Below is an attempt to add automatic tuning of HGCD2_DIV2_METHOD.
On my laptop, method 1 (using div1 in the common case) is 9% faster than
the current default method 2 (bitwise division).
I wonder if it's really useful to support all variants in speed, or if
that could be deleted, to simplify a b
t...@gmplib.org (Torbjörn Granlund) writes:
> ni...@lysator.liu.se (Niels Möller) writes:
>
> What about mpn/generic/hgcd2-div.h?
>
> Sounds OK.
Done.
> I'm leaning towards keeping static for now. Tables aren't that big, and
> most archs don't use them anyway.
>
> You could hide them #if
ni...@lysator.liu.se (Niels Möller) writes:
Replying to myself...
Best answers that way! :-)
> Into gmp-impl.h, or its own file? In the latter case, what's a good name
> and location?
What about mpn/generic/hgcd2-div.h?
Sounds OK.
> How to deal with the tables (used for method 4 a
Replying to myself...
ni...@lysator.liu.se (Niels Möller) writes:
> Into gmp-impl.h, or its own file? In the latter case, what's a good name
> and location?
What about mpn/generic/hgcd2-div.h?
> How to deal with the tables (used for method 4 and 5)? We could keep the
> static declaration (at l
