Mex Memory Interfacing (Open Source Project Floating)

[Arjun Rao]

In short, I would love any and all kind of help in expanding the following project of mine. Keeping formalities short, here’s the relevant description.

MexMemoryInterfacing

Introduction

MATLAB is among the most popular softwares in existence for numerical computation. It is widely used in all manner of scientific investigations including electrical, mechanical, chemical, [you name it] engineering.

As it turns out, most newbies to MATLAB find it challenging to adapt to coding in MATLAB when they realize that the code that they write in C, C++, Java etc., as written in C, C++, Java, runs painfully slowly under MATLAB. The reason for this is MATLAB’s nature as an interpreted language where there is a significant runtime overhead to running every line of code (for e.g., the data type compatibility must be checked, indexing validity must be checked etc.). This puts the requirement on the programmer that a large amount of work is done in each statement i.e. the requirement that the code be “vectorized”.

HOWEVER, as it so happens, some algorithms simply cannot be wrtten in a vectorized manner. A simple example would be to find the cumulative sum of a given vector. Fortunately, matlab includes a function that efficiently executes the above computation. However, there are several other, less fortunate cases where the best one seems to be able to do is write a for loop and lament as MATLAB takes an eternity to execute the code.

Having recognized the above problem, MATLAB provides an interface to use C/C++ to program called MATLAB EXtensions or MEX for short. This interface is an elegant C interface that enables us to access the underlying data in existing MATLAB vectors/matrices and perform the non-vectorizable computation in the FAST C language and return data back to MATLAB in a MATLAB vector/matrix.

However, as C++ programmers, we are used to the neat functionality offered by the C++ Standard Library (e.g. std::vector), static time polymorphism, and especially RAII. These features are entirely missing if we only restrict ourselves to the C interface provided by MEX.

I Have found out that one CANNOT use the C++ standard library to manipulate memory that is returned to MATLAB due to various reasons (e.g. Alignment Requirements and compulsory deallocation at the end of scope).

Thus, I have developed a set of headers in my repository MexMemoryInterfacing which implements two classes MexVector and MexMatrix that allow me to perform a zero-cost wrapping over MATLAB Data. The Additional features are listed below.

Relevant Code

The relevant code can be found in the repository titled MexMemoryInterfacing at the following link.

https://github.com/maharjun/MexMemoryInterfacing

The code in the following repositories uses MexMemoryInterfacing

https://github.com/maharjun/TimeDelNetSim

https://github.com/maharjun/PolychronousGroupFind

https://github.com/maharjun/Grid2D

https://github.com/maharjun/GetAttractionBasin

Current Features

The features that have currently been implemented are as below.

1. MexVector<DataType> can wrap any MATLAB Vector of any data type.
2. MexMatrix<DataType> can wrap any MATLAB Matrix of any data type.
3. MexVector attempts to mirror the interface offered by std::vector
4. MexVector has interfaces that allow it interact with standard library iterators
5. MexVector is capable of wrapping cell arrays as MexVector<MexVector<...>>
6. I have programmed templated input and output functions that perform type inference and type checking. I have emphasized static typing where I felt the need.
7. I have functions that perform Ctrl-C (Interrupt Signal) Handling
8. Include Other Features Later

Current Issues

1. Heavily uses C++11 features and is only compilable under gcc 4.9.2 or greater. (All of my code has been written assuming gcc 5 (i.e. complete C++11 support)). It is also compilable under Visual Studio 13 or 15. This is quite a major problem as MATLAB does not link to the newer version of libstdc++ by default (not a problem in windows).
2. Lack of comprehensive testing and documentation.
3. Requires knowledge of template programming in order to implement new data structures. (Need to simplify process).
4. Input Output functions may need some polishing / refactoring.
5. Implementation of Ownership features is very limited.
6. Havent used a namespace

Why Contribute?

1. Will gain in-depth understanding of C++ and template programming upon reading the source code.
2. MATLAB is widely used and there is no question regarding the utility of this project.
3. Will likely be useful to yourself if you are a regular user of MATLAB and are frustrated by its slowness for your particular application.

Required Skills.

Well, All of the skills can be learned on the job (willing to mentor). But, in order to contribute, you will need to be familiar with the following

1. MATLAB syntax (at least the basic details)
2. MEX Interface. It has an intuitive and clear design. Should not take too long.
3. This is the hardest. C++ Template Metaprogramming (necessary to understand the code, conversely, can be learnt by understanding the code).

​

[Anurag Mundhada]

This seems interesting. If I understand it correctly, the project allows us to utilize C++'s superior performance for the non-vectorizable work we are doing in MATLAB?

Can you expand on specific examples where you found it useful in your work?

--
--
The website for the club is http://wncc-iitb.org/
To post to this group, send email to wncc...@googlegroups.com
---
You received this message because you are subscribed to the Google Groups "Web and Coding Club IIT Bombay" group.
To unsubscribe from this group and stop receiving emails from it, send an email to wncc_iitb+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Kumar Ayush]

Everyone interested can now apply at http://wncc-iitb.org/soc

This project can be viewed at http://wncc-iitb.org/soc_projects/33-MexMemInter-arjun.html

Regards

Kumar Ayush

Manager, WnCC | STAB

[Dilawar Singh]

Have a look here at this project https://sourceforge.net/p/sbml/code/HEAD/tree/trunk/libsbml/src/bindings/matlab/

They binds C++ core library with matlab (among other languages). It will be a good idea to see how they do it.

In case you are not aware of it, gcc has made ABI changes recently since version gcc-5.x.y. Even if gcc-4.x.y supports
c++11, you may not be able to link it with matlab runtime.

Also checkout llvm compilers. If you can

get llvm working, you can support OSX as well (along with unix/cygwin).

best,

     Dilawar

--

--
The website for the club is http://wncc-iitb.org/
To post to this group, send email to wncc...@googlegroups.com
---
You received this message because you are subscribed to the Google Groups "Web and Coding Club IIT Bombay" group.
To unsubscribe from this group and stop receiving emails from it, send an email to wncc_iitb+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Dilawar
NCBS Bangalore

[Arjun Rao]

I am indeed aware of this fact. I had to actually replace the MATLAB libstdc++ to get the code compiled using gcc 5 to work. This is indeed a problem. But the last i tried, my code could be compiled under gcc 4.9.2. Assuming that the ABI hasn't changed in this version, the code should be fine.

[Arjun Rao]

Regarding LLVM, I have managed to get it working under Clang in both Linux and OSX as of two weeks ago. Currently, the build system being used in my projects is CMake (due to Windows/Visual Studio support)