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

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

Parent Directory Parent Directory | Revision Log Revision Log


Revision 3026 - (show annotations)
Mon May 10 06:23:49 2010 UTC (9 years, 5 months ago) by gross
File MIME type: text/plain
File size: 15078 byte(s)
some fix for omp
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: Transport solver with flux corretion (L is row sum zero)
18 *
19 * - Mv_t=Lv v(0)=u
20 *
21 * to return v(dt)
22 *
23 */
24 /**************************************************************/
25
26 /* Author: l.gross@uq.edu.au */
27
28 /**************************************************************/
29
30 #include "FCTSolver.h"
31 #include "Solver.h"
32 #include "PasoUtil.h"
33
34 Paso_Function* Paso_FCTSolver_Function_alloc(Paso_TransportProblem *fctp, Paso_Options* options)
35 {
36 const dim_t n=Paso_SystemMatrix_getTotalNumRows(fctp->transport_matrix);
37 const dim_t blockSize=Paso_TransportProblem_getBlockSize(fctp);
38 Paso_Function * out=NULL;
39 Paso_FCTSolver *more=NULL;
40 out=MEMALLOC(1,Paso_Function);
41 if (! Paso_checkPtr(out)) {
42 out->kind=FCT;
43 out->mpi_info=Paso_MPIInfo_getReference(fctp->mpi_info);
44 out->n=n;
45 out->b=TMPMEMALLOC(n,double); /*b_m */
46 Paso_checkPtr(out->b);
47 out->tmp=TMPMEMALLOC(n,double); /* z_m */
48 Paso_checkPtr(out->tmp);
49
50 more=MEMALLOC(1,Paso_FCTSolver);
51 out->more=(void*) more;
52 if (! Paso_checkPtr(more)) {
53 more->transportproblem=Paso_TransportProblem_getReference(fctp);
54 more->uTilde_n=TMPMEMALLOC(n,double);
55 Paso_checkPtr(more->uTilde_n);
56 more->QN_n=TMPMEMALLOC(n,double);
57 Paso_checkPtr(more->QN_n);
58 more->QP_n=TMPMEMALLOC(n,double);
59 Paso_checkPtr(more->QP_n);
60 more->RN_m=TMPMEMALLOC(n,double);
61 Paso_checkPtr(more->RN_m);
62 more->RP_m=TMPMEMALLOC(n,double);
63 Paso_checkPtr(more->RP_m);
64 more->QN_n_coupler=Paso_Coupler_alloc(Paso_TransportProblem_borrowConnector(fctp),blockSize);
65 more->QP_n_coupler=Paso_Coupler_alloc(Paso_TransportProblem_borrowConnector(fctp),blockSize);
66 more->RN_m_coupler=Paso_Coupler_alloc(Paso_TransportProblem_borrowConnector(fctp),blockSize);
67 more->RP_m_coupler=Paso_Coupler_alloc(Paso_TransportProblem_borrowConnector(fctp),blockSize);
68 more->uTilde_n_coupler=Paso_Coupler_alloc(Paso_TransportProblem_borrowConnector(fctp),blockSize);
69 more->u_m_coupler=Paso_Coupler_alloc(Paso_TransportProblem_borrowConnector(fctp),blockSize);
70 more->flux_matrix_m=Paso_SystemMatrix_alloc(fctp->transport_matrix->type,
71 fctp->transport_matrix->pattern,
72 fctp->transport_matrix->row_block_size,
73 fctp->transport_matrix->col_block_size, TRUE);
74
75 }
76 }
77 if (Paso_noError()) {
78 return out;
79 } else {
80 Paso_FCTSolver_Function_free(out);
81 return NULL;
82 }
83 }
84
85 void Paso_FCTSolver_Function_free(Paso_Function * in)
86 {
87 Paso_FCTSolver *more=NULL;
88 if (in!=NULL) {
89 Paso_MPIInfo_free(in->mpi_info);
90 MEMFREE(in->tmp);
91 MEMFREE(in->b);
92 more=(Paso_FCTSolver *) (in->more);
93 if (NULL !=more) {
94 Paso_TransportProblem_free(more->transportproblem);
95 Paso_SystemMatrix_free(more->flux_matrix_m);
96 TMPMEMFREE(more->uTilde_n);
97 TMPMEMFREE(more->QN_n);
98 TMPMEMFREE(more->QP_n);
99 TMPMEMFREE(more->RN_m);
100 TMPMEMFREE(more->RP_m);
101 Paso_Coupler_free(more->QN_n_coupler);
102 Paso_Coupler_free(more->QP_n_coupler);
103 Paso_Coupler_free(more->RN_m_coupler);
104 Paso_Coupler_free(more->RP_m_coupler);
105 Paso_Coupler_free(more->uTilde_n_coupler);
106 Paso_Coupler_free(more->u_m_coupler);
107 }
108 MEMFREE(in);
109 }
110 }
111 double Paso_FCTSolver_getSafeTimeStepSize(Paso_TransportProblem* fctp)
112 {
113 dim_t i, n;
114 double dt_max=LARGE_POSITIVE_FLOAT, dt_max_loc;
115 register double l_ii,m_i;
116 index_t fail=0, fail_loc;
117 n=Paso_SystemMatrix_getTotalNumRows(fctp->transport_matrix);
118 /* set low order transport operator */
119 Paso_FCTSolver_setLowOrderOperator(fctp);
120
121 if (Paso_noError()) {
122 /*
123 * calculate time step size:
124 */
125 dt_max=LARGE_POSITIVE_FLOAT;
126 fail=0;
127 #pragma omp parallel private(dt_max_loc, fail_loc)
128 {
129 dt_max_loc=LARGE_POSITIVE_FLOAT;
130 fail_loc=0;
131 #pragma omp for schedule(static) private(i,l_ii,m_i)
132 for (i=0;i<n;++i) {
133 l_ii=fctp->main_diagonal_low_order_transport_matrix[i];
134 m_i=fctp->lumped_mass_matrix[i];
135 if ( (m_i > 0) && (l_ii < 0 ) ) {
136 dt_max_loc=MIN(dt_max_loc,-m_i/l_ii);
137 } else {
138 fail_loc=-1;
139 }
140 }
141 #pragma omp critical
142 {
143 dt_max=MIN(dt_max,dt_max_loc);
144 fail=MIN(fail, fail_loc);
145 }
146 }
147 #ifdef PASO_MPI
148 {
149 double rtmp_loc[2], rtmp[2];
150 rtmp_loc[0]=dt_max;
151 rtmp_loc[1]= (double) fail;
152 MPI_Allreduce(rtmp_loc, rtmp, 2, MPI_DOUBLE, MPI_MIN, fctp->mpi_info->comm);
153 dt_max=rtmp[0];
154 fail = rtmp[1] < 0 ? -1 : 0;
155 }
156 #endif
157 if (fail < 0 ) {
158 Paso_setError(VALUE_ERROR, "Paso_FCTSolver_getSafeTimeStepSize: negative mass or positive diffusion term detected.");
159 return -1;
160 } else {
161 if (dt_max<LARGE_POSITIVE_FLOAT) dt_max*=fctp->dt_factor;
162 }
163 }
164 return dt_max;
165 }
166 /*
167 * evaluates F as
168 * (m_i*omega - l_ij)*(F_j/omega)
169 * =(m u^n[i] + dt*(1-theta) sum_{j <> i} l_{ij}*(u^n[j]-u^n[i]))-(m_i*u[i] - dt*theta*sum_{j<>i} l_{ij} (u[j]-u[i]) )- antidiffusion
170 *
171 */
172 err_t Paso_FCTSolver_Function_call(Paso_Function * F,double* value, const double* arg, Paso_Performance *pp)
173 {
174 Paso_Options options;
175 Paso_FCTSolver *more=(Paso_FCTSolver *) (F->more);
176 Paso_Options_setDefaults(&options);
177 options.verbose=TRUE;
178
179 /* set z_m in tmp */
180 Paso_FCTSolver_setUpRightHandSide(more->transportproblem,more->dt,arg,more->u_m_coupler,F->tmp,more->flux_matrix_m,
181 more->uTilde_n_coupler,F->b,
182 more->QN_n_coupler,more->QP_n_coupler,more->RN_m,
183 more->RN_m_coupler,more->RP_m,more->RP_m_coupler,pp);
184 /* apply preconditoner do get du = */
185 /* Paso_Solver(more->transportproblem->iteration_matrix,value,F->tmp,&options,pp); */
186
187
188 Paso_Solver_solvePreconditioner(more->transportproblem->iteration_matrix,value,F->tmp);
189 return NO_ERROR;
190 }
191
192 err_t Paso_FCTSolver_solve(Paso_Function* F, double* u, double dt, Paso_Options* options, Paso_Performance *pp)
193 {
194 const dim_t num_critical_rates_max=3; /* number of rates >=critical_rate accepted before divergence is triggered */
195 const double critical_rate=0.6; /* expected value of convergence rate */
196
197 double norm_u_tilde, ATOL, norm_u, norm_du=LARGE_POSITIVE_FLOAT, norm_du_old, *du=NULL, rate;
198 err_t errorCode=SOLVER_NO_ERROR;
199 Paso_FCTSolver *more=(Paso_FCTSolver *) (F->more);
200 const double omega=1./(dt*Paso_Transport_getTheta(more->transportproblem));
201 Paso_TransportProblem* fctp=more->transportproblem;
202 const dim_t n=Paso_SystemMatrix_getTotalNumRows(fctp->transport_matrix);
203 const double atol=options->absolute_tolerance;
204 const double rtol=options->tolerance;
205 const dim_t max_m=options->iter_max;
206 dim_t m=0, num_critical_rates=0 ;
207 bool_t converged=FALSE, max_m_reached=FALSE,diverged=FALSE;
208
209 /* tolerance? */
210 more->dt=dt;
211 Paso_FCTSolver_setUp(fctp,dt,u, F->b,more->uTilde_n,more->uTilde_n_coupler,more->QN_n,more->QN_n_coupler,more->QP_n,more->QP_n_coupler,
212 options,pp);
213
214
215 norm_u_tilde=Paso_lsup(n,more->uTilde_n,F->mpi_info);
216 ATOL= rtol * norm_u_tilde + atol ;
217 if (options->verbose) printf("Paso_FCTSolver_solve: iteration starts u_tilda lsup = %e (abs. tol = %e)\n",norm_u_tilde,ATOL);
218 if (more->transportproblem->useBackwardEuler && FALSE ) {
219 options->absolute_tolerance = ATOL/omega;
220 options->tolerance=0;
221 errorCode=Paso_Solver_NewtonGMRES(F,u,options,pp);
222 options->absolute_tolerance=atol;
223 options->tolerance=rtol;
224 } else {
225 m=0;
226 du=MEMALLOC(n,double);
227 if (Paso_checkPtr(du)) {
228 errorCode=SOLVER_MEMORY_ERROR;
229 } else {
230 /* tolerance? */
231
232 while ( (!converged) && (!diverged) && (! max_m_reached) && Paso_noError()) {
233 errorCode=Paso_FCTSolver_Function_call(F,du, u, pp);
234 Paso_Update(n,1.,u,omega,du);
235 norm_u=Paso_lsup(n,u, fctp->mpi_info);
236 norm_du_old=norm_du;
237 norm_du=Paso_lsup(n,du, fctp->mpi_info);
238 if (m ==0) {
239 if (options->verbose) printf("Paso_FCTSolver_solve: step %d: increment= %e\n",m+1, norm_du*omega);
240
241 } else {
242 if (norm_du_old > 0.) {
243 rate=norm_du/norm_du_old;
244 } else if (norm_du <= 0.) {
245 rate=0.;
246 } else {
247 rate=LARGE_POSITIVE_FLOAT;
248 }
249 if (options->verbose) printf("Paso_FCTSolver_solve: step %d: increment= %e (rate = %e)\n",m+1, norm_du*omega, rate);
250 num_critical_rates+=( rate<critical_rate ? 0 : 1);
251 }
252
253 max_m_reached=(m>max_m);
254 diverged = (num_critical_rates >= num_critical_rates_max);
255 converged=(norm_du*omega <= ATOL) ;
256 m++;
257 }
258 }
259 MEMFREE(du);
260 if (errorCode == SOLVER_NO_ERROR) {
261 if (converged) {
262 if (options->verbose) printf("Paso_FCTSolver_solve: iteration is completed.\n");
263 errorCode=SOLVER_NO_ERROR;
264 } else if (diverged) {
265 if (options->verbose) printf("Paso_FCTSolver_solve: divergence.\n");
266 errorCode=SOLVER_DIVERGENCE;
267 } else if (max_m_reached) {
268 if (options->verbose) printf("Paso_FCTSolver_solve: maximum number of iteration steps reached.\n");
269 errorCode=SOLVER_MAXITER_REACHED;
270 }
271 }
272 }
273 return errorCode;
274 }
275
276
277
278 err_t Paso_FCTSolver_setUpRightHandSide(Paso_TransportProblem* fctp, const double dt, const double *u_m,
279 Paso_Coupler* u_m_coupler, double * z_m,
280 Paso_SystemMatrix* flux_matrix, Paso_Coupler* uTilde_coupler, const double *b,
281 Paso_Coupler* QN_coupler, Paso_Coupler* QP_coupler,
282 double *RN_m, Paso_Coupler* RN_m_coupler, double* RP_m, Paso_Coupler* RP_m_coupler,
283 Paso_Performance* pp)
284 {
285 const double omega=dt* Paso_Transport_getTheta(fctp);
286 const dim_t n=Paso_SystemMatrix_getTotalNumRows(fctp->transport_matrix);
287 /* distribute u */
288 Paso_Coupler_startCollect(u_m_coupler,u_m);
289 Paso_Coupler_finishCollect(u_m_coupler);
290 /*
291 * set the ant diffusion fluxes:
292 *
293 */
294 Paso_FCTSolver_setAntiDiffusionFlux(dt,fctp,flux_matrix,u_m_coupler);
295 /*
296 * apply pre flux-correction: f_{ij}:=0 if f_{ij}*(\tilde{u}[i]- \tilde{u}[j])<=0
297 *
298 */
299 Paso_FCTSolver_applyPreAntiDiffusionCorrection(flux_matrix,uTilde_coupler);
300 /*
301 * set flux limiters RN_m,RP_m
302 *
303 */
304 Paso_FCTSolver_setRs(flux_matrix,fctp->lumped_mass_matrix,QN_coupler,QP_coupler,RN_m,RP_m);
305 Paso_Coupler_startCollect(RN_m_coupler,RN_m);
306 Paso_Coupler_startCollect(RP_m_coupler,RP_m);
307 /*
308 * z_m[i]=b[i] - (m_i*u_m[i] - dt*theta*sum_{j<>i} l_{ij} (u_m[j]-u_m[i]) )
309 *
310 * note that iteration_matrix stores the negative values of the
311 * low order transport matrix l therefore a=dt*theta is used.
312 */
313 Paso_FCTSolver_setMuPaLu(z_m, fctp->lumped_mass_matrix, u_m_coupler, omega, fctp->iteration_matrix);
314 /*{
315 int kk;
316 for (kk=0;kk<n;kk++) printf("z_m %d : %e %e -> %e\n",kk,z_m[kk], b[kk],z_m[kk]-b[kk]);
317 }
318 */
319 /* z_m=b-z_m */
320 Paso_Update(n,-1.,z_m,1.,b);
321
322 Paso_Coupler_finishCollect(RN_m_coupler);
323 Paso_Coupler_finishCollect(RP_m_coupler);
324 /* add corrected fluxes into z_m */
325 Paso_FCTSolver_addCorrectedFluxes(z_m,flux_matrix,RN_m_coupler,RP_m_coupler);
326 return SOLVER_NO_ERROR;
327 }
328
329 void Paso_FCTSolver_Function_initialize(const double dt, Paso_TransportProblem* fctp, Paso_Options* options, Paso_Performance* pp)
330 {
331 dim_t i;
332 const index_t* main_iptr=Paso_TransportProblem_borrowMainDiagonalPointer(fctp);
333 const dim_t n=Paso_SystemMatrix_getTotalNumRows(fctp->transport_matrix);
334 const double theta = Paso_Transport_getTheta(fctp);
335 const double omega=1./(dt* theta);
336 register double m, rtmp4;
337
338 /*
339 * fctp->iteration_matrix[i,i]=m[i]/(dt theta) -l[i,i]
340 *
341 */
342 Paso_solve_free(fctp->iteration_matrix);
343 #pragma omp parallel for private(i,m,rtmp4)
344 for (i = 0; i < n; ++i) {
345 m=fctp->lumped_mass_matrix[i];
346 rtmp4=m*omega-(fctp->main_diagonal_low_order_transport_matrix[i]);
347 fctp->iteration_matrix->mainBlock->val[main_iptr[i]]=rtmp4;
348 }
349 Performance_startMonitor(pp,PERFORMANCE_PRECONDITIONER_INIT);
350 Paso_Solver_setPreconditioner(fctp->iteration_matrix,options);
351 Performance_stopMonitor(pp,PERFORMANCE_PRECONDITIONER_INIT);
352 }
353
354 void Paso_FCTSolver_setUp(Paso_TransportProblem* fctp, const double dt, const double *u, double* b, double* uTilde,
355 Paso_Coupler* uTilde_coupler, double *QN, Paso_Coupler* QN_coupler, double *QP, Paso_Coupler* QP_coupler,
356 Paso_Options* options, Paso_Performance* pp)
357 {
358 dim_t i;
359 const dim_t n=Paso_SystemMatrix_getTotalNumRows(fctp->transport_matrix);
360 register double m;
361 /* distribute u */
362 Paso_Coupler_startCollect(fctp->u_coupler,u);
363 Paso_Coupler_finishCollect(fctp->u_coupler);
364 /*
365 * b^n[i]=m u^n[i] + dt*(1-theta) sum_{j <> i} l_{ij}*(u^n[j]-u^n[i])
366 *
367 * note that iteration_matrix stores the negative values of the
368 * low order transport matrix l therefore a=-dt*(1-fctp->theta) is used.
369 *
370 */
371 if (fctp->useBackwardEuler) {
372 Paso_FCTSolver_setMuPaLu(b,fctp->lumped_mass_matrix,fctp->u_coupler,0.,fctp->iteration_matrix);
373 } else {
374 Paso_FCTSolver_setMuPaLu(b,fctp->lumped_mass_matrix,fctp->u_coupler,-dt*0.5,fctp->iteration_matrix);
375 }
376 /*
377 * uTilde[i]=b[i]/m[i]
378 *
379 */
380 #pragma omp parallel for private(i,m)
381 for (i = 0; i < n; ++i) {
382 m=fctp->lumped_mass_matrix[i];
383 uTilde[i]=b[i]/m;
384 }
385 /* distribute uTilde: */
386 Paso_Coupler_startCollect(uTilde_coupler,uTilde);
387 Paso_Coupler_finishCollect(uTilde_coupler);
388 /*
389 * calculate QP[i] max_{j} (\tilde{u}[j]- \tilde{u}[i] )
390 * QN[i] min_{j} (\tilde{u}[j]- \tilde{u}[i] )
391 *
392 */
393 Paso_FCTSolver_setQs(uTilde_coupler,QN,QP,fctp->iteration_matrix);
394 Paso_Coupler_startCollect(QN_coupler,QN);
395 Paso_Coupler_startCollect(QP_coupler,QP);
396 Paso_Coupler_finishCollect(QN_coupler);
397 Paso_Coupler_finishCollect(QP_coupler);
398 }

  ViewVC Help
Powered by ViewVC 1.1.26