| 1 |
subroutine kfilt_new(A,z,H,ns,no,varmes,nrsamp) |
| 2 |
|
| 3 |
c*** A is the matrix of ensemble forecasts, z is the observation, |
| 4 |
c*** H is the observation operator, ns is the size of the model |
| 5 |
c*** state, no is the number of observation locations, varmes is |
| 6 |
c*** the observational uncertainty, and nrsamp is the ensemble size. |
| 7 |
|
| 8 |
include 'mpif.h' |
| 9 |
#include "parallel.h" |
| 10 |
|
| 11 |
integer ns, no, nrsamp, MASTER, FROM_MASTER, FROM_WORKER |
| 12 |
parameter (MASTER = 0) |
| 13 |
parameter (FROM_MASTER = 1) |
| 14 |
parameter (FROM_WORKER = 2) |
| 15 |
|
| 16 |
integer taskid, ierr, numtasks, numworkers, snumworkers |
| 17 |
integer source, dest, mtype, num, avenum, extra, offset |
| 18 |
integer ijim |
| 19 |
integer status(MPI_STATUS_SIZE) |
| 20 |
|
| 21 |
real, intent(inout) :: A(ns,nrsamp) |
| 22 |
real, intent(in) :: z(no) |
| 23 |
|
| 24 |
integer i,j,m,mm,info,idum,ibawk |
| 25 |
real rcond |
| 26 |
logical zlz,svd,lud |
| 27 |
real H(no,ns) |
| 28 |
real vave(ns), vvar(ns), varmes(no) |
| 29 |
real temp |
| 30 |
|
| 31 |
real, allocatable :: K(:,:) |
| 32 |
real, allocatable :: RR(:,:) |
| 33 |
real, allocatable :: RRR(:,:) |
| 34 |
real, allocatable :: w(:) |
| 35 |
real, allocatable :: tmp(:,:) |
| 36 |
real, allocatable :: d(:) |
| 37 |
real, allocatable :: pd(:,:) |
| 38 |
real, allocatable :: transA(:,:) |
| 39 |
real, allocatable :: transtmp(:,:) |
| 40 |
real, allocatable :: tmp2(:,:) |
| 41 |
|
| 42 |
zlz=.true. |
| 43 |
svd=.false. |
| 44 |
lud=.false. |
| 45 |
|
| 46 |
allocate (K(no,ns), RR(no,no), RRR(no,no)) |
| 47 |
allocate (transA(nrsamp,ns),transtmp(nrsamp,no)) |
| 48 |
allocate (w(ns), tmp(no,nrsamp)) |
| 49 |
allocate (d(no), pd(1,ns), tmp2(1,ns)) |
| 50 |
|
| 51 |
call MPI_COMM_RANK(MPI_COMM_WORLD,taskid,ierr) |
| 52 |
call MPI_COMM_SIZE(MPI_COMM_WORLD,numtasks,ierr) |
| 53 |
|
| 54 |
numworkers = numtasks-1 |
| 55 |
temp=sqrt(float(numworkers)) |
| 56 |
snumworkers=int(temp) |
| 57 |
|
| 58 |
if(taskid.eq.MASTER)then |
| 59 |
|
| 60 |
c*** initialising the observational error covariance matrix and |
| 61 |
c*** the projector from model space to observational space. |
| 62 |
RR=0.0 |
| 63 |
|
| 64 |
c*** define obs error covariance matrix as a diagonal matrix |
| 65 |
do m=1,no |
| 66 |
RR(m,m)=varmes(m) |
| 67 |
enddo |
| 68 |
|
| 69 |
c*** calculate ensemble mean |
| 70 |
call ranmean2(A,w,ns,nrsamp) |
| 71 |
|
| 72 |
c*** remove ensemble mean from the ensemble members |
| 73 |
do i=1,nrsamp |
| 74 |
A(:,i)=A(:,i)-w(:) |
| 75 |
enddo |
| 76 |
|
| 77 |
endif ! MASTER if |
| 78 |
|
| 79 |
c*** tmp takes on the product of the observation projector and |
| 80 |
c*** the ensemble anomalies. |
| 81 |
#ifdef pmatmul |
| 82 |
if(taskid.eq.MASTER) write(6,*) 'first pmm' |
| 83 |
call pmm(H,A,tmp,no,ns,nrsamp) |
| 84 |
#else |
| 85 |
if(taskid.eq.MASTER)then |
| 86 |
write(6,*) 'first matmul' |
| 87 |
tmp=matmul(H,A) |
| 88 |
write(6,*) 'completed first matmul' |
| 89 |
endif ! MASTER if |
| 90 |
#endif |
| 91 |
|
| 92 |
c*** K takes on the value of H(A-ABAR)(A-ABAR)^T |
| 93 |
#ifdef pmatmul |
| 94 |
if(taskid.eq.MASTER)then |
| 95 |
transA=transpose(A) |
| 96 |
endif ! MASTER if |
| 97 |
|
| 98 |
if(taskid.eq.MASTER) write(6,*) 'second pmm' |
| 99 |
call pmm(tmp,transA,K,no,nrsamp,ns) |
| 100 |
#else |
| 101 |
if(taskid.eq.MASTER)then |
| 102 |
write(6,*) 'second matmul' |
| 103 |
K=matmul(tmp,transpose(A)) |
| 104 |
write(6,*) 'completed second matmul' |
| 105 |
endif ! MASTER if |
| 106 |
#endif |
| 107 |
|
| 108 |
if(taskid.eq.MASTER)then |
| 109 |
|
| 110 |
c*** K is what is called B in eqn (6) |
| 111 |
K=(1.0/(float(nrsamp)-1.0))*K |
| 112 |
|
| 113 |
c*** restore A by adding back ensemble mean value. |
| 114 |
do i=1,nrsamp |
| 115 |
A(:,i)=A(:,i)+w(:) |
| 116 |
enddo |
| 117 |
|
| 118 |
c*** RR is obs error cov matrix, and the forecast error cov is added to |
| 119 |
c*** it, making it the first half of the right hand side of equation (8). |
| 120 |
c*** if one was to wipe out spurious long distance correlations, it would |
| 121 |
c*** happen here. |
| 122 |
#ifdef pmatmul |
| 123 |
transtmp=transpose(tmp) |
| 124 |
|
| 125 |
endif ! MASTER if |
| 126 |
|
| 127 |
if(taskid.eq.MASTER) write(6,*) 'third pmm' |
| 128 |
call pmm(tmp,transtmp,RRR,no,nrsamp,no) |
| 129 |
|
| 130 |
if(taskid.eq.MASTER)then |
| 131 |
RR=RR+(1./(float(nrsamp)-1.))*RRR |
| 132 |
endif ! MASTER if |
| 133 |
#else |
| 134 |
write(6,*) 'third matmul' |
| 135 |
RR=RR+(1./(float(nrsamp)-1.))*matmul(tmp,transpose(tmp)) |
| 136 |
write(6,*) 'completed third matmul' |
| 137 |
endif ! MASTER if |
| 138 |
#endif |
| 139 |
|
| 140 |
if(taskid.eq.MASTER)then |
| 141 |
|
| 142 |
c*** standard solver for coefficients beta |
| 143 |
call spoco(RR,no,no,rcond,w,info) |
| 144 |
|
| 145 |
|
| 146 |
#ifdef pcorrect |
| 147 |
|
| 148 |
endif ! MASTER if |
| 149 |
|
| 150 |
if(taskid.eq.MASTER) then |
| 151 |
|
| 152 |
write(6,*) 'updating ensemble members' |
| 153 |
|
| 154 |
c*** figure out how many ens members go to each processor and look for extras |
| 155 |
avenum = nrsamp/numworkers |
| 156 |
extra = mod(nrsamp, numworkers) |
| 157 |
offset = 1 |
| 158 |
mtype = FROM_MASTER |
| 159 |
|
| 160 |
c*** loop over the worker nodes |
| 161 |
do dest=1, numworkers |
| 162 |
|
| 163 |
c*** pad in the extra columns until they've all been accounted for |
| 164 |
if (dest .le. extra) then |
| 165 |
num = avenum + 1 |
| 166 |
else |
| 167 |
num = avenum |
| 168 |
endif |
| 169 |
|
| 170 |
c*** send worker nodes the offset, number of cols, a and b |
| 171 |
call MPI_SEND(offset, 1, MPI_INTEGER, dest, mtype, |
| 172 |
> MPI_COMM_WORLD, ierr) |
| 173 |
call MPI_SEND(num, 1, MPI_INTEGER, dest, mtype, |
| 174 |
> MPI_COMM_WORLD, ierr) |
| 175 |
call MPI_SEND(z, no, MPI_REAL, dest, mtype, |
| 176 |
> MPI_COMM_WORLD, ierr) |
| 177 |
call MPI_SEND(varmes, no, MPI_REAL, dest, |
| 178 |
> mtype, MPI_COMM_WORLD, ierr) |
| 179 |
call MPI_SEND(RR, no*no, MPI_REAL, dest, |
| 180 |
> mtype, MPI_COMM_WORLD, ierr) |
| 181 |
call MPI_SEND(H, no*ns, MPI_REAL, dest, |
| 182 |
> mtype, MPI_COMM_WORLD, ierr) |
| 183 |
call MPI_SEND(K, no*ns, MPI_REAL, dest, |
| 184 |
> mtype, MPI_COMM_WORLD, ierr) |
| 185 |
call MPI_SEND(A(1,offset), num*ns, MPI_REAL, |
| 186 |
> dest, mtype, MPI_COMM_WORLD, ierr) |
| 187 |
|
| 188 |
c*** increment the offset |
| 189 |
offset = offset + num |
| 190 |
|
| 191 |
enddo |
| 192 |
|
| 193 |
c*** receive results from worker tasks |
| 194 |
mtype = FROM_WORKER |
| 195 |
do i=1, numworkers |
| 196 |
source = i |
| 197 |
call MPI_RECV(offset, 1, MPI_INTEGER, source, |
| 198 |
> mtype, MPI_COMM_WORLD, status, ierr) |
| 199 |
call MPI_RECV(num, 1, MPI_INTEGER, source, |
| 200 |
> mtype, MPI_COMM_WORLD, status, ierr) |
| 201 |
call MPI_RECV(A(1,offset), num*ns, MPI_REAL, source, |
| 202 |
> mtype, MPI_COMM_WORLD, status, ierr) |
| 203 |
enddo |
| 204 |
|
| 205 |
endif ! end of MASTER bits |
| 206 |
|
| 207 |
if(taskid.gt.MASTER)then |
| 208 |
|
| 209 |
c*** receive matrix data from master task |
| 210 |
mtype = FROM_MASTER |
| 211 |
call MPI_RECV(offset, 1, MPI_INTEGER, MASTER, |
| 212 |
> mtype, MPI_COMM_WORLD, status, ierr) |
| 213 |
call MPI_RECV(num, 1, MPI_INTEGER, MASTER, |
| 214 |
> mtype, MPI_COMM_WORLD, status, ierr) |
| 215 |
call MPI_RECV(z, no, MPI_REAL, MASTER, mtype, |
| 216 |
> MPI_COMM_WORLD, status, ierr) |
| 217 |
call MPI_RECV(varmes, no, MPI_REAL, MASTER, mtype, |
| 218 |
> MPI_COMM_WORLD, status, ierr) |
| 219 |
call MPI_RECV(RR, no*no, MPI_REAL, MASTER, mtype, |
| 220 |
> MPI_COMM_WORLD, status, ierr) |
| 221 |
call MPI_RECV(H, no*ns, MPI_REAL, MASTER, mtype, |
| 222 |
> MPI_COMM_WORLD, status, ierr) |
| 223 |
call MPI_RECV(K, no*ns, MPI_REAL, MASTER, mtype, |
| 224 |
> MPI_COMM_WORLD, status, ierr) |
| 225 |
call MPI_RECV(A, num*ns, MPI_REAL, MASTER, mtype, |
| 226 |
> MPI_COMM_WORLD, status, ierr) |
| 227 |
|
| 228 |
|
| 229 |
do j=1,num |
| 230 |
call random2(d,no) |
| 231 |
d=z+sqrt(varmes)*d-matmul(H,A(:,j)) |
| 232 |
call sposl(RR,no,no,d) |
| 233 |
A(:,j)=A(:,j)+matmul(d,K) |
| 234 |
enddo |
| 235 |
|
| 236 |
c*** send results back to master task |
| 237 |
mtype = FROM_WORKER |
| 238 |
call MPI_SEND(offset, 1, MPI_INTEGER, MASTER, mtype, |
| 239 |
> MPI_COMM_WORLD, ierr) |
| 240 |
call MPI_SEND(num, 1, MPI_INTEGER, MASTER, mtype, |
| 241 |
> MPI_COMM_WORLD, ierr) |
| 242 |
call MPI_SEND(A, num*ns, MPI_REAL, MASTER, mtype, |
| 243 |
> MPI_COMM_WORLD, ierr) |
| 244 |
|
| 245 |
endif ! end of SLAVE bits |
| 246 |
|
| 247 |
#else |
| 248 |
write(6,*) 'in correction bit' |
| 249 |
c*** correct each member of the ensemble |
| 250 |
do j=1,nrsamp |
| 251 |
|
| 252 |
call random2(d,no) |
| 253 |
d=z+sqrt(varmes)*d-matmul(H,A(:,j)) |
| 254 |
call sposl(RR,no,no,d) |
| 255 |
A(:,j)=A(:,j)+matmul(d,K) |
| 256 |
|
| 257 |
enddo !A has now been corrected by the data. |
| 258 |
|
| 259 |
write(6,*) 'end of correction bit' |
| 260 |
|
| 261 |
endif ! MASTER if |
| 262 |
|
| 263 |
#endif |
| 264 |
|
| 265 |
deallocate (K, RR, RRR) |
| 266 |
deallocate (transA, transtmp) |
| 267 |
deallocate (w, tmp) |
| 268 |
deallocate (d,pd,tmp2) |
| 269 |
|
| 270 |
return |
| 271 |
end |
| 272 |
|
| 273 |
|
| 274 |
|
| 275 |
|
| 276 |
|
Parent Directory
| 
Revision Log
| 
Revision Graph