mpn_sqrtrem{1,2} - patch for pure C implem

Marco Bodrato bodrato at
Thu Jun 8 05:17:01 UTC 2017


Il Mar, 6 Giugno 2017 9:00 am, paul zimmermann ha scritto:
>        Dear Marco,
>> I'll be happy to receive comments or critics :-)
> I'd be happy to see a comparison with the code of [1]. According to [2],
slide 31, the slow branch of mpfr_sqrt takes 56 cycles for one 64-bit

You mean the code in mpfr/trunk/src/invsqrt_limb.h ?

I'll try.

Then I'll try to mix the tricks:
 - tabling the cube of the initial approximation to save two
multiplications, as the code in MPFR does;
 - adding a little perturbation to some iteration to increase precision
and keep on using a smaller table.

I'll let you know.

Best regards,



More information about the gmp-devel mailing list