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

Contents of /trunk/paso/src/PCG.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1570 - (show annotations)
Sat May 24 21:31:04 2008 UTC (11 years, 2 months ago) by gross
File MIME type: text/plain
File size: 12931 byte(s)
modifications to PCG to support dynamic scheduling
1
2 /* $Id$ */
3
4 /*******************************************************
5 *
6 * Copyright 2003-2007 by ACceSS MNRF
7 * Copyright 2007 by University of Queensland
8 *
9 * http://esscc.uq.edu.au
10 * Primary Business: Queensland, Australia
11 * Licensed under the Open Software License version 3.0
12 * http://www.opensource.org/licenses/osl-3.0.php
13 *
14 *******************************************************/
15
16 /* PCG iterations */
17
18 #include "SystemMatrix.h"
19 #include "Paso.h"
20 #include "Solver.h"
21
22 #ifdef _OPENMP
23 #include <omp.h>
24 #endif
25
26 #ifdef PASO_MPI
27 #include <mpi.h>
28 #endif
29
30 /*
31 *
32 * Purpose
33 * =======
34 *
35 * PCG solves the linear system A*x = b using the
36 * preconditioned conjugate gradient method plus a smoother
37 * A has to be symmetric.
38 *
39 * Convergence test: norm( b - A*x )< TOL.
40 * For other measures, see the above reference.
41 *
42 * Arguments
43 * =========
44 *
45 * r (input) DOUBLE PRECISION array, dimension N.
46 * On entry, residual of inital guess x
47 *
48 * x (input/output) DOUBLE PRECISION array, dimension N.
49 * On input, the initial guess.
50 *
51 * ITER (input/output) INT
52 * On input, the maximum iterations to be performed.
53 * On output, actual number of iterations performed.
54 *
55 * INFO (output) INT
56 *
57 * = SOLVER_NO_ERROR: Successful exit. Iterated approximate solution returned.
58 * = SOLVEr_MAXITER_REACHED
59 * = SOLVER_INPUT_ERROR Illegal parameter:
60 * = SOLVEr_BREAKDOWN: If parameters rHO or OMEGA become smaller
61 * = SOLVER_MEMORY_ERROR : If parameters rHO or OMEGA become smaller
62 *
63 * ==============================================================
64 */
65
66 /* #define PASO_DYNAMIC_SCHEDULING_MVM */
67
68 #if defined PASO_DYNAMIC_SCHEDULING_MVM && defined __OPENMP
69 #define USE_DYNAMIC_SCHEDULING
70 #endif
71
72 err_t Paso_Solver_PCG(
73 Paso_SystemMatrix * A,
74 double * r,
75 double * x,
76 dim_t *iter,
77 double * tolerance,
78 Paso_Performance* pp) {
79
80 /* Local variables */
81 dim_t num_iter=0,maxit,num_iter_global, chunk_size=-1, len,rest, n_chunks, np, ipp;
82 register double ss,ss1;
83 dim_t i0, istart, iend;
84 bool_t breakFlag=FALSE, maxIterFlag=FALSE, convergeFlag=FALSE;
85 err_t status = SOLVER_NO_ERROR;
86 dim_t n = Paso_SystemMatrix_getTotalNumRows(A);
87 double *resid = tolerance, *rs=NULL, *p=NULL, *v=NULL, *x2=NULL ;
88 double tau_old,tau,beta,delta,gamma_1,gamma_2,alpha,sum_1,sum_2,sum_3,sum_4,sum_5,tol;
89 double norm_of_residual,norm_of_residual_global, loc_sum[2], sum[2];
90 register double r_tmp,d,rs_tmp,x2_tmp,x_tmp;
91 char* chksz_chr;
92 np=omp_get_max_threads();
93
94 #ifdef USE_DYNAMIC_SCHEDULING
95 chksz_chr=getenv("PASO_CHUNK_SIZE_PCG");
96 if (chksz_chr!=NULL) sscanf(chksz_chr, "%d",&chunk_size);
97 chunk_size=MIN(MAX(1,chunk_size),n/np);
98 n_chunks=n/chunk_size;
99 if (n_chunks*chunk_size<n) n_chunks+=1;
100 #else
101 len=n/np;
102 rest=n-len*np;
103 #endif
104 /* */
105 /*-----------------------------------------------------------------*/
106 /* */
107 /* Start of Calculation : */
108 /* --------------------- */
109 /* */
110 /* */
111 rs=TMPMEMALLOC(n,double);
112 p=TMPMEMALLOC(n,double);
113 v=TMPMEMALLOC(n,double);
114 x2=TMPMEMALLOC(n,double);
115
116 /* Test the input parameters. */
117
118 if (n < 0) {
119 status = SOLVER_INPUT_ERROR;
120 } else if (rs==NULL || p==NULL || v==NULL || x2==NULL) {
121 status = SOLVER_MEMORY_ERROR;
122 } else {
123 maxit = *iter;
124 tol = *resid;
125 Performance_startMonitor(pp,PERFORMANCE_SOLVER);
126 /* initialize data */
127 #pragma omp parallel (i0, istart, iend, ipp)
128 {
129 #ifdef USE_DYNAMIC_SCHEDULING
130 #pragma omp for schedule(dynamic, 1)
131 for (ipp=0; ipp < n_chunks; ++ipp) {
132 istart=chunk_size*ipp;
133 iend=MIN(istart+chunk_size,n);
134 #else
135 #pragma omp for schedule(static)
136 for (ipp=0; ipp <np; ++ipp) {
137 istart=len*ipp+MIN(ipp,rest);
138 iend=len*(ipp+1)+MIN(ipp+1,rest);
139 #endif
140 #pragma ivdep
141 for (i0=istart;i0<iend;i0++) {
142 rs[i0]=r[i0];
143 x2[i0]=x[i0];
144 p[i0]=0;
145 v[i0]=0;
146 }
147 #ifdef USE_DYNAMIC_SCHEDULING
148 }
149 #else
150 }
151 #endif
152 }
153 num_iter=0;
154 /* start of iteration */
155 while (!(convergeFlag || maxIterFlag || breakFlag)) {
156 ++(num_iter);
157 /* v=prec(r) */
158 Performance_stopMonitor(pp,PERFORMANCE_SOLVER);
159 Performance_startMonitor(pp,PERFORMANCE_PRECONDITIONER);
160 Paso_Solver_solvePreconditioner(A,v,r);
161 Performance_stopMonitor(pp,PERFORMANCE_PRECONDITIONER);
162 Performance_startMonitor(pp,PERFORMANCE_SOLVER);
163 /* tau=v*r */
164 sum_1 = 0;
165 #pragma omp parallel (i0, istart, iend, ipp, ss)
166 {
167 #ifdef USE_DYNAMIC_SCHEDULING
168 #pragma omp for schedule(dynamic, 1)
169 for (ipp=0; ipp < n_chunks; ++ipp) {
170 istart=chunk_size*ipp;
171 iend=MIN(istart+chunk_size,n);
172 #else
173 #pragma omp for schedule(static)
174 for (ipp=0; ipp <np; ++ipp) {
175 istart=len*ipp+MIN(ipp,rest);
176 iend=len*(ipp+1)+MIN(ipp+1,rest);
177 #endif
178 ss=0;
179 #pragma ivdep
180 for (i0=istart;i0<iend;i0++) ss+=v[i0]*r[i0];
181 #pragma critical
182 sum_1+=ss;
183 #ifdef USE_DYNAMIC_SCHEDULING
184 }
185 #else
186 }
187 #endif
188 }
189 #ifdef PASO_MPI
190 /* In case we have many MPI processes, each of which may have several OMP threads:
191 OMP master participates in an MPI reduction to get global sum_1 */
192 loc_saum[0] = sum_1;
193 MPI_Allreduce(loc_sum, &sum_1, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
194 #endif
195 tau_old=tau;
196 tau=sum_1;
197 /* p=v+beta*p */
198 #pragma omp parallel (i0, istart, iend, ipp,beta)
199 {
200 #ifdef USE_DYNAMIC_SCHEDULING
201 #pragma omp for schedule(dynamic, 1)
202 for (ipp=0; ipp < n_chunks; ++ipp) {
203 istart=chunk_size*ipp;
204 iend=MIN(istart+chunk_size,n);
205 #else
206 #pragma omp for schedule(static)
207 for (ipp=0; ipp <np; ++ipp) {
208 istart=len*ipp+MIN(ipp,rest);
209 iend=len*(ipp+1)+MIN(ipp+1,rest);
210 #endif
211 if (num_iter==1) {
212 #pragma ivdep
213 for (i0=istart;i0<iend;i0++) p[i0]=v[i0];
214 } else {
215 beta=tau/tau_old;
216 #pragma ivdep
217 for (i0=istart;i0<iend;i0++) p[i0]=v[i0]+beta*p[i0];
218 }
219 #ifdef USE_DYNAMIC_SCHEDULING
220 }
221 #else
222 }
223 #endif
224 }
225 /* v=A*p */
226 Performance_stopMonitor(pp,PERFORMANCE_SOLVER);
227 Performance_startMonitor(pp,PERFORMANCE_MVM);
228 Paso_SystemMatrix_MatrixVector_CSR_OFFSET0(ONE, A, p,ZERO,v);
229 Performance_stopMonitor(pp,PERFORMANCE_MVM);
230 Performance_startMonitor(pp,PERFORMANCE_SOLVER);
231
232 /* delta=p*v */
233 sum_2 = 0;
234 #pragma omp parallel (i0, istart, iend, ipp,ss)
235 {
236 #ifdef USE_DYNAMIC_SCHEDULING
237 #pragma omp for schedule(dynamic, 1)
238 for (ipp=0; ipp < n_chunks; ++ipp) {
239 istart=chunk_size*ipp;
240 iend=MIN(istart+chunk_size,n);
241 #else
242 #pragma omp for schedule(static)
243 for (ipp=0; ipp <np; ++ipp) {
244 istart=len*ipp+MIN(ipp,rest);
245 iend=len*(ipp+1)+MIN(ipp+1,rest);
246 #endif
247 ss=0;
248 #pragma ivdep
249 for (i0=istart;i0<iend;i0++) ss+=v[i0]*p[i0];
250 #pragma critical
251 sum_2+=ss;
252 #ifdef USE_DYNAMIC_SCHEDULING
253 }
254 #else
255 }
256 #endif
257 }
258 #ifdef PASO_MPI
259 loc_sum[0] = sum_2;
260 MPI_Allreduce(loc_sum, &sum_2, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
261 #endif
262 delta=sum_2;
263
264
265 if (! (breakFlag = (ABS(delta) <= TOLERANCE_FOR_SCALARS))) {
266 /* smoother */
267 sum_3 = 0;
268 sum_4 = 0;
269 #pragma omp parallel (i0, istart, iend, ipp,d, ss, ss1, alpha)
270 {
271 #ifdef USE_DYNAMIC_SCHEDULING
272 #pragma omp for schedule(dynamic, 1)
273 for (ipp=0; ipp < n_chunks; ++ipp) {
274 istart=chunk_size*ipp;
275 iend=MIN(istart+chunk_size,n);
276 #else
277 #pragma omp for schedule(static)
278 for (ipp=0; ipp <np; ++ipp) {
279 istart=len*ipp+MIN(ipp,rest);
280 iend=len*(ipp+1)+MIN(ipp+1,rest);
281 #endif
282 ss=0;
283 ss1=0;
284 alpha=tau/delta;
285 #pragma ivdep
286 for (i0=istart;i0<iend;i0++) {
287 r[i0]-=alpha*v[i0];
288 d=r[i0]-rs[i0];
289 ss+=d*d;
290 ss1+=d*rs[i0];
291 }
292 #pragma critical
293 {
294 sum_3+=ss;
295 sum_4+=ss1;
296 }
297 #ifdef USE_DYNAMIC_SCHEDULING
298 }
299 #else
300 }
301 #endif
302 }
303 #ifdef PASO_MPI
304 loc_sum[0] = sum_3;
305 loc_sum[1] = sum_4;
306 MPI_Allreduce(loc_sum, sum, 2, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
307 sum_3=sum[0];
308 sum_4=sum[1];
309 #endif
310 sum_5 = 0;
311 #pragma omp parallel (i0, istart, iend, ipp, ss, gamma_1,gamma_2)
312 {
313 gamma_1= ( (ABS(sum_3)<= ZERO) ? 0 : -sum_4/sum_3) ;
314 gamma_2= ONE-gamma_1;
315 #ifdef USE_DYNAMIC_SCHEDULING
316 #pragma omp for schedule(dynamic, 1)
317 for (ipp=0; ipp < n_chunks; ++ipp) {
318 istart=chunk_size*ipp;
319 iend=MIN(istart+chunk_size,n);
320 #else
321 #pragma omp for schedule(static)
322 for (ipp=0; ipp <np; ++ipp) {
323 istart=len*ipp+MIN(ipp,rest);
324 iend=len*(ipp+1)+MIN(ipp+1,rest);
325 #endif
326 ss=0;
327 #pragma ivdep
328 for (i0=istart;i0<iend;i0++) {
329 rs[i0]=gamma_2*rs[i0]+gamma_1*r[i0];
330 x2[i0]+=alpha*p[i0];
331 x[i0]=gamma_2*x[i0]+gamma_1*x2[i0];
332 ss+=rs[i0]*rs[i0];
333 }
334 #pragma omp critical
335 sum_5+=ss;
336 #ifdef USE_DYNAMIC_SCHEDULING
337 }
338 #else
339 }
340 #endif
341 }
342 #ifdef PASO_MPI
343 loc_sum[0] = sum_5;
344 MPI_Allreduce(loc_sum, &sum_5, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
345 #endif
346 norm_of_residual=sqrt(sum_5);
347 convergeFlag = norm_of_residual <= tol;
348 maxIterFlag = num_iter == maxit;
349 breakFlag = (ABS(tau) <= TOLERANCE_FOR_SCALARS);
350 }
351 }
352 /* end of iteration */
353 num_iter_global=num_iter;
354 norm_of_residual_global=norm_of_residual;
355 if (maxIterFlag) {
356 status = SOLVER_MAXITER_REACHED;
357 } else if (breakFlag) {
358 status = SOLVER_BREAKDOWN;
359 }
360 Performance_stopMonitor(pp,PERFORMANCE_SOLVER);
361 TMPMEMFREE(rs);
362 TMPMEMFREE(x2);
363 TMPMEMFREE(v);
364 TMPMEMFREE(p);
365 *iter=num_iter_global;
366 *resid=norm_of_residual_global;
367 }
368 /* End of PCG */
369 return status;
370 }

Properties

Name Value
svn:eol-style native
svn:keywords Author Date Id Revision

  ViewVC Help
Powered by ViewVC 1.1.26