| 1 |
C $Header: /u/gcmpack/MITgcm/eesupp/src/global_sum_tile.F,v 1.1 2007/09/04 14:46:31 jmc Exp $ |
| 2 |
C $Name: $ |
| 3 |
|
| 4 |
C-- File global_sum_tile.F: Routines that perform global sum |
| 5 |
C on a tile array |
| 6 |
C Contents |
| 7 |
C o global_sum_tile_rl |
| 8 |
C o global_sum_tile_rs <- not yet coded |
| 9 |
#include "CPP_EEOPTIONS.h" |
| 10 |
|
| 11 |
CBOP |
| 12 |
C !ROUTINE: GLOBAL_SUM_TILE_RL |
| 13 |
|
| 14 |
C !INTERFACE: |
| 15 |
SUBROUTINE GLOBAL_SUM_TILE_RL( |
| 16 |
I phiTile, |
| 17 |
O sumPhi, |
| 18 |
I myThid ) |
| 19 |
IMPLICIT NONE |
| 20 |
C !DESCRIPTION: |
| 21 |
C *==========================================================* |
| 22 |
C | SUBROUTINE GLOBAL\_SUM\_TILE\_RL |
| 23 |
C | o Handle sum for _RL data. |
| 24 |
C *==========================================================* |
| 25 |
C | Apply sum on an array of one value per tile |
| 26 |
C | and operate over all tiles & all the processes. |
| 27 |
C *==========================================================* |
| 28 |
|
| 29 |
C !USES: |
| 30 |
C == Global data == |
| 31 |
#include "SIZE.h" |
| 32 |
#include "EEPARAMS.h" |
| 33 |
#include "EESUPPORT.h" |
| 34 |
#include "GLOBAL_SUM.h" |
| 35 |
|
| 36 |
C !INPUT/OUTPUT PARAMETERS: |
| 37 |
C == Routine arguments == |
| 38 |
C phiTile :: Input array with one value per tile |
| 39 |
C sumPhi :: Result of sum. |
| 40 |
C myThid :: My thread id. |
| 41 |
_RL phiTile(nSx,nSy) |
| 42 |
_RL sumPhi |
| 43 |
INTEGER myThid |
| 44 |
|
| 45 |
C !LOCAL VARIABLES: |
| 46 |
C == Local variables == |
| 47 |
C bi,bj :: Loop counters |
| 48 |
C mpiRC - MPI return code |
| 49 |
C- type declaration of: sumMyPr, sumAllP, localBuf and shareBufGSR8 : |
| 50 |
C all 4 needs to have the same length as MPI_DOUBLE_PRECISION |
| 51 |
INTEGER bi,bj |
| 52 |
#ifdef ALLOW_USE_MPI |
| 53 |
#ifdef GLOBAL_SUM_SEND_RECV |
| 54 |
INTEGER biG, bjG, npe |
| 55 |
INTEGER lbuff, idest, itag, ready_to_receive |
| 56 |
INTEGER istatus(MPI_STATUS_SIZE), ierr |
| 57 |
Real*8 localBuf (nSx,nSy) |
| 58 |
Real*8 globalBuf(nSx*nPx,nSy*nPy) |
| 59 |
#else |
| 60 |
INTEGER mpiRC |
| 61 |
Real*8 sumMyPr |
| 62 |
#endif |
| 63 |
#else /* ALLOW_USE_MPI */ |
| 64 |
Real*8 sumMyPr |
| 65 |
#endif /* ALLOW_USE_MPI */ |
| 66 |
Real*8 sumAllP |
| 67 |
CEOP |
| 68 |
|
| 69 |
C this barrier is not necessary: |
| 70 |
c CALL BAR2( myThid ) |
| 71 |
|
| 72 |
C-- write local sum into shared-buffer array |
| 73 |
DO bj = myByLo(myThid), myByHi(myThid) |
| 74 |
DO bi = myBxLo(myThid), myBxHi(myThid) |
| 75 |
shareBufGSR8(bi,bj) = phiTile(bi,bj) |
| 76 |
ENDDO |
| 77 |
ENDDO |
| 78 |
|
| 79 |
C-- Master thread cannot start until everyone is ready: |
| 80 |
CALL BAR2( myThid ) |
| 81 |
_BEGIN_MASTER( myThid ) |
| 82 |
|
| 83 |
#if (defined (GLOBAL_SUM_SEND_RECV) && defined (ALLOW_USE_MPI) ) |
| 84 |
|
| 85 |
lbuff = nSx*nSy |
| 86 |
idest = 0 |
| 87 |
itag = 0 |
| 88 |
ready_to_receive = 0 |
| 89 |
|
| 90 |
IF ( mpiMyId.NE.0 ) THEN |
| 91 |
|
| 92 |
C-- All proceses except 0 wait to be polled then send local array |
| 93 |
#ifndef DISABLE_MPI_READY_TO_RECEIVE |
| 94 |
CALL MPI_RECV (ready_to_receive, 1, MPI_INTEGER, |
| 95 |
& idest, itag, MPI_COMM_MODEL, istatus, ierr) |
| 96 |
#endif |
| 97 |
CALL MPI_SEND (shareBufGSR8, lbuff, MPI_DOUBLE_PRECISION, |
| 98 |
& idest, itag, MPI_COMM_MODEL, ierr) |
| 99 |
|
| 100 |
C-- All proceses except 0 receive result from process 0 |
| 101 |
CALL MPI_RECV (sumAllP, 1, MPI_DOUBLE_PRECISION, |
| 102 |
& idest, itag, MPI_COMM_MODEL, istatus, ierr) |
| 103 |
|
| 104 |
ELSE |
| 105 |
|
| 106 |
C-- Process 0 fills-in its local data |
| 107 |
npe = 0 |
| 108 |
DO bj=1,nSy |
| 109 |
DO bi=1,nSx |
| 110 |
biG = (mpi_myXGlobalLo(npe+1)-1)/sNx+bi |
| 111 |
bjG = (mpi_myYGlobalLo(npe+1)-1)/sNy+bj |
| 112 |
globalBuf(biG,bjG) = shareBufGSR8(bi,bj) |
| 113 |
ENDDO |
| 114 |
ENDDO |
| 115 |
|
| 116 |
C-- Process 0 polls and receives data from each process in turn |
| 117 |
DO npe = 1, numberOfProcs-1 |
| 118 |
#ifndef DISABLE_MPI_READY_TO_RECEIVE |
| 119 |
CALL MPI_SEND (ready_to_receive, 1, MPI_INTEGER, |
| 120 |
& npe, itag, MPI_COMM_MODEL, ierr) |
| 121 |
#endif |
| 122 |
CALL MPI_RECV (localBuf, lbuff, MPI_DOUBLE_PRECISION, |
| 123 |
& npe, itag, MPI_COMM_MODEL, istatus, ierr) |
| 124 |
|
| 125 |
C-- Process 0 gathers the local arrays into a global array. |
| 126 |
DO bj=1,nSy |
| 127 |
DO bi=1,nSx |
| 128 |
biG = (mpi_myXGlobalLo(npe+1)-1)/sNx+bi |
| 129 |
bjG = (mpi_myYGlobalLo(npe+1)-1)/sNy+bj |
| 130 |
globalBuf(biG,bjG) = localBuf(bi,bj) |
| 131 |
ENDDO |
| 132 |
ENDDO |
| 133 |
ENDDO |
| 134 |
|
| 135 |
C-- Sum over all tiles: |
| 136 |
sumAllP = 0. |
| 137 |
DO bjG = 1,nSy*nPy |
| 138 |
DO biG = 1,nSx*nPx |
| 139 |
sumAllP = sumAllP + globalBuf(biG,bjG) |
| 140 |
ENDDO |
| 141 |
ENDDO |
| 142 |
|
| 143 |
C-- Process 0 sends result to all other processes |
| 144 |
lbuff = 1 |
| 145 |
DO npe = 1, numberOfProcs-1 |
| 146 |
CALL MPI_SEND (sumAllP, 1, MPI_DOUBLE_PRECISION, |
| 147 |
& npe, itag, MPI_COMM_MODEL, ierr) |
| 148 |
ENDDO |
| 149 |
|
| 150 |
|
| 151 |
ENDIF |
| 152 |
|
| 153 |
#else /* not (GLOBAL_SUM_SEND_RECV & ALLOW_USE_MPI) */ |
| 154 |
|
| 155 |
C-- Sum over all tiles (of the same process) first |
| 156 |
sumMyPr = 0. |
| 157 |
DO bj = 1,nSy |
| 158 |
DO bi = 1,nSx |
| 159 |
sumMyPr = sumMyPr + shareBufGSR8(bi,bj) |
| 160 |
ENDDO |
| 161 |
ENDDO |
| 162 |
|
| 163 |
C in case MPI is not used: |
| 164 |
sumAllP = sumMyPr |
| 165 |
|
| 166 |
#ifdef ALLOW_USE_MPI |
| 167 |
#ifndef ALWAYS_USE_MPI |
| 168 |
IF ( usingMPI ) THEN |
| 169 |
#endif |
| 170 |
CALL MPI_Allreduce(sumMyPr,sumAllP,1,MPI_DOUBLE_PRECISION, |
| 171 |
& MPI_SUM,MPI_COMM_MODEL,mpiRC) |
| 172 |
#ifndef ALWAYS_USE_MPI |
| 173 |
ENDIF |
| 174 |
#endif |
| 175 |
#endif /* ALLOW_USE_MPI */ |
| 176 |
|
| 177 |
#endif /* not (GLOBAL_SUM_SEND_RECV & ALLOW_USE_MPI) */ |
| 178 |
|
| 179 |
C-- Write solution to shared buffer (all threads can see it) |
| 180 |
shareBufGSR8(1,1) = sumAllP |
| 181 |
|
| 182 |
_END_MASTER( myThid ) |
| 183 |
C-- Everyone wait for Master thread to be ready |
| 184 |
CALL BAR2( myThid ) |
| 185 |
|
| 186 |
C-- set result for every threads |
| 187 |
sumPhi = shareBufGSR8(1,1) |
| 188 |
|
| 189 |
C- Need a barrier here to prevent thread 1 to modify shareBufGSR8(1,1) |
| 190 |
C (as it would in the following call to this S/R) before all threads get |
| 191 |
C their global-sum result out. |
| 192 |
CALL BAR2( myThid ) |
| 193 |
|
| 194 |
RETURN |
| 195 |
END |
Parent Directory
| 
Revision Log
| 
Revision Graph