I think you are wrong here Ben or at least you are missing one serious problem.
As far as I know the GP fitness evaluation code is a largely branching stuff
that can hardly be vectorized and the SPEs have weak or no branch predictor
at all and small cache and are good precisely at the other kind of code,
namely highly vectorized code with few branching points. I would not be
surprised if it turned out that with the type of code you have to execute to
do GP fitness evaluation the SPEs will be 10 times slower than an Athlon
processor and the whole gain from having 8 such units will be lost.
So it is not just the memory issue -- look at the processing speed as well.
That's another problem, of course. I don't know how your GP evaluation code works - if it runs an interpreter loop on S-expressions or calls functions via pointers or suchlike, it could end up spending most of its time in the resulting pipeline bubbles. Can you compile the expressions to be evaluated into machine code on the fly? If so, will the resulting chunks of code then be doing reasonably vector-friendly stuff?
- Russell
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
