/[escript]/trunk/paso/src/Smoother.c
ViewVC logotype

Annotation of /trunk/paso/src/Smoother.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 3105 - (hide annotations)
Wed Aug 25 04:37:52 2010 UTC (8 years, 6 months ago) by gross
Original Path: trunk/paso/src/Solver_GS.c
File MIME type: text/plain
File size: 14182 byte(s)
bug fixed in the Gauss Seidel
1 jfenwick 1851
2     /*******************************************************
3     *
4 jfenwick 2881 * Copyright (c) 2003-2010 by University of Queensland
5 jfenwick 1851 * 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 gross 3097 /* Paso: Gauss-Seidel */
18 jfenwick 1851
19     /**************************************************************/
20    
21     /* Author: artak@uq.edu.au */
22    
23     /**************************************************************/
24    
25     #include "Paso.h"
26     #include "Solver.h"
27     #include "PasoUtil.h"
28 gross 3097 #include "BlockOps.h"
29 jfenwick 1851
30     #include <stdio.h>
31    
32 gross 3094
33 jfenwick 1851 /**************************************************************/
34    
35     /* free all memory used by GS */
36    
37     void Paso_Solver_GS_free(Paso_Solver_GS * in) {
38     if (in!=NULL) {
39 gross 3094 Paso_Solver_LocalGS_free(in->localGS);
40 jfenwick 1851 MEMFREE(in);
41     }
42     }
43 gross 3094 void Paso_Solver_LocalGS_free(Paso_Solver_LocalGS * in) {
44     if (in!=NULL) {
45     MEMFREE(in->diag);
46     MEMFREE(in->pivot);
47 gross 3097 MEMFREE(in->buffer);
48 gross 3094 MEMFREE(in);
49     }
50     }
51 jfenwick 1851 /**************************************************************/
52    
53 gross 3094 /* constructs the symmetric Gauss-Seidel preconditioner
54 jfenwick 1851
55     */
56 gross 3094 Paso_Solver_GS* Paso_Solver_getGS(Paso_SystemMatrix * A_p, dim_t sweeps, bool_t is_local, bool_t verbose)
57     {
58    
59 jfenwick 1851 /* allocations: */
60     Paso_Solver_GS* out=MEMALLOC(1,Paso_Solver_GS);
61 gross 3094 if (! Paso_checkPtr(out)) {
62     out->localGS=Paso_Solver_getLocalGS(A_p->mainBlock,sweeps,verbose);
63     out->is_local=is_local;
64 jfenwick 1851 }
65 gross 3094 if (Paso_MPIInfo_noError(A_p->mpi_info)) {
66 jfenwick 1851 return out;
67 gross 3094 } else {
68 jfenwick 1851 Paso_Solver_GS_free(out);
69     return NULL;
70     }
71     }
72 gross 3094 Paso_Solver_LocalGS* Paso_Solver_getLocalGS(Paso_SparseMatrix * A_p, dim_t sweeps, bool_t verbose) {
73    
74     dim_t n=A_p->numRows;
75     dim_t n_block=A_p->row_block_size;
76     dim_t block_size=A_p->block_size;
77    
78     double time0=Paso_timer();
79     /* allocations: */
80     Paso_Solver_LocalGS* out=MEMALLOC(1,Paso_Solver_LocalGS);
81     if (! Paso_checkPtr(out)) {
82    
83     out->diag=MEMALLOC( ((size_t) n) * ((size_t) block_size),double);
84     out->pivot=MEMALLOC( ((size_t) n) * ((size_t) n_block), index_t);
85 gross 3097 out->buffer=MEMALLOC( ((size_t) n) * ((size_t) n_block), double);
86 gross 3094 out->sweeps=sweeps;
87    
88     if ( ! ( Paso_checkPtr(out->diag) || Paso_checkPtr(out->pivot) ) ) {
89     Paso_SparseMatrix_invMain(A_p, out->diag, out->pivot );
90     }
91    
92     }
93     time0=Paso_timer()-time0;
94    
95     if (Paso_noError()) {
96     if (verbose) printf("timing: Gauss-Seidel preparation: elemination : %e\n",time0);
97     return out;
98     } else {
99     Paso_Solver_LocalGS_free(out);
100     return NULL;
101     }
102     }
103 jfenwick 1851
104 gross 3097 /*
105 jfenwick 1851
106 gross 3097 performs a few sweeps of the from
107 jfenwick 1851
108 gross 3097 S (x_{k} - x_{k-1}) = b - A x_{k-1}
109 jfenwick 1851
110 gross 3097 where x_{0}=0 and S provides some approximatioon of A.
111 jfenwick 1851
112 gross 3097 Under MPI S is build on using A_p->mainBlock only.
113     if GS is local the defect b - A x_{k-1} is calculated using A_p->mainBlock only.
114 gross 3094
115 jfenwick 1851 */
116    
117 gross 3094 void Paso_Solver_solveGS(Paso_SystemMatrix* A_p, Paso_Solver_GS * gs, double * x, const double * b)
118     {
119     register dim_t i;
120 gross 3097 const dim_t n= (A_p->mainBlock->numRows) * (A_p->mainBlock->row_block_size);
121    
122     double *b_new = gs->localGS->buffer;
123 gross 3094 dim_t sweeps=gs->localGS->sweeps;
124    
125 gross 3097 if (gs->is_local) {
126     Paso_Solver_solveLocalGS(A_p->mainBlock,gs->localGS,x,b);
127     } else {
128     #pragma omp parallel for private(i) schedule(static)
129     for (i=0;i<n;++i) x[i]=b[i];
130    
131     Paso_Solver_localGSSweep(A_p->mainBlock,gs->localGS,x);
132 gross 3105
133     while (sweeps > 1 ) {
134 gross 3097 #pragma omp parallel for private(i) schedule(static)
135     for (i=0;i<n;++i) b_new[i]=b[i];
136    
137     Paso_SystemMatrix_MatrixVector((-1.), A_p, x, 1., b_new); /* b_new = b - A*x */
138    
139     Paso_Solver_localGSSweep(A_p->mainBlock,gs->localGS,b_new);
140    
141     #pragma omp parallel for private(i) schedule(static)
142 gross 3105 for (i=0;i<n;++i) x[i]+=b_new[i];
143 gross 3097 sweeps--;
144     }
145    
146 gross 3094 }
147 gross 3097 }
148     void Paso_Solver_solveLocalGS(Paso_SparseMatrix* A_p, Paso_Solver_LocalGS * gs, double * x, const double * b)
149     {
150     register dim_t i;
151     const dim_t n= (A_p->numRows) * (A_p->row_block_size);
152     double *b_new = gs->buffer;
153     dim_t sweeps=gs->sweeps;
154 gross 3094
155 gross 3097 #pragma omp parallel for private(i) schedule(static)
156     for (i=0;i<n;++i) x[i]=b[i];
157 gross 3105
158 gross 3097 Paso_Solver_localGSSweep(A_p,gs,x);
159 gross 3094
160 gross 3105 while (sweeps > 1 ) {
161 gross 3094 #pragma omp parallel for private(i) schedule(static)
162 gross 3097 for (i=0;i<n;++i) b_new[i]=b[i];
163 gross 3094
164 gross 3097 Paso_SparseMatrix_MatrixVector_CSC_OFFSET0((-1.), A_p, x, 1., b_new); /* b_new = b - A*x */
165    
166     Paso_Solver_localGSSweep(A_p,gs,b_new);
167    
168     #pragma omp parallel for private(i) schedule(static)
169     for (i=0;i<n;++i) x[i]+=b_new[i];
170    
171     sweeps--;
172 gross 3094 }
173     }
174 jfenwick 1851
175 gross 3097 void Paso_Solver_localGSSweep(Paso_SparseMatrix* A, Paso_Solver_LocalGS * gs, double * x)
176 gross 3094 {
177     #ifdef _OPENMP
178     const dim_t nt=omp_get_max_threads();
179     #else
180     const dim_t nt = 1;
181     #endif
182 gross 3098 if (nt < 2) {
183 gross 3097 Paso_Solver_localGSSweep_sequential(A,gs,x);
184     } else {
185     Paso_Solver_localGSSweep_colored(A,gs,x);
186 gross 3094 }
187 jfenwick 1851 }
188    
189 gross 3097 /* inplace Gauss-Seidel sweep in seqential mode: */
190 gross 3094
191 gross 3097 void Paso_Solver_localGSSweep_sequential(Paso_SparseMatrix* A_p, Paso_Solver_LocalGS * gs, double * x)
192     {
193     const dim_t n=A_p->numRows;
194     const dim_t n_block=A_p->row_block_size;
195     const double *diag = gs->diag;
196     /* const index_t* pivot = gs->pivot;
197     const dim_t block_size=A_p->block_size; use for block size >3*/
198 gross 3094
199 gross 3097 register dim_t i,k;
200     register index_t iptr_ik, mm;
201    
202     const index_t* ptr_main = Paso_SparseMatrix_borrowMainDiagonalPointer(A_p);
203     /* forward substitution */
204    
205     if (n_block==1) {
206 gross 3098 Paso_BlockOps_MV_1(&x[0], &diag[0], &x[0]);
207     for (i = 1; i < n; ++i) {
208 gross 3097 mm=ptr_main[i];
209 gross 3105 /* x_i=x_i-a_ik*x_k (with k<i) */
210 gross 3097 for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<mm; ++iptr_ik) {
211 gross 3098 k=A_p->pattern->index[iptr_ik];
212 gross 3097 Paso_BlockOps_SMV_1(&x[i], &A_p->val[iptr_ik], &x[k]);
213     }
214 gross 3098 Paso_BlockOps_MV_1(&x[i], &diag[i], &x[i]);
215 gross 3097 }
216     } else if (n_block==2) {
217 gross 3098 Paso_BlockOps_MV_2(&x[0], &diag[0], &x[0]);
218     for (i = 1; i < n; ++i) {
219 gross 3097 mm=ptr_main[i];
220     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<mm; ++iptr_ik) {
221     k=A_p->pattern->index[iptr_ik];
222     Paso_BlockOps_SMV_2(&x[2*i], &A_p->val[4*iptr_ik], &x[2*k]);
223     }
224 gross 3098 Paso_BlockOps_MV_2(&x[2*i], &diag[4*i], &x[2*i]);
225 gross 3097 }
226     } else if (n_block==3) {
227 gross 3098 Paso_BlockOps_MV_3(&x[0], &diag[0], &x[0]);
228     for (i = 1; i < n; ++i) {
229 gross 3097 mm=ptr_main[i];
230     /* x_i=x_i-a_ik*x_k */
231     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<mm; ++iptr_ik) {
232     k=A_p->pattern->index[iptr_ik];
233     Paso_BlockOps_SMV_3(&x[3*i], &A_p->val[9*iptr_ik], &x[3*k]);
234     }
235 gross 3098 Paso_BlockOps_MV_3(&x[3*i], &diag[9*i], &x[3*i]);
236 gross 3097 }
237     } /* add block size >3 */
238 gross 3105
239 gross 3097 /* backward substitution */
240    
241     if (n_block==1) {
242 gross 3105 for (i = n-2; i > -1; --i) {
243 gross 3098 mm=ptr_main[i];
244     Paso_BlockOps_MV_1(&x[i], &A_p->val[mm], &x[i]);
245     for (iptr_ik=mm+1; iptr_ik < A_p->pattern->ptr[i+1]; ++iptr_ik) {
246     k=A_p->pattern->index[iptr_ik];
247     Paso_BlockOps_SMV_1(&x[i], &A_p->val[iptr_ik], &x[k]);
248     }
249     Paso_BlockOps_MV_1(&x[i], &diag[i], &x[i]);
250 gross 3097 }
251 gross 3098
252 gross 3097 } else if (n_block==2) {
253 gross 3105 for (i = n-2; i > -1; --i) {
254 gross 3098 mm=ptr_main[i];
255     Paso_BlockOps_MV_2(&x[2*i], &A_p->val[4*mm], &x[2*i]);
256     for (iptr_ik=mm+1; iptr_ik < A_p->pattern->ptr[i+1]; ++iptr_ik) {
257     k=A_p->pattern->index[iptr_ik];
258     Paso_BlockOps_SMV_2(&x[2*i], &A_p->val[4*iptr_ik], &x[2*k]);
259     }
260     Paso_BlockOps_MV_2(&x[2*i], &diag[i*4], &x[2*i]);
261 gross 3097 }
262     } else if (n_block==3) {
263 gross 3105 for (i = n-2; i > -1; --i) {
264 gross 3098 mm=ptr_main[i];
265     Paso_BlockOps_MV_3(&x[3*i], &A_p->val[9*mm], &x[3*i]);
266 gross 3097
267 gross 3098 for (iptr_ik=mm+1; iptr_ik < A_p->pattern->ptr[i+1]; ++iptr_ik) {
268     k=A_p->pattern->index[iptr_ik];
269     Paso_BlockOps_SMV_3(&x[3*i], &A_p->val[9*iptr_ik], &x[3*k]);
270     }
271     Paso_BlockOps_MV_3(&x[3*i], &diag[i*9], &x[3*i]);
272 gross 3097 }
273    
274     } /* add block size >3 */
275    
276     return;
277     }
278 gross 3094
279 gross 3097 void Paso_Solver_localGSSweep_colored(Paso_SparseMatrix* A_p, Paso_Solver_LocalGS * gs, double * x)
280 gross 3094 {
281     const dim_t n=A_p->numRows;
282     const dim_t n_block=A_p->row_block_size;
283     const double *diag = gs->diag;
284 gross 3097 index_t* pivot = gs->pivot;
285     const dim_t block_size=A_p->block_size;
286 gross 3094
287     register dim_t i,k;
288     register index_t color,iptr_ik, mm;
289    
290     const index_t* coloring = Paso_Pattern_borrowColoringPointer(A_p->pattern);
291     const dim_t num_colors = Paso_Pattern_getNumColors(A_p->pattern);
292     const index_t* ptr_main = Paso_SparseMatrix_borrowMainDiagonalPointer(A_p);
293    
294     /* forward substitution */
295 gross 3105
296 gross 3094
297     /* color = 0 */
298     if (n_block==1) {
299 gross 3097 #pragma omp parallel for schedule(static) private(i)
300 gross 3094 for (i = 0; i < n; ++i) {
301 gross 3097 if (coloring[i]== 0 ) Paso_BlockOps_MV_1(&x[i], &diag[i], &x[i]);
302     }
303 gross 3094 } else if (n_block==2) {
304 gross 3097 #pragma omp parallel for schedule(static) private(i)
305 gross 3094 for (i = 0; i < n; ++i) {
306 gross 3097 if (coloring[i]== 0 ) Paso_BlockOps_MV_2(&x[2*i], &diag[i*4], &x[2*i]);
307 gross 3094 }
308 gross 3097 } else if (n_block==3) {
309     #pragma omp parallel for schedule(static) private(i)
310 gross 3094 for (i = 0; i < n; ++i) {
311 gross 3097 if (coloring[i]==0) Paso_BlockOps_MV_3(&x[3*i], &diag[i*9], &x[3*i]);
312 gross 3094 }
313 gross 3097 } else {
314     #pragma omp parallel for schedule(static) private(i)
315     for (i = 0; i < n; ++i) {
316     if (coloring[i]==0) Paso_BlockOps_Solve_N(n_block, &x[n_block*i], &diag[block_size*i], &pivot[n_block*i], &x[n_block*i]);
317     }
318     }
319 gross 3094
320     for (color=1;color<num_colors;++color) {
321 gross 3105
322 gross 3094 if (n_block==1) {
323 gross 3097 #pragma omp parallel for schedule(static) private(i,iptr_ik,k)
324 gross 3094 for (i = 0; i < n; ++i) {
325     if (coloring[i]==color) {
326     /* x_i=x_i-a_ik*x_k */
327     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
328 gross 3105 k=A_p->pattern->index[iptr_ik];
329 gross 3097 if (coloring[k]<color) Paso_BlockOps_SMV_1(&x[i], &A_p->val[iptr_ik], &x[k]);
330 gross 3094 }
331 gross 3097 Paso_BlockOps_MV_1(&x[i], &diag[i], &x[i]);
332 gross 3094 }
333     }
334     } else if (n_block==2) {
335 gross 3097 #pragma omp parallel for schedule(static) private(i,iptr_ik,k)
336 gross 3094 for (i = 0; i < n; ++i) {
337     if (coloring[i]==color) {
338     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
339     k=A_p->pattern->index[iptr_ik];
340 gross 3097 if (coloring[k]<color) Paso_BlockOps_SMV_2(&x[2*i], &A_p->val[4*iptr_ik], &x[2*k]);
341 gross 3094 }
342 gross 3097 Paso_BlockOps_MV_2(&x[2*i], &diag[4*i], &x[2*i]);
343 gross 3094 }
344     }
345     } else if (n_block==3) {
346 gross 3097 #pragma omp parallel for schedule(static) private(i,iptr_ik,k)
347 gross 3094 for (i = 0; i < n; ++i) {
348     if (coloring[i]==color) {
349     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
350     k=A_p->pattern->index[iptr_ik];
351 gross 3097 if (coloring[k]<color) Paso_BlockOps_SMV_3(&x[3*i], &A_p->val[9*iptr_ik], &x[3*k]);
352 gross 3094 }
353 gross 3097 Paso_BlockOps_MV_3(&x[3*i], &diag[9*i], &x[3*i]);
354 gross 3094 }
355     }
356 gross 3097 } else {
357     #pragma omp parallel for schedule(static) private(i,iptr_ik,k)
358     for (i = 0; i < n; ++i) {
359     if (coloring[i] == color) {
360     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
361     k=A_p->pattern->index[iptr_ik];
362     if (coloring[k]<color) Paso_BlockOps_SMV_N(n_block, &x[n_block*i], &A_p->val[block_size*iptr_ik], &x[n_block*k]);
363     }
364     Paso_BlockOps_Solve_N(n_block, &x[n_block*i], &diag[block_size*i], &pivot[n_block*i], &x[n_block*i]);
365     }
366     }
367     }
368 gross 3094 } /* end of coloring loop */
369    
370     /* backward substitution */
371     for (color=(num_colors)-2 ;color>-1;--color) { /* Note: color=(num_colors)-1 is not required */
372     if (n_block==1) {
373 gross 3097 #pragma omp parallel for schedule(static) private(mm, i,iptr_ik,k)
374 gross 3094 for (i = 0; i < n; ++i) {
375     if (coloring[i]==color) {
376     mm=ptr_main[i];
377 gross 3097 Paso_BlockOps_MV_1(&x[i], &A_p->val[mm], &x[i]);
378 gross 3094 for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
379     k=A_p->pattern->index[iptr_ik];
380 gross 3097 if (coloring[k]>color) Paso_BlockOps_SMV_1(&x[i], &A_p->val[iptr_ik], &x[k]);
381 gross 3094 }
382 gross 3097 Paso_BlockOps_MV_1(&x[i], &diag[i], &x[i]);
383 gross 3094 }
384     }
385     } else if (n_block==2) {
386 gross 3097 #pragma omp parallel for schedule(static) private(mm, i,iptr_ik,k)
387 gross 3094 for (i = 0; i < n; ++i) {
388     if (coloring[i]==color) {
389     mm=ptr_main[i];
390 gross 3097 Paso_BlockOps_MV_2(&x[2*i], &A_p->val[4*mm], &x[2*i]);
391 gross 3094 for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
392     k=A_p->pattern->index[iptr_ik];
393 gross 3097 if (coloring[k]>color) Paso_BlockOps_SMV_2(&x[2*i], &A_p->val[4*iptr_ik], &x[2*k]);
394 gross 3094 }
395 gross 3097 Paso_BlockOps_MV_2(&x[2*i], &diag[4*i], &x[2*i]);
396 gross 3094 }
397     }
398     } else if (n_block==3) {
399 gross 3097 #pragma omp parallel for schedule(static) private(mm, i,iptr_ik,k)
400 gross 3094 for (i = 0; i < n; ++i) {
401     if (coloring[i]==color) {
402     mm=ptr_main[i];
403 gross 3097 Paso_BlockOps_MV_3(&x[3*i], &A_p->val[9*mm], &x[3*i]);
404 gross 3094 for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
405     k=A_p->pattern->index[iptr_ik];
406 gross 3097 if (coloring[k]>color) Paso_BlockOps_SMV_3(&x[3*i], &A_p->val[9*iptr_ik], &x[3*k]);
407 gross 3094 }
408 gross 3097 Paso_BlockOps_MV_3(&x[3*i], &diag[9*i], &x[3*i]);
409 gross 3094 }
410     }
411 gross 3097 } else {
412     #pragma omp parallel for schedule(static) private(mm, i,iptr_ik,k)
413     for (i = 0; i < n; ++i) {
414     if (coloring[i]==color) {
415     mm=ptr_main[i];
416     Paso_BlockOps_MV_N( n_block, &x[n_block*i], &A_p->val[block_size*mm], &x[n_block*i] );
417     for (iptr_ik=A_p->pattern->ptr[i];iptr_ik<A_p->pattern->ptr[i+1]; ++iptr_ik) {
418     k=A_p->pattern->index[iptr_ik];
419     if (coloring[k]>color) Paso_BlockOps_SMV_N(n_block, &x[n_block*i], &A_p->val[block_size*iptr_ik], &x[n_block*k]);
420     }
421     Paso_BlockOps_Solve_N(n_block, &x[n_block*i], &diag[block_size*i], &pivot[n_block*i], &x[n_block*i]);
422     }
423     }
424     }
425 gross 3094 }
426     return;
427     }
428    
429    
430    

  ViewVC Help
Powered by ViewVC 1.1.26