libMad

libMad is a metapackage used to compile a shared library which contains a c interface for MadNLP. It is currently very much a work in progress.

Installation

libMad can be either installed from source or from a release tarball.

Source build

Given you have the required build software installed (namly Julia 1.12+ and the JuliaC.jl package, then libMad can be installed from source using:

mkdir build && cd build && cmake -DCMAKE_INSTALL_PREFIX=<path/to/install/to> .. && make install

Release Versions

The best way to use libMad currently is by downloading the bundled releases available through the Github Releases page. These tarballs contain the full stand-alone bundle of libraries necessary to use libMad on the three supported platforms. In order to install them simply untar and add the path to the correct search path for your platform: PATH on windows, LD_LIBRARY_PATH on gnu/linux, and DYLD_LIBRARY_PATH on macos. If using libMad from CasADi >=v3.8.0 (which is currently in prerelease) this should be sufficient for the CasADi plugin loader to to load the madnlp and ccopt plugin correctly.

If you wish to use the GPU accelerated solvers in madnlp you need to do same with the -cuda release tarball. This includes the code required to use the CUDSSSolver functionality. You will need to provide system installs of the CUDA runtime, CUDA driver, and CUDA compiler binaries. The GPU accelerated functionality inherits the promises of CUDA.jl insofar that we test primarily on GNU/linux but it should work on windows as well.

Current known issues and their workarounds:

As it currently stands we require several environment variables to be set when using libMad, to handle several issues coming from upstream Julia packages.

General

The following are necessary for any interface:

1. JULIA_HSL_LIBRARY_PATH="path/to/hsl/lib" This is necessary to use the HSL linear system solvers: Ma*7Solver.

CUDA

1. JULIA_CUDA_USE_COMPAT="false" This is necessary to make sure that CUDA_Driver_jll does not attempt to fork a second julia process which fails as the binary does not exist.
2. JULIA_CUDSS_LIBRARY_PATH="path/to/cudss/lib" This may be necessary if the CUDA_Runtime_Discovery package cannot find the cuDSS libraries.

FFI in other languages

And then several language specific issues may occur if the shared library is loaded through, e.g., the CasADi interface for python or Matlab:

Python

If using the library through a python interface on GNU/linux it may be necessary to set the following preload: LD_PRELOAD="/path/to/libmad/bundle/julia/libssl.so". This is necessary as python may load a different version of libssl than the one that libMad was compiled against. This leads to a failure during dynamic loading and a crash.

Matlab

If using the library through a matlab interface on GNU/linux it is necessary to set the following preload:

1. LD_PRELOAD="/path/to/libmad/bundle/julia/libunwind.so" This is necessary because Matlab ships it's own modified version of libunwind which causes segfaults during backtrace generation in libMad. Be warned that this breaks the Matlab debugger, causing hard-crashes when attempting to open the debugger.

Current development

Requires Julia 1.12+ and the JuliaC.jl package. The JuliaC.jl app should be installed and a work around for JuliacLang/JuliaC.jl#13 implemented, i.e., add :$JULIA_LOAD_PATH to the shim where the JULIA_LOAD_PATH is loaded. Checks for these requirements are not currently failing the cmake. To build:

mkdir build
cd build
cmake ..
make

Which makes the library as well as a basic executable.