Installation

If you are lucky to run one of the following operating systems, you can install fimex from a package repository:

Ubuntu Lucid packages are available since 0.28 at: https://launchpad.net/~heiko-klein/+archive/fimex

Ubuntu Lucid packages are available up to 0.25 at: https://launchpad.net/~heiko-klein/+archive/fimex

Fimex requires at least the following libraries to be installed for compilation:

• c99/c++ compiler
• libxml2 >= 2.5.0
• boost library >= 1.32
• proj-4 >= 4.4.9
• udunits 1.12.x or >= 2.1.x

To configure the different file formats it requires:

• NetCDF (netcdf-3 > 3.6)
• Felt (libmi or libfelt (bundled))
• grib_api (> 1.4)

The installation follows the usual autoconf rules, see 'INSTALL' for extensive information. Since Fimex makes some floating-point calculations in large loops, it is advisable to switch on SIMD/SSE operations in your compiler. On a Xeon machine with a x386 machine and gcc, I use the following flags (those are default for x86-64) (see microbenchmark of sse for more information)

Since fimex 0.21

CFLAGS='-O2 -mfpmath=sse -msse2 -ftree-vectorize -fno-math-errno' CXXFLAGS='-O2 -mfpmath=sse -msse2 -ftree-vectorize -fno-math-errno' ./configure --with-netcdf=/usr --with-udunits=/usr
make
make check
make install

In case of compilation against the netcdf4 library with hdf5 support, additional libraries needed by a static netcdf4 need to be added manually in LDFLAGS. Those can be retrieved from the nc-config program, e.g.

 $ /usr/local/netcdf/netcdf-4.1.1/bin/nc-config/nc-config --libs
-L/usr/local/netcdf/netcdf-4.1.1/lib -lnetcdf -L/usr/local/hdf/hdf5-1.8.5-patch1/lib -lhdf5_hl -lhdf5 -lz -lm -lgpfs
 $ export LDFLAGS='-L/usr/local/hdf/hdf5-1.8.5-patch1/lib -lhdf5_hl -lhdf5 -lz -lm -lgpfs'

The following optimizations have been tried with gcc 4.4 and kd-tree interpolation.

• -O2 general optimizations like loop-unrolling/inline
• -mfpmath=sse -msse2 on i386 computers, use math from sse-unit. The build-in FPU is not IEEE conform and performs very slowly (> factor 2) when calculating sqrt(nan). This is already the default on i386-64.
• -ftree-vectorize -fno-math-errno gives tiny performance gain through auto-vecotrization. -fno-math-errno is required to enable vectorization of intrinsic functions like sqrt.
• -fno-trapping-math -fno-signaling-nans fimex does not make use of trapping-math or signaling-nans, so this option can safely be switched on. Unfortunately, this does not give any performance gain.
• -O3 gives no visible performance gain. This should be tested again with newer compiler-versions.
• -fprofile-generate/-fprofile-use gives a performance gain of ~4%. Since it complicates the build, it is not used by default.

Fimex can be build with OpenMP parallelization support with the configure option –enable-openmp. The following operations are parallelized:

• interpolation: fill2d (good parallelization on z-axis)
• interpolation with coord_nearestneighbor (good parallelization in startup)

Due to a bug in libtool, the options required for linking again a parallelized libfimex are not set automatically and need be entered manually by the programmer using libfimex: https://lists.gnu.org/archive/html/bug-libtool/2011-03/msg00022.html

• Download the file with testdata: flth00.dat.gz
• Put this file into the test-catalogue
• gunzip the file
• run 'make check' (make check will work even without the testdata and skip a lot of tests)