| 526 |
|
|
| 527 |
mk_mpi_size() |
mk_mpi_size() |
| 528 |
{ |
{ |
| 529 |
# mk_mpi_size input_file output_file proc_Nb prefer_to_split_X |
# mk_mpi_size input_file output_file proc_Nb threads_Nb_X threads_Nb_Y |
| 530 |
# |
# |
| 531 |
# make new SIZE.h (=output_file) from SIZE.h_mpi (=input_file) |
# make new SIZE.h (=output_file) from SIZE.h_mpi (=input_file) |
| 532 |
# for an MPI build with no more than proc_Nb processors ; |
# for an MPI build with no more than proc_Nb processors ; |
| 533 |
# return the effective number of processors. |
# ensure that enough tiles per proc (nSx,nSy) remain for the given |
| 534 |
|
# number of threads (nTx,nTy) ; |
| 535 |
|
# return the effective number of processors. |
| 536 |
|
|
| 537 |
inp=$1 |
inp=$1 |
| 538 |
out=$2 |
out=$2 |
| 539 |
np=$3 |
np=$3 |
| 540 |
dirX=$4 |
tx=$4 |
| 541 |
|
ty=$5 |
| 542 |
tmp=TTT.$$ |
tmp=TTT.$$ |
| 543 |
|
|
| 544 |
|
# dirX : select with direction to favor in MPI process repartition |
| 545 |
|
# dirX=1 : prefer to put more proc in X direction |
| 546 |
|
# dirX=0 : prefer to put more proc in Y direction |
| 547 |
|
dirX=0 |
| 548 |
|
|
| 549 |
px=`grep "^ & *nPx *=" $inp | sed "s/^ & *nPx *= *//" | sed 's/, *$//'` |
px=`grep "^ & *nPx *=" $inp | sed "s/^ & *nPx *= *//" | sed 's/, *$//'` |
| 550 |
py=`grep "^ & *nPy *=" $inp | sed "s/^ & *nPy *= *//" | sed 's/, *$//'` |
py=`grep "^ & *nPy *=" $inp | sed "s/^ & *nPy *= *//" | sed 's/, *$//'` |
| 551 |
sx=`grep "^ & *nSx *=" $inp | sed "s/^ & *nSx *= *//" | sed 's/, *$//'` |
sx=`grep "^ & *nSx *=" $inp | sed "s/^ & *nSx *= *//" | sed 's/, *$//'` |
| 552 |
sy=`grep "^ & *nSy *=" $inp | sed "s/^ & *nSy *= *//" | sed 's/, *$//'` |
sy=`grep "^ & *nSy *=" $inp | sed "s/^ & *nSy *= *//" | sed 's/, *$//'` |
| 553 |
|
|
| 554 |
|
#- for each direction, assume # of threads is a multiple of total number of tiles |
| 555 |
|
nx=$px |
| 556 |
|
if [ `expr $sx % $tx` -ne 0 -a `expr $sx \* $px % $tx` -eq 0 ] ; then |
| 557 |
|
nx=`expr $sx \* $px / $tx` |
| 558 |
|
if [ $verbose -gt 1 ]; then |
| 559 |
|
echo " change px from $px to $nx to accommodate $tx threads" |
| 560 |
|
fi |
| 561 |
|
fi |
| 562 |
|
ny=$py |
| 563 |
|
if [ `expr $sy % $ty` -ne 0 -a `expr $sy \* $py % $ty` -eq 0 ] ; then |
| 564 |
|
ny=`expr $sy \* $py / $ty` |
| 565 |
|
if [ $verbose -gt 1 ]; then |
| 566 |
|
echo " change py from $py to $ny to accommodate $ty threads" |
| 567 |
|
fi |
| 568 |
|
fi |
| 569 |
#- find the largest divisor of input_file proc Nb, but not larger than $np |
#- find the largest divisor of input_file proc Nb, but not larger than $np |
| 570 |
pp=0 |
pp=0 |
| 571 |
i=1 |
i=1 |
| 572 |
while [ $i -le $px ] ; do |
while [ $i -le $nx ] ; do |
| 573 |
if [ `expr $px % $i` -eq 0 ] ; then |
if [ `expr $nx % $i` -eq 0 ] ; then |
| 574 |
j=1 |
j=1 |
| 575 |
while [ $j -le $py ] ; do |
while [ $j -le $ny ] ; do |
| 576 |
if [ `expr $py % $j` -eq 0 ] ; then |
if [ `expr $ny % $j` -eq 0 ] ; then |
| 577 |
ij=`expr $i \* $j` |
ij=`expr $i \* $j` |
| 578 |
if [ $ij -gt $pp ] ; then |
if [ $ij -gt $pp ] ; then |
| 579 |
flag=1 |
flag=1 |
| 1407 |
fi |
fi |
| 1408 |
|
|
| 1409 |
if test "x$MPI" != "x0" ; then |
if test "x$MPI" != "x0" ; then |
| 1410 |
prefer_X=0 |
ntx=1 ; nty=1 |
| 1411 |
if test "x$MULTI_THREAD" = "xt" ; then |
if test "x$MULTI_THREAD" = "xt" ; then |
| 1412 |
retv=`check_eedata $dir"/input/eedata.mth"` |
ff=$dir"/input/eedata.mth" |
| 1413 |
if test $retv = 1 ; then prefer_X=1 ; fi |
ntx=`grep "^ *nTx *=" $ff | tail -1 | sed 's/^ *nTx *= *//' | sed "s/, *$//"` |
| 1414 |
|
nty=`grep "^ *nTy *=" $ff | tail -1 | sed 's/^ *nTy *= *//' | sed "s/, *$//"` |
| 1415 |
fi |
fi |
| 1416 |
#- create new SIZE.h with no more than '$MPI' Procs |
#- create new SIZE.h with no more than '$MPI' Procs |
| 1417 |
mk_mpi_size $CODE_DIR"/SIZE.h_mpi" $BUILD_DIR"/tr_size.mpi" $MPI $prefer_X |
mk_mpi_size $CODE_DIR"/SIZE.h_mpi" $BUILD_DIR"/tr_size.mpi" $MPI $ntx $nty |
| 1418 |
LOC_NPROC=$? |
LOC_NPROC=$? |
| 1419 |
( cd $BUILD_DIR |
( cd $BUILD_DIR |
| 1420 |
if test -r SIZE.h.mpi ; then |
if test -r SIZE.h.mpi ; then |
Parent Directory
| 
Revision Log
| 
Revision Graph
| 
Patch