Practical ways to reduce overheads
in MATLAB simulations

Abstract

MATLAB is now well-established as an effective tool for performing numerical experiments and graphical simulations. Its simple, high-level programming language allows rapid development of new projects and facilitates debugging. However, a high-level interpreted language such as MATLAB m-code cannot compete in speed and memory efficiency with traditional compiled languages such as Fortran and C. Even on a modern workstation, the capacity to perform complex two and three-dimensional simulations is limited.

In this presentation we will investigate some of the possibilities for alleviating bottlenecks in slow or memory intensive MATLAB programs. We will briefly consider the use of the Profiler in identifying these bottlenecks, the vectorisation of m-files, and the use of the MATLAB Compiler. We will then look in detail at the MATLAB Application Program Interface which allows one to incorporate C and Fortran routines directly into MATLAB programs. I will describe how to write and build MEX files using a simple example and will make a quantitative cost comparison between the different strategies.

Click here to begin the presentation.

Sample MEX files.

Here are two real example MEX applications that demonstrate common implementation issues. The programs are available in both C and Fortran versions and are fully commented. All programs have been tested on the Sun Workshop Compiler Version 4.0.

Example 1: d = find_min_dist(line_x,line_y,x,y). Given a grid (x,y) and a piecewise continuous line (line_x,line_y), this function will compute the minimum distance from each grid node to a point on the line. The underlying algorithm is simple and the function demonstrates basic error validation and the use of MATLAB dense mxArrays (matrices). The Fortran version also demonstrates the use of the %val compiler directive for passing arrays of unspecified length.

• C version - find_min_dist.c.
• Fortran version - find_min_dist.f.

Example 2: L = myluinc(A[,milu[,steps]]). Given a matrix A, this function will compute the level zero incomplete LU factorisation of the matrix. Optional arguments turn on modified incomplete factorisation and specify the number of elimination steps to perform. The algorithm is based on that presented in Owe Axelsson, `Iterative Solution Methods', Cambridge 1996. The program demonstrates how to manipulate sparse matrices, something which can be challenging when unfamiliar. The program also demonstrates the use of optional input and output arguments.

• C version - myluinc.c.
• Fortran version - myluinc.f.

N.B. Fortran version does not compile on versions of gcc below 2.9.x, since before then gcc did not support the %val() directive.