Actual source code: iispai.c
1: /*
2: 3/99 Modified by Stephen Barnard to support SPAI version 3.0
3: */
5: /*
6: Provides an interface to the SPAI Sparse Approximate Inverse Preconditioner
7: Code written by Stephen Barnard.
9: Note: there is some BAD memory bleeding below!
11: This code needs work
13: 1) get rid of all memory bleeding
14: 2) fix PETSc/interface so that it gets if the matrix is symmetric from the matrix
15: rather than having the sp flag for PC_SPAI
16: 3) fix to set the block size based on the matrix block size
18: */
19: #include <petscmacros.h>
20: #if !PetscDefined(SKIP_COMPLEX)
21: #define PETSC_SKIP_COMPLEX /* since spai uses I which conflicts with some complex implementations */
22: #endif
24: #include <petsc/private/pcimpl.h>
26: /*
27: These are the SPAI include files
28: */
29: EXTERN_C_BEGIN
30: #define SPAI_USE_MPI /* required for setting SPAI_Comm correctly in basics.h */
31: #include <spai.h>
32: #include <matrix.h>
33: EXTERN_C_END
35: static PetscErrorCode ConvertMatToMatrix(MPI_Comm, Mat, Mat, matrix **);
36: static PetscErrorCode ConvertMatrixToMat(MPI_Comm, matrix *, Mat *);
38: typedef struct {
39: matrix *B; /* matrix in SPAI format */
40: matrix *BT; /* transpose of matrix in SPAI format */
41: matrix *M; /* the approximate inverse in SPAI format */
43: Mat PM; /* the approximate inverse PETSc format */
45: double epsilon; /* tolerance */
46: int nbsteps; /* max number of "improvement" steps per line */
47: int max; /* max dimensions of is_I, q, etc. */
48: int maxnew; /* max number of new entries per step */
49: int block_size; /* constant block size */
50: int cache_size; /* one of (1,2,3,4,5,6) indicting size of cache */
51: int verbose; /* SPAI prints timing and statistics */
53: int sp; /* symmetric nonzero pattern */
54: MPI_Comm comm_spai; /* communicator to be used with spai */
55: } PC_SPAI;
57: static PetscErrorCode PCSetUp_SPAI(PC pc)
58: {
59: PC_SPAI *ispai = (PC_SPAI *)pc->data;
60: Mat AT;
62: PetscFunctionBegin;
63: init_SPAI();
65: if (ispai->sp) {
66: PetscCall(ConvertMatToMatrix(ispai->comm_spai, pc->pmat, pc->pmat, &ispai->B));
67: } else {
68: /* Use the transpose to get the column nonzero structure. */
69: PetscCall(MatTranspose(pc->pmat, MAT_INITIAL_MATRIX, &AT));
70: PetscCall(ConvertMatToMatrix(ispai->comm_spai, pc->pmat, AT, &ispai->B));
71: PetscCall(MatDestroy(&AT));
72: }
74: /* Destroy the transpose */
75: /* Don't know how to do it. PETSc developers? */
77: /* construct SPAI preconditioner */
78: /* FILE *messages */ /* file for warning messages */
79: /* double epsilon */ /* tolerance */
80: /* int nbsteps */ /* max number of "improvement" steps per line */
81: /* int max */ /* max dimensions of is_I, q, etc. */
82: /* int maxnew */ /* max number of new entries per step */
83: /* int block_size */ /* block_size == 1 specifies scalar elements
84: block_size == n specifies nxn constant-block elements
85: block_size == 0 specifies variable-block elements */
86: /* int cache_size */ /* one of (1,2,3,4,5,6) indicting size of cache. cache_size == 0 indicates no caching */
87: /* int verbose */ /* verbose == 0 specifies that SPAI is silent
88: verbose == 1 prints timing and matrix statistics */
90: PetscCallExternal(bspai, ispai->B, &ispai->M, stdout, ispai->epsilon, ispai->nbsteps, ispai->max, ispai->maxnew, ispai->block_size, ispai->cache_size, ispai->verbose);
92: PetscCall(ConvertMatrixToMat(PetscObjectComm((PetscObject)pc), ispai->M, &ispai->PM));
94: /* free the SPAI matrices */
95: sp_free_matrix(ispai->B);
96: sp_free_matrix(ispai->M);
97: PetscFunctionReturn(PETSC_SUCCESS);
98: }
100: static PetscErrorCode PCApply_SPAI(PC pc, Vec xx, Vec y)
101: {
102: PC_SPAI *ispai = (PC_SPAI *)pc->data;
104: PetscFunctionBegin;
105: /* Now using PETSc's multiply */
106: PetscCall(MatMult(ispai->PM, xx, y));
107: PetscFunctionReturn(PETSC_SUCCESS);
108: }
110: static PetscErrorCode PCMatApply_SPAI(PC pc, Mat X, Mat Y)
111: {
112: PC_SPAI *ispai = (PC_SPAI *)pc->data;
114: PetscFunctionBegin;
115: /* Now using PETSc's multiply */
116: PetscCall(MatMatMult(ispai->PM, X, MAT_REUSE_MATRIX, PETSC_CURRENT, &Y));
117: PetscFunctionReturn(PETSC_SUCCESS);
118: }
120: static PetscErrorCode PCDestroy_SPAI(PC pc)
121: {
122: PC_SPAI *ispai = (PC_SPAI *)pc->data;
124: PetscFunctionBegin;
125: PetscCall(MatDestroy(&ispai->PM));
126: PetscCallMPI(MPI_Comm_free(&ispai->comm_spai));
127: PetscCall(PetscFree(pc->data));
128: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetEpsilon_C", NULL));
129: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetNBSteps_C", NULL));
130: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetMax_C", NULL));
131: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetMaxNew_C", NULL));
132: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetBlockSize_C", NULL));
133: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetCacheSize_C", NULL));
134: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetVerbose_C", NULL));
135: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetSp_C", NULL));
136: PetscFunctionReturn(PETSC_SUCCESS);
137: }
139: static PetscErrorCode PCView_SPAI(PC pc, PetscViewer viewer)
140: {
141: PC_SPAI *ispai = (PC_SPAI *)pc->data;
142: PetscBool isascii;
144: PetscFunctionBegin;
145: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
146: if (isascii) {
147: PetscCall(PetscViewerASCIIPrintf(viewer, " epsilon %g\n", ispai->epsilon));
148: PetscCall(PetscViewerASCIIPrintf(viewer, " nbsteps %d\n", ispai->nbsteps));
149: PetscCall(PetscViewerASCIIPrintf(viewer, " max %d\n", ispai->max));
150: PetscCall(PetscViewerASCIIPrintf(viewer, " maxnew %d\n", ispai->maxnew));
151: PetscCall(PetscViewerASCIIPrintf(viewer, " block_size %d\n", ispai->block_size));
152: PetscCall(PetscViewerASCIIPrintf(viewer, " cache_size %d\n", ispai->cache_size));
153: PetscCall(PetscViewerASCIIPrintf(viewer, " verbose %d\n", ispai->verbose));
154: PetscCall(PetscViewerASCIIPrintf(viewer, " sp %d\n", ispai->sp));
155: }
156: PetscFunctionReturn(PETSC_SUCCESS);
157: }
159: static PetscErrorCode PCSPAISetEpsilon_SPAI(PC pc, PetscReal epsilon1)
160: {
161: PC_SPAI *ispai = (PC_SPAI *)pc->data;
163: PetscFunctionBegin;
164: ispai->epsilon = (double)epsilon1;
165: PetscFunctionReturn(PETSC_SUCCESS);
166: }
168: static PetscErrorCode PCSPAISetNBSteps_SPAI(PC pc, PetscInt nbsteps1)
169: {
170: PC_SPAI *ispai = (PC_SPAI *)pc->data;
172: PetscFunctionBegin;
173: ispai->nbsteps = (int)nbsteps1;
174: PetscFunctionReturn(PETSC_SUCCESS);
175: }
177: /* added 1/7/99 g.h. */
178: static PetscErrorCode PCSPAISetMax_SPAI(PC pc, PetscInt max1)
179: {
180: PC_SPAI *ispai = (PC_SPAI *)pc->data;
182: PetscFunctionBegin;
183: ispai->max = (int)max1;
184: PetscFunctionReturn(PETSC_SUCCESS);
185: }
187: static PetscErrorCode PCSPAISetMaxNew_SPAI(PC pc, PetscInt maxnew1)
188: {
189: PC_SPAI *ispai = (PC_SPAI *)pc->data;
191: PetscFunctionBegin;
192: ispai->maxnew = (int)maxnew1;
193: PetscFunctionReturn(PETSC_SUCCESS);
194: }
196: static PetscErrorCode PCSPAISetBlockSize_SPAI(PC pc, PetscInt block_size1)
197: {
198: PC_SPAI *ispai = (PC_SPAI *)pc->data;
200: PetscFunctionBegin;
201: ispai->block_size = (int)block_size1;
202: PetscFunctionReturn(PETSC_SUCCESS);
203: }
205: static PetscErrorCode PCSPAISetCacheSize_SPAI(PC pc, PetscInt cache_size)
206: {
207: PC_SPAI *ispai = (PC_SPAI *)pc->data;
209: PetscFunctionBegin;
210: ispai->cache_size = (int)cache_size;
211: PetscFunctionReturn(PETSC_SUCCESS);
212: }
214: static PetscErrorCode PCSPAISetVerbose_SPAI(PC pc, PetscInt verbose)
215: {
216: PC_SPAI *ispai = (PC_SPAI *)pc->data;
218: PetscFunctionBegin;
219: ispai->verbose = (int)verbose;
220: PetscFunctionReturn(PETSC_SUCCESS);
221: }
223: static PetscErrorCode PCSPAISetSp_SPAI(PC pc, PetscInt sp)
224: {
225: PC_SPAI *ispai = (PC_SPAI *)pc->data;
227: PetscFunctionBegin;
228: ispai->sp = (int)sp;
229: PetscFunctionReturn(PETSC_SUCCESS);
230: }
232: static PetscErrorCode PCSetFromOptions_SPAI(PC pc, PetscOptionItems PetscOptionsObject)
233: {
234: PC_SPAI *ispai = (PC_SPAI *)pc->data;
235: int nbsteps1, max1, maxnew1, block_size1, cache_size, verbose, sp;
236: double epsilon1;
237: PetscBool flg;
239: PetscFunctionBegin;
240: PetscOptionsHeadBegin(PetscOptionsObject, "SPAI options");
241: PetscCall(PetscOptionsReal("-pc_spai_epsilon", "", "PCSPAISetEpsilon", ispai->epsilon, &epsilon1, &flg));
242: if (flg) PetscCall(PCSPAISetEpsilon(pc, epsilon1));
243: PetscCall(PetscOptionsInt("-pc_spai_nbsteps", "", "PCSPAISetNBSteps", ispai->nbsteps, &nbsteps1, &flg));
244: if (flg) PetscCall(PCSPAISetNBSteps(pc, nbsteps1));
245: /* added 1/7/99 g.h. */
246: PetscCall(PetscOptionsInt("-pc_spai_max", "", "PCSPAISetMax", ispai->max, &max1, &flg));
247: if (flg) PetscCall(PCSPAISetMax(pc, max1));
248: PetscCall(PetscOptionsInt("-pc_spai_maxnew", "", "PCSPAISetMaxNew", ispai->maxnew, &maxnew1, &flg));
249: if (flg) PetscCall(PCSPAISetMaxNew(pc, maxnew1));
250: PetscCall(PetscOptionsInt("-pc_spai_block_size", "", "PCSPAISetBlockSize", ispai->block_size, &block_size1, &flg));
251: if (flg) PetscCall(PCSPAISetBlockSize(pc, block_size1));
252: PetscCall(PetscOptionsInt("-pc_spai_cache_size", "", "PCSPAISetCacheSize", ispai->cache_size, &cache_size, &flg));
253: if (flg) PetscCall(PCSPAISetCacheSize(pc, cache_size));
254: PetscCall(PetscOptionsInt("-pc_spai_verbose", "", "PCSPAISetVerbose", ispai->verbose, &verbose, &flg));
255: if (flg) PetscCall(PCSPAISetVerbose(pc, verbose));
256: PetscCall(PetscOptionsInt("-pc_spai_sp", "", "PCSPAISetSp", ispai->sp, &sp, &flg));
257: if (flg) PetscCall(PCSPAISetSp(pc, sp));
258: PetscOptionsHeadEnd();
259: PetscFunctionReturn(PETSC_SUCCESS);
260: }
262: /*MC
263: PCSPAI - Use the Sparse Approximate Inverse method {cite}`gh97`
265: Options Database Keys:
266: + -pc_spai_epsilon eps - set tolerance
267: . -pc_spai_nbstep n - set nbsteps
268: . -pc_spai_max m - set max
269: . -pc_spai_max_new m - set maxnew
270: . -pc_spai_block_size n - set block size
271: . -pc_spai_cache_size n - set cache size
272: . -pc_spai_sp m - set sp
273: - -pc_spai_set_verbose (true|false) - verbose output
275: Level: beginner
277: Note:
278: This only works with `MATAIJ` matrices.
280: .seealso: [](ch_ksp), `PCCreate()`, `PCSetType()`, `PCType`, `PC`,
281: `PCSPAISetEpsilon()`, `PCSPAISetMax()`, `PCSPAISetMaxNew()`, `PCSPAISetBlockSize()`,
282: `PCSPAISetVerbose()`, `PCSPAISetSp()`, `PCSPAISetNBSteps()`, `PCSPAISetCacheSize()`
283: M*/
285: PETSC_EXTERN PetscErrorCode PCCreate_SPAI(PC pc)
286: {
287: PC_SPAI *ispai;
289: PetscFunctionBegin;
290: PetscCall(PetscNew(&ispai));
291: pc->data = ispai;
293: pc->ops->destroy = PCDestroy_SPAI;
294: pc->ops->apply = PCApply_SPAI;
295: pc->ops->matapply = PCMatApply_SPAI;
296: pc->ops->applyrichardson = 0;
297: pc->ops->setup = PCSetUp_SPAI;
298: pc->ops->view = PCView_SPAI;
299: pc->ops->setfromoptions = PCSetFromOptions_SPAI;
301: ispai->epsilon = .4;
302: ispai->nbsteps = 5;
303: ispai->max = 5000;
304: ispai->maxnew = 5;
305: ispai->block_size = 1;
306: ispai->cache_size = 5;
307: ispai->verbose = 0;
309: ispai->sp = 1;
310: PetscCallMPI(MPI_Comm_dup(PetscObjectComm((PetscObject)pc), &ispai->comm_spai));
312: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetEpsilon_C", PCSPAISetEpsilon_SPAI));
313: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetNBSteps_C", PCSPAISetNBSteps_SPAI));
314: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetMax_C", PCSPAISetMax_SPAI));
315: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetMaxNew_C", PCSPAISetMaxNew_SPAI));
316: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetBlockSize_C", PCSPAISetBlockSize_SPAI));
317: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetCacheSize_C", PCSPAISetCacheSize_SPAI));
318: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetVerbose_C", PCSPAISetVerbose_SPAI));
319: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCSPAISetSp_C", PCSPAISetSp_SPAI));
320: PetscFunctionReturn(PETSC_SUCCESS);
321: }
323: /*
324: Converts from a PETSc matrix to an SPAI matrix
325: */
326: static PetscErrorCode ConvertMatToMatrix(MPI_Comm comm, Mat A, Mat AT, matrix **B)
327: {
328: matrix *M;
329: int i, j, col;
330: int row_indx;
331: int len, pe, local_indx, start_indx;
332: int *mapping;
333: const int *cols;
334: const double *vals;
335: int n, mnl, nnl, nz, rstart, rend;
336: PetscMPIInt size, rank;
337: struct compressed_lines *rows;
339: PetscFunctionBegin;
340: PetscCallMPI(MPI_Comm_size(comm, &size));
341: PetscCallMPI(MPI_Comm_rank(comm, &rank));
342: PetscCall(MatGetSize(A, &n, &n));
343: PetscCall(MatGetLocalSize(A, &mnl, &nnl));
345: /*
346: not sure why a barrier is required. commenting out
347: PetscCallMPI(MPI_Barrier(comm));
348: */
350: M = new_matrix((SPAI_Comm)comm);
352: M->n = n;
353: M->bs = 1;
354: M->max_block_size = 1;
356: M->mnls = (int *)malloc(sizeof(int) * size);
357: M->start_indices = (int *)malloc(sizeof(int) * size);
358: M->pe = (int *)malloc(sizeof(int) * n);
359: M->block_sizes = (int *)malloc(sizeof(int) * n);
360: for (i = 0; i < n; i++) M->block_sizes[i] = 1;
362: PetscCallMPI(MPI_Allgather(&mnl, 1, MPI_INT, M->mnls, 1, MPI_INT, comm));
364: M->start_indices[0] = 0;
365: for (i = 1; i < size; i++) M->start_indices[i] = M->start_indices[i - 1] + M->mnls[i - 1];
367: M->mnl = M->mnls[M->myid];
368: M->my_start_index = M->start_indices[M->myid];
370: for (i = 0; i < size; i++) {
371: start_indx = M->start_indices[i];
372: for (j = 0; j < M->mnls[i]; j++) M->pe[start_indx + j] = i;
373: }
375: if (AT) {
376: M->lines = new_compressed_lines(M->mnls[rank], 1);
377: } else {
378: M->lines = new_compressed_lines(M->mnls[rank], 0);
379: }
381: rows = M->lines;
383: /* Determine the mapping from global indices to pointers */
384: PetscCall(PetscMalloc1(M->n, &mapping));
385: pe = 0;
386: local_indx = 0;
387: for (i = 0; i < M->n; i++) {
388: if (local_indx >= M->mnls[pe]) {
389: pe++;
390: local_indx = 0;
391: }
392: mapping[i] = local_indx + M->start_indices[pe];
393: local_indx++;
394: }
396: /************** Set up the row structure *****************/
398: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
399: for (i = rstart; i < rend; i++) {
400: row_indx = i - rstart;
401: PetscCall(MatGetRow(A, i, &nz, &cols, &vals));
402: /* allocate buffers */
403: rows->ptrs[row_indx] = (int *)malloc(nz * sizeof(int));
404: rows->A[row_indx] = (double *)malloc(nz * sizeof(double));
405: /* copy the matrix */
406: for (j = 0; j < nz; j++) {
407: col = cols[j];
408: len = rows->len[row_indx]++;
410: rows->ptrs[row_indx][len] = mapping[col];
411: rows->A[row_indx][len] = vals[j];
412: }
413: rows->slen[row_indx] = rows->len[row_indx];
415: PetscCall(MatRestoreRow(A, i, &nz, &cols, &vals));
416: }
418: /************** Set up the column structure *****************/
420: if (AT) {
421: for (i = rstart; i < rend; i++) {
422: row_indx = i - rstart;
423: PetscCall(MatGetRow(AT, i, &nz, &cols, &vals));
424: /* allocate buffers */
425: rows->rptrs[row_indx] = (int *)malloc(nz * sizeof(int));
426: /* copy the matrix (i.e., the structure) */
427: for (j = 0; j < nz; j++) {
428: col = cols[j];
429: len = rows->rlen[row_indx]++;
431: rows->rptrs[row_indx][len] = mapping[col];
432: }
433: PetscCall(MatRestoreRow(AT, i, &nz, &cols, &vals));
434: }
435: }
437: PetscCall(PetscFree(mapping));
439: order_pointers(M);
440: M->maxnz = calc_maxnz(M);
441: *B = M;
442: PetscFunctionReturn(PETSC_SUCCESS);
443: }
445: /*
446: Converts from an SPAI matrix B to a PETSc matrix PB.
447: This assumes that the SPAI matrix B is stored in
448: COMPRESSED-ROW format.
449: */
450: static PetscErrorCode ConvertMatrixToMat(MPI_Comm comm, matrix *B, Mat *PB)
451: {
452: PetscMPIInt size, rank;
453: int m, n, M, N;
454: int d_nz, o_nz;
455: int *d_nnz, *o_nnz;
456: int i, k, global_row, global_col, first_diag_col, last_diag_col;
457: PetscScalar val;
459: PetscFunctionBegin;
460: PetscCallMPI(MPI_Comm_size(comm, &size));
461: PetscCallMPI(MPI_Comm_rank(comm, &rank));
463: m = n = B->mnls[rank];
464: d_nz = o_nz = 0;
466: /* Determine preallocation for MatCreateAIJ */
467: PetscCall(PetscMalloc1(m, &d_nnz));
468: PetscCall(PetscMalloc1(m, &o_nnz));
469: for (i = 0; i < m; i++) d_nnz[i] = o_nnz[i] = 0;
470: first_diag_col = B->start_indices[rank];
471: last_diag_col = first_diag_col + B->mnls[rank];
472: for (i = 0; i < B->mnls[rank]; i++) {
473: for (k = 0; k < B->lines->len[i]; k++) {
474: global_col = B->lines->ptrs[i][k];
475: if ((global_col >= first_diag_col) && (global_col < last_diag_col)) d_nnz[i]++;
476: else o_nnz[i]++;
477: }
478: }
480: M = N = B->n;
481: /* Here we only know how to create AIJ format */
482: PetscCall(MatCreate(comm, PB));
483: PetscCall(MatSetSizes(*PB, m, n, M, N));
484: PetscCall(MatSetType(*PB, MATAIJ));
485: PetscCall(MatSeqAIJSetPreallocation(*PB, d_nz, d_nnz));
486: PetscCall(MatMPIAIJSetPreallocation(*PB, d_nz, d_nnz, o_nz, o_nnz));
488: for (i = 0; i < B->mnls[rank]; i++) {
489: global_row = B->start_indices[rank] + i;
490: for (k = 0; k < B->lines->len[i]; k++) {
491: global_col = B->lines->ptrs[i][k];
493: val = B->lines->A[i][k];
494: PetscCall(MatSetValues(*PB, 1, &global_row, 1, &global_col, &val, ADD_VALUES));
495: }
496: }
497: PetscCall(PetscFree(d_nnz));
498: PetscCall(PetscFree(o_nnz));
500: PetscCall(MatAssemblyBegin(*PB, MAT_FINAL_ASSEMBLY));
501: PetscCall(MatAssemblyEnd(*PB, MAT_FINAL_ASSEMBLY));
502: PetscFunctionReturn(PETSC_SUCCESS);
503: }