Re: [eigen] two things

[Benoît Jacob <jacob@xxxxxxxxxxxxxxx>]

By the way, looking at benchVecAdd.cpp... a gigaflop is probably 10^9 flops, 
not 2^30 flops as you coded it...

these results already aren't so shining (as you noted they'd already be much 
better with peeling) so let's not make them even worse! :)

Benoit

On Thursday 26 June 2008 18:12:24 Benoît Jacob wrote:
> Here is the result here after revision 824739 introducing packet(int):
>
> 2.30471s  0.646553 GFlops
> 400 x 400  2.51917s   0.59151 GFlops
> 320 x 500  3.00975s   0.495097 GFlops
> 256 x 625  2.92007s   0.510302 GFlops
> 250 x 640  2.9007s   0.513709 GFlops
> 200 x 800  2.9415s   0.506583 GFlops
> 160 x 1000  2.90815s   0.512393 GFlops
> 128 x 1250  2.73835s   0.544166 GFlops
> 125 x 1280  2.889s   0.515789 GFlops
> 100 x 1600  2.92752s   0.509003 GFlops
> 80 x 2000  2.86383s   0.520322 GFlops
> 64 x 2500  2.9053s   0.512896 GFlops
> 50 x 3200  2.90992s   0.512081 GFlops
> 40 x 4000  2.90439s   0.513056 GFlops
> 32 x 5000  2.80648s   0.530955 GFlops
> 25 x 6400  2.85519s   0.521896 GFlops
> 20 x 8000  2.79833s   0.532503 GFlops
> 16 x 10000  2.8511s   0.522646 GFlops
> 10 x 16000  2.81542s   0.52927 GFlops
> 8 x 20000  2.80733s   0.530795 GFlops
> 5 x 32000  2.76623s   0.53868 GFlops
> 4 x 40000  2.85234s   0.522418 GFlops
> 2.80785s   0.530697 GFlops
>
> So, as expected, this problem is solved.
>
> > hand coded vector with loop peeling:
> > 1.0101 sec 1.47521 GFlops
> >
> > VectorXf(400*400):
> > 1.50368 0.990978 GFlops
>
> So it would be much worth peeling loops. Now that we have a real linear
> path (and could also write a linear path in non-vectorized case) this will
> be much easier and more efficient.
>
> Cheers,
>
> Benoit
>
> PS. There was a crash in benchVecAdd caused by ei_pload on non-aligned
> address. It is tricky as of course with a bit of luck you could get three
> aligned addresses in a row, especially on a 64 bit system like yours...
> fixed in bench/.

Attachment: signature.asc
Description: This is a digitally signed message part.