What's a reasonable size ratio for toom32?

[Niels Möller]

Niels Möller <nisse at lysator.liu.se> writes:

> Paul Zimmermann <Paul.Zimmermann at inria.fr> writes:
>
>>        Dear Niels,
>>
>> ./speed -p 1000000 -c -s 10-200 -f1.1 mpn_mul.0.6 would be more readable,
>> although the change in speed.h would be larger.
>
> See below patch, to support mpn_mul/0.6. Example output:
>
>   $ ./speed -c -r -s 10-50 -t 5 -p 1000000 mpn_mul mpn_mul/0.6 
>   overhead 5.00 cycles, precision 1000000 units of 1.25e-09 secs, CPU freq 798.28 MHz
>                 mpn_mul   mpn_mul/0.6
>   10             215.48       #0.6295
>   15             443.16       #0.6066
>   20             784.50       #0.6138
>   25            1209.57       #0.5957
>   30            1490.59       #0.6934
>   35            1986.51       #0.6918
>   40            2547.99       #0.6981
>   45            3189.27       #0.6633
>   50            3827.87       #0.6734
>
> What do you think? (Also deleted FLAG_RSIZE, which appeared unused).

I pushed this change, then found out it was a bit broken (failed to
recognize old optional .r option, and it seems initialization of the
size_ratio field with memset produces a float value > 0.0). Just pushed
a fix for that.

And now I found that speed's handling of toom22, toom32 and friends
always use fixed ratios, so I have to extend that to be able to get any
interesting benchmarks when varying the ratio.

Regards,
/Niels

-- 
Niels Möller. PGP key CB4962D070D77D7FCB8BA36271D8F1FF368C6677.
Internet email is subject to wholesale government surveillance.