Subject: Re: [boost] different matrix library?
From: DE (satan66613_at_[hidden])
Date: 2009-08-18 12:13:25
on 18.08.2009 at 20:05
joel wrote :
>> fewer instructions doesn't mean faster code
> Thanks to learn me my job.
>> to find ideal ratio you must take intel arch manual and count the clocks
>> or you can actually measure running time for a variety of cases
> Time benchmark is always better. With pipeline in OOCPPU, countign clock
> in a linear fashion is dumb.
>> i naively (lack of time) implemented nrc (how it is spelled btw?) for
>> 2d case and it performs worse then computed indeces in general
>> i'll try to reimplement it properly and measure again
> It shouldn't. 2D is worse case and perform like only 3-5% less than
> index. Proper test case is up to 3D.
ah, ok
you might be right (actually i have no chance to prove the contrary)
so why not compute indeces for 2d case and use nrc for N>2?
i think it's trivial to implement
-- Pavel