Re: [Boost-users] [multi-array] Performance? Difference from GIL or uBLAS?

From: joel falcou (joel.falcou_at_[hidden])
Date: 2008-08-07 13:19:06

John C. Femiani a écrit :
> Did you mean to say 'index or
pointer'? 
Yes of course

> It seems like the trend in
image 
> processing is to avoid using indexes in the inner loop if
at all possible.
Which is a non sensical way of doing stuff cause
code using pointer arithmetic and "leet" stuff like *(ptr++ +
x*width) makes code hard to decipher and to parallelize afterward.
I
wish you a good deal of fun to block a pointer based IP code into SMP or
SIMD :/ 
It's preferable to stick to indexes as it works fine on both
side (parallelization and performances).

> Second, the
results of the operation were never read, so the compiler 
>
removed the whole 'Pointers' test.
I always fall into this trap
myself even after years ;)

I got similar result on my box too.
But now, turn to fortran indexing and it gets worse even with indexes.

Side notes : a better metric than absolute time or cycles is cycle
per element as it allows us to detect cache miss and to have performances
which can be comapred through various architecture at the metric is
independant of most of classic pitfall from other performances measures.