/[escript]/branches/doubleplusgood/paso/src/RILU.cpp
ViewVC logotype

Contents of /branches/doubleplusgood/paso/src/RILU.cpp

Parent Directory Parent Directory | Revision Log Revision Log


Revision 3259 - (show annotations)
Mon Oct 11 01:48:14 2010 UTC (8 years, 7 months ago) by jfenwick
Original Path: trunk/paso/src/RILU.c
File MIME type: text/plain
File size: 15976 byte(s)
Merging dudley and scons updates from branches

1
2 /*******************************************************
3 *
4 * Copyright (c) 2003-2010 by University of Queensland
5 * Earth Systems Science Computational Center (ESSCC)
6 * http://www.uq.edu.au/esscc
7 *
8 * Primary Business: Queensland, Australia
9 * Licensed under the Open Software License version 3.0
10 * http://www.opensource.org/licenses/osl-3.0.php
11 *
12 *******************************************************/
13
14
15 /**************************************************************/
16
17 /* Paso: RILU preconditioner with reordering */
18
19 /**************************************************************/
20
21 /* Author: Lutz Gross, l.gross@uq.edu.au */
22
23 /**************************************************************/
24
25 #include "Paso.h"
26 #include "Preconditioner.h"
27 #include "PasoUtil.h"
28 #include "BlockOps.h"
29
30 /**************************************************************/
31
32 /* free all memory used by RILU */
33
34 void Paso_Solver_RILU_free(Paso_Solver_RILU * in) {
35 if (in!=NULL) {
36 Paso_Solver_RILU_free(in->RILU_of_Schur);
37 MEMFREE(in->inv_A_FF);
38 MEMFREE(in->A_FF_pivot);
39 Paso_SparseMatrix_free(in->A_FC);
40 Paso_SparseMatrix_free(in->A_CF);
41 MEMFREE(in->rows_in_F);
42 MEMFREE(in->rows_in_C);
43 MEMFREE(in->mask_F);
44 MEMFREE(in->mask_C);
45 MEMFREE(in->x_F);
46 MEMFREE(in->b_F);
47 MEMFREE(in->x_C);
48 MEMFREE(in->b_C);
49 MEMFREE(in);
50 }
51 }
52
53 /**************************************************************/
54
55 /* constructs the block-block factorization of
56
57 [ A_FF A_FC ]
58 A_p=
59 [ A_CF A_FF ]
60
61 to
62
63 [ I 0 ] [ A_FF 0 ] [ I invA_FF*A_FF ]
64 [ A_CF*invA_FF I ] [ 0 S ] [ 0 I ]
65
66
67 where S=A_FF-ACF*invA_FF*A_FC within the shape of S
68
69 then RILU is applied to S again until S becomes empty
70
71 */
72 Paso_Solver_RILU* Paso_Solver_getRILU(Paso_SparseMatrix *A_p,bool_t verbose) {
73 Paso_Solver_RILU* out=NULL;
74 dim_t n=A_p->numRows;
75 dim_t n_block=A_p->row_block_size;
76 index_t* mis_marker=NULL;
77 index_t* counter=NULL;
78 index_t iPtr,*index, *where_p;
79 dim_t i,k;
80 Paso_SparseMatrix * schur=NULL;
81 double A11,A12,A13,A21,A22,A23,A31,A32,A33,D,time0=0,time1=0,time2=0;
82
83
84 /* identify independend set of rows/columns */
85 mis_marker=TMPMEMALLOC(n,index_t);
86 counter=TMPMEMALLOC(n,index_t);
87 out=MEMALLOC(1,Paso_Solver_RILU);
88 out->RILU_of_Schur=NULL;
89 out->inv_A_FF=NULL;
90 out->A_FF_pivot=NULL;
91 out->A_FC=NULL;
92 out->A_CF=NULL;
93 out->rows_in_F=NULL;
94 out->rows_in_C=NULL;
95 out->mask_F=NULL;
96 out->mask_C=NULL;
97 out->x_F=NULL;
98 out->b_F=NULL;
99 out->x_C=NULL;
100 out->b_C=NULL;
101
102 if ( !(Esys_checkPtr(mis_marker) || Esys_checkPtr(out) || Esys_checkPtr(counter) ) ) {
103 /* identify independend set of rows/columns */
104 time0=Esys_timer();
105 #pragma omp parallel for private(i) schedule(static)
106 for (i=0;i<n;++i) mis_marker[i]=-1;
107 Paso_Pattern_mis(A_p->pattern,mis_marker);
108 time2=Esys_timer()-time0;
109 if (Esys_noError()) {
110 #pragma omp parallel for private(i) schedule(static)
111 for (i = 0; i < n; ++i) counter[i]=mis_marker[i];
112 out->n=n;
113 out->n_block=n_block;
114 out->n_F=Paso_Util_cumsum(n,counter);
115 out->mask_F=MEMALLOC(n,index_t);
116 out->rows_in_F=MEMALLOC(out->n_F,index_t);
117 out->inv_A_FF=MEMALLOC(n_block*n_block*out->n_F,double);
118 out->A_FF_pivot=NULL; /* later use for block size>3 */
119 if (! (Esys_checkPtr(out->mask_F) || Esys_checkPtr(out->inv_A_FF) || Esys_checkPtr(out->rows_in_F) ) ) {
120 #pragma omp parallel
121 {
122 /* creates an index for F from mask */
123 #pragma omp for private(i) schedule(static)
124 for (i = 0; i < out->n_F; ++i) out->rows_in_F[i]=-1;
125 #pragma omp for private(i) schedule(static)
126 for (i = 0; i < n; ++i) {
127 if (mis_marker[i]) {
128 out->rows_in_F[counter[i]]=i;
129 out->mask_F[i]=counter[i];
130 } else {
131 out->mask_F[i]=-1;
132 }
133 }
134 #pragma omp for private(i, where_p,iPtr,A11,A12,A13,A21,A22,A23,A31,A32,A33,D,index) schedule(static)
135 for (i = 0; i < out->n_F; i++) {
136 /* find main diagonal */
137 iPtr=A_p->pattern->ptr[out->rows_in_F[i]];
138 index=&(A_p->pattern->index[iPtr]);
139 where_p=(index_t*)bsearch(&out->rows_in_F[i],
140 index,
141 A_p->pattern->ptr[out->rows_in_F[i] + 1]-A_p->pattern->ptr[out->rows_in_F[i]],
142 sizeof(index_t),
143 Paso_comparIndex);
144 if (where_p==NULL) {
145 Esys_setError(VALUE_ERROR, "Paso_Solver_getRILU: main diagonal element missing.");
146 } else {
147 iPtr+=(index_t)(where_p-index);
148 /* get inverse of A_FF block: */
149 if (n_block==1) {
150 if (ABS(A_p->val[iPtr])>0.) {
151 out->inv_A_FF[i]=1./A_p->val[iPtr];
152 } else {
153 Esys_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU: Break-down in RILU decomposition: non-regular main diagonal block.");
154 }
155 } else if (n_block==2) {
156 A11=A_p->val[iPtr*4];
157 A21=A_p->val[iPtr*4+1];
158 A12=A_p->val[iPtr*4+2];
159 A22=A_p->val[iPtr*4+3];
160 D = A11*A22-A12*A21;
161 if (ABS(D) > 0 ){
162 D=1./D;
163 out->inv_A_FF[i*4]= A22*D;
164 out->inv_A_FF[i*4+1]=-A21*D;
165 out->inv_A_FF[i*4+2]=-A12*D;
166 out->inv_A_FF[i*4+3]= A11*D;
167 } else {
168 Esys_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU:Break-down in RILU decomposition: non-regular main diagonal block.");
169 }
170 } else if (n_block==3) {
171 A11=A_p->val[iPtr*9 ];
172 A21=A_p->val[iPtr*9+1];
173 A31=A_p->val[iPtr*9+2];
174 A12=A_p->val[iPtr*9+3];
175 A22=A_p->val[iPtr*9+4];
176 A32=A_p->val[iPtr*9+5];
177 A13=A_p->val[iPtr*9+6];
178 A23=A_p->val[iPtr*9+7];
179 A33=A_p->val[iPtr*9+8];
180 D = A11*(A22*A33-A23*A32)+ A12*(A31*A23-A21*A33)+A13*(A21*A32-A31*A22);
181 if (ABS(D) > 0 ){
182 D=1./D;
183 out->inv_A_FF[i*9 ]=(A22*A33-A23*A32)*D;
184 out->inv_A_FF[i*9+1]=(A31*A23-A21*A33)*D;
185 out->inv_A_FF[i*9+2]=(A21*A32-A31*A22)*D;
186 out->inv_A_FF[i*9+3]=(A13*A32-A12*A33)*D;
187 out->inv_A_FF[i*9+4]=(A11*A33-A31*A13)*D;
188 out->inv_A_FF[i*9+5]=(A12*A31-A11*A32)*D;
189 out->inv_A_FF[i*9+6]=(A12*A23-A13*A22)*D;
190 out->inv_A_FF[i*9+7]=(A13*A21-A11*A23)*D;
191 out->inv_A_FF[i*9+8]=(A11*A22-A12*A21)*D;
192 } else {
193 Esys_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU:Break-down in RILU decomposition: non-regular main diagonal block.");
194 }
195 }
196 }
197 }
198 } /* end parallel region */
199
200 if( Esys_noError()) {
201 /* if there are no nodes in the coarse level there is no more work to do */
202 out->n_C=n-out->n_F;
203 if (out->n_C>0) {
204 out->rows_in_C=MEMALLOC(out->n_C,index_t);
205 out->mask_C=MEMALLOC(n,index_t);
206 if (! (Esys_checkPtr(out->mask_C) || Esys_checkPtr(out->rows_in_C) ) ) {
207 /* creates an index for C from mask */
208 #pragma omp parallel for private(i) schedule(static)
209 for (i = 0; i < n; ++i) counter[i]=! mis_marker[i];
210 Paso_Util_cumsum(n,counter);
211 #pragma omp parallel
212 {
213 #pragma omp for private(i) schedule(static)
214 for (i = 0; i < out->n_C; ++i) out->rows_in_C[i]=-1;
215 #pragma omp for private(i) schedule(static)
216 for (i = 0; i < n; ++i) {
217 if (! mis_marker[i]) {
218 out->rows_in_C[counter[i]]=i;
219 out->mask_C[i]=counter[i];
220 } else {
221 out->mask_C[i]=-1;
222 }
223 }
224 } /* end parallel region */
225 /* get A_CF block: */
226 out->A_CF=Paso_SparseMatrix_getSubmatrix(A_p,out->n_C,out->n_F,out->rows_in_C,out->mask_F);
227 if (Esys_noError()) {
228 /* get A_FC block: */
229 out->A_FC=Paso_SparseMatrix_getSubmatrix(A_p,out->n_F,out->n_C,out->rows_in_F,out->mask_C);
230 /* get A_FF block: */
231 if (Esys_noError()) {
232 schur=Paso_SparseMatrix_getSubmatrix(A_p,out->n_C,out->n_C,out->rows_in_C,out->mask_C);
233 time0=Esys_timer()-time0;
234 if (Esys_noError()) {
235 time1=Esys_timer();
236 /* update A_CC block to get Schur complement and then apply RILU to it */
237 Paso_Solver_updateIncompleteSchurComplement(schur,out->A_CF,out->inv_A_FF,out->A_FF_pivot,out->A_FC);
238 time1=Esys_timer()-time1;
239 out->RILU_of_Schur=Paso_Solver_getRILU(schur,verbose);
240 Paso_SparseMatrix_free(schur);
241 }
242 /* allocate work arrays for RILU application */
243 if (Esys_noError()) {
244 out->x_F=MEMALLOC(n_block*out->n_F,double);
245 out->b_F=MEMALLOC(n_block*out->n_F,double);
246 out->x_C=MEMALLOC(n_block*out->n_C,double);
247 out->b_C=MEMALLOC(n_block*out->n_C,double);
248 if (! (Esys_checkPtr(out->x_F) || Esys_checkPtr(out->b_F) || Esys_checkPtr(out->x_C) || Esys_checkPtr(out->b_C) ) ) {
249 #pragma omp parallel
250 {
251 #pragma omp for private(i,k) schedule(static)
252 for (i = 0; i < out->n_F; ++i) {
253 for (k=0; k<n_block;++k) {
254 out->x_F[i*n_block+k]=0.;
255 out->b_F[i*n_block+k]=0.;
256 }
257 }
258 #pragma omp for private(i,k) schedule(static)
259 for (i = 0; i < out->n_C; ++i) {
260 for (k=0; k<n_block;++k) {
261 out->x_C[i*n_block+k]=0.;
262 out->b_C[i*n_block+k]=0.;
263 }
264 }
265 } /* end parallel region */
266 }
267 }
268 }
269 }
270 }
271 }
272 }
273 }
274 }
275 }
276 TMPMEMFREE(mis_marker);
277 TMPMEMFREE(counter);
278 if (Esys_noError()) {
279 /*
280 if (verbose) {
281 printf("RILU: %d unknowns eliminated. %d left.\n",out->n_F,n-out->n_F);
282 if (out->n_C>0) {
283 printf("timing: RILU: MIS/reordering/elemination : %e/%e/%e\n",time2,time0,time1);
284 } else {
285 printf("timing: RILU: MIS: %e\n",time2);
286 }
287 }
288 */
289 return out;
290 } else {
291 Paso_Solver_RILU_free(out);
292 return NULL;
293 }
294 }
295
296 /**************************************************************/
297
298 /* apply RILU precondition b-> x
299
300 in fact it solves
301
302 [ I 0 ] [ A_FF 0 ] [ I invA_FF*A_FF ] [ x_F ] = [b_F]
303 [ A_CF*invA_FF I ] [ 0 S ] [ 0 I ] [ x_C ] = [b_C]
304
305 in the form
306
307 b->[b_F,b_C]
308 x_F=invA_FF*b_F
309 b_C=b_C-A_CF*x_F
310 x_C=RILU(b_C)
311 b_F=b_F-A_FC*x_C
312 x_F=invA_FF*b_F
313 x<-[x_F,x_C]
314
315 should be called within a parallel region
316 barrier synconization should be performed to make sure that the input vector available
317
318 */
319
320 void Paso_Solver_solveRILU(Paso_Solver_RILU * rilu, double * x, double * b) {
321 dim_t i,k;
322 dim_t n_block=rilu->n_block;
323
324 if (rilu->n_C==0) {
325 /* x=invA_FF*b */
326 Paso_Copy(n_block*rilu->n_F, x, b);
327 Paso_BlockOps_allMV(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,x);
328 } else {
329 /* b->[b_F,b_C] */
330 if (n_block==1) {
331 #pragma omp parallel for private(i) schedule(static)
332 for (i=0;i<rilu->n_F;++i) rilu->b_F[i]=b[rilu->rows_in_F[i]];
333 #pragma omp parallel for private(i) schedule(static)
334 for (i=0;i<rilu->n_C;++i) rilu->b_C[i]=b[rilu->rows_in_C[i]];
335 } else {
336 #pragma omp parallel for private(i,k) schedule(static)
337 for (i=0;i<rilu->n_F;++i)
338 for (k=0;k<n_block;k++) rilu->b_F[rilu->n_block*i+k]=b[n_block*rilu->rows_in_F[i]+k];
339 #pragma omp parallel for private(i,k) schedule(static)
340 for (i=0;i<rilu->n_C;++i)
341 for (k=0;k<n_block;k++) rilu->b_C[rilu->n_block*i+k]=b[n_block*rilu->rows_in_C[i]+k];
342 }
343 /* x_F=invA_FF*b_F */
344 Paso_Copy(n_block*rilu->n_F, rilu->x_F,rilu->b_F);
345 Paso_BlockOps_allMV(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,rilu->x_F);
346 /* b_C=b_C-A_CF*x_F */
347 Paso_SparseMatrix_MatrixVector_CSR_OFFSET0(-1.,rilu->A_CF,rilu->x_F,1.,rilu->b_C);
348 /* x_C=RILU(b_C) */
349 Paso_Solver_solveRILU(rilu->RILU_of_Schur,rilu->x_C,rilu->b_C);
350 /* b_F=b_F-A_FC*x_C */
351 Paso_SparseMatrix_MatrixVector_CSR_OFFSET0(-1.,rilu->A_FC,rilu->x_C,1.,rilu->b_F);
352 /* x_F=invA_FF*b_F */
353 Paso_Copy(n_block*rilu->n_F, rilu->x_F,rilu->b_F);
354 Paso_BlockOps_allMV(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,rilu->x_F);
355 /* x<-[x_F,x_C] */
356 if (n_block==1) {
357 #pragma omp parallel for private(i) schedule(static)
358 for (i=0;i<rilu->n;++i) {
359 if (rilu->mask_C[i]>-1) {
360 x[i]=rilu->x_C[rilu->mask_C[i]];
361 } else {
362 x[i]=rilu->x_F[rilu->mask_F[i]];
363 }
364 }
365 } else {
366 #pragma omp parallel for private(i,k) schedule(static)
367 for (i=0;i<rilu->n;++i) {
368 if (rilu->mask_C[i]>-1) {
369 for (k=0;k<n_block;k++) x[n_block*i+k]=rilu->x_C[n_block*rilu->mask_C[i]+k];
370 } else {
371 for (k=0;k<n_block;k++) x[n_block*i+k]=rilu->x_F[n_block*rilu->mask_F[i]+k];
372 }
373 }
374 }
375 /* all done */
376 }
377 return;
378 }
379
380 /*
381 * $Log$
382 *
383 */

  ViewVC Help
Powered by ViewVC 1.1.26