Actual source code: imumps.c
1: /*
2: Provides an interface to the MUMPS sparse solver
3: */
4: #include <petscpkg_version.h>
5: #include <petscsf.h>
6: #include <../src/mat/impls/aij/mpi/mpiaij.h>
7: #include <../src/mat/impls/sbaij/mpi/mpisbaij.h>
8: #include <../src/mat/impls/sell/mpi/mpisell.h>
9: #include <petsc/private/vecimpl.h>
11: #define MUMPS_MANUALS "(see users manual https://mumps-solver.org/index.php?page=doc \"Error and warning diagnostics\")"
13: EXTERN_C_BEGIN
14: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
15: #include <cmumps_c.h>
16: #include <zmumps_c.h>
17: #include <smumps_c.h>
18: #include <dmumps_c.h>
19: #else
20: #if defined(PETSC_USE_COMPLEX)
21: #if defined(PETSC_USE_REAL_SINGLE)
22: #include <cmumps_c.h>
23: #define MUMPS_c cmumps_c
24: #define MumpsScalar CMUMPS_COMPLEX
25: #else
26: #include <zmumps_c.h>
27: #define MUMPS_c zmumps_c
28: #define MumpsScalar ZMUMPS_COMPLEX
29: #endif
30: #else
31: #if defined(PETSC_USE_REAL_SINGLE)
32: #include <smumps_c.h>
33: #define MUMPS_c smumps_c
34: #define MumpsScalar SMUMPS_REAL
35: #else
36: #include <dmumps_c.h>
37: #define MUMPS_c dmumps_c
38: #define MumpsScalar DMUMPS_REAL
39: #endif
40: #endif
41: #endif
42: #if defined(PETSC_USE_COMPLEX)
43: #if defined(PETSC_USE_REAL_SINGLE)
44: #define MUMPS_STRUC_C CMUMPS_STRUC_C
45: #else
46: #define MUMPS_STRUC_C ZMUMPS_STRUC_C
47: #endif
48: #else
49: #if defined(PETSC_USE_REAL_SINGLE)
50: #define MUMPS_STRUC_C SMUMPS_STRUC_C
51: #else
52: #define MUMPS_STRUC_C DMUMPS_STRUC_C
53: #endif
54: #endif
55: EXTERN_C_END
57: #define JOB_INIT -1
58: #define JOB_NULL 0
59: #define JOB_FACTSYMBOLIC 1
60: #define JOB_FACTNUMERIC 2
61: #define JOB_SOLVE 3
62: #define JOB_END -2
64: /* MUMPS uses MUMPS_INT for nonzero indices such as irn/jcn, irn_loc/jcn_loc and uses int64_t for
65: number of nonzeros such as nnz, nnz_loc. We typedef MUMPS_INT to PetscMUMPSInt to follow the
66: naming convention in PetscMPIInt, PetscBLASInt etc.
67: */
68: typedef MUMPS_INT PetscMUMPSInt;
70: #if PETSC_PKG_MUMPS_VERSION_GE(5, 3, 0)
71: #if defined(MUMPS_INTSIZE64) /* MUMPS_INTSIZE64 is in MUMPS headers if it is built in full 64-bit mode, therefore the macro is more reliable */
72: #error "PETSc has not been tested with full 64-bit MUMPS and we choose to error out"
73: #endif
74: #else
75: #if defined(INTSIZE64) /* INTSIZE64 is a command line macro one used to build MUMPS in full 64-bit mode */
76: #error "PETSc has not been tested with full 64-bit MUMPS and we choose to error out"
77: #endif
78: #endif
80: #define MPIU_MUMPSINT MPI_INT
81: #define PETSC_MUMPS_INT_MAX 2147483647
82: #define PETSC_MUMPS_INT_MIN -2147483648
84: /* Cast PetscInt to PetscMUMPSInt. Usually there is no overflow since <a> is row/col indices or some small integers*/
85: static inline PetscErrorCode PetscMUMPSIntCast(PetscCount a, PetscMUMPSInt *b)
86: {
87: PetscFunctionBegin;
88: #if PetscDefined(USE_64BIT_INDICES)
89: PetscAssert(a <= PETSC_MUMPS_INT_MAX && a >= PETSC_MUMPS_INT_MIN, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "PetscInt too long for PetscMUMPSInt");
90: #endif
91: *b = (PetscMUMPSInt)a;
92: PetscFunctionReturn(PETSC_SUCCESS);
93: }
95: /* Put these utility routines here since they are only used in this file */
96: static inline PetscErrorCode PetscOptionsMUMPSInt_Private(PetscOptionItems PetscOptionsObject, const char opt[], const char text[], const char man[], PetscMUMPSInt currentvalue, PetscMUMPSInt *value, PetscBool *set, PetscMUMPSInt lb, PetscMUMPSInt ub)
97: {
98: PetscInt myval;
99: PetscBool myset;
101: PetscFunctionBegin;
102: /* PetscInt's size should be always >= PetscMUMPSInt's. It is safe to call PetscOptionsInt_Private to read a PetscMUMPSInt */
103: PetscCall(PetscOptionsInt_Private(PetscOptionsObject, opt, text, man, (PetscInt)currentvalue, &myval, &myset, lb, ub));
104: if (myset) PetscCall(PetscMUMPSIntCast(myval, value));
105: if (set) *set = myset;
106: PetscFunctionReturn(PETSC_SUCCESS);
107: }
108: #define PetscOptionsMUMPSInt(a, b, c, d, e, f) PetscOptionsMUMPSInt_Private(PetscOptionsObject, a, b, c, d, e, f, PETSC_MUMPS_INT_MIN, PETSC_MUMPS_INT_MAX)
110: // An abstract type for specific MUMPS types {S,D,C,Z}MUMPS_STRUC_C.
111: //
112: // With the abstract (outer) type, we can write shared code. We call MUMPS through a type-to-be-determined inner field within the abstract type.
113: // Before/after calling MUMPS, we need to copy in/out fields between the outer and the inner, which seems expensive. But note that the large fixed size
114: // arrays within the types are directly linked. At the end, we only need to copy ~20 integers/pointers, which is doable. See PreMumpsCall()/PostMumpsCall().
115: //
116: // Not all fields in the specific types are exposed in the abstract type. We only need those used by the PETSc/MUMPS interface.
117: // Notably, DMUMPS_COMPLEX* and DMUMPS_REAL* fields are now declared as void *. Their type will be determined by the the actual precision to be used.
118: // Also note that we added some *_len fields not in specific types to track sizes of those MumpsScalar buffers.
119: typedef struct {
120: PetscPrecision precision; // precision used by MUMPS
121: void *internal_id; // the data structure passed to MUMPS, whose actual type {S,D,C,Z}MUMPS_STRUC_C is to be decided by precision and PETSc's use of complex
123: // aliased fields from internal_id, so that we can use XMUMPS_STRUC_C to write shared code across different precisions.
124: MUMPS_INT sym, par, job;
125: MUMPS_INT comm_fortran; /* Fortran communicator */
126: MUMPS_INT *icntl;
127: void *cntl; // MumpsReal, fixed size array
128: MUMPS_INT n;
129: MUMPS_INT nblk;
131: /* Assembled entry */
132: MUMPS_INT8 nnz;
133: MUMPS_INT *irn;
134: MUMPS_INT *jcn;
135: void *a; // MumpsScalar, centralized input
136: PetscCount a_len;
138: /* Distributed entry */
139: MUMPS_INT8 nnz_loc;
140: MUMPS_INT *irn_loc;
141: MUMPS_INT *jcn_loc;
142: void *a_loc; // MumpsScalar, distributed input
143: PetscCount a_loc_len;
145: /* Matrix by blocks */
146: MUMPS_INT *blkptr;
147: MUMPS_INT *blkvar;
149: /* Ordering, if given by user */
150: MUMPS_INT *perm_in;
152: /* RHS, solution, ouptput data and statistics */
153: void *rhs, *redrhs, *rhs_sparse, *sol_loc, *rhs_loc; // MumpsScalar buffers
154: PetscCount rhs_len, redrhs_len, rhs_sparse_len, sol_loc_len, rhs_loc_len; // length of buffers (in MumpsScalar) IF allocated in a different precision than PetscScalar
156: MUMPS_INT *irhs_sparse, *irhs_ptr, *isol_loc, *irhs_loc;
157: MUMPS_INT nrhs, lrhs, lredrhs, nz_rhs, lsol_loc, nloc_rhs, lrhs_loc;
158: // MUMPS_INT nsol_loc; // introduced in MUMPS-5.7, but PETSc doesn't use it; would cause compile errors with the widely used 5.6. If you add it, must also update PreMumpsCall() and guard this with #if PETSC_PKG_MUMPS_VERSION_GE(5, 7, 0)
159: MUMPS_INT schur_lld;
160: MUMPS_INT *info, *infog; // fixed size array
161: void *rinfo, *rinfog; // MumpsReal, fixed size array
163: /* Null space */
164: MUMPS_INT *pivnul_list; // allocated by MUMPS!
165: MUMPS_INT *mapping; // allocated by MUMPS!
167: /* Schur */
168: MUMPS_INT size_schur;
169: MUMPS_INT *listvar_schur;
170: void *schur; // MumpsScalar
171: PetscCount schur_len;
173: /* For out-of-core */
174: char *ooc_tmpdir; // fixed size array
175: char *ooc_prefix; // fixed size array
176: } XMUMPS_STRUC_C;
178: // Note: fixed-size arrays are allocated by MUMPS; redirect them to the outer struct
179: #define AllocateInternalID(MUMPS_STRUC_T, outer) \
180: do { \
181: MUMPS_STRUC_T *inner; \
182: PetscCall(PetscNew(&inner)); \
183: outer->icntl = inner->icntl; \
184: outer->cntl = inner->cntl; \
185: outer->info = inner->info; \
186: outer->infog = inner->infog; \
187: outer->rinfo = inner->rinfo; \
188: outer->rinfog = inner->rinfog; \
189: outer->ooc_tmpdir = inner->ooc_tmpdir; \
190: outer->ooc_prefix = inner->ooc_prefix; \
191: /* the three field should never change after init */ \
192: inner->comm_fortran = outer->comm_fortran; \
193: inner->par = outer->par; \
194: inner->sym = outer->sym; \
195: outer->internal_id = inner; \
196: } while (0)
198: // Allocate the internal [SDCZ]MUMPS_STRUC_C ID data structure in the given , and link fields of the outer and the inner
199: static inline PetscErrorCode MatMumpsAllocateInternalID(XMUMPS_STRUC_C *outer, PetscPrecision precision)
200: {
201: PetscFunctionBegin;
202: outer->precision = precision;
203: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
204: #if defined(PETSC_USE_COMPLEX)
205: if (precision == PETSC_PRECISION_SINGLE) AllocateInternalID(CMUMPS_STRUC_C, outer);
206: else AllocateInternalID(ZMUMPS_STRUC_C, outer);
207: #else
208: if (precision == PETSC_PRECISION_SINGLE) AllocateInternalID(SMUMPS_STRUC_C, outer);
209: else AllocateInternalID(DMUMPS_STRUC_C, outer);
210: #endif
211: #else
212: AllocateInternalID(MUMPS_STRUC_C, outer);
213: #endif
214: PetscFunctionReturn(PETSC_SUCCESS);
215: }
217: #define FreeInternalIDFields(MUMPS_STRUC_T, outer) \
218: do { \
219: MUMPS_STRUC_T *inner = (MUMPS_STRUC_T *)(outer)->internal_id; \
220: PetscCall(PetscFree(inner->a)); \
221: PetscCall(PetscFree(inner->a_loc)); \
222: PetscCall(PetscFree(inner->rhs)); \
223: PetscCall(PetscFree(inner->rhs_sparse)); \
224: PetscCall(PetscFree(inner->rhs_loc)); \
225: PetscCall(PetscFree(inner->sol_loc)); \
226: } while (0)
228: static inline PetscErrorCode MatMumpsFreeInternalID(XMUMPS_STRUC_C *outer)
229: {
230: PetscFunctionBegin;
231: if (outer->internal_id) { // sometimes, the inner is never created before we destroy the outer
232: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
233: const PetscPrecision mumps_precision = outer->precision;
234: if (mumps_precision != PETSC_SCALAR_PRECISION) { // Free internal buffers if we used mixed precision
235: #if defined(PETSC_USE_COMPLEX)
236: if (mumps_precision == PETSC_PRECISION_SINGLE) FreeInternalIDFields(CMUMPS_STRUC_C, outer);
237: else FreeInternalIDFields(ZMUMPS_STRUC_C, outer);
238: #else
239: if (mumps_precision == PETSC_PRECISION_SINGLE) FreeInternalIDFields(SMUMPS_STRUC_C, outer);
240: else FreeInternalIDFields(DMUMPS_STRUC_C, outer);
241: #endif
242: }
243: #endif
244: PetscCall(PetscFree(outer->internal_id));
245: }
246: PetscFunctionReturn(PETSC_SUCCESS);
247: }
249: // Make a companion MumpsScalar array (with a given PetscScalar array), to hold at least MumpsScalars in the given and return the address at .
250: // indicates if we need to convert PetscScalars to MumpsScalars after allocating the MumpsScalar array.
251: // (For brevity, we use for array address and for its length in MumpsScalar, though in code they should be <*ma> and <*m>)
252: // If already points to a buffer/array, on input should be its length. Note the buffer might be freed if it is not big enough for this request.
253: //
254: // The returned array is a companion, so how it is created depends on if PetscScalar and MumpsScalar are the same.
255: // 1) If they are different, a separate array will be made and its length and address will be provided at and on output.
256: // 2) Otherwise, will be returned in , and will be zero on output.
257: //
258: //
259: // Input parameters:
260: // + convert - whether to do PetscScalar to MumpsScalar conversion
261: // . n - length of the PetscScalar array
262: // . pa - [n]], points to the PetscScalar array
263: // . precision - precision of MumpsScalar
264: // . m - on input, length of an existing MumpsScalar array if any, otherwise *m is just zero.
265: // - ma - on input, an existing MumpsScalar array if any.
266: //
267: // Output parameters:
268: // + m - length of the MumpsScalar buffer at if MumpsScalar is different from PetscScalar, otherwise 0
269: // . ma - the MumpsScalar array, which could be an alias of when the two types are the same.
270: //
271: // Note:
272: // New memory, if allocated, is done via PetscMalloc1(), and is owned by caller.
273: static PetscErrorCode MatMumpsMakeMumpsScalarArray(PetscBool convert, PetscCount n, const PetscScalar *pa, PetscPrecision precision, PetscCount *m, void **ma)
274: {
275: PetscFunctionBegin;
276: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
277: const PetscPrecision mumps_precision = precision;
278: PetscCheck(precision == PETSC_PRECISION_SINGLE || precision == PETSC_PRECISION_DOUBLE, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unsupported precicison (%d). Must be single or double", (int)precision);
279: #if defined(PETSC_USE_COMPLEX)
280: if (mumps_precision != PETSC_SCALAR_PRECISION) {
281: if (mumps_precision == PETSC_PRECISION_SINGLE) {
282: if (*m < n) {
283: PetscCall(PetscFree(*ma));
284: PetscCall(PetscMalloc1(n, (CMUMPS_COMPLEX **)ma));
285: *m = n;
286: }
287: if (convert) {
288: CMUMPS_COMPLEX *b = *(CMUMPS_COMPLEX **)ma;
289: for (PetscCount i = 0; i < n; i++) {
290: b[i].r = PetscRealPart(pa[i]);
291: b[i].i = PetscImaginaryPart(pa[i]);
292: }
293: }
294: } else {
295: if (*m < n) {
296: PetscCall(PetscFree(*ma));
297: PetscCall(PetscMalloc1(n, (ZMUMPS_COMPLEX **)ma));
298: *m = n;
299: }
300: if (convert) {
301: ZMUMPS_COMPLEX *b = *(ZMUMPS_COMPLEX **)ma;
302: for (PetscCount i = 0; i < n; i++) {
303: b[i].r = PetscRealPart(pa[i]);
304: b[i].i = PetscImaginaryPart(pa[i]);
305: }
306: }
307: }
308: }
309: #else
310: if (mumps_precision != PETSC_SCALAR_PRECISION) {
311: if (mumps_precision == PETSC_PRECISION_SINGLE) {
312: if (*m < n) {
313: PetscCall(PetscFree(*ma));
314: PetscCall(PetscMalloc1(n, (SMUMPS_REAL **)ma));
315: *m = n;
316: }
317: if (convert) {
318: SMUMPS_REAL *b = *(SMUMPS_REAL **)ma;
319: for (PetscCount i = 0; i < n; i++) b[i] = pa[i];
320: }
321: } else {
322: if (*m < n) {
323: PetscCall(PetscFree(*ma));
324: PetscCall(PetscMalloc1(n, (DMUMPS_REAL **)ma));
325: *m = n;
326: }
327: if (convert) {
328: DMUMPS_REAL *b = *(DMUMPS_REAL **)ma;
329: for (PetscCount i = 0; i < n; i++) b[i] = pa[i];
330: }
331: }
332: }
333: #endif
334: else
335: #endif
336: {
337: if (*m != 0) PetscCall(PetscFree(*ma)); // free existing buffer if any
338: *ma = (void *)pa; // same precision, make them alias
339: *m = 0;
340: }
341: PetscFunctionReturn(PETSC_SUCCESS);
342: }
344: // Cast a MumpsScalar array in to a PetscScalar array at address .
345: //
346: // 1) If the two types are different, cast array elements.
347: // 2) Otherwise, this works as a memcpy; of course, if the two addresses are equal, it is a no-op.
348: static PetscErrorCode MatMumpsCastMumpsScalarArray(PetscCount n, PetscPrecision mumps_precision, const void *ma, PetscScalar *pa)
349: {
350: PetscFunctionBegin;
351: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
352: if (mumps_precision != PETSC_SCALAR_PRECISION) {
353: #if defined(PETSC_USE_COMPLEX)
354: if (mumps_precision == PETSC_PRECISION_SINGLE) {
355: PetscReal *a = (PetscReal *)pa;
356: const SMUMPS_REAL *b = (const SMUMPS_REAL *)ma;
357: for (PetscCount i = 0; i < 2 * n; i++) a[i] = b[i];
358: } else {
359: PetscReal *a = (PetscReal *)pa;
360: const DMUMPS_REAL *b = (const DMUMPS_REAL *)ma;
361: for (PetscCount i = 0; i < 2 * n; i++) a[i] = b[i];
362: }
363: #else
364: if (mumps_precision == PETSC_PRECISION_SINGLE) {
365: const SMUMPS_REAL *b = (const SMUMPS_REAL *)ma;
366: for (PetscCount i = 0; i < n; i++) pa[i] = b[i];
367: } else {
368: const DMUMPS_REAL *b = (const DMUMPS_REAL *)ma;
369: for (PetscCount i = 0; i < n; i++) pa[i] = b[i];
370: }
371: #endif
372: } else
373: #endif
374: PetscCall(PetscArraycpy((PetscScalar *)pa, (PetscScalar *)ma, n));
375: PetscFunctionReturn(PETSC_SUCCESS);
376: }
378: // Cast a PetscScalar array to a MumpsScalar array in the given at address .
379: //
380: // 1) If the two types are different, cast array elements.
381: // 2) Otherwise, this works as a memcpy; of course, if the two addresses are equal, it is a no-op.
382: static PetscErrorCode MatMumpsCastPetscScalarArray(PetscCount n, const PetscScalar *pa, PetscPrecision mumps_precision, const void *ma)
383: {
384: PetscFunctionBegin;
385: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
386: if (mumps_precision != PETSC_SCALAR_PRECISION) {
387: #if defined(PETSC_USE_COMPLEX)
388: if (mumps_precision == PETSC_PRECISION_SINGLE) {
389: CMUMPS_COMPLEX *b = (CMUMPS_COMPLEX *)ma;
390: for (PetscCount i = 0; i < n; i++) {
391: b[i].r = PetscRealPart(pa[i]);
392: b[i].i = PetscImaginaryPart(pa[i]);
393: }
394: } else {
395: ZMUMPS_COMPLEX *b = (ZMUMPS_COMPLEX *)ma;
396: for (PetscCount i = 0; i < n; i++) {
397: b[i].r = PetscRealPart(pa[i]);
398: b[i].i = PetscImaginaryPart(pa[i]);
399: }
400: }
401: #else
402: if (mumps_precision == PETSC_PRECISION_SINGLE) {
403: SMUMPS_REAL *b = (SMUMPS_REAL *)ma;
404: for (PetscCount i = 0; i < n; i++) b[i] = pa[i];
405: } else {
406: DMUMPS_REAL *b = (DMUMPS_REAL *)ma;
407: for (PetscCount i = 0; i < n; i++) b[i] = pa[i];
408: }
409: #endif
410: } else
411: #endif
412: PetscCall(PetscArraycpy((PetscScalar *)ma, (PetscScalar *)pa, n));
413: PetscFunctionReturn(PETSC_SUCCESS);
414: }
416: static inline MPI_Datatype MPIU_MUMPSREAL(const XMUMPS_STRUC_C *id)
417: {
418: return id->precision == PETSC_PRECISION_DOUBLE ? MPI_DOUBLE : MPI_FLOAT;
419: }
421: #define PreMumpsCall(inner, outer, mumpsscalar) \
422: do { \
423: inner->job = outer->job; \
424: inner->n = outer->n; \
425: inner->nblk = outer->nblk; \
426: inner->nnz = outer->nnz; \
427: inner->irn = outer->irn; \
428: inner->jcn = outer->jcn; \
429: inner->a = (mumpsscalar *)outer->a; \
430: inner->nnz_loc = outer->nnz_loc; \
431: inner->irn_loc = outer->irn_loc; \
432: inner->jcn_loc = outer->jcn_loc; \
433: inner->a_loc = (mumpsscalar *)outer->a_loc; \
434: inner->blkptr = outer->blkptr; \
435: inner->blkvar = outer->blkvar; \
436: inner->perm_in = outer->perm_in; \
437: inner->rhs = (mumpsscalar *)outer->rhs; \
438: inner->redrhs = (mumpsscalar *)outer->redrhs; \
439: inner->rhs_sparse = (mumpsscalar *)outer->rhs_sparse; \
440: inner->sol_loc = (mumpsscalar *)outer->sol_loc; \
441: inner->rhs_loc = (mumpsscalar *)outer->rhs_loc; \
442: inner->irhs_sparse = outer->irhs_sparse; \
443: inner->irhs_ptr = outer->irhs_ptr; \
444: inner->isol_loc = outer->isol_loc; \
445: inner->irhs_loc = outer->irhs_loc; \
446: inner->nrhs = outer->nrhs; \
447: inner->lrhs = outer->lrhs; \
448: inner->lredrhs = outer->lredrhs; \
449: inner->nz_rhs = outer->nz_rhs; \
450: inner->lsol_loc = outer->lsol_loc; \
451: inner->nloc_rhs = outer->nloc_rhs; \
452: inner->lrhs_loc = outer->lrhs_loc; \
453: inner->schur_lld = outer->schur_lld; \
454: inner->size_schur = outer->size_schur; \
455: inner->listvar_schur = outer->listvar_schur; \
456: inner->schur = (mumpsscalar *)outer->schur; \
457: } while (0)
459: #define PostMumpsCall(inner, outer) \
460: do { \
461: outer->pivnul_list = inner->pivnul_list; \
462: outer->mapping = inner->mapping; \
463: } while (0)
465: // Entry for PETSc to call mumps
466: static inline PetscErrorCode PetscCallMumps_Private(XMUMPS_STRUC_C *outer)
467: {
468: PetscFunctionBegin;
469: #if defined(PETSC_HAVE_MUMPS_MIXED_PRECISION)
470: #if defined(PETSC_USE_COMPLEX)
471: if (outer->precision == PETSC_PRECISION_SINGLE) {
472: CMUMPS_STRUC_C *inner = (CMUMPS_STRUC_C *)outer->internal_id;
473: PreMumpsCall(inner, outer, CMUMPS_COMPLEX);
474: PetscCallExternalVoid("cmumps_c", cmumps_c(inner));
475: PostMumpsCall(inner, outer);
476: } else {
477: ZMUMPS_STRUC_C *inner = (ZMUMPS_STRUC_C *)outer->internal_id;
478: PreMumpsCall(inner, outer, ZMUMPS_COMPLEX);
479: PetscCallExternalVoid("zmumps_c", zmumps_c(inner));
480: PostMumpsCall(inner, outer);
481: }
482: #else
483: if (outer->precision == PETSC_PRECISION_SINGLE) {
484: SMUMPS_STRUC_C *inner = (SMUMPS_STRUC_C *)outer->internal_id;
485: PreMumpsCall(inner, outer, SMUMPS_REAL);
486: PetscCallExternalVoid("smumps_c", smumps_c(inner));
487: PostMumpsCall(inner, outer);
488: } else {
489: DMUMPS_STRUC_C *inner = (DMUMPS_STRUC_C *)outer->internal_id;
490: PreMumpsCall(inner, outer, DMUMPS_REAL);
491: PetscCallExternalVoid("dmumps_c", dmumps_c(inner));
492: PostMumpsCall(inner, outer);
493: }
494: #endif
495: #else
496: MUMPS_STRUC_C *inner = (MUMPS_STRUC_C *)outer->internal_id;
497: PreMumpsCall(inner, outer, MumpsScalar);
498: PetscCallExternalVoid(PetscStringize(MUMPS_c), MUMPS_c(inner));
499: PostMumpsCall(inner, outer);
500: #endif
501: PetscFunctionReturn(PETSC_SUCCESS);
502: }
504: /* macros s.t. indices match MUMPS documentation */
505: #define ICNTL(I) icntl[(I) - 1]
506: #define INFOG(I) infog[(I) - 1]
507: #define INFO(I) info[(I) - 1]
509: // Get a value from a MumpsScalar array, which is the field in the struct of MUMPS_STRUC_C. The value is convertible to PetscScalar. Note no minus 1 on I!
510: #if defined(PETSC_USE_COMPLEX)
511: #define ID_FIELD_GET(ID, F, I) ((ID).precision == PETSC_PRECISION_SINGLE ? ((CMUMPS_COMPLEX *)(ID).F)[I].r + PETSC_i * ((CMUMPS_COMPLEX *)(ID).F)[I].i : ((ZMUMPS_COMPLEX *)(ID).F)[I].r + PETSC_i * ((ZMUMPS_COMPLEX *)(ID).F)[I].i)
512: #else
513: #define ID_FIELD_GET(ID, F, I) ((ID).precision == PETSC_PRECISION_SINGLE ? ((float *)(ID).F)[I] : ((double *)(ID).F)[I])
514: #endif
516: // Get a value from MumpsReal arrays. The value is convertible to PetscReal.
517: #define ID_CNTL_GET(ID, I) ((ID).precision == PETSC_PRECISION_SINGLE ? ((float *)(ID).cntl)[(I) - 1] : ((double *)(ID).cntl)[(I) - 1])
518: #define ID_RINFOG_GET(ID, I) ((ID).precision == PETSC_PRECISION_SINGLE ? ((float *)(ID).rinfog)[(I) - 1] : ((double *)(ID).rinfog)[(I) - 1])
519: #define ID_RINFO_GET(ID, I) ((ID).precision == PETSC_PRECISION_SINGLE ? ((float *)(ID).rinfo)[(I) - 1] : ((double *)(ID).rinfo)[(I) - 1])
521: // Set the I-th entry of the MumpsReal array id.cntl[] with a PetscReal
522: #define ID_CNTL_SET(ID, I, VAL) \
523: do { \
524: if ((ID).precision == PETSC_PRECISION_SINGLE) ((float *)(ID).cntl)[(I) - 1] = (VAL); \
525: else ((double *)(ID).cntl)[(I) - 1] = (VAL); \
526: } while (0)
528: /* if using PETSc OpenMP support, we only call MUMPS on master ranks. Before/after the call, we change/restore CPUs the master ranks can run on */
529: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
530: #define PetscMUMPS_c(mumps) \
531: do { \
532: if (mumps->use_petsc_omp_support) { \
533: if (mumps->is_omp_master) { \
534: PetscCall(PetscOmpCtrlOmpRegionOnMasterBegin(mumps->omp_ctrl)); \
535: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); \
536: PetscCall(PetscCallMumps_Private(&mumps->id)); \
537: PetscCall(PetscFPTrapPop()); \
538: PetscCall(PetscOmpCtrlOmpRegionOnMasterEnd(mumps->omp_ctrl)); \
539: } \
540: PetscCall(PetscOmpCtrlBarrier(mumps->omp_ctrl)); \
541: /* Global info is same on all processes so we Bcast it within omp_comm. Local info is specific \
542: to processes, so we only Bcast info[1], an error code and leave others (since they do not have \
543: an easy translation between omp_comm and petsc_comm). See MUMPS-5.1.2 manual p82. \
544: omp_comm is a small shared memory communicator, hence doing multiple Bcast as shown below is OK. \
545: */ \
546: MUMPS_STRUC_C tmp; /* All MUMPS_STRUC_C types have same lengths on these info arrays */ \
547: PetscCallMPI(MPI_Bcast(mumps->id.infog, PETSC_STATIC_ARRAY_LENGTH(tmp.infog), MPIU_MUMPSINT, 0, mumps->omp_comm)); \
548: PetscCallMPI(MPI_Bcast(mumps->id.info, PETSC_STATIC_ARRAY_LENGTH(tmp.info), MPIU_MUMPSINT, 0, mumps->omp_comm)); \
549: PetscCallMPI(MPI_Bcast(mumps->id.rinfog, PETSC_STATIC_ARRAY_LENGTH(tmp.rinfog), MPIU_MUMPSREAL(&mumps->id), 0, mumps->omp_comm)); \
550: PetscCallMPI(MPI_Bcast(mumps->id.rinfo, PETSC_STATIC_ARRAY_LENGTH(tmp.rinfo), MPIU_MUMPSREAL(&mumps->id), 0, mumps->omp_comm)); \
551: } else { \
552: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); \
553: PetscCall(PetscCallMumps_Private(&mumps->id)); \
554: PetscCall(PetscFPTrapPop()); \
555: } \
556: } while (0)
557: #else
558: #define PetscMUMPS_c(mumps) \
559: do { \
560: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); \
561: PetscCall(PetscCallMumps_Private(&mumps->id)); \
562: PetscCall(PetscFPTrapPop()); \
563: } while (0)
564: #endif
566: typedef struct Mat_MUMPS Mat_MUMPS;
567: struct Mat_MUMPS {
568: XMUMPS_STRUC_C id;
570: MatStructure matstruc;
571: PetscMPIInt myid, petsc_size;
572: PetscMUMPSInt *irn, *jcn; /* the (i,j,v) triplets passed to mumps. */
573: PetscScalar *val, *val_alloc; /* For some matrices, we can directly access their data array without a buffer. For others, we need a buffer. So comes val_alloc. */
574: PetscCount nnz; /* number of nonzeros. The type is called selective 64-bit in mumps */
575: PetscMUMPSInt sym;
576: MPI_Comm mumps_comm;
577: PetscMUMPSInt *ICNTL_pre;
578: PetscReal *CNTL_pre;
579: PetscMUMPSInt ICNTL9_pre; /* check if ICNTL(9) is changed from previous MatSolve */
580: VecScatter scat_rhs, scat_sol; /* used by MatSolve() */
581: PetscMUMPSInt ICNTL20; /* use centralized (0) or distributed (10) dense RHS */
582: PetscMUMPSInt ICNTL26;
583: PetscMUMPSInt lrhs_loc, nloc_rhs, *irhs_loc;
584: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
585: PetscInt *rhs_nrow, max_nrhs;
586: PetscMPIInt *rhs_recvcounts, *rhs_disps;
587: PetscScalar *rhs_loc, *rhs_recvbuf;
588: #endif
589: Vec b_seq, x_seq;
590: PetscInt ninfo, *info; /* which INFO to display */
591: PetscInt sizeredrhs;
592: PetscScalar *schur_sol;
593: PetscInt schur_sizesol;
594: PetscScalar *redrhs; // buffer in PetscScalar in case MumpsScalar is in a different precision
595: PetscMUMPSInt *ia_alloc, *ja_alloc; /* work arrays used for the CSR struct for sparse rhs */
596: PetscCount cur_ilen, cur_jlen; /* current len of ia_alloc[], ja_alloc[] */
597: PetscErrorCode (*ConvertToTriples)(Mat, PetscInt, MatReuse, Mat_MUMPS *);
599: /* Support for MATNEST */
600: PetscErrorCode (**nest_convert_to_triples)(Mat, PetscInt, MatReuse, Mat_MUMPS *);
601: PetscCount *nest_vals_start;
602: PetscScalar *nest_vals;
604: /* stuff used by petsc/mumps OpenMP support*/
605: PetscBool use_petsc_omp_support;
606: PetscOmpCtrl omp_ctrl; /* an OpenMP controller that blocked processes will release their CPU (MPI_Barrier does not have this guarantee) */
607: MPI_Comm petsc_comm, omp_comm; /* petsc_comm is PETSc matrix's comm */
608: PetscCount *recvcount; /* a collection of nnz on omp_master */
609: PetscMPIInt tag, omp_comm_size;
610: PetscBool is_omp_master; /* is this rank the master of omp_comm */
611: MPI_Request *reqs;
612: };
614: /* Cast a 1-based CSR represented by (nrow, ia, ja) of type PetscInt to a CSR of type PetscMUMPSInt.
615: Here, nrow is number of rows, ia[] is row pointer and ja[] is column indices.
616: */
617: static PetscErrorCode PetscMUMPSIntCSRCast(PETSC_UNUSED Mat_MUMPS *mumps, PetscInt nrow, PetscInt *ia, PetscInt *ja, PetscMUMPSInt **ia_mumps, PetscMUMPSInt **ja_mumps, PetscMUMPSInt *nnz_mumps)
618: {
619: PetscInt nnz = ia[nrow] - 1; /* mumps uses 1-based indices. Uses PetscInt instead of PetscCount since mumps only uses PetscMUMPSInt for rhs */
621: PetscFunctionBegin;
622: #if defined(PETSC_USE_64BIT_INDICES)
623: {
624: if (nrow + 1 > mumps->cur_ilen) { /* realloc ia_alloc/ja_alloc to fit ia/ja */
625: PetscCall(PetscFree(mumps->ia_alloc));
626: PetscCall(PetscMalloc1(nrow + 1, &mumps->ia_alloc));
627: mumps->cur_ilen = nrow + 1;
628: }
629: if (nnz > mumps->cur_jlen) {
630: PetscCall(PetscFree(mumps->ja_alloc));
631: PetscCall(PetscMalloc1(nnz, &mumps->ja_alloc));
632: mumps->cur_jlen = nnz;
633: }
634: for (PetscInt i = 0; i < nrow + 1; i++) PetscCall(PetscMUMPSIntCast(ia[i], &mumps->ia_alloc[i]));
635: for (PetscInt i = 0; i < nnz; i++) PetscCall(PetscMUMPSIntCast(ja[i], &mumps->ja_alloc[i]));
636: *ia_mumps = mumps->ia_alloc;
637: *ja_mumps = mumps->ja_alloc;
638: }
639: #else
640: *ia_mumps = ia;
641: *ja_mumps = ja;
642: #endif
643: PetscCall(PetscMUMPSIntCast(nnz, nnz_mumps));
644: PetscFunctionReturn(PETSC_SUCCESS);
645: }
647: static PetscErrorCode MatMumpsEnsureSchurArray_Private(Mat F)
648: {
649: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
651: PetscFunctionBegin;
652: if (F->schur && !mumps->id.schur) {
653: const PetscScalar *array;
654: PetscCount size = mumps->id.size_schur;
656: PetscCall(MatDenseGetArrayRead(F->schur, &array));
657: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_FALSE, size * size, array, mumps->id.precision, &mumps->id.schur_len, &mumps->id.schur));
658: PetscCall(MatDenseRestoreArrayRead(F->schur, &array));
659: }
660: PetscFunctionReturn(PETSC_SUCCESS);
661: }
663: static PetscErrorCode MatMumpsResetSchur_Private(Mat F)
664: {
665: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
667: PetscFunctionBegin;
668: PetscCall(PetscFree(mumps->id.listvar_schur));
669: PetscCall(PetscFree(mumps->schur_sol));
670: if (mumps->redrhs != mumps->id.redrhs) PetscCall(PetscFree(mumps->id.redrhs));
671: else mumps->id.redrhs = NULL;
672: PetscCall(PetscFree(mumps->redrhs));
673: mumps->id.redrhs_len = 0;
674: mumps->id.schur_len = 0;
675: mumps->id.lredrhs = 0;
676: mumps->sizeredrhs = 0;
677: mumps->id.size_schur = 0;
678: mumps->id.schur_lld = 0;
679: if (mumps->id.internal_id) mumps->id.ICNTL(19) = 0; // sometimes, the inner id is yet built
680: if (F->schur) {
681: const PetscScalar *array;
683: PetscCall(MatDenseGetArrayRead(F->schur, &array));
684: if (array != mumps->id.schur) PetscCall(PetscFree(mumps->id.schur));
685: else mumps->id.schur = NULL;
686: PetscCall(MatDenseRestoreArrayRead(F->schur, &array));
687: }
688: PetscCall(MatDestroy(&F->schur));
689: if (mumps->id.icntl) mumps->id.ICNTL(26) = 0;
690: else mumps->ICNTL26 = 0;
691: PetscFunctionReturn(PETSC_SUCCESS);
692: }
694: /* solve with rhs in mumps->id.redrhs and return in the same location */
695: static PetscErrorCode MatMumpsSolveSchur_Private(Mat F)
696: {
697: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
698: Mat S, B, X; // solve S*X = B; all three matrices are dense
699: MatFactorSchurStatus schurstatus;
700: PetscInt sizesol;
701: const PetscScalar *xarray;
703: PetscFunctionBegin;
704: PetscCall(MatFactorFactorizeSchurComplement(F));
705: PetscCall(MatFactorGetSchurComplement(F, &S, &schurstatus));
706: PetscCall(MatMumpsCastMumpsScalarArray(mumps->sizeredrhs, mumps->id.precision, mumps->id.redrhs, mumps->redrhs));
708: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mumps->id.size_schur, mumps->id.nrhs, mumps->redrhs, &B));
709: PetscCall(MatSetType(B, ((PetscObject)S)->type_name));
710: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
711: PetscCall(MatBindToCPU(B, S->boundtocpu));
712: #endif
713: switch (schurstatus) {
714: case MAT_FACTOR_SCHUR_FACTORED:
715: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mumps->id.size_schur, mumps->id.nrhs, mumps->redrhs, &X));
716: PetscCall(MatSetType(X, ((PetscObject)S)->type_name));
717: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
718: PetscCall(MatBindToCPU(X, S->boundtocpu));
719: #endif
720: if (!mumps->id.ICNTL(9)) { /* transpose solve */
721: PetscCall(MatMatSolveTranspose(S, B, X));
722: } else {
723: PetscCall(MatMatSolve(S, B, X));
724: }
725: break;
726: case MAT_FACTOR_SCHUR_INVERTED:
727: sizesol = mumps->id.nrhs * mumps->id.size_schur;
728: if (!mumps->schur_sol || sizesol > mumps->schur_sizesol) {
729: PetscCall(PetscFree(mumps->schur_sol));
730: PetscCall(PetscMalloc1(sizesol, &mumps->schur_sol));
731: mumps->schur_sizesol = sizesol;
732: }
733: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mumps->id.size_schur, mumps->id.nrhs, mumps->schur_sol, &X));
734: PetscCall(MatSetType(X, ((PetscObject)S)->type_name));
735: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
736: PetscCall(MatBindToCPU(X, S->boundtocpu));
737: #endif
738: PetscCall(MatProductCreateWithMat(S, B, NULL, X));
739: if (!mumps->id.ICNTL(9)) { /* transpose solve */
740: PetscCall(MatProductSetType(X, MATPRODUCT_AtB));
741: } else {
742: PetscCall(MatProductSetType(X, MATPRODUCT_AB));
743: }
744: PetscCall(MatProductSetFromOptions(X));
745: PetscCall(MatProductSymbolic(X));
746: PetscCall(MatProductNumeric(X));
748: PetscCall(MatCopy(X, B, SAME_NONZERO_PATTERN));
749: break;
750: default:
751: SETERRQ(PetscObjectComm((PetscObject)F), PETSC_ERR_SUP, "Unhandled MatFactorSchurStatus %d", F->schur_status);
752: }
753: // MUST get the array from X (not B), though they share the same host array. We can only guarantee X has the correct data on device.
754: PetscCall(MatDenseGetArrayRead(X, &xarray)); // xarray should be mumps->redrhs, but using MatDenseGetArrayRead is safer with GPUs.
755: PetscCall(MatMumpsCastPetscScalarArray(mumps->sizeredrhs, xarray, mumps->id.precision, mumps->id.redrhs));
756: PetscCall(MatDenseRestoreArrayRead(X, &xarray));
757: PetscCall(MatFactorRestoreSchurComplement(F, &S, schurstatus));
758: PetscCall(MatDestroy(&B));
759: PetscCall(MatDestroy(&X));
760: PetscFunctionReturn(PETSC_SUCCESS);
761: }
763: static PetscErrorCode MatMumpsHandleSchur_Private(Mat F, PetscBool expansion)
764: {
765: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
767: PetscFunctionBegin;
768: if (!mumps->id.ICNTL(19)) { /* do nothing when Schur complement has not been computed */
769: PetscFunctionReturn(PETSC_SUCCESS);
770: }
771: if (!expansion) { /* prepare for the condensation step */
772: PetscInt sizeredrhs = mumps->id.nrhs * mumps->id.size_schur;
773: /* allocate MUMPS internal array to store reduced right-hand sides */
774: if (!mumps->id.redrhs || sizeredrhs > mumps->sizeredrhs) {
775: mumps->id.lredrhs = mumps->id.size_schur;
776: mumps->sizeredrhs = mumps->id.nrhs * mumps->id.lredrhs;
777: if (mumps->id.redrhs_len) PetscCall(PetscFree(mumps->id.redrhs));
778: PetscCall(PetscFree(mumps->redrhs));
779: PetscCall(PetscMalloc1(mumps->sizeredrhs, &mumps->redrhs));
780: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_FALSE, mumps->sizeredrhs, mumps->redrhs, mumps->id.precision, &mumps->id.redrhs_len, &mumps->id.redrhs));
781: }
782: } else { /* prepare for the expansion step */
783: PetscCall(MatMumpsSolveSchur_Private(F)); /* solve Schur complement, put solution in id.redrhs (this has to be done by the MUMPS user, so basically us) */
784: mumps->id.ICNTL(26) = 2; /* expansion phase */
785: PetscMUMPS_c(mumps);
786: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in solve: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
787: /* restore defaults */
788: mumps->id.ICNTL(26) = -1;
789: /* free MUMPS internal array for redrhs if we have solved for multiple rhs in order to save memory space */
790: if (mumps->id.nrhs > 1) {
791: if (mumps->redrhs != mumps->id.redrhs) PetscCall(PetscFree(mumps->id.redrhs));
792: else mumps->id.redrhs = NULL;
793: PetscCall(PetscFree(mumps->redrhs));
794: mumps->id.redrhs_len = 0;
795: mumps->id.lredrhs = 0;
796: mumps->sizeredrhs = 0;
797: }
798: }
799: PetscFunctionReturn(PETSC_SUCCESS);
800: }
802: /*
803: MatConvertToTriples_A_B - convert PETSc matrix to triples: row[nz], col[nz], val[nz]
805: input:
806: A - matrix in aij,baij or sbaij format
807: shift - 0: C style output triple; 1: Fortran style output triple.
808: reuse - MAT_INITIAL_MATRIX: spaces are allocated and values are set for the triple
809: MAT_REUSE_MATRIX: only the values in v array are updated
810: output:
811: nnz - dim of r, c, and v (number of local nonzero entries of A)
812: r, c, v - row and col index, matrix values (matrix triples)
814: The returned values r, c, and sometimes v are obtained in a single PetscMalloc(). Then in MatDestroy_MUMPS() it is
815: freed with PetscFree(mumps->irn); This is not ideal code, the fact that v is ONLY sometimes part of mumps->irn means
816: that the PetscMalloc() cannot easily be replaced with a PetscMalloc3().
818: */
820: static PetscErrorCode MatConvertToTriples_seqaij_seqaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
821: {
822: const PetscScalar *av;
823: const PetscInt *ai, *aj, *ajj, M = A->rmap->n;
824: PetscCount nz, rnz, k;
825: PetscMUMPSInt *row, *col;
826: Mat_SeqAIJ *aa = (Mat_SeqAIJ *)A->data;
828: PetscFunctionBegin;
829: PetscCall(MatSeqAIJGetArrayRead(A, &av));
830: if (reuse == MAT_INITIAL_MATRIX) {
831: nz = aa->nz;
832: ai = aa->i;
833: aj = aa->j;
834: PetscCall(PetscMalloc2(nz, &row, nz, &col));
835: for (PetscCount i = k = 0; i < M; i++) {
836: rnz = ai[i + 1] - ai[i];
837: ajj = aj + ai[i];
838: for (PetscCount j = 0; j < rnz; j++) {
839: PetscCall(PetscMUMPSIntCast(i + shift, &row[k]));
840: PetscCall(PetscMUMPSIntCast(ajj[j] + shift, &col[k]));
841: k++;
842: }
843: }
844: mumps->val = (PetscScalar *)av;
845: mumps->irn = row;
846: mumps->jcn = col;
847: mumps->nnz = nz;
848: } else if (mumps->nest_vals) PetscCall(PetscArraycpy(mumps->val, av, aa->nz)); /* MatConvertToTriples_nest_xaij() allocates mumps->val outside of MatConvertToTriples_seqaij_seqaij(), so one needs to copy the memory */
849: else mumps->val = (PetscScalar *)av; /* in the default case, mumps->val is never allocated, one just needs to update the mumps->val pointer */
850: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
851: PetscFunctionReturn(PETSC_SUCCESS);
852: }
854: static PetscErrorCode MatConvertToTriples_seqsell_seqaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
855: {
856: PetscCount nz, i, j, k, r;
857: Mat_SeqSELL *a = (Mat_SeqSELL *)A->data;
858: PetscMUMPSInt *row, *col;
860: PetscFunctionBegin;
861: nz = a->sliidx[a->totalslices];
862: if (reuse == MAT_INITIAL_MATRIX) {
863: PetscCall(PetscMalloc2(nz, &row, nz, &col));
864: for (i = k = 0; i < a->totalslices; i++) {
865: for (j = a->sliidx[i], r = 0; j < a->sliidx[i + 1]; j++, r = ((r + 1) & 0x07)) PetscCall(PetscMUMPSIntCast(8 * i + r + shift, &row[k++]));
866: }
867: for (i = 0; i < nz; i++) PetscCall(PetscMUMPSIntCast(a->colidx[i] + shift, &col[i]));
868: mumps->irn = row;
869: mumps->jcn = col;
870: mumps->nnz = nz;
871: mumps->val = a->val;
872: } else if (mumps->nest_vals) PetscCall(PetscArraycpy(mumps->val, a->val, nz)); /* MatConvertToTriples_nest_xaij() allocates mumps->val outside of MatConvertToTriples_seqsell_seqaij(), so one needs to copy the memory */
873: else mumps->val = a->val; /* in the default case, mumps->val is never allocated, one just needs to update the mumps->val pointer */
874: PetscFunctionReturn(PETSC_SUCCESS);
875: }
877: static PetscErrorCode MatConvertToTriples_seqbaij_seqaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
878: {
879: Mat_SeqBAIJ *aa = (Mat_SeqBAIJ *)A->data;
880: const PetscInt *ai, *aj, *ajj, bs2 = aa->bs2;
881: PetscCount M, nz = bs2 * aa->nz, idx = 0, rnz, i, j, k, m;
882: PetscInt bs;
883: PetscMUMPSInt *row, *col;
885: PetscFunctionBegin;
886: if (reuse == MAT_INITIAL_MATRIX) {
887: PetscCall(MatGetBlockSize(A, &bs));
888: M = A->rmap->N / bs;
889: ai = aa->i;
890: aj = aa->j;
891: PetscCall(PetscMalloc2(nz, &row, nz, &col));
892: for (i = 0; i < M; i++) {
893: ajj = aj + ai[i];
894: rnz = ai[i + 1] - ai[i];
895: for (k = 0; k < rnz; k++) {
896: for (j = 0; j < bs; j++) {
897: for (m = 0; m < bs; m++) {
898: PetscCall(PetscMUMPSIntCast(i * bs + m + shift, &row[idx]));
899: PetscCall(PetscMUMPSIntCast(bs * ajj[k] + j + shift, &col[idx]));
900: idx++;
901: }
902: }
903: }
904: }
905: mumps->irn = row;
906: mumps->jcn = col;
907: mumps->nnz = nz;
908: mumps->val = aa->a;
909: } else if (mumps->nest_vals) PetscCall(PetscArraycpy(mumps->val, aa->a, nz)); /* MatConvertToTriples_nest_xaij() allocates mumps->val outside of MatConvertToTriples_seqbaij_seqaij(), so one needs to copy the memory */
910: else mumps->val = aa->a; /* in the default case, mumps->val is never allocated, one just needs to update the mumps->val pointer */
911: PetscFunctionReturn(PETSC_SUCCESS);
912: }
914: static PetscErrorCode MatConvertToTriples_seqsbaij_seqsbaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
915: {
916: const PetscInt *ai, *aj, *ajj;
917: PetscInt bs;
918: PetscCount nz, rnz, i, j, k, m;
919: PetscMUMPSInt *row, *col;
920: PetscScalar *val;
921: Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ *)A->data;
922: const PetscInt bs2 = aa->bs2, mbs = aa->mbs;
923: #if defined(PETSC_USE_COMPLEX)
924: PetscBool isset, hermitian;
925: #endif
927: PetscFunctionBegin;
928: #if defined(PETSC_USE_COMPLEX)
929: PetscCall(MatIsHermitianKnown(A, &isset, &hermitian));
930: PetscCheck(!isset || !hermitian, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MUMPS does not support Hermitian symmetric matrices for Choleksy");
931: #endif
932: ai = aa->i;
933: aj = aa->j;
934: PetscCall(MatGetBlockSize(A, &bs));
935: if (reuse == MAT_INITIAL_MATRIX) {
936: const PetscCount alloc_size = aa->nz * bs2;
938: PetscCall(PetscMalloc2(alloc_size, &row, alloc_size, &col));
939: if (bs > 1) {
940: PetscCall(PetscMalloc1(alloc_size, &mumps->val_alloc));
941: mumps->val = mumps->val_alloc;
942: } else {
943: mumps->val = aa->a;
944: }
945: mumps->irn = row;
946: mumps->jcn = col;
947: } else {
948: row = mumps->irn;
949: col = mumps->jcn;
950: }
951: val = mumps->val;
953: nz = 0;
954: if (bs > 1) {
955: for (i = 0; i < mbs; i++) {
956: rnz = ai[i + 1] - ai[i];
957: ajj = aj + ai[i];
958: for (j = 0; j < rnz; j++) {
959: for (k = 0; k < bs; k++) {
960: for (m = 0; m < bs; m++) {
961: if (ajj[j] > i || k >= m) {
962: if (reuse == MAT_INITIAL_MATRIX) {
963: PetscCall(PetscMUMPSIntCast(i * bs + m + shift, &row[nz]));
964: PetscCall(PetscMUMPSIntCast(ajj[j] * bs + k + shift, &col[nz]));
965: }
966: val[nz++] = aa->a[(ai[i] + j) * bs2 + m + k * bs];
967: }
968: }
969: }
970: }
971: }
972: } else if (reuse == MAT_INITIAL_MATRIX) {
973: for (i = 0; i < mbs; i++) {
974: rnz = ai[i + 1] - ai[i];
975: ajj = aj + ai[i];
976: for (j = 0; j < rnz; j++) {
977: PetscCall(PetscMUMPSIntCast(i + shift, &row[nz]));
978: PetscCall(PetscMUMPSIntCast(ajj[j] + shift, &col[nz]));
979: nz++;
980: }
981: }
982: PetscCheck(nz == aa->nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Different numbers of nonzeros %" PetscCount_FMT " != %" PetscInt_FMT, nz, aa->nz);
983: } else if (mumps->nest_vals)
984: PetscCall(PetscArraycpy(mumps->val, aa->a, aa->nz)); /* bs == 1 and MAT_REUSE_MATRIX, MatConvertToTriples_nest_xaij() allocates mumps->val outside of MatConvertToTriples_seqsbaij_seqsbaij(), so one needs to copy the memory */
985: else mumps->val = aa->a; /* in the default case, mumps->val is never allocated, one just needs to update the mumps->val pointer */
986: if (reuse == MAT_INITIAL_MATRIX) mumps->nnz = nz;
987: PetscFunctionReturn(PETSC_SUCCESS);
988: }
990: static PetscErrorCode MatConvertToTriples_seqaij_seqsbaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
991: {
992: const PetscInt *ai, *aj, *ajj, *adiag, M = A->rmap->n;
993: PetscCount nz, rnz, i, j;
994: const PetscScalar *av, *v1;
995: PetscScalar *val;
996: PetscMUMPSInt *row, *col;
997: Mat_SeqAIJ *aa = (Mat_SeqAIJ *)A->data;
998: PetscBool diagDense;
999: #if defined(PETSC_USE_COMPLEX)
1000: PetscBool hermitian, isset;
1001: #endif
1003: PetscFunctionBegin;
1004: #if defined(PETSC_USE_COMPLEX)
1005: PetscCall(MatIsHermitianKnown(A, &isset, &hermitian));
1006: PetscCheck(!isset || !hermitian, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MUMPS does not support Hermitian symmetric matrices for Choleksy");
1007: #endif
1008: PetscCall(MatSeqAIJGetArrayRead(A, &av));
1009: ai = aa->i;
1010: aj = aa->j;
1011: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, &diagDense));
1012: if (reuse == MAT_INITIAL_MATRIX) {
1013: /* count nz in the upper triangular part of A */
1014: nz = 0;
1015: if (!diagDense) {
1016: for (i = 0; i < M; i++) {
1017: if (PetscUnlikely(adiag[i] >= ai[i + 1])) {
1018: for (j = ai[i]; j < ai[i + 1]; j++) {
1019: if (aj[j] < i) continue;
1020: nz++;
1021: }
1022: } else {
1023: nz += ai[i + 1] - adiag[i];
1024: }
1025: }
1026: } else {
1027: for (i = 0; i < M; i++) nz += ai[i + 1] - adiag[i];
1028: }
1029: PetscCall(PetscMalloc2(nz, &row, nz, &col));
1030: PetscCall(PetscMalloc1(nz, &val));
1031: mumps->nnz = nz;
1032: mumps->irn = row;
1033: mumps->jcn = col;
1034: mumps->val = mumps->val_alloc = val;
1036: nz = 0;
1037: if (!diagDense) {
1038: for (i = 0; i < M; i++) {
1039: if (PetscUnlikely(adiag[i] >= ai[i + 1])) {
1040: for (j = ai[i]; j < ai[i + 1]; j++) {
1041: if (aj[j] < i) continue;
1042: PetscCall(PetscMUMPSIntCast(i + shift, &row[nz]));
1043: PetscCall(PetscMUMPSIntCast(aj[j] + shift, &col[nz]));
1044: val[nz] = av[j];
1045: nz++;
1046: }
1047: } else {
1048: rnz = ai[i + 1] - adiag[i];
1049: ajj = aj + adiag[i];
1050: v1 = av + adiag[i];
1051: for (j = 0; j < rnz; j++) {
1052: PetscCall(PetscMUMPSIntCast(i + shift, &row[nz]));
1053: PetscCall(PetscMUMPSIntCast(ajj[j] + shift, &col[nz]));
1054: val[nz++] = v1[j];
1055: }
1056: }
1057: }
1058: } else {
1059: for (i = 0; i < M; i++) {
1060: rnz = ai[i + 1] - adiag[i];
1061: ajj = aj + adiag[i];
1062: v1 = av + adiag[i];
1063: for (j = 0; j < rnz; j++) {
1064: PetscCall(PetscMUMPSIntCast(i + shift, &row[nz]));
1065: PetscCall(PetscMUMPSIntCast(ajj[j] + shift, &col[nz]));
1066: val[nz++] = v1[j];
1067: }
1068: }
1069: }
1070: } else {
1071: nz = 0;
1072: val = mumps->val;
1073: if (!diagDense) {
1074: for (i = 0; i < M; i++) {
1075: if (PetscUnlikely(adiag[i] >= ai[i + 1])) {
1076: for (j = ai[i]; j < ai[i + 1]; j++) {
1077: if (aj[j] < i) continue;
1078: val[nz++] = av[j];
1079: }
1080: } else {
1081: rnz = ai[i + 1] - adiag[i];
1082: v1 = av + adiag[i];
1083: for (j = 0; j < rnz; j++) val[nz++] = v1[j];
1084: }
1085: }
1086: } else {
1087: for (i = 0; i < M; i++) {
1088: rnz = ai[i + 1] - adiag[i];
1089: v1 = av + adiag[i];
1090: for (j = 0; j < rnz; j++) val[nz++] = v1[j];
1091: }
1092: }
1093: }
1094: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
1095: PetscFunctionReturn(PETSC_SUCCESS);
1096: }
1098: static PetscErrorCode MatConvertToTriples_mpisbaij_mpisbaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1099: {
1100: const PetscInt *ai, *aj, *bi, *bj, *garray, *ajj, *bjj;
1101: PetscInt bs;
1102: PetscCount rstart, nz, i, j, k, m, jj, irow, countA, countB;
1103: PetscMUMPSInt *row, *col;
1104: const PetscScalar *av, *bv, *v1, *v2;
1105: PetscScalar *val;
1106: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ *)A->data;
1107: Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ *)mat->A->data;
1108: Mat_SeqBAIJ *bb = (Mat_SeqBAIJ *)mat->B->data;
1109: const PetscInt bs2 = aa->bs2, mbs = aa->mbs;
1110: #if defined(PETSC_USE_COMPLEX)
1111: PetscBool hermitian, isset;
1112: #endif
1114: PetscFunctionBegin;
1115: #if defined(PETSC_USE_COMPLEX)
1116: PetscCall(MatIsHermitianKnown(A, &isset, &hermitian));
1117: PetscCheck(!isset || !hermitian, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MUMPS does not support Hermitian symmetric matrices for Choleksy");
1118: #endif
1119: PetscCall(MatGetBlockSize(A, &bs));
1120: rstart = A->rmap->rstart;
1121: ai = aa->i;
1122: aj = aa->j;
1123: bi = bb->i;
1124: bj = bb->j;
1125: av = aa->a;
1126: bv = bb->a;
1128: garray = mat->garray;
1130: if (reuse == MAT_INITIAL_MATRIX) {
1131: nz = (aa->nz + bb->nz) * bs2; /* just a conservative estimate */
1132: PetscCall(PetscMalloc2(nz, &row, nz, &col));
1133: PetscCall(PetscMalloc1(nz, &val));
1134: /* can not decide the exact mumps->nnz now because of the SBAIJ */
1135: mumps->irn = row;
1136: mumps->jcn = col;
1137: mumps->val = mumps->val_alloc = val;
1138: } else {
1139: val = mumps->val;
1140: }
1142: jj = 0;
1143: irow = rstart;
1144: for (i = 0; i < mbs; i++) {
1145: ajj = aj + ai[i]; /* ptr to the beginning of this row */
1146: countA = ai[i + 1] - ai[i];
1147: countB = bi[i + 1] - bi[i];
1148: bjj = bj + bi[i];
1149: v1 = av + ai[i] * bs2;
1150: v2 = bv + bi[i] * bs2;
1152: if (bs > 1) {
1153: /* A-part */
1154: for (j = 0; j < countA; j++) {
1155: for (k = 0; k < bs; k++) {
1156: for (m = 0; m < bs; m++) {
1157: if (rstart + ajj[j] * bs > irow || k >= m) {
1158: if (reuse == MAT_INITIAL_MATRIX) {
1159: PetscCall(PetscMUMPSIntCast(irow + m + shift, &row[jj]));
1160: PetscCall(PetscMUMPSIntCast(rstart + ajj[j] * bs + k + shift, &col[jj]));
1161: }
1162: val[jj++] = v1[j * bs2 + m + k * bs];
1163: }
1164: }
1165: }
1166: }
1168: /* B-part */
1169: for (j = 0; j < countB; j++) {
1170: for (k = 0; k < bs; k++) {
1171: for (m = 0; m < bs; m++) {
1172: if (reuse == MAT_INITIAL_MATRIX) {
1173: PetscCall(PetscMUMPSIntCast(irow + m + shift, &row[jj]));
1174: PetscCall(PetscMUMPSIntCast(garray[bjj[j]] * bs + k + shift, &col[jj]));
1175: }
1176: val[jj++] = v2[j * bs2 + m + k * bs];
1177: }
1178: }
1179: }
1180: } else {
1181: /* A-part */
1182: for (j = 0; j < countA; j++) {
1183: if (reuse == MAT_INITIAL_MATRIX) {
1184: PetscCall(PetscMUMPSIntCast(irow + shift, &row[jj]));
1185: PetscCall(PetscMUMPSIntCast(rstart + ajj[j] + shift, &col[jj]));
1186: }
1187: val[jj++] = v1[j];
1188: }
1190: /* B-part */
1191: for (j = 0; j < countB; j++) {
1192: if (reuse == MAT_INITIAL_MATRIX) {
1193: PetscCall(PetscMUMPSIntCast(irow + shift, &row[jj]));
1194: PetscCall(PetscMUMPSIntCast(garray[bjj[j]] + shift, &col[jj]));
1195: }
1196: val[jj++] = v2[j];
1197: }
1198: }
1199: irow += bs;
1200: }
1201: if (reuse == MAT_INITIAL_MATRIX) mumps->nnz = jj;
1202: PetscFunctionReturn(PETSC_SUCCESS);
1203: }
1205: static PetscErrorCode MatConvertToTriples_mpiaij_mpiaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1206: {
1207: const PetscInt *ai, *aj, *bi, *bj, *garray, m = A->rmap->n, *ajj, *bjj;
1208: PetscCount rstart, cstart, nz, i, j, jj, irow, countA, countB;
1209: PetscMUMPSInt *row, *col;
1210: const PetscScalar *av, *bv, *v1, *v2;
1211: PetscScalar *val;
1212: Mat Ad, Ao;
1213: Mat_SeqAIJ *aa;
1214: Mat_SeqAIJ *bb;
1216: PetscFunctionBegin;
1217: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &garray));
1218: PetscCall(MatSeqAIJGetArrayRead(Ad, &av));
1219: PetscCall(MatSeqAIJGetArrayRead(Ao, &bv));
1221: aa = (Mat_SeqAIJ *)Ad->data;
1222: bb = (Mat_SeqAIJ *)Ao->data;
1223: ai = aa->i;
1224: aj = aa->j;
1225: bi = bb->i;
1226: bj = bb->j;
1228: rstart = A->rmap->rstart;
1229: cstart = A->cmap->rstart;
1231: if (reuse == MAT_INITIAL_MATRIX) {
1232: nz = (PetscCount)aa->nz + bb->nz; /* make sure the sum won't overflow PetscInt */
1233: PetscCall(PetscMalloc2(nz, &row, nz, &col));
1234: PetscCall(PetscMalloc1(nz, &val));
1235: mumps->nnz = nz;
1236: mumps->irn = row;
1237: mumps->jcn = col;
1238: mumps->val = mumps->val_alloc = val;
1239: } else {
1240: val = mumps->val;
1241: }
1243: jj = 0;
1244: irow = rstart;
1245: for (i = 0; i < m; i++) {
1246: ajj = aj + ai[i]; /* ptr to the beginning of this row */
1247: countA = ai[i + 1] - ai[i];
1248: countB = bi[i + 1] - bi[i];
1249: bjj = bj + bi[i];
1250: v1 = av + ai[i];
1251: v2 = bv + bi[i];
1253: /* A-part */
1254: for (j = 0; j < countA; j++) {
1255: if (reuse == MAT_INITIAL_MATRIX) {
1256: PetscCall(PetscMUMPSIntCast(irow + shift, &row[jj]));
1257: PetscCall(PetscMUMPSIntCast(cstart + ajj[j] + shift, &col[jj]));
1258: }
1259: val[jj++] = v1[j];
1260: }
1262: /* B-part */
1263: for (j = 0; j < countB; j++) {
1264: if (reuse == MAT_INITIAL_MATRIX) {
1265: PetscCall(PetscMUMPSIntCast(irow + shift, &row[jj]));
1266: PetscCall(PetscMUMPSIntCast(garray[bjj[j]] + shift, &col[jj]));
1267: }
1268: val[jj++] = v2[j];
1269: }
1270: irow++;
1271: }
1272: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &av));
1273: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &bv));
1274: PetscFunctionReturn(PETSC_SUCCESS);
1275: }
1277: static PetscErrorCode MatConvertToTriples_mpibaij_mpiaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1278: {
1279: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ *)A->data;
1280: Mat_SeqBAIJ *aa = (Mat_SeqBAIJ *)mat->A->data;
1281: Mat_SeqBAIJ *bb = (Mat_SeqBAIJ *)mat->B->data;
1282: const PetscInt *ai = aa->i, *bi = bb->i, *aj = aa->j, *bj = bb->j, *ajj, *bjj;
1283: const PetscInt *garray = mat->garray, mbs = mat->mbs, rstart = A->rmap->rstart, cstart = A->cmap->rstart;
1284: const PetscInt bs2 = mat->bs2;
1285: PetscInt bs;
1286: PetscCount nz, i, j, k, n, jj, irow, countA, countB, idx;
1287: PetscMUMPSInt *row, *col;
1288: const PetscScalar *av = aa->a, *bv = bb->a, *v1, *v2;
1289: PetscScalar *val;
1291: PetscFunctionBegin;
1292: PetscCall(MatGetBlockSize(A, &bs));
1293: if (reuse == MAT_INITIAL_MATRIX) {
1294: nz = bs2 * (aa->nz + bb->nz);
1295: PetscCall(PetscMalloc2(nz, &row, nz, &col));
1296: PetscCall(PetscMalloc1(nz, &val));
1297: mumps->nnz = nz;
1298: mumps->irn = row;
1299: mumps->jcn = col;
1300: mumps->val = mumps->val_alloc = val;
1301: } else {
1302: val = mumps->val;
1303: }
1305: jj = 0;
1306: irow = rstart;
1307: for (i = 0; i < mbs; i++) {
1308: countA = ai[i + 1] - ai[i];
1309: countB = bi[i + 1] - bi[i];
1310: ajj = aj + ai[i];
1311: bjj = bj + bi[i];
1312: v1 = av + bs2 * ai[i];
1313: v2 = bv + bs2 * bi[i];
1315: idx = 0;
1316: /* A-part */
1317: for (k = 0; k < countA; k++) {
1318: for (j = 0; j < bs; j++) {
1319: for (n = 0; n < bs; n++) {
1320: if (reuse == MAT_INITIAL_MATRIX) {
1321: PetscCall(PetscMUMPSIntCast(irow + n + shift, &row[jj]));
1322: PetscCall(PetscMUMPSIntCast(cstart + bs * ajj[k] + j + shift, &col[jj]));
1323: }
1324: val[jj++] = v1[idx++];
1325: }
1326: }
1327: }
1329: idx = 0;
1330: /* B-part */
1331: for (k = 0; k < countB; k++) {
1332: for (j = 0; j < bs; j++) {
1333: for (n = 0; n < bs; n++) {
1334: if (reuse == MAT_INITIAL_MATRIX) {
1335: PetscCall(PetscMUMPSIntCast(irow + n + shift, &row[jj]));
1336: PetscCall(PetscMUMPSIntCast(bs * garray[bjj[k]] + j + shift, &col[jj]));
1337: }
1338: val[jj++] = v2[idx++];
1339: }
1340: }
1341: }
1342: irow += bs;
1343: }
1344: PetscFunctionReturn(PETSC_SUCCESS);
1345: }
1347: static PetscErrorCode MatConvertToTriples_mpiaij_mpisbaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1348: {
1349: const PetscInt *ai, *aj, *adiag, *bi, *bj, *garray, m = A->rmap->n, *ajj, *bjj;
1350: PetscCount rstart, nz, nza, nzb, i, j, jj, irow, countA, countB;
1351: PetscMUMPSInt *row, *col;
1352: const PetscScalar *av, *bv, *v1, *v2;
1353: PetscScalar *val;
1354: Mat Ad, Ao;
1355: Mat_SeqAIJ *aa;
1356: Mat_SeqAIJ *bb;
1357: #if defined(PETSC_USE_COMPLEX)
1358: PetscBool hermitian, isset;
1359: #endif
1361: PetscFunctionBegin;
1362: #if defined(PETSC_USE_COMPLEX)
1363: PetscCall(MatIsHermitianKnown(A, &isset, &hermitian));
1364: PetscCheck(!isset || !hermitian, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MUMPS does not support Hermitian symmetric matrices for Choleksy");
1365: #endif
1366: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &garray));
1367: PetscCall(MatSeqAIJGetArrayRead(Ad, &av));
1368: PetscCall(MatSeqAIJGetArrayRead(Ao, &bv));
1370: aa = (Mat_SeqAIJ *)Ad->data;
1371: bb = (Mat_SeqAIJ *)Ao->data;
1372: ai = aa->i;
1373: aj = aa->j;
1374: bi = bb->i;
1375: bj = bb->j;
1376: PetscCall(MatGetDiagonalMarkers_SeqAIJ(Ad, &adiag, NULL));
1377: rstart = A->rmap->rstart;
1379: if (reuse == MAT_INITIAL_MATRIX) {
1380: nza = 0; /* num of upper triangular entries in mat->A, including diagonals */
1381: nzb = 0; /* num of upper triangular entries in mat->B */
1382: for (i = 0; i < m; i++) {
1383: nza += (ai[i + 1] - adiag[i]);
1384: countB = bi[i + 1] - bi[i];
1385: bjj = bj + bi[i];
1386: for (j = 0; j < countB; j++) {
1387: if (garray[bjj[j]] > rstart) nzb++;
1388: }
1389: }
1391: nz = nza + nzb; /* total nz of upper triangular part of mat */
1392: PetscCall(PetscMalloc2(nz, &row, nz, &col));
1393: PetscCall(PetscMalloc1(nz, &val));
1394: mumps->nnz = nz;
1395: mumps->irn = row;
1396: mumps->jcn = col;
1397: mumps->val = mumps->val_alloc = val;
1398: } else {
1399: val = mumps->val;
1400: }
1402: jj = 0;
1403: irow = rstart;
1404: for (i = 0; i < m; i++) {
1405: ajj = aj + adiag[i]; /* ptr to the beginning of the diagonal of this row */
1406: v1 = av + adiag[i];
1407: countA = ai[i + 1] - adiag[i];
1408: countB = bi[i + 1] - bi[i];
1409: bjj = bj + bi[i];
1410: v2 = bv + bi[i];
1412: /* A-part */
1413: for (j = 0; j < countA; j++) {
1414: if (reuse == MAT_INITIAL_MATRIX) {
1415: PetscCall(PetscMUMPSIntCast(irow + shift, &row[jj]));
1416: PetscCall(PetscMUMPSIntCast(rstart + ajj[j] + shift, &col[jj]));
1417: }
1418: val[jj++] = v1[j];
1419: }
1421: /* B-part */
1422: for (j = 0; j < countB; j++) {
1423: if (garray[bjj[j]] > rstart) {
1424: if (reuse == MAT_INITIAL_MATRIX) {
1425: PetscCall(PetscMUMPSIntCast(irow + shift, &row[jj]));
1426: PetscCall(PetscMUMPSIntCast(garray[bjj[j]] + shift, &col[jj]));
1427: }
1428: val[jj++] = v2[j];
1429: }
1430: }
1431: irow++;
1432: }
1433: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &av));
1434: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &bv));
1435: PetscFunctionReturn(PETSC_SUCCESS);
1436: }
1438: static PetscErrorCode MatConvertToTriples_diagonal_xaij(Mat A, PETSC_UNUSED PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1439: {
1440: const PetscScalar *av;
1441: const PetscInt M = A->rmap->n;
1442: PetscCount i;
1443: PetscMUMPSInt *row, *col;
1444: Vec v;
1446: PetscFunctionBegin;
1447: PetscCall(MatDiagonalGetDiagonal(A, &v));
1448: PetscCall(VecGetArrayRead(v, &av));
1449: if (reuse == MAT_INITIAL_MATRIX) {
1450: PetscCall(PetscMalloc2(M, &row, M, &col));
1451: for (i = 0; i < M; i++) {
1452: PetscCall(PetscMUMPSIntCast(i + A->rmap->rstart, &row[i]));
1453: col[i] = row[i];
1454: }
1455: mumps->val = (PetscScalar *)av;
1456: mumps->irn = row;
1457: mumps->jcn = col;
1458: mumps->nnz = M;
1459: } else if (mumps->nest_vals) PetscCall(PetscArraycpy(mumps->val, av, M)); /* MatConvertToTriples_nest_xaij() allocates mumps->val outside of MatConvertToTriples_diagonal_xaij(), so one needs to copy the memory */
1460: else mumps->val = (PetscScalar *)av; /* in the default case, mumps->val is never allocated, one just needs to update the mumps->val pointer */
1461: PetscCall(VecRestoreArrayRead(v, &av));
1462: PetscFunctionReturn(PETSC_SUCCESS);
1463: }
1465: static PetscErrorCode MatConvertToTriples_dense_xaij(Mat A, PETSC_UNUSED PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1466: {
1467: PetscScalar *v;
1468: const PetscInt m = A->rmap->n, N = A->cmap->N;
1469: PetscInt lda;
1470: PetscCount i, j;
1471: PetscMUMPSInt *row, *col;
1473: PetscFunctionBegin;
1474: PetscCall(MatDenseGetArray(A, &v));
1475: PetscCall(MatDenseGetLDA(A, &lda));
1476: if (reuse == MAT_INITIAL_MATRIX) {
1477: PetscCall(PetscMalloc2(m * N, &row, m * N, &col));
1478: for (i = 0; i < m; i++) {
1479: col[i] = 0;
1480: PetscCall(PetscMUMPSIntCast(i + A->rmap->rstart, &row[i]));
1481: }
1482: for (j = 1; j < N; j++) {
1483: for (i = 0; i < m; i++) PetscCall(PetscMUMPSIntCast(j, col + i + m * j));
1484: PetscCall(PetscArraycpy(row + m * j, row + m * (j - 1), m));
1485: }
1486: if (lda == m) mumps->val = v;
1487: else {
1488: PetscCall(PetscMalloc1(m * N, &mumps->val));
1489: mumps->val_alloc = mumps->val;
1490: for (j = 0; j < N; j++) PetscCall(PetscArraycpy(mumps->val + m * j, v + lda * j, m));
1491: }
1492: mumps->irn = row;
1493: mumps->jcn = col;
1494: mumps->nnz = m * N;
1495: } else {
1496: if (lda == m && !mumps->nest_vals) mumps->val = v;
1497: else {
1498: for (j = 0; j < N; j++) PetscCall(PetscArraycpy(mumps->val + m * j, v + lda * j, m));
1499: }
1500: }
1501: PetscCall(MatDenseRestoreArray(A, &v));
1502: PetscFunctionReturn(PETSC_SUCCESS);
1503: }
1505: // If the input Mat (sub) is either MATTRANSPOSEVIRTUAL or MATHERMITIANTRANSPOSEVIRTUAL, this function gets the parent Mat until it is not a
1506: // MATTRANSPOSEVIRTUAL or MATHERMITIANTRANSPOSEVIRTUAL itself and returns the appropriate shift, scaling, and whether the parent Mat should be conjugated
1507: // and its rows and columns permuted
1508: // TODO FIXME: this should not be in this file and should instead be refactored where the same logic applies, e.g., MatAXPY_Dense_Nest()
1509: static PetscErrorCode MatGetTranspose_TransposeVirtual(Mat *sub, PetscBool *conjugate, PetscScalar *vshift, PetscScalar *vscale, PetscBool *swap)
1510: {
1511: Mat A;
1512: PetscScalar s[2];
1513: PetscBool isTrans, isHTrans, compare;
1515: PetscFunctionBegin;
1516: do {
1517: PetscCall(PetscObjectTypeCompare((PetscObject)*sub, MATTRANSPOSEVIRTUAL, &isTrans));
1518: if (isTrans) {
1519: PetscCall(MatTransposeGetMat(*sub, &A));
1520: isHTrans = PETSC_FALSE;
1521: } else {
1522: PetscCall(PetscObjectTypeCompare((PetscObject)*sub, MATHERMITIANTRANSPOSEVIRTUAL, &isHTrans));
1523: if (isHTrans) PetscCall(MatHermitianTransposeGetMat(*sub, &A));
1524: }
1525: compare = (PetscBool)(isTrans || isHTrans);
1526: if (compare) {
1527: if (vshift && vscale) {
1528: PetscCall(MatShellGetScalingShifts(*sub, s, s + 1, (Vec *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Mat *)MAT_SHELL_NOT_ALLOWED, (IS *)MAT_SHELL_NOT_ALLOWED, (IS *)MAT_SHELL_NOT_ALLOWED));
1529: if (!*conjugate) {
1530: *vshift += s[0] * *vscale;
1531: *vscale *= s[1];
1532: } else {
1533: *vshift += PetscConj(s[0]) * *vscale;
1534: *vscale *= PetscConj(s[1]);
1535: }
1536: }
1537: if (swap) *swap = (PetscBool)!*swap;
1538: if (isHTrans && conjugate) *conjugate = (PetscBool)!*conjugate;
1539: *sub = A;
1540: }
1541: } while (compare);
1542: PetscFunctionReturn(PETSC_SUCCESS);
1543: }
1545: static PetscErrorCode MatConvertToTriples_nest_xaij(Mat A, PetscInt shift, MatReuse reuse, Mat_MUMPS *mumps)
1546: {
1547: Mat **mats;
1548: PetscInt nr, nc;
1549: PetscBool chol = mumps->sym ? PETSC_TRUE : PETSC_FALSE;
1551: PetscFunctionBegin;
1552: PetscCall(MatNestGetSubMats(A, &nr, &nc, &mats));
1553: if (reuse == MAT_INITIAL_MATRIX) {
1554: PetscMUMPSInt *irns, *jcns;
1555: PetscScalar *vals;
1556: PetscCount totnnz, cumnnz, maxnnz;
1557: PetscInt *pjcns_w, Mbs = 0;
1558: IS *rows, *cols;
1559: PetscInt **rows_idx, **cols_idx;
1561: cumnnz = 0;
1562: maxnnz = 0;
1563: PetscCall(PetscMalloc2(nr * nc + 1, &mumps->nest_vals_start, nr * nc, &mumps->nest_convert_to_triples));
1564: for (PetscInt r = 0; r < nr; r++) {
1565: for (PetscInt c = 0; c < nc; c++) {
1566: Mat sub = mats[r][c];
1568: mumps->nest_convert_to_triples[r * nc + c] = NULL;
1569: if (chol && c < r) continue; /* skip lower-triangular block for Cholesky */
1570: if (sub) {
1571: PetscErrorCode (*convert_to_triples)(Mat, PetscInt, MatReuse, Mat_MUMPS *) = NULL;
1572: PetscBool isSeqAIJ, isMPIAIJ, isSeqBAIJ, isMPIBAIJ, isSeqSBAIJ, isMPISBAIJ, isDiag, isDense;
1573: MatInfo info;
1575: PetscCall(MatGetTranspose_TransposeVirtual(&sub, NULL, NULL, NULL, NULL));
1576: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATSEQAIJ, &isSeqAIJ));
1577: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATMPIAIJ, &isMPIAIJ));
1578: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATSEQBAIJ, &isSeqBAIJ));
1579: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATMPIBAIJ, &isMPIBAIJ));
1580: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATSEQSBAIJ, &isSeqSBAIJ));
1581: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATMPISBAIJ, &isMPISBAIJ));
1582: PetscCall(PetscObjectTypeCompare((PetscObject)sub, MATDIAGONAL, &isDiag));
1583: PetscCall(PetscObjectTypeCompareAny((PetscObject)sub, &isDense, MATSEQDENSE, MATMPIDENSE, NULL));
1585: if (chol) {
1586: if (r == c) {
1587: if (isSeqAIJ) convert_to_triples = MatConvertToTriples_seqaij_seqsbaij;
1588: else if (isMPIAIJ) convert_to_triples = MatConvertToTriples_mpiaij_mpisbaij;
1589: else if (isSeqSBAIJ) convert_to_triples = MatConvertToTriples_seqsbaij_seqsbaij;
1590: else if (isMPISBAIJ) convert_to_triples = MatConvertToTriples_mpisbaij_mpisbaij;
1591: else if (isDiag) convert_to_triples = MatConvertToTriples_diagonal_xaij;
1592: else if (isDense) convert_to_triples = MatConvertToTriples_dense_xaij;
1593: } else {
1594: if (isSeqAIJ) convert_to_triples = MatConvertToTriples_seqaij_seqaij;
1595: else if (isMPIAIJ) convert_to_triples = MatConvertToTriples_mpiaij_mpiaij;
1596: else if (isSeqBAIJ) convert_to_triples = MatConvertToTriples_seqbaij_seqaij;
1597: else if (isMPIBAIJ) convert_to_triples = MatConvertToTriples_mpibaij_mpiaij;
1598: else if (isDiag) convert_to_triples = MatConvertToTriples_diagonal_xaij;
1599: else if (isDense) convert_to_triples = MatConvertToTriples_dense_xaij;
1600: }
1601: } else {
1602: if (isSeqAIJ) convert_to_triples = MatConvertToTriples_seqaij_seqaij;
1603: else if (isMPIAIJ) convert_to_triples = MatConvertToTriples_mpiaij_mpiaij;
1604: else if (isSeqBAIJ) convert_to_triples = MatConvertToTriples_seqbaij_seqaij;
1605: else if (isMPIBAIJ) convert_to_triples = MatConvertToTriples_mpibaij_mpiaij;
1606: else if (isDiag) convert_to_triples = MatConvertToTriples_diagonal_xaij;
1607: else if (isDense) convert_to_triples = MatConvertToTriples_dense_xaij;
1608: }
1609: PetscCheck(convert_to_triples, PetscObjectComm((PetscObject)sub), PETSC_ERR_SUP, "Not for block of type %s", ((PetscObject)sub)->type_name);
1610: mumps->nest_convert_to_triples[r * nc + c] = convert_to_triples;
1611: PetscCall(MatGetInfo(sub, MAT_LOCAL, &info));
1612: cumnnz += (PetscCount)info.nz_used; /* can be overestimated for Cholesky */
1613: maxnnz = PetscMax(maxnnz, info.nz_used);
1614: }
1615: }
1616: }
1618: /* Allocate total COO */
1619: totnnz = cumnnz;
1620: PetscCall(PetscMalloc2(totnnz, &irns, totnnz, &jcns));
1621: PetscCall(PetscMalloc1(totnnz, &vals));
1623: /* Handle rows and column maps
1624: We directly map rows and use an SF for the columns */
1625: PetscCall(PetscMalloc4(nr, &rows, nc, &cols, nr, &rows_idx, nc, &cols_idx));
1626: PetscCall(MatNestGetISs(A, rows, cols));
1627: for (PetscInt r = 0; r < nr; r++) PetscCall(ISGetIndices(rows[r], (const PetscInt **)&rows_idx[r]));
1628: for (PetscInt c = 0; c < nc; c++) PetscCall(ISGetIndices(cols[c], (const PetscInt **)&cols_idx[c]));
1629: if (PetscDefined(USE_64BIT_INDICES)) PetscCall(PetscMalloc1(maxnnz, &pjcns_w));
1630: else (void)maxnnz;
1632: cumnnz = 0;
1633: for (PetscInt r = 0; r < nr; r++) {
1634: for (PetscInt c = 0; c < nc; c++) {
1635: Mat sub = mats[r][c];
1636: const PetscInt *ridx = rows_idx[r];
1637: const PetscInt *cidx = cols_idx[c];
1638: PetscScalar vscale = 1.0, vshift = 0.0;
1639: PetscInt rst, size, bs;
1640: PetscSF csf;
1641: PetscBool conjugate = PETSC_FALSE, swap = PETSC_FALSE;
1642: PetscLayout cmap;
1643: PetscInt innz;
1645: mumps->nest_vals_start[r * nc + c] = cumnnz;
1646: if (c == r) {
1647: PetscCall(ISGetSize(rows[r], &size));
1648: if (!mumps->nest_convert_to_triples[r * nc + c]) {
1649: for (PetscInt c = 0; c < nc && !sub; ++c) sub = mats[r][c]; // diagonal Mat is NULL, so start over from the beginning of the current row
1650: }
1651: PetscCall(MatGetBlockSize(sub, &bs));
1652: Mbs += size / bs;
1653: }
1654: if (!mumps->nest_convert_to_triples[r * nc + c]) continue;
1656: /* Extract inner blocks if needed */
1657: PetscCall(MatGetTranspose_TransposeVirtual(&sub, &conjugate, &vshift, &vscale, &swap));
1658: PetscCheck(vshift == 0.0, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Nonzero shift in parent MatShell");
1660: /* Get column layout to map off-process columns */
1661: PetscCall(MatGetLayouts(sub, NULL, &cmap));
1663: /* Get row start to map on-process rows */
1664: PetscCall(MatGetOwnershipRange(sub, &rst, NULL));
1666: /* Directly use the mumps datastructure and use C ordering for now */
1667: PetscCall((*mumps->nest_convert_to_triples[r * nc + c])(sub, 0, MAT_INITIAL_MATRIX, mumps));
1669: /* Swap the role of rows and columns indices for transposed blocks
1670: since we need values with global final ordering */
1671: if (swap) {
1672: cidx = rows_idx[r];
1673: ridx = cols_idx[c];
1674: }
1676: /* Communicate column indices
1677: This could have been done with a single SF but it would have complicated the code a lot.
1678: But since we do it only once, we pay the price of setting up an SF for each block */
1679: if (PetscDefined(USE_64BIT_INDICES)) {
1680: for (PetscInt k = 0; k < mumps->nnz; k++) pjcns_w[k] = mumps->jcn[k];
1681: } else pjcns_w = (PetscInt *)mumps->jcn; /* This cast is needed only to silence warnings for 64bit integers builds */
1682: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &csf));
1683: PetscCall(PetscIntCast(mumps->nnz, &innz));
1684: PetscCall(PetscSFSetGraphLayout(csf, cmap, innz, NULL, PETSC_OWN_POINTER, pjcns_w));
1685: PetscCall(PetscSFBcastBegin(csf, MPIU_INT, cidx, pjcns_w, MPI_REPLACE));
1686: PetscCall(PetscSFBcastEnd(csf, MPIU_INT, cidx, pjcns_w, MPI_REPLACE));
1687: PetscCall(PetscSFDestroy(&csf));
1689: /* Import indices: use direct map for rows and mapped indices for columns */
1690: if (swap) {
1691: for (PetscInt k = 0; k < mumps->nnz; k++) {
1692: PetscCall(PetscMUMPSIntCast(ridx[mumps->irn[k] - rst] + shift, &jcns[cumnnz + k]));
1693: PetscCall(PetscMUMPSIntCast(pjcns_w[k] + shift, &irns[cumnnz + k]));
1694: }
1695: } else {
1696: for (PetscInt k = 0; k < mumps->nnz; k++) {
1697: PetscCall(PetscMUMPSIntCast(ridx[mumps->irn[k] - rst] + shift, &irns[cumnnz + k]));
1698: PetscCall(PetscMUMPSIntCast(pjcns_w[k] + shift, &jcns[cumnnz + k]));
1699: }
1700: }
1702: /* Import values to full COO */
1703: if (conjugate) { /* conjugate the entries */
1704: PetscScalar *v = vals + cumnnz;
1705: for (PetscInt k = 0; k < mumps->nnz; k++) v[k] = vscale * PetscConj(mumps->val[k]);
1706: } else if (vscale != 1.0) {
1707: PetscScalar *v = vals + cumnnz;
1708: for (PetscInt k = 0; k < mumps->nnz; k++) v[k] = vscale * mumps->val[k];
1709: } else PetscCall(PetscArraycpy(vals + cumnnz, mumps->val, mumps->nnz));
1711: /* Shift new starting point and sanity check */
1712: cumnnz += mumps->nnz;
1713: PetscCheck(cumnnz <= totnnz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected number of nonzeros %" PetscCount_FMT " != %" PetscCount_FMT, cumnnz, totnnz);
1715: /* Free scratch memory */
1716: PetscCall(PetscFree2(mumps->irn, mumps->jcn));
1717: PetscCall(PetscFree(mumps->val_alloc));
1718: mumps->val = NULL;
1719: mumps->nnz = 0;
1720: }
1721: }
1722: if (mumps->id.ICNTL(15) == 1) {
1723: if (Mbs != A->rmap->N) {
1724: PetscMPIInt rank, size;
1726: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)A), &rank));
1727: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1728: if (rank == 0) {
1729: PetscInt shift = 0;
1731: PetscCall(PetscMUMPSIntCast(Mbs, &mumps->id.nblk));
1732: PetscCall(PetscFree(mumps->id.blkptr));
1733: PetscCall(PetscMalloc1(Mbs + 1, &mumps->id.blkptr));
1734: mumps->id.blkptr[0] = 1;
1735: for (PetscInt i = 0; i < size; ++i) {
1736: for (PetscInt r = 0; r < nr; r++) {
1737: Mat sub = mats[r][r];
1738: const PetscInt *ranges;
1739: PetscInt bs;
1741: for (PetscInt c = 0; c < nc && !sub; ++c) sub = mats[r][c]; // diagonal Mat is NULL, so start over from the beginning of the current row
1742: PetscCall(MatGetOwnershipRanges(sub, &ranges));
1743: PetscCall(MatGetBlockSize(sub, &bs));
1744: for (PetscInt j = 0, start = mumps->id.blkptr[shift] + bs; j < ranges[i + 1] - ranges[i]; j += bs) PetscCall(PetscMUMPSIntCast(start + j, mumps->id.blkptr + shift + j / bs + 1));
1745: shift += (ranges[i + 1] - ranges[i]) / bs;
1746: }
1747: }
1748: }
1749: } else mumps->id.ICNTL(15) = 0;
1750: }
1751: if (PetscDefined(USE_64BIT_INDICES)) PetscCall(PetscFree(pjcns_w));
1752: for (PetscInt r = 0; r < nr; r++) PetscCall(ISRestoreIndices(rows[r], (const PetscInt **)&rows_idx[r]));
1753: for (PetscInt c = 0; c < nc; c++) PetscCall(ISRestoreIndices(cols[c], (const PetscInt **)&cols_idx[c]));
1754: PetscCall(PetscFree4(rows, cols, rows_idx, cols_idx));
1755: if (!chol) PetscCheck(cumnnz == totnnz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Different number of nonzeros %" PetscCount_FMT " != %" PetscCount_FMT, cumnnz, totnnz);
1756: mumps->nest_vals_start[nr * nc] = cumnnz;
1758: /* Set pointers for final MUMPS data structure */
1759: mumps->nest_vals = vals;
1760: mumps->val_alloc = NULL; /* do not use val_alloc since it may be reallocated with the OMP callpath */
1761: mumps->val = vals;
1762: mumps->irn = irns;
1763: mumps->jcn = jcns;
1764: mumps->nnz = cumnnz;
1765: } else {
1766: PetscScalar *oval = mumps->nest_vals;
1767: for (PetscInt r = 0; r < nr; r++) {
1768: for (PetscInt c = 0; c < nc; c++) {
1769: PetscBool conjugate = PETSC_FALSE;
1770: Mat sub = mats[r][c];
1771: PetscScalar vscale = 1.0, vshift = 0.0;
1772: PetscInt midx = r * nc + c;
1774: if (!mumps->nest_convert_to_triples[midx]) continue;
1775: PetscCall(MatGetTranspose_TransposeVirtual(&sub, &conjugate, &vshift, &vscale, NULL));
1776: PetscCheck(vshift == 0.0, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Nonzero shift in parent MatShell");
1777: mumps->val = oval + mumps->nest_vals_start[midx];
1778: PetscCall((*mumps->nest_convert_to_triples[midx])(sub, shift, MAT_REUSE_MATRIX, mumps));
1779: if (conjugate) {
1780: PetscCount nnz = mumps->nest_vals_start[midx + 1] - mumps->nest_vals_start[midx];
1781: for (PetscCount k = 0; k < nnz; k++) mumps->val[k] = vscale * PetscConj(mumps->val[k]);
1782: } else if (vscale != 1.0) {
1783: PetscCount nnz = mumps->nest_vals_start[midx + 1] - mumps->nest_vals_start[midx];
1784: for (PetscCount k = 0; k < nnz; k++) mumps->val[k] *= vscale;
1785: }
1786: }
1787: }
1788: mumps->val = oval;
1789: }
1790: PetscFunctionReturn(PETSC_SUCCESS);
1791: }
1793: static PetscErrorCode MatDestroy_MUMPS(Mat F)
1794: {
1795: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
1797: PetscFunctionBegin;
1798: PetscCall(MatMumpsResetSchur_Private(F));
1799: PetscCall(PetscFree(mumps->id.isol_loc));
1800: PetscCall(VecScatterDestroy(&mumps->scat_rhs));
1801: PetscCall(VecScatterDestroy(&mumps->scat_sol));
1802: PetscCall(VecDestroy(&mumps->b_seq));
1803: PetscCall(VecDestroy(&mumps->x_seq));
1804: PetscCall(PetscFree(mumps->id.perm_in));
1805: PetscCall(PetscFree(mumps->id.blkvar));
1806: PetscCall(PetscFree(mumps->id.blkptr));
1807: PetscCall(PetscFree2(mumps->irn, mumps->jcn));
1808: PetscCall(PetscFree(mumps->val_alloc));
1809: PetscCall(PetscFree(mumps->info));
1810: PetscCall(PetscFree(mumps->ICNTL_pre));
1811: PetscCall(PetscFree(mumps->CNTL_pre));
1812: if (mumps->id.job != JOB_NULL) { /* cannot call PetscMUMPS_c() if JOB_INIT has never been called for this instance */
1813: mumps->id.job = JOB_END;
1814: PetscMUMPS_c(mumps);
1815: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in termination: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
1816: if (mumps->mumps_comm != MPI_COMM_NULL) {
1817: if (PetscDefined(HAVE_OPENMP_SUPPORT) && mumps->use_petsc_omp_support) PetscCallMPI(MPI_Comm_free(&mumps->mumps_comm));
1818: else PetscCall(PetscCommRestoreComm(PetscObjectComm((PetscObject)F), &mumps->mumps_comm));
1819: }
1820: }
1821: PetscCall(MatMumpsFreeInternalID(&mumps->id));
1822: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
1823: if (mumps->use_petsc_omp_support) {
1824: PetscCall(PetscOmpCtrlDestroy(&mumps->omp_ctrl));
1825: PetscCall(PetscFree2(mumps->rhs_loc, mumps->rhs_recvbuf));
1826: PetscCall(PetscFree3(mumps->rhs_nrow, mumps->rhs_recvcounts, mumps->rhs_disps));
1827: }
1828: #endif
1829: PetscCall(PetscFree(mumps->ia_alloc));
1830: PetscCall(PetscFree(mumps->ja_alloc));
1831: PetscCall(PetscFree(mumps->recvcount));
1832: PetscCall(PetscFree(mumps->reqs));
1833: PetscCall(PetscFree(mumps->irhs_loc));
1834: PetscCall(PetscFree2(mumps->nest_vals_start, mumps->nest_convert_to_triples));
1835: PetscCall(PetscFree(mumps->nest_vals));
1836: PetscCall(PetscFree(F->data));
1838: /* clear composed functions */
1839: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatFactorGetSolverType_C", NULL));
1840: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatFactorSetSchurIS_C", NULL));
1841: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatFactorCreateSchurComplement_C", NULL));
1842: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsSetIcntl_C", NULL));
1843: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetIcntl_C", NULL));
1844: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsSetCntl_C", NULL));
1845: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetCntl_C", NULL));
1846: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetInfo_C", NULL));
1847: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetInfog_C", NULL));
1848: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetRinfo_C", NULL));
1849: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetRinfog_C", NULL));
1850: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetNullPivots_C", NULL));
1851: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetInverse_C", NULL));
1852: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsGetInverseTranspose_C", NULL));
1853: PetscCall(PetscObjectComposeFunction((PetscObject)F, "MatMumpsSetBlk_C", NULL));
1854: PetscFunctionReturn(PETSC_SUCCESS);
1855: }
1857: /* Set up the distributed RHS info for MUMPS. <nrhs> is the number of RHS. <array> points to start of RHS on the local processor. */
1858: static PetscErrorCode MatMumpsSetUpDistRHSInfo(Mat A, PetscInt nrhs, const PetscScalar *array)
1859: {
1860: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
1861: const PetscMPIInt ompsize = mumps->omp_comm_size;
1862: PetscInt i, m, M, rstart;
1864: PetscFunctionBegin;
1865: PetscCall(MatGetSize(A, &M, NULL));
1866: PetscCall(MatGetLocalSize(A, &m, NULL));
1867: PetscCheck(M <= PETSC_MUMPS_INT_MAX, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "PetscInt too long for PetscMUMPSInt");
1868: if (ompsize == 1) {
1869: if (!mumps->irhs_loc) {
1870: mumps->nloc_rhs = (PetscMUMPSInt)m;
1871: PetscCall(PetscMalloc1(m, &mumps->irhs_loc));
1872: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
1873: for (i = 0; i < m; i++) PetscCall(PetscMUMPSIntCast(rstart + i + 1, &mumps->irhs_loc[i])); /* use 1-based indices */
1874: }
1875: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, m * nrhs, array, mumps->id.precision, &mumps->id.rhs_loc_len, &mumps->id.rhs_loc));
1876: } else {
1877: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
1878: const PetscInt *ranges;
1879: PetscMPIInt j, k, sendcount, *petsc_ranks, *omp_ranks;
1880: MPI_Group petsc_group, omp_group;
1881: PetscScalar *recvbuf = NULL;
1883: if (mumps->is_omp_master) {
1884: /* Lazily initialize the omp stuff for distributed rhs */
1885: if (!mumps->irhs_loc) {
1886: PetscCall(PetscMalloc2(ompsize, &omp_ranks, ompsize, &petsc_ranks));
1887: PetscCall(PetscMalloc3(ompsize, &mumps->rhs_nrow, ompsize, &mumps->rhs_recvcounts, ompsize, &mumps->rhs_disps));
1888: PetscCallMPI(MPI_Comm_group(mumps->petsc_comm, &petsc_group));
1889: PetscCallMPI(MPI_Comm_group(mumps->omp_comm, &omp_group));
1890: for (j = 0; j < ompsize; j++) omp_ranks[j] = j;
1891: PetscCallMPI(MPI_Group_translate_ranks(omp_group, ompsize, omp_ranks, petsc_group, petsc_ranks));
1893: /* Populate mumps->irhs_loc[], rhs_nrow[] */
1894: mumps->nloc_rhs = 0;
1895: PetscCall(MatGetOwnershipRanges(A, &ranges));
1896: for (j = 0; j < ompsize; j++) {
1897: mumps->rhs_nrow[j] = ranges[petsc_ranks[j] + 1] - ranges[petsc_ranks[j]];
1898: mumps->nloc_rhs += mumps->rhs_nrow[j];
1899: }
1900: PetscCall(PetscMalloc1(mumps->nloc_rhs, &mumps->irhs_loc));
1901: for (j = k = 0; j < ompsize; j++) {
1902: for (i = ranges[petsc_ranks[j]]; i < ranges[petsc_ranks[j] + 1]; i++, k++) PetscCall(PetscMUMPSIntCast(i + 1, &mumps->irhs_loc[k])); /* uses 1-based indices */
1903: }
1905: PetscCall(PetscFree2(omp_ranks, petsc_ranks));
1906: PetscCallMPI(MPI_Group_free(&petsc_group));
1907: PetscCallMPI(MPI_Group_free(&omp_group));
1908: }
1910: /* Realloc buffers when current nrhs is bigger than what we have met */
1911: if (nrhs > mumps->max_nrhs) {
1912: PetscCall(PetscFree2(mumps->rhs_loc, mumps->rhs_recvbuf));
1913: PetscCall(PetscMalloc2(mumps->nloc_rhs * nrhs, &mumps->rhs_loc, mumps->nloc_rhs * nrhs, &mumps->rhs_recvbuf));
1914: mumps->max_nrhs = nrhs;
1915: }
1917: /* Setup recvcounts[], disps[], recvbuf on omp rank 0 for the upcoming MPI_Gatherv */
1918: for (j = 0; j < ompsize; j++) PetscCall(PetscMPIIntCast(mumps->rhs_nrow[j] * nrhs, &mumps->rhs_recvcounts[j]));
1919: mumps->rhs_disps[0] = 0;
1920: for (j = 1; j < ompsize; j++) {
1921: mumps->rhs_disps[j] = mumps->rhs_disps[j - 1] + mumps->rhs_recvcounts[j - 1];
1922: PetscCheck(mumps->rhs_disps[j] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "PetscMPIInt overflow!");
1923: }
1924: recvbuf = (nrhs == 1) ? mumps->rhs_loc : mumps->rhs_recvbuf; /* Directly use rhs_loc[] as recvbuf. Single rhs is common in Ax=b */
1925: }
1927: PetscCall(PetscMPIIntCast(m * nrhs, &sendcount));
1928: PetscCallMPI(MPI_Gatherv(array, sendcount, MPIU_SCALAR, recvbuf, mumps->rhs_recvcounts, mumps->rhs_disps, MPIU_SCALAR, 0, mumps->omp_comm));
1930: if (mumps->is_omp_master) {
1931: if (nrhs > 1) { /* Copy & re-arrange data from rhs_recvbuf[] to mumps->rhs_loc[] only when there are multiple rhs */
1932: PetscScalar *dst, *dstbase = mumps->rhs_loc;
1933: for (j = 0; j < ompsize; j++) {
1934: const PetscScalar *src = mumps->rhs_recvbuf + mumps->rhs_disps[j];
1935: dst = dstbase;
1936: for (i = 0; i < nrhs; i++) {
1937: PetscCall(PetscArraycpy(dst, src, mumps->rhs_nrow[j]));
1938: src += mumps->rhs_nrow[j];
1939: dst += mumps->nloc_rhs;
1940: }
1941: dstbase += mumps->rhs_nrow[j];
1942: }
1943: }
1944: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nloc_rhs * nrhs, mumps->rhs_loc, mumps->id.precision, &mumps->id.rhs_loc_len, &mumps->id.rhs_loc));
1945: }
1946: #endif /* PETSC_HAVE_OPENMP_SUPPORT */
1947: }
1948: mumps->id.nrhs = (PetscMUMPSInt)nrhs;
1949: mumps->id.nloc_rhs = (PetscMUMPSInt)mumps->nloc_rhs;
1950: mumps->id.lrhs_loc = mumps->nloc_rhs;
1951: mumps->id.irhs_loc = mumps->irhs_loc;
1952: PetscFunctionReturn(PETSC_SUCCESS);
1953: }
1955: static PetscErrorCode MatSolve_MUMPS(Mat A, Vec b, Vec x)
1956: {
1957: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
1958: const PetscScalar *barray = NULL;
1959: PetscScalar *array;
1960: IS is_iden, is_petsc;
1961: PetscBool second_solve = PETSC_FALSE;
1962: static PetscBool cite1 = PETSC_FALSE, cite2 = PETSC_FALSE;
1964: PetscFunctionBegin;
1965: PetscCall(PetscCitationsRegister("@article{MUMPS01,\n author = {P.~R. Amestoy and I.~S. Duff and J.-Y. L'Excellent and J. Koster},\n title = {A fully asynchronous multifrontal solver using distributed dynamic scheduling},\n journal = {SIAM "
1966: "Journal on Matrix Analysis and Applications},\n volume = {23},\n number = {1},\n pages = {15--41},\n year = {2001}\n}\n",
1967: &cite1));
1968: PetscCall(PetscCitationsRegister("@article{MUMPS02,\n author = {P.~R. Amestoy and A. Guermouche and J.-Y. L'Excellent and S. Pralet},\n title = {Hybrid scheduling for the parallel solution of linear systems},\n journal = {Parallel "
1969: "Computing},\n volume = {32},\n number = {2},\n pages = {136--156},\n year = {2006}\n}\n",
1970: &cite2));
1972: PetscCall(VecFlag(x, A->factorerrortype));
1973: if (A->factorerrortype) {
1974: PetscCall(PetscInfo(A, "MatSolve is called with singular matrix factor, INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
1975: PetscFunctionReturn(PETSC_SUCCESS);
1976: }
1978: mumps->id.nrhs = 1;
1979: if (mumps->petsc_size > 1) {
1980: if (mumps->ICNTL20 == 10) {
1981: mumps->id.ICNTL(20) = 10; /* dense distributed RHS, need to set rhs_loc[], irhs_loc[] */
1982: PetscCall(VecGetArrayRead(b, &barray));
1983: PetscCall(MatMumpsSetUpDistRHSInfo(A, 1, barray));
1984: } else {
1985: mumps->id.ICNTL(20) = 0; /* dense centralized RHS; Scatter b into a sequential b_seq vector*/
1986: PetscCall(VecScatterBegin(mumps->scat_rhs, b, mumps->b_seq, INSERT_VALUES, SCATTER_FORWARD));
1987: PetscCall(VecScatterEnd(mumps->scat_rhs, b, mumps->b_seq, INSERT_VALUES, SCATTER_FORWARD));
1988: if (!mumps->myid) {
1989: PetscCall(VecGetArray(mumps->b_seq, &array));
1990: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->b_seq->map->n, array, mumps->id.precision, &mumps->id.rhs_len, &mumps->id.rhs));
1991: }
1992: }
1993: } else { /* petsc_size == 1, use MUMPS's dense centralized RHS feature, so that we don't need to bother with isol_loc[] to get the solution */
1994: mumps->id.ICNTL(20) = 0;
1995: PetscCall(VecCopy(b, x));
1996: PetscCall(VecGetArray(x, &array));
1997: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, x->map->n, array, mumps->id.precision, &mumps->id.rhs_len, &mumps->id.rhs));
1998: }
2000: /*
2001: handle condensation step of Schur complement (if any)
2002: We set by default ICNTL(26) == -1 when Schur indices have been provided by the user.
2003: According to MUMPS (5.0.0) manual, any value should be harmful during the factorization phase
2004: Unless the user provides a valid value for ICNTL(26), MatSolve and MatMatSolve routines solve the full system.
2005: This requires an extra call to PetscMUMPS_c and the computation of the factors for S
2006: */
2007: if (mumps->id.size_schur > 0) {
2008: PetscCheck(mumps->petsc_size <= 1, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Parallel Schur complements not yet supported from PETSc");
2009: if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) {
2010: second_solve = PETSC_TRUE;
2011: PetscCall(MatMumpsHandleSchur_Private(A, PETSC_FALSE)); // allocate id.redrhs
2012: mumps->id.ICNTL(26) = 1; /* condensation phase */
2013: } else if (mumps->id.ICNTL(26) == 1) PetscCall(MatMumpsHandleSchur_Private(A, PETSC_FALSE));
2014: }
2016: mumps->id.job = JOB_SOLVE;
2017: PetscMUMPS_c(mumps); // reduced solve, put solution in id.redrhs
2018: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in solve: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
2020: /* handle expansion step of Schur complement (if any) */
2021: if (second_solve) PetscCall(MatMumpsHandleSchur_Private(A, PETSC_TRUE));
2022: else if (mumps->id.ICNTL(26) == 1) { // condense the right hand side
2023: PetscCall(MatMumpsSolveSchur_Private(A));
2024: for (PetscInt i = 0; i < mumps->id.size_schur; ++i) array[mumps->id.listvar_schur[i] - 1] = ID_FIELD_GET(mumps->id, redrhs, i);
2025: }
2027: if (mumps->petsc_size > 1) { /* convert mumps distributed solution to PETSc mpi x */
2028: if (mumps->scat_sol && mumps->ICNTL9_pre != mumps->id.ICNTL(9)) {
2029: /* when id.ICNTL(9) changes, the contents of ilsol_loc may change (not its size, lsol_loc), recreates scat_sol */
2030: PetscCall(VecScatterDestroy(&mumps->scat_sol));
2031: }
2032: if (!mumps->scat_sol) { /* create scatter scat_sol */
2033: PetscInt *isol2_loc = NULL;
2034: PetscCall(ISCreateStride(PETSC_COMM_SELF, mumps->id.lsol_loc, 0, 1, &is_iden)); /* from */
2035: PetscCall(PetscMalloc1(mumps->id.lsol_loc, &isol2_loc));
2036: for (PetscInt i = 0; i < mumps->id.lsol_loc; i++) isol2_loc[i] = mumps->id.isol_loc[i] - 1; /* change Fortran style to C style */
2037: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, mumps->id.lsol_loc, isol2_loc, PETSC_OWN_POINTER, &is_petsc)); /* to */
2038: PetscCall(VecScatterCreate(mumps->x_seq, is_iden, x, is_petsc, &mumps->scat_sol));
2039: PetscCall(ISDestroy(&is_iden));
2040: PetscCall(ISDestroy(&is_petsc));
2041: mumps->ICNTL9_pre = mumps->id.ICNTL(9); /* save current value of id.ICNTL(9) */
2042: }
2044: PetscScalar *xarray;
2045: PetscCall(VecGetArray(mumps->x_seq, &xarray));
2046: PetscCall(MatMumpsCastMumpsScalarArray(mumps->id.lsol_loc, mumps->id.precision, mumps->id.sol_loc, xarray));
2047: PetscCall(VecRestoreArray(mumps->x_seq, &xarray));
2048: PetscCall(VecScatterBegin(mumps->scat_sol, mumps->x_seq, x, INSERT_VALUES, SCATTER_FORWARD));
2049: PetscCall(VecScatterEnd(mumps->scat_sol, mumps->x_seq, x, INSERT_VALUES, SCATTER_FORWARD));
2051: if (mumps->ICNTL20 == 10) { // distributed RHS
2052: PetscCall(VecRestoreArrayRead(b, &barray));
2053: } else if (!mumps->myid) { // centralized RHS
2054: PetscCall(VecRestoreArray(mumps->b_seq, &array));
2055: }
2056: } else {
2057: // id.rhs has the solution in mumps precision
2058: PetscCall(MatMumpsCastMumpsScalarArray(x->map->n, mumps->id.precision, mumps->id.rhs, array));
2059: PetscCall(VecRestoreArray(x, &array));
2060: }
2062: PetscCall(PetscLogFlops(2.0 * PetscMax(0, (mumps->id.INFO(28) >= 0 ? mumps->id.INFO(28) : -1000000 * mumps->id.INFO(28)) - A->cmap->n)));
2063: PetscFunctionReturn(PETSC_SUCCESS);
2064: }
2066: static PetscErrorCode MatSolveTranspose_MUMPS(Mat A, Vec b, Vec x)
2067: {
2068: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
2069: const PetscMUMPSInt value = mumps->id.ICNTL(9);
2071: PetscFunctionBegin;
2072: mumps->id.ICNTL(9) = 0;
2073: PetscCall(MatSolve_MUMPS(A, b, x));
2074: mumps->id.ICNTL(9) = value;
2075: PetscFunctionReturn(PETSC_SUCCESS);
2076: }
2078: static PetscErrorCode MatMatSolve_MUMPS(Mat A, Mat B, Mat X)
2079: {
2080: Mat Bt = NULL;
2081: PetscBool denseX, denseB, flg, flgT;
2082: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
2083: PetscInt i, nrhs, M, nrhsM;
2084: PetscScalar *array;
2085: const PetscScalar *barray;
2086: PetscInt lsol_loc, nlsol_loc, *idxx, iidx = 0;
2087: PetscMUMPSInt *isol_loc, *isol_loc_save;
2088: PetscScalar *sol_loc;
2089: void *sol_loc_save;
2090: PetscCount sol_loc_len_save;
2091: IS is_to, is_from;
2092: PetscInt k, proc, j, m, myrstart;
2093: const PetscInt *rstart;
2094: Vec v_mpi, msol_loc;
2095: VecScatter scat_sol;
2096: Vec b_seq;
2097: VecScatter scat_rhs;
2098: PetscScalar *aa;
2099: PetscInt spnr, *ia, *ja;
2100: Mat_MPIAIJ *b = NULL;
2102: PetscFunctionBegin;
2103: PetscCall(PetscObjectTypeCompareAny((PetscObject)X, &denseX, MATSEQDENSE, MATMPIDENSE, NULL));
2104: PetscCheck(denseX, PetscObjectComm((PetscObject)X), PETSC_ERR_ARG_WRONG, "Matrix X must be MATDENSE matrix");
2106: PetscCall(PetscObjectTypeCompareAny((PetscObject)B, &denseB, MATSEQDENSE, MATMPIDENSE, NULL));
2108: if (denseB) {
2109: PetscCheck(B->rmap->n == X->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Matrix B and X must have same row distribution");
2110: mumps->id.ICNTL(20) = 0; /* dense RHS */
2111: } else { /* sparse B */
2112: PetscCheck(X != B, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_IDN, "X and B must be different matrices");
2113: PetscCall(PetscObjectTypeCompare((PetscObject)B, MATTRANSPOSEVIRTUAL, &flgT));
2114: PetscCheck(flgT, PetscObjectComm((PetscObject)B), PETSC_ERR_ARG_WRONG, "Matrix B must be MATTRANSPOSEVIRTUAL matrix");
2115: PetscCall(MatShellGetScalingShifts(B, (PetscScalar *)MAT_SHELL_NOT_ALLOWED, (PetscScalar *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Mat *)MAT_SHELL_NOT_ALLOWED, (IS *)MAT_SHELL_NOT_ALLOWED, (IS *)MAT_SHELL_NOT_ALLOWED));
2116: /* input B is transpose of actual RHS matrix,
2117: because mumps requires sparse compressed COLUMN storage! See MatMatTransposeSolve_MUMPS() */
2118: PetscCall(MatTransposeGetMat(B, &Bt));
2119: mumps->id.ICNTL(20) = 1; /* sparse RHS */
2120: }
2122: PetscCall(MatGetSize(B, &M, &nrhs));
2123: PetscCall(PetscIntMultError(nrhs, M, &nrhsM));
2124: mumps->id.nrhs = (PetscMUMPSInt)nrhs;
2125: mumps->id.lrhs = (PetscMUMPSInt)M;
2127: if (mumps->petsc_size == 1) { // handle this easy case specially and return early
2128: PetscScalar *aa;
2129: PetscInt spnr, *ia, *ja;
2130: PetscBool second_solve = PETSC_FALSE;
2132: PetscCall(MatDenseGetArray(X, &array));
2133: if (denseB) {
2134: /* copy B to X */
2135: PetscCall(MatDenseGetArrayRead(B, &barray));
2136: PetscCall(PetscArraycpy(array, barray, nrhsM));
2137: PetscCall(MatDenseRestoreArrayRead(B, &barray));
2138: } else { /* sparse B */
2139: PetscCall(MatSeqAIJGetArray(Bt, &aa));
2140: PetscCall(MatGetRowIJ(Bt, 1, PETSC_FALSE, PETSC_FALSE, &spnr, (const PetscInt **)&ia, (const PetscInt **)&ja, &flg));
2141: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot get IJ structure");
2142: PetscCall(PetscMUMPSIntCSRCast(mumps, spnr, ia, ja, &mumps->id.irhs_ptr, &mumps->id.irhs_sparse, &mumps->id.nz_rhs));
2143: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->id.nz_rhs, aa, mumps->id.precision, &mumps->id.rhs_sparse_len, &mumps->id.rhs_sparse));
2144: }
2145: PetscCall(MatMumpsMakeMumpsScalarArray(denseB, nrhsM, array, mumps->id.precision, &mumps->id.rhs_len, &mumps->id.rhs));
2147: /* handle condensation step of Schur complement (if any) */
2148: if (mumps->id.size_schur > 0) {
2149: if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) {
2150: second_solve = PETSC_TRUE;
2151: PetscCall(MatMumpsHandleSchur_Private(A, PETSC_FALSE)); // allocate id.redrhs
2152: mumps->id.ICNTL(26) = 1; /* condensation phase, i.e, to solve id.redrhs */
2153: } else if (mumps->id.ICNTL(26) == 1) PetscCall(MatMumpsHandleSchur_Private(A, PETSC_FALSE));
2154: }
2156: mumps->id.job = JOB_SOLVE;
2157: PetscMUMPS_c(mumps);
2158: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in solve: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
2160: /* handle expansion step of Schur complement (if any) */
2161: if (second_solve) PetscCall(MatMumpsHandleSchur_Private(A, PETSC_TRUE));
2162: else if (mumps->id.ICNTL(26) == 1) { // condense the right hand side
2163: PetscCall(MatMumpsSolveSchur_Private(A));
2164: for (j = 0; j < nrhs; ++j)
2165: for (i = 0; i < mumps->id.size_schur; ++i) array[mumps->id.listvar_schur[i] - 1 + j * M] = ID_FIELD_GET(mumps->id, redrhs, i + j * mumps->id.lredrhs);
2166: }
2168: if (!denseB) { /* sparse B, restore ia, ja */
2169: PetscCall(MatSeqAIJRestoreArray(Bt, &aa));
2170: PetscCall(MatRestoreRowIJ(Bt, 1, PETSC_FALSE, PETSC_FALSE, &spnr, (const PetscInt **)&ia, (const PetscInt **)&ja, &flg));
2171: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot restore IJ structure");
2172: }
2174: // no matter dense B or sparse B, solution is in id.rhs; convert it to array of X.
2175: PetscCall(MatMumpsCastMumpsScalarArray(nrhsM, mumps->id.precision, mumps->id.rhs, array));
2176: PetscCall(MatDenseRestoreArray(X, &array));
2177: PetscFunctionReturn(PETSC_SUCCESS);
2178: }
2180: /* parallel case: MUMPS requires rhs B to be centralized on the host! */
2181: PetscCheck(!mumps->id.ICNTL(19), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Parallel Schur complements not yet supported from PETSc");
2183: /* create msol_loc to hold mumps local solution */
2184: isol_loc_save = mumps->id.isol_loc; /* save these, as we want to reuse them in MatSolve() */
2185: sol_loc_save = mumps->id.sol_loc;
2186: sol_loc_len_save = mumps->id.sol_loc_len;
2187: mumps->id.isol_loc = NULL; // an init state
2188: mumps->id.sol_loc = NULL;
2189: mumps->id.sol_loc_len = 0;
2191: lsol_loc = mumps->id.lsol_loc;
2192: PetscCall(PetscIntMultError(nrhs, lsol_loc, &nlsol_loc)); /* length of sol_loc */
2193: PetscCall(PetscMalloc2(nlsol_loc, &sol_loc, lsol_loc, &isol_loc));
2194: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_FALSE, nlsol_loc, sol_loc, mumps->id.precision, &mumps->id.sol_loc_len, &mumps->id.sol_loc));
2195: mumps->id.isol_loc = isol_loc;
2197: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, nlsol_loc, (PetscScalar *)sol_loc, &msol_loc));
2199: if (denseB) {
2200: if (mumps->ICNTL20 == 10) {
2201: mumps->id.ICNTL(20) = 10; /* dense distributed RHS */
2202: PetscCall(MatDenseGetArrayRead(B, &barray));
2203: PetscCall(MatMumpsSetUpDistRHSInfo(A, nrhs, barray)); // put barray to rhs_loc
2204: PetscCall(MatDenseRestoreArrayRead(B, &barray));
2205: PetscCall(MatGetLocalSize(B, &m, NULL));
2206: PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)B), 1, nrhs * m, nrhsM, NULL, &v_mpi)); // will scatter the solution to v_mpi, which wraps X
2207: } else {
2208: mumps->id.ICNTL(20) = 0; /* dense centralized RHS */
2209: /* TODO: Because of non-contiguous indices, the created vecscatter scat_rhs is not done in MPI_Gather, resulting in
2210: very inefficient communication. An optimization is to use VecScatterCreateToZero to gather B to rank 0. Then on rank
2211: 0, re-arrange B into desired order, which is a local operation.
2212: */
2214: /* scatter v_mpi to b_seq because MUMPS before 5.3.0 only supports centralized rhs */
2215: /* wrap dense rhs matrix B into a vector v_mpi */
2216: PetscCall(MatGetLocalSize(B, &m, NULL));
2217: PetscCall(MatDenseGetArrayRead(B, &barray));
2218: PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)B), 1, nrhs * m, nrhsM, barray, &v_mpi));
2219: PetscCall(MatDenseRestoreArrayRead(B, &barray));
2221: /* scatter v_mpi to b_seq in proc[0]. With ICNTL(20) = 0, MUMPS requires rhs to be centralized on the host! */
2222: if (!mumps->myid) {
2223: PetscInt *idx;
2224: /* idx: maps from k-th index of v_mpi to (i,j)-th global entry of B */
2225: PetscCall(PetscMalloc1(nrhsM, &idx));
2226: PetscCall(MatGetOwnershipRanges(B, &rstart));
2227: for (proc = 0, k = 0; proc < mumps->petsc_size; proc++) {
2228: for (j = 0; j < nrhs; j++) {
2229: for (i = rstart[proc]; i < rstart[proc + 1]; i++) idx[k++] = j * M + i;
2230: }
2231: }
2233: PetscCall(VecCreateSeq(PETSC_COMM_SELF, nrhsM, &b_seq));
2234: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, nrhsM, idx, PETSC_OWN_POINTER, &is_to));
2235: PetscCall(ISCreateStride(PETSC_COMM_SELF, nrhsM, 0, 1, &is_from));
2236: } else {
2237: PetscCall(VecCreateSeq(PETSC_COMM_SELF, 0, &b_seq));
2238: PetscCall(ISCreateStride(PETSC_COMM_SELF, 0, 0, 1, &is_to));
2239: PetscCall(ISCreateStride(PETSC_COMM_SELF, 0, 0, 1, &is_from));
2240: }
2242: PetscCall(VecScatterCreate(v_mpi, is_from, b_seq, is_to, &scat_rhs));
2243: PetscCall(VecScatterBegin(scat_rhs, v_mpi, b_seq, INSERT_VALUES, SCATTER_FORWARD));
2244: PetscCall(ISDestroy(&is_to));
2245: PetscCall(ISDestroy(&is_from));
2246: PetscCall(VecScatterEnd(scat_rhs, v_mpi, b_seq, INSERT_VALUES, SCATTER_FORWARD));
2248: if (!mumps->myid) { /* define rhs on the host */
2249: PetscCall(VecGetArrayRead(b_seq, &barray));
2250: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, nrhsM, barray, mumps->id.precision, &mumps->id.rhs_len, &mumps->id.rhs));
2251: PetscCall(VecRestoreArrayRead(b_seq, &barray));
2252: }
2253: }
2254: } else { /* sparse B */
2255: b = (Mat_MPIAIJ *)Bt->data;
2257: /* wrap dense X into a vector v_mpi */
2258: PetscCall(MatGetLocalSize(X, &m, NULL));
2259: PetscCall(MatDenseGetArrayRead(X, &barray));
2260: PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)X), 1, nrhs * m, nrhsM, barray, &v_mpi));
2261: PetscCall(MatDenseRestoreArrayRead(X, &barray));
2263: if (!mumps->myid) {
2264: PetscCall(MatSeqAIJGetArray(b->A, &aa));
2265: PetscCall(MatGetRowIJ(b->A, 1, PETSC_FALSE, PETSC_FALSE, &spnr, (const PetscInt **)&ia, (const PetscInt **)&ja, &flg));
2266: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot get IJ structure");
2267: PetscCall(PetscMUMPSIntCSRCast(mumps, spnr, ia, ja, &mumps->id.irhs_ptr, &mumps->id.irhs_sparse, &mumps->id.nz_rhs));
2268: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, ((Mat_SeqAIJ *)b->A->data)->nz, aa, mumps->id.precision, &mumps->id.rhs_sparse_len, &mumps->id.rhs_sparse));
2269: } else {
2270: mumps->id.irhs_ptr = NULL;
2271: mumps->id.irhs_sparse = NULL;
2272: mumps->id.nz_rhs = 0;
2273: if (mumps->id.rhs_sparse_len) {
2274: PetscCall(PetscFree(mumps->id.rhs_sparse));
2275: mumps->id.rhs_sparse_len = 0;
2276: }
2277: }
2278: }
2280: /* solve phase */
2281: mumps->id.job = JOB_SOLVE;
2282: PetscMUMPS_c(mumps);
2283: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in solve: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
2285: /* scatter mumps distributed solution to PETSc vector v_mpi, which shares local arrays with solution matrix X */
2286: PetscCall(MatDenseGetArray(X, &array));
2287: PetscCall(VecPlaceArray(v_mpi, array));
2289: /* create scatter scat_sol */
2290: PetscCall(MatGetOwnershipRanges(X, &rstart));
2291: /* iidx: index for scatter mumps solution to PETSc X */
2293: PetscCall(ISCreateStride(PETSC_COMM_SELF, nlsol_loc, 0, 1, &is_from));
2294: PetscCall(PetscMalloc1(nlsol_loc, &idxx));
2295: for (i = 0; i < lsol_loc; i++) {
2296: isol_loc[i] -= 1; /* change Fortran style to C style. isol_loc[i+j*lsol_loc] contains x[isol_loc[i]] in j-th vector */
2298: for (proc = 0; proc < mumps->petsc_size; proc++) {
2299: if (isol_loc[i] >= rstart[proc] && isol_loc[i] < rstart[proc + 1]) {
2300: myrstart = rstart[proc];
2301: k = isol_loc[i] - myrstart; /* local index on 1st column of PETSc vector X */
2302: iidx = k + myrstart * nrhs; /* maps mumps isol_loc[i] to PETSc index in X */
2303: m = rstart[proc + 1] - rstart[proc]; /* rows of X for this proc */
2304: break;
2305: }
2306: }
2308: for (j = 0; j < nrhs; j++) idxx[i + j * lsol_loc] = iidx + j * m;
2309: }
2310: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, nlsol_loc, idxx, PETSC_COPY_VALUES, &is_to));
2311: PetscCall(MatMumpsCastMumpsScalarArray(nlsol_loc, mumps->id.precision, mumps->id.sol_loc, sol_loc)); // Vec msol_loc is created with sol_loc[]
2312: PetscCall(VecScatterCreate(msol_loc, is_from, v_mpi, is_to, &scat_sol));
2313: PetscCall(VecScatterBegin(scat_sol, msol_loc, v_mpi, INSERT_VALUES, SCATTER_FORWARD));
2314: PetscCall(ISDestroy(&is_from));
2315: PetscCall(ISDestroy(&is_to));
2316: PetscCall(VecScatterEnd(scat_sol, msol_loc, v_mpi, INSERT_VALUES, SCATTER_FORWARD));
2317: PetscCall(MatDenseRestoreArray(X, &array));
2319: if (mumps->id.sol_loc_len) { // in case we allocated intermediate buffers
2320: mumps->id.sol_loc_len = 0;
2321: PetscCall(PetscFree(mumps->id.sol_loc));
2322: }
2324: // restore old values
2325: mumps->id.sol_loc = sol_loc_save;
2326: mumps->id.sol_loc_len = sol_loc_len_save;
2327: mumps->id.isol_loc = isol_loc_save;
2329: PetscCall(PetscFree2(sol_loc, isol_loc));
2330: PetscCall(PetscFree(idxx));
2331: PetscCall(VecDestroy(&msol_loc));
2332: PetscCall(VecDestroy(&v_mpi));
2333: if (!denseB) {
2334: if (!mumps->myid) {
2335: b = (Mat_MPIAIJ *)Bt->data;
2336: PetscCall(MatSeqAIJRestoreArray(b->A, &aa));
2337: PetscCall(MatRestoreRowIJ(b->A, 1, PETSC_FALSE, PETSC_FALSE, &spnr, (const PetscInt **)&ia, (const PetscInt **)&ja, &flg));
2338: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot restore IJ structure");
2339: }
2340: } else {
2341: if (mumps->ICNTL20 == 0) {
2342: PetscCall(VecDestroy(&b_seq));
2343: PetscCall(VecScatterDestroy(&scat_rhs));
2344: }
2345: }
2346: PetscCall(VecScatterDestroy(&scat_sol));
2347: PetscCall(PetscLogFlops(nrhs * PetscMax(0, 2.0 * (mumps->id.INFO(28) >= 0 ? mumps->id.INFO(28) : -1000000 * mumps->id.INFO(28)) - A->cmap->n)));
2348: PetscFunctionReturn(PETSC_SUCCESS);
2349: }
2351: static PetscErrorCode MatMatSolveTranspose_MUMPS(Mat A, Mat B, Mat X)
2352: {
2353: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
2354: const PetscMUMPSInt value = mumps->id.ICNTL(9);
2356: PetscFunctionBegin;
2357: mumps->id.ICNTL(9) = 0;
2358: PetscCall(MatMatSolve_MUMPS(A, B, X));
2359: mumps->id.ICNTL(9) = value;
2360: PetscFunctionReturn(PETSC_SUCCESS);
2361: }
2363: static PetscErrorCode MatMatTransposeSolve_MUMPS(Mat A, Mat Bt, Mat X)
2364: {
2365: PetscBool flg;
2366: Mat B;
2368: PetscFunctionBegin;
2369: PetscCall(PetscObjectTypeCompareAny((PetscObject)Bt, &flg, MATSEQAIJ, MATMPIAIJ, NULL));
2370: PetscCheck(flg, PetscObjectComm((PetscObject)Bt), PETSC_ERR_ARG_WRONG, "Matrix Bt must be MATAIJ matrix");
2372: /* Create B=Bt^T that uses Bt's data structure */
2373: PetscCall(MatCreateTranspose(Bt, &B));
2375: PetscCall(MatMatSolve_MUMPS(A, B, X));
2376: PetscCall(MatDestroy(&B));
2377: PetscFunctionReturn(PETSC_SUCCESS);
2378: }
2380: #if !defined(PETSC_USE_COMPLEX)
2381: /*
2382: input:
2383: F: numeric factor
2384: output:
2385: nneg: total number of negative pivots
2386: nzero: total number of zero pivots
2387: npos: (global dimension of F) - nneg - nzero
2388: */
2389: static PetscErrorCode MatGetInertia_SBAIJMUMPS(Mat F, PetscInt *nneg, PetscInt *nzero, PetscInt *npos)
2390: {
2391: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
2392: PetscMPIInt size;
2394: PetscFunctionBegin;
2395: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)F), &size));
2396: /* MUMPS 4.3.1 calls ScaLAPACK when ICNTL(13)=0 (default), which does not offer the possibility to compute the inertia of a dense matrix. Set ICNTL(13)=1 to skip ScaLAPACK */
2397: PetscCheck(size <= 1 || mumps->id.ICNTL(13) == 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "ICNTL(13)=%d. -mat_mumps_icntl_13 must be set as 1 for correct global matrix inertia", mumps->id.INFOG(13));
2399: if (nneg) *nneg = mumps->id.INFOG(12);
2400: if (nzero || npos) {
2401: PetscCheck(mumps->id.ICNTL(24) == 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-mat_mumps_icntl_24 must be set as 1 for null pivot row detection");
2402: if (nzero) *nzero = mumps->id.INFOG(28);
2403: if (npos) *npos = F->rmap->N - (mumps->id.INFOG(12) + mumps->id.INFOG(28));
2404: }
2405: PetscFunctionReturn(PETSC_SUCCESS);
2406: }
2407: #endif
2409: static PetscErrorCode MatMumpsGatherNonzerosOnMaster(MatReuse reuse, Mat_MUMPS *mumps)
2410: {
2411: PetscMPIInt nreqs;
2412: PetscMUMPSInt *irn, *jcn;
2413: PetscMPIInt count;
2414: PetscCount totnnz, remain;
2415: const PetscInt osize = mumps->omp_comm_size;
2416: PetscScalar *val;
2418: PetscFunctionBegin;
2419: if (osize > 1) {
2420: if (reuse == MAT_INITIAL_MATRIX) {
2421: /* master first gathers counts of nonzeros to receive */
2422: if (mumps->is_omp_master) PetscCall(PetscMalloc1(osize, &mumps->recvcount));
2423: PetscCallMPI(MPI_Gather(&mumps->nnz, 1, MPIU_INT64, mumps->recvcount, 1, MPIU_INT64, 0 /*master*/, mumps->omp_comm));
2425: /* Then each computes number of send/recvs */
2426: if (mumps->is_omp_master) {
2427: /* Start from 1 since self communication is not done in MPI */
2428: nreqs = 0;
2429: for (PetscMPIInt i = 1; i < osize; i++) nreqs += (mumps->recvcount[i] + PETSC_MPI_INT_MAX - 1) / PETSC_MPI_INT_MAX;
2430: } else {
2431: nreqs = (PetscMPIInt)(((mumps->nnz + PETSC_MPI_INT_MAX - 1) / PETSC_MPI_INT_MAX));
2432: }
2433: PetscCall(PetscMalloc1(nreqs * 3, &mumps->reqs)); /* Triple the requests since we send irn, jcn and val separately */
2435: /* The following code is doing a very simple thing: omp_master rank gathers irn/jcn/val from others.
2436: MPI_Gatherv would be enough if it supports big counts > 2^31-1. Since it does not, and mumps->nnz
2437: might be a prime number > 2^31-1, we have to slice the message. Note omp_comm_size
2438: is very small, the current approach should have no extra overhead compared to MPI_Gatherv.
2439: */
2440: nreqs = 0; /* counter for actual send/recvs */
2441: if (mumps->is_omp_master) {
2442: totnnz = 0;
2444: for (PetscMPIInt i = 0; i < osize; i++) totnnz += mumps->recvcount[i]; /* totnnz = sum of nnz over omp_comm */
2445: PetscCall(PetscMalloc2(totnnz, &irn, totnnz, &jcn));
2446: PetscCall(PetscMalloc1(totnnz, &val));
2448: /* Self communication */
2449: PetscCall(PetscArraycpy(irn, mumps->irn, mumps->nnz));
2450: PetscCall(PetscArraycpy(jcn, mumps->jcn, mumps->nnz));
2451: PetscCall(PetscArraycpy(val, mumps->val, mumps->nnz));
2453: /* Replace mumps->irn/jcn etc on master with the newly allocated bigger arrays */
2454: PetscCall(PetscFree2(mumps->irn, mumps->jcn));
2455: PetscCall(PetscFree(mumps->val_alloc));
2456: mumps->nnz = totnnz;
2457: mumps->irn = irn;
2458: mumps->jcn = jcn;
2459: mumps->val = mumps->val_alloc = val;
2461: irn += mumps->recvcount[0]; /* recvcount[0] is old mumps->nnz on omp rank 0 */
2462: jcn += mumps->recvcount[0];
2463: val += mumps->recvcount[0];
2465: /* Remote communication */
2466: for (PetscMPIInt i = 1; i < osize; i++) {
2467: count = (PetscMPIInt)PetscMin(mumps->recvcount[i], (PetscMPIInt)PETSC_MPI_INT_MAX);
2468: remain = mumps->recvcount[i] - count;
2469: while (count > 0) {
2470: PetscCallMPI(MPIU_Irecv(irn, count, MPIU_MUMPSINT, i, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2471: PetscCallMPI(MPIU_Irecv(jcn, count, MPIU_MUMPSINT, i, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2472: PetscCallMPI(MPIU_Irecv(val, count, MPIU_SCALAR, i, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2473: irn += count;
2474: jcn += count;
2475: val += count;
2476: count = (PetscMPIInt)PetscMin(remain, (PetscMPIInt)PETSC_MPI_INT_MAX);
2477: remain -= count;
2478: }
2479: }
2480: } else {
2481: irn = mumps->irn;
2482: jcn = mumps->jcn;
2483: val = mumps->val;
2484: count = (PetscMPIInt)PetscMin(mumps->nnz, (PetscMPIInt)PETSC_MPI_INT_MAX);
2485: remain = mumps->nnz - count;
2486: while (count > 0) {
2487: PetscCallMPI(MPIU_Isend(irn, count, MPIU_MUMPSINT, 0, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2488: PetscCallMPI(MPIU_Isend(jcn, count, MPIU_MUMPSINT, 0, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2489: PetscCallMPI(MPIU_Isend(val, count, MPIU_SCALAR, 0, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2490: irn += count;
2491: jcn += count;
2492: val += count;
2493: count = (PetscMPIInt)PetscMin(remain, (PetscMPIInt)PETSC_MPI_INT_MAX);
2494: remain -= count;
2495: }
2496: }
2497: } else {
2498: nreqs = 0;
2499: if (mumps->is_omp_master) {
2500: val = mumps->val + mumps->recvcount[0];
2501: for (PetscMPIInt i = 1; i < osize; i++) { /* Remote communication only since self data is already in place */
2502: count = (PetscMPIInt)PetscMin(mumps->recvcount[i], (PetscMPIInt)PETSC_MPI_INT_MAX);
2503: remain = mumps->recvcount[i] - count;
2504: while (count > 0) {
2505: PetscCallMPI(MPIU_Irecv(val, count, MPIU_SCALAR, i, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2506: val += count;
2507: count = (PetscMPIInt)PetscMin(remain, (PetscMPIInt)PETSC_MPI_INT_MAX);
2508: remain -= count;
2509: }
2510: }
2511: } else {
2512: val = mumps->val;
2513: count = (PetscMPIInt)PetscMin(mumps->nnz, (PetscMPIInt)PETSC_MPI_INT_MAX);
2514: remain = mumps->nnz - count;
2515: while (count > 0) {
2516: PetscCallMPI(MPIU_Isend(val, count, MPIU_SCALAR, 0, mumps->tag, mumps->omp_comm, &mumps->reqs[nreqs++]));
2517: val += count;
2518: count = (PetscMPIInt)PetscMin(remain, (PetscMPIInt)PETSC_MPI_INT_MAX);
2519: remain -= count;
2520: }
2521: }
2522: }
2523: PetscCallMPI(MPI_Waitall(nreqs, mumps->reqs, MPI_STATUSES_IGNORE));
2524: mumps->tag++; /* It is totally fine for above send/recvs to share one mpi tag */
2525: }
2526: PetscFunctionReturn(PETSC_SUCCESS);
2527: }
2529: static PetscErrorCode MatFactorNumeric_MUMPS(Mat F, Mat A, PETSC_UNUSED const MatFactorInfo *info)
2530: {
2531: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
2533: PetscFunctionBegin;
2534: if (mumps->id.INFOG(1) < 0 && !(mumps->id.INFOG(1) == -16 && mumps->id.INFOG(1) == 0)) {
2535: if (mumps->id.INFOG(1) == -6) PetscCall(PetscInfo(A, "MatFactorNumeric is called with singular matrix structure, INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2536: PetscCall(PetscInfo(A, "MatFactorNumeric is called after analysis phase fails, INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2537: PetscFunctionReturn(PETSC_SUCCESS);
2538: }
2540: PetscCall((*mumps->ConvertToTriples)(A, 1, MAT_REUSE_MATRIX, mumps));
2541: PetscCall(MatMumpsGatherNonzerosOnMaster(MAT_REUSE_MATRIX, mumps));
2542: PetscCall(MatMumpsEnsureSchurArray_Private(F));
2544: /* numerical factorization phase */
2545: mumps->id.job = JOB_FACTNUMERIC;
2546: if (!mumps->id.ICNTL(18)) { /* A is centralized */
2547: if (!mumps->myid) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_len, &mumps->id.a));
2548: } else {
2549: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_loc_len, &mumps->id.a_loc));
2550: }
2552: if (mumps->id.ICNTL(22)) PetscCall(PetscStrncpy(mumps->id.ooc_prefix, ((PetscObject)F)->prefix, sizeof(((MUMPS_STRUC_C *)NULL)->ooc_prefix)));
2553: if (A->rmap->N && A->cmap->N) PetscMUMPS_c(mumps);
2554: if (mumps->id.INFOG(1) < 0) {
2555: PetscCheck(!A->erroriffailure, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in numerical factorization: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
2556: if (mumps->id.INFOG(1) == -10) {
2557: PetscCall(PetscInfo(F, "MUMPS error in numerical factorization: matrix is numerically singular, INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2558: F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
2559: } else if (mumps->id.INFOG(1) == -13) {
2560: PetscCall(PetscInfo(F, "MUMPS error in numerical factorization: INFOG(1)=%d, cannot allocate required memory %d megabytes\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2561: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
2562: } else if (mumps->id.INFOG(1) == -8 || mumps->id.INFOG(1) == -9 || (-16 < mumps->id.INFOG(1) && mumps->id.INFOG(1) < -10)) {
2563: PetscCall(PetscInfo(F, "MUMPS error in numerical factorization: INFOG(1)=%d, INFO(2)=%d, problem with work array\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2564: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
2565: } else {
2566: PetscCall(PetscInfo(F, "MUMPS error in numerical factorization: INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2567: F->factorerrortype = MAT_FACTOR_OTHER;
2568: }
2569: }
2570: PetscCheck(mumps->myid || mumps->id.ICNTL(16) <= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in numerical factorization: ICNTL(16)=%d " MUMPS_MANUALS, mumps->id.INFOG(16));
2572: F->assembled = PETSC_TRUE;
2574: if (F->schur) { /* reset Schur status to unfactored */
2575: #if defined(PETSC_HAVE_CUDA)
2576: F->schur->offloadmask = PETSC_OFFLOAD_CPU;
2577: #endif
2578: PetscScalar *array;
2579: PetscCall(MatDenseGetArray(F->schur, &array));
2580: PetscCall(MatMumpsCastMumpsScalarArray(mumps->id.size_schur * mumps->id.size_schur, mumps->id.precision, mumps->id.schur, array));
2581: PetscCall(MatDenseRestoreArray(F->schur, &array));
2582: if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
2583: mumps->id.ICNTL(19) = 2;
2584: PetscCall(MatTranspose(F->schur, MAT_INPLACE_MATRIX, &F->schur));
2585: }
2586: PetscCall(MatFactorRestoreSchurComplement(F, NULL, MAT_FACTOR_SCHUR_UNFACTORED));
2587: }
2589: /* just to be sure that ICNTL(19) value returned by a call from MatMumpsGetIcntl is always consistent */
2590: if (!mumps->sym && mumps->id.ICNTL(19) && mumps->id.ICNTL(19) != 1) mumps->id.ICNTL(19) = 3;
2592: if (!mumps->is_omp_master) mumps->id.INFO(23) = 0;
2593: // MUMPS userguide: ISOL_loc should be allocated by the user between the factorization and the
2594: // solve phases. On exit from the solve phase, ISOL_loc(i) contains the index of the variables for
2595: // which the solution (in SOL_loc) is available on the local processor.
2596: // If successive calls to the solve phase (JOB= 3) are performed for a given matrix, ISOL_loc will
2597: // normally have the same contents for each of these calls. The only exception is the case of
2598: // unsymmetric matrices (SYM=1) when the transpose option is changed (see ICNTL(9)) and non
2599: // symmetric row/column exchanges (see ICNTL(6)) have occurred before the solve phase.
2600: if (mumps->petsc_size > 1) {
2601: PetscInt lsol_loc;
2602: PetscScalar *array;
2604: /* distributed solution; Create x_seq=sol_loc for repeated use */
2605: if (mumps->x_seq) {
2606: PetscCall(VecScatterDestroy(&mumps->scat_sol));
2607: PetscCall(PetscFree(mumps->id.isol_loc));
2608: PetscCall(VecDestroy(&mumps->x_seq));
2609: }
2610: lsol_loc = mumps->id.INFO(23); /* length of sol_loc */
2611: PetscCall(PetscMalloc1(lsol_loc, &mumps->id.isol_loc));
2612: PetscCall(VecCreateSeq(PETSC_COMM_SELF, lsol_loc, &mumps->x_seq));
2613: PetscCall(VecGetArray(mumps->x_seq, &array));
2614: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_FALSE, lsol_loc, array, mumps->id.precision, &mumps->id.sol_loc_len, &mumps->id.sol_loc));
2615: PetscCall(VecRestoreArray(mumps->x_seq, &array));
2616: mumps->id.lsol_loc = (PetscMUMPSInt)lsol_loc;
2617: }
2618: PetscCall(PetscLogFlops((double)ID_RINFO_GET(mumps->id, 2)));
2619: PetscFunctionReturn(PETSC_SUCCESS);
2620: }
2622: /* Sets MUMPS options from the options database */
2623: static PetscErrorCode MatSetFromOptions_MUMPS(Mat F, Mat A)
2624: {
2625: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
2626: PetscReal cntl;
2627: PetscMUMPSInt icntl = 0, size, *listvar_schur;
2628: PetscInt info[80], i, ninfo = 80, rbs, cbs;
2629: PetscBool flg = PETSC_FALSE;
2630: PetscBool schur = mumps->id.icntl ? (PetscBool)(mumps->id.ICNTL(26) == -1) : (PetscBool)(mumps->ICNTL26 == -1);
2631: void *arr;
2633: PetscFunctionBegin;
2634: PetscOptionsBegin(PetscObjectComm((PetscObject)F), ((PetscObject)F)->prefix, "MUMPS Options", "Mat");
2635: if (mumps->id.job == JOB_NULL) { /* MatSetFromOptions_MUMPS() has never been called before */
2636: PetscPrecision precision = PetscDefined(USE_REAL_SINGLE) ? PETSC_PRECISION_SINGLE : PETSC_PRECISION_DOUBLE;
2637: PetscInt nthreads = 0;
2638: PetscInt nCNTL_pre = mumps->CNTL_pre ? mumps->CNTL_pre[0] : 0;
2639: PetscInt nICNTL_pre = mumps->ICNTL_pre ? mumps->ICNTL_pre[0] : 0;
2640: PetscMUMPSInt nblk, *blkvar, *blkptr;
2642: mumps->petsc_comm = PetscObjectComm((PetscObject)A);
2643: PetscCallMPI(MPI_Comm_size(mumps->petsc_comm, &mumps->petsc_size));
2644: PetscCallMPI(MPI_Comm_rank(mumps->petsc_comm, &mumps->myid)); /* "if (!myid)" still works even if mumps_comm is different */
2646: PetscCall(PetscOptionsName("-mat_mumps_use_omp_threads", "Convert MPI processes into OpenMP threads", "None", &mumps->use_petsc_omp_support));
2647: if (mumps->use_petsc_omp_support) nthreads = -1; /* -1 will let PetscOmpCtrlCreate() guess a proper value when user did not supply one */
2648: /* do not use PetscOptionsInt() so that the option -mat_mumps_use_omp_threads is not displayed twice in the help */
2649: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)F)->prefix, "-mat_mumps_use_omp_threads", &nthreads, NULL));
2650: if (mumps->use_petsc_omp_support) {
2651: PetscCheck(!schur, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot use -%smat_mumps_use_omp_threads with the Schur complement feature", ((PetscObject)F)->prefix ? ((PetscObject)F)->prefix : "");
2652: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
2653: PetscCall(PetscOmpCtrlCreate(mumps->petsc_comm, nthreads, &mumps->omp_ctrl));
2654: PetscCall(PetscOmpCtrlGetOmpComms(mumps->omp_ctrl, &mumps->omp_comm, &mumps->mumps_comm, &mumps->is_omp_master));
2655: #else
2656: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP_SYS, "The system does not have PETSc OpenMP support but you added the -%smat_mumps_use_omp_threads option. Configure PETSc with --with-openmp --download-hwloc (or --with-hwloc) to enable it, see more in MATSOLVERMUMPS manual",
2657: ((PetscObject)F)->prefix ? ((PetscObject)F)->prefix : "");
2658: #endif
2659: } else {
2660: mumps->omp_comm = PETSC_COMM_SELF;
2661: mumps->mumps_comm = mumps->petsc_comm;
2662: mumps->is_omp_master = PETSC_TRUE;
2663: }
2664: PetscCallMPI(MPI_Comm_size(mumps->omp_comm, &mumps->omp_comm_size));
2665: mumps->reqs = NULL;
2666: mumps->tag = 0;
2668: if (mumps->mumps_comm != MPI_COMM_NULL) {
2669: if (PetscDefined(HAVE_OPENMP_SUPPORT) && mumps->use_petsc_omp_support) {
2670: /* It looks like MUMPS does not dup the input comm. Dup a new comm for MUMPS to avoid any tag mismatches. */
2671: MPI_Comm comm;
2672: PetscCallMPI(MPI_Comm_dup(mumps->mumps_comm, &comm));
2673: mumps->mumps_comm = comm;
2674: } else PetscCall(PetscCommGetComm(mumps->petsc_comm, &mumps->mumps_comm));
2675: }
2677: mumps->id.comm_fortran = MPI_Comm_c2f(mumps->mumps_comm);
2678: mumps->id.job = JOB_INIT;
2679: mumps->id.par = 1; /* host participates factorizaton and solve */
2680: mumps->id.sym = mumps->sym;
2682: size = mumps->id.size_schur;
2683: arr = mumps->id.schur;
2684: listvar_schur = mumps->id.listvar_schur;
2685: nblk = mumps->id.nblk;
2686: blkvar = mumps->id.blkvar;
2687: blkptr = mumps->id.blkptr;
2688: if (PetscDefined(USE_DEBUG)) {
2689: for (PetscInt i = 0; i < size; i++)
2690: PetscCheck(listvar_schur[i] - 1 >= 0 && listvar_schur[i] - 1 < A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_USER, "Invalid Schur index at position %" PetscInt_FMT "! %" PetscInt_FMT " must be in [0, %" PetscInt_FMT ")", i, (PetscInt)listvar_schur[i] - 1,
2691: A->rmap->N);
2692: }
2694: PetscCall(PetscOptionsEnum("-pc_precision", "Precision used by MUMPS", "MATSOLVERMUMPS", PetscPrecisionTypes, (PetscEnum)precision, (PetscEnum *)&precision, NULL));
2695: PetscCheck(precision == PETSC_PRECISION_SINGLE || precision == PETSC_PRECISION_DOUBLE, PetscObjectComm((PetscObject)F), PETSC_ERR_SUP, "MUMPS does not support %s precision", PetscPrecisionTypes[precision]);
2696: PetscCheck(precision == PETSC_SCALAR_PRECISION || PetscDefined(HAVE_MUMPS_MIXED_PRECISION), PetscObjectComm((PetscObject)F), PETSC_ERR_USER, "Your MUMPS library does not support mixed precision, but which is needed with your specified PetscScalar");
2697: PetscCall(MatMumpsAllocateInternalID(&mumps->id, precision));
2699: PetscMUMPS_c(mumps);
2700: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
2702: /* set PETSc-MUMPS default options - override MUMPS default */
2703: mumps->id.ICNTL(3) = 0;
2704: mumps->id.ICNTL(4) = 0;
2705: if (mumps->petsc_size == 1) {
2706: mumps->id.ICNTL(18) = 0; /* centralized assembled matrix input */
2707: mumps->id.ICNTL(7) = 7; /* automatic choice of ordering done by the package */
2708: } else {
2709: mumps->id.ICNTL(18) = 3; /* distributed assembled matrix input */
2710: mumps->id.ICNTL(21) = 1; /* distributed solution */
2711: }
2712: if (nblk && blkptr) {
2713: mumps->id.ICNTL(15) = 1;
2714: mumps->id.nblk = nblk;
2715: mumps->id.blkvar = blkvar;
2716: mumps->id.blkptr = blkptr;
2717: } else mumps->id.ICNTL(15) = 0;
2719: /* restore cached ICNTL and CNTL values */
2720: for (icntl = 0; icntl < nICNTL_pre; ++icntl) mumps->id.ICNTL(mumps->ICNTL_pre[1 + 2 * icntl]) = mumps->ICNTL_pre[2 + 2 * icntl];
2721: for (icntl = 0; icntl < nCNTL_pre; ++icntl) ID_CNTL_SET(mumps->id, (PetscInt)mumps->CNTL_pre[1 + 2 * icntl], mumps->CNTL_pre[2 + 2 * icntl]);
2723: PetscCall(PetscFree(mumps->ICNTL_pre));
2724: PetscCall(PetscFree(mumps->CNTL_pre));
2726: if (schur) {
2727: mumps->id.size_schur = size;
2728: mumps->id.schur_lld = size;
2729: mumps->id.schur = arr;
2730: mumps->id.listvar_schur = listvar_schur;
2731: if (mumps->petsc_size > 1) {
2732: PetscBool gs; /* gs is false if any rank other than root has non-empty IS */
2734: mumps->id.ICNTL(19) = 1; /* MUMPS returns Schur centralized on the host */
2735: gs = mumps->myid ? (mumps->id.size_schur ? PETSC_FALSE : PETSC_TRUE) : PETSC_TRUE; /* always true on root; false on others if their size != 0 */
2736: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &gs, 1, MPI_C_BOOL, MPI_LAND, mumps->petsc_comm));
2737: PetscCheck(gs, PETSC_COMM_SELF, PETSC_ERR_SUP, "MUMPS distributed parallel Schur complements not yet supported from PETSc");
2738: } else {
2739: if (F->factortype == MAT_FACTOR_LU) {
2740: mumps->id.ICNTL(19) = 3; /* MUMPS returns full matrix */
2741: } else {
2742: mumps->id.ICNTL(19) = 2; /* MUMPS returns lower triangular part */
2743: }
2744: }
2745: mumps->id.ICNTL(26) = -1;
2746: }
2748: /* copy MUMPS default control values from master to slaves. Although slaves do not call MUMPS, they may access these values in code.
2749: For example, ICNTL(9) is initialized to 1 by MUMPS and slaves check ICNTL(9) in MatSolve_MUMPS.
2750: */
2751: PetscCallMPI(MPI_Bcast(mumps->id.icntl, 40, MPI_INT, 0, mumps->omp_comm));
2752: PetscCallMPI(MPI_Bcast(mumps->id.cntl, 15, MPIU_MUMPSREAL(&mumps->id), 0, mumps->omp_comm));
2754: mumps->scat_rhs = NULL;
2755: mumps->scat_sol = NULL;
2756: }
2757: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_1", "ICNTL(1): output stream for error messages", "None", mumps->id.ICNTL(1), &icntl, &flg));
2758: if (flg) mumps->id.ICNTL(1) = icntl;
2759: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_2", "ICNTL(2): output stream for diagnostic printing, statistics, and warning", "None", mumps->id.ICNTL(2), &icntl, &flg));
2760: if (flg) mumps->id.ICNTL(2) = icntl;
2761: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_3", "ICNTL(3): output stream for global information, collected on the host", "None", mumps->id.ICNTL(3), &icntl, &flg));
2762: if (flg) mumps->id.ICNTL(3) = icntl;
2764: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_4", "ICNTL(4): level of printing (0 to 4)", "None", mumps->id.ICNTL(4), &icntl, &flg));
2765: if (flg) mumps->id.ICNTL(4) = icntl;
2766: if (mumps->id.ICNTL(4) || PetscLogPrintInfo) mumps->id.ICNTL(3) = 6; /* resume MUMPS default id.ICNTL(3) = 6 */
2768: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_6", "ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)", "None", mumps->id.ICNTL(6), &icntl, &flg));
2769: if (flg) mumps->id.ICNTL(6) = icntl;
2771: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_7", "ICNTL(7): computes a symmetric permutation in sequential analysis. 0=AMD, 2=AMF, 3=Scotch, 4=PORD, 5=Metis, 6=QAMD, and 7=auto(default)", "None", mumps->id.ICNTL(7), &icntl, &flg));
2772: if (flg) {
2773: PetscCheck(icntl != 1 && icntl >= 0 && icntl <= 7, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Valid values are 0=AMD, 2=AMF, 3=Scotch, 4=PORD, 5=Metis, 6=QAMD, and 7=auto");
2774: mumps->id.ICNTL(7) = icntl;
2775: }
2777: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_8", "ICNTL(8): scaling strategy (-2 to 8 or 77)", "None", mumps->id.ICNTL(8), &mumps->id.ICNTL(8), NULL));
2778: /* PetscCall(PetscOptionsInt("-mat_mumps_icntl_9","ICNTL(9): computes the solution using A or A^T","None",mumps->id.ICNTL(9),&mumps->id.ICNTL(9),NULL)); handled by MatSolveTranspose_MUMPS() */
2779: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_10", "ICNTL(10): max num of refinements", "None", mumps->id.ICNTL(10), &mumps->id.ICNTL(10), NULL));
2780: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_11", "ICNTL(11): statistics related to an error analysis (via -ksp_view)", "None", mumps->id.ICNTL(11), &mumps->id.ICNTL(11), NULL));
2781: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_12", "ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)", "None", mumps->id.ICNTL(12), &mumps->id.ICNTL(12), NULL));
2782: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_13", "ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting", "None", mumps->id.ICNTL(13), &mumps->id.ICNTL(13), NULL));
2783: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_14", "ICNTL(14): percentage increase in the estimated working space", "None", mumps->id.ICNTL(14), &mumps->id.ICNTL(14), NULL));
2784: PetscCall(MatGetBlockSizes(A, &rbs, &cbs));
2785: if (rbs == cbs && rbs > 1) mumps->id.ICNTL(15) = (PetscMUMPSInt)-rbs;
2786: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_15", "ICNTL(15): compression of the input matrix resulting from a block format", "None", mumps->id.ICNTL(15), &mumps->id.ICNTL(15), &flg));
2787: if (flg) {
2788: if (mumps->id.ICNTL(15) < 0) PetscCheck((-mumps->id.ICNTL(15) % cbs == 0) && (-mumps->id.ICNTL(15) % rbs == 0), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "The opposite of -mat_mumps_icntl_15 must be a multiple of the column and row blocksizes");
2789: else if (mumps->id.ICNTL(15) > 0) {
2790: const PetscInt *bsizes;
2791: PetscInt nblocks, p, *blkptr = NULL;
2792: PetscMPIInt *recvcounts, *displs, n;
2793: PetscMPIInt rank, size = 0;
2795: PetscCall(MatGetVariableBlockSizes(A, &nblocks, &bsizes));
2796: flg = PETSC_TRUE;
2797: for (p = 0; p < nblocks; ++p) {
2798: if (bsizes[p] > 1) break;
2799: }
2800: if (p == nblocks) flg = PETSC_FALSE;
2801: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &flg, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)A)));
2802: if (flg) { // if at least one process supplies variable block sizes and they are not all set to 1
2803: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)A), &rank));
2804: if (rank == 0) PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
2805: PetscCall(PetscCalloc2(size, &recvcounts, size + 1, &displs));
2806: PetscCall(PetscMPIIntCast(nblocks, &n));
2807: PetscCallMPI(MPI_Gather(&n, 1, MPI_INT, recvcounts, 1, MPI_INT, 0, PetscObjectComm((PetscObject)A)));
2808: for (PetscInt p = 0; p < size; ++p) displs[p + 1] = displs[p] + recvcounts[p];
2809: PetscCall(PetscMalloc1(displs[size] + 1, &blkptr));
2810: PetscCallMPI(MPI_Bcast(displs + size, 1, MPIU_INT, 0, PetscObjectComm((PetscObject)A)));
2811: PetscCallMPI(MPI_Gatherv(bsizes, n, MPIU_INT, blkptr + 1, recvcounts, displs, MPIU_INT, 0, PetscObjectComm((PetscObject)A)));
2812: if (rank == 0) {
2813: blkptr[0] = 1;
2814: for (PetscInt p = 0; p < n; ++p) blkptr[p + 1] += blkptr[p];
2815: PetscCall(MatMumpsSetBlk(F, displs[size], NULL, blkptr));
2816: }
2817: PetscCall(PetscFree2(recvcounts, displs));
2818: PetscCall(PetscFree(blkptr));
2819: }
2820: }
2821: }
2822: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_19", "ICNTL(19): computes the Schur complement", "None", mumps->id.ICNTL(19), &mumps->id.ICNTL(19), NULL));
2823: if (mumps->id.ICNTL(19) <= 0 || mumps->id.ICNTL(19) > 3) { /* reset any Schur data (if any) */
2824: PetscCall(MatMumpsResetSchur_Private(F));
2825: }
2827: /* Two MPICH Fortran MPI_IN_PLACE binding bugs prevented the use of 'mpich + mumps'. One happened with "mpi4py + mpich + mumps",
2828: and was reported by Firedrake. See https://bitbucket.org/mpi4py/mpi4py/issues/162/mpi4py-initialization-breaks-fortran
2829: and a petsc-maint mailing list thread with subject 'MUMPS segfaults in parallel because of ...'
2830: This bug was fixed by https://github.com/pmodels/mpich/pull/4149. But the fix brought a new bug,
2831: see https://github.com/pmodels/mpich/issues/5589. This bug was fixed by https://github.com/pmodels/mpich/pull/5590.
2832: In short, we could not use distributed RHS until with MPICH v4.0b1 or we enabled a workaround in mumps-5.6.2+
2833: */
2834: mumps->ICNTL20 = 10; /* Distributed dense RHS, by default */
2835: #if PETSC_PKG_MUMPS_VERSION_LT(5, 3, 0) || (PetscDefined(HAVE_MPICH) && MPICH_NUMVERSION < 40000101) || PetscDefined(HAVE_MSMPI)
2836: mumps->ICNTL20 = 0; /* Centralized dense RHS, if need be */
2837: #endif
2838: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_20", "ICNTL(20): give mumps centralized (0) or distributed (10) dense right-hand sides", "None", mumps->ICNTL20, &mumps->ICNTL20, &flg));
2839: PetscCheck(!flg || mumps->ICNTL20 == 10 || mumps->ICNTL20 == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "ICNTL(20)=%d is not supported by the PETSc/MUMPS interface. Allowed values are 0, 10", (int)mumps->ICNTL20);
2840: #if PETSC_PKG_MUMPS_VERSION_LT(5, 3, 0)
2841: PetscCheck(!flg || mumps->ICNTL20 != 10, PETSC_COMM_SELF, PETSC_ERR_SUP, "ICNTL(20)=10 is not supported before MUMPS-5.3.0");
2842: #endif
2843: /* PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_21","ICNTL(21): the distribution (centralized or distributed) of the solution vectors","None",mumps->id.ICNTL(21),&mumps->id.ICNTL(21),NULL)); we only use distributed solution vector */
2845: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_22", "ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)", "None", mumps->id.ICNTL(22), &mumps->id.ICNTL(22), &flg));
2846: if (flg && mumps->id.ICNTL(22) != 1) mumps->id.ICNTL(22) = 0; // MUMPS treats values other than 1 as 0. Normalize it so we can safely use 'if (mumps->id.ICNTL(22))'
2847: if (mumps->id.ICNTL(22)) {
2848: // MUMPS will use the /tmp directory if -mat_mumps_ooc_tmpdir is not set by user.
2849: // We don't provide option -mat_mumps_ooc_prefix, as we use F's prefix as OOC_PREFIX, which is set later during MatFactorNumeric_MUMPS() to also handle cases where users enable OOC via MatMumpsSetIcntl().
2850: PetscCall(PetscOptionsString("-mat_mumps_ooc_tmpdir", "Out of core directory", "None", mumps->id.ooc_tmpdir, mumps->id.ooc_tmpdir, sizeof(((MUMPS_STRUC_C *)NULL)->ooc_tmpdir), NULL));
2851: }
2852: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_23", "ICNTL(23): max size of the working memory (MB) that can allocate per processor", "None", mumps->id.ICNTL(23), &mumps->id.ICNTL(23), NULL));
2853: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_24", "ICNTL(24): detection of null pivot rows (0 or 1)", "None", mumps->id.ICNTL(24), &mumps->id.ICNTL(24), NULL));
2854: if (mumps->id.ICNTL(24)) mumps->id.ICNTL(13) = 1; /* turn-off ScaLAPACK to help with the correct detection of null pivots */
2856: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_25", "ICNTL(25): computes a solution of a deficient matrix and a null space basis", "None", mumps->id.ICNTL(25), &mumps->id.ICNTL(25), NULL));
2857: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_26", "ICNTL(26): drives the solution phase if a Schur complement matrix", "None", mumps->id.ICNTL(26), &mumps->id.ICNTL(26), NULL));
2858: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_27", "ICNTL(27): controls the blocking size for multiple right-hand sides", "None", mumps->id.ICNTL(27), &mumps->id.ICNTL(27), NULL));
2859: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_28", "ICNTL(28): use 1 for sequential analysis and ICNTL(7) ordering, or 2 for parallel analysis and ICNTL(29) ordering", "None", mumps->id.ICNTL(28), &mumps->id.ICNTL(28), NULL));
2860: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_29", "ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis", "None", mumps->id.ICNTL(29), &mumps->id.ICNTL(29), NULL));
2861: /* PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_30","ICNTL(30): compute user-specified set of entries in inv(A)","None",mumps->id.ICNTL(30),&mumps->id.ICNTL(30),NULL)); */ /* call MatMumpsGetInverse() directly */
2862: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_31", "ICNTL(31): indicates which factors may be discarded during factorization", "None", mumps->id.ICNTL(31), &mumps->id.ICNTL(31), NULL));
2863: /* PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_32","ICNTL(32): performs the forward elimination of the right-hand sides during factorization","None",mumps->id.ICNTL(32),&mumps->id.ICNTL(32),NULL)); -- not supported by PETSc API */
2864: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_33", "ICNTL(33): compute determinant", "None", mumps->id.ICNTL(33), &mumps->id.ICNTL(33), NULL));
2865: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_35", "ICNTL(35): activates Block Low Rank (BLR) based factorization", "None", mumps->id.ICNTL(35), &mumps->id.ICNTL(35), NULL));
2866: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_36", "ICNTL(36): choice of BLR factorization variant", "None", mumps->id.ICNTL(36), &mumps->id.ICNTL(36), NULL));
2867: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_37", "ICNTL(37): compression of the contribution blocks (CB)", "None", mumps->id.ICNTL(37), &mumps->id.ICNTL(37), NULL));
2868: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_38", "ICNTL(38): estimated compression rate of LU factors with BLR", "None", mumps->id.ICNTL(38), &mumps->id.ICNTL(38), NULL));
2869: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_40", "ICNTL(40): adaptive BLR precision feature", "None", mumps->id.ICNTL(40), &mumps->id.ICNTL(40), NULL));
2870: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_47", "ICNTL(47): single precision factorization in a double precision instance", "None", mumps->id.ICNTL(47), &mumps->id.ICNTL(47), NULL));
2871: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_48", "ICNTL(48): multithreading with tree parallelism", "None", mumps->id.ICNTL(48), &mumps->id.ICNTL(48), NULL));
2872: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_49", "ICNTL(49): compact workarray at the end of factorization phase", "None", mumps->id.ICNTL(49), &mumps->id.ICNTL(49), NULL));
2873: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_56", "ICNTL(56): postponing and rank-revealing factorization", "None", mumps->id.ICNTL(56), &mumps->id.ICNTL(56), NULL));
2874: PetscCall(PetscOptionsMUMPSInt("-mat_mumps_icntl_58", "ICNTL(58): defines options for symbolic factorization", "None", mumps->id.ICNTL(58), &mumps->id.ICNTL(58), NULL));
2876: PetscCall(PetscOptionsReal("-mat_mumps_cntl_1", "CNTL(1): relative pivoting threshold", "None", (PetscReal)ID_CNTL_GET(mumps->id, 1), &cntl, &flg));
2877: if (flg) ID_CNTL_SET(mumps->id, 1, cntl);
2878: PetscCall(PetscOptionsReal("-mat_mumps_cntl_2", "CNTL(2): stopping criterion of refinement", "None", (PetscReal)ID_CNTL_GET(mumps->id, 2), &cntl, &flg));
2879: if (flg) ID_CNTL_SET(mumps->id, 2, cntl);
2880: PetscCall(PetscOptionsReal("-mat_mumps_cntl_3", "CNTL(3): absolute pivoting threshold", "None", (PetscReal)ID_CNTL_GET(mumps->id, 3), &cntl, &flg));
2881: if (flg) ID_CNTL_SET(mumps->id, 3, cntl);
2882: PetscCall(PetscOptionsReal("-mat_mumps_cntl_4", "CNTL(4): value for static pivoting", "None", (PetscReal)ID_CNTL_GET(mumps->id, 4), &cntl, &flg));
2883: if (flg) ID_CNTL_SET(mumps->id, 4, cntl);
2884: PetscCall(PetscOptionsReal("-mat_mumps_cntl_5", "CNTL(5): fixation for null pivots", "None", (PetscReal)ID_CNTL_GET(mumps->id, 5), &cntl, &flg));
2885: if (flg) ID_CNTL_SET(mumps->id, 5, cntl);
2886: PetscCall(PetscOptionsReal("-mat_mumps_cntl_7", "CNTL(7): dropping parameter used during BLR", "None", (PetscReal)ID_CNTL_GET(mumps->id, 7), &cntl, &flg));
2887: if (flg) ID_CNTL_SET(mumps->id, 7, cntl);
2889: PetscCall(PetscOptionsIntArray("-mat_mumps_view_info", "request INFO local to each processor", "", info, &ninfo, NULL));
2890: if (ninfo) {
2891: PetscCheck(ninfo <= 80, PETSC_COMM_SELF, PETSC_ERR_USER, "number of INFO %" PetscInt_FMT " must <= 80", ninfo);
2892: PetscCall(PetscMalloc1(ninfo, &mumps->info));
2893: mumps->ninfo = ninfo;
2894: for (i = 0; i < ninfo; i++) {
2895: PetscCheck(info[i] >= 0 && info[i] <= 80, PETSC_COMM_SELF, PETSC_ERR_USER, "index of INFO %" PetscInt_FMT " must between 1 and 80", ninfo);
2896: mumps->info[i] = info[i];
2897: }
2898: }
2899: PetscOptionsEnd();
2900: PetscFunctionReturn(PETSC_SUCCESS);
2901: }
2903: static PetscErrorCode MatFactorSymbolic_MUMPS_ReportIfError(Mat F, Mat A, PETSC_UNUSED const MatFactorInfo *info, Mat_MUMPS *mumps)
2904: {
2905: PetscFunctionBegin;
2906: if (mumps->id.INFOG(1) < 0) {
2907: PetscCheck(!A->erroriffailure, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in analysis: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
2908: if (mumps->id.INFOG(1) == -6) {
2909: PetscCall(PetscInfo(F, "MUMPS error in analysis: matrix is singular, INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2910: F->factorerrortype = MAT_FACTOR_STRUCT_ZEROPIVOT;
2911: } else if (mumps->id.INFOG(1) == -5 || mumps->id.INFOG(1) == -7) {
2912: PetscCall(PetscInfo(F, "MUMPS error in analysis: problem with work array, INFOG(1)=%d, INFO(2)=%d\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2913: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
2914: } else {
2915: PetscCall(PetscInfo(F, "MUMPS error in analysis: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS "\n", mumps->id.INFOG(1), mumps->id.INFO(2)));
2916: F->factorerrortype = MAT_FACTOR_OTHER;
2917: }
2918: }
2919: if (!mumps->id.n) F->factorerrortype = MAT_FACTOR_NOERROR;
2920: PetscFunctionReturn(PETSC_SUCCESS);
2921: }
2923: static PetscErrorCode MatLUFactorSymbolic_AIJMUMPS(Mat F, Mat A, IS r, PETSC_UNUSED IS c, const MatFactorInfo *info)
2924: {
2925: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
2926: Vec b;
2927: const PetscInt M = A->rmap->N;
2929: PetscFunctionBegin;
2930: if (mumps->matstruc == SAME_NONZERO_PATTERN) {
2931: /* F is assembled by a previous call of MatLUFactorSymbolic_AIJMUMPS() */
2932: PetscFunctionReturn(PETSC_SUCCESS);
2933: }
2935: /* Set MUMPS options from the options database */
2936: PetscCall(MatSetFromOptions_MUMPS(F, A));
2938: PetscCall((*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, mumps));
2939: PetscCall(MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX, mumps));
2940: PetscCall(MatMumpsEnsureSchurArray_Private(F));
2942: /* analysis phase */
2943: mumps->id.job = JOB_FACTSYMBOLIC;
2944: PetscCall(PetscMUMPSIntCast(M, &mumps->id.n));
2945: switch (mumps->id.ICNTL(18)) {
2946: case 0: /* centralized assembled matrix input */
2947: if (!mumps->myid) {
2948: mumps->id.nnz = mumps->nnz;
2949: mumps->id.irn = mumps->irn;
2950: mumps->id.jcn = mumps->jcn;
2951: if (1 < mumps->id.ICNTL(6) && mumps->id.ICNTL(6) < 7) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_len, &mumps->id.a));
2952: if (r && mumps->id.ICNTL(7) == 7 && !F->schur) {
2953: mumps->id.ICNTL(7) = 1;
2954: if (!mumps->myid) {
2955: const PetscInt *idx;
2957: PetscCall(PetscMalloc1(M, &mumps->id.perm_in));
2958: PetscCall(ISGetIndices(r, &idx));
2959: for (PetscInt i = 0; i < M; i++) PetscCall(PetscMUMPSIntCast(idx[i] + 1, &mumps->id.perm_in[i])); /* perm_in[]: start from 1, not 0! */
2960: PetscCall(ISRestoreIndices(r, &idx));
2961: }
2962: }
2963: }
2964: break;
2965: case 3: /* distributed assembled matrix input (size>1) */
2966: mumps->id.nnz_loc = mumps->nnz;
2967: mumps->id.irn_loc = mumps->irn;
2968: mumps->id.jcn_loc = mumps->jcn;
2969: if (1 < mumps->id.ICNTL(6) && mumps->id.ICNTL(6) < 7) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_loc_len, &mumps->id.a_loc));
2970: if (mumps->ICNTL20 == 0) { /* Centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
2971: PetscCall(MatCreateVecs(A, NULL, &b));
2972: PetscCall(VecScatterCreateToZero(b, &mumps->scat_rhs, &mumps->b_seq));
2973: PetscCall(VecDestroy(&b));
2974: }
2975: break;
2976: }
2977: if (A->rmap->N && A->cmap->N) {
2978: PetscMUMPS_c(mumps);
2979: PetscCall(MatFactorSymbolic_MUMPS_ReportIfError(F, A, info, mumps));
2980: }
2981: F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
2982: F->ops->solve = MatSolve_MUMPS;
2983: F->ops->solvetranspose = MatSolveTranspose_MUMPS;
2984: F->ops->matsolve = MatMatSolve_MUMPS;
2985: F->ops->mattransposesolve = MatMatTransposeSolve_MUMPS;
2986: F->ops->matsolvetranspose = MatMatSolveTranspose_MUMPS;
2988: mumps->matstruc = SAME_NONZERO_PATTERN;
2989: PetscFunctionReturn(PETSC_SUCCESS);
2990: }
2992: /* Note the PETSc r and c permutations are ignored */
2993: static PetscErrorCode MatLUFactorSymbolic_BAIJMUMPS(Mat F, Mat A, PETSC_UNUSED IS r, PETSC_UNUSED IS c, const MatFactorInfo *info)
2994: {
2995: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
2996: Vec b;
2997: const PetscInt M = A->rmap->N;
2999: PetscFunctionBegin;
3000: if (mumps->matstruc == SAME_NONZERO_PATTERN) {
3001: /* F is assembled by a previous call of MatLUFactorSymbolic_BAIJMUMPS() */
3002: PetscFunctionReturn(PETSC_SUCCESS);
3003: }
3005: /* Set MUMPS options from the options database */
3006: PetscCall(MatSetFromOptions_MUMPS(F, A));
3008: PetscCall((*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, mumps));
3009: PetscCall(MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX, mumps));
3010: PetscCall(MatMumpsEnsureSchurArray_Private(F));
3012: /* analysis phase */
3013: mumps->id.job = JOB_FACTSYMBOLIC;
3014: PetscCall(PetscMUMPSIntCast(M, &mumps->id.n));
3015: switch (mumps->id.ICNTL(18)) {
3016: case 0: /* centralized assembled matrix input */
3017: if (!mumps->myid) {
3018: mumps->id.nnz = mumps->nnz;
3019: mumps->id.irn = mumps->irn;
3020: mumps->id.jcn = mumps->jcn;
3021: if (1 < mumps->id.ICNTL(6) && mumps->id.ICNTL(6) < 7) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_len, &mumps->id.a));
3022: }
3023: break;
3024: case 3: /* distributed assembled matrix input (size>1) */
3025: mumps->id.nnz_loc = mumps->nnz;
3026: mumps->id.irn_loc = mumps->irn;
3027: mumps->id.jcn_loc = mumps->jcn;
3028: if (1 < mumps->id.ICNTL(6) && mumps->id.ICNTL(6) < 7) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_loc_len, &mumps->id.a_loc));
3029: if (mumps->ICNTL20 == 0) { /* Centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
3030: PetscCall(MatCreateVecs(A, NULL, &b));
3031: PetscCall(VecScatterCreateToZero(b, &mumps->scat_rhs, &mumps->b_seq));
3032: PetscCall(VecDestroy(&b));
3033: }
3034: break;
3035: }
3036: if (A->rmap->N && A->cmap->N) {
3037: PetscMUMPS_c(mumps);
3038: PetscCall(MatFactorSymbolic_MUMPS_ReportIfError(F, A, info, mumps));
3039: }
3040: F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
3041: F->ops->solve = MatSolve_MUMPS;
3042: F->ops->solvetranspose = MatSolveTranspose_MUMPS;
3043: F->ops->matsolve = MatMatSolve_MUMPS;
3044: F->ops->mattransposesolve = MatMatTransposeSolve_MUMPS;
3045: F->ops->matsolvetranspose = MatMatSolveTranspose_MUMPS;
3047: mumps->matstruc = SAME_NONZERO_PATTERN;
3048: PetscFunctionReturn(PETSC_SUCCESS);
3049: }
3051: /* Note the PETSc r permutation and factor info are ignored */
3052: static PetscErrorCode MatCholeskyFactorSymbolic_MUMPS(Mat F, Mat A, PETSC_UNUSED IS r, const MatFactorInfo *info)
3053: {
3054: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3055: Vec b;
3056: const PetscInt M = A->rmap->N;
3058: PetscFunctionBegin;
3059: if (mumps->matstruc == SAME_NONZERO_PATTERN) {
3060: /* F is assembled by a previous call of MatCholeskyFactorSymbolic_MUMPS() */
3061: PetscFunctionReturn(PETSC_SUCCESS);
3062: }
3064: /* Set MUMPS options from the options database */
3065: PetscCall(MatSetFromOptions_MUMPS(F, A));
3067: PetscCall((*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, mumps));
3068: PetscCall(MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX, mumps));
3069: PetscCall(MatMumpsEnsureSchurArray_Private(F));
3071: /* analysis phase */
3072: mumps->id.job = JOB_FACTSYMBOLIC;
3073: PetscCall(PetscMUMPSIntCast(M, &mumps->id.n));
3074: switch (mumps->id.ICNTL(18)) {
3075: case 0: /* centralized assembled matrix input */
3076: if (!mumps->myid) {
3077: mumps->id.nnz = mumps->nnz;
3078: mumps->id.irn = mumps->irn;
3079: mumps->id.jcn = mumps->jcn;
3080: if (1 < mumps->id.ICNTL(6) && mumps->id.ICNTL(6) < 7) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_len, &mumps->id.a));
3081: }
3082: break;
3083: case 3: /* distributed assembled matrix input (size>1) */
3084: mumps->id.nnz_loc = mumps->nnz;
3085: mumps->id.irn_loc = mumps->irn;
3086: mumps->id.jcn_loc = mumps->jcn;
3087: if (1 < mumps->id.ICNTL(6) && mumps->id.ICNTL(6) < 7) PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, mumps->nnz, mumps->val, mumps->id.precision, &mumps->id.a_loc_len, &mumps->id.a_loc));
3088: if (mumps->ICNTL20 == 0) { /* Centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
3089: PetscCall(MatCreateVecs(A, NULL, &b));
3090: PetscCall(VecScatterCreateToZero(b, &mumps->scat_rhs, &mumps->b_seq));
3091: PetscCall(VecDestroy(&b));
3092: }
3093: break;
3094: }
3095: if (A->rmap->N && A->cmap->N) {
3096: PetscMUMPS_c(mumps);
3097: PetscCall(MatFactorSymbolic_MUMPS_ReportIfError(F, A, info, mumps));
3098: }
3099: F->ops->choleskyfactornumeric = MatFactorNumeric_MUMPS;
3100: F->ops->solve = MatSolve_MUMPS;
3101: F->ops->solvetranspose = MatSolve_MUMPS;
3102: F->ops->matsolve = MatMatSolve_MUMPS;
3103: F->ops->mattransposesolve = MatMatTransposeSolve_MUMPS;
3104: F->ops->matsolvetranspose = MatMatSolveTranspose_MUMPS;
3105: #if defined(PETSC_USE_COMPLEX)
3106: F->ops->getinertia = NULL;
3107: #else
3108: F->ops->getinertia = MatGetInertia_SBAIJMUMPS;
3109: #endif
3111: mumps->matstruc = SAME_NONZERO_PATTERN;
3112: PetscFunctionReturn(PETSC_SUCCESS);
3113: }
3115: static PetscErrorCode MatView_MUMPS(Mat A, PetscViewer viewer)
3116: {
3117: PetscBool isascii;
3118: PetscViewerFormat format;
3119: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
3121: PetscFunctionBegin;
3122: /* check if matrix is mumps type */
3123: if (A->ops->solve != MatSolve_MUMPS) PetscFunctionReturn(PETSC_SUCCESS);
3125: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
3126: if (isascii) {
3127: PetscCall(PetscViewerGetFormat(viewer, &format));
3128: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
3129: PetscCall(PetscViewerASCIIPrintf(viewer, "MUMPS run parameters:\n"));
3130: PetscCall(PetscViewerASCIIPrintf(viewer, " precision: %s\n", PetscPrecisionTypes[mumps->id.precision]));
3131: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
3132: PetscCall(PetscViewerASCIIPrintf(viewer, " SYM (matrix type): %d\n", mumps->id.sym));
3133: PetscCall(PetscViewerASCIIPrintf(viewer, " PAR (host participation): %d\n", mumps->id.par));
3134: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(1) (output for error): %d\n", mumps->id.ICNTL(1)));
3135: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(2) (output of diagnostic msg): %d\n", mumps->id.ICNTL(2)));
3136: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(3) (output for global info): %d\n", mumps->id.ICNTL(3)));
3137: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(4) (level of printing): %d\n", mumps->id.ICNTL(4)));
3138: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(5) (input mat struct): %d\n", mumps->id.ICNTL(5)));
3139: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(6) (matrix prescaling): %d\n", mumps->id.ICNTL(6)));
3140: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(7) (sequential matrix ordering):%d\n", mumps->id.ICNTL(7)));
3141: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(8) (scaling strategy): %d\n", mumps->id.ICNTL(8)));
3142: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(10) (max num of refinements): %d\n", mumps->id.ICNTL(10)));
3143: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(11) (error analysis): %d\n", mumps->id.ICNTL(11)));
3144: if (mumps->id.ICNTL(11) > 0) {
3145: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(4) (inf norm of input mat): %g\n", (double)ID_RINFOG_GET(mumps->id, 4)));
3146: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(5) (inf norm of solution): %g\n", (double)ID_RINFOG_GET(mumps->id, 5)));
3147: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(6) (inf norm of residual): %g\n", (double)ID_RINFOG_GET(mumps->id, 6)));
3148: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(7),RINFOG(8) (backward error est): %g, %g\n", (double)ID_RINFOG_GET(mumps->id, 7), (double)ID_RINFOG_GET(mumps->id, 8)));
3149: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(9) (error estimate): %g\n", (double)ID_RINFOG_GET(mumps->id, 9)));
3150: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(10),RINFOG(11)(condition numbers): %g, %g\n", (double)ID_RINFOG_GET(mumps->id, 10), (double)ID_RINFOG_GET(mumps->id, 11)));
3151: }
3152: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(12) (efficiency control): %d\n", mumps->id.ICNTL(12)));
3153: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(13) (sequential factorization of the root node): %d\n", mumps->id.ICNTL(13)));
3154: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(14) (percentage of estimated workspace increase): %d\n", mumps->id.ICNTL(14)));
3155: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(15) (compression of the input matrix): %d\n", mumps->id.ICNTL(15)));
3156: /* ICNTL(15-17) not used */
3157: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(18) (input mat struct): %d\n", mumps->id.ICNTL(18)));
3158: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(19) (Schur complement info): %d\n", mumps->id.ICNTL(19)));
3159: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(20) (RHS sparse pattern): %d\n", mumps->id.ICNTL(20)));
3160: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(21) (solution struct): %d\n", mumps->id.ICNTL(21)));
3161: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(22) (in-core/out-of-core facility): %d\n", mumps->id.ICNTL(22)));
3162: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(23) (max size of memory allocated locally): %d\n", mumps->id.ICNTL(23)));
3164: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(24) (detection of null pivot rows): %d\n", mumps->id.ICNTL(24)));
3165: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(25) (computation of a null space basis): %d\n", mumps->id.ICNTL(25)));
3166: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(26) (Schur options for RHS or solution): %d\n", mumps->id.ICNTL(26)));
3167: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(27) (blocking size for multiple RHS): %d\n", mumps->id.ICNTL(27)));
3168: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(28) (use parallel or sequential ordering): %d\n", mumps->id.ICNTL(28)));
3169: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(29) (parallel ordering): %d\n", mumps->id.ICNTL(29)));
3171: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(30) (user-specified set of entries in inv(A)): %d\n", mumps->id.ICNTL(30)));
3172: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(31) (factors is discarded in the solve phase): %d\n", mumps->id.ICNTL(31)));
3173: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(33) (compute determinant): %d\n", mumps->id.ICNTL(33)));
3174: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(35) (activate BLR based factorization): %d\n", mumps->id.ICNTL(35)));
3175: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(36) (choice of BLR factorization variant): %d\n", mumps->id.ICNTL(36)));
3176: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(37) (compression of the contribution blocks): %d\n", mumps->id.ICNTL(37)));
3177: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(38) (estimated compression rate of LU factors): %d\n", mumps->id.ICNTL(38)));
3178: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(40) (adaptive BLR precision feature): %d\n", mumps->id.ICNTL(40)));
3179: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(47) (single precision factorization in a double precision instance): %d\n", mumps->id.ICNTL(47)));
3180: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(48) (multithreading with tree parallelism): %d\n", mumps->id.ICNTL(48)));
3181: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(49) (compact workarray at the end of factorization phase):%d\n", mumps->id.ICNTL(49)));
3182: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(56) (postponing and rank-revealing factorization):%d\n", mumps->id.ICNTL(56)));
3183: PetscCall(PetscViewerASCIIPrintf(viewer, " ICNTL(58) (options for symbolic factorization): %d\n", mumps->id.ICNTL(58)));
3185: PetscCall(PetscViewerASCIIPrintf(viewer, " CNTL(1) (relative pivoting threshold): %g\n", (double)ID_CNTL_GET(mumps->id, 1)));
3186: PetscCall(PetscViewerASCIIPrintf(viewer, " CNTL(2) (stopping criterion of refinement): %g\n", (double)ID_CNTL_GET(mumps->id, 2)));
3187: PetscCall(PetscViewerASCIIPrintf(viewer, " CNTL(3) (absolute pivoting threshold): %g\n", (double)ID_CNTL_GET(mumps->id, 3)));
3188: PetscCall(PetscViewerASCIIPrintf(viewer, " CNTL(4) (value of static pivoting): %g\n", (double)ID_CNTL_GET(mumps->id, 4)));
3189: PetscCall(PetscViewerASCIIPrintf(viewer, " CNTL(5) (fixation for null pivots): %g\n", (double)ID_CNTL_GET(mumps->id, 5)));
3190: PetscCall(PetscViewerASCIIPrintf(viewer, " CNTL(7) (dropping parameter for BLR): %g\n", (double)ID_CNTL_GET(mumps->id, 7)));
3192: /* information local to each processor */
3193: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFO(1) (local estimated flops for the elimination after analysis):\n"));
3194: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
3195: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %g\n", mumps->myid, (double)ID_RINFO_GET(mumps->id, 1)));
3196: PetscCall(PetscViewerFlush(viewer));
3197: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFO(2) (local estimated flops for the assembly after factorization):\n"));
3198: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %g\n", mumps->myid, (double)ID_RINFO_GET(mumps->id, 2)));
3199: PetscCall(PetscViewerFlush(viewer));
3200: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFO(3) (local estimated flops for the elimination after factorization):\n"));
3201: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %g\n", mumps->myid, (double)ID_RINFO_GET(mumps->id, 3)));
3202: PetscCall(PetscViewerFlush(viewer));
3204: PetscCall(PetscViewerASCIIPrintf(viewer, " INFO(15) (estimated size of (in MB) MUMPS internal data for running numerical factorization):\n"));
3205: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %d\n", mumps->myid, mumps->id.INFO(15)));
3206: PetscCall(PetscViewerFlush(viewer));
3208: PetscCall(PetscViewerASCIIPrintf(viewer, " INFO(16) (size of (in MB) MUMPS internal data used during numerical factorization):\n"));
3209: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %d\n", mumps->myid, mumps->id.INFO(16)));
3210: PetscCall(PetscViewerFlush(viewer));
3212: PetscCall(PetscViewerASCIIPrintf(viewer, " INFO(23) (num of pivots eliminated on this processor after factorization):\n"));
3213: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %d\n", mumps->myid, mumps->id.INFO(23)));
3214: PetscCall(PetscViewerFlush(viewer));
3216: if (mumps->ninfo && mumps->ninfo <= 80) {
3217: for (PetscInt i = 0; i < mumps->ninfo; i++) {
3218: PetscCall(PetscViewerASCIIPrintf(viewer, " INFO(%" PetscInt_FMT "):\n", mumps->info[i]));
3219: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] %d\n", mumps->myid, mumps->id.INFO(mumps->info[i])));
3220: PetscCall(PetscViewerFlush(viewer));
3221: }
3222: }
3223: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
3224: } else PetscCall(PetscViewerASCIIPrintf(viewer, " Use -%sksp_view ::ascii_info_detail to display information for all processes\n", ((PetscObject)A)->prefix ? ((PetscObject)A)->prefix : ""));
3226: if (mumps->myid == 0) { /* information from the host */
3227: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(1) (global estimated flops for the elimination after analysis): %g\n", (double)ID_RINFOG_GET(mumps->id, 1)));
3228: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(2) (global estimated flops for the assembly after factorization): %g\n", (double)ID_RINFOG_GET(mumps->id, 2)));
3229: PetscCall(PetscViewerASCIIPrintf(viewer, " RINFOG(3) (global estimated flops for the elimination after factorization): %g\n", (double)ID_RINFOG_GET(mumps->id, 3)));
3230: PetscCall(PetscViewerASCIIPrintf(viewer, " (RINFOG(12) RINFOG(13))*2^INFOG(34) (determinant): (%g,%g)*(2^%d)\n", (double)ID_RINFOG_GET(mumps->id, 12), (double)ID_RINFOG_GET(mumps->id, 13), mumps->id.INFOG(34)));
3232: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(3) (estimated real workspace for factors on all processors after analysis): %d\n", mumps->id.INFOG(3)));
3233: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(4) (estimated integer workspace for factors on all processors after analysis): %d\n", mumps->id.INFOG(4)));
3234: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(5) (estimated maximum front size in the complete tree): %d\n", mumps->id.INFOG(5)));
3235: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(6) (number of nodes in the complete tree): %d\n", mumps->id.INFOG(6)));
3236: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(7) (ordering option effectively used after analysis): %d\n", mumps->id.INFOG(7)));
3237: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(8) (structural symmetry in percent of the permuted matrix after analysis): %d\n", mumps->id.INFOG(8)));
3238: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(9) (total real/complex workspace to store the matrix factors after factorization): %d\n", mumps->id.INFOG(9)));
3239: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(10) (total integer space store the matrix factors after factorization): %d\n", mumps->id.INFOG(10)));
3240: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(11) (order of largest frontal matrix after factorization): %d\n", mumps->id.INFOG(11)));
3241: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(12) (number of off-diagonal pivots): %d\n", mumps->id.INFOG(12)));
3242: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(13) (number of delayed pivots after factorization): %d\n", mumps->id.INFOG(13)));
3243: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(14) (number of memory compress after factorization): %d\n", mumps->id.INFOG(14)));
3244: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(15) (number of steps of iterative refinement after solution): %d\n", mumps->id.INFOG(15)));
3245: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(16) (estimated size (in MB) of all MUMPS internal data for factorization after analysis: value on the most memory consuming processor): %d\n", mumps->id.INFOG(16)));
3246: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(17) (estimated size of all MUMPS internal data for factorization after analysis: sum over all processors): %d\n", mumps->id.INFOG(17)));
3247: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(18) (size of all MUMPS internal data allocated during factorization: value on the most memory consuming processor): %d\n", mumps->id.INFOG(18)));
3248: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(19) (size of all MUMPS internal data allocated during factorization: sum over all processors): %d\n", mumps->id.INFOG(19)));
3249: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(20) (estimated number of entries in the factors): %d\n", mumps->id.INFOG(20)));
3250: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(21) (size in MB of memory effectively used during factorization - value on the most memory consuming processor): %d\n", mumps->id.INFOG(21)));
3251: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(22) (size in MB of memory effectively used during factorization - sum over all processors): %d\n", mumps->id.INFOG(22)));
3252: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(23) (after analysis: value of ICNTL(6) effectively used): %d\n", mumps->id.INFOG(23)));
3253: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(24) (after analysis: value of ICNTL(12) effectively used): %d\n", mumps->id.INFOG(24)));
3254: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(25) (after factorization: number of pivots modified by static pivoting): %d\n", mumps->id.INFOG(25)));
3255: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(28) (after factorization: number of null pivots encountered): %d\n", mumps->id.INFOG(28)));
3256: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(29) (after factorization: effective number of entries in the factors (sum over all processors)): %d\n", mumps->id.INFOG(29)));
3257: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(30, 31) (after solution: size in Mbytes of memory used during solution phase): %d, %d\n", mumps->id.INFOG(30), mumps->id.INFOG(31)));
3258: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(32) (after analysis: type of analysis done): %d\n", mumps->id.INFOG(32)));
3259: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(33) (value used for ICNTL(8)): %d\n", mumps->id.INFOG(33)));
3260: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(34) (exponent of the determinant if determinant is requested): %d\n", mumps->id.INFOG(34)));
3261: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(35) (after factorization: number of entries taking into account BLR factor compression - sum over all processors): %d\n", mumps->id.INFOG(35)));
3262: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(36) (after analysis: estimated size of all MUMPS internal data for running BLR in-core - value on the most memory consuming processor): %d\n", mumps->id.INFOG(36)));
3263: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(37) (after analysis: estimated size of all MUMPS internal data for running BLR in-core - sum over all processors): %d\n", mumps->id.INFOG(37)));
3264: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(38) (after analysis: estimated size of all MUMPS internal data for running BLR out-of-core - value on the most memory consuming processor): %d\n", mumps->id.INFOG(38)));
3265: PetscCall(PetscViewerASCIIPrintf(viewer, " INFOG(39) (after analysis: estimated size of all MUMPS internal data for running BLR out-of-core - sum over all processors): %d\n", mumps->id.INFOG(39)));
3266: }
3267: }
3268: }
3269: PetscFunctionReturn(PETSC_SUCCESS);
3270: }
3272: static PetscErrorCode MatGetInfo_MUMPS(Mat A, PETSC_UNUSED MatInfoType flag, MatInfo *info)
3273: {
3274: Mat_MUMPS *mumps = (Mat_MUMPS *)A->data;
3276: PetscFunctionBegin;
3277: info->block_size = 1.0;
3278: info->nz_allocated = mumps->id.INFOG(20) >= 0 ? mumps->id.INFOG(20) : -1000000 * mumps->id.INFOG(20);
3279: info->nz_used = mumps->id.INFOG(20) >= 0 ? mumps->id.INFOG(20) : -1000000 * mumps->id.INFOG(20);
3280: info->nz_unneeded = 0.0;
3281: info->assemblies = 0.0;
3282: info->mallocs = 0.0;
3283: info->memory = 0.0;
3284: info->fill_ratio_given = 0;
3285: info->fill_ratio_needed = 0;
3286: info->factor_mallocs = 0;
3287: PetscFunctionReturn(PETSC_SUCCESS);
3288: }
3290: static PetscErrorCode MatFactorSetSchurIS_MUMPS(Mat F, IS is)
3291: {
3292: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3293: const PetscInt *idxs;
3294: PetscInt size, i;
3296: PetscFunctionBegin;
3297: PetscCall(MatMumpsResetSchur_Private(F));
3298: if (!is) PetscFunctionReturn(PETSC_SUCCESS);
3299: /* Schur complement matrix */
3300: PetscCall(ISGetLocalSize(is, &size));
3301: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, size, size, NULL, &F->schur));
3302: // don't allocate mumps->id.schur[] now as its precision is yet to be known
3303: PetscCall(PetscMUMPSIntCast(size, &mumps->id.size_schur));
3304: PetscCall(PetscMUMPSIntCast(size, &mumps->id.schur_lld));
3305: if (mumps->sym == 1) PetscCall(MatSetOption(F->schur, MAT_SPD, PETSC_TRUE));
3307: /* MUMPS expects Fortran style indices */
3308: PetscCall(PetscMalloc1(size, &mumps->id.listvar_schur));
3309: PetscCall(ISGetIndices(is, &idxs));
3310: for (i = 0; i < size; i++) PetscCall(PetscMUMPSIntCast(idxs[i] + 1, &mumps->id.listvar_schur[i]));
3311: PetscCall(ISRestoreIndices(is, &idxs));
3312: /* set a special value of ICNTL (not handled my MUMPS) to be used in the solve phase by PETSc */
3313: if (mumps->id.icntl) mumps->id.ICNTL(26) = -1;
3314: else mumps->ICNTL26 = -1;
3315: PetscFunctionReturn(PETSC_SUCCESS);
3316: }
3318: static PetscErrorCode MatFactorCreateSchurComplement_MUMPS(Mat F, Mat *S)
3319: {
3320: Mat St;
3321: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3322: PetscScalar *array;
3323: PetscInt i, j, N = mumps->id.size_schur;
3325: PetscFunctionBegin;
3326: PetscCheck(mumps->id.ICNTL(19), PetscObjectComm((PetscObject)F), PETSC_ERR_ORDER, "Schur complement mode not selected! Call MatFactorSetSchurIS() to enable it");
3327: PetscCall(MatCreate(PETSC_COMM_SELF, &St));
3328: PetscCall(MatSetSizes(St, PETSC_DECIDE, PETSC_DECIDE, mumps->id.size_schur, mumps->id.size_schur));
3329: PetscCall(MatSetType(St, MATDENSE));
3330: PetscCall(MatSetUp(St));
3331: PetscCall(MatDenseGetArray(St, &array));
3332: PetscCall(MatMumpsEnsureSchurArray_Private(F));
3333: if (!mumps->sym) { /* MUMPS always returns a full matrix */
3334: if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
3335: for (i = 0; i < N; i++) {
3336: for (j = 0; j < N; j++) array[j * N + i] = ID_FIELD_GET(mumps->id, schur, i * N + j);
3337: }
3338: } else { /* stored by columns */
3339: PetscCall(MatMumpsCastMumpsScalarArray(N * N, mumps->id.precision, mumps->id.schur, array));
3340: }
3341: } else { /* either full or lower-triangular (not packed) */
3342: if (mumps->id.ICNTL(19) == 2) { /* lower triangular stored by columns */
3343: for (i = 0; i < N; i++) {
3344: for (j = i; j < N; j++) array[i * N + j] = array[j * N + i] = ID_FIELD_GET(mumps->id, schur, i * N + j);
3345: }
3346: } else if (mumps->id.ICNTL(19) == 3) { /* full matrix */
3347: PetscCall(MatMumpsCastMumpsScalarArray(N * N, mumps->id.precision, mumps->id.schur, array));
3348: } else { /* ICNTL(19) == 1 lower triangular stored by rows */
3349: for (i = 0; i < N; i++) {
3350: for (j = 0; j < i + 1; j++) array[i * N + j] = array[j * N + i] = ID_FIELD_GET(mumps->id, schur, i * N + j);
3351: }
3352: }
3353: }
3354: PetscCall(MatDenseRestoreArray(St, &array));
3355: *S = St;
3356: PetscFunctionReturn(PETSC_SUCCESS);
3357: }
3359: static PetscErrorCode MatMumpsSetIcntl_MUMPS(Mat F, PetscInt icntl, PetscInt ival)
3360: {
3361: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3363: PetscFunctionBegin;
3364: PetscCheck((icntl >= 1 && icntl <= 38) || icntl == 40 || icntl == 47 || icntl == 48 || icntl == 49 || icntl == 56 || icntl == 58, PetscObjectComm((PetscObject)F), PETSC_ERR_ARG_WRONG, "Unsupported ICNTL value %" PetscInt_FMT, icntl);
3365: if (mumps->id.job == JOB_NULL) { /* need to cache icntl and ival since PetscMUMPS_c() has never been called */
3366: PetscMUMPSInt i, nICNTL_pre = mumps->ICNTL_pre ? mumps->ICNTL_pre[0] : 0; /* number of already cached ICNTL */
3367: for (i = 0; i < nICNTL_pre; ++i)
3368: if (mumps->ICNTL_pre[1 + 2 * i] == icntl) break; /* is this ICNTL already cached? */
3369: if (i == nICNTL_pre) { /* not already cached */
3370: if (i > 0) PetscCall(PetscRealloc(sizeof(PetscMUMPSInt) * (2 * nICNTL_pre + 3), &mumps->ICNTL_pre));
3371: else PetscCall(PetscCalloc(sizeof(PetscMUMPSInt) * 3, &mumps->ICNTL_pre));
3372: mumps->ICNTL_pre[0]++;
3373: }
3374: mumps->ICNTL_pre[1 + 2 * i] = (PetscMUMPSInt)icntl;
3375: PetscCall(PetscMUMPSIntCast(ival, mumps->ICNTL_pre + 2 + 2 * i));
3376: } else PetscCall(PetscMUMPSIntCast(ival, &mumps->id.ICNTL(icntl)));
3377: PetscFunctionReturn(PETSC_SUCCESS);
3378: }
3380: static PetscErrorCode MatMumpsGetIcntl_MUMPS(Mat F, PetscInt icntl, PetscInt *ival)
3381: {
3382: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3384: PetscFunctionBegin;
3385: PetscCheck((icntl >= 1 && icntl <= 38) || icntl == 40 || icntl == 47 || icntl == 48 || icntl == 49 || icntl == 56 || icntl == 58, PetscObjectComm((PetscObject)F), PETSC_ERR_ARG_WRONG, "Unsupported ICNTL value %" PetscInt_FMT, icntl);
3386: if (mumps->id.job == JOB_NULL) {
3387: PetscInt i, nICNTL_pre = mumps->ICNTL_pre ? mumps->ICNTL_pre[0] : 0;
3388: *ival = 0;
3389: for (i = 0; i < nICNTL_pre; ++i) {
3390: if (mumps->ICNTL_pre[1 + 2 * i] == icntl) *ival = mumps->ICNTL_pre[2 + 2 * i];
3391: }
3392: } else *ival = mumps->id.ICNTL(icntl);
3393: PetscFunctionReturn(PETSC_SUCCESS);
3394: }
3396: static PetscErrorCode MatMumpsSetCntl_MUMPS(Mat F, PetscInt icntl, PetscReal val)
3397: {
3398: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3400: PetscFunctionBegin;
3401: PetscCheck(icntl >= 1 && icntl <= 7, PetscObjectComm((PetscObject)F), PETSC_ERR_ARG_WRONG, "Unsupported CNTL value %" PetscInt_FMT, icntl);
3402: if (mumps->id.job == JOB_NULL) {
3403: PetscInt i, nCNTL_pre = mumps->CNTL_pre ? mumps->CNTL_pre[0] : 0;
3404: for (i = 0; i < nCNTL_pre; ++i)
3405: if (mumps->CNTL_pre[1 + 2 * i] == icntl) break;
3406: if (i == nCNTL_pre) {
3407: if (i > 0) PetscCall(PetscRealloc(sizeof(PetscReal) * (2 * nCNTL_pre + 3), &mumps->CNTL_pre));
3408: else PetscCall(PetscCalloc(sizeof(PetscReal) * 3, &mumps->CNTL_pre));
3409: mumps->CNTL_pre[0]++;
3410: }
3411: mumps->CNTL_pre[1 + 2 * i] = icntl;
3412: mumps->CNTL_pre[2 + 2 * i] = val;
3413: } else ID_CNTL_SET(mumps->id, icntl, val);
3414: PetscFunctionReturn(PETSC_SUCCESS);
3415: }
3417: static PetscErrorCode MatMumpsGetCntl_MUMPS(Mat F, PetscInt icntl, PetscReal *val)
3418: {
3419: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3421: PetscFunctionBegin;
3422: PetscCheck(icntl >= 1 && icntl <= 7, PetscObjectComm((PetscObject)F), PETSC_ERR_ARG_WRONG, "Unsupported CNTL value %" PetscInt_FMT, icntl);
3423: if (mumps->id.job == JOB_NULL) {
3424: PetscInt i, nCNTL_pre = mumps->CNTL_pre ? mumps->CNTL_pre[0] : 0;
3425: *val = 0.0;
3426: for (i = 0; i < nCNTL_pre; ++i) {
3427: if (mumps->CNTL_pre[1 + 2 * i] == icntl) *val = mumps->CNTL_pre[2 + 2 * i];
3428: }
3429: } else *val = ID_CNTL_GET(mumps->id, icntl);
3430: PetscFunctionReturn(PETSC_SUCCESS);
3431: }
3433: /*@C
3434: MatMumpsSetOocTmpDir - Set MUMPS out-of-core `OOC_TMPDIR` <https://mumps-solver.org/index.php?page=doc>
3436: Logically Collective
3438: Input Parameters:
3439: + F - the factored matrix obtained by calling `MatGetFactor()` with a `MatSolverType` of `MATSOLVERMUMPS` and a `MatFactorType` of `MAT_FACTOR_LU` or `MAT_FACTOR_CHOLESKY`.
3440: - tmpdir - temporary directory for out-of-core facility.
3442: Level: beginner
3444: Note:
3445: To make it effective, this routine must be called before the numeric factorization, i.e., `PCSetUp()`.
3446: If `ooc_tmpdir` is not set, MUMPS will also check the environment variable `MUMPS_OOC_TMPDIR`. But if neither was defined, it will use /tmp by default.
3448: .seealso: [](ch_matrices), `Mat`, `MatGetFactor()`, `MatMumpsGetOocTmpDir`, `MatMumpsSetIcntl()`, `MatMumpsGetIcntl()`, `MatMumpsSetCntl()`, `MatMumpsGetInfo()`, `MatMumpsGetInfog()`, `MatMumpsGetRinfo()`, `MatMumpsGetRinfog()`
3449: @*/
3450: PetscErrorCode MatMumpsSetOocTmpDir(Mat F, const char *tmpdir)
3451: {
3452: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3454: PetscFunctionBegin;
3457: PetscCall(PetscStrncpy(mumps->id.ooc_tmpdir, tmpdir, sizeof(((MUMPS_STRUC_C *)NULL)->ooc_tmpdir)));
3458: PetscFunctionReturn(PETSC_SUCCESS);
3459: }
3461: /*@C
3462: MatMumpsGetOocTmpDir - Get MUMPS out-of-core `OOC_TMPDIR` <https://mumps-solver.org/index.php?page=doc>
3464: Logically Collective
3466: Input Parameter:
3467: . F - the factored matrix obtained by calling `MatGetFactor()` with a `MatSolverType` of `MATSOLVERMUMPS` and a `MatFactorType` of `MAT_FACTOR_LU` or `MAT_FACTOR_CHOLESKY`.
3469: Output Parameter:
3470: . tmpdir - temporary directory for out-of-core facility.
3472: Level: beginner
3474: Note:
3475: The returned string is read-only and user should not try to change it.
3477: .seealso: [](ch_matrices), `Mat`, `MatGetFactor()`, `MatMumpsSetOocTmpDir`, `MatMumpsSetIcntl()`, `MatMumpsGetIcntl()`, `MatMumpsSetCntl()`, `MatMumpsGetInfo()`, `MatMumpsGetInfog()`, `MatMumpsGetRinfo()`, `MatMumpsGetRinfog()`
3478: @*/
3479: PetscErrorCode MatMumpsGetOocTmpDir(Mat F, const char *tmpdir[])
3480: {
3481: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3483: PetscFunctionBegin;
3486: if (tmpdir) *tmpdir = mumps->id.ooc_tmpdir;
3487: PetscFunctionReturn(PETSC_SUCCESS);
3488: }
3490: static PetscErrorCode MatMumpsGetInfo_MUMPS(Mat F, PetscInt icntl, PetscInt *info)
3491: {
3492: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3494: PetscFunctionBegin;
3495: *info = mumps->id.INFO(icntl);
3496: PetscFunctionReturn(PETSC_SUCCESS);
3497: }
3499: static PetscErrorCode MatMumpsGetInfog_MUMPS(Mat F, PetscInt icntl, PetscInt *infog)
3500: {
3501: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3503: PetscFunctionBegin;
3504: *infog = mumps->id.INFOG(icntl);
3505: PetscFunctionReturn(PETSC_SUCCESS);
3506: }
3508: static PetscErrorCode MatMumpsGetRinfo_MUMPS(Mat F, PetscInt icntl, PetscReal *rinfo)
3509: {
3510: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3512: PetscFunctionBegin;
3513: *rinfo = ID_RINFO_GET(mumps->id, icntl);
3514: PetscFunctionReturn(PETSC_SUCCESS);
3515: }
3517: static PetscErrorCode MatMumpsGetRinfog_MUMPS(Mat F, PetscInt icntl, PetscReal *rinfog)
3518: {
3519: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3521: PetscFunctionBegin;
3522: *rinfog = ID_RINFOG_GET(mumps->id, icntl);
3523: PetscFunctionReturn(PETSC_SUCCESS);
3524: }
3526: static PetscErrorCode MatMumpsGetNullPivots_MUMPS(Mat F, PetscInt *size, PetscInt **array)
3527: {
3528: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3530: PetscFunctionBegin;
3531: PetscCheck(mumps->id.ICNTL(24) == 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-mat_mumps_icntl_24 must be set as 1 for null pivot row detection");
3532: *size = 0;
3533: *array = NULL;
3534: if (!mumps->myid) {
3535: *size = mumps->id.INFOG(28);
3536: PetscCall(PetscMalloc1(*size, array));
3537: for (int i = 0; i < *size; i++) (*array)[i] = mumps->id.pivnul_list[i] - 1;
3538: }
3539: PetscFunctionReturn(PETSC_SUCCESS);
3540: }
3542: static PetscErrorCode MatMumpsGetInverse_MUMPS(Mat F, Mat spRHS)
3543: {
3544: Mat Bt = NULL, Btseq = NULL;
3545: PetscBool flg;
3546: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3547: PetscScalar *aa;
3548: PetscInt spnr, *ia, *ja, M, nrhs;
3550: PetscFunctionBegin;
3551: PetscAssertPointer(spRHS, 2);
3552: PetscCall(PetscObjectTypeCompare((PetscObject)spRHS, MATTRANSPOSEVIRTUAL, &flg));
3553: PetscCheck(flg, PetscObjectComm((PetscObject)spRHS), PETSC_ERR_ARG_WRONG, "Matrix spRHS must be type MATTRANSPOSEVIRTUAL matrix");
3554: PetscCall(MatShellGetScalingShifts(spRHS, (PetscScalar *)MAT_SHELL_NOT_ALLOWED, (PetscScalar *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Vec *)MAT_SHELL_NOT_ALLOWED, (Mat *)MAT_SHELL_NOT_ALLOWED, (IS *)MAT_SHELL_NOT_ALLOWED, (IS *)MAT_SHELL_NOT_ALLOWED));
3555: PetscCall(MatTransposeGetMat(spRHS, &Bt));
3557: PetscCall(MatMumpsSetIcntl(F, 30, 1));
3559: if (mumps->petsc_size > 1) {
3560: Mat_MPIAIJ *b = (Mat_MPIAIJ *)Bt->data;
3561: Btseq = b->A;
3562: } else {
3563: Btseq = Bt;
3564: }
3566: PetscCall(MatGetSize(spRHS, &M, &nrhs));
3567: mumps->id.nrhs = (PetscMUMPSInt)nrhs;
3568: PetscCall(PetscMUMPSIntCast(M, &mumps->id.lrhs));
3569: mumps->id.rhs = NULL;
3571: if (!mumps->myid) {
3572: PetscCall(MatSeqAIJGetArray(Btseq, &aa));
3573: PetscCall(MatGetRowIJ(Btseq, 1, PETSC_FALSE, PETSC_FALSE, &spnr, (const PetscInt **)&ia, (const PetscInt **)&ja, &flg));
3574: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot get IJ structure");
3575: PetscCall(PetscMUMPSIntCSRCast(mumps, spnr, ia, ja, &mumps->id.irhs_ptr, &mumps->id.irhs_sparse, &mumps->id.nz_rhs));
3576: PetscCall(MatMumpsMakeMumpsScalarArray(PETSC_TRUE, ((Mat_SeqAIJ *)Btseq->data)->nz, aa, mumps->id.precision, &mumps->id.rhs_sparse_len, &mumps->id.rhs_sparse));
3577: } else {
3578: mumps->id.irhs_ptr = NULL;
3579: mumps->id.irhs_sparse = NULL;
3580: mumps->id.nz_rhs = 0;
3581: if (mumps->id.rhs_sparse_len) {
3582: PetscCall(PetscFree(mumps->id.rhs_sparse));
3583: mumps->id.rhs_sparse_len = 0;
3584: }
3585: }
3586: mumps->id.ICNTL(20) = 1; /* rhs is sparse */
3587: mumps->id.ICNTL(21) = 0; /* solution is in assembled centralized format */
3589: /* solve phase */
3590: mumps->id.job = JOB_SOLVE;
3591: PetscMUMPS_c(mumps);
3592: PetscCheck(mumps->id.INFOG(1) >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "MUMPS error in solve: INFOG(1)=%d, INFO(2)=%d " MUMPS_MANUALS, mumps->id.INFOG(1), mumps->id.INFO(2));
3594: if (!mumps->myid) {
3595: PetscCall(MatSeqAIJRestoreArray(Btseq, &aa));
3596: PetscCall(MatRestoreRowIJ(Btseq, 1, PETSC_FALSE, PETSC_FALSE, &spnr, (const PetscInt **)&ia, (const PetscInt **)&ja, &flg));
3597: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot restore IJ structure");
3598: }
3599: PetscFunctionReturn(PETSC_SUCCESS);
3600: }
3602: static PetscErrorCode MatMumpsGetInverseTranspose_MUMPS(Mat F, Mat spRHST)
3603: {
3604: Mat spRHS;
3606: PetscFunctionBegin;
3607: PetscCall(MatCreateTranspose(spRHST, &spRHS));
3608: PetscCall(MatMumpsGetInverse_MUMPS(F, spRHS));
3609: PetscCall(MatDestroy(&spRHS));
3610: PetscFunctionReturn(PETSC_SUCCESS);
3611: }
3613: static PetscErrorCode MatMumpsSetBlk_MUMPS(Mat F, PetscInt nblk, const PetscInt blkvar[], const PetscInt blkptr[])
3614: {
3615: Mat_MUMPS *mumps = (Mat_MUMPS *)F->data;
3617: PetscFunctionBegin;
3618: if (nblk) {
3619: PetscAssertPointer(blkptr, 4);
3620: PetscCall(PetscMUMPSIntCast(nblk, &mumps->id.nblk));
3621: PetscCall(PetscFree(mumps->id.blkptr));
3622: PetscCall(PetscMalloc1(nblk + 1, &mumps->id.blkptr));
3623: for (PetscInt i = 0; i < nblk + 1; ++i) PetscCall(PetscMUMPSIntCast(blkptr[i], mumps->id.blkptr + i));
3624: // mumps->id.icntl[] might have not been allocated, which is done in MatSetFromOptions_MUMPS(). So we don't assign ICNTL(15).
3625: // We use id.nblk and id.blkptr to know what values to set to ICNTL(15) in MatSetFromOptions_MUMPS().
3626: // mumps->id.ICNTL(15) = 1;
3627: if (blkvar) {
3628: PetscCall(PetscFree(mumps->id.blkvar));
3629: PetscCall(PetscMalloc1(F->rmap->N, &mumps->id.blkvar));
3630: for (PetscInt i = 0; i < F->rmap->N; ++i) PetscCall(PetscMUMPSIntCast(blkvar[i], mumps->id.blkvar + i));
3631: }
3632: } else {
3633: PetscCall(PetscFree(mumps->id.blkptr));
3634: PetscCall(PetscFree(mumps->id.blkvar));
3635: // mumps->id.ICNTL(15) = 0;
3636: mumps->id.nblk = 0;
3637: }
3638: PetscFunctionReturn(PETSC_SUCCESS);
3639: }
3641: /*MC
3642: MATSOLVERMUMPS - A matrix type providing direct solvers (LU and Cholesky) for
3643: MPI distributed and sequential matrices via the external package MUMPS <https://mumps-solver.org/index.php?page=doc>
3645: Works with `MATAIJ` and `MATSBAIJ` matrices
3647: Use ./configure --download-mumps --download-scalapack --download-parmetis --download-metis --download-ptscotch to have PETSc installed with MUMPS
3649: Use ./configure --with-openmp --download-hwloc (or --with-hwloc) to enable running MUMPS in MPI+OpenMP hybrid mode and non-MUMPS in flat-MPI mode.
3650: See details below.
3652: Use `-pc_type cholesky` or `lu` `-pc_factor_mat_solver_type mumps` to use this direct solver
3654: Options Database Keys:
3655: + -mat_mumps_icntl_1 - ICNTL(1): output stream for error messages
3656: . -mat_mumps_icntl_2 - ICNTL(2): output stream for diagnostic printing, statistics, and warning
3657: . -mat_mumps_icntl_3 - ICNTL(3): output stream for global information, collected on the host
3658: . -mat_mumps_icntl_4 - ICNTL(4): level of printing (0 to 4)
3659: . -mat_mumps_icntl_6 - ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)
3660: . -mat_mumps_icntl_7 - ICNTL(7): computes a symmetric permutation in sequential analysis, 0=AMD, 2=AMF, 3=Scotch, 4=PORD, 5=Metis, 6=QAMD, and 7=auto
3661: Use -pc_factor_mat_ordering_type type to have PETSc perform the ordering (sequential only)
3662: . -mat_mumps_icntl_8 - ICNTL(8): scaling strategy (-2 to 8 or 77)
3663: . -mat_mumps_icntl_10 - ICNTL(10): max num of refinements
3664: . -mat_mumps_icntl_11 - ICNTL(11): statistics related to an error analysis (via -ksp_view)
3665: . -mat_mumps_icntl_12 - ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)
3666: . -mat_mumps_icntl_13 - ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting
3667: . -mat_mumps_icntl_14 - ICNTL(14): percentage increase in the estimated working space
3668: . -mat_mumps_icntl_15 - ICNTL(15): compression of the input matrix resulting from a block format
3669: . -mat_mumps_icntl_19 - ICNTL(19): computes the Schur complement
3670: . -mat_mumps_icntl_20 - ICNTL(20): give MUMPS centralized (0) or distributed (10) dense RHS
3671: . -mat_mumps_icntl_22 - ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)
3672: . -mat_mumps_icntl_23 - ICNTL(23): max size of the working memory (MB) that can allocate per processor
3673: . -mat_mumps_icntl_24 - ICNTL(24): detection of null pivot rows (0 or 1)
3674: . -mat_mumps_icntl_25 - ICNTL(25): compute a solution of a deficient matrix and a null space basis
3675: . -mat_mumps_icntl_26 - ICNTL(26): drives the solution phase if a Schur complement matrix
3676: . -mat_mumps_icntl_28 - ICNTL(28): use 1 for sequential analysis and ICNTL(7) ordering, or 2 for parallel analysis and ICNTL(29) ordering
3677: . -mat_mumps_icntl_29 - ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis
3678: . -mat_mumps_icntl_30 - ICNTL(30): compute user-specified set of entries in inv(A)
3679: . -mat_mumps_icntl_31 - ICNTL(31): indicates which factors may be discarded during factorization
3680: . -mat_mumps_icntl_33 - ICNTL(33): compute determinant
3681: . -mat_mumps_icntl_35 - ICNTL(35): level of activation of BLR (Block Low-Rank) feature
3682: . -mat_mumps_icntl_36 - ICNTL(36): controls the choice of BLR factorization variant
3683: . -mat_mumps_icntl_37 - ICNTL(37): compression of the contribution blocks (CB)
3684: . -mat_mumps_icntl_38 - ICNTL(38): sets the estimated compression rate of LU factors with BLR
3685: . -mat_mumps_icntl_40 - ICNTL(40): adaptive BLR precision feature
3686: . -mat_mumps_icntl_47 - ICNTL(47): single precision factorization in a double precision instance
3687: . -mat_mumps_icntl_48 - ICNTL(48): multithreading with tree parallelism
3688: . -mat_mumps_icntl_49 - ICNTL(49): compact workarray at the end of factorization phase
3689: . -mat_mumps_icntl_58 - ICNTL(58): options for symbolic factorization
3690: . -mat_mumps_cntl_1 - CNTL(1): relative pivoting threshold
3691: . -mat_mumps_cntl_2 - CNTL(2): stopping criterion of refinement
3692: . -mat_mumps_cntl_3 - CNTL(3): absolute pivoting threshold
3693: . -mat_mumps_cntl_4 - CNTL(4): value for static pivoting
3694: . -mat_mumps_cntl_5 - CNTL(5): fixation for null pivots
3695: . -mat_mumps_cntl_7 - CNTL(7): precision of the dropping parameter used during BLR factorization
3696: - -mat_mumps_use_omp_threads m - run MUMPS in MPI+OpenMP hybrid mode as if omp_set_num_threads(m) is called before calling MUMPS.
3697: Default might be the number of cores per CPU package (socket) as reported by hwloc and suggested by the MUMPS manual.
3699: Level: beginner
3701: Notes:
3702: MUMPS Cholesky does not handle (complex) Hermitian matrices (see User's Guide at <https://mumps-solver.org/index.php?page=doc>) so using it will
3703: error if the matrix is Hermitian.
3705: When used within a `KSP`/`PC` solve the options are prefixed with that of the `PC`. Otherwise one can set the options prefix by calling
3706: `MatSetOptionsPrefixFactor()` on the matrix from which the factor was obtained or `MatSetOptionsPrefix()` on the factor matrix.
3708: When a MUMPS factorization fails inside a KSP solve, for example with a `KSP_DIVERGED_PC_FAILED`, one can find the MUMPS information about
3709: the failure with
3710: .vb
3711: KSPGetPC(ksp,&pc);
3712: PCFactorGetMatrix(pc,&mat);
3713: MatMumpsGetInfo(mat,....);
3714: MatMumpsGetInfog(mat,....); etc.
3715: .ve
3716: Or run with `-ksp_error_if_not_converged` and the program will be stopped and the information printed in the error message.
3718: MUMPS provides 64-bit integer support in two build modes:
3719: full 64-bit: here MUMPS is built with C preprocessing flag -DINTSIZE64 and Fortran compiler option -i8, -fdefault-integer-8 or equivalent, and
3720: requires all dependent libraries MPI, ScaLAPACK, LAPACK and BLAS built the same way with 64-bit integers (for example ILP64 Intel MKL and MPI).
3722: selective 64-bit: with the default MUMPS build, 64-bit integers have been introduced where needed. In compressed sparse row (CSR) storage of matrices,
3723: MUMPS stores column indices in 32-bit, but row offsets in 64-bit, so you can have a huge number of non-zeros, but must have less than 2^31 rows and
3724: columns. This can lead to significant memory and performance gains with respect to a full 64-bit integer MUMPS version. This requires a regular (32-bit
3725: integer) build of all dependent libraries MPI, ScaLAPACK, LAPACK and BLAS.
3727: With --download-mumps=1, PETSc always build MUMPS in selective 64-bit mode, which can be used by both --with-64-bit-indices=0/1 variants of PETSc.
3729: Two modes to run MUMPS/PETSc with OpenMP
3730: .vb
3731: Set `OMP_NUM_THREADS` and run with fewer MPI ranks than cores. For example, if you want to have 16 OpenMP
3732: threads per rank, then you may use "export `OMP_NUM_THREADS` = 16 && mpiexec -n 4 ./test".
3733: .ve
3735: .vb
3736: `-mat_mumps_use_omp_threads` [m] and run your code with as many MPI ranks as the number of cores. For example,
3737: if a compute node has 32 cores and you run on two nodes, you may use "mpiexec -n 64 ./test -mat_mumps_use_omp_threads 16"
3738: .ve
3740: To run MUMPS in MPI+OpenMP hybrid mode (i.e., enable multithreading in MUMPS), but still run the non-MUMPS part
3741: (i.e., PETSc part) of your code in the so-called flat-MPI (aka pure-MPI) mode, you need to configure PETSc with `--with-openmp` `--download-hwloc`
3742: (or `--with-hwloc`), and have an MPI that supports MPI-3.0's process shared memory (which is usually available). Since MUMPS calls BLAS
3743: libraries, to really get performance, you should have multithreaded BLAS libraries such as Intel MKL, AMD ACML, Cray libSci or OpenBLAS
3744: (PETSc will automatically try to utilized a threaded BLAS if `--with-openmp` is provided).
3746: If you run your code through a job submission system, there are caveats in MPI rank mapping. We use MPI_Comm_split_type() to obtain MPI
3747: processes on each compute node. Listing the processes in rank ascending order, we split processes on a node into consecutive groups of
3748: size m and create a communicator called omp_comm for each group. Rank 0 in an omp_comm is called the master rank, and others in the omp_comm
3749: are called slave ranks (or slaves). Only master ranks are seen to MUMPS and slaves are not. We will free CPUs assigned to slaves (might be set
3750: by CPU binding policies in job scripts) and make the CPUs available to the master so that OMP threads spawned by MUMPS can run on the CPUs.
3751: In a multi-socket compute node, MPI rank mapping is an issue. Still use the above example and suppose your compute node has two sockets,
3752: if you interleave MPI ranks on the two sockets, in other words, even ranks are placed on socket 0, and odd ranks are on socket 1, and bind
3753: MPI ranks to cores, then with `-mat_mumps_use_omp_threads` 16, a master rank (and threads it spawns) will use half cores in socket 0, and half
3754: cores in socket 1, that definitely hurts locality. On the other hand, if you map MPI ranks consecutively on the two sockets, then the
3755: problem will not happen. Therefore, when you use `-mat_mumps_use_omp_threads`, you need to keep an eye on your MPI rank mapping and CPU binding.
3756: For example, with the Slurm job scheduler, one can use srun `--cpu-bind`=verbose -m block:block to map consecutive MPI ranks to sockets and
3757: examine the mapping result.
3759: PETSc does not control thread binding in MUMPS. So to get best performance, one still has to set `OMP_PROC_BIND` and `OMP_PLACES` in job scripts,
3760: for example, export `OMP_PLACES`=threads and export `OMP_PROC_BIND`=spread. One does not need to export `OMP_NUM_THREADS`=m in job scripts as PETSc
3761: calls `omp_set_num_threads`(m) internally before calling MUMPS.
3763: See {cite}`heroux2011bi` and {cite}`gutierrez2017accommodating`
3765: .seealso: [](ch_matrices), `Mat`, `PCFactorSetMatSolverType()`, `MatSolverType`, `MatMumpsSetIcntl()`, `MatMumpsGetIcntl()`, `MatMumpsSetCntl()`, `MatMumpsGetCntl()`, `MatMumpsGetInfo()`, `MatMumpsGetInfog()`, `MatMumpsGetRinfo()`, `MatMumpsGetRinfog()`, `MatMumpsSetBlk()`, `KSPGetPC()`, `PCFactorGetMatrix()`
3766: M*/
3768: static PetscErrorCode MatFactorGetSolverType_mumps(PETSC_UNUSED Mat A, MatSolverType *type)
3769: {
3770: PetscFunctionBegin;
3771: *type = MATSOLVERMUMPS;
3772: PetscFunctionReturn(PETSC_SUCCESS);
3773: }
3775: /* MatGetFactor for Seq and MPI AIJ matrices */
3776: static PetscErrorCode MatGetFactor_aij_mumps(Mat A, MatFactorType ftype, Mat *F)
3777: {
3778: Mat B;
3779: Mat_MUMPS *mumps;
3780: PetscBool isSeqAIJ, isDiag, isDense;
3781: PetscMPIInt size;
3783: PetscFunctionBegin;
3784: if (PetscDefined(USE_COMPLEX) && ftype == MAT_FACTOR_CHOLESKY && A->hermitian == PETSC_BOOL3_TRUE && A->symmetric != PETSC_BOOL3_TRUE) {
3785: PetscCall(PetscInfo(A, "Hermitian MAT_FACTOR_CHOLESKY is not supported. Use MAT_FACTOR_LU instead.\n"));
3786: *F = NULL;
3787: PetscFunctionReturn(PETSC_SUCCESS);
3788: }
3789: /* Create the factorization matrix */
3790: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isSeqAIJ));
3791: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATDIAGONAL, &isDiag));
3792: PetscCall(PetscObjectTypeCompareAny((PetscObject)A, &isDense, MATSEQDENSE, MATMPIDENSE, NULL));
3793: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
3794: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
3795: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &((PetscObject)B)->type_name));
3796: PetscCall(MatSetUp(B));
3798: PetscCall(PetscNew(&mumps));
3800: B->ops->view = MatView_MUMPS;
3801: B->ops->getinfo = MatGetInfo_MUMPS;
3803: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mumps));
3804: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MUMPS));
3805: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorCreateSchurComplement_C", MatFactorCreateSchurComplement_MUMPS));
3806: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetIcntl_C", MatMumpsSetIcntl_MUMPS));
3807: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetIcntl_C", MatMumpsGetIcntl_MUMPS));
3808: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetCntl_C", MatMumpsSetCntl_MUMPS));
3809: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetCntl_C", MatMumpsGetCntl_MUMPS));
3810: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfo_C", MatMumpsGetInfo_MUMPS));
3811: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfog_C", MatMumpsGetInfog_MUMPS));
3812: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfo_C", MatMumpsGetRinfo_MUMPS));
3813: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfog_C", MatMumpsGetRinfog_MUMPS));
3814: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetNullPivots_C", MatMumpsGetNullPivots_MUMPS));
3815: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverse_C", MatMumpsGetInverse_MUMPS));
3816: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverseTranspose_C", MatMumpsGetInverseTranspose_MUMPS));
3817: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetBlk_C", MatMumpsSetBlk_MUMPS));
3819: if (ftype == MAT_FACTOR_LU) {
3820: B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
3821: B->factortype = MAT_FACTOR_LU;
3822: if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqaij;
3823: else if (isDiag) mumps->ConvertToTriples = MatConvertToTriples_diagonal_xaij;
3824: else if (isDense) mumps->ConvertToTriples = MatConvertToTriples_dense_xaij;
3825: else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpiaij;
3826: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_LU]));
3827: mumps->sym = 0;
3828: } else {
3829: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
3830: B->factortype = MAT_FACTOR_CHOLESKY;
3831: if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqsbaij;
3832: else if (isDiag) mumps->ConvertToTriples = MatConvertToTriples_diagonal_xaij;
3833: else if (isDense) mumps->ConvertToTriples = MatConvertToTriples_dense_xaij;
3834: else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpisbaij;
3835: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_CHOLESKY]));
3836: if (PetscDefined(USE_COMPLEX)) mumps->sym = 2;
3837: else if (A->spd == PETSC_BOOL3_TRUE) mumps->sym = 1;
3838: else mumps->sym = 2;
3839: }
3841: /* set solvertype */
3842: PetscCall(PetscFree(B->solvertype));
3843: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &B->solvertype));
3844: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
3845: if (size == 1) {
3846: /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
3847: B->canuseordering = PETSC_TRUE;
3848: }
3849: B->ops->destroy = MatDestroy_MUMPS;
3850: B->data = (void *)mumps;
3852: *F = B;
3853: mumps->id.job = JOB_NULL;
3854: mumps->ICNTL_pre = NULL;
3855: mumps->CNTL_pre = NULL;
3856: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
3857: PetscFunctionReturn(PETSC_SUCCESS);
3858: }
3860: /* MatGetFactor for Seq and MPI SBAIJ matrices */
3861: static PetscErrorCode MatGetFactor_sbaij_mumps(Mat A, PETSC_UNUSED MatFactorType ftype, Mat *F)
3862: {
3863: Mat B;
3864: Mat_MUMPS *mumps;
3865: PetscBool isSeqSBAIJ;
3866: PetscMPIInt size;
3868: PetscFunctionBegin;
3869: if (PetscDefined(USE_COMPLEX) && ftype == MAT_FACTOR_CHOLESKY && A->hermitian == PETSC_BOOL3_TRUE && A->symmetric != PETSC_BOOL3_TRUE) {
3870: PetscCall(PetscInfo(A, "Hermitian MAT_FACTOR_CHOLESKY is not supported. Use MAT_FACTOR_LU instead.\n"));
3871: *F = NULL;
3872: PetscFunctionReturn(PETSC_SUCCESS);
3873: }
3874: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
3875: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
3876: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &((PetscObject)B)->type_name));
3877: PetscCall(MatSetUp(B));
3879: PetscCall(PetscNew(&mumps));
3880: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQSBAIJ, &isSeqSBAIJ));
3881: if (isSeqSBAIJ) {
3882: mumps->ConvertToTriples = MatConvertToTriples_seqsbaij_seqsbaij;
3883: } else {
3884: mumps->ConvertToTriples = MatConvertToTriples_mpisbaij_mpisbaij;
3885: }
3887: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
3888: B->ops->view = MatView_MUMPS;
3889: B->ops->getinfo = MatGetInfo_MUMPS;
3891: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mumps));
3892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MUMPS));
3893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorCreateSchurComplement_C", MatFactorCreateSchurComplement_MUMPS));
3894: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetIcntl_C", MatMumpsSetIcntl_MUMPS));
3895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetIcntl_C", MatMumpsGetIcntl_MUMPS));
3896: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetCntl_C", MatMumpsSetCntl_MUMPS));
3897: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetCntl_C", MatMumpsGetCntl_MUMPS));
3898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfo_C", MatMumpsGetInfo_MUMPS));
3899: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfog_C", MatMumpsGetInfog_MUMPS));
3900: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfo_C", MatMumpsGetRinfo_MUMPS));
3901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfog_C", MatMumpsGetRinfog_MUMPS));
3902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetNullPivots_C", MatMumpsGetNullPivots_MUMPS));
3903: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverse_C", MatMumpsGetInverse_MUMPS));
3904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverseTranspose_C", MatMumpsGetInverseTranspose_MUMPS));
3905: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetBlk_C", MatMumpsSetBlk_MUMPS));
3907: B->factortype = MAT_FACTOR_CHOLESKY;
3908: if (PetscDefined(USE_COMPLEX)) mumps->sym = 2;
3909: else if (A->spd == PETSC_BOOL3_TRUE) mumps->sym = 1;
3910: else mumps->sym = 2;
3912: /* set solvertype */
3913: PetscCall(PetscFree(B->solvertype));
3914: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &B->solvertype));
3915: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
3916: if (size == 1) {
3917: /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
3918: B->canuseordering = PETSC_TRUE;
3919: }
3920: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_CHOLESKY]));
3921: B->ops->destroy = MatDestroy_MUMPS;
3922: B->data = (void *)mumps;
3924: *F = B;
3925: mumps->id.job = JOB_NULL;
3926: mumps->ICNTL_pre = NULL;
3927: mumps->CNTL_pre = NULL;
3928: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
3929: PetscFunctionReturn(PETSC_SUCCESS);
3930: }
3932: static PetscErrorCode MatGetFactor_baij_mumps(Mat A, MatFactorType ftype, Mat *F)
3933: {
3934: Mat B;
3935: Mat_MUMPS *mumps;
3936: PetscBool isSeqBAIJ;
3937: PetscMPIInt size;
3939: PetscFunctionBegin;
3940: /* Create the factorization matrix */
3941: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQBAIJ, &isSeqBAIJ));
3942: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
3943: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
3944: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &((PetscObject)B)->type_name));
3945: PetscCall(MatSetUp(B));
3947: PetscCall(PetscNew(&mumps));
3948: PetscCheck(ftype == MAT_FACTOR_LU, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot use PETSc BAIJ matrices with MUMPS Cholesky, use SBAIJ or AIJ matrix instead");
3949: B->ops->lufactorsymbolic = MatLUFactorSymbolic_BAIJMUMPS;
3950: B->factortype = MAT_FACTOR_LU;
3951: if (isSeqBAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqbaij_seqaij;
3952: else mumps->ConvertToTriples = MatConvertToTriples_mpibaij_mpiaij;
3953: mumps->sym = 0;
3954: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_LU]));
3956: B->ops->view = MatView_MUMPS;
3957: B->ops->getinfo = MatGetInfo_MUMPS;
3959: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mumps));
3960: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MUMPS));
3961: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorCreateSchurComplement_C", MatFactorCreateSchurComplement_MUMPS));
3962: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetIcntl_C", MatMumpsSetIcntl_MUMPS));
3963: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetIcntl_C", MatMumpsGetIcntl_MUMPS));
3964: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetCntl_C", MatMumpsSetCntl_MUMPS));
3965: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetCntl_C", MatMumpsGetCntl_MUMPS));
3966: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfo_C", MatMumpsGetInfo_MUMPS));
3967: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfog_C", MatMumpsGetInfog_MUMPS));
3968: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfo_C", MatMumpsGetRinfo_MUMPS));
3969: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfog_C", MatMumpsGetRinfog_MUMPS));
3970: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetNullPivots_C", MatMumpsGetNullPivots_MUMPS));
3971: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverse_C", MatMumpsGetInverse_MUMPS));
3972: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverseTranspose_C", MatMumpsGetInverseTranspose_MUMPS));
3973: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetBlk_C", MatMumpsSetBlk_MUMPS));
3975: /* set solvertype */
3976: PetscCall(PetscFree(B->solvertype));
3977: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &B->solvertype));
3978: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
3979: if (size == 1) {
3980: /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
3981: B->canuseordering = PETSC_TRUE;
3982: }
3983: B->ops->destroy = MatDestroy_MUMPS;
3984: B->data = (void *)mumps;
3986: *F = B;
3987: mumps->id.job = JOB_NULL;
3988: mumps->ICNTL_pre = NULL;
3989: mumps->CNTL_pre = NULL;
3990: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
3991: PetscFunctionReturn(PETSC_SUCCESS);
3992: }
3994: /* MatGetFactor for Seq and MPI SELL matrices */
3995: static PetscErrorCode MatGetFactor_sell_mumps(Mat A, MatFactorType ftype, Mat *F)
3996: {
3997: Mat B;
3998: Mat_MUMPS *mumps;
3999: PetscBool isSeqSELL;
4000: PetscMPIInt size;
4002: PetscFunctionBegin;
4003: /* Create the factorization matrix */
4004: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQSELL, &isSeqSELL));
4005: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
4006: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
4007: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &((PetscObject)B)->type_name));
4008: PetscCall(MatSetUp(B));
4010: PetscCall(PetscNew(&mumps));
4012: B->ops->view = MatView_MUMPS;
4013: B->ops->getinfo = MatGetInfo_MUMPS;
4015: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mumps));
4016: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MUMPS));
4017: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorCreateSchurComplement_C", MatFactorCreateSchurComplement_MUMPS));
4018: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetIcntl_C", MatMumpsSetIcntl_MUMPS));
4019: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetIcntl_C", MatMumpsGetIcntl_MUMPS));
4020: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetCntl_C", MatMumpsSetCntl_MUMPS));
4021: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetCntl_C", MatMumpsGetCntl_MUMPS));
4022: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfo_C", MatMumpsGetInfo_MUMPS));
4023: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfog_C", MatMumpsGetInfog_MUMPS));
4024: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfo_C", MatMumpsGetRinfo_MUMPS));
4025: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfog_C", MatMumpsGetRinfog_MUMPS));
4026: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetNullPivots_C", MatMumpsGetNullPivots_MUMPS));
4028: PetscCheck(ftype == MAT_FACTOR_LU, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "To be implemented");
4029: B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
4030: B->factortype = MAT_FACTOR_LU;
4031: PetscCheck(isSeqSELL, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "To be implemented");
4032: mumps->ConvertToTriples = MatConvertToTriples_seqsell_seqaij;
4033: mumps->sym = 0;
4034: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_LU]));
4036: /* set solvertype */
4037: PetscCall(PetscFree(B->solvertype));
4038: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &B->solvertype));
4039: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
4040: if (size == 1) {
4041: /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
4042: B->canuseordering = PETSC_TRUE;
4043: }
4044: B->ops->destroy = MatDestroy_MUMPS;
4045: B->data = (void *)mumps;
4047: *F = B;
4048: mumps->id.job = JOB_NULL;
4049: mumps->ICNTL_pre = NULL;
4050: mumps->CNTL_pre = NULL;
4051: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
4052: PetscFunctionReturn(PETSC_SUCCESS);
4053: }
4055: /* MatGetFactor for MATNEST matrices */
4056: static PetscErrorCode MatGetFactor_nest_mumps(Mat A, MatFactorType ftype, Mat *F)
4057: {
4058: Mat B, **mats;
4059: Mat_MUMPS *mumps;
4060: PetscInt nr, nc;
4061: PetscMPIInt size;
4062: PetscBool flg = PETSC_TRUE;
4064: PetscFunctionBegin;
4065: if (PetscDefined(USE_COMPLEX) && ftype == MAT_FACTOR_CHOLESKY && A->hermitian == PETSC_BOOL3_TRUE && A->symmetric != PETSC_BOOL3_TRUE) {
4066: PetscCall(PetscInfo(A, "Hermitian MAT_FACTOR_CHOLESKY is not supported. Use MAT_FACTOR_LU instead.\n"));
4067: *F = NULL;
4068: PetscFunctionReturn(PETSC_SUCCESS);
4069: }
4071: /* Return if some condition is not satisfied */
4072: *F = NULL;
4073: PetscCall(MatNestGetSubMats(A, &nr, &nc, &mats));
4074: if (ftype == MAT_FACTOR_CHOLESKY) {
4075: IS *rows, *cols;
4076: PetscInt *m, *M;
4078: PetscCheck(nr == nc, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MAT_FACTOR_CHOLESKY not supported for nest sizes %" PetscInt_FMT " != %" PetscInt_FMT ". Use MAT_FACTOR_LU.", nr, nc);
4079: PetscCall(PetscMalloc2(nr, &rows, nc, &cols));
4080: PetscCall(MatNestGetISs(A, rows, cols));
4081: for (PetscInt r = 0; flg && r < nr; r++) PetscCall(ISEqualUnsorted(rows[r], cols[r], &flg));
4082: if (!flg) {
4083: PetscCall(PetscFree2(rows, cols));
4084: PetscCall(PetscInfo(A, "MAT_FACTOR_CHOLESKY not supported for unequal row and column maps. Use MAT_FACTOR_LU.\n"));
4085: PetscFunctionReturn(PETSC_SUCCESS);
4086: }
4087: PetscCall(PetscMalloc2(nr, &m, nr, &M));
4088: for (PetscInt r = 0; r < nr; r++) PetscCall(ISGetMinMax(rows[r], &m[r], &M[r]));
4089: for (PetscInt r = 0; flg && r < nr; r++)
4090: for (PetscInt k = r + 1; flg && k < nr; k++)
4091: if ((m[k] <= m[r] && m[r] <= M[k]) || (m[k] <= M[r] && M[r] <= M[k])) flg = PETSC_FALSE;
4092: PetscCall(PetscFree2(m, M));
4093: PetscCall(PetscFree2(rows, cols));
4094: if (!flg) {
4095: PetscCall(PetscInfo(A, "MAT_FACTOR_CHOLESKY not supported for intersecting row maps. Use MAT_FACTOR_LU.\n"));
4096: PetscFunctionReturn(PETSC_SUCCESS);
4097: }
4098: }
4100: for (PetscInt r = 0; r < nr; r++) {
4101: for (PetscInt c = 0; c < nc; c++) {
4102: Mat sub = mats[r][c];
4103: PetscBool isSeqAIJ, isMPIAIJ, isSeqBAIJ, isMPIBAIJ, isSeqSBAIJ, isMPISBAIJ, isDiag, isDense;
4105: if (!sub || (ftype == MAT_FACTOR_CHOLESKY && c < r)) continue;
4106: PetscCall(MatGetTranspose_TransposeVirtual(&sub, NULL, NULL, NULL, NULL));
4107: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATSEQAIJ, &isSeqAIJ));
4108: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATMPIAIJ, &isMPIAIJ));
4109: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATSEQBAIJ, &isSeqBAIJ));
4110: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATMPIBAIJ, &isMPIBAIJ));
4111: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATSEQSBAIJ, &isSeqSBAIJ));
4112: PetscCall(PetscObjectBaseTypeCompare((PetscObject)sub, MATMPISBAIJ, &isMPISBAIJ));
4113: PetscCall(PetscObjectTypeCompare((PetscObject)sub, MATDIAGONAL, &isDiag));
4114: PetscCall(PetscObjectTypeCompareAny((PetscObject)sub, &isDense, MATSEQDENSE, MATMPIDENSE, NULL));
4115: if (ftype == MAT_FACTOR_CHOLESKY) {
4116: if (r == c) {
4117: if (!isSeqAIJ && !isMPIAIJ && !isSeqBAIJ && !isMPIBAIJ && !isSeqSBAIJ && !isMPISBAIJ && !isDiag && !isDense) {
4118: PetscCall(PetscInfo(sub, "MAT_FACTOR_CHOLESKY not supported for diagonal block of type %s.\n", ((PetscObject)sub)->type_name));
4119: flg = PETSC_FALSE;
4120: }
4121: } else if (!isSeqAIJ && !isMPIAIJ && !isSeqBAIJ && !isMPIBAIJ && !isDiag && !isDense) {
4122: PetscCall(PetscInfo(sub, "MAT_FACTOR_CHOLESKY not supported for off-diagonal block of type %s.\n", ((PetscObject)sub)->type_name));
4123: flg = PETSC_FALSE;
4124: }
4125: } else if (!isSeqAIJ && !isMPIAIJ && !isSeqBAIJ && !isMPIBAIJ && !isDiag && !isDense) {
4126: PetscCall(PetscInfo(sub, "MAT_FACTOR_LU not supported for block of type %s.\n", ((PetscObject)sub)->type_name));
4127: flg = PETSC_FALSE;
4128: }
4129: }
4130: }
4131: if (!flg) PetscFunctionReturn(PETSC_SUCCESS);
4133: /* Create the factorization matrix */
4134: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
4135: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
4136: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &((PetscObject)B)->type_name));
4137: PetscCall(MatSetUp(B));
4139: PetscCall(PetscNew(&mumps));
4141: B->ops->view = MatView_MUMPS;
4142: B->ops->getinfo = MatGetInfo_MUMPS;
4144: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mumps));
4145: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MUMPS));
4146: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorCreateSchurComplement_C", MatFactorCreateSchurComplement_MUMPS));
4147: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetIcntl_C", MatMumpsSetIcntl_MUMPS));
4148: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetIcntl_C", MatMumpsGetIcntl_MUMPS));
4149: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetCntl_C", MatMumpsSetCntl_MUMPS));
4150: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetCntl_C", MatMumpsGetCntl_MUMPS));
4151: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfo_C", MatMumpsGetInfo_MUMPS));
4152: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInfog_C", MatMumpsGetInfog_MUMPS));
4153: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfo_C", MatMumpsGetRinfo_MUMPS));
4154: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetRinfog_C", MatMumpsGetRinfog_MUMPS));
4155: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetNullPivots_C", MatMumpsGetNullPivots_MUMPS));
4156: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverse_C", MatMumpsGetInverse_MUMPS));
4157: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsGetInverseTranspose_C", MatMumpsGetInverseTranspose_MUMPS));
4158: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMumpsSetBlk_C", MatMumpsSetBlk_MUMPS));
4160: if (ftype == MAT_FACTOR_LU) {
4161: B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
4162: B->factortype = MAT_FACTOR_LU;
4163: mumps->sym = 0;
4164: } else {
4165: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
4166: B->factortype = MAT_FACTOR_CHOLESKY;
4167: if (PetscDefined(USE_COMPLEX)) mumps->sym = 2;
4168: else if (A->spd == PETSC_BOOL3_TRUE) mumps->sym = 1;
4169: else mumps->sym = 2;
4170: }
4171: mumps->ConvertToTriples = MatConvertToTriples_nest_xaij;
4172: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[ftype]));
4174: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
4175: if (size == 1) {
4176: /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
4177: B->canuseordering = PETSC_TRUE;
4178: }
4180: /* set solvertype */
4181: PetscCall(PetscFree(B->solvertype));
4182: PetscCall(PetscStrallocpy(MATSOLVERMUMPS, &B->solvertype));
4183: B->ops->destroy = MatDestroy_MUMPS;
4184: B->data = (void *)mumps;
4186: *F = B;
4187: mumps->id.job = JOB_NULL;
4188: mumps->ICNTL_pre = NULL;
4189: mumps->CNTL_pre = NULL;
4190: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
4191: PetscFunctionReturn(PETSC_SUCCESS);
4192: }
4194: PETSC_INTERN PetscErrorCode MatSolverTypeRegister_MUMPS(void)
4195: {
4196: PetscFunctionBegin;
4197: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPIAIJ, MAT_FACTOR_LU, MatGetFactor_aij_mumps));
4198: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPIAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mumps));
4199: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPIBAIJ, MAT_FACTOR_LU, MatGetFactor_baij_mumps));
4200: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPIBAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_baij_mumps));
4201: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPISBAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_sbaij_mumps));
4202: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQAIJ, MAT_FACTOR_LU, MatGetFactor_aij_mumps));
4203: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mumps));
4204: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQBAIJ, MAT_FACTOR_LU, MatGetFactor_baij_mumps));
4205: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQBAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_baij_mumps));
4206: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQSBAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_sbaij_mumps));
4207: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQSELL, MAT_FACTOR_LU, MatGetFactor_sell_mumps));
4208: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATDIAGONAL, MAT_FACTOR_LU, MatGetFactor_aij_mumps));
4209: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATDIAGONAL, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mumps));
4210: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQDENSE, MAT_FACTOR_LU, MatGetFactor_aij_mumps));
4211: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATSEQDENSE, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mumps));
4212: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPIDENSE, MAT_FACTOR_LU, MatGetFactor_aij_mumps));
4213: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATMPIDENSE, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mumps));
4214: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATNEST, MAT_FACTOR_LU, MatGetFactor_nest_mumps));
4215: PetscCall(MatSolverTypeRegister(MATSOLVERMUMPS, MATNEST, MAT_FACTOR_CHOLESKY, MatGetFactor_nest_mumps));
4216: PetscFunctionReturn(PETSC_SUCCESS);
4217: }