nisse at lysator.liu.se
Mon Jul 5 19:28:52 UTC 2021
Torbjörn Granlund <tg at gmplib.org> writes:
> I suppose it would be somewhat interesting to know how close the
> measured best and measured 2nd best methods are, or how much better the
> measured best methods is than the configured method.
Part 1 is done by the "runner up" logic I added to tuneup a while ago. In
the tuneup output, it looks like
#define DIV_QR_1N_PI1_METHOD 4 /* 13.42% faster than 2 */
But not surfaced in an obvious way on the threshold web pages.
Part 2 would be relevant too, in particular when the configured method
is neither best or the runner up, according to tuneup.
> And, when there is asm, the scary performance ratio between C and asm
> which would suggest that C-to-C comparisons of tight GMP loops might not
> be terribly relevant. :-)
That's a more subtle comparison. Would also be good to annotate which C
method (if any) the assembly routine is based on.
Niels Möller. PGP-encrypted email is preferred. Keyid 368C6677.
Internet email is subject to wholesale government surveillance.
More information about the gmp-devel