Bill Allombert on Fri, 15 May 2009 22:40:06 +0200

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

Re: Another problem with matrix inversion

On Fri, May 15, 2009 at 05:34:09AM +0200, Lorenz Minder wrote:
> Hi,
> I take it, the API for FpM_gauss() must not change, right?

Yes, it should stay as documented, though you can change FpM_gauss_sp
as long as it does not affect FpM_gauss().

> > At this point the only trouble is to pick the fastest algorithm depending 
> > on the matrices entries.
> I did some tests to see how this behaves for t_INTMOD over small moduli.
> I took m = 2^8 - 1 = 3*5*17 as an example. I created random invertible
> matrices mod m of size n * n.
> Here are the results:
> matadjoint technique:
>   n = 150, stack usage between 128MB and 256MB, running time 67s.
>   n = 170, stack usage between 256MB and 512MB, running time 112s.
> chinese remaindering:
>   n = 150, stack usage < 8MB, running time 0.15s.
>   n = 170, stack usage < 8MB, running time 0.227s.
> For me, the most worrying thing here is the excessive memory usage,
> which suggests that something with n = 200 variables or so will be
> difficult to solve on a typical PC.
> The numbers suggest that for t_INTMOD, the splitting trick, or possibly
> adapting matsolvemod(), will fare much better, unless matadjoint() can
> be sped up by orders of magnitude, and its memory usage can be contained.

This is in part because RgM_inv calls Flm_inv which is optimized for modular
arithmetic with long integer, instead of computing with INTMOD. 
Flm are between 7 and 9 times smaller that INTMOD matrices and operations
can be done "in place".