Math Forum / Math Software / Scilab / December 2008

I'm writing a Kalman filter application that involves going through big
piles of data.  The Kalman filter itself has numerous intermediate steps
to simulate the plant, which in turn has embedded calculations using
quaternions, for which I had to work up a Scilab data type.

It runs slooowly.  Each second of real-world data takes about 1.3 to 1.5
seconds to execute.  As I'm getting to a stage of the game where I need
to run this over and over again while I tune and tweak, execution time is
suddenly important to me.

What general strategies can I take to speed this up?  

I have already "vectorized" the code to a great extent; I may be able to
do more but only at great cost to readability and flexibility -- I'd
rather spend more time in simulations at this point than in struggling
with oddball code.

My understanding is that when I load a function into Scilab it is
automatically compiled -- is that correct, or are there opportunities for
speed gains with the code I have?

Beyond this, my understanding is that I need to identify the most often-
called functions and write them in C (I'm already doing this for another
project, so I know how).  Is this it?  Does Scilab have any profiling
tools to help me decide what functions to flog first?  I'd love it if I
could print out a report that said what percent of the processor's time
was spent in which function -- is there any way to extract this
information, and if so how?

Comments and direction are welcome.  Thanks.