall(ISRestoreIndices(is, &lindices));
723: PetscCall(PetscFree2(sizes, offsets));
725: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, N, indices, PETSC_OWN_POINTER, isout));
726: }
727: PetscFunctionReturn(PETSC_SUCCESS);
728: }
730: /*@C
731: ISAllGatherColors - Given a set of colors on each processor, generates a large
732: set (same on each processor) by concatenating together each processors colors
734: Collective
736: Input Parameters:
737: + comm - communicator to share the indices
738: . n - local size of set
739: - lindices - local colors
741: Output Parameters:
742: + outN - total number of indices
743: - outindices - all of the colors
745: Level: intermediate
747: Note:
748: `ISAllGatherColors()` is clearly not scalable for large index sets.
750: .seealso: `ISColoringValue`, `ISColoring()`, `ISCreateGeneral()`, `ISCreateStride()`, `ISCreateBlock()`, `ISAllGather()`
751: @*/
752: PetscErrorCode ISAllGatherColors(MPI_Comm comm, PetscInt n, ISColoringValue lindices[], PetscInt *outN, ISColoringValue *outindices[])
753: {
754: ISColoringValue *indices;
755: PetscInt N;
756: PetscMPIInt size, *offsets = NULL, *sizes = NULL, nn;
758: PetscFunctionBegin;
759: PetscCall(PetscMPIIntCast(n, &nn));
760: PetscCallMPI(MPI_Comm_size(comm, &size));
761: PetscCall(PetscMalloc2(size, &sizes, size, &offsets));
763: PetscCallMPI(MPI_Allgather(&nn, 1, MPI_INT, sizes, 1, MPI_INT, comm));
764: offsets[0] = 0;
765: for (PetscMPIInt i = 1; i < size; i++) offsets[i] = offsets[i - 1] + sizes[i - 1];
766: N = offsets[size - 1] + sizes[size - 1];
767: PetscCall(PetscFree2(sizes, offsets));
769: PetscCall(PetscMalloc1(N + 1, &indices));
770: PetscCallMPI(MPI_Allgatherv(lindices, nn, MPIU_COLORING_VALUE, indices, sizes, offsets, MPIU_COLORING_VALUE, comm));
772: *outindices = indices;
773: if (outN) *outN = N;
774: PetscFunctionReturn(PETSC_SUCCESS);
775: }
777: /*@
778: ISComplement - Given an index set `IS` generates the complement index set. That is
779: all indices that are NOT in the given set.
781: Collective
783: Input Parameters:
784: + is - the index set
785: . nmin - the first index desired in the local part of the complement
786: - nmax - the largest index desired in the local part of the complement (note that all indices in `is` must be greater or equal to `nmin` and less than `nmax`)
788: Output Parameter:
789: . isout - the complement
791: Level: intermediate
793: Notes:
794: The communicator for `isout` is the same as for the input `is`
796: For a parallel `is`, this will generate the local part of the complement on each process
798: To generate the entire complement (on each process) of a parallel `is`, first call `ISAllGather()` and then
799: call this routine.
801: .seealso: [](sec_scatter), `IS`, `ISCreateGeneral()`, `ISCreateStride()`, `ISCreateBlock()`, `ISAllGather()`
802: @*/
803: PetscErrorCode ISComplement(IS is, PetscInt nmin, PetscInt nmax, IS *isout)
804: {
805: const PetscInt *indices;
806: PetscInt n, i, j, unique, cnt, *nindices;
807: PetscBool sorted;
809: PetscFunctionBegin;
811: PetscAssertPointer(isout, 4);
812: PetscCheck(nmin >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nmin %" PetscInt_FMT " cannot be negative", nmin);
813: PetscCheck(nmin <= nmax, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nmin %" PetscInt_FMT " cannot be greater than nmax %" PetscInt_FMT, nmin, nmax);
814: PetscCall(ISSorted(is, &sorted));
815: PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Index set must be sorted");
817: PetscCall(ISGetLocalSize(is, &n));
818: PetscCall(ISGetIndices(is, &indices));
819: if (PetscDefined(USE_DEBUG)) {
820: for (i = 0; i < n; i++) {
821: PetscCheck(indices[i] >= nmin, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Index %" PetscInt_FMT "'s value %" PetscInt_FMT " is smaller than minimum given %" PetscInt_FMT, i, indices[i], nmin);
822: PetscCheck(indices[i] < nmax, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Index %" PetscInt_FMT "'s value %" PetscInt_FMT " is larger than maximum given %" PetscInt_FMT, i, indices[i], nmax);
823: }
824: }
825: /* Count number of unique entries */
826: unique = (n > 0);
827: for (i = 0; i < n - 1; i++) {
828: if (indices[i + 1] != indices[i]) unique++;
829: }
830: PetscCall(PetscMalloc1(nmax - nmin - unique, &nindices));
831: cnt = 0;
832: for (i = nmin, j = 0; i < nmax; i++) {
833: if (j < n && i == indices[j]) do {
834: j++;
835: } while (j < n && i == indices[j]);
836: else nindices[cnt++] = i;
837: }
838: PetscCheck(cnt == nmax - nmin - unique, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Number of entries found in complement %" PetscInt_FMT " does not match expected %" PetscInt_FMT, cnt, nmax - nmin - unique);
839: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)is), cnt, nindices, PETSC_OWN_POINTER, isout));
840: PetscCall(ISSetInfo(*isout, IS_SORTED, IS_GLOBAL, PETSC_FALSE, PETSC_TRUE));
841: PetscCall(ISRestoreIndices(is, &indices));
842: PetscFunctionReturn(PETSC_SUCCESS);
843: }