1..2..3 ways of integrating MATLAB with the .NET

• Three methods test Binary - 39 Kb
• Three methods test Source - 193 Kb
• MATNET library Binary - 40.6 Kb
• MATNET library Source & Doc - 420 Kb

Abstract

In this article, I want to describe three possible techniques of integration of the .NET Framework with MATLAB, then I'll chose one of them and create a library for accessing MATLAB functionalities. The final objective is to create a C# component that encapsulates an M-file, much like MATLAB R13 COM Builder does. The only requirement of this project is to have MATLAB installed.

Introduction

MATLAB is a powerful scientific tool for math manipulation, specifically matrix manipulations, with its own scripting language, where files are .M files. Usually it's used as a standalone program, but can be integrated in two ways using other languages like Java, C and FORTRAN. With the first usage, we declare a new function or component in a native language that extends the functionalities of MATLAB, usually to improve the execution speed; this approach uses the MEX system. The second usage, described in this article, uses MATLAB functionalities from an external program, this time a program in the .NET framework. In this schema, the Java language has a special role because, from the R12 (a.k.a. 6.0), MATLAB is Java-based (specifically for the multiplatform GUI functionalities) and you can easily use Java classes in both ways. To use MATLAB from an external program, there are three possible solutions:

1. low level C API
2. DDE
3. COM

I'll describe these solutions, implement in C#, from the easiest and highest level to the most difficult and efficient. Then I'll use the C API interface, to create first a wrapping class that calls the C APIs. Finally I'll show a tool to generate a C# class that encapsulates easily, a MATLAB algorithm, stored in a m-file, and publish the functionality as an object oriented set of methods. For example, given the following m-file that uses the built-in function magic:

function y = mymagic(x)
    y = magic(x);

the idea is to create a .NET library that gives the possibility to write a component like:

class MyMagic { 
   public static Matrix mymagic(Matrix m);
}

Possible MATLAB - .NET bindings

Solution 1 - COM approach

MATLAB exposes itself as a COM Automation Server, so we can use the .NET COM Interoperability to use its functions: we can work directly using a Type Library or using only the Dispatch interface. The progid of the server is Matlab.Application.

This code creates the server and executes a command cmd:

m = CreateObject("Matlab.Application")
m.Execute("cmd");

All the data transfer from MATLAB and our program is done through COM data types, and matrices are represented as SAFEARRAY(double) pointers, talking in COM terms. Using the COM interoperability, we can use double arrays.

double [,] MReal;
double [] MImag;
Result = Matlab.Execute("a = [ 1 2 3 ; 4 5 6 ; 7 8 9];");
call m.GetFullMatrix("a", "base", MReal, MImag);
call m.PutFullMatrix("b", "....)

The second parameter base is the workspace where to find the matrix variable a. According to the documentation, we have two main workspaces: base and global.

This solution is quite slow because of the COM type exchange, so it's better that we go ahead finding another way. Anyway, it is present in the article source code under the folder COMMATLib that exposes a Matrix class.

Solution 2 - DDE

The Dynamic Data Exchange is a quite old but powerful service of Windows that lets applications communicate and exchange data. It's the ancestor of the whole COM and so it's the grandfather of the whole .NET. Again, as with the COM solution, we have a client, our .NET program, and the server, the MATLAB DDE Server, that communicates.

A DDE communication is made with a Service Name, like Matlab or WinWord, about a topic of discussion, like System or Engine, and is based on the exchange of data elements called items.

The problem with the DDE exchange is the data types used to exchange information, because it uses the Windows clipboard data types. Specifically MATLAB supports:

• Text for operation results, and matrix data exchange (gulp!);
• Metafilepict for image transfer;
• XLTable for MATLAB-Excel data exchange;

This time we create a DDE communication with MATLAB and evaluates expressions using the EngEvalString item; to get or set Matrix data, we use an item with the same name of the matrix.

I've created a wrapper class called DDEMatlab with the same interface as COMMatlab, but this time we can exchange pictures to, like the one returned by the plot command. Anyway, the very problem with DDE is the matrix data exchange that is all text based.

Solution 3 - C API

The direct access to the MATLAB C API is the best solution in terms of performance and features, just let use P/Invoke and some unsafe pointer operations.

The C library internally uses Unix pipes or the COM to communicate with the main MATLAB instance, but in this case the data transfer is done using blocks of raw memory instead of the SAFEARRAY. So we can use the unsafe capabilities of C# to transfer efficiently matrices from and to MATLAB, it's just necessary to remember that MATLAB stores data in column wise order.

The structure of this solution is described later, with some examples of usage.

Comparison of the solutions

I've tested the three solutions too see the performance differences between them, but it's easy to imagine the results ahead: the test was done writing and reading a matrix of 200 by 200 elements. The following table synthesizes the results:

The matnet library

The matnet library encapsulates the access to the MATLAB engine using the classic P/Invoke mechanism, exposed by the class MATInvoke, and provides the following features:

• direct access to matrices, stored with different formats (Matrix class)
• access to MAT files to load data stored in binary format (MATFile class)
• evaluation of expressions (Evaluate method of the EngMATAccess class)

Examples and Usage

The basic usage of the matnet library requires to create an instance of the EngMATAccess that opens the communication with a MATLAB instance, or if not present, starts it up. Then the user can create instances of the Matrix class to store its data, and send evaluation commands to MATLAB.

using (EngMATAccess mat = new EngMATAccess())
{
    mat.Evaluate("A = [ 1 2 3; 3 5 6]"));
    double [,] mx = null;
    mat.GetMatrix("A", ref mx);
}

There are also two more interesting examples of usage.

Interaction with the Imaging Toolbox and Bitmaps

The mateng library gives also the possibility to interact with Bitmap as MATLAB matrices, and use the MATLAB Imaging Toolbox. The program ImageDemo in the demo source code loads an image and processes it using the imnoise function using MATLAB, in this case the problem is the startup time if there are no MATLAB instances running. Internally MATLAB stores images as byte type three dimensional matrices in column wise order, as shown in the following diagram. Maybe the integration with the Bitmap class could be put inside a matnet.imaging library.

Access with DLL

One of the interesting features of MATLAB is the translation of M-files into DLL libraries that can be used to distribute an algorithm in a very efficient form. The matnet library can be used to access also a custom DLL library (see example code). When an M-file is translated into a DLL, a function named AA is exposed as mlfAA, with Matrix parameters. These functions can be accessed from P/Invoke, for example an M-file with the function AA compiled as AAX.dll:

function x = AA(b)
    x = b + 2

Can be accessed with the class:

class MyDLL
{
    [DllImport("AAX.dll")]
    static extern IntPtr mlfAA(IntPtr a);

    static public Matrix mlfAA(Matrix m)
    {
        return new Matrix(mlfAA((IntPtr)m));
    }
}

Conclusions and Future Ideas

This article introduces the matnet .NET library as a method to access MATLAB functionalities from .NET applications in a very efficient way. This solution has been compared with the other two solutions (DDE & COM) available from MATLAB. The matnet library can be used to work directly with MATLAB or to expose M-files as components, it's also possible to interact with the imaging toolbox and bitmaps.

There are some things still missing:

• a tool, M2NET, that encapsulates automatically an M-file as a .NET component or a user custom library
• the extraction of figure and plots as Bitmap objects
• make it work on Linux with Mono (sorry, but I haven't MATLAB for Linux)