| 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 |
/* |
| 17 |
Crude modifications and translations for Paso by Matt Davies and Lutz Gross |
| 18 |
*/ |
| 19 |
|
| 20 |
#include "Paso.h" |
| 21 |
#include "SystemMatrix.h" |
| 22 |
#include "Solver.h" |
| 23 |
#ifdef _OPENMP |
| 24 |
#include <omp.h> |
| 25 |
#endif |
| 26 |
|
| 27 |
/* -- Iterative template routine -- |
| 28 |
* Univ. of Tennessee and Oak Ridge National Laboratory |
| 29 |
* October 1, 1993 |
| 30 |
* Details of this algorithm are described in "Templates for the |
| 31 |
* Solution of Linear Systems: Building Blocks for Iterative |
| 32 |
* Methods", Barrett, Berry, Chan, Demmel, Donato, Dongarra, |
| 33 |
* Eijkhout, Pozo, Romine, and van der Vorst, SIAM Publications, |
| 34 |
* 1993. (ftp netlib2.cs.utk.edu; cd linalg; get templates.ps). |
| 35 |
* |
| 36 |
* Purpose |
| 37 |
* ======= |
| 38 |
* |
| 39 |
* BICGSTAB solves the linear system A*x = b using the |
| 40 |
* BiConjugate Gradient Stabilized iterative method with |
| 41 |
* preconditioning. |
| 42 |
* |
| 43 |
* Convergence test: norm( b - A*x )< TOL. |
| 44 |
* For other measures, see the above reference. |
| 45 |
* |
| 46 |
* Arguments |
| 47 |
* ========= |
| 48 |
* |
| 49 |
* A (input) |
| 50 |
* |
| 51 |
* R (input) DOUBLE PRECISION array, dimension N. |
| 52 |
* On entry, residual of inital guess X |
| 53 |
* |
| 54 |
* X (input/output) DOUBLE PRECISION array, dimension N. |
| 55 |
* On input, the initial guess. |
| 56 |
* |
| 57 |
* ITER (input/output) INT |
| 58 |
* On input, the maximum iterations to be performed. |
| 59 |
* On output, actual number of iterations performed. |
| 60 |
* |
| 61 |
* RESID (input/output) DOUBLE PRECISION |
| 62 |
* On input, the allowable convergence measure for |
| 63 |
* norm( b - A*x ) |
| 64 |
* On output, the final value of this measure. |
| 65 |
* |
| 66 |
* return value |
| 67 |
* |
| 68 |
* = SOLVER_NO_ERROR: Successful exit. Iterated approximate solution returned. |
| 69 |
* = SOLVER_MAXITER_REACHED |
| 70 |
* = SOLVER_INPUT_ERROR Illegal parameter: |
| 71 |
* = SOLVER_BREAKDOWN: If parameters RHO or OMEGA become smaller |
| 72 |
* = SOLVER_MEMORY_ERROR : If parameters RHO or OMEGA become smaller |
| 73 |
* |
| 74 |
* ============================================================== |
| 75 |
*/ |
| 76 |
|
| 77 |
err_t Paso_Solver_BiCGStab( |
| 78 |
Paso_SystemMatrix * A, |
| 79 |
double * r, |
| 80 |
double * x, |
| 81 |
dim_t *iter, |
| 82 |
double * tolerance, |
| 83 |
Paso_Performance* pp) { |
| 84 |
|
| 85 |
|
| 86 |
/* Local variables */ |
| 87 |
double *rtld=NULL,*p=NULL,*v=NULL,*t=NULL,*phat=NULL,*shat=NULL,*s=NULL, *buf1=NULL, *buf0=NULL; |
| 88 |
double beta,norm_of_residual,sum_1,sum_2,sum_3,sum_4,norm_of_residual_global; |
| 89 |
double alpha, omega, omegaNumtr, omegaDenumtr, rho, tol, rho1, loc_sum[2], sum[2]; |
| 90 |
dim_t num_iter=0,maxit,num_iter_global; |
| 91 |
dim_t i0; |
| 92 |
bool_t breakFlag=FALSE, maxIterFlag=FALSE, convergeFlag=FALSE; |
| 93 |
dim_t status = SOLVER_NO_ERROR; |
| 94 |
double *resid = tolerance; |
| 95 |
dim_t n = Paso_SystemMatrix_getTotalNumRows(A); |
| 96 |
|
| 97 |
/* Executable Statements */ |
| 98 |
|
| 99 |
/* allocate memory: */ |
| 100 |
rtld=TMPMEMALLOC(n,double); |
| 101 |
p=TMPMEMALLOC(n,double); |
| 102 |
v=TMPMEMALLOC(n,double); |
| 103 |
t=TMPMEMALLOC(n,double); |
| 104 |
phat=TMPMEMALLOC(n,double); |
| 105 |
shat=TMPMEMALLOC(n,double); |
| 106 |
s=TMPMEMALLOC(n,double); |
| 107 |
/* Test the input parameters. */ |
| 108 |
|
| 109 |
if (n < 0) { |
| 110 |
status = SOLVER_INPUT_ERROR; |
| 111 |
} else if (rtld==NULL || p==NULL || v==NULL || t==NULL || phat==NULL || shat==NULL || s==NULL) { |
| 112 |
status = SOLVER_MEMORY_ERROR; |
| 113 |
} else { |
| 114 |
|
| 115 |
/* now bicgstab starts : */ |
| 116 |
maxit = *iter; |
| 117 |
tol = *resid; |
| 118 |
|
| 119 |
#pragma omp parallel firstprivate(maxit,tol) \ |
| 120 |
private(rho,omega,num_iter,norm_of_residual,beta,alpha,rho1, convergeFlag,maxIterFlag,breakFlag) |
| 121 |
{ |
| 122 |
num_iter =0; |
| 123 |
convergeFlag=FALSE; |
| 124 |
maxIterFlag=FALSE; |
| 125 |
breakFlag=FALSE; |
| 126 |
|
| 127 |
/* initialize arrays */ |
| 128 |
|
| 129 |
#pragma omp for private(i0) schedule(static) |
| 130 |
for (i0 = 0; i0 < n; i0++) { |
| 131 |
rtld[i0]=0; |
| 132 |
p[i0]=0; |
| 133 |
v[i0]=0; |
| 134 |
t[i0]=0; |
| 135 |
phat[i0]=0; |
| 136 |
shat[i0]=0; |
| 137 |
} |
| 138 |
#pragma omp for private(i0) schedule(static) |
| 139 |
for (i0 = 0; i0 < n; i0++) rtld[i0] = r[i0]; |
| 140 |
|
| 141 |
/* Perform BiConjugate Gradient Stabilized iteration. */ |
| 142 |
|
| 143 |
L10: |
| 144 |
++(num_iter); |
| 145 |
#pragma omp barrier |
| 146 |
#pragma omp master |
| 147 |
{ |
| 148 |
sum_1 = 0; |
| 149 |
sum_2 = 0; |
| 150 |
sum_3 = 0; |
| 151 |
sum_4 = 0; |
| 152 |
omegaNumtr = 0.0; |
| 153 |
omegaDenumtr = 0.0; |
| 154 |
} |
| 155 |
#pragma omp barrier |
| 156 |
#pragma omp for private(i0) reduction(+:sum_1) schedule(static) |
| 157 |
for (i0 = 0; i0 < n; i0++) sum_1 += rtld[i0] * r[i0]; |
| 158 |
#ifdef PASO_MPI |
| 159 |
#pragma omp master |
| 160 |
{ |
| 161 |
loc_sum[0] = sum_1; |
| 162 |
MPI_Allreduce(loc_sum, &sum_1, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm); |
| 163 |
} |
| 164 |
#endif |
| 165 |
rho = sum_1; |
| 166 |
|
| 167 |
if (! (breakFlag = (ABS(rho) <= TOLERANCE_FOR_SCALARS))) { |
| 168 |
/* Compute vector P. */ |
| 169 |
|
| 170 |
if (num_iter > 1) { |
| 171 |
beta = rho / rho1 * (alpha / omega); |
| 172 |
#pragma omp for private(i0) schedule(static) |
| 173 |
for (i0 = 0; i0 < n; i0++) p[i0] = r[i0] + beta * (p[i0] - omega * v[i0]); |
| 174 |
} else { |
| 175 |
#pragma omp for private(i0) schedule(static) |
| 176 |
for (i0 = 0; i0 < n; i0++) p[i0] = r[i0]; |
| 177 |
} |
| 178 |
|
| 179 |
/* Compute direction adjusting vector PHAT and scalar ALPHA. */ |
| 180 |
|
| 181 |
Paso_Solver_solvePreconditioner(A,&phat[0], &p[0]); |
| 182 |
Paso_SystemMatrix_MatrixVector_CSR_OFFSET0(ONE, A, &phat[0],ZERO, &v[0]); |
| 183 |
|
| 184 |
#pragma omp for private(i0) reduction(+:sum_2) schedule(static) |
| 185 |
for (i0 = 0; i0 < n; i0++) sum_2 += rtld[i0] * v[i0]; |
| 186 |
#ifdef PASO_MPI |
| 187 |
#pragma omp master |
| 188 |
{ |
| 189 |
loc_sum[0] = sum_2; |
| 190 |
MPI_Allreduce(loc_sum, &sum_2, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm); |
| 191 |
} |
| 192 |
#endif |
| 193 |
if (! (breakFlag = (ABS(sum_2) <= TOLERANCE_FOR_SCALARS))) { |
| 194 |
alpha = rho / sum_2; |
| 195 |
|
| 196 |
#pragma omp for private(i0) reduction(+:sum_3) schedule(static) |
| 197 |
for (i0 = 0; i0 < n; i0++) { |
| 198 |
r[i0] -= alpha * v[i0]; |
| 199 |
s[i0] = r[i0]; |
| 200 |
sum_3 += s[i0] * s[i0]; |
| 201 |
} |
| 202 |
#ifdef PASO_MPI |
| 203 |
#pragma omp master |
| 204 |
{ |
| 205 |
loc_sum[0] = sum_3; |
| 206 |
MPI_Allreduce(loc_sum, &sum_3, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm); |
| 207 |
} |
| 208 |
#endif |
| 209 |
norm_of_residual = sqrt(sum_3); |
| 210 |
|
| 211 |
/* Early check for tolerance. */ |
| 212 |
if ( (convergeFlag = (norm_of_residual <= tol)) ) { |
| 213 |
#pragma omp for private(i0) schedule(static) |
| 214 |
for (i0 = 0; i0 < n; i0++) x[i0] += alpha * phat[i0]; |
| 215 |
maxIterFlag = FALSE; |
| 216 |
breakFlag = FALSE; |
| 217 |
} else { |
| 218 |
/* Compute stabilizer vector SHAT and scalar OMEGA. */ |
| 219 |
Paso_Solver_solvePreconditioner(A,&shat[0], &s[0]); |
| 220 |
Paso_SystemMatrix_MatrixVector_CSR_OFFSET0(ONE, A, &shat[0],ZERO,&t[0]); |
| 221 |
|
| 222 |
#pragma omp for private(i0) reduction(+:omegaNumtr,omegaDenumtr) schedule(static) |
| 223 |
for (i0 = 0; i0 < n; i0++) { |
| 224 |
omegaNumtr +=t[i0] * s[i0]; |
| 225 |
omegaDenumtr += t[i0] * t[i0]; |
| 226 |
} |
| 227 |
#ifdef PASO_MPI |
| 228 |
#pragma omp master |
| 229 |
{ |
| 230 |
loc_sum[0] = omegaNumtr; |
| 231 |
loc_sum[1] = omegaDenumtr; |
| 232 |
MPI_Allreduce(loc_sum, sum, 2, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm); |
| 233 |
omegaNumtr=sum[0]; |
| 234 |
omegaDenumtr=sum[1]; |
| 235 |
} |
| 236 |
#endif |
| 237 |
if (! (breakFlag = (ABS(omegaDenumtr) <= TOLERANCE_FOR_SCALARS))) { |
| 238 |
omega = omegaNumtr / omegaDenumtr; |
| 239 |
|
| 240 |
#pragma omp for private(i0) reduction(+:sum_4) schedule(static) |
| 241 |
for (i0 = 0; i0 < n; i0++) { |
| 242 |
x[i0] += alpha * phat[i0] + omega * shat[i0]; |
| 243 |
r[i0] = s[i0]-omega * t[i0]; |
| 244 |
sum_4 += r[i0] * r[i0]; |
| 245 |
} |
| 246 |
#ifdef PASO_MPI |
| 247 |
#pragma omp master |
| 248 |
{ |
| 249 |
loc_sum[0] = sum_4; |
| 250 |
MPI_Allreduce(loc_sum, &sum_4, 1, MPI_DOUBLE, MPI_SUM, A->mpi_info->comm); |
| 251 |
} |
| 252 |
#endif |
| 253 |
norm_of_residual = sqrt(sum_4); |
| 254 |
convergeFlag = norm_of_residual <= tol; |
| 255 |
maxIterFlag = num_iter == maxit; |
| 256 |
breakFlag = (ABS(omega) <= TOLERANCE_FOR_SCALARS); |
| 257 |
} |
| 258 |
} |
| 259 |
} |
| 260 |
if (!(convergeFlag || maxIterFlag || breakFlag)) { |
| 261 |
rho1 = rho; |
| 262 |
goto L10; |
| 263 |
} |
| 264 |
} |
| 265 |
/* end of iteration */ |
| 266 |
#pragma omp master |
| 267 |
{ |
| 268 |
num_iter_global=num_iter; |
| 269 |
norm_of_residual_global=norm_of_residual; |
| 270 |
if (maxIterFlag) { |
| 271 |
status = SOLVER_MAXITER_REACHED; |
| 272 |
} else if (breakFlag) { |
| 273 |
status = SOLVER_BREAKDOWN; |
| 274 |
} |
| 275 |
} |
| 276 |
} /* end of parallel region */ |
| 277 |
} |
| 278 |
TMPMEMFREE(rtld); |
| 279 |
TMPMEMFREE(p); |
| 280 |
TMPMEMFREE(v); |
| 281 |
TMPMEMFREE(t); |
| 282 |
TMPMEMFREE(phat); |
| 283 |
TMPMEMFREE(shat); |
| 284 |
TMPMEMFREE(s); |
| 285 |
*iter=num_iter_global; |
| 286 |
*resid=norm_of_residual_global; |
| 287 |
|
| 288 |
/* End of BICGSTAB */ |
| 289 |
return status; |
| 290 |
} |
Parent Directory
| 
Revision Log