Peter Bruin on Thu, 08 Oct 2015 17:51:17 +0200

[Date Prev] [Date Next] [Thread Prev] [Thread Next] [Date Index] [Thread Index]

Re: Faster digits and fromdigits in base 2^k, and FlxqM_mul_Kronecker

Hi Bill,

> On Thu, Oct 08, 2015 at 02:53:30PM +0200, Peter Bruin wrote:
>> Bonjour,
>> My motivation for this was mostly to speed up multiplication of matrices
>> over non-prime finite fields of small characteristic (FlxqM).  I am
>> attaching a second patch (FlxqM_mul_Kronecker.patch) that enhances
>> Kronecker multiplication for FlxqM to use the above functions for
>> packing into and unpacking from integer matrices in cases where it
>> allows us to pack the entries in fewer words per entry than the
>> (half-)word-aligned packing that is used now.
> Hello Peter,
> Thanks for your patch!
> Is that method applicable to ZX_mul and Flx_mul ?
> (What about ZX_eval2BILspec/Z_mod2BIL_ZX (and
> Flx_mulspec_mulii_inflate/Z_mod2BIL_Flx)) ?

In principle it is, but it is harder to make it pay off because the
overhead remains non-negligible.  I spent some time trying to apply this
method to Flx_mul; so far I haven't managed to get a systematic speed-up
over the existing code, but I am still optimistic that some improvement
may be possible.  As for ZX_mul, I haven't looked at that yet.