jlcall: Call Julia from MATLAB.

jlcall is a MATLAB MEX function that embeds Julia in MATLAB.

Why jlcall?

MATLAB is a tool for technical computing with a well-polished interface, many useful toolboxes, and a large number of users who have established efficient workflows around it. However, MATLAB's language has limitations that lead users to turn to writing MEX extensions in FORTRAN, C or C++ when efficient implementations are infeasible in MATLAB itself.

Julia is a relatively new language for technical computing. Its designers have striven to bring together into one place the most important positive attributes of a variety of languages that one might find being used in today's scientific computing projects, including: a familiar, unintrusive syntax and a large toolbox of scientific software, like MATLAB; convenient system-level "glue" capabilities, like Python; the possibility of achieving excellent runtime performance, like FORTRAN or C. Julia's present is already formidable, and its future is even brighter, but it still lacks an interface as clean and feature-rich as MATLAB's.

The purpose of jlcall is to bring MATLAB and Julia together in a way that mitigates the weaknesses of each by exploiting the strengths of the other, benefiting the users of each. jlcall lets MATLAB users extend MATLAB's functionality through the MEX interface with a language capable of achieving performance comparable to FORTRAN or C, but in a way that should make them feel more at home. Julia users benefit by being able to use MATLAB's polished front-end, as well as having a means to collaborate seamlessly with colleagues who work primarily in MATLAB.

jlcall is not the only way to bridge MATLAB and Julia. MATLAB.jl is a Julia package that enables calling MATLAB from Julia through the MATLAB Engine interface. Indeed, jlcall itself takes advantage of the data marshaling functionality provided in MATLAB.jl. However, MATLAB.jl itself does not provide a means for MATLAB to call into Julia. jlcall complements MATLAB.jl in this manner.

Nor is jlcall the first attempt to facilitate calling Julia from MATLAB. For instance, julia-matlab provides a means for calling Julia from MATLAB through a ZeroMQ connection and a clever data codec. For julia-matlab the communication is inter-process. As such, all data passed between the two involves copying. In contrast, with jlcall, MATLAB and Julia coexist in the same process, sharing a common address space. In this case, one can avoid data copying in certain cases when crossing the language border. The time and space savings can be significant when working with large chunks of memory, as is frequently done in technical computing.

Getting started with jlcall

Prerequisites

It goes without saying that using jlcall requires having both MATLAB and Julia (>=v.0.4) installed on your system. Additionally, to build the MEX function from source, you need a C++ compiler configured to work with MATLAB's mex function. You can ensure this is done by executing:

>> mex -setup C++

from the MATLAB command prompt. Further, jlcall relies on the MATLAB.jl package being available to Julia. You can install it by executing the following command in Julia:

julia> Pkg.add("MATLAB")

Configuration

Now you are ready to set up jlcall. Clone it, if you haven't already. Then start MATLAB and navigate to the jlcall/m directory. Once there, run the jlconfig command:

>> jlconfig

You will be prompted to select a julia executable. Once you have done so, jlconfig will:

1. interrogate your system for settings,
2. write a jlconfig.mat file containing information necessary to properly initialize jlcall,
3. build the jlcall MEX function from source,
4. add the jlcall/m directory to your MATLAB path.

Simple uses of jlcall

The jlcall project provides Jl.m, which contains a high-level interface to the jlcall MEX function. Aside from hiding the details of Julia runtime initialization, Jl.m provides some useful member functions.

Jl.eval

Using Jl.eval, one can evaluate Julia expressions captured in MATLAB strings:

>> Jl.eval('2+2')

ans =

                    4

Jl.call

With Jl.call, one can call any Julia function whose arguments can be marshaled to Julia via MATLAB.jl's jvariable function and whose return value can be marshaled to MATLAB via mxarray. The first argument to Jl.call is the name of the function as a string. The remaining arguments are the values to be passed to the named function:

>> Jl.call('factorial', 10)

ans =

     3628800

Jl.include

One can load new Julia code by calling Jl.include. Suppose the file double_it.jl contains:

function double_it(outs::Vector{Ptr{Void}}, ins::Vector{Ptr{Void}})
  try
    mex_return(outs, [ 2*v for v in mex_args(ins) ]...)
  catch e
    mex_showerror(e)
  end
end

This MEX-like function can be loaded into the Julia runtime as follows:

>> Jl.include('double_it.jl')

Jl.mex

MEX-like functions can be invoked with Jl.mex. The first argument is a string naming the function to be called. Remaining arguments are passed to that function. For instance, now that it is loaded, we can invoke double_it:

>> a = rand(5,5)

a =

    0.7577    0.7060    0.8235    0.4387    0.4898
    0.7431    0.0318    0.6948    0.3816    0.4456
    0.3922    0.2769    0.3171    0.7655    0.6463
    0.6555    0.0462    0.9502    0.7952    0.7094
    0.1712    0.0971    0.0344    0.1869    0.7547

>> Jl.mex('double_it', a)

ans =

    1.5155    1.4121    1.6469    0.8775    0.9795
    1.4863    0.0637    1.3897    0.7631    0.8912
    0.7845    0.5538    0.6342    1.5310    1.2926
    1.3110    0.0923    1.9004    1.5904    1.4187
    0.3424    0.1943    0.0689    0.3737    1.5094

It is worth noting that Jl.mex is a constant of the Jl class that is equal to @jlcall, a handle to the jlcall MEX function. We recommend always referring to Jl.mex instead of jlcall, as the initialization of the Jl.mex constant forces the initialization of the Julia runtime. If one were to use jlcall directly, one would have to separately ensure initialization occurred.

Under the hood: the jlcall MEX function

The principal way that the jlcall MEX function works is, given its first argument is a string, to call the Julia function with the given name, passing in the the (remaining) MEX function arguments, packaged as two values of type Vector{Ptr{Void}}. The first argument is the array of outputs, while the second is the array of inputs. Thus, jlcall call allows any Julia function with this MEX-like signature to be invoked from MATLAB. In particular, the functionality underlying Jl.eval and Jl.call is provided by two such Julia functions.

Of course, there are no functions in Julia's Base library that match the above criteria, so one also needs, at a minimum, the ability to load code into the Julia runtime. Further, provision must be made for initializing the Julia runtime in the first place. To accomodate, jlcall can also be called in other ways. Invoking jlcall:

• with no arguments returns a logical representing whether the Julia runtime has been initialized;
• with the empty string as its first argument signals initialization, at which point the second and third arguments are interpreted as strings representing the paths of JULIA_HOME and the desired Julia system image, respectively;
• with a non-string first argument signals that all remaining string arguments are to be passed in as arguments to jl_eval_string.

The jlcall MEX interface is sparse by design. It is meant to provide the minimal functionality necessary to connect MATLAB and Julia through the MEX interface, and nothing more. Higher level functionality is left to be implemented on either side of the interface, each of which provides a language that is more appropriate for that purpose than C++.

Limitations and caveats

• The input arguments passed in to jlcall from MATLAB must not be modified on the Julia side. If it happens, expect MATLAB to crash. This is a requirement of the MEX interface, and is not specific to jlcall.

• Currently, Julia's STDOUT and STDERR are not redirected to the MATLAB GUI console. However, it should still be visible in the console from which MATLAB was started. For Windows users, starting MATLAB from the Windows terminal (cmd) is insufficient to see output, however, launching MATLAB from the cygwin or msys prompt should work.

• This project is very new and has been tested on only one machine, let alone across multiple machines, platforms, MATLAB versions, etc. Expect bugs, and please report them when they occur. PRs that extend functionality to new platforms are particularly welcome. The author would like to see this project succeed across the three platforms on which Julia and MATLAB are supported, and will make every effort consistent with his other many, many obligations to assist users in getting jlcall functioning in their environment.