Difference between revisions of "Useful Makefiles"

From CUC3
Jump to navigation Jump to search
import>Cen1001
import>Jss43
Line 243: Line 243:
   
 
Having to set the c compiler for parallel compilation is not mentioned in the docs...
  
Having to set the c compiler for parallel compilation is not mentioned in the docs...
  +
This should also work with openMPI.
   
 
== mek-quake ==
  
== mek-quake ==

Revision as of 11:39, 24 July 2008

This page is a place to stick Makefiles or other config files for common codes that you've got working on the local compute servers

Tardis

cp2k with intel compilers

Setting the FORT_C_NAME variable to 'intel' helps cp2k's build system select the right compiler.

You need the mkl, mpi/mvapich/intel, blacs/mvapich/intel64, and scalapack/intel modules loaded 

# The following settings worked for:
# - AMD64 Opteron cluster
# - SUSE Linux 10.0 (x86_64)
# - Intel(R) Fortran Compiler for Intel(R) EM64T-based applications, Version 9.1.037
# - Intel(R) Cluster Math Kernel Library v7.2 for Linux
# - MVAPICH
# - BLACS and ScaLAPACK compiled for Intel
#
PERL     = perl
CC       = cc
CPP      = cpp
FC       = mpif90 -FR
LD       = mpif90
AR       = ar -r
DFLAGS   = -D__INTEL -D__FFTSG\
          -D__parallel -D__BLACS -D__SCALAPACK\
          -Dfftwnd_f77=fftwnd_f77_\
          -Dfftwnd_f77_one=fftwnd_f77_one_\
          -Dfftw3d_f77_create_plan=fftw3d_f77_create_plan_\
          -Dfftw2d_f77_create_plan=fftw2d_f77_create_plan_\
          -Dfftwnd_f77_destroy_plan=fftwnd_f77_destroy_plan_\
          -Dfftw_f77_create_plan=fftw_f77_create_plan_\
          -Dfftw_f77=fftw_f77_\
          -Dfftw_f77_destroy_plan=fftw_f77_destroy_plan_
CPPFLAGS = -traditional -C $(DFLAGS) -P
FCFLAGS  = $(DFLAGS) -O2
MKLPATH  = /usr/local/Cluster-Apps/intel/mkl/8.0/lib/em64t
LDFLAGS  = $(FCFLAGS) -i-static
LIBS     = \
          -L/usr/local/Cluster-Apps/scalapack/intel/lib64 -lscalapack \
          $(MKLPATH)/libmkl_lapack.a \
          -L/usr/local/Cluster-Apps/blacs/mvapich/intel/lib64 -lblacsF77init -lblacs \
          $(MKLPATH)/libmkl_em64t.a \
          $(MKLPATH)/libguide.a\
          -lpthread

OBJECTS_ARCHITECTURE = machine_intel.o

#           -D__FFTW\


Jochen's CPMD with Portland compilers and MVAPICH

You need the mpi/mvapich/pgi module but ACML comes in automatically with PGI. 

#----------------------------------------------------------------------------
# Makefile for cpmd.x (plane wave electronic calculation)
# Configuration: PGI-AMD64-MPI
# Creation of Makefile: Dec  5 2006
# on Linux tardis 2.6.15.1-clustervision-128_cvos #1 SMP Mon Sep 25 12:05:46 CEST 2006   x86_64 x86_64 x86_64 GNU/Linux
# Author: jb376
#----------------------------------------------------------------------------
#
SHELL = /bin/sh
#
#--------------- Default Configuration for PGI-AMD64-MPI ---------------
SRC  = .
DEST = .
BIN  = .
#QMMM_FLAGS = -D__QMECHCOUPL
#QMMM_LIBS  = -L. -lmm
FFLAGS = -r8 -pc=64 -Msignextend  
#LFLAGS = -Bstatic -L. -latlas_x86-64 $(QMMM_LIBS)
#LFLAGS = -Bstatic -L. -latlas_x86_64 $(QMMM_LIBS)
LFLAGS = -lacml $(QMMM_LIBS)
CFLAGS = 
CPP = /lib/cpp -P -C -traditional
#CPPFLAGS = -D__Linux -D__PGI -DLAPACK -DFFT_DEFAULT -DPOINTER8 -D__pgf90 \
#               -DPARALLEL -DMP_LIBRARY=__MPI -DMYRINET  
CPPFLAGS = -D__Linux -D__PGI -DLAPACK -DFFT_DEFAULT -DPOINTER8 -D__pgf90 \
              -DPARALLEL -DMP_LIBRARY=__MPI 
NOOPT_FLAG = 
CC = mpicc -O2 -Wall -m64
FC = mpif77 -c -fastsse -tp k8-64
LD = mpif77 -fastsse -tp k8-64
AR = 
#----------------------------------------------------------------------------

I had problems compiling CPMD v3.11.1 using mpif77: the compiler complained about some valid Fortran statements (e.g. append and cycle). Using mpif90 instead resolved this.--james 11:56, 8 August 2007 (BST)

The recent upgrade to tardis has changed how some modules work. In particular, mpicc now points to pgcc rather than gcc if the portland environment module is loaded. This is, Catherine and I think, the sane approach. The above CC options cause make to barf, as pgcc uses different flags to gcc. Change "-O2 -Wall -m64" to "-O2 -Minform=inform -pc=64" to give pgcc the equivalent options.--james 19:37, 7 March 2008 (GMT)

NAMD2 with Intel compilers

This one requires the openmpi/intel64 module

Step 1: charm++

./build charm++ mpi-linux-amd64 icc ifort

my src/arch/mpi-linux-amd64/conv-mach.sh 

CMK_REAL_COMPILER=`mpiCC -show 2>/dev/null | cut -d' ' -f1 `
case "$CMK_REAL_COMPILER" in
 g++) CMK_AMD64="-m64 -fPIC" ;;
esac

CMK_CPP_CHARM="/lib/cpp -P"
CMK_CPP_C="mpicc -E"
CMK_CC="mpicc $CMK_AMD64 "
CMK_CXX="mpiCC $CMK_AMD64 "
CMK_CXXPP="mpiCC -E $CMK_AMD64 "

#CMK_SYSLIBS="-lmpich" 
# -lmpich is not needed as we replace 'icc' with 'mpicc' in the cc-icc.sh file
CMK_SYSLIBS=" "
CMK_LIBS="-lckqt $CMK_SYSLIBS "
CMK_LD_LIBRARY_PATH="-Wl,-rpath,$CHARMLIBSO/"

CMK_NATIVE_CC="gcc $CMK_AMD64 "
CMK_NATIVE_LD="gcc $CMK_AMD64 "
CMK_NATIVE_CXX="g++ $CMK_AMD64 "
CMK_NATIVE_LDXX="g++ $CMK_AMD64 "
CMK_NATIVE_LIBS=""

# fortran compiler 
CMK_CF77="f77"
CMK_CF90="f90"
CMK_F90LIBS=" "
CMK_F77LIBS=" "
CMK_MOD_NAME_ALLCAPS=1
CMK_MOD_EXT="mod"
CMK_F90_USE_MODDIR=1
CMK_F90_MODINC="-p"

CMK_QT='generic64'
CMK_RANLIB="ranlib"

and src/arch/common/cc-icc.sh 

# Changed all the C/C++ compilers and linkers to the mpi compiler version s/icc/mpicc/ s/icpc/mpiCC/
CMK_CPP_C='mpicc -E '
CMK_CC="mpicc -fpic "
CMK_CXX="mpiCC -fpic "
CMK_CXXPP='mpiCC -E '
CMK_LD='mpicc -i_dynamic '
CMK_LDXX='mpiCC -i_dynamic '
CMK_LD_LIBRARY_PATH="-Wl,-rpath,$CHARMLIBSO/"
# The F90 needed changing to ifort and -fPIC adding
CMK_CF90='ifort -auto -fPIC '
CMK_CF90_FIXED="$CMK_CF90 -132 -FI "
CMK_NATIVE_F90="$CMK_CF90"
CMK_NATIVE_CC="$CMK_CC"
CMK_NATIVE_CXX="$CMK_CXX"
CMK_NATIVE_LD="$CMK_LD"
CMK_NATIVE_LDXX="$CMK_LDXX"
# I removed a bunch of bogus -L options pointing to an ancient and nonexistent ifc installation
CMK_F90LIBS='-lintrins -lIEPCF90 -lPEPCF90 -lF90 -lintrins -limf  '
CMK_MOD_NAME_ALLCAPS=1
CMK_MOD_EXT="mod"
CMK_F90_USE_MODDIR=""

Since tardis's IB stack was updated I can't get the charm++ built with Intel compilers to pass all the tests anymore. One built with gcc seems to do fine though. Load the environment/64-bit/openmpi/gnu64 module. Unpack fresh charm++ source and cd to it.

Edit src/arch/mpi-linux-amd64/conv-mach.sh and change the 

 CMK_SYSLIBS="-lmpich"

to 

CMK_SYSLIBS=" "

then 

./build charm++ mpi-linux-amd64 --no-shared -O -DCMK_OPTIMIZE

The best test (according to the NAMD people) is to cd tests/charm++/megatest; build and run that one. This does pass.

Step 2: NAMD

Load the environment/64-bit/openmpi/intel64 module 

./config tcl fftw Linux-amd64-MPI-icc
cd Linux-amd64-MPI-icc

my arch/Linux-amd64.tcl: 

TCLDIR=/usr
TCLINCL=-I$(TCLDIR)/include 
TCLLIB=-L$(TCLDIR)/lib -ltcl8.4 -ldl
TCLFLAGS=-DNAMD_TCL -DUSE_NON_CONST 
TCL=$(TCLINCL) $(TCLFLAGS)


my arch/Linux-amd64.fftw: 

FFTDIR=/usr/local/fftw2/intel/64/2.1.5
FFTINCL=-I$(FFTDIR)/include -I$(HOME)/fftw/include
FFTLIB=-L$(FFTDIR)/lib -L$(HOME)/fftw/lib -lsrfftw -lsfftw
FFTFLAGS=-DNAMD_FFTW
FFT=$(FFTINCL) $(FFTFLAGS)

my Make.charmm 

CHARMBASE = /usr/local/charm++/charm-5.9-openmpi-gcc

my Linux-amd64-MPI-icc.arch which owes a great deal to Jochen's below 

NAMD_ARCH = Linux-amd64
CHARMARCH = mpi-linux-amd64
FLOATOPTS = -ip -fno-rtti
CXX = mpiCC
CXXOPTS =  -i-static -static-libcxa -O2 -unroll $(FLOATOPTS)
CXXNOALIASOPTS = -O2 -unroll -fno-alias $(FLOATOPTS)
CC = mpicc
COPTS = -i-static -static-libcxa -O2 $(FLOATOPTS)

now 

make

Test it by doing an interactive qsub and 

mpirun ./namd2 src/alanin

Jochen's Linux-amd64-MPI-icc.arch (Vastly improves performance over the defaults): 

NAMD_ARCH = Linux-amd64
# If using the gcc compiled charm++ you want to uncomment the line immediately below this
# CHARMARCH = mpi-linux-amd64
# and comment out the one below this - CEN
CHARMARCH = mpi-linux-amd64-icc
FLOATOPTS = -fno-rtti
CXX = /usr/local/Cluster-Apps/ofed/1.0/mpi/intel/mvapich-0.9.7-mlx2.1.0/bin/mpicxx
# This is a little odd as -tpp6 is a Portland option - CEN
CXXOPTS = -tpp6 -pc64 -i-static -static-libcxa -O2 -unroll $(FLOATOPTS)
CXXNOALIASOPTS = -O2 -unroll -fno-alias $(FLOATOPTS)
CC = /usr/local/Cluster-Apps/ofed/1.0/mpi/intel/mvapich-0.9.7-mlx2.1.0/bin/mpicc
COPTS = -i-static -static-libcxa -O2 $(FLOATOPTS)

FFTW 2.1.5 with MVAPICH

This is not really well-documented elsewhere. FFTW 2.1.5 is used in (e.g.) CPMD as an alternative to the default FFT engine supplied. It is trivial to compile in serial (FFTW 3 is even easier, but sadly the parallel version is in alpha and incompatible with the widely-used FFTW2). On tardis, in fftw-2.1.5 directory formed by extracting the tarball:

  1. env CC=mpicc F77=mpif90 ./configure --prefix=`pwd` --enable-mpi --enable-sse
  2. make
  3. make install

Having to set the c compiler for parallel compilation is not mentioned in the docs... This should also work with openMPI.

mek-quake

GAMESS-US

This will probably work on clust and nimbus too as they are very similar machines. 

module add pgi64/7.1-6

You need to edit the scripts comp, compall, lked, and ddi/compddi. The TARGET in all of these should be linux64 and the fortran compiler (FORTRAN) set to pgf77 within the appropriate section. Leave the C compiler (CCOMP) as gcc. It will automatically link in the Portland copy of ACML so you will get a fast blas library. We did not need to change any other options.

In the DDI compilation set COMM to sockets .

Retrieved from "http://wikis.ch.cam.ac.uk/cuc3/wiki/index.php?title=Useful_Makefiles&oldid=1734"