General Advice About Matlab Hacking

• You want to have the graphics windows from Matlab opening up on your local machine, even when you are running Matlab remotely. Use ssh-agent for this - read the instructions from earlier in the course about using ssh.
• Start up matlab without all of the Java GUI windows. My personal alias (using tcsh) is

          alias matlab '/usr/local/bin/matlab  -nojvm -nosplash'
  

  where the first option avoids costly start up of Java virtual machine(s), and the second prevents the advertisement screen on startup.
• Avoid using loops whenever possible. Array operations are essential to decent performance. Example: matrix-matrix multiply of 200x200 matrices is over 1000 times faster using A*B instead of the usual triply nested loops. "Think vector" is a rule for Matlab, and it also applies to high performance computing (HPC) in general. The reason vectorization helps in HPC is that it generally maps well to hardware pipelines, or allows expressing the operations in terms of optimized BLAS libraries.
• Use script files; interactive operations should be for examining variables, debugging, experimenting with graphics options, making small changes.
• Use the diary function to keep track of session and to avoid having to retype in useful commands later. Beware, however: the diary file will be appended with each new Matlab session.
• Commands useful at top of every Matlab script file:
  • !rm session Removes the file "session" from the current directory
  • diary session Creates a new file "session" to store commands, results
  • format short e Makes Matlab print all variables using exponential notation. Otherwise, a single entry large in absolute value in an array can cause all the other entries to be printed as 0.0000
  • format compact You may have noticed that Matlab prints a blank line before and after every command/result. This supresses those blank lines.
  • dbstop if error causes Matlab to stop inside any function that causes an error, so that you can examine local variables inside that function. Otherwise, when an error occurs, Matlab returns from the function and you cannot find out what values were held inside the function. Use "return" to go back to calling function, and "dbquit" to exit the debugger. [This may be the default now in Matlab 6.x, to drop into the debugger when an error occurs.]
  Most of these things can be done via a file ".matlabrc", but by putting them into the scripts they will take effect wherever they are run. That is, you may ship your code around to other machines or give them to other people, and not have your personalized .matlabrc file available.
• When creating arrays in a loop, first preallocate the space using zeros(). The loop

  for k = 1:n
      C(1:n,k) = v(1:n);
  end
  

  makes n calls for memory allocation from operating system, and it requires copying over the part of C built up so far to the newly allocated space each time. Instead use

  C = zeros(n,n);
  for k = 1:n
      C(1:n,k) = v(1:n);
  end
  

  Better yet, use: C = v(1:n)*ones(1,n);
• Generally, manuals are out of date before being printed. The online help facility will show many useful commands not documented elsewhere. The demo command may also point out useful capabilities.
• Conformality problems: x = 1:n sets up a vector of length n, but is it a row vector or column vector? In the first case, y = A(1:n,1:n)*x gives an error message, while in the second it is valid. Handle this by converting it into a column vector: x = x(:). This makes x a column vector whether it is a row or column vector to begin with. Check and see what that does for a two-dimensional matrix. Also see reshape command.
• Inquiry functions are useful for debugging: size, length, whos.
• You can use shell escape to edit a file: !emacs filename, or to do other stuff if you have only one window.
• Plotting: some graphics commands do not produce the result until the end of a script file is reached, or a command changing the view or axes occurs. Use drawnow to force drawing. The reason for this is that Matlab was single-threaded for many years, so it could only do one thing at a time.
• Performance evaluation: use clock or cputime. Many older Matlab scripts have a "flop" command, which gives an estimate of the number of floating point operations performed so far. However, that is gone now from Matlab 6.0 - you'll have to count or estimate your own flops. If you need to use an older one, you may find it useful to write a flop() function that returns a zero value to prevent error messages.
• Sparse matrices should be used whenever possible. Convert to a sparse matrix by means of A = sparse(A). Note that many routines ( eig, svd, qr, ...) require A to be dense; in that case use eig(full(A)), etc. We'll cover sparse matrices later, just keep this as a rule to bear in mind then.

• Next page: More Plotting
• Class home page: A321