Actual source code: classical.c
1: #include <../src/ksp/pc/impls/gamg/gamg.h>
2: #include <petscsf.h>
4: static PetscFunctionList PCGAMGClassicalProlongatorList = NULL;
5: static PetscBool PCGAMGClassicalPackageInitialized = PETSC_FALSE;
7: typedef struct {
8: PetscReal interp_threshold; /* interpolation threshold */
9: char prolongtype[256];
10: PetscInt nsmooths; /* number of jacobi smoothings on the prolongator */
11: } PC_GAMG_Classical;
13: /*@
14: PCGAMGClassicalSetType - Sets the type of classical interpolation to use with `PCGAMG`
16: Collective
18: Input Parameters:
19: + pc - the preconditioner context
20: - type - the interpolation to use, see `PCGAMGClassicalType()`
22: Options Database Key:
23: . -pc_gamg_classical_type (direct|standard) - set type of classical AMG prolongation
25: Level: intermediate
27: .seealso: [](ch_ksp), `PCGAMG`, `PCGAMGClassicalType`, `PCGAMGClassicalGetType()`
28: @*/
29: PetscErrorCode PCGAMGClassicalSetType(PC pc, PCGAMGClassicalType type)
30: {
31: PetscFunctionBegin;
33: PetscTryMethod(pc, "PCGAMGClassicalSetType_C", (PC, PCGAMGClassicalType), (pc, type));
34: PetscFunctionReturn(PETSC_SUCCESS);
35: }
37: /*@
38: PCGAMGClassicalGetType - Gets the type of classical interpolation to use with `PCGAMG`
40: Collective
42: Input Parameter:
43: . pc - the preconditioner context
45: Output Parameter:
46: . type - the type used, see `PCGAMGClassicalType()`
48: Level: intermediate
50: .seealso: [](ch_ksp), `PCGAMG`, `PCGAMGClassicalType`, `PCGAMGClassicalSetType()`
51: @*/
52: PetscErrorCode PCGAMGClassicalGetType(PC pc, PCGAMGClassicalType *type)
53: {
54: PetscFunctionBegin;
56: PetscUseMethod(pc, "PCGAMGClassicalGetType_C", (PC, PCGAMGClassicalType *), (pc, type));
57: PetscFunctionReturn(PETSC_SUCCESS);
58: }
60: static PetscErrorCode PCGAMGClassicalSetType_GAMG(PC pc, PCGAMGClassicalType type)
61: {
62: PC_MG *mg = (PC_MG *)pc->data;
63: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
64: PC_GAMG_Classical *cls = (PC_GAMG_Classical *)pc_gamg->subctx;
66: PetscFunctionBegin;
67: PetscCall(PetscStrncpy(cls->prolongtype, type, sizeof(cls->prolongtype)));
68: PetscFunctionReturn(PETSC_SUCCESS);
69: }
71: static PetscErrorCode PCGAMGClassicalGetType_GAMG(PC pc, PCGAMGClassicalType *type)
72: {
73: PC_MG *mg = (PC_MG *)pc->data;
74: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
75: PC_GAMG_Classical *cls = (PC_GAMG_Classical *)pc_gamg->subctx;
77: PetscFunctionBegin;
78: *type = cls->prolongtype;
79: PetscFunctionReturn(PETSC_SUCCESS);
80: }
82: static PetscErrorCode PCGAMGCreateGraph_Classical(PC pc, Mat A, Mat *G)
83: {
84: PetscInt s, f, n, idx, lidx, gidx;
85: PetscInt r, c, ncols;
86: const PetscInt *rcol;
87: const PetscScalar *rval;
88: PetscInt *gcol;
89: PetscScalar *gval;
90: PetscReal rmax;
91: PetscInt cmax = 0;
92: PC_MG *mg = (PC_MG *)pc->data;
93: PC_GAMG *gamg = (PC_GAMG *)mg->innerctx;
94: PetscInt *gsparse, *lsparse;
95: PetscScalar *Amax;
96: MatType mtype;
98: PetscFunctionBegin;
99: PetscCall(MatGetOwnershipRange(A, &s, &f));
100: n = f - s;
101: PetscCall(PetscMalloc3(n, &lsparse, n, &gsparse, n, &Amax));
103: for (r = 0; r < n; r++) {
104: lsparse[r] = 0;
105: gsparse[r] = 0;
106: }
108: for (r = s; r < f; r++) {
109: /* determine the maximum off-diagonal in each row */
110: rmax = 0.;
111: PetscCall(MatGetRow(A, r, &ncols, &rcol, &rval));
112: for (c = 0; c < ncols; c++) {
113: if (PetscRealPart(-rval[c]) > rmax && rcol[c] != r) rmax = PetscRealPart(-rval[c]);
114: }
115: Amax[r - s] = rmax;
116: if (ncols > cmax) cmax = ncols;
117: lidx = 0;
118: gidx = 0;
119: /* create the local and global sparsity patterns */
120: for (c = 0; c < ncols; c++) {
121: if (PetscRealPart(-rval[c]) > gamg->threshold[0] * PetscRealPart(Amax[r - s]) || rcol[c] == r) {
122: if (rcol[c] < f && rcol[c] >= s) {
123: lidx++;
124: } else {
125: gidx++;
126: }
127: }
128: }
129: PetscCall(MatRestoreRow(A, r, &ncols, &rcol, &rval));
130: lsparse[r - s] = lidx;
131: gsparse[r - s] = gidx;
132: }
133: PetscCall(PetscMalloc2(cmax, &gval, cmax, &gcol));
135: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), G));
136: PetscCall(MatGetType(A, &mtype));
137: PetscCall(MatSetType(*G, mtype));
138: PetscCall(MatSetSizes(*G, n, n, PETSC_DETERMINE, PETSC_DETERMINE));
139: PetscCall(MatMPIAIJSetPreallocation(*G, 0, lsparse, 0, gsparse));
140: PetscCall(MatSeqAIJSetPreallocation(*G, 0, lsparse));
141: for (r = s; r < f; r++) {
142: PetscCall(MatGetRow(A, r, &ncols, &rcol, &rval));
143: idx = 0;
144: for (c = 0; c < ncols; c++) {
145: /* classical strength of connection */
146: if (PetscRealPart(-rval[c]) > gamg->threshold[0] * PetscRealPart(Amax[r - s]) || rcol[c] == r) {
147: gcol[idx] = rcol[c];
148: gval[idx] = rval[c];
149: idx++;
150: }
151: }
152: PetscCall(MatSetValues(*G, 1, &r, idx, gcol, gval, INSERT_VALUES));
153: PetscCall(MatRestoreRow(A, r, &ncols, &rcol, &rval));
154: }
155: PetscCall(MatAssemblyBegin(*G, MAT_FINAL_ASSEMBLY));
156: PetscCall(MatAssemblyEnd(*G, MAT_FINAL_ASSEMBLY));
158: PetscCall(PetscFree2(gval, gcol));
159: PetscCall(PetscFree3(lsparse, gsparse, Amax));
160: PetscFunctionReturn(PETSC_SUCCESS);
161: }
163: static PetscErrorCode PCGAMGCoarsen_Classical(PC pc, Mat *G, PetscCoarsenData **agg_lists)
164: {
165: MatCoarsen crs;
166: MPI_Comm fcomm = ((PetscObject)pc)->comm;
167: const char *prefix;
169: PetscFunctionBegin;
170: PetscCheck(G, fcomm, PETSC_ERR_ARG_WRONGSTATE, "Must set Graph in PC in PCGAMG before coarsening");
172: PetscCall(MatCoarsenCreate(fcomm, &crs));
173: PetscCall(PetscObjectGetOptionsPrefix((PetscObject)pc, &prefix));
174: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)crs, prefix));
175: PetscCall(PetscObjectAppendOptionsPrefix((PetscObject)crs, "pc_gamg_"));
176: PetscCall(MatCoarsenSetFromOptions(crs));
177: PetscCall(MatCoarsenSetAdjacency(crs, *G));
178: PetscCall(MatCoarsenSetStrictAggs(crs, PETSC_TRUE));
179: PetscCall(MatCoarsenApply(crs));
180: PetscCall(MatCoarsenGetData(crs, agg_lists));
181: PetscCall(MatCoarsenDestroy(&crs));
182: PetscFunctionReturn(PETSC_SUCCESS);
183: }
185: static PetscErrorCode PCGAMGProlongator_Classical_Direct(PC pc, Mat A, PetscCoarsenData *agg_lists, Mat *P)
186: {
187: PC_MG *mg = (PC_MG *)pc->data;
188: PC_GAMG *gamg = (PC_GAMG *)mg->innerctx;
189: PetscBool iscoarse, isMPIAIJ, isSEQAIJ;
190: PetscInt fn, cn, fs, fe, cs, ce, i, j, ncols, col, row_f, row_c, cmax = 0, idx, noff;
191: PetscInt *lcid, *gcid, *lsparse, *gsparse, *pcols;
192: const PetscInt *rcol;
193: PetscReal *Amax_pos, *Amax_neg;
194: PetscScalar g_pos, g_neg, a_pos, a_neg, diag, invdiag, alpha, beta, pij;
195: PetscScalar *pvals;
196: const PetscScalar *rval;
197: Mat lA, gA = NULL;
198: MatType mtype;
199: Vec C, lvec;
200: PetscSF sf;
201: Mat_MPIAIJ *mpiaij;
203: PetscFunctionBegin;
204: PetscCall(MatGetOwnershipRange(A, &fs, &fe));
205: fn = fe - fs;
206: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &isMPIAIJ));
207: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJ, &isSEQAIJ));
208: PetscCheck(isMPIAIJ || isSEQAIJ, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Classical AMG requires MPIAIJ matrix");
209: if (isMPIAIJ) {
210: mpiaij = (Mat_MPIAIJ *)A->data;
211: lA = mpiaij->A;
212: gA = mpiaij->B;
213: lvec = mpiaij->lvec;
214: PetscCall(VecGetSize(lvec, &noff));
215: PetscCall(MatGetMultPetscSF(A, &sf));
216: PetscCall(PetscMalloc1(noff, &gcid));
217: } else {
218: lA = A;
219: }
220: PetscCall(PetscMalloc5(fn, &lsparse, fn, &gsparse, fn, &lcid, fn, &Amax_pos, fn, &Amax_neg));
222: /* count the number of coarse unknowns */
223: cn = 0;
224: for (i = 0; i < fn; i++) {
225: /* filter out singletons */
226: PetscCall(PetscCDIsEmptyAt(agg_lists, i, &iscoarse));
227: lcid[i] = -1;
228: if (!iscoarse) cn++;
229: }
231: /* create the coarse vector */
232: PetscCall(VecCreateMPI(PetscObjectComm((PetscObject)A), cn, PETSC_DECIDE, &C));
233: PetscCall(VecGetOwnershipRange(C, &cs, &ce));
235: cn = 0;
236: for (i = 0; i < fn; i++) {
237: PetscCall(PetscCDIsEmptyAt(agg_lists, i, &iscoarse));
238: if (!iscoarse) {
239: lcid[i] = cs + cn;
240: cn++;
241: } else {
242: lcid[i] = -1;
243: }
244: }
246: if (gA) {
247: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, lcid, gcid, MPI_REPLACE));
248: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, lcid, gcid, MPI_REPLACE));
249: }
251: /* determine the largest off-diagonal entries in each row */
252: for (i = fs; i < fe; i++) {
253: Amax_pos[i - fs] = 0.;
254: Amax_neg[i - fs] = 0.;
255: PetscCall(MatGetRow(A, i, &ncols, &rcol, &rval));
256: for (j = 0; j < ncols; j++) {
257: if ((PetscRealPart(-rval[j]) > Amax_neg[i - fs]) && i != rcol[j]) Amax_neg[i - fs] = PetscAbsScalar(rval[j]);
258: if ((PetscRealPart(rval[j]) > Amax_pos[i - fs]) && i != rcol[j]) Amax_pos[i - fs] = PetscAbsScalar(rval[j]);
259: }
260: if (ncols > cmax) cmax = ncols;
261: PetscCall(MatRestoreRow(A, i, &ncols, &rcol, &rval));
262: }
263: PetscCall(PetscMalloc2(cmax, &pcols, cmax, &pvals));
264: PetscCall(VecDestroy(&C));
266: /* count the on and off processor sparsity patterns for the prolongator */
267: for (i = 0; i < fn; i++) {
268: /* on */
269: lsparse[i] = 0;
270: gsparse[i] = 0;
271: if (lcid[i] >= 0) {
272: lsparse[i] = 1;
273: gsparse[i] = 0;
274: } else {
275: PetscCall(MatGetRow(lA, i, &ncols, &rcol, &rval));
276: for (j = 0; j < ncols; j++) {
277: col = rcol[j];
278: if (lcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold[0] * Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold[0] * Amax_neg[i])) lsparse[i] += 1;
279: }
280: PetscCall(MatRestoreRow(lA, i, &ncols, &rcol, &rval));
281: /* off */
282: if (gA) {
283: PetscCall(MatGetRow(gA, i, &ncols, &rcol, &rval));
284: for (j = 0; j < ncols; j++) {
285: col = rcol[j];
286: if (gcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold[0] * Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold[0] * Amax_neg[i])) gsparse[i] += 1;
287: }
288: PetscCall(MatRestoreRow(gA, i, &ncols, &rcol, &rval));
289: }
290: }
291: }
293: /* preallocate and create the prolongator */
294: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), P));
295: PetscCall(MatGetType(A, &mtype));
296: PetscCall(MatSetType(*P, mtype));
297: PetscCall(MatSetSizes(*P, fn, cn, PETSC_DETERMINE, PETSC_DETERMINE));
298: PetscCall(MatMPIAIJSetPreallocation(*P, 0, lsparse, 0, gsparse));
299: PetscCall(MatSeqAIJSetPreallocation(*P, 0, lsparse));
301: /* loop over local fine nodes -- get the diagonal, the sum of positive and negative strong and weak weights, and set up the row */
302: for (i = 0; i < fn; i++) {
303: /* determine on or off */
304: row_f = i + fs;
305: row_c = lcid[i];
306: if (row_c >= 0) {
307: pij = 1.;
308: PetscCall(MatSetValues(*P, 1, &row_f, 1, &row_c, &pij, INSERT_VALUES));
309: } else {
310: g_pos = 0.;
311: g_neg = 0.;
312: a_pos = 0.;
313: a_neg = 0.;
314: diag = 0.;
316: /* local connections */
317: PetscCall(MatGetRow(lA, i, &ncols, &rcol, &rval));
318: for (j = 0; j < ncols; j++) {
319: col = rcol[j];
320: if (lcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold[0] * Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold[0] * Amax_neg[i])) {
321: if (PetscRealPart(rval[j]) > 0.) {
322: g_pos += rval[j];
323: } else {
324: g_neg += rval[j];
325: }
326: }
327: if (col != i) {
328: if (PetscRealPart(rval[j]) > 0.) {
329: a_pos += rval[j];
330: } else {
331: a_neg += rval[j];
332: }
333: } else {
334: diag = rval[j];
335: }
336: }
337: PetscCall(MatRestoreRow(lA, i, &ncols, &rcol, &rval));
339: /* ghosted connections */
340: if (gA) {
341: PetscCall(MatGetRow(gA, i, &ncols, &rcol, &rval));
342: for (j = 0; j < ncols; j++) {
343: col = rcol[j];
344: if (gcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold[0] * Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold[0] * Amax_neg[i])) {
345: if (PetscRealPart(rval[j]) > 0.) {
346: g_pos += rval[j];
347: } else {
348: g_neg += rval[j];
349: }
350: }
351: if (PetscRealPart(rval[j]) > 0.) {
352: a_pos += rval[j];
353: } else {
354: a_neg += rval[j];
355: }
356: }
357: PetscCall(MatRestoreRow(gA, i, &ncols, &rcol, &rval));
358: }
360: if (g_neg == 0.) {
361: alpha = 0.;
362: } else {
363: alpha = -a_neg / g_neg;
364: }
366: if (g_pos == 0.) {
367: diag += a_pos;
368: beta = 0.;
369: } else {
370: beta = -a_pos / g_pos;
371: }
372: if (diag == 0.) {
373: invdiag = 0.;
374: } else invdiag = 1. / diag;
375: /* on */
376: PetscCall(MatGetRow(lA, i, &ncols, &rcol, &rval));
377: idx = 0;
378: for (j = 0; j < ncols; j++) {
379: col = rcol[j];
380: if (lcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold[0] * Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold[0] * Amax_neg[i])) {
381: row_f = i + fs;
382: row_c = lcid[col];
383: /* set the values for on-processor ones */
384: if (PetscRealPart(rval[j]) < 0.) {
385: pij = rval[j] * alpha * invdiag;
386: } else {
387: pij = rval[j] * beta * invdiag;
388: }
389: if (PetscAbsScalar(pij) != 0.) {
390: pvals[idx] = pij;
391: pcols[idx] = row_c;
392: idx++;
393: }
394: }
395: }
396: PetscCall(MatRestoreRow(lA, i, &ncols, &rcol, &rval));
397: /* off */
398: if (gA) {
399: PetscCall(MatGetRow(gA, i, &ncols, &rcol, &rval));
400: for (j = 0; j < ncols; j++) {
401: col = rcol[j];
402: if (gcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold[0] * Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold[0] * Amax_neg[i])) {
403: row_f = i + fs;
404: row_c = gcid[col];
405: /* set the values for on-processor ones */
406: if (PetscRealPart(rval[j]) < 0.) {
407: pij = rval[j] * alpha * invdiag;
408: } else {
409: pij = rval[j] * beta * invdiag;
410: }
411: if (PetscAbsScalar(pij) != 0.) {
412: pvals[idx] = pij;
413: pcols[idx] = row_c;
414: idx++;
415: }
416: }
417: }
418: PetscCall(MatRestoreRow(gA, i, &ncols, &rcol, &rval));
419: }
420: PetscCall(MatSetValues(*P, 1, &row_f, idx, pcols, pvals, INSERT_VALUES));
421: }
422: }
424: PetscCall(MatAssemblyBegin(*P, MAT_FINAL_ASSEMBLY));
425: PetscCall(MatAssemblyEnd(*P, MAT_FINAL_ASSEMBLY));
427: PetscCall(PetscFree5(lsparse, gsparse, lcid, Amax_pos, Amax_neg));
429: PetscCall(PetscFree2(pcols, pvals));
430: if (gA) PetscCall(PetscFree(gcid));
431: PetscFunctionReturn(PETSC_SUCCESS);
432: }
434: static PetscErrorCode PCGAMGTruncateProlongator_Private(PC pc, Mat *P)
435: {
436: PetscInt j, i, ps, pf, pn, pcs, pcf, pcn, idx, cmax;
437: const PetscScalar *pval;
438: const PetscInt *pcol;
439: PetscScalar *pnval;
440: PetscInt *pncol;
441: PetscInt ncols;
442: Mat Pnew;
443: PetscInt *lsparse, *gsparse;
444: PetscReal pmax_pos, pmax_neg, ptot_pos, ptot_neg, pthresh_pos, pthresh_neg;
445: PC_MG *mg = (PC_MG *)pc->data;
446: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
447: PC_GAMG_Classical *cls = (PC_GAMG_Classical *)pc_gamg->subctx;
448: MatType mtype;
450: PetscFunctionBegin;
451: /* trim and rescale with reallocation */
452: PetscCall(MatGetOwnershipRange(*P, &ps, &pf));
453: PetscCall(MatGetOwnershipRangeColumn(*P, &pcs, &pcf));
454: pn = pf - ps;
455: pcn = pcf - pcs;
456: PetscCall(PetscMalloc2(pn, &lsparse, pn, &gsparse));
457: /* allocate */
458: cmax = 0;
459: for (i = ps; i < pf; i++) {
460: lsparse[i - ps] = 0;
461: gsparse[i - ps] = 0;
462: PetscCall(MatGetRow(*P, i, &ncols, &pcol, &pval));
463: if (ncols > cmax) cmax = ncols;
464: pmax_pos = 0.;
465: pmax_neg = 0.;
466: for (j = 0; j < ncols; j++) {
467: if (PetscRealPart(pval[j]) > pmax_pos) {
468: pmax_pos = PetscRealPart(pval[j]);
469: } else if (PetscRealPart(pval[j]) < pmax_neg) {
470: pmax_neg = PetscRealPart(pval[j]);
471: }
472: }
473: for (j = 0; j < ncols; j++) {
474: if (PetscRealPart(pval[j]) >= pmax_pos * cls->interp_threshold || PetscRealPart(pval[j]) <= pmax_neg * cls->interp_threshold) {
475: if (pcol[j] >= pcs && pcol[j] < pcf) {
476: lsparse[i - ps]++;
477: } else {
478: gsparse[i - ps]++;
479: }
480: }
481: }
482: PetscCall(MatRestoreRow(*P, i, &ncols, &pcol, &pval));
483: }
485: PetscCall(PetscMalloc2(cmax, &pnval, cmax, &pncol));
487: PetscCall(MatGetType(*P, &mtype));
488: PetscCall(MatCreate(PetscObjectComm((PetscObject)*P), &Pnew));
489: PetscCall(MatSetType(Pnew, mtype));
490: PetscCall(MatSetSizes(Pnew, pn, pcn, PETSC_DETERMINE, PETSC_DETERMINE));
491: PetscCall(MatSeqAIJSetPreallocation(Pnew, 0, lsparse));
492: PetscCall(MatMPIAIJSetPreallocation(Pnew, 0, lsparse, 0, gsparse));
494: for (i = ps; i < pf; i++) {
495: PetscCall(MatGetRow(*P, i, &ncols, &pcol, &pval));
496: pmax_pos = 0.;
497: pmax_neg = 0.;
498: for (j = 0; j < ncols; j++) {
499: if (PetscRealPart(pval[j]) > pmax_pos) {
500: pmax_pos = PetscRealPart(pval[j]);
501: } else if (PetscRealPart(pval[j]) < pmax_neg) {
502: pmax_neg = PetscRealPart(pval[j]);
503: }
504: }
505: pthresh_pos = 0.;
506: pthresh_neg = 0.;
507: ptot_pos = 0.;
508: ptot_neg = 0.;
509: for (j = 0; j < ncols; j++) {
510: if (PetscRealPart(pval[j]) >= cls->interp_threshold * pmax_pos) {
511: pthresh_pos += PetscRealPart(pval[j]);
512: } else if (PetscRealPart(pval[j]) <= cls->interp_threshold * pmax_neg) {
513: pthresh_neg += PetscRealPart(pval[j]);
514: }
515: if (PetscRealPart(pval[j]) > 0.) {
516: ptot_pos += PetscRealPart(pval[j]);
517: } else {
518: ptot_neg += PetscRealPart(pval[j]);
519: }
520: }
521: if (PetscAbsReal(pthresh_pos) > 0.) ptot_pos /= pthresh_pos;
522: if (PetscAbsReal(pthresh_neg) > 0.) ptot_neg /= pthresh_neg;
523: idx = 0;
524: for (j = 0; j < ncols; j++) {
525: if (PetscRealPart(pval[j]) >= pmax_pos * cls->interp_threshold) {
526: pnval[idx] = ptot_pos * pval[j];
527: pncol[idx] = pcol[j];
528: idx++;
529: } else if (PetscRealPart(pval[j]) <= pmax_neg * cls->interp_threshold) {
530: pnval[idx] = ptot_neg * pval[j];
531: pncol[idx] = pcol[j];
532: idx++;
533: }
534: }
535: PetscCall(MatRestoreRow(*P, i, &ncols, &pcol, &pval));
536: PetscCall(MatSetValues(Pnew, 1, &i, idx, pncol, pnval, INSERT_VALUES));
537: }
539: PetscCall(MatAssemblyBegin(Pnew, MAT_FINAL_ASSEMBLY));
540: PetscCall(MatAssemblyEnd(Pnew, MAT_FINAL_ASSEMBLY));
541: PetscCall(MatDestroy(P));
543: *P = Pnew;
544: PetscCall(PetscFree2(lsparse, gsparse));
545: PetscCall(PetscFree2(pnval, pncol));
546: PetscFunctionReturn(PETSC_SUCCESS);
547: }
549: static PetscErrorCode PCGAMGProlongator_Classical_Standard(PC pc, Mat A, PetscCoarsenData *agg_lists, Mat *P)
550: {
551: Mat lA, *lAs;
552: MatType mtype;
553: Vec cv;
554: PetscInt *gcid, *lcid, *lsparse, *gsparse, *picol;
555: PetscInt fs, fe, cs, ce, nl, i, j, k, li, lni, ci, ncols, maxcols, fn, cn, cid;
556: PetscMPIInt size;
557: const PetscInt *lidx, *icol, *gidx;
558: PetscBool iscoarse;
559: PetscScalar vi, pentry, pjentry;
560: PetscScalar *pcontrib, *pvcol;
561: const PetscScalar *vcol;
562: PetscReal diag, jdiag, jwttotal;
563: PetscInt pncols;
564: PetscSF sf;
565: PetscLayout clayout;
566: IS lis;
568: PetscFunctionBegin;
569: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
570: PetscCall(MatGetOwnershipRange(A, &fs, &fe));
571: fn = fe - fs;
572: PetscCall(ISCreateStride(PETSC_COMM_SELF, fe - fs, fs, 1, &lis));
573: if (size > 1) {
574: PetscCall(MatGetLayouts(A, NULL, &clayout));
575: /* increase the overlap by two to get neighbors of neighbors */
576: PetscCall(MatIncreaseOverlap(A, 1, &lis, 2));
577: PetscCall(ISSort(lis));
578: /* get the local part of A */
579: PetscCall(MatCreateSubMatrices(A, 1, &lis, &lis, MAT_INITIAL_MATRIX, &lAs));
580: lA = lAs[0];
581: /* build an SF out of it */
582: PetscCall(ISGetLocalSize(lis, &nl));
583: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
584: PetscCall(ISGetIndices(lis, &lidx));
585: PetscCall(PetscSFSetGraphLayout(sf, clayout, nl, NULL, PETSC_COPY_VALUES, lidx));
586: PetscCall(ISRestoreIndices(lis, &lidx));
587: } else {
588: lA = A;
589: nl = fn;
590: }
591: /* create a communication structure for the overlapped portion and transmit coarse indices */
592: PetscCall(PetscMalloc3(fn, &lsparse, fn, &gsparse, nl, &pcontrib));
593: /* create coarse vector */
594: cn = 0;
595: for (i = 0; i < fn; i++) {
596: PetscCall(PetscCDIsEmptyAt(agg_lists, i, &iscoarse));
597: if (!iscoarse) cn++;
598: }
599: PetscCall(PetscMalloc1(fn, &gcid));
600: PetscCall(VecCreateMPI(PetscObjectComm((PetscObject)A), cn, PETSC_DECIDE, &cv));
601: PetscCall(VecGetOwnershipRange(cv, &cs, &ce));
602: cn = 0;
603: for (i = 0; i < fn; i++) {
604: PetscCall(PetscCDIsEmptyAt(agg_lists, i, &iscoarse));
605: if (!iscoarse) {
606: gcid[i] = cs + cn;
607: cn++;
608: } else {
609: gcid[i] = -1;
610: }
611: }
612: if (size > 1) {
613: PetscCall(PetscMalloc1(nl, &lcid));
614: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, gcid, lcid, MPI_REPLACE));
615: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, gcid, lcid, MPI_REPLACE));
616: } else {
617: lcid = gcid;
618: }
619: /* count to preallocate the prolongator */
620: PetscCall(ISGetIndices(lis, &gidx));
621: maxcols = 0;
622: /* count the number of unique contributing coarse cells for each fine */
623: for (i = 0; i < nl; i++) {
624: pcontrib[i] = 0.;
625: PetscCall(MatGetRow(lA, i, &ncols, &icol, NULL));
626: if (gidx[i] >= fs && gidx[i] < fe) {
627: li = gidx[i] - fs;
628: lsparse[li] = 0;
629: gsparse[li] = 0;
630: cid = lcid[i];
631: if (cid >= 0) {
632: lsparse[li] = 1;
633: } else {
634: for (j = 0; j < ncols; j++) {
635: if (lcid[icol[j]] >= 0) {
636: pcontrib[icol[j]] = 1.;
637: } else {
638: ci = icol[j];
639: PetscCall(MatRestoreRow(lA, i, &ncols, &icol, NULL));
640: PetscCall(MatGetRow(lA, ci, &ncols, &icol, NULL));
641: for (k = 0; k < ncols; k++) {
642: if (lcid[icol[k]] >= 0) pcontrib[icol[k]] = 1.;
643: }
644: PetscCall(MatRestoreRow(lA, ci, &ncols, &icol, NULL));
645: PetscCall(MatGetRow(lA, i, &ncols, &icol, NULL));
646: }
647: }
648: for (j = 0; j < ncols; j++) {
649: if (lcid[icol[j]] >= 0 && pcontrib[icol[j]] != 0.) {
650: lni = lcid[icol[j]];
651: if (lni >= cs && lni < ce) {
652: lsparse[li]++;
653: } else {
654: gsparse[li]++;
655: }
656: pcontrib[icol[j]] = 0.;
657: } else {
658: ci = icol[j];
659: PetscCall(MatRestoreRow(lA, i, &ncols, &icol, NULL));
660: PetscCall(MatGetRow(lA, ci, &ncols, &icol, NULL));
661: for (k = 0; k < ncols; k++) {
662: if (lcid[icol[k]] >= 0 && pcontrib[icol[k]] != 0.) {
663: lni = lcid[icol[k]];
664: if (lni >= cs && lni < ce) {
665: lsparse[li]++;
666: } else {
667: gsparse[li]++;
668: }
669: pcontrib[icol[k]] = 0.;
670: }
671: }
672: PetscCall(MatRestoreRow(lA, ci, &ncols, &icol, NULL));
673: PetscCall(MatGetRow(lA, i, &ncols, &icol, NULL));
674: }
675: }
676: }
677: if (lsparse[li] + gsparse[li] > maxcols) maxcols = lsparse[li] + gsparse[li];
678: }
679: PetscCall(MatRestoreRow(lA, i, &ncols, &icol, &vcol));
680: }
681: PetscCall(PetscMalloc2(maxcols, &picol, maxcols, &pvcol));
682: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), P));
683: PetscCall(MatGetType(A, &mtype));
684: PetscCall(MatSetType(*P, mtype));
685: PetscCall(MatSetSizes(*P, fn, cn, PETSC_DETERMINE, PETSC_DETERMINE));
686: PetscCall(MatMPIAIJSetPreallocation(*P, 0, lsparse, 0, gsparse));
687: PetscCall(MatSeqAIJSetPreallocation(*P, 0, lsparse));
688: for (i = 0; i < nl; i++) {
689: diag = 0.;
690: if (gidx[i] >= fs && gidx[i] < fe) {
691: pncols = 0;
692: cid = lcid[i];
693: if (cid >= 0) {
694: pncols = 1;
695: picol[0] = cid;
696: pvcol[0] = 1.;
697: } else {
698: PetscCall(MatGetRow(lA, i, &ncols, &icol, &vcol));
699: for (j = 0; j < ncols; j++) {
700: pentry = vcol[j];
701: if (lcid[icol[j]] >= 0) {
702: /* coarse neighbor */
703: pcontrib[icol[j]] += pentry;
704: } else if (icol[j] != i) {
705: /* the neighbor is a strongly connected fine node */
706: ci = icol[j];
707: vi = vcol[j];
708: PetscCall(MatRestoreRow(lA, i, &ncols, &icol, &vcol));
709: PetscCall(MatGetRow(lA, ci, &ncols, &icol, &vcol));
710: jwttotal = 0.;
711: jdiag = 0.;
712: for (k = 0; k < ncols; k++) {
713: if (ci == icol[k]) jdiag = PetscRealPart(vcol[k]);
714: }
715: for (k = 0; k < ncols; k++) {
716: if (lcid[icol[k]] >= 0 && jdiag * PetscRealPart(vcol[k]) < 0.) {
717: pjentry = vcol[k];
718: jwttotal += PetscRealPart(pjentry);
719: }
720: }
721: if (jwttotal != 0.) {
722: jwttotal = PetscRealPart(vi) / jwttotal;
723: for (k = 0; k < ncols; k++) {
724: if (lcid[icol[k]] >= 0 && jdiag * PetscRealPart(vcol[k]) < 0.) {
725: pjentry = vcol[k] * jwttotal;
726: pcontrib[icol[k]] += pjentry;
727: }
728: }
729: } else {
730: diag += PetscRealPart(vi);
731: }
732: PetscCall(MatRestoreRow(lA, ci, &ncols, &icol, &vcol));
733: PetscCall(MatGetRow(lA, i, &ncols, &icol, &vcol));
734: } else {
735: diag += PetscRealPart(vcol[j]);
736: }
737: }
738: if (diag != 0.) {
739: diag = 1. / diag;
740: for (j = 0; j < ncols; j++) {
741: if (lcid[icol[j]] >= 0 && pcontrib[icol[j]] != 0.) {
742: /* the neighbor is a coarse node */
743: if (PetscAbsScalar(pcontrib[icol[j]]) > 0.0) {
744: lni = lcid[icol[j]];
745: pvcol[pncols] = -pcontrib[icol[j]] * diag;
746: picol[pncols] = lni;
747: pncols++;
748: }
749: pcontrib[icol[j]] = 0.;
750: } else {
751: /* the neighbor is a strongly connected fine node */
752: ci = icol[j];
753: PetscCall(MatRestoreRow(lA, i, &ncols, &icol, &vcol));
754: PetscCall(MatGetRow(lA, ci, &ncols, &icol, &vcol));
755: for (k = 0; k < ncols; k++) {
756: if (lcid[icol[k]] >= 0 && pcontrib[icol[k]] != 0.) {
757: if (PetscAbsScalar(pcontrib[icol[k]]) > 0.0) {
758: lni = lcid[icol[k]];
759: pvcol[pncols] = -pcontrib[icol[k]] * diag;
760: picol[pncols] = lni;
761: pncols++;
762: }
763: pcontrib[icol[k]] = 0.;
764: }
765: }
766: PetscCall(MatRestoreRow(lA, ci, &ncols, &icol, &vcol));
767: PetscCall(MatGetRow(lA, i, &ncols, &icol, &vcol));
768: }
769: pcontrib[icol[j]] = 0.;
770: }
771: PetscCall(MatRestoreRow(lA, i, &ncols, &icol, &vcol));
772: }
773: }
774: ci = gidx[i];
775: if (pncols > 0) PetscCall(MatSetValues(*P, 1, &ci, pncols, picol, pvcol, INSERT_VALUES));
776: }
777: }
778: PetscCall(ISRestoreIndices(lis, &gidx));
779: PetscCall(PetscFree2(picol, pvcol));
780: PetscCall(PetscFree3(lsparse, gsparse, pcontrib));
781: PetscCall(ISDestroy(&lis));
782: PetscCall(PetscFree(gcid));
783: if (size > 1) {
784: PetscCall(PetscFree(lcid));
785: PetscCall(MatDestroyMatrices(1, &lAs));
786: PetscCall(PetscSFDestroy(&sf));
787: }
788: PetscCall(VecDestroy(&cv));
789: PetscCall(MatAssemblyBegin(*P, MAT_FINAL_ASSEMBLY));
790: PetscCall(MatAssemblyEnd(*P, MAT_FINAL_ASSEMBLY));
791: PetscFunctionReturn(PETSC_SUCCESS);
792: }
794: static PetscErrorCode PCGAMGOptProlongator_Classical_Jacobi(PC pc, Mat A, Mat *P)
795: {
796: PetscInt f, s, n, cf, cs, i, idx;
797: PetscInt *coarserows;
798: PetscInt ncols;
799: const PetscInt *pcols;
800: const PetscScalar *pvals;
801: Mat Pnew;
802: Vec diag;
803: PC_MG *mg = (PC_MG *)pc->data;
804: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
805: PC_GAMG_Classical *cls = (PC_GAMG_Classical *)pc_gamg->subctx;
807: PetscFunctionBegin;
808: if (cls->nsmooths == 0) {
809: PetscCall(PCGAMGTruncateProlongator_Private(pc, P));
810: PetscFunctionReturn(PETSC_SUCCESS);
811: }
812: PetscCall(MatGetOwnershipRange(*P, &s, &f));
813: n = f - s;
814: PetscCall(MatGetOwnershipRangeColumn(*P, &cs, &cf));
815: PetscCall(PetscMalloc1(n, &coarserows));
816: /* identify the rows corresponding to coarse unknowns */
817: idx = 0;
818: for (i = s; i < f; i++) {
819: PetscCall(MatGetRow(*P, i, &ncols, &pcols, &pvals));
820: /* assume, for now, that it's a coarse unknown if it has a single unit entry */
821: if (ncols == 1) {
822: if (pvals[0] == 1.) {
823: coarserows[idx] = i;
824: idx++;
825: }
826: }
827: PetscCall(MatRestoreRow(*P, i, &ncols, &pcols, &pvals));
828: }
829: PetscCall(MatCreateVecs(A, &diag, NULL));
830: PetscCall(MatGetDiagonal(A, diag));
831: PetscCall(VecReciprocal(diag));
832: for (i = 0; i < cls->nsmooths; i++) {
833: PetscCall(MatMatMult(A, *P, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Pnew));
834: PetscCall(MatZeroRows(Pnew, idx, coarserows, 0., NULL, NULL));
835: PetscCall(MatDiagonalScale(Pnew, diag, NULL));
836: PetscCall(MatAYPX(Pnew, -1.0, *P, DIFFERENT_NONZERO_PATTERN));
837: PetscCall(MatDestroy(P));
838: *P = Pnew;
839: Pnew = NULL;
840: }
841: PetscCall(VecDestroy(&diag));
842: PetscCall(PetscFree(coarserows));
843: PetscCall(PCGAMGTruncateProlongator_Private(pc, P));
844: PetscFunctionReturn(PETSC_SUCCESS);
845: }
847: static PetscErrorCode PCGAMGProlongator_Classical(PC pc, Mat A, PetscCoarsenData *agg_lists, Mat *P)
848: {
849: PetscErrorCode (*f)(PC, Mat, PetscCoarsenData *, Mat *);
850: PC_MG *mg = (PC_MG *)pc->data;
851: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
852: PC_GAMG_Classical *cls = (PC_GAMG_Classical *)pc_gamg->subctx;
854: PetscFunctionBegin;
855: PetscCall(PetscFunctionListFind(PCGAMGClassicalProlongatorList, cls->prolongtype, &f));
856: PetscCheck(f, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONGSTATE, "Cannot find PCGAMG Classical prolongator type");
857: PetscCall((*f)(pc, A, agg_lists, P));
858: PetscFunctionReturn(PETSC_SUCCESS);
859: }
861: static PetscErrorCode PCGAMGDestroy_Classical(PC pc)
862: {
863: PC_MG *mg = (PC_MG *)pc->data;
864: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
866: PetscFunctionBegin;
867: PetscCall(PetscFree(pc_gamg->subctx));
868: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCGAMGClassicalSetType_C", NULL));
869: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCGAMGClassicalGetType_C", NULL));
870: PetscFunctionReturn(PETSC_SUCCESS);
871: }
873: static PetscErrorCode PCGAMGSetFromOptions_Classical(PC pc, PetscOptionItems PetscOptionsObject)
874: {
875: PC_MG *mg = (PC_MG *)pc->data;
876: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
877: PC_GAMG_Classical *cls = (PC_GAMG_Classical *)pc_gamg->subctx;
878: char tname[256];
879: PetscBool flg;
881: PetscFunctionBegin;
882: PetscOptionsHeadBegin(PetscOptionsObject, "GAMG-Classical options");
883: PetscCall(PetscOptionsFList("-pc_gamg_classical_type", "Type of Classical AMG prolongation", "PCGAMGClassicalSetType", PCGAMGClassicalProlongatorList, cls->prolongtype, tname, sizeof(tname), &flg));
884: if (flg) PetscCall(PCGAMGClassicalSetType(pc, tname));
885: PetscCall(PetscOptionsReal("-pc_gamg_classical_interp_threshold", "Threshold for classical interpolator entries", "", cls->interp_threshold, &cls->interp_threshold, NULL));
886: PetscCall(PetscOptionsInt("-pc_gamg_classical_nsmooths", "Threshold for classical interpolator entries", "", cls->nsmooths, &cls->nsmooths, NULL));
887: PetscOptionsHeadEnd();
888: PetscFunctionReturn(PETSC_SUCCESS);
889: }
891: static PetscErrorCode PCGAMGSetData_Classical(PC pc, Mat A)
892: {
893: PC_MG *mg = (PC_MG *)pc->data;
894: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
896: PetscFunctionBegin;
897: /* no data for classical AMG */
898: pc_gamg->data = NULL;
899: pc_gamg->data_cell_cols = 0;
900: pc_gamg->data_cell_rows = 0;
901: pc_gamg->data_sz = 0;
902: PetscFunctionReturn(PETSC_SUCCESS);
903: }
905: static PetscErrorCode PCGAMGClassicalFinalizePackage(void)
906: {
907: PetscFunctionBegin;
908: PCGAMGClassicalPackageInitialized = PETSC_FALSE;
909: PetscCall(PetscFunctionListDestroy(&PCGAMGClassicalProlongatorList));
910: PetscFunctionReturn(PETSC_SUCCESS);
911: }
913: static PetscErrorCode PCGAMGClassicalInitializePackage(void)
914: {
915: PetscFunctionBegin;
916: if (PCGAMGClassicalPackageInitialized) PetscFunctionReturn(PETSC_SUCCESS);
917: PetscCall(PetscFunctionListAdd(&PCGAMGClassicalProlongatorList, PCGAMGCLASSICALDIRECT, PCGAMGProlongator_Classical_Direct));
918: PetscCall(PetscFunctionListAdd(&PCGAMGClassicalProlongatorList, PCGAMGCLASSICALSTANDARD, PCGAMGProlongator_Classical_Standard));
919: PetscCall(PetscRegisterFinalize(PCGAMGClassicalFinalizePackage));
920: PetscFunctionReturn(PETSC_SUCCESS);
921: }
923: PetscErrorCode PCCreateGAMG_Classical(PC pc)
924: {
925: PC_MG *mg = (PC_MG *)pc->data;
926: PC_GAMG *pc_gamg = (PC_GAMG *)mg->innerctx;
927: PC_GAMG_Classical *pc_gamg_classical;
929: PetscFunctionBegin;
930: PetscCall(PCGAMGClassicalInitializePackage());
931: if (pc_gamg->subctx) {
932: /* call base class */
933: PetscCall(PCDestroy_GAMG(pc));
934: }
936: /* create sub context for SA */
937: PetscCall(PetscNew(&pc_gamg_classical));
938: pc_gamg->subctx = pc_gamg_classical;
939: pc->ops->setfromoptions = PCGAMGSetFromOptions_Classical;
940: /* reset does not do anything; setup not virtual */
942: /* set internal function pointers */
943: pc_gamg->ops->destroy = PCGAMGDestroy_Classical;
944: pc_gamg->ops->creategraph = PCGAMGCreateGraph_Classical;
945: pc_gamg->ops->coarsen = PCGAMGCoarsen_Classical;
946: pc_gamg->ops->prolongator = PCGAMGProlongator_Classical;
947: pc_gamg->ops->optprolongator = PCGAMGOptProlongator_Classical_Jacobi;
948: pc_gamg->ops->setfromoptions = PCGAMGSetFromOptions_Classical;
950: pc_gamg->ops->createdefaultdata = PCGAMGSetData_Classical;
951: pc_gamg_classical->interp_threshold = 0.2;
952: pc_gamg_classical->nsmooths = 0;
953: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCGAMGClassicalSetType_C", PCGAMGClassicalSetType_GAMG));
954: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCGAMGClassicalGetType_C", PCGAMGClassicalGetType_GAMG));
955: PetscCall(PCGAMGClassicalSetType(pc, PCGAMGCLASSICALSTANDARD));
956: PetscFunctionReturn(PETSC_SUCCESS);
957: }