/[escript]/trunk/paso/src/RILU.cpp
ViewVC logotype

Annotation of /trunk/paso/src/RILU.cpp

Parent Directory Parent Directory | Revision Log Revision Log


Revision 4803 - (hide annotations)
Wed Mar 26 06:52:28 2014 UTC (5 years, 10 months ago) by caltinay
File size: 16127 byte(s)
Removed obsolete wrappers for malloc and friends.
Paso_Pattern -> paso::Pattern

1 ksteube 1312
2 jfenwick 3981 /*****************************************************************************
3 ksteube 1811 *
4 jfenwick 4657 * Copyright (c) 2003-2014 by University of Queensland
5 jfenwick 3981 * http://www.uq.edu.au
6 ksteube 1811 *
7     * Primary Business: Queensland, Australia
8     * Licensed under the Open Software License version 3.0
9     * http://www.opensource.org/licenses/osl-3.0.php
10     *
11 jfenwick 3981 * Development until 2012 by Earth Systems Science Computational Center (ESSCC)
12 jfenwick 4657 * Development 2012-2013 by School of Earth Sciences
13     * Development from 2014 by Centre for Geoscience Computing (GeoComp)
14 jfenwick 3981 *
15     *****************************************************************************/
16 dhawcroft 631
17 ksteube 1811
18 jfenwick 3981 /************************************************************************************/
19 gross 430
20     /* Paso: RILU preconditioner with reordering */
21    
22 jfenwick 3981 /************************************************************************************/
23 gross 430
24 caltinay 3642 /* Author: Lutz Gross, l.gross@uq.edu.au */
25 gross 430
26 jfenwick 3981 /************************************************************************************/
27 gross 430
28 gross 700 #include "Paso.h"
29 gross 3120 #include "Preconditioner.h"
30 gross 700 #include "PasoUtil.h"
31 gross 3120 #include "BlockOps.h"
32 gross 430
33 jfenwick 3981 /************************************************************************************/
34 gross 430
35 caltinay 3642 /* free all memory used by RILU */
36 gross 430
37     void Paso_Solver_RILU_free(Paso_Solver_RILU * in) {
38     if (in!=NULL) {
39     Paso_Solver_RILU_free(in->RILU_of_Schur);
40 jfenwick 4280 delete[] in->inv_A_FF;
41     delete[] in->A_FF_pivot;
42 caltinay 4797 paso::SparseMatrix_free(in->A_FC);
43     paso::SparseMatrix_free(in->A_CF);
44 jfenwick 4280 delete[] in->rows_in_F;
45     delete[] in->rows_in_C;
46     delete[] in->mask_F;
47     delete[] in->mask_C;
48     delete[] in->x_F;
49     delete[] in->b_F;
50     delete[] in->x_C;
51     delete[] in->b_C;
52     delete in;
53 gross 430 }
54     }
55    
56 jfenwick 3981 /************************************************************************************/
57 gross 430
58     /* constructs the block-block factorization of
59    
60     [ A_FF A_FC ]
61     A_p=
62     [ A_CF A_FF ]
63    
64     to
65    
66     [ I 0 ] [ A_FF 0 ] [ I invA_FF*A_FF ]
67     [ A_CF*invA_FF I ] [ 0 S ] [ 0 I ]
68    
69    
70     where S=A_FF-ACF*invA_FF*A_FC within the shape of S
71    
72     then RILU is applied to S again until S becomes empty
73    
74     */
75 caltinay 4797 Paso_Solver_RILU* Paso_Solver_getRILU(paso::SparseMatrix *A_p,bool verbose) {
76 gross 430 Paso_Solver_RILU* out=NULL;
77 ksteube 1312 dim_t n=A_p->numRows;
78 gross 430 dim_t n_block=A_p->row_block_size;
79     index_t* mis_marker=NULL;
80     index_t* counter=NULL;
81     index_t iPtr,*index, *where_p;
82     dim_t i,k;
83 caltinay 4797 paso::SparseMatrix * schur=NULL;
84 caltinay 3489 double A11,A12,A13,A21,A22,A23,A31,A32,A33,D,time0=0,time1=0;/*,time2=0;*/
85 gross 430
86    
87 jfenwick 4280 mis_marker=new index_t[n];
88     counter=new index_t[n];
89     out=new Paso_Solver_RILU;
90 gross 430 out->RILU_of_Schur=NULL;
91     out->inv_A_FF=NULL;
92     out->A_FF_pivot=NULL;
93     out->A_FC=NULL;
94     out->A_CF=NULL;
95     out->rows_in_F=NULL;
96     out->rows_in_C=NULL;
97     out->mask_F=NULL;
98     out->mask_C=NULL;
99     out->x_F=NULL;
100     out->b_F=NULL;
101     out->x_C=NULL;
102     out->b_C=NULL;
103    
104 jfenwick 3259 if ( !(Esys_checkPtr(mis_marker) || Esys_checkPtr(out) || Esys_checkPtr(counter) ) ) {
105 caltinay 3642 /* identify independent set of rows/columns */
106 jfenwick 3259 time0=Esys_timer();
107 gross 431 #pragma omp parallel for private(i) schedule(static)
108     for (i=0;i<n;++i) mis_marker[i]=-1;
109 caltinay 4803 paso::Pattern_mis(A_p->pattern,mis_marker);
110 caltinay 3489 /*time2=Esys_timer()-time0;*/
111 jfenwick 3259 if (Esys_noError()) {
112 gross 430 #pragma omp parallel for private(i) schedule(static)
113     for (i = 0; i < n; ++i) counter[i]=mis_marker[i];
114     out->n=n;
115     out->n_block=n_block;
116     out->n_F=Paso_Util_cumsum(n,counter);
117 jfenwick 4280 out->mask_F=new index_t[n];
118     out->rows_in_F=new index_t[out->n_F];
119     out->inv_A_FF=new double[n_block*n_block*out->n_F];
120 gross 430 out->A_FF_pivot=NULL; /* later use for block size>3 */
121 jfenwick 3259 if (! (Esys_checkPtr(out->mask_F) || Esys_checkPtr(out->inv_A_FF) || Esys_checkPtr(out->rows_in_F) ) ) {
122 gross 430 #pragma omp parallel
123     {
124     /* creates an index for F from mask */
125     #pragma omp for private(i) schedule(static)
126     for (i = 0; i < out->n_F; ++i) out->rows_in_F[i]=-1;
127     #pragma omp for private(i) schedule(static)
128     for (i = 0; i < n; ++i) {
129     if (mis_marker[i]) {
130     out->rows_in_F[counter[i]]=i;
131     out->mask_F[i]=counter[i];
132     } else {
133     out->mask_F[i]=-1;
134     }
135     }
136     #pragma omp for private(i, where_p,iPtr,A11,A12,A13,A21,A22,A23,A31,A32,A33,D,index) schedule(static)
137     for (i = 0; i < out->n_F; i++) {
138     /* find main diagonal */
139     iPtr=A_p->pattern->ptr[out->rows_in_F[i]];
140     index=&(A_p->pattern->index[iPtr]);
141     where_p=(index_t*)bsearch(&out->rows_in_F[i],
142     index,
143     A_p->pattern->ptr[out->rows_in_F[i] + 1]-A_p->pattern->ptr[out->rows_in_F[i]],
144     sizeof(index_t),
145 caltinay 4803 paso::comparIndex);
146 gross 430 if (where_p==NULL) {
147 jfenwick 3259 Esys_setError(VALUE_ERROR, "Paso_Solver_getRILU: main diagonal element missing.");
148 gross 430 } else {
149     iPtr+=(index_t)(where_p-index);
150     /* get inverse of A_FF block: */
151     if (n_block==1) {
152     if (ABS(A_p->val[iPtr])>0.) {
153     out->inv_A_FF[i]=1./A_p->val[iPtr];
154     } else {
155 jfenwick 3259 Esys_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU: Break-down in RILU decomposition: non-regular main diagonal block.");
156 gross 430 }
157     } else if (n_block==2) {
158     A11=A_p->val[iPtr*4];
159     A21=A_p->val[iPtr*4+1];
160     A12=A_p->val[iPtr*4+2];
161     A22=A_p->val[iPtr*4+3];
162     D = A11*A22-A12*A21;
163     if (ABS(D) > 0 ){
164     D=1./D;
165     out->inv_A_FF[i*4]= A22*D;
166     out->inv_A_FF[i*4+1]=-A21*D;
167     out->inv_A_FF[i*4+2]=-A12*D;
168     out->inv_A_FF[i*4+3]= A11*D;
169     } else {
170 caltinay 3642 Esys_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU: Break-down in RILU decomposition: non-regular main diagonal block.");
171 gross 430 }
172     } else if (n_block==3) {
173     A11=A_p->val[iPtr*9 ];
174     A21=A_p->val[iPtr*9+1];
175     A31=A_p->val[iPtr*9+2];
176     A12=A_p->val[iPtr*9+3];
177     A22=A_p->val[iPtr*9+4];
178     A32=A_p->val[iPtr*9+5];
179     A13=A_p->val[iPtr*9+6];
180     A23=A_p->val[iPtr*9+7];
181     A33=A_p->val[iPtr*9+8];
182     D = A11*(A22*A33-A23*A32)+ A12*(A31*A23-A21*A33)+A13*(A21*A32-A31*A22);
183     if (ABS(D) > 0 ){
184     D=1./D;
185     out->inv_A_FF[i*9 ]=(A22*A33-A23*A32)*D;
186     out->inv_A_FF[i*9+1]=(A31*A23-A21*A33)*D;
187     out->inv_A_FF[i*9+2]=(A21*A32-A31*A22)*D;
188     out->inv_A_FF[i*9+3]=(A13*A32-A12*A33)*D;
189     out->inv_A_FF[i*9+4]=(A11*A33-A31*A13)*D;
190     out->inv_A_FF[i*9+5]=(A12*A31-A11*A32)*D;
191     out->inv_A_FF[i*9+6]=(A12*A23-A13*A22)*D;
192     out->inv_A_FF[i*9+7]=(A13*A21-A11*A23)*D;
193     out->inv_A_FF[i*9+8]=(A11*A22-A12*A21)*D;
194     } else {
195 caltinay 3642 Esys_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU: Break-down in RILU decomposition: non-regular main diagonal block.");
196 gross 430 }
197     }
198     }
199     }
200     } /* end parallel region */
201    
202 jfenwick 3259 if( Esys_noError()) {
203 gross 430 /* if there are no nodes in the coarse level there is no more work to do */
204     out->n_C=n-out->n_F;
205     if (out->n_C>0) {
206 jfenwick 4280 out->rows_in_C=new index_t[out->n_C];
207     out->mask_C=new index_t[n];
208 jfenwick 3259 if (! (Esys_checkPtr(out->mask_C) || Esys_checkPtr(out->rows_in_C) ) ) {
209 gross 430 /* creates an index for C from mask */
210     #pragma omp parallel for private(i) schedule(static)
211     for (i = 0; i < n; ++i) counter[i]=! mis_marker[i];
212     Paso_Util_cumsum(n,counter);
213     #pragma omp parallel
214     {
215     #pragma omp for private(i) schedule(static)
216     for (i = 0; i < out->n_C; ++i) out->rows_in_C[i]=-1;
217     #pragma omp for private(i) schedule(static)
218     for (i = 0; i < n; ++i) {
219     if (! mis_marker[i]) {
220     out->rows_in_C[counter[i]]=i;
221     out->mask_C[i]=counter[i];
222     } else {
223     out->mask_C[i]=-1;
224     }
225     }
226     } /* end parallel region */
227     /* get A_CF block: */
228 caltinay 4797 out->A_CF=paso::SparseMatrix_getSubmatrix(A_p,out->n_C,out->n_F,out->rows_in_C,out->mask_F);
229 jfenwick 3259 if (Esys_noError()) {
230 gross 430 /* get A_FC block: */
231 caltinay 4797 out->A_FC=paso::SparseMatrix_getSubmatrix(A_p,out->n_F,out->n_C,out->rows_in_F,out->mask_C);
232 gross 430 /* get A_FF block: */
233 jfenwick 3259 if (Esys_noError()) {
234 caltinay 4797 schur=paso::SparseMatrix_getSubmatrix(A_p,out->n_C,out->n_C,out->rows_in_C,out->mask_C);
235 jfenwick 3259 time0=Esys_timer()-time0;
236     if (Esys_noError()) {
237     time1=Esys_timer();
238 gross 430 /* update A_CC block to get Schur complement and then apply RILU to it */
239     Paso_Solver_updateIncompleteSchurComplement(schur,out->A_CF,out->inv_A_FF,out->A_FF_pivot,out->A_FC);
240 jfenwick 3259 time1=Esys_timer()-time1;
241 gross 430 out->RILU_of_Schur=Paso_Solver_getRILU(schur,verbose);
242 caltinay 4797 paso::SparseMatrix_free(schur);
243 gross 430 }
244     /* allocate work arrays for RILU application */
245 jfenwick 3259 if (Esys_noError()) {
246 jfenwick 4280 out->x_F=new double[n_block*out->n_F];
247     out->b_F=new double[n_block*out->n_F];
248     out->x_C=new double[n_block*out->n_C];
249     out->b_C=new double[n_block*out->n_C];
250 jfenwick 3259 if (! (Esys_checkPtr(out->x_F) || Esys_checkPtr(out->b_F) || Esys_checkPtr(out->x_C) || Esys_checkPtr(out->b_C) ) ) {
251 gross 430 #pragma omp parallel
252     {
253     #pragma omp for private(i,k) schedule(static)
254     for (i = 0; i < out->n_F; ++i) {
255     for (k=0; k<n_block;++k) {
256     out->x_F[i*n_block+k]=0.;
257     out->b_F[i*n_block+k]=0.;
258     }
259     }
260     #pragma omp for private(i,k) schedule(static)
261     for (i = 0; i < out->n_C; ++i) {
262     for (k=0; k<n_block;++k) {
263     out->x_C[i*n_block+k]=0.;
264     out->b_C[i*n_block+k]=0.;
265     }
266     }
267     } /* end parallel region */
268     }
269     }
270     }
271     }
272     }
273     }
274     }
275     }
276     }
277     }
278 jfenwick 4280 delete[] mis_marker;
279     delete[] counter;
280 jfenwick 3259 if (Esys_noError()) {
281 gross 2100 /*
282 gross 430 if (verbose) {
283     printf("RILU: %d unknowns eliminated. %d left.\n",out->n_F,n-out->n_F);
284     if (out->n_C>0) {
285 caltinay 3642 printf("timing: RILU: MIS/reordering/elimination : %e/%e/%e\n",time2,time0,time1);
286 gross 430 } else {
287     printf("timing: RILU: MIS: %e\n",time2);
288     }
289     }
290 gross 2100 */
291 gross 430 return out;
292     } else {
293     Paso_Solver_RILU_free(out);
294     return NULL;
295     }
296     }
297    
298 jfenwick 3981 /************************************************************************************/
299 gross 430
300     /* apply RILU precondition b-> x
301    
302     in fact it solves
303    
304     [ I 0 ] [ A_FF 0 ] [ I invA_FF*A_FF ] [ x_F ] = [b_F]
305     [ A_CF*invA_FF I ] [ 0 S ] [ 0 I ] [ x_C ] = [b_C]
306    
307     in the form
308    
309     b->[b_F,b_C]
310     x_F=invA_FF*b_F
311     b_C=b_C-A_CF*x_F
312     x_C=RILU(b_C)
313     b_F=b_F-A_FC*x_C
314     x_F=invA_FF*b_F
315     x<-[x_F,x_C]
316    
317 caltinay 3642 Should be called within a parallel region.
318     Barrier synchronization should be performed to make sure that the input
319     vector is available.
320 gross 430 */
321    
322     void Paso_Solver_solveRILU(Paso_Solver_RILU * rilu, double * x, double * b) {
323     dim_t i,k;
324     dim_t n_block=rilu->n_block;
325    
326     if (rilu->n_C==0) {
327     /* x=invA_FF*b */
328 gross 3160 Paso_Copy(n_block*rilu->n_F, x, b);
329 gross 3402 Paso_BlockOps_solveAll(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,x);
330 gross 430 } else {
331     /* b->[b_F,b_C] */
332     if (n_block==1) {
333 gross 1556 #pragma omp parallel for private(i) schedule(static)
334 gross 430 for (i=0;i<rilu->n_F;++i) rilu->b_F[i]=b[rilu->rows_in_F[i]];
335 gross 1556 #pragma omp parallel for private(i) schedule(static)
336 gross 430 for (i=0;i<rilu->n_C;++i) rilu->b_C[i]=b[rilu->rows_in_C[i]];
337     } else {
338 gross 1556 #pragma omp parallel for private(i,k) schedule(static)
339 gross 430 for (i=0;i<rilu->n_F;++i)
340     for (k=0;k<n_block;k++) rilu->b_F[rilu->n_block*i+k]=b[n_block*rilu->rows_in_F[i]+k];
341 gross 1556 #pragma omp parallel for private(i,k) schedule(static)
342 gross 430 for (i=0;i<rilu->n_C;++i)
343     for (k=0;k<n_block;k++) rilu->b_C[rilu->n_block*i+k]=b[n_block*rilu->rows_in_C[i]+k];
344     }
345     /* x_F=invA_FF*b_F */
346 gross 3158 Paso_Copy(n_block*rilu->n_F, rilu->x_F,rilu->b_F);
347 gross 3402 Paso_BlockOps_solveAll(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,rilu->x_F);
348 gross 430 /* b_C=b_C-A_CF*x_F */
349 caltinay 4795 paso::SparseMatrix_MatrixVector_CSR_OFFSET0(-1.,rilu->A_CF,rilu->x_F,1.,rilu->b_C);
350 gross 430 /* x_C=RILU(b_C) */
351     Paso_Solver_solveRILU(rilu->RILU_of_Schur,rilu->x_C,rilu->b_C);
352     /* b_F=b_F-A_FC*x_C */
353 caltinay 4795 paso::SparseMatrix_MatrixVector_CSR_OFFSET0(-1.,rilu->A_FC,rilu->x_C,1.,rilu->b_F);
354 gross 430 /* x_F=invA_FF*b_F */
355 gross 3158 Paso_Copy(n_block*rilu->n_F, rilu->x_F,rilu->b_F);
356 gross 3402 Paso_BlockOps_solveAll(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,rilu->x_F);
357 gross 430 /* x<-[x_F,x_C] */
358     if (n_block==1) {
359 gross 1556 #pragma omp parallel for private(i) schedule(static)
360 gross 430 for (i=0;i<rilu->n;++i) {
361     if (rilu->mask_C[i]>-1) {
362     x[i]=rilu->x_C[rilu->mask_C[i]];
363     } else {
364     x[i]=rilu->x_F[rilu->mask_F[i]];
365     }
366     }
367     } else {
368 gross 1556 #pragma omp parallel for private(i,k) schedule(static)
369 gross 430 for (i=0;i<rilu->n;++i) {
370     if (rilu->mask_C[i]>-1) {
371     for (k=0;k<n_block;k++) x[n_block*i+k]=rilu->x_C[n_block*rilu->mask_C[i]+k];
372     } else {
373     for (k=0;k<n_block;k++) x[n_block*i+k]=rilu->x_F[n_block*rilu->mask_F[i]+k];
374     }
375     }
376     }
377     /* all done */
378     }
379     return;
380     }
381    
382     /*
383     * $Log$
384     *
385     */

Properties

Name Value
svn:mergeinfo /branches/amg_from_3530/paso/src/RILU.cpp:3531-3826 /branches/lapack2681/paso/src/RILU.cpp:2682-2741 /branches/pasowrap/paso/src/RILU.cpp:3661-3674 /branches/py3_attempt2/paso/src/RILU.cpp:3871-3891 /branches/restext/paso/src/RILU.cpp:2610-2624 /branches/ripleygmg_from_3668/paso/src/RILU.cpp:3669-3791 /branches/stage3.0/paso/src/RILU.cpp:2569-2590 /branches/symbolic_from_3470/paso/src/RILU.cpp:3471-3974 /branches/symbolic_from_3470/ripley/test/python/paso/src/RILU.cpp:3517-3974 /release/3.0/paso/src/RILU.cpp:2591-2601 /trunk/paso/src/RILU.cpp:4257-4344 /trunk/ripley/test/python/paso/src/RILU.cpp:3480-3515

  ViewVC Help
Powered by ViewVC 1.1.26