MINC/Tools/emma/emma-speed

Speeding up MATLAB

One of the frequent complaints that we hear from people just being introduced to MATLAB is that it is terribly slow. There is a speed trade-off inherent in MATLAB. Since it is an interpreted language instead of a compiled language (such as C or FORTRAN), its speed will almost always lag behind that of a custom program written in a language like C. However, this speed loss is made up for by the ease of code development.

Fortunately, the speed of MATLAB can be increased significantly by careful construction of the scripts. There is one basic rule to follow when writing a MATLAB script for speed:

• Under all circumstances, try to avoid the use of for loops in MATLAB.

This rule has a simple to remember corollary:

• Vectorize, vectorize, vectorize.

In other words, whenever possible use vector operations. For many operations, with some thought applied to vectorization, code speed can be increased by several orders of magnitude. Unfortunately, vectorization can be difficult for someone new to MATLAB to visualize. Since most people are introduced to MATLAB after they already have a familiarization with traditional programming languages, they tend to use for loops automatically.

Let's consider a simple example. We want to find how many points in an image have a value greater than 2000. In a traditional programming language, we might take the following approach:

j=0;
for i=1:16384;
    if (PET(i)>2000)
        j=j 1;
    end
end

This takes about a second to execute under MATLAB. A vectorized approach is to use the MATLAB find function:

length(find(PET>2000));

This takes about 0.07 seconds to execute, and is therefore approximately 15 times faster than the traditional approach. Now, consider the following related problem. We wish to create a mask based on the image we were just manipulating, where any point whose value is greater than 2000 is set to one, and all other points are set to zero. The traditional solution to this might look something like:

mask=zeros(16384,1);
for i=1:16384;
    if (PET(i)>2000)
        mask(i)=1;
    end;
end;

This takes about 0.81 seconds to execute. A faster, vectorized approach is as follows:

mask2=zeros(16384,1);
index=find(PET>2000);
mask2(index)=ones(size(index));

This takes about 0.05 seconds to execute, and is therefore about 16 times faster than the traditional approach. It also introduces an important technique that is useful when vectorizing routines: using a vector as an index to another vector. In the above code fragment, the variable index contains the indices of values.

One other thing to keep in mind when trying to speed up MATLAB code is that it is possible to call FORTRAN or C routines from within MATLAB. These functions are called either FMEX or CMEX functions depending on which language they are written in, and details on creating them can be found in the MATLAB External Interface Guide.

The EMMA toolbox contains several CMEX versions of popular MATLAB functions:

1. ntrapz is a replacement for the MATLAB trapz trapezoidal integration function.
2. nfmins is a replacement for the MATLAB fmins function minimization routine.
3. lookup performs 1-D linear interpolation, similar to MATLAB's interp1 routine.
4. rescale is useful for scaling a large matrix by a constant, or by another matrix of the same size and shape. rescale doesn't offer a significant speed advantage over the .* operator, but it performs the arithmetic in place, so no copy is made of the target matrix. This can result in significant memory savings.

These routines are functionally identical to the MATLAB routines, and simply offer an increase in speed.