[vox-tech] a comment on Pete's fast-loop question

[Peter Jay Salzman]

On Sun 18 Jun 06, 12:06 AM, Norm Matloff <matloff at cs.ucdavis.edu> said:
> Pete Salzman asked about optimizing a loop in terms of execution speed,
> by having the loop index go from high value to low instead of vice
> versa.  Some discussion ensued in terms of what machine instructions a
> compiler could take advantage of in this manner.
> 
> One point that you might consider, Pete, is that these considerations
> are kind of nickel-and-dime in comparison to things like memory
> hierarchy issue.  There is much better payoff potential in writing code
> in such a way as to minimize cache misses, which cause major time
> penalties, and page faults, which cause catastrophic time penalties.
 

Yeah, I know.  My company hired me for algorithmic optimizations --- my
knowledge of devising a set of algorithms to accomplish a given task (e.g.,
solving dense matrices, choosing a particular Monte Carlo for a given
problem).

I guess my attitude towards this was more along the lines of "while I'm in
the neighborhood, why don't I stop by and say hello".  It certainly wouldn't
be prudent to re-implement code to use nickel and dime optimization, but
while thoughts in my neurons travel down my CNS and translate into my
fingers tapping on keys, I might as well tap on keys in such a way that the
nickel and dime optimizations appear on vim's terminal.  The price of
admission is free, so I might as well.

OK.  I took that nutty metaphor as far as it can go.  Hope it made sense.


> There is even a book on this, I believe in the Intel Press series.

Definitely sounds like a book I should pick up.  I'll Google for it!

Thanks,
Pete