REQUIREMENTS:

The build process for CSDP uses make, so your system must have make

installed. The GNU make works quite well, but CSDP has also been

built with other versions of make. The make files are very simple.

In order to build CSDP, you will need an ANSI C compiler. The GNU C

compiler gcc works quite well, but the code has also been compiled

with Intel's icc, IBM's xlc, Sun's cc, and many other compilers.

Although CSDP itself is written in C, the BLAS/LAPACK libraries used

by CSDP were originally written in Fortran. Combining C and Fortran

is generally straight forward, but you may have to install a Fortran

compiler in order to get run time libraries needed by the Fortran code

in BLAS/LAPACK. In particular, on Linux systems using gcc and gfortran

you will link CSDP with the -lgfortran library.

Use of the CSDP library and the solver and theta functions requires

the BLAS library and a few routines from the LAPACK library. Using

BLAS and LAPACK routines that have been optimized for your computer is

critical to getting good performance from CSDP, since the code

typically spends nearly all of its time in these routines. On the

same hardware, it's not uncommon to find that optimized BLAS and

LAPACK routines are an order of magnitude faster than unoptimized BLAS

and LAPACK routines.

The original authors of BLAS and LAPACK have provided "reference

implementations" that are freely available but not very well

optimized. If your system has BLAS and LAPACK libraries in /usr/lib,

there's a good chance that these are the reference implementations.

Using them is an easy way to start, but you may well find that you

need faster routines to get acceptable performance.

Most hardware manufacturers have developed optimized BLAS/LAPACK

libraries for their systems, such as Apple's veclib for G5

systems, Intel's Math Kernel Library (MKL), AMD's Core Mathematical

Library (ACML), IBM's extended scientific subroutine library (ESSL),

and Sun's sunperf library. These typically have restrictive licenses and

may be expensive to purchase, but can provide very good performance.

An open source library that I recommend for use with CSDP is OpenBLAS.

See

http://www.openblas.net/

You will generally need to compile OpenBLAS from source rather than installing

a binary version that is not optimized for your system.

CSDP also includes a set of routines for interfacing CSDP to MATLAB

and Octave. This interface is entirely optional- CSDP works fine

without it. In order to use this interface, you will need MATLAB (6.5

or later) or Octave (2.9.5 or later.) The README file in the matlab

directory contains instructions for installing the matlab interface

and testing it on a sample problem.

DOWNLOADING CSDP:

The most current version of CSDP can be obtained using subversion from

the COIN-OR web site. To do this, first install the subversion tools,

then issue the command

svn co https://projects.coin-or.org/svn/Csdp/trunk

Or use a GUI to access the repository.

Alternatively, you can download a .tar or .zip archive of the source code

for the most recent stable release of CSDP from the project web site.

INSTALLING CSDP:

The following instructions assume that you're using a Linux system

with gcc, gfortran, LAPACK, and BLAS installed. The make files will

have to be altered for other systems, but the basic process of

building the software will be similar.

After you've downloaded the source code, unpack the tar archive if you

downloaded a .tar archive of CSDP, and then go into the csdp

directory. You will find subdirectories, lib, solver, theta,

example, and doc. These contain the CSDP library, the command line

solver, an example program that computes the Lovasz theta number of

a graph, and a small example program showing how to call CSDP. There

are Makefiles within these subdirectories.

There are two lines in the top level Makefile that will likely need to

be edited:

1. CFLAGS. The CFLAGS variable specifies command line arguments to the

C compiler. The default value of CFLAGS is appropriate for 64 bit Linux

systems using the gcc compiler. It is:

CFLAGS=-m64 -march=native -mtune=native -Ofast, -fopenmp -ansi -Wall -DBIT64

-DUSEOPENMP -DSETNUMTHREADS -DUSESIGTERM -DUSEGETTIME -I../include

The flags -m64 through -fopenmp tell gcc to produce optimized 64 bit

multithreaded code using the OpenMP features. The flags -ansi and -Wall

tell gcc to produce warnings for any code that vioates the ANSI C standard.

The -D... flags specify options within the code as described below.

The -I../include flag specifies the location of CSDP's .h files.

The -D flags that are recommended for CSDP on 64 bit Linux are

-DBIT64 Used on 64 bit systems

-DUSEOPENMP Used to build a parallel threaded version of CSDP

-DSETNUMTHREADS Use OpenMP to set the number of threads to use.

-DUSESIGTERM Terminate CSDP nicely after SIGTERM or other signals

-DUSEGETTIME Use ANSI C gettime() to get wall clock running time.

For Windows 64 bit systems using the mingw64 compiler, recommended CFLAGS

are:

CFLAGS=-m64 -march=native -mtune=native -Ofast, -fopenmp -ansi -Wall -DBIT64

-DUSEOPENMP -DSETNUMTHREADS -DUSEGETTIME -I../include

The only difference here is that -DUSESIGTERM is left out because Windows

doesn't support Linux/unix signals.

There are some flags that may rarely be useful on other systems:

-DCAPSBLAS if BLAS routine names are capitalized

-DCAPSLAPACK if LAPACK routine names are capitalized

-DNOUNDERBLAS if BLAS routine names have no underscore

-DNOUNDERLAPACK if LAPACK routine names have no underscore

-DHIDDENSTRLEN if Fortran string lengths are hidden arguments

2. LIBS. The LIBS= line of the Makefile specifies the location of the

CSDP library and the BLAS and LAPACK libraries. The distributed version

of the Makefile has values suitable for a 64 bit Linux system with BLAS

and LAPACK installed in the system library directories:

LIBS=-static -L../lib -lsdp -llapack -lblas -lm

-static produce a statically linked executable

-L../lib -lsdp libsdp.a is in the lib subdirectory

-llapack -lblas LAPACK and BLAS libraries

-lm C standard math library

In many cases it will be necessary to add the Fortran run time library:

-lgfortran

The LIBS variable will also have to be edited if your LAPACK and BLAS libraries

are not in the standard location or have different names.

For example, on my system, the OpenBLAS library is located in

/opt/OpenBLAS/lib/libopenblas.a. This OpenBLAS library contains both

LAPACK and BLAS routines, so I don't need to specify a spearate LAPACK

library. Thus I use the flags:

LIBS=-static -L../lib -L/opt/OpenBLAS/lib -lopenblas -lm

Once you've set the LIBS and CFLAGS variables in the top level Makefile,

issue the command

> make

to build CSDP. Make will go into the lib, solver, theta, and example

subdirectories and compile the C code.

If the build fails, it is important to start by identifying where the

build failed. Failures in building the CSDP library are extremely

rare. Most reported failures have occured in the solver and theta

directories. The most common problem in practice is that one or more

of the required libraries (blas or lapack) is missing. In that

case you must install the required library before continuing with the

installation of CSDP. If the build fails for some other reason, feel

free to contact the author for help- we're interested in learning about

problems on different systems.

If for some reason the build fails, it's a good idea to remove all

of the binaries before modifying the make files and rebuilding. To do

this, issue the command

> make clean

Once the build appears to be successful, you can test the code with

> make unitTest

This will run tests of the stand alone solver csdp and Lovasz theta

program theta. Compare the .out files produced by the tests with the

corresponding .correct files. The actual values will typically vary

because of small differences in floating point round-off, compiler

optimizations, and so on. However, the optimal objective values

should match to at least six digits, the relative primal and dual

infeasibilities should be smaller than 1.0e-7, and all DIMACS errors

should be smaller than 5.0e-7.

If either of the tests fail, please contact the author.

Once you're satisfied with tests, you can become root and issue the

command

> make install

This will copy csdp, rand_graph, complement, theta, and graphtoprob

into the /usr/local/bin directory. If you use the C shell, remember

to "rehash" so that the shell will know that these programs have been

added to the /usr/local/bin directory.

The matlab directory contains .m files that provide a matlab interface

to CSDP solver. To install these .m files they must be added to your

matlab path. This requires use of the path(...) command in

matlab. See '>help path' for instructions on adding a new directory to

your matlab path (you can put the .m files in any directory you wish,

then add that directory to your matlab path).