1 |
C $Id: cg2d.F,v 1.2 2006/05/12 22:21:21 ce107 Exp $ |
2 |
SUBROUTINE CG2D |
3 |
C /==========================================================\ |
4 |
C | SUBROUTINE CG2D | |
5 |
C | o Two-dimensional grid problem conjugate-gradient | |
6 |
C | inverter (with preconditioner). | |
7 |
C |==========================================================| |
8 |
C | Con. grad is an iterative procedure for solving Ax = b. | |
9 |
C | It requires the A be symmetric. | |
10 |
C | This implementation assumes A is a five-diagonal | |
11 |
C | matrix of the form that arises in the discrete | |
12 |
C | representation of the del^2 operator in a | |
13 |
C | two-dimensional space. | |
14 |
C | Notes: | |
15 |
C | ====== | |
16 |
C | This implementation can support shared-memory | |
17 |
C | multi-threaded execution. In order to do this COMMON | |
18 |
C | blocks are used for many of the arrays - even ones that | |
19 |
C | are only used for intermedaite results. This design is | |
20 |
C | OK if you want to all the threads to collaborate on | |
21 |
C | solving the same problem. On the other hand if you want | |
22 |
C | the threads to solve several different problems | |
23 |
C | concurrently this implementation will not work. | |
24 |
C \==========================================================/ |
25 |
IMPLICIT NONE |
26 |
|
27 |
C === Global data === |
28 |
#include "SIZE.h" |
29 |
#include "EEPARAMS.h" |
30 |
#include "PARAMS.h" |
31 |
#include "CG2D.h" |
32 |
#if defined(USE_PAPI_FLOPS) || defined(USE_PAPI_FLIPS) |
33 |
#if defined(PAPI_PER_ITERATION) || defined(PAPI_PER_TIMESTEP) |
34 |
#include "PAPI.h" |
35 |
#endif |
36 |
#endif |
37 |
|
38 |
C === Routine arguments === |
39 |
C myThid - Thread on which I am working. |
40 |
INTEGER myThid |
41 |
|
42 |
C === Local variables ==== |
43 |
C actualIts - Number of iterations taken |
44 |
C actualResidual - residual |
45 |
C bi - Block index in X and Y. |
46 |
C bj |
47 |
C etaN - Used in computing search directions |
48 |
C etaNM1 suffix N and NM1 denote current and |
49 |
C beta previous iterations respectively. |
50 |
C alpha |
51 |
C sumRHS - Sum of right-hand-side. Sometimes this is a |
52 |
C useful debuggin/trouble shooting diagnostic. |
53 |
C For neumann problems sumRHS needs to be ~0. |
54 |
C or they converge at a non-zero residual. |
55 |
C err - Measure of residual of Ax - b, usually the norm. |
56 |
C I, J, N - Loop counters ( N counts CG iterations ) |
57 |
INTEGER actualIts |
58 |
INTEGER bi, bj |
59 |
INTEGER I, J, N |
60 |
#ifdef USE_MIXED_PRECISION |
61 |
REAL*8 actualResidual |
62 |
REAL*8 err |
63 |
REAL*8 errSum |
64 |
REAL*8 etaN |
65 |
REAL*8 etaNM1 |
66 |
REAL*8 etaNSum |
67 |
REAL*8 beta |
68 |
REAL*8 alpha |
69 |
REAL*8 alphaSum |
70 |
REAL*8 sumRHS |
71 |
REAL*8 temp |
72 |
#else |
73 |
Real actualResidual |
74 |
Real err |
75 |
Real errSum |
76 |
Real etaN |
77 |
Real etaNM1 |
78 |
Real etaNSum |
79 |
Real beta |
80 |
Real alpha |
81 |
Real alphaSum |
82 |
Real sumRHS |
83 |
Real temp |
84 |
#endif |
85 |
|
86 |
C-- Initialise inverter |
87 |
etaNM1 = 1. _d 0 |
88 |
|
89 |
C-- Initial residual calculation |
90 |
err = 0. _d 0 |
91 |
sumRHS = 0. _d 0 |
92 |
DO J=1,sNy |
93 |
DO I=1,sNx |
94 |
cg2d_s(I,J) = 0. _d 0 |
95 |
cg2d_r(I,J) = cg2d_b(I,J) - |
96 |
& ( aW2d(I ,J )*cg2d_x(I-1,J )+aW2d(I+1,J )*cg2d_x(I+1,J ) |
97 |
& +aS2d(I ,J )*cg2d_x(I ,J-1)+aS2d(I ,J+1)*cg2d_x(I ,J+1) |
98 |
& -aW2d(I ,J )*cg2d_x(I ,J )-aW2d(I+1,J )*cg2d_x(I ,J ) |
99 |
& -aS2d(I ,J )*cg2d_x(I ,J )-aS2d(I ,J+1)*cg2d_x(I ,J ) |
100 |
& ) |
101 |
err = err + cg2d_r(I,J)*cg2d_r(I,J) |
102 |
sumRHS = sumRHS + cg2d_b(I,J) |
103 |
ENDDO |
104 |
ENDDO |
105 |
CALL EXCH_XY_R8( cg2d_r ) |
106 |
CALL EXCH_XY_R8( cg2d_s ) |
107 |
CALL GSUM_R8( temp, err ) |
108 |
err = temp |
109 |
CALL GSUM_R8( temp, sumRHS ) |
110 |
sumRHS = temp |
111 |
|
112 |
actualIts = 0 |
113 |
actualResidual = SQRT(err) |
114 |
WRITE(6,*) ' CG2D iters, err = ', actualIts, actualResidual |
115 |
IF ( actualResidual .EQ. 0. _d 0) STOP 'ABNORMAL END: RESIDUAL 0' |
116 |
|
117 |
C >>>>>>>>>>>>>>> BEGIN SOLVER <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<< |
118 |
DO 10 N=1, cg2dMaxIters |
119 |
C-- Solve preconditioning equation and update |
120 |
C-- conjugate direction vector "s". |
121 |
etaN = 0. _d 0 |
122 |
DO J=1,sNy |
123 |
DO I=1,sNx |
124 |
cg2d_q(I,J) = |
125 |
& pW(I ,J )*cg2d_r(I-1,J )+pW(I+1,J )*cg2d_r(I+1,J ) |
126 |
& +pS(I ,J )*cg2d_r(I ,J-1)+pS(I ,J+1)*cg2d_r(I ,J+1) |
127 |
& +pC(I ,J )*cg2d_r(I ,J ) |
128 |
etaN = etaN+cg2d_q(I,J)*cg2d_r(I,J) |
129 |
ENDDO |
130 |
ENDDO |
131 |
|
132 |
CALL GSUM_R8( temp, etaN ) |
133 |
etaN = temp |
134 |
beta = etaN/etaNM1 |
135 |
etaNM1 = etaN |
136 |
|
137 |
DO J=1,sNy |
138 |
DO I=1,sNx |
139 |
cg2d_s(I,J) = cg2d_q(I,J) + beta*cg2d_s(I,J) |
140 |
ENDDO |
141 |
ENDDO |
142 |
|
143 |
C-- Do exchanges that require messages i.e. between |
144 |
C-- processes. |
145 |
CALL EXCH_XY_R8( cg2d_s ) |
146 |
|
147 |
C-- Ten extra exchanges |
148 |
#ifdef TEN_EXTRA_EXCHS |
149 |
DO J=1,10 |
150 |
CALL EXCH_XY_R8( cg2d_s ) |
151 |
ENDDO |
152 |
#endif |
153 |
|
154 |
C== Evaluate laplace operator on conjugate gradient vector |
155 |
C== q = A.s |
156 |
alpha = 0. _d 0 |
157 |
DO J=1,sNy |
158 |
DO I=1,sNx |
159 |
cg2d_q(I,J) = |
160 |
& aW2d(I ,J )*cg2d_s(I-1,J )+aW2d(I+1,J )*cg2d_s(I+1,J ) |
161 |
& +aS2d(I ,J )*cg2d_s(I ,J-1)+aS2d(I ,J+1)*cg2d_s(I ,J+1) |
162 |
& -aW2d(I ,J )*cg2d_s(I ,J )-aW2d(I+1,J )*cg2d_s(I ,J ) |
163 |
& -aS2d(I ,J )*cg2d_s(I ,J )-aS2d(I ,J+1)*cg2d_s(I ,J ) |
164 |
alpha = alpha+cg2d_s(I,J)*cg2d_q(I,J) |
165 |
ENDDO |
166 |
ENDDO |
167 |
CALL GSUM_R8( temp, alpha ) |
168 |
|
169 |
#ifdef HUNDRED_EXTRA_SUMS |
170 |
C-- Hundred extra global sums |
171 |
DO J=1,100 |
172 |
CALL GSUM_R8( temp, alpha ) |
173 |
ENDDO |
174 |
#endif |
175 |
|
176 |
alpha = temp |
177 |
alpha = etaN/alpha |
178 |
|
179 |
C== Update solution and residual vectors |
180 |
C Now compute "interior" points. |
181 |
err = 0. _d 0 |
182 |
DO J=1,sNy |
183 |
DO I=1,sNx |
184 |
cg2d_x(I,J)=cg2d_x(I,J)+alpha*cg2d_s(I,J) |
185 |
cg2d_r(I,J)=cg2d_r(I,J)-alpha*cg2d_q(I,J) |
186 |
err = err+cg2d_r(I,J)*cg2d_r(I,J) |
187 |
ENDDO |
188 |
ENDDO |
189 |
|
190 |
CALL GSUM_R8( temp, err ) |
191 |
|
192 |
err = temp |
193 |
err = SQRT(err) |
194 |
actualIts = N |
195 |
actualResidual = err |
196 |
#ifdef RESIDUAL_PER_ITERATION |
197 |
WRITE(6,*) ' CG2D iters, err = ', actualIts, actualResidual |
198 |
#endif |
199 |
IF ( err .LT. cg2dTargetResidual ) GOTO 11 |
200 |
CALL EXCH_XY_R8(cg2d_r ) |
201 |
#ifdef PAPI_PER_ITERATION |
202 |
#ifdef USE_PAPI_FLOPS |
203 |
call PAPIF_flops(real_time, proc_time, flpops, mflops, check) |
204 |
WRITE(6,'(F10.3,A7,F10.3,A37,I8)') |
205 |
$ mflops, ' user ', mflops*proc_time/real_time, |
206 |
$ ' wallclock Mflop/s during iteration ', N |
207 |
#else |
208 |
#ifdef USE_PAPI_FLIPS |
209 |
call PAPIF_flips(real_time, proc_time, flpops, mflops, check) |
210 |
WRITE(6,'(F10.3,A7,F10.3,A37,I8)') |
211 |
$ mflops, ' user ', mflops*proc_time/real_time, |
212 |
$ ' wallclock Mflip/s during iteration ', N |
213 |
#endif |
214 |
#endif |
215 |
#endif |
216 |
10 CONTINUE |
217 |
11 CONTINUE |
218 |
CALL EXCH_XY_R8(cg2d_x ) |
219 |
#ifdef PAPI_PER_TIMESTEP |
220 |
#ifdef USE_PAPI_FLOPS |
221 |
call PAPIF_flops(real_time, proc_time, flpops, mflops, check) |
222 |
WRITE(6,'(F10.3,A7,F10.3,A37,I8)') |
223 |
$ mflops, ' user ', mflops*proc_time/real_time, |
224 |
$ ' wallclock Mflop/s during iteration ', N |
225 |
#else |
226 |
#ifdef USE_PAPI_FLIPS |
227 |
call PAPIF_flips(real_time, proc_time, flpops, mflops, check) |
228 |
WRITE(6,'(F10.3,A7,F10.3,A37,I8)') |
229 |
$ mflops, ' user ', mflops*proc_time/real_time, |
230 |
$ ' wallclock Mflip/s during iteration ', N |
231 |
#endif |
232 |
#endif |
233 |
#endif |
234 |
WRITE(6,*) ' CG2D iters, err = ', actualIts, actualResidual |
235 |
|
236 |
C Calc Ax to check result |
237 |
DO J=1,sNy |
238 |
DO I=1,sNx |
239 |
cg2d_Ax(I,J) = |
240 |
& ( aW2d(I ,J )*cg2d_x(I-1,J )+aW2d(I+1,J )*cg2d_x(I+1,J ) |
241 |
& +aS2d(I ,J )*cg2d_x(I ,J-1)+aS2d(I ,J+1)*cg2d_x(I ,J+1) |
242 |
& -aW2d(I ,J )*cg2d_x(I ,J )-aW2d(I+1,J )*cg2d_x(I ,J ) |
243 |
& -aS2d(I ,J )*cg2d_x(I ,J )-aS2d(I ,J+1)*cg2d_x(I ,J ) |
244 |
& ) |
245 |
cg2d_r(I,J) = cg2d_b(I,J)-cg2d_Ax(I,J) |
246 |
ENDDO |
247 |
ENDDO |
248 |
CALL EXCH_XY_R8(cg2d_Ax ) |
249 |
CALL EXCH_XY_R8(cg2d_r ) |
250 |
|
251 |
END |