Actual source code: ml.c
1: /*
2: Provides an interface to the ML smoothed Aggregation
3: Note: Something non-obvious breaks -pc_mg_type ADDITIVE for parallel runs
4: Jed Brown, see [PETSC #18321, #18449].
5: */
6: #include <petsc/private/pcimpl.h>
7: #include <petsc/private/pcmgimpl.h>
8: #include <../src/mat/impls/aij/seq/aij.h>
9: #include <../src/mat/impls/aij/mpi/mpiaij.h>
10: #include <petscdm.h>
12: EXTERN_C_BEGIN
13: /* HAVE_CONFIG_H flag is required by ML include files */
14: #ifndef HAVE_CONFIG_H
15: #define HAVE_CONFIG_H
16: #endif
17: #include <ml_include.h>
18: #include <ml_viz_stats.h>
19: EXTERN_C_END
21: typedef enum {
22: PCML_NULLSPACE_AUTO,
23: PCML_NULLSPACE_USER,
24: PCML_NULLSPACE_BLOCK,
25: PCML_NULLSPACE_SCALAR
26: } PCMLNullSpaceType;
27: static const char *const PCMLNullSpaceTypes[] = {"AUTO", "USER", "BLOCK", "SCALAR", "PCMLNullSpaceType", "PCML_NULLSPACE_", 0};
29: /* The context (data structure) at each grid level */
30: typedef struct {
31: Vec x, b, r; /* global vectors */
32: Mat A, P, R;
33: KSP ksp;
34: Vec coords; /* projected by ML, if PCSetCoordinates is called; values packed by node */
35: } GridCtx;
37: /* The context used to input PETSc matrix into ML at fine grid */
38: typedef struct {
39: Mat A; /* Petsc matrix in aij format */
40: Mat Aloc; /* local portion of A to be used by ML */
41: Vec x, y;
42: ML_Operator *mlmat;
43: PetscScalar *pwork; /* tmp array used by PetscML_comm() */
44: } FineGridCtx;
46: /* The context associates a ML matrix with a PETSc shell matrix */
47: typedef struct {
48: Mat A; /* PETSc shell matrix associated with mlmat */
49: ML_Operator *mlmat; /* ML matrix assorciated with A */
50: } Mat_MLShell;
52: /* Private context for the ML preconditioner */
53: typedef struct {
54: ML *ml_object;
55: ML_Aggregate *agg_object;
56: GridCtx *gridctx;
57: FineGridCtx *PetscMLdata;
58: PetscInt Nlevels, MaxNlevels, MaxCoarseSize, CoarsenScheme, EnergyMinimization, MinPerProc, PutOnSingleProc, RepartitionType, ZoltanScheme;
59: PetscReal Threshold, DampingFactor, EnergyMinimizationDropTol, MaxMinRatio, AuxThreshold;
60: PetscBool SpectralNormScheme_Anorm, BlockScaling, EnergyMinimizationCheap, Symmetrize, OldHierarchy, KeepAggInfo, Reusable, Repartition, Aux;
61: PetscBool reuse_interpolation;
62: PCMLNullSpaceType nulltype;
63: PetscMPIInt size; /* size of communicator for pc->pmat */
64: PetscInt dim; /* data from PCSetCoordinates(_ML) */
65: PetscInt nloc;
66: PetscReal *coords; /* ML has a grid object for each level: the finest grid will point into coords */
67: } PC_ML;
69: static int PetscML_getrow(ML_Operator *ML_data, int N_requested_rows, int requested_rows[], int allocated_space, int columns[], double values[], int row_lengths[])
70: {
71: PetscInt m, i, j, k = 0, row, *aj;
72: PetscScalar *aa;
73: FineGridCtx *ml = (FineGridCtx *)ML_Get_MyGetrowData(ML_data);
74: Mat_SeqAIJ *a = (Mat_SeqAIJ *)ml->Aloc->data;
76: if (MatGetSize(ml->Aloc, &m, NULL)) return 0;
77: for (i = 0; i < N_requested_rows; i++) {
78: row = requested_rows[i];
79: row_lengths[i] = a->ilen[row];
80: if (allocated_space < k + row_lengths[i]) return 0;
81: if ((row >= 0) || (row <= (m - 1))) {
82: aj = a->j + a->i[row];
83: aa = a->a + a->i[row];
84: for (j = 0; j < row_lengths[i]; j++) {
85: columns[k] = aj[j];
86: values[k++] = aa[j];
87: }
88: }
89: }
90: return 1;
91: }
93: static PetscErrorCode PetscML_comm(double p[], void *ML_data)
94: {
95: FineGridCtx *ml = (FineGridCtx *)ML_data;
96: Mat A = ml->A;
97: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
98: PetscMPIInt size;
99: PetscInt i, in_length = A->rmap->n, out_length = ml->Aloc->cmap->n;
100: const PetscScalar *array;
102: PetscFunctionBegin;
103: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
104: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
106: PetscCall(VecPlaceArray(ml->y, p));
107: PetscCall(VecScatterBegin(a->Mvctx, ml->y, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
108: PetscCall(VecScatterEnd(a->Mvctx, ml->y, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
109: PetscCall(VecResetArray(ml->y));
110: PetscCall(VecGetArrayRead(a->lvec, &array));
111: for (i = in_length; i < out_length; i++) p[i] = array[i - in_length];
112: PetscCall(VecRestoreArrayRead(a->lvec, &array));
113: PetscFunctionReturn(PETSC_SUCCESS);
114: }
116: /*
117: Needed since ML expects an int (*)(double *, void *) but PetscErrorCode may be an
118: enum. Instead of modifying PetscML_comm() it is easier to just wrap it
119: */
120: static int ML_PetscML_comm(double p[], void *ML_data)
121: {
122: return (int)PetscML_comm(p, ML_data);
123: }
125: static int PetscML_matvec(ML_Operator *ML_data, int in_length, double p[], int out_length, double ap[])
126: {
127: FineGridCtx *ml = (FineGridCtx *)ML_Get_MyMatvecData(ML_data);
128: Mat A = ml->A, Aloc = ml->Aloc;
129: PetscMPIInt size;
130: PetscScalar *pwork = ml->pwork;
131: PetscInt i;
133: PetscFunctionBegin;
134: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
135: if (size == 1) {
136: PetscCall(VecPlaceArray(ml->x, p));
137: } else {
138: for (i = 0; i < in_length; i++) pwork[i] = p[i];
139: PetscCall(PetscML_comm(pwork, ml));
140: PetscCall(VecPlaceArray(ml->x, pwork));
141: }
142: PetscCall(VecPlaceArray(ml->y, ap));
143: PetscCall(MatMult(Aloc, ml->x, ml->y));
144: PetscCall(VecResetArray(ml->x));
145: PetscCall(VecResetArray(ml->y));
146: PetscFunctionReturn(PETSC_SUCCESS);
147: }
149: static PetscErrorCode MatMult_ML(Mat A, Vec x, Vec y)
150: {
151: Mat_MLShell *shell;
152: PetscScalar *yarray;
153: const PetscScalar *xarray;
154: PetscInt x_length, y_length;
156: PetscFunctionBegin;
157: PetscCall(MatShellGetContext(A, &shell));
158: PetscCall(VecGetArrayRead(x, &xarray));
159: PetscCall(VecGetArray(y, &yarray));
160: x_length = shell->mlmat->invec_leng;
161: y_length = shell->mlmat->outvec_leng;
162: PetscStackCallExternalVoid("ML_Operator_Apply", ML_Operator_Apply(shell->mlmat, x_length, (PetscScalar *)xarray, y_length, yarray));
163: PetscCall(VecRestoreArrayRead(x, &xarray));
164: PetscCall(VecRestoreArray(y, &yarray));
165: PetscFunctionReturn(PETSC_SUCCESS);
166: }
168: /* newtype is ignored since only handles one case */
169: static PetscErrorCode MatConvert_MPIAIJ_ML(Mat A, MatType newtype, MatReuse scall, Mat *Aloc)
170: {
171: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
172: Mat_SeqAIJ *mat, *a = (Mat_SeqAIJ *)mpimat->A->data, *b = (Mat_SeqAIJ *)mpimat->B->data;
173: PetscInt *ai = a->i, *aj = a->j, *bi = b->i, *bj = b->j;
174: PetscScalar *aa, *ba, *ca;
175: PetscInt am = A->rmap->n, an = A->cmap->n, i, j, k;
176: PetscInt *ci, *cj, ncols;
178: PetscFunctionBegin;
179: PetscCheck(am == an, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "A must have a square diagonal portion, am: %d != an: %d", am, an);
180: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, (const PetscScalar **)&aa));
181: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, (const PetscScalar **)&ba));
182: if (scall == MAT_INITIAL_MATRIX) {
183: PetscCall(PetscMalloc1(1 + am, &ci));
184: ci[0] = 0;
185: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
186: PetscCall(PetscMalloc1(1 + ci[am], &cj));
187: PetscCall(PetscMalloc1(1 + ci[am], &ca));
189: k = 0;
190: for (i = 0; i < am; i++) {
191: /* diagonal portion of A */
192: ncols = ai[i + 1] - ai[i];
193: for (j = 0; j < ncols; j++) {
194: cj[k] = *aj++;
195: ca[k++] = *aa++;
196: }
197: /* off-diagonal portion of A */
198: ncols = bi[i + 1] - bi[i];
199: for (j = 0; j < ncols; j++) {
200: cj[k] = an + (*bj);
201: bj++;
202: ca[k++] = *ba++;
203: }
204: }
205: PetscCheck(k == ci[am], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "k: %d != ci[am]: %d", k, ci[am]);
207: /* put together the new matrix */
208: an = mpimat->A->cmap->n + mpimat->B->cmap->n;
209: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, an, ci, cj, ca, Aloc));
211: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
212: /* Since these are PETSc arrays, change flags to free them as necessary. */
213: mat = (Mat_SeqAIJ *)(*Aloc)->data;
214: mat->free_a = PETSC_TRUE;
215: mat->free_ij = PETSC_TRUE;
217: mat->nonew = 0;
218: } else if (scall == MAT_REUSE_MATRIX) {
219: mat = (Mat_SeqAIJ *)(*Aloc)->data;
220: ci = mat->i;
221: cj = mat->j;
222: ca = mat->a;
223: for (i = 0; i < am; i++) {
224: /* diagonal portion of A */
225: ncols = ai[i + 1] - ai[i];
226: for (j = 0; j < ncols; j++) *ca++ = *aa++;
227: /* off-diagonal portion of A */
228: ncols = bi[i + 1] - bi[i];
229: for (j = 0; j < ncols; j++) *ca++ = *ba++;
230: }
231: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
232: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, (const PetscScalar **)&aa));
233: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, (const PetscScalar **)&ba));
234: PetscFunctionReturn(PETSC_SUCCESS);
235: }
237: static PetscErrorCode MatDestroy_ML(Mat A)
238: {
239: Mat_MLShell *shell;
241: PetscFunctionBegin;
242: PetscCall(MatShellGetContext(A, &shell));
243: PetscCall(PetscFree(shell));
244: PetscFunctionReturn(PETSC_SUCCESS);
245: }
247: static PetscErrorCode MatWrapML_SeqAIJ(ML_Operator *mlmat, MatReuse reuse, Mat *newmat)
248: {
249: struct ML_CSR_MSRdata *matdata = (struct ML_CSR_MSRdata *)mlmat->data;
250: PetscInt m = mlmat->outvec_leng, n = mlmat->invec_leng, *nnz = NULL, nz_max;
251: PetscInt *ml_cols = matdata->columns, *ml_rowptr = matdata->rowptr, *aj, i;
252: PetscScalar *ml_vals = matdata->values, *aa;
254: PetscFunctionBegin;
255: PetscCheck(mlmat->getrow, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "mlmat->getrow = NULL");
256: if (m != n) { /* ML Pmat and Rmat are in CSR format. Pass array pointers into SeqAIJ matrix */
257: if (reuse) {
258: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)(*newmat)->data;
259: aij->i = ml_rowptr;
260: aij->j = ml_cols;
261: aij->a = ml_vals;
262: } else {
263: /* sort ml_cols and ml_vals */
264: PetscCall(PetscMalloc1(m + 1, &nnz));
265: for (i = 0; i < m; i++) nnz[i] = ml_rowptr[i + 1] - ml_rowptr[i];
266: aj = ml_cols;
267: aa = ml_vals;
268: for (i = 0; i < m; i++) {
269: PetscCall(PetscSortIntWithScalarArray(nnz[i], aj, aa));
270: aj += nnz[i];
271: aa += nnz[i];
272: }
273: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, ml_rowptr, ml_cols, ml_vals, newmat));
274: PetscCall(PetscFree(nnz));
275: }
276: PetscCall(MatAssemblyBegin(*newmat, MAT_FINAL_ASSEMBLY));
277: PetscCall(MatAssemblyEnd(*newmat, MAT_FINAL_ASSEMBLY));
278: PetscFunctionReturn(PETSC_SUCCESS);
279: }
281: nz_max = PetscMax(1, mlmat->max_nz_per_row);
282: PetscCall(PetscMalloc2(nz_max, &aa, nz_max, &aj));
283: if (!reuse) {
284: PetscCall(MatCreate(PETSC_COMM_SELF, newmat));
285: PetscCall(MatSetSizes(*newmat, m, n, PETSC_DECIDE, PETSC_DECIDE));
286: PetscCall(MatSetType(*newmat, MATSEQAIJ));
287: /* keep track of block size for A matrices */
288: PetscCall(MatSetBlockSize(*newmat, mlmat->num_PDEs));
290: PetscCall(PetscMalloc1(m, &nnz));
291: for (i = 0; i < m; i++) PetscStackCallExternalVoid("ML_Operator_Getrow", ML_Operator_Getrow(mlmat, 1, &i, nz_max, aj, aa, &nnz[i]));
292: PetscCall(MatSeqAIJSetPreallocation(*newmat, 0, nnz));
293: }
294: for (i = 0; i < m; i++) {
295: PetscInt ncols;
297: PetscStackCallExternalVoid("ML_Operator_Getrow", ML_Operator_Getrow(mlmat, 1, &i, nz_max, aj, aa, &ncols));
298: PetscCall(MatSetValues(*newmat, 1, &i, ncols, aj, aa, INSERT_VALUES));
299: }
300: PetscCall(MatAssemblyBegin(*newmat, MAT_FINAL_ASSEMBLY));
301: PetscCall(MatAssemblyEnd(*newmat, MAT_FINAL_ASSEMBLY));
303: PetscCall(PetscFree2(aa, aj));
304: PetscCall(PetscFree(nnz));
305: PetscFunctionReturn(PETSC_SUCCESS);
306: }
308: static PetscErrorCode MatWrapML_SHELL(ML_Operator *mlmat, MatReuse reuse, Mat *newmat)
309: {
310: PetscInt m, n;
311: ML_Comm *MLcomm;
312: Mat_MLShell *shellctx;
314: PetscFunctionBegin;
315: m = mlmat->outvec_leng;
316: n = mlmat->invec_leng;
318: if (reuse) {
319: PetscCall(MatShellGetContext(*newmat, &shellctx));
320: shellctx->mlmat = mlmat;
321: PetscFunctionReturn(PETSC_SUCCESS);
322: }
324: MLcomm = mlmat->comm;
326: PetscCall(PetscNew(&shellctx));
327: PetscCall(MatCreateShell(MLcomm->USR_comm, m, n, PETSC_DETERMINE, PETSC_DETERMINE, shellctx, newmat));
328: PetscCall(MatShellSetOperation(*newmat, MATOP_MULT, (void (*)(void))MatMult_ML));
329: PetscCall(MatShellSetOperation(*newmat, MATOP_DESTROY, (void (*)(void))MatDestroy_ML));
331: shellctx->A = *newmat;
332: shellctx->mlmat = mlmat;
333: PetscFunctionReturn(PETSC_SUCCESS);
334: }
336: static PetscErrorCode MatWrapML_MPIAIJ(ML_Operator *mlmat, MatReuse reuse, Mat *newmat)
337: {
338: PetscInt *aj;
339: PetscScalar *aa;
340: PetscInt i, j, *gordering;
341: PetscInt m = mlmat->outvec_leng, n, nz_max, row;
342: Mat A;
344: PetscFunctionBegin;
345: PetscCheck(mlmat->getrow, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "mlmat->getrow = NULL");
346: n = mlmat->invec_leng;
347: PetscCheck(m == n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "m %d must equal to n %d", m, n);
349: /* create global row numbering for a ML_Operator */
350: PetscStackCallExternalVoid("ML_build_global_numbering", ML_build_global_numbering(mlmat, &gordering, "rows"));
352: nz_max = PetscMax(1, mlmat->max_nz_per_row) + 1;
353: PetscCall(PetscMalloc2(nz_max, &aa, nz_max, &aj));
354: if (reuse) {
355: A = *newmat;
356: } else {
357: PetscInt *nnzA, *nnzB, *nnz;
358: PetscInt rstart;
359: PetscCall(MatCreate(mlmat->comm->USR_comm, &A));
360: PetscCall(MatSetSizes(A, m, n, PETSC_DECIDE, PETSC_DECIDE));
361: PetscCall(MatSetType(A, MATMPIAIJ));
362: /* keep track of block size for A matrices */
363: PetscCall(MatSetBlockSize(A, mlmat->num_PDEs));
364: PetscCall(PetscMalloc3(m, &nnzA, m, &nnzB, m, &nnz));
365: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, mlmat->comm->USR_comm));
366: rstart -= m;
368: for (i = 0; i < m; i++) {
369: row = gordering[i] - rstart;
370: PetscStackCallExternalVoid("ML_Operator_Getrow", ML_Operator_Getrow(mlmat, 1, &i, nz_max, aj, aa, &nnz[i]));
371: nnzA[row] = 0;
372: for (j = 0; j < nnz[i]; j++) {
373: if (aj[j] < m) nnzA[row]++;
374: }
375: nnzB[row] = nnz[i] - nnzA[row];
376: }
377: PetscCall(MatMPIAIJSetPreallocation(A, 0, nnzA, 0, nnzB));
378: PetscCall(PetscFree3(nnzA, nnzB, nnz));
379: }
380: for (i = 0; i < m; i++) {
381: PetscInt ncols;
382: row = gordering[i];
384: PetscStackCallExternalVoid(",ML_Operator_Getrow", ML_Operator_Getrow(mlmat, 1, &i, nz_max, aj, aa, &ncols));
385: for (j = 0; j < ncols; j++) aj[j] = gordering[aj[j]];
386: PetscCall(MatSetValues(A, 1, &row, ncols, aj, aa, INSERT_VALUES));
387: }
388: PetscStackCallExternalVoid("ML_free", ML_free(gordering));
389: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
390: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
391: *newmat = A;
393: PetscCall(PetscFree2(aa, aj));
394: PetscFunctionReturn(PETSC_SUCCESS);
395: }
397: /*
398: PCSetCoordinates_ML
400: Input Parameter:
401: . pc - the preconditioner context
402: */
403: static PetscErrorCode PCSetCoordinates_ML(PC pc, PetscInt ndm, PetscInt a_nloc, PetscReal *coords)
404: {
405: PC_MG *mg = (PC_MG *)pc->data;
406: PC_ML *pc_ml = (PC_ML *)mg->innerctx;
407: PetscInt arrsz, oldarrsz, bs, my0, kk, ii, nloc, Iend, aloc;
408: Mat Amat = pc->pmat;
410: /* this function copied and modified from PCSetCoordinates_GEO -TGI */
411: PetscFunctionBegin;
413: PetscCall(MatGetBlockSize(Amat, &bs));
415: PetscCall(MatGetOwnershipRange(Amat, &my0, &Iend));
416: aloc = (Iend - my0);
417: nloc = (Iend - my0) / bs;
419: PetscCheck((nloc == a_nloc) || (aloc == a_nloc), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Number of local blocks %" PetscInt_FMT " must be %" PetscInt_FMT " or %" PetscInt_FMT ".", a_nloc, nloc, aloc);
421: oldarrsz = pc_ml->dim * pc_ml->nloc;
422: pc_ml->dim = ndm;
423: pc_ml->nloc = nloc;
424: arrsz = ndm * nloc;
426: /* create data - syntactic sugar that should be refactored at some point */
427: if (pc_ml->coords == 0 || (oldarrsz != arrsz)) {
428: PetscCall(PetscFree(pc_ml->coords));
429: PetscCall(PetscMalloc1(arrsz, &pc_ml->coords));
430: }
431: for (kk = 0; kk < arrsz; kk++) pc_ml->coords[kk] = -999.;
432: /* copy data in - column-oriented */
433: if (nloc == a_nloc) {
434: for (kk = 0; kk < nloc; kk++) {
435: for (ii = 0; ii < ndm; ii++) pc_ml->coords[ii * nloc + kk] = coords[kk * ndm + ii];
436: }
437: } else { /* assumes the coordinates are blocked */
438: for (kk = 0; kk < nloc; kk++) {
439: for (ii = 0; ii < ndm; ii++) pc_ml->coords[ii * nloc + kk] = coords[bs * kk * ndm + ii];
440: }
441: }
442: PetscFunctionReturn(PETSC_SUCCESS);
443: }
445: extern PetscErrorCode PCReset_MG(PC);
446: static PetscErrorCode PCReset_ML(PC pc)
447: {
448: PC_MG *mg = (PC_MG *)pc->data;
449: PC_ML *pc_ml = (PC_ML *)mg->innerctx;
450: PetscInt level, fine_level = pc_ml->Nlevels - 1, dim = pc_ml->dim;
452: PetscFunctionBegin;
453: if (dim) {
454: for (level = 0; level <= fine_level; level++) PetscCall(VecDestroy(&pc_ml->gridctx[level].coords));
455: if (pc_ml->ml_object && pc_ml->ml_object->Grid) {
456: ML_Aggregate_Viz_Stats *grid_info = (ML_Aggregate_Viz_Stats *)pc_ml->ml_object->Grid[0].Grid;
457: grid_info->x = 0; /* do this so ML doesn't try to free coordinates */
458: grid_info->y = 0;
459: grid_info->z = 0;
460: PetscStackCallExternalVoid("ML_Operator_Getrow", ML_Aggregate_VizAndStats_Clean(pc_ml->ml_object));
461: }
462: }
463: PetscStackCallExternalVoid("ML_Aggregate_Destroy", ML_Aggregate_Destroy(&pc_ml->agg_object));
464: PetscStackCallExternalVoid("ML_Aggregate_Destroy", ML_Destroy(&pc_ml->ml_object));
466: if (pc_ml->PetscMLdata) {
467: PetscCall(PetscFree(pc_ml->PetscMLdata->pwork));
468: PetscCall(MatDestroy(&pc_ml->PetscMLdata->Aloc));
469: PetscCall(VecDestroy(&pc_ml->PetscMLdata->x));
470: PetscCall(VecDestroy(&pc_ml->PetscMLdata->y));
471: }
472: PetscCall(PetscFree(pc_ml->PetscMLdata));
474: if (pc_ml->gridctx) {
475: for (level = 0; level < fine_level; level++) {
476: if (pc_ml->gridctx[level].A) PetscCall(MatDestroy(&pc_ml->gridctx[level].A));
477: if (pc_ml->gridctx[level].P) PetscCall(MatDestroy(&pc_ml->gridctx[level].P));
478: if (pc_ml->gridctx[level].R) PetscCall(MatDestroy(&pc_ml->gridctx[level].R));
479: if (pc_ml->gridctx[level].x) PetscCall(VecDestroy(&pc_ml->gridctx[level].x));
480: if (pc_ml->gridctx[level].b) PetscCall(VecDestroy(&pc_ml->gridctx[level].b));
481: if (pc_ml->gridctx[level + 1].r) PetscCall(VecDestroy(&pc_ml->gridctx[level + 1].r));
482: }
483: }
484: PetscCall(PetscFree(pc_ml->gridctx));
485: PetscCall(PetscFree(pc_ml->coords));
487: pc_ml->dim = 0;
488: pc_ml->nloc = 0;
489: PetscCall(PCReset_MG(pc));
490: PetscFunctionReturn(PETSC_SUCCESS);
491: }
493: /*
494: PCSetUp_ML - Prepares for the use of the ML preconditioner
495: by setting data structures and options.
497: Input Parameter:
498: . pc - the preconditioner context
500: Application Interface Routine: PCSetUp()
502: Note:
503: The interface routine PCSetUp() is not usually called directly by
504: the user, but instead is called by PCApply() if necessary.
505: */
506: extern PetscErrorCode PCSetFromOptions_MG(PC, PetscOptionItems *PetscOptionsObject);
507: extern PetscErrorCode PCReset_MG(PC);
509: static PetscErrorCode PCSetUp_ML(PC pc)
510: {
511: PetscMPIInt size;
512: FineGridCtx *PetscMLdata;
513: ML *ml_object;
514: ML_Aggregate *agg_object;
515: ML_Operator *mlmat;
516: PetscInt nlocal_allcols, Nlevels, mllevel, level, level1, m, fine_level, bs;
517: Mat A, Aloc;
518: GridCtx *gridctx;
519: PC_MG *mg = (PC_MG *)pc->data;
520: PC_ML *pc_ml = (PC_ML *)mg->innerctx;
521: PetscBool isSeq, isMPI;
522: KSP smoother;
523: PC subpc;
524: PetscInt mesh_level, old_mesh_level;
525: MatInfo info;
526: static PetscBool cite = PETSC_FALSE;
528: PetscFunctionBegin;
529: PetscCall(PetscCitationsRegister("@TechReport{ml_users_guide,\n author = {M. Sala and J.J. Hu and R.S. Tuminaro},\n title = {{ML}3.1 {S}moothed {A}ggregation {U}ser's {G}uide},\n institution = {Sandia National Laboratories},\n number = "
530: "{SAND2004-4821},\n year = 2004\n}\n",
531: &cite));
532: A = pc->pmat;
533: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
535: if (pc->setupcalled) {
536: if (pc->flag == SAME_NONZERO_PATTERN && pc_ml->reuse_interpolation) {
537: /*
538: Reuse interpolaton instead of recomputing aggregates and updating the whole hierarchy. This is less expensive for
539: multiple solves in which the matrix is not changing too quickly.
540: */
541: ml_object = pc_ml->ml_object;
542: gridctx = pc_ml->gridctx;
543: Nlevels = pc_ml->Nlevels;
544: fine_level = Nlevels - 1;
545: gridctx[fine_level].A = A;
547: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isSeq));
548: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isMPI));
549: PetscCheck(isMPI || isSeq, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.", ((PetscObject)A)->type_name);
550: if (isMPI) {
551: PetscCall(MatConvert_MPIAIJ_ML(A, NULL, MAT_INITIAL_MATRIX, &Aloc));
552: } else {
553: Aloc = A;
554: PetscCall(PetscObjectReference((PetscObject)Aloc));
555: }
557: PetscCall(MatGetSize(Aloc, &m, &nlocal_allcols));
558: PetscMLdata = pc_ml->PetscMLdata;
559: PetscCall(MatDestroy(&PetscMLdata->Aloc));
560: PetscMLdata->A = A;
561: PetscMLdata->Aloc = Aloc;
562: PetscStackCallExternalVoid("ML_Aggregate_Destroy", ML_Init_Amatrix(ml_object, 0, m, m, PetscMLdata));
563: PetscStackCallExternalVoid("ML_Set_Amatrix_Matvec", ML_Set_Amatrix_Matvec(ml_object, 0, PetscML_matvec));
565: mesh_level = ml_object->ML_finest_level;
566: while (ml_object->SingleLevel[mesh_level].Rmat->to) {
567: old_mesh_level = mesh_level;
568: mesh_level = ml_object->SingleLevel[mesh_level].Rmat->to->levelnum;
570: /* clean and regenerate A */
571: mlmat = &ml_object->Amat[mesh_level];
572: PetscStackCallExternalVoid("ML_Operator_Clean", ML_Operator_Clean(mlmat));
573: PetscStackCallExternalVoid("ML_Operator_Init", ML_Operator_Init(mlmat, ml_object->comm));
574: PetscStackCallExternalVoid("ML_Gen_AmatrixRAP", ML_Gen_AmatrixRAP(ml_object, old_mesh_level, mesh_level));
575: }
577: level = fine_level - 1;
578: if (size == 1) { /* convert ML P, R and A into seqaij format */
579: for (mllevel = 1; mllevel < Nlevels; mllevel++) {
580: mlmat = &ml_object->Amat[mllevel];
581: PetscCall(MatWrapML_SeqAIJ(mlmat, MAT_REUSE_MATRIX, &gridctx[level].A));
582: level--;
583: }
584: } else { /* convert ML P and R into shell format, ML A into mpiaij format */
585: for (mllevel = 1; mllevel < Nlevels; mllevel++) {
586: mlmat = &ml_object->Amat[mllevel];
587: PetscCall(MatWrapML_MPIAIJ(mlmat, MAT_REUSE_MATRIX, &gridctx[level].A));
588: level--;
589: }
590: }
592: for (level = 0; level < fine_level; level++) {
593: if (level > 0) PetscCall(PCMGSetResidual(pc, level, PCMGResidualDefault, gridctx[level].A));
594: PetscCall(KSPSetOperators(gridctx[level].ksp, gridctx[level].A, gridctx[level].A));
595: }
596: PetscCall(PCMGSetResidual(pc, fine_level, PCMGResidualDefault, gridctx[fine_level].A));
597: PetscCall(KSPSetOperators(gridctx[fine_level].ksp, gridctx[level].A, gridctx[fine_level].A));
599: PetscCall(PCSetUp_MG(pc));
600: PetscFunctionReturn(PETSC_SUCCESS);
601: } else {
602: /* since ML can change the size of vectors/matrices at any level we must destroy everything */
603: PetscCall(PCReset_ML(pc));
604: }
605: }
607: /* setup special features of PCML */
608: /* convert A to Aloc to be used by ML at fine grid */
609: pc_ml->size = size;
610: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isSeq));
611: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isMPI));
612: PetscCheck(isMPI || isSeq, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.", ((PetscObject)A)->type_name);
613: if (isMPI) {
614: PetscCall(MatConvert_MPIAIJ_ML(A, NULL, MAT_INITIAL_MATRIX, &Aloc));
615: } else {
616: Aloc = A;
617: PetscCall(PetscObjectReference((PetscObject)Aloc));
618: }
620: /* create and initialize struct 'PetscMLdata' */
621: PetscCall(PetscNew(&PetscMLdata));
622: pc_ml->PetscMLdata = PetscMLdata;
623: PetscCall(PetscMalloc1(Aloc->cmap->n + 1, &PetscMLdata->pwork));
625: PetscCall(MatCreateVecs(Aloc, &PetscMLdata->x, &PetscMLdata->y));
627: PetscMLdata->A = A;
628: PetscMLdata->Aloc = Aloc;
629: if (pc_ml->dim) { /* create vecs around the coordinate data given */
630: PetscInt i, j, dim = pc_ml->dim;
631: PetscInt nloc = pc_ml->nloc, nlocghost;
632: PetscReal *ghostedcoords;
634: PetscCall(MatGetBlockSize(A, &bs));
635: nlocghost = Aloc->cmap->n / bs;
636: PetscCall(PetscMalloc1(dim * nlocghost, &ghostedcoords));
637: for (i = 0; i < dim; i++) {
638: /* copy coordinate values into first component of pwork */
639: for (j = 0; j < nloc; j++) PetscMLdata->pwork[bs * j] = pc_ml->coords[nloc * i + j];
640: /* get the ghost values */
641: PetscCall(PetscML_comm(PetscMLdata->pwork, PetscMLdata));
642: /* write into the vector */
643: for (j = 0; j < nlocghost; j++) ghostedcoords[i * nlocghost + j] = PetscMLdata->pwork[bs * j];
644: }
645: /* replace the original coords with the ghosted coords, because these are
646: * what ML needs */
647: PetscCall(PetscFree(pc_ml->coords));
648: pc_ml->coords = ghostedcoords;
649: }
651: /* create ML discretization matrix at fine grid */
652: /* ML requires input of fine-grid matrix. It determines nlevels. */
653: PetscCall(MatGetSize(Aloc, &m, &nlocal_allcols));
654: PetscCall(MatGetBlockSize(A, &bs));
655: PetscStackCallExternalVoid("ML_Create", ML_Create(&ml_object, pc_ml->MaxNlevels));
656: PetscStackCallExternalVoid("ML_Comm_Set_UsrComm", ML_Comm_Set_UsrComm(ml_object->comm, PetscObjectComm((PetscObject)A)));
657: pc_ml->ml_object = ml_object;
658: PetscStackCallExternalVoid("ML_Init_Amatrix", ML_Init_Amatrix(ml_object, 0, m, m, PetscMLdata));
659: PetscStackCallExternalVoid("ML_Set_Amatrix_Getrow", ML_Set_Amatrix_Getrow(ml_object, 0, PetscML_getrow, ML_PetscML_comm, nlocal_allcols));
660: PetscStackCallExternalVoid("ML_Set_Amatrix_Matvec", ML_Set_Amatrix_Matvec(ml_object, 0, PetscML_matvec));
662: PetscStackCallExternalVoid("ML_Set_Symmetrize", ML_Set_Symmetrize(ml_object, pc_ml->Symmetrize ? ML_YES : ML_NO));
664: /* aggregation */
665: PetscStackCallExternalVoid("ML_Aggregate_Create", ML_Aggregate_Create(&agg_object));
666: pc_ml->agg_object = agg_object;
668: {
669: MatNullSpace mnull;
670: PetscCall(MatGetNearNullSpace(A, &mnull));
671: if (pc_ml->nulltype == PCML_NULLSPACE_AUTO) {
672: if (mnull) pc_ml->nulltype = PCML_NULLSPACE_USER;
673: else if (bs > 1) pc_ml->nulltype = PCML_NULLSPACE_BLOCK;
674: else pc_ml->nulltype = PCML_NULLSPACE_SCALAR;
675: }
676: switch (pc_ml->nulltype) {
677: case PCML_NULLSPACE_USER: {
678: PetscScalar *nullvec;
679: const PetscScalar *v;
680: PetscBool has_const;
681: PetscInt i, j, mlocal, nvec, M;
682: const Vec *vecs;
684: PetscCheck(mnull, PetscObjectComm((PetscObject)pc), PETSC_ERR_USER, "Must provide explicit null space using MatSetNearNullSpace() to use user-specified null space");
685: PetscCall(MatGetSize(A, &M, NULL));
686: PetscCall(MatGetLocalSize(Aloc, &mlocal, NULL));
687: PetscCall(MatNullSpaceGetVecs(mnull, &has_const, &nvec, &vecs));
688: PetscCall(PetscMalloc1((nvec + !!has_const) * mlocal, &nullvec));
689: if (has_const)
690: for (i = 0; i < mlocal; i++) nullvec[i] = 1.0 / M;
691: for (i = 0; i < nvec; i++) {
692: PetscCall(VecGetArrayRead(vecs[i], &v));
693: for (j = 0; j < mlocal; j++) nullvec[(i + !!has_const) * mlocal + j] = v[j];
694: PetscCall(VecRestoreArrayRead(vecs[i], &v));
695: }
696: PetscStackCallExternalVoid("ML_Aggregate_Create", ML_Aggregate_Set_NullSpace(agg_object, bs, nvec + !!has_const, nullvec, mlocal));
697: PetscCall(PetscFree(nullvec));
698: } break;
699: case PCML_NULLSPACE_BLOCK:
700: PetscStackCallExternalVoid("ML_Aggregate_Set_NullSpace", ML_Aggregate_Set_NullSpace(agg_object, bs, bs, 0, 0));
701: break;
702: case PCML_NULLSPACE_SCALAR:
703: break;
704: default:
705: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Unknown null space type");
706: }
707: }
708: PetscStackCallExternalVoid("ML_Aggregate_Set_MaxCoarseSize", ML_Aggregate_Set_MaxCoarseSize(agg_object, pc_ml->MaxCoarseSize));
709: /* set options */
710: switch (pc_ml->CoarsenScheme) {
711: case 1:
712: PetscStackCallExternalVoid("ML_Aggregate_Set_CoarsenScheme_Coupled", ML_Aggregate_Set_CoarsenScheme_Coupled(agg_object));
713: break;
714: case 2:
715: PetscStackCallExternalVoid("ML_Aggregate_Set_CoarsenScheme_MIS", ML_Aggregate_Set_CoarsenScheme_MIS(agg_object));
716: break;
717: case 3:
718: PetscStackCallExternalVoid("ML_Aggregate_Set_CoarsenScheme_METIS", ML_Aggregate_Set_CoarsenScheme_METIS(agg_object));
719: break;
720: }
721: PetscStackCallExternalVoid("ML_Aggregate_Set_Threshold", ML_Aggregate_Set_Threshold(agg_object, pc_ml->Threshold));
722: PetscStackCallExternalVoid("ML_Aggregate_Set_DampingFactor", ML_Aggregate_Set_DampingFactor(agg_object, pc_ml->DampingFactor));
723: if (pc_ml->SpectralNormScheme_Anorm) PetscStackCallExternalVoid("ML_Set_SpectralNormScheme_Anorm", ML_Set_SpectralNormScheme_Anorm(ml_object));
724: agg_object->keep_agg_information = (int)pc_ml->KeepAggInfo;
725: agg_object->keep_P_tentative = (int)pc_ml->Reusable;
726: agg_object->block_scaled_SA = (int)pc_ml->BlockScaling;
727: agg_object->minimizing_energy = (int)pc_ml->EnergyMinimization;
728: agg_object->minimizing_energy_droptol = (double)pc_ml->EnergyMinimizationDropTol;
729: agg_object->cheap_minimizing_energy = (int)pc_ml->EnergyMinimizationCheap;
731: if (pc_ml->Aux) {
732: PetscCheck(pc_ml->dim, PetscObjectComm((PetscObject)pc), PETSC_ERR_USER, "Auxiliary matrix requires coordinates");
733: ml_object->Amat[0].aux_data->threshold = pc_ml->AuxThreshold;
734: ml_object->Amat[0].aux_data->enable = 1;
735: ml_object->Amat[0].aux_data->max_level = 10;
736: ml_object->Amat[0].num_PDEs = bs;
737: }
739: PetscCall(MatGetInfo(A, MAT_LOCAL, &info));
740: ml_object->Amat[0].N_nonzeros = (int)info.nz_used;
742: if (pc_ml->dim) {
743: PetscInt i, dim = pc_ml->dim;
744: ML_Aggregate_Viz_Stats *grid_info;
745: PetscInt nlocghost;
747: PetscCall(MatGetBlockSize(A, &bs));
748: nlocghost = Aloc->cmap->n / bs;
750: PetscStackCallExternalVoid("ML_Aggregate_VizAndStats_Setup(", ML_Aggregate_VizAndStats_Setup(ml_object)); /* create ml info for coords */
751: grid_info = (ML_Aggregate_Viz_Stats *)ml_object->Grid[0].Grid;
752: for (i = 0; i < dim; i++) {
753: /* set the finest level coordinates to point to the column-order array
754: * in pc_ml */
755: /* NOTE: must point away before VizAndStats_Clean so ML doesn't free */
756: switch (i) {
757: case 0:
758: grid_info->x = pc_ml->coords + nlocghost * i;
759: break;
760: case 1:
761: grid_info->y = pc_ml->coords + nlocghost * i;
762: break;
763: case 2:
764: grid_info->z = pc_ml->coords + nlocghost * i;
765: break;
766: default:
767: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_SIZ, "PCML coordinate dimension must be <= 3");
768: }
769: }
770: grid_info->Ndim = dim;
771: }
773: /* repartitioning */
774: if (pc_ml->Repartition) {
775: PetscStackCallExternalVoid("ML_Repartition_Activate", ML_Repartition_Activate(ml_object));
776: PetscStackCallExternalVoid("ML_Repartition_Set_LargestMinMaxRatio", ML_Repartition_Set_LargestMinMaxRatio(ml_object, pc_ml->MaxMinRatio));
777: PetscStackCallExternalVoid("ML_Repartition_Set_MinPerProc", ML_Repartition_Set_MinPerProc(ml_object, pc_ml->MinPerProc));
778: PetscStackCallExternalVoid("ML_Repartition_Set_PutOnSingleProc", ML_Repartition_Set_PutOnSingleProc(ml_object, pc_ml->PutOnSingleProc));
779: #if 0 /* Function not yet defined in ml-6.2 */
780: /* I'm not sure what compatibility issues might crop up if we partitioned
781: * on the finest level, so to be safe repartition starts on the next
782: * finest level (reflection default behavior in
783: * ml_MultiLevelPreconditioner) */
784: PetscStackCallExternalVoid("ML_Repartition_Set_StartLevel",ML_Repartition_Set_StartLevel(ml_object,1));
785: #endif
787: if (!pc_ml->RepartitionType) {
788: PetscInt i;
790: PetscCheck(pc_ml->dim, PetscObjectComm((PetscObject)pc), PETSC_ERR_USER, "ML Zoltan repartitioning requires coordinates");
791: PetscStackCallExternalVoid("ML_Repartition_Set_Partitioner", ML_Repartition_Set_Partitioner(ml_object, ML_USEZOLTAN));
792: PetscStackCallExternalVoid("ML_Aggregate_Set_Dimensions", ML_Aggregate_Set_Dimensions(agg_object, pc_ml->dim));
794: for (i = 0; i < ml_object->ML_num_levels; i++) {
795: ML_Aggregate_Viz_Stats *grid_info = (ML_Aggregate_Viz_Stats *)ml_object->Grid[i].Grid;
796: grid_info->zoltan_type = pc_ml->ZoltanScheme + 1; /* ml numbers options 1, 2, 3 */
797: /* defaults from ml_agg_info.c */
798: grid_info->zoltan_estimated_its = 40; /* only relevant to hypergraph / fast hypergraph */
799: grid_info->zoltan_timers = 0;
800: grid_info->smoothing_steps = 4; /* only relevant to hypergraph / fast hypergraph */
801: }
802: } else {
803: PetscStackCallExternalVoid("ML_Repartition_Set_Partitioner", ML_Repartition_Set_Partitioner(ml_object, ML_USEPARMETIS));
804: }
805: }
807: if (pc_ml->OldHierarchy) {
808: PetscStackCallExternalVoid("ML_Gen_MGHierarchy_UsingAggregation", Nlevels = ML_Gen_MGHierarchy_UsingAggregation(ml_object, 0, ML_INCREASING, agg_object));
809: } else {
810: PetscStackCallExternalVoid("ML_Gen_MultiLevelHierarchy_UsingAggregation", Nlevels = ML_Gen_MultiLevelHierarchy_UsingAggregation(ml_object, 0, ML_INCREASING, agg_object));
811: }
812: PetscCheck(Nlevels > 0, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_OUTOFRANGE, "Nlevels %d must > 0", Nlevels);
813: pc_ml->Nlevels = Nlevels;
814: fine_level = Nlevels - 1;
816: PetscCall(PCMGSetLevels(pc, Nlevels, NULL));
817: /* set default smoothers */
818: for (level = 1; level <= fine_level; level++) {
819: PetscCall(PCMGGetSmoother(pc, level, &smoother));
820: PetscCall(KSPSetType(smoother, KSPRICHARDSON));
821: PetscCall(KSPGetPC(smoother, &subpc));
822: PetscCall(PCSetType(subpc, PCSOR));
823: }
824: PetscObjectOptionsBegin((PetscObject)pc);
825: PetscCall(PCSetFromOptions_MG(pc, PetscOptionsObject)); /* should be called in PCSetFromOptions_ML(), but cannot be called prior to PCMGSetLevels() */
826: PetscOptionsEnd();
828: PetscCall(PetscMalloc1(Nlevels, &gridctx));
830: pc_ml->gridctx = gridctx;
832: /* wrap ML matrices by PETSc shell matrices at coarsened grids.
833: Level 0 is the finest grid for ML, but coarsest for PETSc! */
834: gridctx[fine_level].A = A;
836: level = fine_level - 1;
837: /* TODO: support for GPUs */
838: if (size == 1) { /* convert ML P, R and A into seqaij format */
839: for (mllevel = 1; mllevel < Nlevels; mllevel++) {
840: mlmat = &ml_object->Pmat[mllevel];
841: PetscCall(MatWrapML_SeqAIJ(mlmat, MAT_INITIAL_MATRIX, &gridctx[level].P));
842: mlmat = &ml_object->Rmat[mllevel - 1];
843: PetscCall(MatWrapML_SeqAIJ(mlmat, MAT_INITIAL_MATRIX, &gridctx[level].R));
845: mlmat = &ml_object->Amat[mllevel];
846: PetscCall(MatWrapML_SeqAIJ(mlmat, MAT_INITIAL_MATRIX, &gridctx[level].A));
847: level--;
848: }
849: } else { /* convert ML P and R into shell format, ML A into mpiaij format */
850: for (mllevel = 1; mllevel < Nlevels; mllevel++) {
851: mlmat = &ml_object->Pmat[mllevel];
852: PetscCall(MatWrapML_SHELL(mlmat, MAT_INITIAL_MATRIX, &gridctx[level].P));
853: mlmat = &ml_object->Rmat[mllevel - 1];
854: PetscCall(MatWrapML_SHELL(mlmat, MAT_INITIAL_MATRIX, &gridctx[level].R));
856: mlmat = &ml_object->Amat[mllevel];
857: PetscCall(MatWrapML_MPIAIJ(mlmat, MAT_INITIAL_MATRIX, &gridctx[level].A));
858: level--;
859: }
860: }
862: /* create vectors and ksp at all levels */
863: for (level = 0; level < fine_level; level++) {
864: level1 = level + 1;
866: PetscCall(MatCreateVecs(gridctx[level].A, &gridctx[level].x, &gridctx[level].b));
867: PetscCall(MatCreateVecs(gridctx[level1].A, NULL, &gridctx[level1].r));
868: PetscCall(PCMGSetX(pc, level, gridctx[level].x));
869: PetscCall(PCMGSetRhs(pc, level, gridctx[level].b));
870: PetscCall(PCMGSetR(pc, level1, gridctx[level1].r));
872: if (level == 0) {
873: PetscCall(PCMGGetCoarseSolve(pc, &gridctx[level].ksp));
874: } else {
875: PetscCall(PCMGGetSmoother(pc, level, &gridctx[level].ksp));
876: }
877: }
878: PetscCall(PCMGGetSmoother(pc, fine_level, &gridctx[fine_level].ksp));
880: /* create coarse level and the interpolation between the levels */
881: for (level = 0; level < fine_level; level++) {
882: level1 = level + 1;
884: PetscCall(PCMGSetInterpolation(pc, level1, gridctx[level].P));
885: PetscCall(PCMGSetRestriction(pc, level1, gridctx[level].R));
886: if (level > 0) PetscCall(PCMGSetResidual(pc, level, PCMGResidualDefault, gridctx[level].A));
887: PetscCall(KSPSetOperators(gridctx[level].ksp, gridctx[level].A, gridctx[level].A));
888: }
889: PetscCall(PCMGSetResidual(pc, fine_level, PCMGResidualDefault, gridctx[fine_level].A));
890: PetscCall(KSPSetOperators(gridctx[fine_level].ksp, gridctx[level].A, gridctx[fine_level].A));
892: /* put coordinate info in levels */
893: if (pc_ml->dim) {
894: PetscInt i, j, dim = pc_ml->dim;
895: PetscInt bs, nloc;
896: PC subpc;
897: PetscReal *array;
899: level = fine_level;
900: for (mllevel = 0; mllevel < Nlevels; mllevel++) {
901: ML_Aggregate_Viz_Stats *grid_info = (ML_Aggregate_Viz_Stats *)ml_object->Amat[mllevel].to->Grid->Grid;
902: MPI_Comm comm = ((PetscObject)gridctx[level].A)->comm;
904: PetscCall(MatGetBlockSize(gridctx[level].A, &bs));
905: PetscCall(MatGetLocalSize(gridctx[level].A, NULL, &nloc));
906: nloc /= bs; /* number of local nodes */
908: PetscCall(VecCreate(comm, &gridctx[level].coords));
909: PetscCall(VecSetSizes(gridctx[level].coords, dim * nloc, PETSC_DECIDE));
910: PetscCall(VecSetType(gridctx[level].coords, VECMPI));
911: PetscCall(VecGetArray(gridctx[level].coords, &array));
912: for (j = 0; j < nloc; j++) {
913: for (i = 0; i < dim; i++) {
914: switch (i) {
915: case 0:
916: array[dim * j + i] = grid_info->x[j];
917: break;
918: case 1:
919: array[dim * j + i] = grid_info->y[j];
920: break;
921: case 2:
922: array[dim * j + i] = grid_info->z[j];
923: break;
924: default:
925: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_SIZ, "PCML coordinate dimension must be <= 3");
926: }
927: }
928: }
930: /* passing coordinates to smoothers/coarse solver, should they need them */
931: PetscCall(KSPGetPC(gridctx[level].ksp, &subpc));
932: PetscCall(PCSetCoordinates(subpc, dim, nloc, array));
933: PetscCall(VecRestoreArray(gridctx[level].coords, &array));
934: level--;
935: }
936: }
938: /* setupcalled is set to 0 so that MG is setup from scratch */
939: pc->setupcalled = 0;
940: PetscCall(PCSetUp_MG(pc));
941: PetscFunctionReturn(PETSC_SUCCESS);
942: }
944: /*
945: PCDestroy_ML - Destroys the private context for the ML preconditioner
946: that was created with PCCreate_ML().
948: Input Parameter:
949: . pc - the preconditioner context
951: Application Interface Routine: PCDestroy()
952: */
953: static PetscErrorCode PCDestroy_ML(PC pc)
954: {
955: PC_MG *mg = (PC_MG *)pc->data;
956: PC_ML *pc_ml = (PC_ML *)mg->innerctx;
958: PetscFunctionBegin;
959: PetscCall(PCReset_ML(pc));
960: PetscCall(PetscFree(pc_ml));
961: PetscCall(PCDestroy_MG(pc));
962: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSetCoordinates_C", NULL));
963: PetscFunctionReturn(PETSC_SUCCESS);
964: }
966: static PetscErrorCode PCSetFromOptions_ML(PC pc, PetscOptionItems *PetscOptionsObject)
967: {
968: PetscInt indx, PrintLevel, partindx;
969: const char *scheme[] = {"Uncoupled", "Coupled", "MIS", "METIS"};
970: const char *part[] = {"Zoltan", "ParMETIS"};
971: #if defined(HAVE_ML_ZOLTAN)
972: const char *zscheme[] = {"RCB", "hypergraph", "fast_hypergraph"};
973: #endif
974: PC_MG *mg = (PC_MG *)pc->data;
975: PC_ML *pc_ml = (PC_ML *)mg->innerctx;
976: PetscMPIInt size;
977: MPI_Comm comm;
979: PetscFunctionBegin;
980: PetscCall(PetscObjectGetComm((PetscObject)pc, &comm));
981: PetscCallMPI(MPI_Comm_size(comm, &size));
982: PetscOptionsHeadBegin(PetscOptionsObject, "ML options");
984: PrintLevel = 0;
985: indx = 0;
986: partindx = 0;
988: PetscCall(PetscOptionsInt("-pc_ml_PrintLevel", "Print level", "ML_Set_PrintLevel", PrintLevel, &PrintLevel, NULL));
989: PetscStackCallExternalVoid("ML_Set_PrintLevel", ML_Set_PrintLevel(PrintLevel));
990: PetscCall(PetscOptionsInt("-pc_ml_maxNlevels", "Maximum number of levels", "None", pc_ml->MaxNlevels, &pc_ml->MaxNlevels, NULL));
991: PetscCall(PetscOptionsInt("-pc_ml_maxCoarseSize", "Maximum coarsest mesh size", "ML_Aggregate_Set_MaxCoarseSize", pc_ml->MaxCoarseSize, &pc_ml->MaxCoarseSize, NULL));
992: PetscCall(PetscOptionsEList("-pc_ml_CoarsenScheme", "Aggregate Coarsen Scheme", "ML_Aggregate_Set_CoarsenScheme_*", scheme, 4, scheme[0], &indx, NULL));
994: pc_ml->CoarsenScheme = indx;
996: PetscCall(PetscOptionsReal("-pc_ml_DampingFactor", "P damping factor", "ML_Aggregate_Set_DampingFactor", pc_ml->DampingFactor, &pc_ml->DampingFactor, NULL));
997: PetscCall(PetscOptionsReal("-pc_ml_Threshold", "Smoother drop tol", "ML_Aggregate_Set_Threshold", pc_ml->Threshold, &pc_ml->Threshold, NULL));
998: PetscCall(PetscOptionsBool("-pc_ml_SpectralNormScheme_Anorm", "Method used for estimating spectral radius", "ML_Set_SpectralNormScheme_Anorm", pc_ml->SpectralNormScheme_Anorm, &pc_ml->SpectralNormScheme_Anorm, NULL));
999: PetscCall(PetscOptionsBool("-pc_ml_Symmetrize", "Symmetrize aggregation", "ML_Set_Symmetrize", pc_ml->Symmetrize, &pc_ml->Symmetrize, NULL));
1000: PetscCall(PetscOptionsBool("-pc_ml_BlockScaling", "Scale all dofs at each node together", "None", pc_ml->BlockScaling, &pc_ml->BlockScaling, NULL));
1001: PetscCall(PetscOptionsEnum("-pc_ml_nullspace", "Which type of null space information to use", "None", PCMLNullSpaceTypes, (PetscEnum)pc_ml->nulltype, (PetscEnum *)&pc_ml->nulltype, NULL));
1002: PetscCall(PetscOptionsInt("-pc_ml_EnergyMinimization", "Energy minimization norm type (0=no minimization; see ML manual for 1,2,3; -1 and 4 undocumented)", "None", pc_ml->EnergyMinimization, &pc_ml->EnergyMinimization, NULL));
1003: PetscCall(PetscOptionsBool("-pc_ml_reuse_interpolation", "Reuse the interpolation operators when possible (cheaper, weaker when matrix entries change a lot)", "None", pc_ml->reuse_interpolation, &pc_ml->reuse_interpolation, NULL));
1004: /*
1005: The following checks a number of conditions. If we let this stuff slip by, then ML's error handling will take over.
1006: This is suboptimal because it amounts to calling exit(1) so we check for the most common conditions.
1008: We also try to set some sane defaults when energy minimization is activated, otherwise it's hard to find a working
1009: combination of options and ML's exit(1) explanations don't help matters.
1010: */
1011: PetscCheck(pc_ml->EnergyMinimization >= -1 && pc_ml->EnergyMinimization <= 4, comm, PETSC_ERR_ARG_OUTOFRANGE, "EnergyMinimization must be in range -1..4");
1012: PetscCheck(pc_ml->EnergyMinimization != 4 || size == 1, comm, PETSC_ERR_SUP, "Energy minimization type 4 does not work in parallel");
1013: if (pc_ml->EnergyMinimization == 4) PetscCall(PetscInfo(pc, "Mandel's energy minimization scheme is experimental and broken in ML-6.2\n"));
1014: if (pc_ml->EnergyMinimization) PetscCall(PetscOptionsReal("-pc_ml_EnergyMinimizationDropTol", "Energy minimization drop tolerance", "None", pc_ml->EnergyMinimizationDropTol, &pc_ml->EnergyMinimizationDropTol, NULL));
1015: if (pc_ml->EnergyMinimization == 2) {
1016: /* According to ml_MultiLevelPreconditioner.cpp, this option is only meaningful for norm type (2) */
1017: PetscCall(PetscOptionsBool("-pc_ml_EnergyMinimizationCheap", "Use cheaper variant of norm type 2", "None", pc_ml->EnergyMinimizationCheap, &pc_ml->EnergyMinimizationCheap, NULL));
1018: }
1019: /* energy minimization sometimes breaks if this is turned off, the more classical stuff should be okay without it */
1020: if (pc_ml->EnergyMinimization) pc_ml->KeepAggInfo = PETSC_TRUE;
1021: PetscCall(PetscOptionsBool("-pc_ml_KeepAggInfo", "Allows the preconditioner to be reused, or auxiliary matrices to be generated", "None", pc_ml->KeepAggInfo, &pc_ml->KeepAggInfo, NULL));
1022: /* Option (-1) doesn't work at all (calls exit(1)) if the tentative restriction operator isn't stored. */
1023: if (pc_ml->EnergyMinimization == -1) pc_ml->Reusable = PETSC_TRUE;
1024: PetscCall(PetscOptionsBool("-pc_ml_Reusable", "Store intermedaiate data structures so that the multilevel hierarchy is reusable", "None", pc_ml->Reusable, &pc_ml->Reusable, NULL));
1025: /*
1026: ML's C API is severely underdocumented and lacks significant functionality. The C++ API calls
1027: ML_Gen_MultiLevelHierarchy_UsingAggregation() which is a modified copy (!?) of the documented function
1028: ML_Gen_MGHierarchy_UsingAggregation(). This modification, however, does not provide a strict superset of the
1029: functionality in the old function, so some users may still want to use it. Note that many options are ignored in
1030: this context, but ML doesn't provide a way to find out which ones.
1031: */
1032: PetscCall(PetscOptionsBool("-pc_ml_OldHierarchy", "Use old routine to generate hierarchy", "None", pc_ml->OldHierarchy, &pc_ml->OldHierarchy, NULL));
1033: PetscCall(PetscOptionsBool("-pc_ml_repartition", "Allow ML to repartition levels of the hierarchy", "ML_Repartition_Activate", pc_ml->Repartition, &pc_ml->Repartition, NULL));
1034: if (pc_ml->Repartition) {
1035: PetscCall(PetscOptionsReal("-pc_ml_repartitionMaxMinRatio", "Acceptable ratio of repartitioned sizes", "ML_Repartition_Set_LargestMinMaxRatio", pc_ml->MaxMinRatio, &pc_ml->MaxMinRatio, NULL));
1036: PetscCall(PetscOptionsInt("-pc_ml_repartitionMinPerProc", "Smallest repartitioned size", "ML_Repartition_Set_MinPerProc", pc_ml->MinPerProc, &pc_ml->MinPerProc, NULL));
1037: PetscCall(PetscOptionsInt("-pc_ml_repartitionPutOnSingleProc", "Problem size automatically repartitioned to one processor", "ML_Repartition_Set_PutOnSingleProc", pc_ml->PutOnSingleProc, &pc_ml->PutOnSingleProc, NULL));
1038: #if defined(HAVE_ML_ZOLTAN)
1039: partindx = 0;
1040: PetscCall(PetscOptionsEList("-pc_ml_repartitionType", "Repartitioning library to use", "ML_Repartition_Set_Partitioner", part, 2, part[0], &partindx, NULL));
1042: pc_ml->RepartitionType = partindx;
1043: if (!partindx) {
1044: PetscInt zindx = 0;
1046: PetscCall(PetscOptionsEList("-pc_ml_repartitionZoltanScheme", "Repartitioning scheme to use", "None", zscheme, 3, zscheme[0], &zindx, NULL));
1048: pc_ml->ZoltanScheme = zindx;
1049: }
1050: #else
1051: partindx = 1;
1052: PetscCall(PetscOptionsEList("-pc_ml_repartitionType", "Repartitioning library to use", "ML_Repartition_Set_Partitioner", part, 2, part[1], &partindx, NULL));
1053: pc_ml->RepartitionType = partindx;
1054: PetscCheck(partindx, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP_SYS, "ML not compiled with Zoltan");
1055: #endif
1056: PetscCall(PetscOptionsBool("-pc_ml_Aux", "Aggregate using auxiliary coordinate-based laplacian", "None", pc_ml->Aux, &pc_ml->Aux, NULL));
1057: PetscCall(PetscOptionsReal("-pc_ml_AuxThreshold", "Auxiliary smoother drop tol", "None", pc_ml->AuxThreshold, &pc_ml->AuxThreshold, NULL));
1058: }
1059: PetscOptionsHeadEnd();
1060: PetscFunctionReturn(PETSC_SUCCESS);
1061: }
1063: /*
1064: PCCreate_ML - Creates a ML preconditioner context, PC_ML,
1065: and sets this as the private data within the generic preconditioning
1066: context, PC, that was created within PCCreate().
1068: Input Parameter:
1069: . pc - the preconditioner context
1071: Application Interface Routine: PCCreate()
1072: */
1074: /*MC
1075: PCML - Use the SNL ML algebraic multigrid preconditioner.
1077: Options Database Keys:
1078: Multigrid options(inherited):
1079: + -pc_mg_cycle_type <v> - v for V cycle, w for W-cycle (`PCMGSetCycleType()`)
1080: . -pc_mg_distinct_smoothup - Should one configure the up and down smoothers separately (`PCMGSetDistinctSmoothUp()`)
1081: - -pc_mg_type <multiplicative> - (one of) additive multiplicative full kascade
1083: ML Options Database Key:
1084: + -pc_ml_PrintLevel <0> - Print level (`ML_Set_PrintLevel()`)
1085: . -pc_ml_maxNlevels <10> - Maximum number of levels (None)
1086: . -pc_ml_maxCoarseSize <1> - Maximum coarsest mesh size (`ML_Aggregate_Set_MaxCoarseSize()`)
1087: . -pc_ml_CoarsenScheme <Uncoupled> - (one of) Uncoupled Coupled MIS METIS
1088: . -pc_ml_DampingFactor <1.33333> - P damping factor (`ML_Aggregate_Set_DampingFactor()`)
1089: . -pc_ml_Threshold <0> - Smoother drop tol (`ML_Aggregate_Set_Threshold()`)
1090: . -pc_ml_SpectralNormScheme_Anorm <false> - Method used for estimating spectral radius (`ML_Set_SpectralNormScheme_Anorm()`)
1091: . -pc_ml_repartition <false> - Allow ML to repartition levels of the hierarchy (`ML_Repartition_Activate()`)
1092: . -pc_ml_repartitionMaxMinRatio <1.3> - Acceptable ratio of repartitioned sizes (`ML_Repartition_Set_LargestMinMaxRatio()`)
1093: . -pc_ml_repartitionMinPerProc <512> - Smallest repartitioned size (`ML_Repartition_Set_MinPerProc()`)
1094: . -pc_ml_repartitionPutOnSingleProc <5000> - Problem size automatically repartitioned to one processor (`ML_Repartition_Set_PutOnSingleProc()`)
1095: . -pc_ml_repartitionType <Zoltan> - Repartitioning library to use (`ML_Repartition_Set_Partitioner()`)
1096: . -pc_ml_repartitionZoltanScheme <RCB> - Repartitioning scheme to use (None)
1097: . -pc_ml_Aux <false> - Aggregate using auxiliary coordinate-based Laplacian (None)
1098: - -pc_ml_AuxThreshold <0.0> - Auxiliary smoother drop tol (None)
1100: Level: intermediate
1102: Developer Note:
1103: The coarser grid matrices and restriction/interpolation
1104: operators are computed by ML, with the matrices converted to PETSc matrices in `MATAIJ` format
1105: and the restriction/interpolation operators wrapped as PETSc shell matrices.
1107: .seealso: [](ch_ksp), `PCCreate()`, `PCSetType()`, `PCType`, `PC`, `PCMGType`, `PCMG`, `PCHYPRE`, `PCGAMG`,
1108: `PCMGSetLevels()`, `PCMGGetLevels()`, `PCMGSetType()`, `MPSetCycles()`, `PCMGSetDistinctSmoothUp()`,
1109: `PCMGGetCoarseSolve()`, `PCMGSetResidual()`, `PCMGSetInterpolation()`,
1110: `PCMGSetRestriction()`, `PCMGGetSmoother()`, `PCMGGetSmootherUp()`, `PCMGGetSmootherDown()`,
1111: `PCMGSetCycleTypeOnLevel()`, `PCMGSetRhs()`, `PCMGSetX()`, `PCMGSetR()`
1112: M*/
1114: PETSC_EXTERN PetscErrorCode PCCreate_ML(PC pc)
1115: {
1116: PC_ML *pc_ml;
1117: PC_MG *mg;
1119: PetscFunctionBegin;
1120: /* PCML is an inherited class of PCMG. Initialize pc as PCMG */
1121: PetscCall(PCSetType(pc, PCMG)); /* calls PCCreate_MG() and MGCreate_Private() */
1122: PetscCall(PetscObjectChangeTypeName((PetscObject)pc, PCML));
1123: /* Since PCMG tries to use DM associated with PC must delete it */
1124: PetscCall(DMDestroy(&pc->dm));
1125: PetscCall(PCMGSetGalerkin(pc, PC_MG_GALERKIN_EXTERNAL));
1126: mg = (PC_MG *)pc->data;
1128: /* create a supporting struct and attach it to pc */
1129: PetscCall(PetscNew(&pc_ml));
1130: mg->innerctx = pc_ml;
1132: pc_ml->ml_object = 0;
1133: pc_ml->agg_object = 0;
1134: pc_ml->gridctx = 0;
1135: pc_ml->PetscMLdata = 0;
1136: pc_ml->Nlevels = -1;
1137: pc_ml->MaxNlevels = 10;
1138: pc_ml->MaxCoarseSize = 1;
1139: pc_ml->CoarsenScheme = 1;
1140: pc_ml->Threshold = 0.0;
1141: pc_ml->DampingFactor = 4.0 / 3.0;
1142: pc_ml->SpectralNormScheme_Anorm = PETSC_FALSE;
1143: pc_ml->size = 0;
1144: pc_ml->dim = 0;
1145: pc_ml->nloc = 0;
1146: pc_ml->coords = 0;
1147: pc_ml->Repartition = PETSC_FALSE;
1148: pc_ml->MaxMinRatio = 1.3;
1149: pc_ml->MinPerProc = 512;
1150: pc_ml->PutOnSingleProc = 5000;
1151: pc_ml->RepartitionType = 0;
1152: pc_ml->ZoltanScheme = 0;
1153: pc_ml->Aux = PETSC_FALSE;
1154: pc_ml->AuxThreshold = 0.0;
1156: /* allow for coordinates to be passed */
1157: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSetCoordinates_C", PCSetCoordinates_ML));
1159: /* overwrite the pointers of PCMG by the functions of PCML */
1160: pc->ops->setfromoptions = PCSetFromOptions_ML;
1161: pc->ops->setup = PCSetUp_ML;
1162: pc->ops->reset = PCReset_ML;
1163: pc->ops->destroy = PCDestroy_ML;
1164: PetscFunctionReturn(PETSC_SUCCESS);
1165: }