Actual source code: superlu_dist.c
petsc-3.14.6 2021-03-30
1: /*
2: Provides an interface to the SuperLU_DIST sparse solver
3: */
5: #include <../src/mat/impls/aij/seq/aij.h>
6: #include <../src/mat/impls/aij/mpi/mpiaij.h>
7: #include <petscpkg_version.h>
9: EXTERN_C_BEGIN
10: #if defined(PETSC_USE_COMPLEX)
11: #include <superlu_zdefs.h>
12: #if PETSC_PKG_SUPERLU_DIST_VERSION_GE(6,3,0)
13: #define LUstructInit zLUstructInit
14: #define ScalePermstructInit zScalePermstructInit
15: #define ScalePermstructFree zScalePermstructFree
16: #define LUstructFree zLUstructFree
17: #define Destroy_LU zDestroy_LU
18: #define ScalePermstruct_t zScalePermstruct_t
19: #define LUstruct_t zLUstruct_t
20: #define SOLVEstruct_t zSOLVEstruct_t
21: #endif
22: #else
23: #include <superlu_ddefs.h>
24: #if PETSC_PKG_SUPERLU_DIST_VERSION_GE(6,3,0)
25: #define LUstructInit dLUstructInit
26: #define ScalePermstructInit dScalePermstructInit
27: #define ScalePermstructFree dScalePermstructFree
28: #define LUstructFree dLUstructFree
29: #define Destroy_LU dDestroy_LU
30: #define ScalePermstruct_t dScalePermstruct_t
31: #define LUstruct_t dLUstruct_t
32: #define SOLVEstruct_t dSOLVEstruct_t
33: #endif
34: #endif
35: EXTERN_C_END
37: typedef struct {
38: int_t nprow,npcol,*row,*col;
39: gridinfo_t grid;
40: superlu_dist_options_t options;
41: SuperMatrix A_sup;
42: ScalePermstruct_t ScalePermstruct;
43: LUstruct_t LUstruct;
44: int StatPrint;
45: SOLVEstruct_t SOLVEstruct;
46: fact_t FactPattern;
47: MPI_Comm comm_superlu;
48: #if defined(PETSC_USE_COMPLEX)
49: doublecomplex *val;
50: #else
51: double *val;
52: #endif
53: PetscBool matsolve_iscalled,matmatsolve_iscalled;
54: PetscBool CleanUpSuperLU_Dist; /* Flag to clean up (non-global) SuperLU objects during Destroy */
55: } Mat_SuperLU_DIST;
58: PetscErrorCode MatSuperluDistGetDiagU_SuperLU_DIST(Mat F,PetscScalar *diagU)
59: {
60: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)F->data;
63: #if defined(PETSC_USE_COMPLEX)
64: PetscStackCall("SuperLU_DIST:pzGetDiagU",pzGetDiagU(F->rmap->N,&lu->LUstruct,&lu->grid,(doublecomplex*)diagU));
65: #else
66: PetscStackCall("SuperLU_DIST:pdGetDiagU",pdGetDiagU(F->rmap->N,&lu->LUstruct,&lu->grid,diagU));
67: #endif
68: return(0);
69: }
71: PetscErrorCode MatSuperluDistGetDiagU(Mat F,PetscScalar *diagU)
72: {
77: PetscTryMethod(F,"MatSuperluDistGetDiagU_C",(Mat,PetscScalar*),(F,diagU));
78: return(0);
79: }
81: /* This allows reusing the Superlu_DIST communicator and grid when only a single SuperLU_DIST matrix is used at a time */
82: typedef struct {
83: MPI_Comm comm;
84: PetscBool busy;
85: gridinfo_t grid;
86: } PetscSuperLU_DIST;
87: static PetscMPIInt Petsc_Superlu_dist_keyval = MPI_KEYVAL_INVALID;
89: PETSC_EXTERN PetscMPIInt MPIAPI Petsc_Superlu_dist_keyval_Delete_Fn(MPI_Comm comm,PetscMPIInt keyval,void *attr_val,void *extra_state)
90: {
91: PetscErrorCode ierr;
92: PetscSuperLU_DIST *context = (PetscSuperLU_DIST *) attr_val;
95: if (keyval != Petsc_Superlu_dist_keyval) SETERRMPI(PETSC_COMM_SELF,PETSC_ERR_ARG_CORRUPT,"Unexpected keyval");
96: PetscInfo(NULL,"Removing Petsc_Superlu_dist_keyval attribute from communicator that is being freed\n");
97: PetscStackCall("SuperLU_DIST:superlu_gridexit",superlu_gridexit(&context->grid));
98: MPI_Comm_free(&context->comm);
99: PetscFree(context);
100: PetscFunctionReturn(MPI_SUCCESS);
101: }
103: /*
104: Performs MPI_Comm_free_keyval() on Petsc_Superlu_dist_keyval but keeps the global variable for
105: users who do not destroy all PETSc objects before PetscFinalize().
107: The value Petsc_Superlu_dist_keyval is retained so that Petsc_Superlu_dist_keyval_Delete_Fn()
108: can still check that the keyval associated with the MPI communicator is correct when the MPI
109: communicator is destroyed.
111: This is called in PetscFinalize()
112: */
113: static PetscErrorCode Petsc_Superlu_dist_keyval_free(void)
114: {
116: PetscMPIInt Petsc_Superlu_dist_keyval_temp = Petsc_Superlu_dist_keyval;
119: PetscInfo(NULL,"Freeing Petsc_Superlu_dist_keyval\n");
120: MPI_Comm_free_keyval(&Petsc_Superlu_dist_keyval_temp);
121: return(0);
122: }
124: static PetscErrorCode MatDestroy_SuperLU_DIST(Mat A)
125: {
126: PetscErrorCode ierr;
127: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->data;
130: if (lu->CleanUpSuperLU_Dist) {
131: /* Deallocate SuperLU_DIST storage */
132: PetscStackCall("SuperLU_DIST:Destroy_CompRowLoc_Matrix_dist",Destroy_CompRowLoc_Matrix_dist(&lu->A_sup));
133: if (lu->options.SolveInitialized) {
134: #if defined(PETSC_USE_COMPLEX)
135: PetscStackCall("SuperLU_DIST:zSolveFinalize",zSolveFinalize(&lu->options, &lu->SOLVEstruct));
136: #else
137: PetscStackCall("SuperLU_DIST:dSolveFinalize",dSolveFinalize(&lu->options, &lu->SOLVEstruct));
138: #endif
139: }
140: PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(A->cmap->N, &lu->grid, &lu->LUstruct));
141: PetscStackCall("SuperLU_DIST:ScalePermstructFree",ScalePermstructFree(&lu->ScalePermstruct));
142: PetscStackCall("SuperLU_DIST:LUstructFree",LUstructFree(&lu->LUstruct));
144: /* Release the SuperLU_DIST process grid. Only if the matrix has its own copy, this is it is not in the communicator context */
145: if (lu->comm_superlu) {
146: PetscStackCall("SuperLU_DIST:superlu_gridexit",superlu_gridexit(&lu->grid));
147: MPI_Comm_free(&(lu->comm_superlu));
148: } else {
149: PetscSuperLU_DIST *context;
150: MPI_Comm comm;
151: PetscMPIInt flg;
153: PetscObjectGetComm((PetscObject)A,&comm);
154: MPI_Comm_get_attr(comm,Petsc_Superlu_dist_keyval,&context,&flg);
155: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Communicator does not have expected Petsc_Superlu_dist_keyval attribute");
156: context->busy = PETSC_FALSE;
157: }
158: }
159: PetscFree(A->data);
160: /* clear composed functions */
161: PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverType_C",NULL);
162: PetscObjectComposeFunction((PetscObject)A,"MatSuperluDistGetDiagU_C",NULL);
164: return(0);
165: }
167: static PetscErrorCode MatSolve_SuperLU_DIST(Mat A,Vec b_mpi,Vec x)
168: {
169: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->data;
170: PetscErrorCode ierr;
171: PetscInt m=A->rmap->n;
172: SuperLUStat_t stat;
173: double berr[1];
174: PetscScalar *bptr=NULL;
175: int info; /* SuperLU_Dist info code is ALWAYS an int, even with long long indices */
176: static PetscBool cite = PETSC_FALSE;
179: if (lu->options.Fact != FACTORED) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"SuperLU_DIST options.Fact must equal FACTORED");
180: PetscCitationsRegister("@article{lidemmel03,\n author = {Xiaoye S. Li and James W. Demmel},\n title = {{SuperLU_DIST}: A Scalable Distributed-Memory Sparse Direct\n Solver for Unsymmetric Linear Systems},\n journal = {ACM Trans. Mathematical Software},\n volume = {29},\n number = {2},\n pages = {110-140},\n year = 2003\n}\n",&cite);
182: if (lu->options.SolveInitialized && !lu->matsolve_iscalled) {
183: /* see comments in MatMatSolve() */
184: #if defined(PETSC_USE_COMPLEX)
185: PetscStackCall("SuperLU_DIST:zSolveFinalize",zSolveFinalize(&lu->options, &lu->SOLVEstruct));
186: #else
187: PetscStackCall("SuperLU_DIST:dSolveFinalize",dSolveFinalize(&lu->options, &lu->SOLVEstruct));
188: #endif
189: lu->options.SolveInitialized = NO;
190: }
191: VecCopy(b_mpi,x);
192: VecGetArray(x,&bptr);
194: PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat)); /* Initialize the statistics variables. */
195: #if defined(PETSC_USE_COMPLEX)
196: PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,(doublecomplex*)bptr,m,1,&lu->grid,&lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
197: #else
198: PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,bptr,m,1,&lu->grid,&lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
199: #endif
200: if (info) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pdgssvx fails, info: %d\n",info);
202: if (lu->options.PrintStat) PetscStackCall("SuperLU_DIST:PStatPrint",PStatPrint(&lu->options, &stat, &lu->grid)); /* Print the statistics. */
203: PetscStackCall("SuperLU_DIST:PStatFree",PStatFree(&stat));
205: VecRestoreArray(x,&bptr);
206: lu->matsolve_iscalled = PETSC_TRUE;
207: lu->matmatsolve_iscalled = PETSC_FALSE;
208: return(0);
209: }
211: static PetscErrorCode MatMatSolve_SuperLU_DIST(Mat A,Mat B_mpi,Mat X)
212: {
213: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->data;
214: PetscErrorCode ierr;
215: PetscInt m=A->rmap->n,nrhs;
216: SuperLUStat_t stat;
217: double berr[1];
218: PetscScalar *bptr;
219: int info; /* SuperLU_Dist info code is ALWAYS an int, even with long long indices */
220: PetscBool flg;
223: if (lu->options.Fact != FACTORED) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"SuperLU_DIST options.Fact must equal FACTORED");
224: PetscObjectTypeCompareAny((PetscObject)B_mpi,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
225: if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix");
226: if (X != B_mpi) {
227: PetscObjectTypeCompareAny((PetscObject)X,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
228: if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix");
229: }
231: if (lu->options.SolveInitialized && !lu->matmatsolve_iscalled) {
232: /* communication pattern of SOLVEstruct is unlikely created for matmatsolve,
233: thus destroy it and create a new SOLVEstruct.
234: Otherwise it may result in memory corruption or incorrect solution
235: See src/mat/tests/ex125.c */
236: #if defined(PETSC_USE_COMPLEX)
237: PetscStackCall("SuperLU_DIST:zSolveFinalize",zSolveFinalize(&lu->options, &lu->SOLVEstruct));
238: #else
239: PetscStackCall("SuperLU_DIST:dSolveFinalize",dSolveFinalize(&lu->options, &lu->SOLVEstruct));
240: #endif
241: lu->options.SolveInitialized = NO;
242: }
243: if (X != B_mpi) {
244: MatCopy(B_mpi,X,SAME_NONZERO_PATTERN);
245: }
247: MatGetSize(B_mpi,NULL,&nrhs);
249: PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat)); /* Initialize the statistics variables. */
250: MatDenseGetArray(X,&bptr);
252: #if defined(PETSC_USE_COMPLEX)
253: PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,(doublecomplex*)bptr,m,nrhs,&lu->grid, &lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
254: #else
255: PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,bptr,m,nrhs,&lu->grid,&lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
256: #endif
258: if (info) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pdgssvx fails, info: %d\n",info);
259: MatDenseRestoreArray(X,&bptr);
261: if (lu->options.PrintStat) PetscStackCall("SuperLU_DIST:PStatPrint",PStatPrint(&lu->options, &stat, &lu->grid)); /* Print the statistics. */
262: PetscStackCall("SuperLU_DIST:PStatFree",PStatFree(&stat));
263: lu->matsolve_iscalled = PETSC_FALSE;
264: lu->matmatsolve_iscalled = PETSC_TRUE;
265: return(0);
266: }
268: /*
269: input:
270: F: numeric Cholesky factor
271: output:
272: nneg: total number of negative pivots
273: nzero: total number of zero pivots
274: npos: (global dimension of F) - nneg - nzero
275: */
276: static PetscErrorCode MatGetInertia_SuperLU_DIST(Mat F,PetscInt *nneg,PetscInt *nzero,PetscInt *npos)
277: {
278: PetscErrorCode ierr;
279: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)F->data;
280: PetscScalar *diagU=NULL;
281: PetscInt M,i,neg=0,zero=0,pos=0;
282: PetscReal r;
285: if (!F->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Matrix factor F is not assembled");
286: if (lu->options.RowPerm != NOROWPERM) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Must set NOROWPERM");
287: MatGetSize(F,&M,NULL);
288: PetscMalloc1(M,&diagU);
289: MatSuperluDistGetDiagU(F,diagU);
290: for (i=0; i<M; i++) {
291: #if defined(PETSC_USE_COMPLEX)
292: r = PetscImaginaryPart(diagU[i])/10.0;
293: if (r< -PETSC_MACHINE_EPSILON || r>PETSC_MACHINE_EPSILON) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"diagU[%d]=%g + i %g is non-real",i,PetscRealPart(diagU[i]),r*10.0);
294: r = PetscRealPart(diagU[i]);
295: #else
296: r = diagU[i];
297: #endif
298: if (r > 0) {
299: pos++;
300: } else if (r < 0) {
301: neg++;
302: } else zero++;
303: }
305: PetscFree(diagU);
306: if (nneg) *nneg = neg;
307: if (nzero) *nzero = zero;
308: if (npos) *npos = pos;
309: return(0);
310: }
312: static PetscErrorCode MatLUFactorNumeric_SuperLU_DIST(Mat F,Mat A,const MatFactorInfo *info)
313: {
314: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)F->data;
315: Mat Aloc;
316: const PetscScalar *av;
317: const PetscInt *ai=NULL,*aj=NULL;
318: PetscInt nz,dummy;
319: int sinfo; /* SuperLU_Dist info flag is always an int even with long long indices */
320: SuperLUStat_t stat;
321: double *berr=0;
322: PetscBool ismpiaij,isseqaij,flg;
323: PetscErrorCode ierr;
326: PetscObjectBaseTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);
327: PetscObjectBaseTypeCompare((PetscObject)A,MATMPIAIJ,&ismpiaij);
328: if (ismpiaij) {
329: MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&Aloc);
330: } else if (isseqaij) {
331: PetscObjectReference((PetscObject)A);
332: Aloc = A;
333: } else SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Not for type %s",((PetscObject)A)->type_name);
335: MatGetRowIJ(Aloc,0,PETSC_FALSE,PETSC_FALSE,&dummy,&ai,&aj,&flg);
336: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"GetRowIJ failed");
337: MatSeqAIJGetArrayRead(Aloc,&av);
338: nz = ai[Aloc->rmap->n];
340: /* Allocations for A_sup */
341: if (lu->options.Fact == DOFACT) { /* first numeric factorization */
342: #if defined(PETSC_USE_COMPLEX)
343: PetscStackCall("SuperLU_DIST:zallocateA_dist",zallocateA_dist(Aloc->rmap->n, nz, &lu->val, &lu->col, &lu->row));
344: #else
345: PetscStackCall("SuperLU_DIST:dallocateA_dist",dallocateA_dist(Aloc->rmap->n, nz, &lu->val, &lu->col, &lu->row));
346: #endif
347: } else { /* successive numeric factorization, sparsity pattern and perm_c are reused. */
348: if (lu->FactPattern == SamePattern_SameRowPerm) {
349: lu->options.Fact = SamePattern_SameRowPerm; /* matrix has similar numerical values */
350: } else if (lu->FactPattern == SamePattern) {
351: PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(A->rmap->N, &lu->grid, &lu->LUstruct)); /* Deallocate L and U matrices. */
352: lu->options.Fact = SamePattern;
353: } else if (lu->FactPattern == DOFACT) {
354: PetscStackCall("SuperLU_DIST:Destroy_CompRowLoc_Matrix_dist",Destroy_CompRowLoc_Matrix_dist(&lu->A_sup));
355: PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(A->rmap->N, &lu->grid, &lu->LUstruct));
356: lu->options.Fact = DOFACT;
358: #if defined(PETSC_USE_COMPLEX)
359: PetscStackCall("SuperLU_DIST:zallocateA_dist",zallocateA_dist(Aloc->rmap->n, nz, &lu->val, &lu->col, &lu->row));
360: #else
361: PetscStackCall("SuperLU_DIST:dallocateA_dist",dallocateA_dist(Aloc->rmap->n, nz, &lu->val, &lu->col, &lu->row));
362: #endif
363: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"options.Fact must be one of SamePattern SamePattern_SameRowPerm DOFACT");
364: }
366: /* Copy AIJ matrix to superlu_dist matrix */
367: PetscArraycpy(lu->row,ai,Aloc->rmap->n+1);
368: PetscArraycpy(lu->col,aj,nz);
369: PetscArraycpy(lu->val,av,nz);
370: MatRestoreRowIJ(Aloc,0,PETSC_FALSE,PETSC_FALSE,&dummy,&ai,&aj,&flg);
371: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"RestoreRowIJ failed");
372: MatSeqAIJRestoreArrayRead(Aloc,&av);
373: MatDestroy(&Aloc);
375: /* Create and setup A_sup */
376: if (lu->options.Fact == DOFACT) {
377: #if defined(PETSC_USE_COMPLEX)
378: PetscStackCall("SuperLU_DIST:zCreate_CompRowLoc_Matrix_dist",zCreate_CompRowLoc_Matrix_dist(&lu->A_sup, A->rmap->N, A->cmap->N, nz, A->rmap->n, A->rmap->rstart, lu->val, lu->col, lu->row, SLU_NR_loc, SLU_Z, SLU_GE));
379: #else
380: PetscStackCall("SuperLU_DIST:dCreate_CompRowLoc_Matrix_dist",dCreate_CompRowLoc_Matrix_dist(&lu->A_sup, A->rmap->N, A->cmap->N, nz, A->rmap->n, A->rmap->rstart, lu->val, lu->col, lu->row, SLU_NR_loc, SLU_D, SLU_GE));
381: #endif
382: }
384: /* Factor the matrix. */
385: PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat)); /* Initialize the statistics variables. */
386: #if defined(PETSC_USE_COMPLEX)
387: PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, A->rmap->n, 0, &lu->grid, &lu->LUstruct, &lu->SOLVEstruct, berr, &stat, &sinfo));
388: #else
389: PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, A->rmap->n, 0, &lu->grid, &lu->LUstruct, &lu->SOLVEstruct, berr, &stat, &sinfo));
390: #endif
392: if (sinfo > 0) {
393: if (A->erroriffailure) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot in row %D",sinfo);
394: else {
395: if (sinfo <= lu->A_sup.ncol) {
396: F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
397: PetscInfo1(F,"U(i,i) is exactly zero, i= %D\n",sinfo);
398: } else if (sinfo > lu->A_sup.ncol) {
399: /*
400: number of bytes allocated when memory allocation
401: failure occurred, plus A->ncol.
402: */
403: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
404: PetscInfo1(F,"Number of bytes allocated when memory allocation fails %D\n",sinfo);
405: }
406: }
407: } else if (sinfo < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB, "info = %D, argument in p*gssvx() had an illegal value", sinfo);
409: if (lu->options.PrintStat) {
410: PetscStackCall("SuperLU_DIST:PStatPrint",PStatPrint(&lu->options, &stat, &lu->grid)); /* Print the statistics. */
411: }
412: PetscStackCall("SuperLU_DIST:PStatFree",PStatFree(&stat));
413: F->assembled = PETSC_TRUE;
414: F->preallocated = PETSC_TRUE;
415: lu->options.Fact = FACTORED; /* The factored form of A is supplied. Local option used by this func. only */
416: return(0);
417: }
419: /* Note the Petsc r and c permutations are ignored */
420: static PetscErrorCode MatLUFactorSymbolic_SuperLU_DIST(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
421: {
422: Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)F->data;
423: PetscInt M = A->rmap->N,N=A->cmap->N;
426: /* Initialize ScalePermstruct and LUstruct. */
427: PetscStackCall("SuperLU_DIST:ScalePermstructInit",ScalePermstructInit(M, N, &lu->ScalePermstruct));
428: PetscStackCall("SuperLU_DIST:LUstructInit",LUstructInit(N, &lu->LUstruct));
429: F->ops->lufactornumeric = MatLUFactorNumeric_SuperLU_DIST;
430: F->ops->solve = MatSolve_SuperLU_DIST;
431: F->ops->matsolve = MatMatSolve_SuperLU_DIST;
432: F->ops->getinertia = NULL;
434: if (A->symmetric || A->hermitian) F->ops->getinertia = MatGetInertia_SuperLU_DIST;
435: lu->CleanUpSuperLU_Dist = PETSC_TRUE;
436: return(0);
437: }
439: static PetscErrorCode MatCholeskyFactorSymbolic_SuperLU_DIST(Mat F,Mat A,IS r,const MatFactorInfo *info)
440: {
444: MatLUFactorSymbolic_SuperLU_DIST(F,A,r,r,info);
445: F->ops->choleskyfactornumeric = MatLUFactorNumeric_SuperLU_DIST;
446: return(0);
447: }
449: static PetscErrorCode MatFactorGetSolverType_aij_superlu_dist(Mat A,MatSolverType *type)
450: {
452: *type = MATSOLVERSUPERLU_DIST;
453: return(0);
454: }
456: static PetscErrorCode MatView_Info_SuperLU_DIST(Mat A,PetscViewer viewer)
457: {
458: Mat_SuperLU_DIST *lu=(Mat_SuperLU_DIST*)A->data;
459: superlu_dist_options_t options;
460: PetscErrorCode ierr;
463: /* check if matrix is superlu_dist type */
464: if (A->ops->solve != MatSolve_SuperLU_DIST) return(0);
466: options = lu->options;
467: PetscViewerASCIIPrintf(viewer,"SuperLU_DIST run parameters:\n");
468: PetscViewerASCIIPrintf(viewer," Process grid nprow %D x npcol %D \n",lu->nprow,lu->npcol);
469: PetscViewerASCIIPrintf(viewer," Equilibrate matrix %s \n",PetscBools[options.Equil != NO]);
470: PetscViewerASCIIPrintf(viewer," Replace tiny pivots %s \n",PetscBools[options.ReplaceTinyPivot != NO]);
471: PetscViewerASCIIPrintf(viewer," Use iterative refinement %s \n",PetscBools[options.IterRefine == SLU_DOUBLE]);
472: PetscViewerASCIIPrintf(viewer," Processors in row %d col partition %d \n",lu->nprow,lu->npcol);
474: switch (options.RowPerm) {
475: case NOROWPERM:
476: PetscViewerASCIIPrintf(viewer," Row permutation NOROWPERM\n");
477: break;
478: case LargeDiag_MC64:
479: PetscViewerASCIIPrintf(viewer," Row permutation LargeDiag_MC64\n");
480: break;
481: case LargeDiag_AWPM:
482: PetscViewerASCIIPrintf(viewer," Row permutation LargeDiag_AWPM\n");
483: break;
484: case MY_PERMR:
485: PetscViewerASCIIPrintf(viewer," Row permutation MY_PERMR\n");
486: break;
487: default:
488: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown column permutation");
489: }
491: switch (options.ColPerm) {
492: case NATURAL:
493: PetscViewerASCIIPrintf(viewer," Column permutation NATURAL\n");
494: break;
495: case MMD_AT_PLUS_A:
496: PetscViewerASCIIPrintf(viewer," Column permutation MMD_AT_PLUS_A\n");
497: break;
498: case MMD_ATA:
499: PetscViewerASCIIPrintf(viewer," Column permutation MMD_ATA\n");
500: break;
501: /* Even though this is called METIS, the SuperLU_DIST code sets this by default if PARMETIS is defined, not METIS */
502: case METIS_AT_PLUS_A:
503: PetscViewerASCIIPrintf(viewer," Column permutation METIS_AT_PLUS_A\n");
504: break;
505: case PARMETIS:
506: PetscViewerASCIIPrintf(viewer," Column permutation PARMETIS\n");
507: break;
508: default:
509: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown column permutation");
510: }
512: PetscViewerASCIIPrintf(viewer," Parallel symbolic factorization %s \n",PetscBools[options.ParSymbFact != NO]);
514: if (lu->FactPattern == SamePattern) {
515: PetscViewerASCIIPrintf(viewer," Repeated factorization SamePattern\n");
516: } else if (lu->FactPattern == SamePattern_SameRowPerm) {
517: PetscViewerASCIIPrintf(viewer," Repeated factorization SamePattern_SameRowPerm\n");
518: } else if (lu->FactPattern == DOFACT) {
519: PetscViewerASCIIPrintf(viewer," Repeated factorization DOFACT\n");
520: } else {
521: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown factorization pattern");
522: }
523: return(0);
524: }
526: static PetscErrorCode MatView_SuperLU_DIST(Mat A,PetscViewer viewer)
527: {
528: PetscErrorCode ierr;
529: PetscBool iascii;
530: PetscViewerFormat format;
533: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
534: if (iascii) {
535: PetscViewerGetFormat(viewer,&format);
536: if (format == PETSC_VIEWER_ASCII_INFO) {
537: MatView_Info_SuperLU_DIST(A,viewer);
538: }
539: }
540: return(0);
541: }
543: static PetscErrorCode MatGetFactor_aij_superlu_dist(Mat A,MatFactorType ftype,Mat *F)
544: {
545: Mat B;
546: Mat_SuperLU_DIST *lu;
547: PetscErrorCode ierr;
548: PetscInt M=A->rmap->N,N=A->cmap->N,indx;
549: PetscMPIInt size;
550: superlu_dist_options_t options;
551: PetscBool flg;
552: const char *colperm[] = {"NATURAL","MMD_AT_PLUS_A","MMD_ATA","METIS_AT_PLUS_A","PARMETIS"};
553: const char *rowperm[] = {"NOROWPERM","LargeDiag_MC64","LargeDiag_AWPM","MY_PERMR"};
554: const char *factPattern[] = {"SamePattern","SamePattern_SameRowPerm","DOFACT"};
555: PetscBool set;
558: /* Create the factorization matrix */
559: MatCreate(PetscObjectComm((PetscObject)A),&B);
560: MatSetSizes(B,A->rmap->n,A->cmap->n,M,N);
561: PetscStrallocpy("superlu_dist",&((PetscObject)B)->type_name);
562: MatSetUp(B);
563: B->ops->getinfo = MatGetInfo_External;
564: B->ops->view = MatView_SuperLU_DIST;
565: B->ops->destroy = MatDestroy_SuperLU_DIST;
567: /* Set the default input options:
568: options.Fact = DOFACT;
569: options.Equil = YES;
570: options.ParSymbFact = NO;
571: options.ColPerm = METIS_AT_PLUS_A;
572: options.RowPerm = LargeDiag_MC64;
573: options.ReplaceTinyPivot = YES;
574: options.IterRefine = DOUBLE;
575: options.Trans = NOTRANS;
576: options.SolveInitialized = NO; -hold the communication pattern used MatSolve() and MatMatSolve()
577: options.RefineInitialized = NO;
578: options.PrintStat = YES;
579: options.SymPattern = NO;
580: */
581: set_default_options_dist(&options);
583: if (ftype == MAT_FACTOR_LU) {
584: B->factortype = MAT_FACTOR_LU;
585: B->ops->lufactorsymbolic = MatLUFactorSymbolic_SuperLU_DIST;
586: } else {
587: B->factortype = MAT_FACTOR_CHOLESKY;
588: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SuperLU_DIST;
589: options.SymPattern = YES;
590: }
592: /* set solvertype */
593: PetscFree(B->solvertype);
594: PetscStrallocpy(MATSOLVERSUPERLU_DIST,&B->solvertype);
596: PetscNewLog(B,&lu);
597: B->data = lu;
598: MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
600: {
601: PetscMPIInt flg;
602: MPI_Comm comm;
603: PetscSuperLU_DIST *context = NULL;
605: PetscObjectGetComm((PetscObject)A,&comm);
606: if (Petsc_Superlu_dist_keyval == MPI_KEYVAL_INVALID) {
607: MPI_Comm_create_keyval(MPI_COMM_NULL_COPY_FN,Petsc_Superlu_dist_keyval_Delete_Fn,&Petsc_Superlu_dist_keyval,(void*)0);
608: PetscRegisterFinalize(Petsc_Superlu_dist_keyval_free);
609: }
610: MPI_Comm_get_attr(comm,Petsc_Superlu_dist_keyval,&context,&flg);
611: if (!flg || context->busy) {
612: if (!flg) {
613: PetscNew(&context);
614: context->busy = PETSC_TRUE;
615: MPI_Comm_dup(comm,&context->comm);
616: MPI_Comm_set_attr(comm,Petsc_Superlu_dist_keyval,context);
617: } else {
618: MPI_Comm_dup(comm,&lu->comm_superlu);
619: }
621: /* Default num of process columns and rows */
622: lu->nprow = (int_t) (0.5 + PetscSqrtReal((PetscReal)size));
623: if (!lu->nprow) lu->nprow = 1;
624: while (lu->nprow > 0) {
625: lu->npcol = (int_t) (size/lu->nprow);
626: if (size == lu->nprow * lu->npcol) break;
627: lu->nprow--;
628: }
629: PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"SuperLU_Dist Options","Mat");
630: PetscOptionsInt("-mat_superlu_dist_r","Number rows in processor partition","None",lu->nprow,(PetscInt*)&lu->nprow,NULL);
631: PetscOptionsInt("-mat_superlu_dist_c","Number columns in processor partition","None",lu->npcol,(PetscInt*)&lu->npcol,NULL);
632: PetscOptionsEnd();
633: if (size != lu->nprow * lu->npcol) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number of processes %d must equal to nprow %d * npcol %d",size,lu->nprow,lu->npcol);
634: PetscStackCall("SuperLU_DIST:superlu_gridinit",superlu_gridinit(context ? context->comm : lu->comm_superlu, lu->nprow, lu->npcol, &lu->grid));
635: if (context) context->grid = lu->grid;
636: PetscInfo(NULL,"Duplicating a communicator for SuperLU_DIST and calling superlu_gridinit()\n");
637: if (!flg) {
638: PetscInfo(NULL,"Storing communicator and SuperLU_DIST grid in communicator attribute\n");
639: } else {
640: PetscInfo(NULL,"Communicator attribute already in use so not saving communicator and SuperLU_DIST grid in communicator attribute \n");
641: }
642: } else {
643: PetscInfo(NULL,"Reusing communicator and superlu_gridinit() for SuperLU_DIST from communicator attribute.");
644: context->busy = PETSC_TRUE;
645: lu->grid = context->grid;
646: }
647: }
649: PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"SuperLU_Dist Options","Mat");
650: PetscOptionsBool("-mat_superlu_dist_equil","Equilibrate matrix","None",options.Equil ? PETSC_TRUE : PETSC_FALSE,&flg,&set);
651: if (set && !flg) options.Equil = NO;
653: PetscOptionsEList("-mat_superlu_dist_rowperm","Row permutation","None",rowperm,4,rowperm[1],&indx,&flg);
654: if (flg) {
655: switch (indx) {
656: case 0:
657: options.RowPerm = NOROWPERM;
658: break;
659: case 1:
660: options.RowPerm = LargeDiag_MC64;
661: break;
662: case 2:
663: options.RowPerm = LargeDiag_AWPM;
664: break;
665: case 3:
666: options.RowPerm = MY_PERMR;
667: break;
668: default:
669: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown row permutation");
670: }
671: }
673: PetscOptionsEList("-mat_superlu_dist_colperm","Column permutation","None",colperm,5,colperm[3],&indx,&flg);
674: if (flg) {
675: switch (indx) {
676: case 0:
677: options.ColPerm = NATURAL;
678: break;
679: case 1:
680: options.ColPerm = MMD_AT_PLUS_A;
681: break;
682: case 2:
683: options.ColPerm = MMD_ATA;
684: break;
685: case 3:
686: options.ColPerm = METIS_AT_PLUS_A;
687: break;
688: case 4:
689: options.ColPerm = PARMETIS; /* only works for np>1 */
690: break;
691: default:
692: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown column permutation");
693: }
694: }
696: options.ReplaceTinyPivot = NO;
697: PetscOptionsBool("-mat_superlu_dist_replacetinypivot","Replace tiny pivots","None",options.ReplaceTinyPivot ? PETSC_TRUE : PETSC_FALSE,&flg,&set);
698: if (set && flg) options.ReplaceTinyPivot = YES;
700: options.ParSymbFact = NO;
701: PetscOptionsBool("-mat_superlu_dist_parsymbfact","Parallel symbolic factorization","None",PETSC_FALSE,&flg,&set);
702: if (set && flg && size>1) {
703: #if defined(PETSC_HAVE_PARMETIS)
704: options.ParSymbFact = YES;
705: options.ColPerm = PARMETIS; /* in v2.2, PARMETIS is forced for ParSymbFact regardless of user ordering setting */
706: #else
707: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"parsymbfact needs PARMETIS");
708: #endif
709: }
711: lu->FactPattern = SamePattern;
712: PetscOptionsEList("-mat_superlu_dist_fact","Sparsity pattern for repeated matrix factorization","None",factPattern,3,factPattern[0],&indx,&flg);
713: if (flg) {
714: switch (indx) {
715: case 0:
716: lu->FactPattern = SamePattern;
717: break;
718: case 1:
719: lu->FactPattern = SamePattern_SameRowPerm;
720: break;
721: case 2:
722: lu->FactPattern = DOFACT;
723: break;
724: }
725: }
727: options.IterRefine = NOREFINE;
728: PetscOptionsBool("-mat_superlu_dist_iterrefine","Use iterative refinement","None",options.IterRefine == NOREFINE ? PETSC_FALSE : PETSC_TRUE ,&flg,&set);
729: if (set) {
730: if (flg) options.IterRefine = SLU_DOUBLE;
731: else options.IterRefine = NOREFINE;
732: }
734: if (PetscLogPrintInfo) options.PrintStat = YES;
735: else options.PrintStat = NO;
736: PetscOptionsBool("-mat_superlu_dist_statprint","Print factorization information","None",(PetscBool)options.PrintStat,(PetscBool*)&options.PrintStat,NULL);
737: PetscOptionsEnd();
739: lu->options = options;
740: lu->options.Fact = DOFACT;
741: lu->matsolve_iscalled = PETSC_FALSE;
742: lu->matmatsolve_iscalled = PETSC_FALSE;
744: PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_aij_superlu_dist);
745: PetscObjectComposeFunction((PetscObject)B,"MatSuperluDistGetDiagU_C",MatSuperluDistGetDiagU_SuperLU_DIST);
747: *F = B;
748: return(0);
749: }
751: PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_SuperLU_DIST(void)
752: {
755: MatSolverTypeRegister(MATSOLVERSUPERLU_DIST,MATMPIAIJ,MAT_FACTOR_LU,MatGetFactor_aij_superlu_dist);
756: MatSolverTypeRegister(MATSOLVERSUPERLU_DIST,MATSEQAIJ,MAT_FACTOR_LU,MatGetFactor_aij_superlu_dist);
757: MatSolverTypeRegister(MATSOLVERSUPERLU_DIST,MATMPIAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_superlu_dist);
758: MatSolverTypeRegister(MATSOLVERSUPERLU_DIST,MATSEQAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_superlu_dist);
759: return(0);
760: }
762: /*MC
763: MATSOLVERSUPERLU_DIST - Parallel direct solver package for LU factorization
765: Use ./configure --download-superlu_dist --download-parmetis --download-metis --download-ptscotch to have PETSc installed with SuperLU_DIST
767: Use -pc_type lu -pc_factor_mat_solver_type superlu_dist to use this direct solver
769: Works with AIJ matrices
771: Options Database Keys:
772: + -mat_superlu_dist_r <n> - number of rows in processor partition
773: . -mat_superlu_dist_c <n> - number of columns in processor partition
774: . -mat_superlu_dist_equil - equilibrate the matrix
775: . -mat_superlu_dist_rowperm <NOROWPERM,LargeDiag_MC64,LargeDiag_AWPM,MY_PERMR> - row permutation
776: . -mat_superlu_dist_colperm <NATURAL,MMD_AT_PLUS_A,MMD_ATA,METIS_AT_PLUS_A,PARMETIS> - column permutation
777: . -mat_superlu_dist_replacetinypivot - replace tiny pivots
778: . -mat_superlu_dist_fact <SamePattern> - (choose one of) SamePattern SamePattern_SameRowPerm DOFACT
779: . -mat_superlu_dist_iterrefine - use iterative refinement
780: - -mat_superlu_dist_statprint - print factorization information
782: Level: beginner
784: .seealso: PCLU
786: .seealso: PCFactorSetMatSolverType(), MatSolverType
788: M*/