/[escript]/branches/arrexp_2137_win/paso/src/PCG.c
ViewVC logotype

Annotation of /branches/arrexp_2137_win/paso/src/PCG.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1556 - (hide annotations)
Mon May 12 00:54:58 2008 UTC (11 years, 1 month ago) by gross
Original Path: trunk/paso/src/PCG.c
File MIME type: text/plain
File size: 8022 byte(s)
Modification to allow mixed mode execution. 
In order to keep the code portable accross platform all MPI calls within
parallel regions have been moved. 


1 ksteube 1312
2 jgs 150 /* $Id$ */
3    
4 ksteube 1312 /*******************************************************
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 gross 495
16 jgs 150 /* PCG iterations */
17    
18 gross 700 #include "SystemMatrix.h"
19     #include "Paso.h"
20 jgs 150 #include "Solver.h"
21 woo409 757
22 jgs 150 #ifdef _OPENMP
23     #include <omp.h>
24     #endif
25    
26 ksteube 1312 #ifdef PASO_MPI
27     #include <mpi.h>
28     #endif
29    
30 jgs 150 /*
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     err_t Paso_Solver_PCG(
67     Paso_SystemMatrix * A,
68     double * r,
69     double * x,
70     dim_t *iter,
71 gross 584 double * tolerance,
72     Paso_Performance* pp) {
73 jgs 150
74    
75     /* Local variables */
76     dim_t num_iter=0,maxit,num_iter_global;
77     dim_t i0;
78     bool_t breakFlag=FALSE, maxIterFlag=FALSE, convergeFlag=FALSE;
79     err_t status = SOLVER_NO_ERROR;
80 ksteube 1312 dim_t n = Paso_SystemMatrix_getTotalNumRows(A);
81 jgs 150 double *resid = tolerance, *rs=NULL, *p=NULL, *v=NULL, *x2=NULL ;
82     double tau_old,tau,beta,delta,gamma_1,gamma_2,alpha,sum_1,sum_2,sum_3,sum_4,sum_5,tol;
83 ksteube 1312 double norm_of_residual,norm_of_residual_global, loc_sum[2], sum[2];
84 gross 495 register double r_tmp,d,rs_tmp,x2_tmp,x_tmp;
85 jgs 150
86     /* */
87     /*-----------------------------------------------------------------*/
88     /* */
89     /* Start of Calculation : */
90     /* --------------------- */
91     /* */
92     /* */
93     rs=TMPMEMALLOC(n,double);
94     p=TMPMEMALLOC(n,double);
95     v=TMPMEMALLOC(n,double);
96     x2=TMPMEMALLOC(n,double);
97    
98     /* Test the input parameters. */
99    
100     if (n < 0) {
101     status = SOLVER_INPUT_ERROR;
102     } else if (rs==NULL || p==NULL || v==NULL || x2==NULL) {
103     status = SOLVER_MEMORY_ERROR;
104     } else {
105     maxit = *iter;
106     tol = *resid;
107 gross 1556 Performance_startMonitor(pp,PERFORMANCE_SOLVER);
108     /* initialize data */
109     #pragma omp parallel
110 jgs 150 {
111 gross 1556 #pragma omp for private(i0) schedule(static)
112     for (i0=0;i0<n;i0++) {
113     rs[i0]=r[i0];
114     x2[i0]=x[i0];
115     p[i0]=0;
116     v[i0]=0;
117     }
118     }
119     num_iter=0;
120     /* start of iteration */
121     while (!(convergeFlag || maxIterFlag || breakFlag)) {
122 jgs 150 ++(num_iter);
123 gross 1556 sum_1 = 0;
124     sum_2 = 0;
125     sum_3 = 0;
126     sum_4 = 0;
127     sum_5 = 0;
128 jgs 150 /* v=prec(r) */
129 gross 584 Performance_stopMonitor(pp,PERFORMANCE_SOLVER);
130     Performance_startMonitor(pp,PERFORMANCE_PRECONDITIONER);
131 jgs 150 Paso_Solver_solvePreconditioner(A,v,r);
132 gross 584 Performance_stopMonitor(pp,PERFORMANCE_PRECONDITIONER);
133     Performance_startMonitor(pp,PERFORMANCE_SOLVER);
134 jgs 150 /* tau=v*r */
135 gross 1556 #pragma omp parallel for private(i0) reduction(+:sum_1) schedule(static)
136 ksteube 1312 for (i0=0;i0<n;i0++) sum_1+=v[i0]*r[i0]; /* Limit to local values of v[] and r[] */
137     #ifdef PASO_MPI
138     /* In case we have many MPI processes, each of which may have several OMP threads:
139     OMP master participates in an MPI reduction to get global sum_1 */
140 gross 1556 loc_sum[0] = sum_1;
141     MPI_Allreduce(loc_sum, &sum_1, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
142 ksteube 1312 #endif
143 jgs 150 tau_old=tau;
144     tau=sum_1;
145     /* p=v+beta*p */
146     if (num_iter==1) {
147 gross 1556 #pragma omp parallel for private(i0) schedule(static)
148 jgs 150 for (i0=0;i0<n;i0++) p[i0]=v[i0];
149     } else {
150     beta=tau/tau_old;
151 gross 1556 #pragma omp parallel for private(i0) schedule(static)
152 jgs 150 for (i0=0;i0<n;i0++) p[i0]=v[i0]+beta*p[i0];
153     }
154     /* v=A*p */
155 gross 584 Performance_stopMonitor(pp,PERFORMANCE_SOLVER);
156     Performance_startMonitor(pp,PERFORMANCE_MVM);
157 gross 415 Paso_SystemMatrix_MatrixVector_CSR_OFFSET0(ONE, A, p,ZERO,v);
158 gross 584 Performance_stopMonitor(pp,PERFORMANCE_MVM);
159     Performance_startMonitor(pp,PERFORMANCE_SOLVER);
160 jgs 150 /* delta=p*v */
161 gross 1556 #pragma omp parallel for private(i0) reduction(+:sum_2) schedule(static)
162 jgs 150 for (i0=0;i0<n;i0++) sum_2+=v[i0]*p[i0];
163 ksteube 1312 #ifdef PASO_MPI
164 gross 1556 loc_sum[0] = sum_2;
165     MPI_Allreduce(loc_sum, &sum_2, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
166 ksteube 1312 #endif
167 jgs 150 delta=sum_2;
168    
169    
170     if (! (breakFlag = (ABS(delta) <= TOLERANCE_FOR_SCALARS))) {
171     alpha=tau/delta;
172     /* smoother */
173 gross 1556 #pragma omp parallel
174     {
175     #pragma omp for private(i0) schedule(static)
176     for (i0=0;i0<n;i0++) r[i0]-=alpha*v[i0];
177     #pragma omp for private(i0,d) reduction(+:sum_3,sum_4) schedule(static)
178     for (i0=0;i0<n;i0++) {
179     d=r[i0]-rs[i0];
180     sum_3+=d*d;
181     sum_4+=d*rs[i0];
182     }
183 ksteube 1312 }
184     #ifdef PASO_MPI
185 gross 1556 loc_sum[0] = sum_3;
186     loc_sum[1] = sum_4;
187     MPI_Allreduce(loc_sum, sum, 2, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
188     sum_3=sum[0];
189     sum_4=sum[1];
190 ksteube 1312 #endif
191 jgs 150 gamma_1= ( (ABS(sum_3)<= ZERO) ? 0 : -sum_4/sum_3) ;
192     gamma_2= ONE-gamma_1;
193 gross 1556 #pragma omp parallel
194     {
195     #pragma omp for private(i0,x2_tmp,x_tmp,rs_tmp) schedule(static)
196     for (i0=0;i0<n;++i0) {
197     rs[i0]=gamma_2*rs[i0]+gamma_1*r[i0];
198     x2[i0]+=alpha*p[i0];
199     x[i0]=gamma_2*x[i0]+gamma_1*x2[i0];
200     }
201     #pragma omp for private(i0) reduction(+:sum_5) schedule(static)
202     for (i0=0;i0<n;++i0) sum_5+=rs[i0]*rs[i0];
203 gross 495 }
204 ksteube 1312 #ifdef PASO_MPI
205 gross 1556 loc_sum[0] = sum_5;
206     MPI_Allreduce(loc_sum, &sum_5, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm);
207 ksteube 1312 #endif
208 gross 495 norm_of_residual=sqrt(sum_5);
209     convergeFlag = norm_of_residual <= tol;
210     maxIterFlag = num_iter == maxit;
211     breakFlag = (ABS(tau) <= TOLERANCE_FOR_SCALARS);
212     }
213 gross 1556 }
214     /* end of iteration */
215     num_iter_global=num_iter;
216     norm_of_residual_global=norm_of_residual;
217     if (maxIterFlag) {
218     status = SOLVER_MAXITER_REACHED;
219     } else if (breakFlag) {
220     status = SOLVER_BREAKDOWN;
221     }
222     Performance_stopMonitor(pp,PERFORMANCE_SOLVER);
223 jgs 150 TMPMEMFREE(rs);
224     TMPMEMFREE(x2);
225     TMPMEMFREE(v);
226     TMPMEMFREE(p);
227     *iter=num_iter_global;
228     *resid=norm_of_residual_global;
229     }
230     /* End of PCG */
231     return status;
232     }

Properties

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

  ViewVC Help
Powered by ViewVC 1.1.26