FAQ: how to perform jit for function evaluations of my optimization problem? - casadi/casadi GitHub Wiki

Take any optimization problem such as rosenbrock from https://github.com/casadi/casadi/blob/master/docs/examples/

nlpsol takes a fourth optional argument with options. If you are using opti, you pass options as second argument to opti.solver In Matlab, options are given by a struct, in Python options are given by a dictionary.

The jit option will have the following effect:

any helper Function that a solver requests (e.g. nlp_jac_g to evaluate the constraint Jacobian) will be code-generated to the disk
the code will be compiled into a shared library
the shared library will be loaded back into CasADi, like External

By default, a shipped clang-compiler is used, but it appears to be buggy. We recommend to use your system compiler (shell).

Examples:

options = struct;
options.jit = true;
options.compiler = 'shell';
options.jit_options.flags = {'-O3'};
options.jit_options.verbose = true;

jit_options = {"flags": ["-O3"], "verbose": True}
options = {"jit": True, "compiler": "shell", "jit_options": jit_options}

On Linux/Mac, this just work if you have gcc installed. For Windows, Matlab/Python will need to be started from an environment that has Visual Studio compiler set up. For example, launch Matlab/Python from x64 Native Tools Command Propmt for VS (Requires Visual C++ components or Build Tools for Visual Studio, available from Visual Studio installer. You also need SDK libraries in order to access stdio and math.).

Which compiler flags to use to make it really fast?
A good choice for GCC is -Ofast -march=native.
TODO why.

JIT has additional options like

opts.jit_cleanup = false;
opts.jit_temp_suffix = false;

TODO what do they mean?