Actual source code: bddcfetidp.c
petsc-3.14.6 2021-03-30
1: #include <../src/ksp/pc/impls/bddc/bddc.h>
2: #include <../src/ksp/pc/impls/bddc/bddcprivate.h>
3: #include <petscblaslapack.h>
5: static PetscErrorCode MatMult_BDdelta_deluxe_nonred(Mat A, Vec x, Vec y)
6: {
7: BDdelta_DN ctx;
11: MatShellGetContext(A,(void**)&ctx);
12: MatMultTranspose(ctx->BD,x,ctx->work);
13: KSPSolveTranspose(ctx->kBD,ctx->work,y);
14: /* No PC so cannot propagate up failure in KSPSolveTranspose() */
15: return(0);
16: }
18: static PetscErrorCode MatMultTranspose_BDdelta_deluxe_nonred(Mat A, Vec x, Vec y)
19: {
20: BDdelta_DN ctx;
24: MatShellGetContext(A,(void**)&ctx);
25: KSPSolve(ctx->kBD,x,ctx->work);
26: /* No PC so cannot propagate up failure in KSPSolve() */
27: MatMult(ctx->BD,ctx->work,y);
28: return(0);
29: }
31: static PetscErrorCode MatDestroy_BDdelta_deluxe_nonred(Mat A)
32: {
33: BDdelta_DN ctx;
37: MatShellGetContext(A,(void**)&ctx);
38: MatDestroy(&ctx->BD);
39: KSPDestroy(&ctx->kBD);
40: VecDestroy(&ctx->work);
41: PetscFree(ctx);
42: return(0);
43: }
46: PetscErrorCode PCBDDCCreateFETIDPMatContext(PC pc, FETIDPMat_ctx *fetidpmat_ctx)
47: {
48: FETIDPMat_ctx newctx;
52: PetscNew(&newctx);
53: /* increase the reference count for BDDC preconditioner */
54: PetscObjectReference((PetscObject)pc);
55: newctx->pc = pc;
56: *fetidpmat_ctx = newctx;
57: return(0);
58: }
60: PetscErrorCode PCBDDCCreateFETIDPPCContext(PC pc, FETIDPPC_ctx *fetidppc_ctx)
61: {
62: FETIDPPC_ctx newctx;
66: PetscNew(&newctx);
67: /* increase the reference count for BDDC preconditioner */
68: PetscObjectReference((PetscObject)pc);
69: newctx->pc = pc;
70: *fetidppc_ctx = newctx;
71: return(0);
72: }
74: PetscErrorCode PCBDDCDestroyFETIDPMat(Mat A)
75: {
76: FETIDPMat_ctx mat_ctx;
80: MatShellGetContext(A,(void**)&mat_ctx);
81: VecDestroy(&mat_ctx->lambda_local);
82: VecDestroy(&mat_ctx->temp_solution_D);
83: VecDestroy(&mat_ctx->temp_solution_B);
84: MatDestroy(&mat_ctx->B_delta);
85: MatDestroy(&mat_ctx->B_Ddelta);
86: MatDestroy(&mat_ctx->B_BB);
87: MatDestroy(&mat_ctx->B_BI);
88: MatDestroy(&mat_ctx->Bt_BB);
89: MatDestroy(&mat_ctx->Bt_BI);
90: MatDestroy(&mat_ctx->C);
91: VecDestroy(&mat_ctx->rhs_flip);
92: VecDestroy(&mat_ctx->vP);
93: VecDestroy(&mat_ctx->xPg);
94: VecDestroy(&mat_ctx->yPg);
95: VecScatterDestroy(&mat_ctx->l2g_lambda);
96: VecScatterDestroy(&mat_ctx->l2g_lambda_only);
97: VecScatterDestroy(&mat_ctx->l2g_p);
98: VecScatterDestroy(&mat_ctx->g2g_p);
99: PCDestroy(&mat_ctx->pc); /* decrease PCBDDC reference count */
100: ISDestroy(&mat_ctx->pressure);
101: ISDestroy(&mat_ctx->lagrange);
102: PetscFree(mat_ctx);
103: return(0);
104: }
106: PetscErrorCode PCBDDCDestroyFETIDPPC(PC pc)
107: {
108: FETIDPPC_ctx pc_ctx;
112: PCShellGetContext(pc,(void**)&pc_ctx);
113: VecDestroy(&pc_ctx->lambda_local);
114: MatDestroy(&pc_ctx->B_Ddelta);
115: VecScatterDestroy(&pc_ctx->l2g_lambda);
116: MatDestroy(&pc_ctx->S_j);
117: PCDestroy(&pc_ctx->pc); /* decrease PCBDDC reference count */
118: VecDestroy(&pc_ctx->xPg);
119: VecDestroy(&pc_ctx->yPg);
120: PetscFree(pc_ctx);
121: return(0);
122: }
124: PetscErrorCode PCBDDCSetupFETIDPMatContext(FETIDPMat_ctx fetidpmat_ctx)
125: {
127: PC_IS *pcis=(PC_IS*)fetidpmat_ctx->pc->data;
128: PC_BDDC *pcbddc=(PC_BDDC*)fetidpmat_ctx->pc->data;
129: PCBDDCGraph mat_graph=pcbddc->mat_graph;
130: Mat_IS *matis = (Mat_IS*)fetidpmat_ctx->pc->pmat->data;
131: MPI_Comm comm;
132: Mat ScalingMat,BD1,BD2;
133: Vec fetidp_global;
134: IS IS_l2g_lambda;
135: IS subset,subset_mult,subset_n,isvert;
136: PetscBool skip_node,fully_redundant;
137: PetscInt i,j,k,s,n_boundary_dofs,n_global_lambda,n_vertices,partial_sum;
138: PetscInt cum,n_local_lambda,n_lambda_for_dof,dual_size,n_neg_values,n_pos_values;
139: PetscMPIInt rank,size,buf_size,neigh;
140: PetscScalar scalar_value;
141: const PetscInt *vertex_indices;
142: PetscInt *dual_dofs_boundary_indices,*aux_local_numbering_1;
143: const PetscInt *aux_global_numbering;
144: PetscInt *aux_sums,*cols_B_delta,*l2g_indices;
145: PetscScalar *array,*scaling_factors,*vals_B_delta;
146: PetscScalar **all_factors;
147: PetscInt *aux_local_numbering_2;
148: PetscLayout llay;
150: /* saddlepoint */
151: ISLocalToGlobalMapping l2gmap_p;
152: PetscLayout play;
153: IS gP,pP;
154: PetscInt nPl,nPg,nPgl;
157: PetscObjectGetComm((PetscObject)(fetidpmat_ctx->pc),&comm);
158: MPI_Comm_rank(comm,&rank);
159: MPI_Comm_size(comm,&size);
161: /* saddlepoint */
162: nPl = 0;
163: nPg = 0;
164: nPgl = 0;
165: gP = NULL;
166: pP = NULL;
167: l2gmap_p = NULL;
168: play = NULL;
169: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_pP",(PetscObject*)&pP);
170: if (pP) { /* saddle point */
171: /* subdomain pressures in global numbering */
172: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_gP",(PetscObject*)&gP);
173: if (!gP) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"gP not present");
174: ISGetLocalSize(gP,&nPl);
175: VecCreate(PETSC_COMM_SELF,&fetidpmat_ctx->vP);
176: VecSetSizes(fetidpmat_ctx->vP,nPl,nPl);
177: VecSetType(fetidpmat_ctx->vP,VECSTANDARD);
178: VecSetUp(fetidpmat_ctx->vP);
180: /* pressure matrix */
181: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_C",(PetscObject*)&fetidpmat_ctx->C);
182: if (!fetidpmat_ctx->C) { /* null pressure block, compute layout and global numbering for pressures */
183: IS Pg;
185: ISRenumber(gP,NULL,&nPg,&Pg);
186: ISLocalToGlobalMappingCreateIS(Pg,&l2gmap_p);
187: ISDestroy(&Pg);
188: PetscLayoutCreate(comm,&play);
189: PetscLayoutSetBlockSize(play,1);
190: PetscLayoutSetSize(play,nPg);
191: ISGetLocalSize(pP,&nPgl);
192: PetscLayoutSetLocalSize(play,nPgl);
193: PetscLayoutSetUp(play);
194: } else {
195: PetscObjectReference((PetscObject)fetidpmat_ctx->C);
196: MatGetLocalToGlobalMapping(fetidpmat_ctx->C,&l2gmap_p,NULL);
197: PetscObjectReference((PetscObject)l2gmap_p);
198: MatGetSize(fetidpmat_ctx->C,&nPg,NULL);
199: MatGetLocalSize(fetidpmat_ctx->C,NULL,&nPgl);
200: MatGetLayouts(fetidpmat_ctx->C,NULL,&llay);
201: PetscLayoutReference(llay,&play);
202: }
203: VecCreateMPIWithArray(comm,1,nPgl,nPg,NULL,&fetidpmat_ctx->xPg);
204: VecCreateMPIWithArray(comm,1,nPgl,nPg,NULL,&fetidpmat_ctx->yPg);
206: /* import matrices for pressures coupling */
207: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_B_BI",(PetscObject*)&fetidpmat_ctx->B_BI);
208: if (!fetidpmat_ctx->B_BI) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"B_BI not present");
209: PetscObjectReference((PetscObject)fetidpmat_ctx->B_BI);
211: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_B_BB",(PetscObject*)&fetidpmat_ctx->B_BB);
212: if (!fetidpmat_ctx->B_BB) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"B_BB not present");
213: PetscObjectReference((PetscObject)fetidpmat_ctx->B_BB);
215: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_Bt_BI",(PetscObject*)&fetidpmat_ctx->Bt_BI);
216: if (!fetidpmat_ctx->Bt_BI) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Bt_BI not present");
217: PetscObjectReference((PetscObject)fetidpmat_ctx->Bt_BI);
219: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_Bt_BB",(PetscObject*)&fetidpmat_ctx->Bt_BB);
220: if (!fetidpmat_ctx->Bt_BB) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Bt_BB not present");
221: PetscObjectReference((PetscObject)fetidpmat_ctx->Bt_BB);
223: PetscObjectQuery((PetscObject)fetidpmat_ctx->pc,"__KSPFETIDP_flip" ,(PetscObject*)&fetidpmat_ctx->rhs_flip);
224: if (fetidpmat_ctx->rhs_flip) {
225: PetscObjectReference((PetscObject)fetidpmat_ctx->rhs_flip);
226: }
227: }
229: /* Default type of lagrange multipliers is non-redundant */
230: fully_redundant = fetidpmat_ctx->fully_redundant;
232: /* Evaluate local and global number of lagrange multipliers */
233: VecSet(pcis->vec1_N,0.0);
234: n_local_lambda = 0;
235: partial_sum = 0;
236: n_boundary_dofs = 0;
237: s = 0;
239: /* Get Vertices used to define the BDDC */
240: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&isvert);
241: ISGetLocalSize(isvert,&n_vertices);
242: ISGetIndices(isvert,&vertex_indices);
244: dual_size = pcis->n_B-n_vertices;
245: PetscMalloc1(dual_size,&dual_dofs_boundary_indices);
246: PetscMalloc1(dual_size,&aux_local_numbering_1);
247: PetscMalloc1(dual_size,&aux_local_numbering_2);
249: VecGetArray(pcis->vec1_N,&array);
250: for (i=0;i<pcis->n;i++){
251: j = mat_graph->count[i]; /* RECALL: mat_graph->count[i] does not count myself */
252: if (j > 0) n_boundary_dofs++;
253: skip_node = PETSC_FALSE;
254: if (s < n_vertices && vertex_indices[s] == i) { /* it works for a sorted set of vertices */
255: skip_node = PETSC_TRUE;
256: s++;
257: }
258: if (j < 1) skip_node = PETSC_TRUE;
259: if (mat_graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) skip_node = PETSC_TRUE;
260: if (!skip_node) {
261: if (fully_redundant) {
262: /* fully redundant set of lagrange multipliers */
263: n_lambda_for_dof = (j*(j+1))/2;
264: } else {
265: n_lambda_for_dof = j;
266: }
267: n_local_lambda += j;
268: /* needed to evaluate global number of lagrange multipliers */
269: array[i]=(1.0*n_lambda_for_dof)/(j+1.0); /* already scaled for the next global sum */
270: /* store some data needed */
271: dual_dofs_boundary_indices[partial_sum] = n_boundary_dofs-1;
272: aux_local_numbering_1[partial_sum] = i;
273: aux_local_numbering_2[partial_sum] = n_lambda_for_dof;
274: partial_sum++;
275: }
276: }
277: VecRestoreArray(pcis->vec1_N,&array);
278: ISRestoreIndices(isvert,&vertex_indices);
279: PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&isvert);
280: dual_size = partial_sum;
282: /* compute global ordering of lagrange multipliers and associate l2g map */
283: ISCreateGeneral(comm,partial_sum,aux_local_numbering_1,PETSC_COPY_VALUES,&subset_n);
284: ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);
285: ISDestroy(&subset_n);
286: ISCreateGeneral(comm,partial_sum,aux_local_numbering_2,PETSC_OWN_POINTER,&subset_mult);
287: ISRenumber(subset,subset_mult,&fetidpmat_ctx->n_lambda,&subset_n);
288: ISDestroy(&subset);
290: if (PetscDefined(USE_DEBUG)) {
291: VecSet(pcis->vec1_global,0.0);
292: VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
293: VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
294: VecSum(pcis->vec1_global,&scalar_value);
295: i = (PetscInt)PetscRealPart(scalar_value);
296: if (i != fetidpmat_ctx->n_lambda) SETERRQ2(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"Global number of multipliers mismatch! (%D != %D)",fetidpmat_ctx->n_lambda,i);
297: }
299: /* init data for scaling factors exchange */
300: if (!pcbddc->use_deluxe_scaling) {
301: PetscInt *ptrs_buffer,neigh_position;
302: PetscScalar *send_buffer,*recv_buffer;
303: MPI_Request *send_reqs,*recv_reqs;
305: partial_sum = 0;
306: PetscMalloc1(pcis->n_neigh,&ptrs_buffer);
307: PetscMalloc1(PetscMax(pcis->n_neigh-1,0),&send_reqs);
308: PetscMalloc1(PetscMax(pcis->n_neigh-1,0),&recv_reqs);
309: PetscMalloc1(pcis->n+1,&all_factors);
310: if (pcis->n_neigh > 0) ptrs_buffer[0]=0;
311: for (i=1;i<pcis->n_neigh;i++) {
312: partial_sum += pcis->n_shared[i];
313: ptrs_buffer[i] = ptrs_buffer[i-1]+pcis->n_shared[i];
314: }
315: PetscMalloc1(partial_sum,&send_buffer);
316: PetscMalloc1(partial_sum,&recv_buffer);
317: PetscMalloc1(partial_sum,&all_factors[0]);
318: for (i=0;i<pcis->n-1;i++) {
319: j = mat_graph->count[i];
320: all_factors[i+1]=all_factors[i]+j;
321: }
323: /* scatter B scaling to N vec */
324: VecScatterBegin(pcis->N_to_B,pcis->D,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);
325: VecScatterEnd(pcis->N_to_B,pcis->D,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);
326: /* communications */
327: VecGetArrayRead(pcis->vec1_N,(const PetscScalar**)&array);
328: for (i=1;i<pcis->n_neigh;i++) {
329: for (j=0;j<pcis->n_shared[i];j++) {
330: send_buffer[ptrs_buffer[i-1]+j]=array[pcis->shared[i][j]];
331: }
332: PetscMPIIntCast(ptrs_buffer[i]-ptrs_buffer[i-1],&buf_size);
333: PetscMPIIntCast(pcis->neigh[i],&neigh);
334: MPI_Isend(&send_buffer[ptrs_buffer[i-1]],buf_size,MPIU_SCALAR,neigh,0,comm,&send_reqs[i-1]);
335: MPI_Irecv(&recv_buffer[ptrs_buffer[i-1]],buf_size,MPIU_SCALAR,neigh,0,comm,&recv_reqs[i-1]);
336: }
337: VecRestoreArrayRead(pcis->vec1_N,(const PetscScalar**)&array);
338: if (pcis->n_neigh > 0) {
339: MPI_Waitall(pcis->n_neigh-1,recv_reqs,MPI_STATUSES_IGNORE);
340: }
341: /* put values in correct places */
342: for (i=1;i<pcis->n_neigh;i++) {
343: for (j=0;j<pcis->n_shared[i];j++) {
344: k = pcis->shared[i][j];
345: neigh_position = 0;
346: while (mat_graph->neighbours_set[k][neigh_position] != pcis->neigh[i]) {neigh_position++;}
347: all_factors[k][neigh_position]=recv_buffer[ptrs_buffer[i-1]+j];
348: }
349: }
350: if (pcis->n_neigh > 0) {
351: MPI_Waitall(pcis->n_neigh-1,send_reqs,MPI_STATUSES_IGNORE);
352: }
353: PetscFree(send_reqs);
354: PetscFree(recv_reqs);
355: PetscFree(send_buffer);
356: PetscFree(recv_buffer);
357: PetscFree(ptrs_buffer);
358: }
360: /* Compute B and B_delta (local actions) */
361: PetscMalloc1(pcis->n_neigh,&aux_sums);
362: PetscMalloc1(n_local_lambda,&l2g_indices);
363: PetscMalloc1(n_local_lambda,&vals_B_delta);
364: PetscMalloc1(n_local_lambda,&cols_B_delta);
365: if (!pcbddc->use_deluxe_scaling) {
366: PetscMalloc1(n_local_lambda,&scaling_factors);
367: } else {
368: scaling_factors = NULL;
369: all_factors = NULL;
370: }
371: ISGetIndices(subset_n,&aux_global_numbering);
372: partial_sum=0;
373: cum = 0;
374: for (i=0;i<dual_size;i++) {
375: n_global_lambda = aux_global_numbering[cum];
376: j = mat_graph->count[aux_local_numbering_1[i]];
377: aux_sums[0]=0;
378: for (s=1;s<j;s++) {
379: aux_sums[s]=aux_sums[s-1]+j-s+1;
380: }
381: if (all_factors) array = all_factors[aux_local_numbering_1[i]];
382: n_neg_values = 0;
383: while (n_neg_values < j && mat_graph->neighbours_set[aux_local_numbering_1[i]][n_neg_values] < rank) {n_neg_values++;}
384: n_pos_values = j - n_neg_values;
385: if (fully_redundant) {
386: for (s=0;s<n_neg_values;s++) {
387: l2g_indices [partial_sum+s]=aux_sums[s]+n_neg_values-s-1+n_global_lambda;
388: cols_B_delta [partial_sum+s]=dual_dofs_boundary_indices[i];
389: vals_B_delta [partial_sum+s]=-1.0;
390: if (!pcbddc->use_deluxe_scaling) scaling_factors[partial_sum+s]=array[s];
391: }
392: for (s=0;s<n_pos_values;s++) {
393: l2g_indices [partial_sum+s+n_neg_values]=aux_sums[n_neg_values]+s+n_global_lambda;
394: cols_B_delta [partial_sum+s+n_neg_values]=dual_dofs_boundary_indices[i];
395: vals_B_delta [partial_sum+s+n_neg_values]=1.0;
396: if (!pcbddc->use_deluxe_scaling) scaling_factors[partial_sum+s+n_neg_values]=array[s+n_neg_values];
397: }
398: partial_sum += j;
399: } else {
400: /* l2g_indices and default cols and vals of B_delta */
401: for (s=0;s<j;s++) {
402: l2g_indices [partial_sum+s]=n_global_lambda+s;
403: cols_B_delta [partial_sum+s]=dual_dofs_boundary_indices[i];
404: vals_B_delta [partial_sum+s]=0.0;
405: }
406: /* B_delta */
407: if (n_neg_values > 0) { /* there's a rank next to me to the left */
408: vals_B_delta [partial_sum+n_neg_values-1]=-1.0;
409: }
410: if (n_neg_values < j) { /* there's a rank next to me to the right */
411: vals_B_delta [partial_sum+n_neg_values]=1.0;
412: }
413: /* scaling as in Klawonn-Widlund 1999 */
414: if (!pcbddc->use_deluxe_scaling) {
415: for (s=0;s<n_neg_values;s++) {
416: scalar_value = 0.0;
417: for (k=0;k<s+1;k++) scalar_value += array[k];
418: scaling_factors[partial_sum+s] = -scalar_value;
419: }
420: for (s=0;s<n_pos_values;s++) {
421: scalar_value = 0.0;
422: for (k=s+n_neg_values;k<j;k++) scalar_value += array[k];
423: scaling_factors[partial_sum+s+n_neg_values] = scalar_value;
424: }
425: }
426: partial_sum += j;
427: }
428: cum += aux_local_numbering_2[i];
429: }
430: ISRestoreIndices(subset_n,&aux_global_numbering);
431: ISDestroy(&subset_mult);
432: ISDestroy(&subset_n);
433: PetscFree(aux_sums);
434: PetscFree(aux_local_numbering_1);
435: PetscFree(dual_dofs_boundary_indices);
436: if (all_factors) {
437: PetscFree(all_factors[0]);
438: PetscFree(all_factors);
439: }
441: /* Create local part of B_delta */
442: MatCreate(PETSC_COMM_SELF,&fetidpmat_ctx->B_delta);
443: MatSetSizes(fetidpmat_ctx->B_delta,n_local_lambda,pcis->n_B,n_local_lambda,pcis->n_B);
444: MatSetType(fetidpmat_ctx->B_delta,MATSEQAIJ);
445: MatSeqAIJSetPreallocation(fetidpmat_ctx->B_delta,1,NULL);
446: MatSetOption(fetidpmat_ctx->B_delta,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);
447: for (i=0;i<n_local_lambda;i++) {
448: MatSetValue(fetidpmat_ctx->B_delta,i,cols_B_delta[i],vals_B_delta[i],INSERT_VALUES);
449: }
450: PetscFree(vals_B_delta);
451: MatAssemblyBegin(fetidpmat_ctx->B_delta,MAT_FINAL_ASSEMBLY);
452: MatAssemblyEnd(fetidpmat_ctx->B_delta,MAT_FINAL_ASSEMBLY);
454: BD1 = NULL;
455: BD2 = NULL;
456: if (fully_redundant) {
457: if (pcbddc->use_deluxe_scaling) SETERRQ(comm,PETSC_ERR_SUP,"Deluxe FETIDP with fully-redundant multipliers to be implemented");
458: MatCreate(PETSC_COMM_SELF,&ScalingMat);
459: MatSetSizes(ScalingMat,n_local_lambda,n_local_lambda,n_local_lambda,n_local_lambda);
460: MatSetType(ScalingMat,MATSEQAIJ);
461: MatSeqAIJSetPreallocation(ScalingMat,1,NULL);
462: for (i=0;i<n_local_lambda;i++) {
463: MatSetValue(ScalingMat,i,i,scaling_factors[i],INSERT_VALUES);
464: }
465: MatAssemblyBegin(ScalingMat,MAT_FINAL_ASSEMBLY);
466: MatAssemblyEnd(ScalingMat,MAT_FINAL_ASSEMBLY);
467: MatMatMult(ScalingMat,fetidpmat_ctx->B_delta,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&fetidpmat_ctx->B_Ddelta);
468: MatDestroy(&ScalingMat);
469: } else {
470: MatCreate(PETSC_COMM_SELF,&fetidpmat_ctx->B_Ddelta);
471: MatSetSizes(fetidpmat_ctx->B_Ddelta,n_local_lambda,pcis->n_B,n_local_lambda,pcis->n_B);
472: if (!pcbddc->use_deluxe_scaling || !pcbddc->sub_schurs) {
473: MatSetType(fetidpmat_ctx->B_Ddelta,MATSEQAIJ);
474: MatSeqAIJSetPreallocation(fetidpmat_ctx->B_Ddelta,1,NULL);
475: for (i=0;i<n_local_lambda;i++) {
476: MatSetValue(fetidpmat_ctx->B_Ddelta,i,cols_B_delta[i],scaling_factors[i],INSERT_VALUES);
477: }
478: MatAssemblyBegin(fetidpmat_ctx->B_Ddelta,MAT_FINAL_ASSEMBLY);
479: MatAssemblyEnd(fetidpmat_ctx->B_Ddelta,MAT_FINAL_ASSEMBLY);
480: } else {
481: /* scaling as in Klawonn-Widlund 1999 */
482: PCBDDCDeluxeScaling deluxe_ctx = pcbddc->deluxe_ctx;
483: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
484: Mat T;
485: PetscScalar *W,lwork,*Bwork;
486: const PetscInt *idxs = NULL;
487: PetscInt cum,mss,*nnz;
488: PetscBLASInt *pivots,B_lwork,B_N,B_ierr;
490: if (!pcbddc->deluxe_singlemat) SETERRQ(comm,PETSC_ERR_USER,"Cannot compute B_Ddelta! rerun with -pc_bddc_deluxe_singlemat");
491: mss = 0;
492: PetscCalloc1(pcis->n_B,&nnz);
493: if (sub_schurs->is_Ej_all) {
494: ISGetIndices(sub_schurs->is_Ej_all,&idxs);
495: for (i=0,cum=0;i<sub_schurs->n_subs;i++) {
496: PetscInt subset_size;
498: ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);
499: for (j=0;j<subset_size;j++) nnz[idxs[j+cum]] = subset_size;
500: mss = PetscMax(mss,subset_size);
501: cum += subset_size;
502: }
503: }
504: MatCreate(PETSC_COMM_SELF,&T);
505: MatSetSizes(T,pcis->n_B,pcis->n_B,pcis->n_B,pcis->n_B);
506: MatSetType(T,MATSEQAIJ);
507: MatSeqAIJSetPreallocation(T,0,nnz);
508: PetscFree(nnz);
510: /* workspace allocation */
511: B_lwork = 0;
512: if (mss) {
513: PetscScalar dummy = 1;
515: B_lwork = -1;
516: PetscBLASIntCast(mss,&B_N);
517: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
518: PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,&dummy,&B_N,&B_N,&lwork,&B_lwork,&B_ierr));
519: PetscFPTrapPop();
520: if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr);
521: PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);
522: }
523: PetscMalloc3(mss*mss,&W,mss,&pivots,B_lwork,&Bwork);
525: for (i=0,cum=0;i<sub_schurs->n_subs;i++) {
526: const PetscScalar *M;
527: PetscInt subset_size;
529: ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);
530: PetscBLASIntCast(subset_size,&B_N);
531: MatDenseGetArrayRead(deluxe_ctx->seq_mat[i],&M);
532: PetscArraycpy(W,M,subset_size*subset_size);
533: MatDenseRestoreArrayRead(deluxe_ctx->seq_mat[i],&M);
534: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
535: PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,W,&B_N,pivots,&B_ierr));
536: if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr);
537: PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,W,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
538: if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr);
539: PetscFPTrapPop();
540: /* silent static analyzer */
541: if (!idxs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"IDXS not present");
542: MatSetValues(T,subset_size,idxs+cum,subset_size,idxs+cum,W,INSERT_VALUES);
543: cum += subset_size;
544: }
545: MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);
546: MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);
547: MatMatTransposeMult(T,fetidpmat_ctx->B_delta,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&BD1);
548: MatMatMult(fetidpmat_ctx->B_delta,BD1,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&BD2);
549: MatDestroy(&T);
550: PetscFree3(W,pivots,Bwork);
551: if (sub_schurs->is_Ej_all) {
552: ISRestoreIndices(sub_schurs->is_Ej_all,&idxs);
553: }
554: }
555: }
556: PetscFree(scaling_factors);
557: PetscFree(cols_B_delta);
559: /* Layout of multipliers */
560: PetscLayoutCreate(comm,&llay);
561: PetscLayoutSetBlockSize(llay,1);
562: PetscLayoutSetSize(llay,fetidpmat_ctx->n_lambda);
563: PetscLayoutSetUp(llay);
564: PetscLayoutGetLocalSize(llay,&fetidpmat_ctx->n);
566: /* Local work vector of multipliers */
567: VecCreate(PETSC_COMM_SELF,&fetidpmat_ctx->lambda_local);
568: VecSetSizes(fetidpmat_ctx->lambda_local,n_local_lambda,n_local_lambda);
569: VecSetType(fetidpmat_ctx->lambda_local,VECSEQ);
571: if (BD2) {
572: ISLocalToGlobalMapping l2g;
573: Mat T,TA,*pT;
574: IS is;
575: PetscInt nl,N;
576: BDdelta_DN ctx;
578: PetscLayoutGetLocalSize(llay,&nl);
579: PetscLayoutGetSize(llay,&N);
580: MatCreate(comm,&T);
581: MatSetSizes(T,nl,nl,N,N);
582: MatSetType(T,MATIS);
583: ISLocalToGlobalMappingCreate(comm,1,n_local_lambda,l2g_indices,PETSC_COPY_VALUES,&l2g);
584: MatSetLocalToGlobalMapping(T,l2g,l2g);
585: ISLocalToGlobalMappingDestroy(&l2g);
586: MatISSetLocalMat(T,BD2);
587: MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);
588: MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);
589: MatDestroy(&BD2);
590: MatConvert(T,MATAIJ,MAT_INITIAL_MATRIX,&TA);
591: MatDestroy(&T);
592: ISCreateGeneral(comm,n_local_lambda,l2g_indices,PETSC_USE_POINTER,&is);
593: MatCreateSubMatrices(TA,1,&is,&is,MAT_INITIAL_MATRIX,&pT);
594: MatDestroy(&TA);
595: ISDestroy(&is);
596: BD2 = pT[0];
597: PetscFree(pT);
599: /* B_Ddelta for non-redundant multipliers with deluxe scaling */
600: PetscNew(&ctx);
601: MatSetType(fetidpmat_ctx->B_Ddelta,MATSHELL);
602: MatShellSetContext(fetidpmat_ctx->B_Ddelta,(void *)ctx);
603: MatShellSetOperation(fetidpmat_ctx->B_Ddelta,MATOP_MULT,(void (*)(void))MatMult_BDdelta_deluxe_nonred);
604: MatShellSetOperation(fetidpmat_ctx->B_Ddelta,MATOP_MULT_TRANSPOSE,(void (*)(void))MatMultTranspose_BDdelta_deluxe_nonred);
605: MatShellSetOperation(fetidpmat_ctx->B_Ddelta,MATOP_DESTROY,(void (*)(void))MatDestroy_BDdelta_deluxe_nonred);
606: MatSetUp(fetidpmat_ctx->B_Ddelta);
608: PetscObjectReference((PetscObject)BD1);
609: ctx->BD = BD1;
610: KSPCreate(PETSC_COMM_SELF,&ctx->kBD);
611: KSPSetOperators(ctx->kBD,BD2,BD2);
612: VecDuplicate(fetidpmat_ctx->lambda_local,&ctx->work);
613: fetidpmat_ctx->deluxe_nonred = PETSC_TRUE;
614: }
615: MatDestroy(&BD1);
616: MatDestroy(&BD2);
618: /* fetidpmat sizes */
619: fetidpmat_ctx->n += nPgl;
620: fetidpmat_ctx->N = fetidpmat_ctx->n_lambda+nPg;
622: /* Global vector for FETI-DP linear system */
623: VecCreate(comm,&fetidp_global);
624: VecSetSizes(fetidp_global,fetidpmat_ctx->n,fetidpmat_ctx->N);
625: VecSetType(fetidp_global,VECMPI);
626: VecSetUp(fetidp_global);
628: /* Decide layout for fetidp dofs: if it is a saddle point problem
629: pressure is ordered first in the local part of the global vector
630: of the FETI-DP linear system */
631: if (nPg) {
632: Vec v;
633: IS IS_l2g_p,ais;
634: PetscLayout alay;
635: const PetscInt *idxs,*pranges,*aranges,*lranges;
636: PetscInt *l2g_indices_p,rst;
637: PetscMPIInt rank;
639: PetscMalloc1(nPl,&l2g_indices_p);
640: VecGetLayout(fetidp_global,&alay);
641: PetscLayoutGetRanges(alay,&aranges);
642: PetscLayoutGetRanges(play,&pranges);
643: PetscLayoutGetRanges(llay,&lranges);
645: MPI_Comm_rank(PetscObjectComm((PetscObject)fetidp_global),&rank);
646: ISCreateStride(PetscObjectComm((PetscObject)fetidp_global),pranges[rank+1]-pranges[rank],aranges[rank],1,&fetidpmat_ctx->pressure);
647: PetscObjectSetName((PetscObject)fetidpmat_ctx->pressure,"F_P");
648: ISCreateStride(PetscObjectComm((PetscObject)fetidp_global),lranges[rank+1]-lranges[rank],aranges[rank]+pranges[rank+1]-pranges[rank],1,&fetidpmat_ctx->lagrange);
649: PetscObjectSetName((PetscObject)fetidpmat_ctx->lagrange,"F_L");
650: ISLocalToGlobalMappingGetIndices(l2gmap_p,&idxs);
651: /* shift local to global indices for pressure */
652: for (i=0;i<nPl;i++) {
653: PetscMPIInt owner;
655: PetscLayoutFindOwner(play,idxs[i],&owner);
656: l2g_indices_p[i] = idxs[i]-pranges[owner]+aranges[owner];
657: }
658: ISLocalToGlobalMappingRestoreIndices(l2gmap_p,&idxs);
659: ISCreateGeneral(comm,nPl,l2g_indices_p,PETSC_OWN_POINTER,&IS_l2g_p);
661: /* local to global scatter for pressure */
662: VecScatterCreate(fetidpmat_ctx->vP,NULL,fetidp_global,IS_l2g_p,&fetidpmat_ctx->l2g_p);
663: ISDestroy(&IS_l2g_p);
665: /* scatter for lagrange multipliers only */
666: VecCreate(comm,&v);
667: VecSetType(v,VECSTANDARD);
668: VecSetLayout(v,llay);
669: VecSetUp(v);
670: ISCreateGeneral(comm,n_local_lambda,l2g_indices,PETSC_COPY_VALUES,&ais);
671: VecScatterCreate(fetidpmat_ctx->lambda_local,NULL,v,ais,&fetidpmat_ctx->l2g_lambda_only);
672: ISDestroy(&ais);
673: VecDestroy(&v);
675: /* shift local to global indices for multipliers */
676: for (i=0;i<n_local_lambda;i++) {
677: PetscInt ps;
678: PetscMPIInt owner;
680: PetscLayoutFindOwner(llay,l2g_indices[i],&owner);
681: ps = pranges[owner+1]-pranges[owner];
682: l2g_indices[i] = l2g_indices[i]-lranges[owner]+aranges[owner]+ps;
683: }
685: /* scatter from alldofs to pressures global fetidp vector */
686: PetscLayoutGetRange(alay,&rst,NULL);
687: ISCreateStride(comm,nPgl,rst,1,&ais);
688: VecScatterCreate(pcis->vec1_global,pP,fetidp_global,ais,&fetidpmat_ctx->g2g_p);
689: ISDestroy(&ais);
690: }
691: PetscLayoutDestroy(&llay);
692: PetscLayoutDestroy(&play);
693: ISCreateGeneral(comm,n_local_lambda,l2g_indices,PETSC_OWN_POINTER,&IS_l2g_lambda);
695: /* scatter from local to global multipliers */
696: VecScatterCreate(fetidpmat_ctx->lambda_local,NULL,fetidp_global,IS_l2g_lambda,&fetidpmat_ctx->l2g_lambda);
697: ISDestroy(&IS_l2g_lambda);
698: ISLocalToGlobalMappingDestroy(&l2gmap_p);
699: VecDestroy(&fetidp_global);
701: /* Create some work vectors needed by fetidp */
702: VecDuplicate(pcis->vec1_B,&fetidpmat_ctx->temp_solution_B);
703: VecDuplicate(pcis->vec1_D,&fetidpmat_ctx->temp_solution_D);
704: return(0);
705: }
708: PetscErrorCode PCBDDCSetupFETIDPPCContext(Mat fetimat, FETIDPPC_ctx fetidppc_ctx)
709: {
710: FETIDPMat_ctx mat_ctx;
711: PC_BDDC *pcbddc = (PC_BDDC*)fetidppc_ctx->pc->data;
712: PC_IS *pcis = (PC_IS*)fetidppc_ctx->pc->data;
713: PetscBool lumped = PETSC_FALSE;
717: MatShellGetContext(fetimat,(void**)&mat_ctx);
718: /* get references from objects created when setting up feti mat context */
719: PetscObjectReference((PetscObject)mat_ctx->lambda_local);
720: fetidppc_ctx->lambda_local = mat_ctx->lambda_local;
721: PetscObjectReference((PetscObject)mat_ctx->B_Ddelta);
722: fetidppc_ctx->B_Ddelta = mat_ctx->B_Ddelta;
723: if (mat_ctx->deluxe_nonred) {
724: PC pc,mpc;
725: BDdelta_DN ctx;
726: MatSolverType solver;
727: const char *prefix;
729: MatShellGetContext(mat_ctx->B_Ddelta,&ctx);
730: KSPSetType(ctx->kBD,KSPPREONLY);
731: KSPGetPC(ctx->kBD,&mpc);
732: KSPGetPC(pcbddc->ksp_D,&pc);
733: PCSetType(mpc,PCLU);
734: PCFactorGetMatSolverType(pc,(MatSolverType*)&solver);
735: if (solver) {
736: PCFactorSetMatSolverType(mpc,solver);
737: }
738: MatGetOptionsPrefix(fetimat,&prefix);
739: KSPSetOptionsPrefix(ctx->kBD,prefix);
740: KSPAppendOptionsPrefix(ctx->kBD,"bddelta_");
741: KSPSetFromOptions(ctx->kBD);
742: }
744: if (mat_ctx->l2g_lambda_only) {
745: PetscObjectReference((PetscObject)mat_ctx->l2g_lambda_only);
746: fetidppc_ctx->l2g_lambda = mat_ctx->l2g_lambda_only;
747: } else {
748: PetscObjectReference((PetscObject)mat_ctx->l2g_lambda);
749: fetidppc_ctx->l2g_lambda = mat_ctx->l2g_lambda;
750: }
751: /* Dirichlet preconditioner */
752: PetscOptionsGetBool(NULL,((PetscObject)fetimat)->prefix,"-pc_lumped",&lumped,NULL);
753: if (!lumped) {
754: IS iV;
755: PetscBool discrete_harmonic = PETSC_FALSE;
757: PetscObjectQuery((PetscObject)fetidppc_ctx->pc,"__KSPFETIDP_iV",(PetscObject*)&iV);
758: if (iV) {
759: PetscOptionsGetBool(NULL,((PetscObject)fetimat)->prefix,"-pc_discrete_harmonic",&discrete_harmonic,NULL);
760: }
761: if (discrete_harmonic) {
762: KSP sksp;
763: PC pc;
764: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
765: Mat A_II,A_IB,A_BI;
766: IS iP = NULL;
767: PetscBool isshell,reuse = PETSC_FALSE;
768: KSPType ksptype;
769: const char *prefix;
771: /*
772: We constructs a Schur complement for
774: | A_II A_ID |
775: | A_DI A_DD |
777: instead of
779: | A_II B^t_II A_ID |
780: | B_II -C_II B_ID |
781: | A_DI B^t_ID A_DD |
783: */
784: if (sub_schurs && sub_schurs->reuse_solver) {
785: PetscObjectQuery((PetscObject)sub_schurs->A,"__KSPFETIDP_iP",(PetscObject*)&iP);
786: if (iP) reuse = PETSC_TRUE;
787: }
788: if (!reuse) {
789: IS aB;
790: PetscInt nb;
791: ISGetLocalSize(pcis->is_B_local,&nb);
792: ISCreateStride(PetscObjectComm((PetscObject)pcis->A_II),nb,0,1,&aB);
793: MatCreateSubMatrix(pcis->A_II,iV,iV,MAT_INITIAL_MATRIX,&A_II);
794: MatCreateSubMatrix(pcis->A_IB,iV,aB,MAT_INITIAL_MATRIX,&A_IB);
795: MatCreateSubMatrix(pcis->A_BI,aB,iV,MAT_INITIAL_MATRIX,&A_BI);
796: ISDestroy(&aB);
797: } else {
798: MatCreateSubMatrix(sub_schurs->A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);
799: MatCreateSubMatrix(sub_schurs->A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);
800: PetscObjectReference((PetscObject)pcis->A_II);
801: A_II = pcis->A_II;
802: }
803: MatCreateSchurComplement(A_II,A_II,A_IB,A_BI,pcis->A_BB,&fetidppc_ctx->S_j);
805: /* propagate settings of solver */
806: MatSchurComplementGetKSP(fetidppc_ctx->S_j,&sksp);
807: KSPGetType(pcis->ksp_D,&ksptype);
808: KSPSetType(sksp,ksptype);
809: KSPGetPC(pcis->ksp_D,&pc);
810: PetscObjectTypeCompare((PetscObject)pc,PCSHELL,&isshell);
811: if (!isshell) {
812: MatSolverType solver;
813: PCType pctype;
815: PCGetType(pc,&pctype);
816: PCFactorGetMatSolverType(pc,(MatSolverType*)&solver);
817: KSPGetPC(sksp,&pc);
818: PCSetType(pc,pctype);
819: if (solver) {
820: PCFactorSetMatSolverType(pc,solver);
821: }
822: } else {
823: KSPGetPC(sksp,&pc);
824: PCSetType(pc,PCLU);
825: }
826: MatDestroy(&A_II);
827: MatDestroy(&A_IB);
828: MatDestroy(&A_BI);
829: MatGetOptionsPrefix(fetimat,&prefix);
830: KSPSetOptionsPrefix(sksp,prefix);
831: KSPAppendOptionsPrefix(sksp,"harmonic_");
832: KSPSetFromOptions(sksp);
833: if (reuse) {
834: KSPSetPC(sksp,sub_schurs->reuse_solver->interior_solver);
835: PetscObjectIncrementTabLevel((PetscObject)sub_schurs->reuse_solver->interior_solver,(PetscObject)sksp,0);
836: }
837: } else { /* default Dirichlet preconditioner is pde-harmonic */
838: MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&fetidppc_ctx->S_j);
839: MatSchurComplementSetKSP(fetidppc_ctx->S_j,pcis->ksp_D);
840: }
841: } else {
842: PetscObjectReference((PetscObject)pcis->A_BB);
843: fetidppc_ctx->S_j = pcis->A_BB;
844: }
845: /* saddle-point */
846: if (mat_ctx->xPg) {
847: PetscObjectReference((PetscObject)mat_ctx->xPg);
848: fetidppc_ctx->xPg = mat_ctx->xPg;
849: PetscObjectReference((PetscObject)mat_ctx->yPg);
850: fetidppc_ctx->yPg = mat_ctx->yPg;
851: }
852: return(0);
853: }
855: PetscErrorCode FETIDPMatMult_Kernel(Mat fetimat, Vec x, Vec y, PetscBool trans)
856: {
857: FETIDPMat_ctx mat_ctx;
858: PC_BDDC *pcbddc;
859: PC_IS *pcis;
863: MatShellGetContext(fetimat,(void**)&mat_ctx);
864: pcis = (PC_IS*)mat_ctx->pc->data;
865: pcbddc = (PC_BDDC*)mat_ctx->pc->data;
866: /* Application of B_delta^T */
867: VecSet(pcis->vec1_B,0.);
868: VecScatterBegin(mat_ctx->l2g_lambda,x,mat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);
869: VecScatterEnd(mat_ctx->l2g_lambda,x,mat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);
870: MatMultTranspose(mat_ctx->B_delta,mat_ctx->lambda_local,pcis->vec1_B);
872: /* Add contribution from saddle point */
873: if (mat_ctx->l2g_p) {
874: VecScatterBegin(mat_ctx->l2g_p,x,mat_ctx->vP,INSERT_VALUES,SCATTER_REVERSE);
875: VecScatterEnd(mat_ctx->l2g_p,x,mat_ctx->vP,INSERT_VALUES,SCATTER_REVERSE);
876: if (pcbddc->switch_static) {
877: if (trans) {
878: MatMultTranspose(mat_ctx->B_BI,mat_ctx->vP,pcis->vec1_D);
879: } else {
880: MatMult(mat_ctx->Bt_BI,mat_ctx->vP,pcis->vec1_D);
881: }
882: }
883: if (trans) {
884: MatMultTransposeAdd(mat_ctx->B_BB,mat_ctx->vP,pcis->vec1_B,pcis->vec1_B);
885: } else {
886: MatMultAdd(mat_ctx->Bt_BB,mat_ctx->vP,pcis->vec1_B,pcis->vec1_B);
887: }
888: } else {
889: if (pcbddc->switch_static) {
890: VecSet(pcis->vec1_D,0.0);
891: }
892: }
893: /* Application of \widetilde{S}^-1 */
894: PetscArrayzero(pcbddc->benign_p0,pcbddc->benign_n);
895: PCBDDCApplyInterfacePreconditioner(mat_ctx->pc,trans);
896: PetscArrayzero(pcbddc->benign_p0,pcbddc->benign_n);
897: VecSet(y,0.0);
898: /* Application of B_delta */
899: MatMult(mat_ctx->B_delta,pcis->vec1_B,mat_ctx->lambda_local);
900: /* Contribution from boundary pressures */
901: if (mat_ctx->C) {
902: const PetscScalar *lx;
903: PetscScalar *ly;
905: /* pressure ordered first in the local part of x and y */
906: VecGetArrayRead(x,&lx);
907: VecGetArray(y,&ly);
908: VecPlaceArray(mat_ctx->xPg,lx);
909: VecPlaceArray(mat_ctx->yPg,ly);
910: if (trans) {
911: MatMultTranspose(mat_ctx->C,mat_ctx->xPg,mat_ctx->yPg);
912: } else {
913: MatMult(mat_ctx->C,mat_ctx->xPg,mat_ctx->yPg);
914: }
915: VecResetArray(mat_ctx->xPg);
916: VecResetArray(mat_ctx->yPg);
917: VecRestoreArrayRead(x,&lx);
918: VecRestoreArray(y,&ly);
919: }
920: /* Add contribution from saddle point */
921: if (mat_ctx->l2g_p) {
922: if (trans) {
923: MatMultTranspose(mat_ctx->Bt_BB,pcis->vec1_B,mat_ctx->vP);
924: } else {
925: MatMult(mat_ctx->B_BB,pcis->vec1_B,mat_ctx->vP);
926: }
927: if (pcbddc->switch_static) {
928: if (trans) {
929: MatMultTransposeAdd(mat_ctx->Bt_BI,pcis->vec1_D,mat_ctx->vP,mat_ctx->vP);
930: } else {
931: MatMultAdd(mat_ctx->B_BI,pcis->vec1_D,mat_ctx->vP,mat_ctx->vP);
932: }
933: }
934: VecScatterBegin(mat_ctx->l2g_p,mat_ctx->vP,y,ADD_VALUES,SCATTER_FORWARD);
935: VecScatterEnd(mat_ctx->l2g_p,mat_ctx->vP,y,ADD_VALUES,SCATTER_FORWARD);
936: }
937: VecScatterBegin(mat_ctx->l2g_lambda,mat_ctx->lambda_local,y,ADD_VALUES,SCATTER_FORWARD);
938: VecScatterEnd(mat_ctx->l2g_lambda,mat_ctx->lambda_local,y,ADD_VALUES,SCATTER_FORWARD);
939: return(0);
940: }
942: PetscErrorCode FETIDPMatMult(Mat fetimat, Vec x, Vec y)
943: {
947: FETIDPMatMult_Kernel(fetimat,x,y,PETSC_FALSE);
948: return(0);
949: }
951: PetscErrorCode FETIDPMatMultTranspose(Mat fetimat, Vec x, Vec y)
952: {
956: FETIDPMatMult_Kernel(fetimat,x,y,PETSC_TRUE);
957: return(0);
958: }
960: PetscErrorCode FETIDPPCApply_Kernel(PC fetipc, Vec x, Vec y, PetscBool trans)
961: {
962: FETIDPPC_ctx pc_ctx;
963: PC_IS *pcis;
967: PCShellGetContext(fetipc,(void**)&pc_ctx);
968: pcis = (PC_IS*)pc_ctx->pc->data;
969: /* Application of B_Ddelta^T */
970: VecScatterBegin(pc_ctx->l2g_lambda,x,pc_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);
971: VecScatterEnd(pc_ctx->l2g_lambda,x,pc_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);
972: VecSet(pcis->vec2_B,0.0);
973: MatMultTranspose(pc_ctx->B_Ddelta,pc_ctx->lambda_local,pcis->vec2_B);
974: /* Application of local Schur complement */
975: if (trans) {
976: MatMultTranspose(pc_ctx->S_j,pcis->vec2_B,pcis->vec1_B);
977: } else {
978: MatMult(pc_ctx->S_j,pcis->vec2_B,pcis->vec1_B);
979: }
980: /* Application of B_Ddelta */
981: MatMult(pc_ctx->B_Ddelta,pcis->vec1_B,pc_ctx->lambda_local);
982: VecSet(y,0.0);
983: VecScatterBegin(pc_ctx->l2g_lambda,pc_ctx->lambda_local,y,ADD_VALUES,SCATTER_FORWARD);
984: VecScatterEnd(pc_ctx->l2g_lambda,pc_ctx->lambda_local,y,ADD_VALUES,SCATTER_FORWARD);
985: return(0);
986: }
988: PetscErrorCode FETIDPPCApply(PC pc, Vec x, Vec y)
989: {
993: FETIDPPCApply_Kernel(pc,x,y,PETSC_FALSE);
994: return(0);
995: }
997: PetscErrorCode FETIDPPCApplyTranspose(PC pc, Vec x, Vec y)
998: {
1002: FETIDPPCApply_Kernel(pc,x,y,PETSC_TRUE);
1003: return(0);
1004: }
1006: PetscErrorCode FETIDPPCView(PC pc, PetscViewer viewer)
1007: {
1008: FETIDPPC_ctx pc_ctx;
1009: PetscBool iascii;
1010: PetscViewer sviewer;
1011: PetscErrorCode ierr;
1014: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1015: if (iascii) {
1016: PetscMPIInt rank;
1017: PetscBool isschur,isshell;
1019: PCShellGetContext(pc,(void**)&pc_ctx);
1020: MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);
1021: PetscObjectTypeCompare((PetscObject)pc_ctx->S_j,MATSCHURCOMPLEMENT,&isschur);
1022: if (isschur) {
1023: PetscViewerASCIIPrintf(viewer," Dirichlet preconditioner (just from rank 0)\n");
1024: } else {
1025: PetscViewerASCIIPrintf(viewer," Lumped preconditioner (just from rank 0)\n");
1026: }
1027: PetscViewerGetSubViewer(viewer,PetscObjectComm((PetscObject)pc_ctx->S_j),&sviewer);
1028: if (!rank) {
1029: PetscViewerPushFormat(sviewer,PETSC_VIEWER_ASCII_INFO);
1030: PetscViewerASCIIPushTab(sviewer);
1031: MatView(pc_ctx->S_j,sviewer);
1032: PetscViewerASCIIPopTab(sviewer);
1033: PetscViewerPopFormat(sviewer);
1034: }
1035: PetscViewerRestoreSubViewer(viewer,PetscObjectComm((PetscObject)pc_ctx->S_j),&sviewer);
1036: PetscObjectTypeCompare((PetscObject)pc_ctx->B_Ddelta,MATSHELL,&isshell);
1037: if (isshell) {
1038: BDdelta_DN ctx;
1039: PetscViewerASCIIPrintf(viewer," FETI-DP BDdelta: DB^t * (B D^-1 B^t)^-1 for deluxe scaling (just from rank 0)\n");
1040: MatShellGetContext(pc_ctx->B_Ddelta,&ctx);
1041: PetscViewerGetSubViewer(viewer,PetscObjectComm((PetscObject)pc_ctx->S_j),&sviewer);
1042: if (!rank) {
1043: PetscInt tl;
1045: PetscViewerASCIIGetTab(sviewer,&tl);
1046: PetscObjectSetTabLevel((PetscObject)ctx->kBD,tl);
1047: KSPView(ctx->kBD,sviewer);
1048: PetscViewerPushFormat(sviewer,PETSC_VIEWER_ASCII_INFO);
1049: MatView(ctx->BD,sviewer);
1050: PetscViewerPopFormat(sviewer);
1051: }
1052: PetscViewerRestoreSubViewer(viewer,PetscObjectComm((PetscObject)pc_ctx->S_j),&sviewer);
1053: }
1054: PetscViewerFlush(viewer);
1055: }
1056: return(0);
1057: }