Quick build¶

For users on 100% compatible systems

We assume that you have successfully and unpacked the package.

   cd sphinx
   ./setup
   ./configure --disable-debug
   make all
   make install

If this does not work, please read the rest of this page.

You may want to run the testbed after compilation.

Prerequisites¶

• C++ compiler
• automake
• libtool
• flex, bison
• Additional libraries (see also 3rd-party packages alter Link nach sxlib =: 3rd-party packages)
  • BLAS+LAPACK library
    • ATLAS+LAPACK or
    • mkl or
    • acml or
    • libGOTO or
    • ESSL
  • FFT library
    • FFTW (recommended)
    • ACML (we encountered many bugs...)
    • mkl
  • netCDF

CentOS 5.6	yum install gcc44-c++ gcc44-gfortran automake libtool flex bison
Fedora 15	yum install gcc-c++, gcc-gfortran, automake, libtool, flex, bison
Debian 6	apt-get install g++ gfortran automake libtool flex bison
MacOS X 10.6	X-Code (install from Apple installation DVD)

Please make sure that CPU throttling is deactivated when compiling the third-party libraries (specifically: ATLAS!).

Building¶

Depending on the scope of your work with S/PHI/nX please choose the best fitting building approach:

	Users	Testers	Developers
build S/PHI/nX release	once	once	frequently
build S/PHI/nX debug		once	frequently
build SxAccelerate			once - frequently
build numlibs			once - rarely

• On machines with N CPUs: you may use parallel make via make -j 2N

configure options (selection)¶

--disable-atlas --disable-lapack --enable-mkl --with-mklpath=...	use MKL for BLAS/LAPACK (support for selected versions)
--disable-atlas --disable-lapack --enable-acml	use ACML for BLAS/LAPACK
--enable-mpi	compile MPI version (experimental)
--enable-openmp	compile using openmp, try to use openmp numerical libraries (experimental)
--with-numlibs=...	all external numerical libraries can be found in these paths (/path/1/:/path/2/:/path/3/)
--without-numlibs	external numerical libraries can be found in standard paths or via specified LDFLAGS and CPPFLAGS
CXX=...	set C++ compiler
CXXFLAGS=...	set additional C++ compiler flags
LDFLAGS=...	set additional linker flags (see compiler/linker docu)
CPPFLAGS=...	set additional preparser flags (see compiler docu)
CPPFLAGS='-I...'	set additional search path for external includes

S/PHI/nX users¶

(Build S/PHI/nX only once to create an optimized executable.)

1. cd sphinx
2. ./setup
3. ./configure --disable-debug --prefix=/instpath/to/sphinx
4. make all install

If --prefix=/instpath/to/sphinx is omitted, SPHInX will be installed into the current directory.

S/PHI/nX test users¶

(Build S/PHI/nX in debug and release only once to create an optimized executable as well as a version that can generate a memory core file.)

1. cd sphinx
2. ./setup
3. ./configure --prefix=/instpath/to/sphinx --with-debug --with-release
4. make all install

Read more about being a S/PHI/nX alpha tester.

S/PHI/nX developers¶

Easy build instructions¶

• As a simple alternative, you may use the build instruction for test users.
• Properly configured build directories for debug and release mode will be found as subdirectories in sphinx/sxaccelerate/ and sphinx/, respectively.
• While developing, there is no need to install (make all install -> make all).
• You can compile all build directories by a top-level make, ...
• ... but also each build directory separately (but beware: when changes are made in sxaccelerate, you typically need to rebuild sxaccelerate first, and then sphinx!).
• For developers, we recommend to build the external libraries separately, as described in the following. Precompiled numlibs avoid repeated installation of the numlibs when top-level make is invoked, and can be reused for concurrent SPHInX versions.
• use parallel make (make -j ...)
  *use top-level sxmake wrapper instead of make

Advanced build instructions¶

1. cd sphinx
2. ./setup
3. Possibly switch to a separate build directory (on a local file system)

    mkdir -p /build/path/sphinx
    mkdir -p /build/path/sphinx/sxaccelerate/3rd-party
    cd /build/path/sphinx

If a build path is used, replace ./configure by the corresponding absolute filenames to the configure scripts in the source tree.

4. Building 3rd-party numeric libraries

1. cd sxaccelerate/3rd-party
2. ./configure --prefix=/instpath/to/numlibs
   where /instpath/to/numlibs refers to the folder where the compiled libraries will be installed in.
3. make all install
4. cd ../..
5. optionally clean up:
   rm -rf sxaccelerate/3rd-party

If --prefix=/instpath/to/numlibs is omitted, numlibs will be installed in sxaccelerate/3rd-party (NO CLEANUP!)

5. Compile S/PHI/nX

1. ./setup
2. ./configure --prefix=/instpath/to/sphinx --with-debug --with-release --with-numlibs=/instpath/to/numlibs
3. where /instpath/to/sxaccelerate is where the libraries and headers will be installed to.
4. make

S/PHI/nX core developers¶

(Suitable only for freaks)

More configure options¶

--enable-debug	compile in debug mode (default)
--disable-debug	compile in release mode
--enable-shared/--disable-shared	enable/disable shared objects/linkage (default: enable)
--enable-static/--disable-static	enable/disable static objects/linkage (default: disable)
--enable-numlibschecks	check the numlibs install

Advanced build instructions¶

1. Call ./setup in /src/path/to/sphinx
2. Build 3rd-party numeric libraries (see instructions for developers)
3. Building sxaccelerate

   mkdir -p /build/path/sxaccelerate
   cd /build/path/sxaccelerate
   /src/path/to/sphinx/sxaccelerate/src/configure [options] --with-numlibs=/instpath/to/numlibs

For available configure options, see configure --help. You may build debug and release versions side by side using the sxaccelerate/configure --with-debug --with-release - note the omission of the src before the configure in this case!

4. Building sphinx on top of an existing sxaccelerate build (or install)

    mkdir -p /build/path/sphinx
    cd /build/path/sphinx
    /src/path/to/sphinx/src/configure --with-sxaccelerate=/build/path/sxaccelerate

Note that you cannot override any configure options set in the sxaccelerate that affect compiler/linker flags or numlibs selection.