New thresholds in table

2011-11-12 Thread Torbjorn Granlund
I have added HGCD_APPR_THRESHOLD and HGCD_REDUCE_THRESHOLD to http://gmplib.org/devel/thresholds.html. -- Torbjörn ___ gmp-devel mailing list gmp-devel@gmplib.org http://gmplib.org/mailman/listinfo/gmp-devel

Re: Cancellation with hgcd / unbalanced mulmod_bnm1

2011-11-12 Thread bodrato
Ciao Niels! Il Ven, 11 Novembre 2011 4:21 pm, Niels ha scritto: The input sizes to mpn_mulmod_bnm1 is then 3n x n. How well suited is current mpn_mulmod_bnm1 to these unbalanced operands? The cost of current mpn_mulmod_bnm1 implementation does not really depend on how much balanced the

Re: Cancellation with hgcd / unbalanced mulmod_bnm1

2011-11-12 Thread Torbjorn Granlund
ni...@lysator.liu.se (Niels Möller) writes: Any other tricks? I think you forgot the save-the-transform trick. Since b is implicitly invariant over several a_i pieces, evaluating b once will be a slight win. -- Torbjörn ___ gmp-devel mailing list

Re: Cancellation with hgcd / unbalanced mulmod_bnm1

2011-11-12 Thread Torbjorn Granlund
ni...@lysator.liu.se (Niels Möller) writes: Torbjorn Granlund t...@gmplib.org writes: I think you forgot the save-the-transform trick. Since b is implicitly invariant over several a_i pieces, evaluating b once will be a slight win. That's a bit difficult to do with current

Re: Cancellation with hgcd / unbalanced mulmod_bnm1

2011-11-12 Thread Niels Möller
bodr...@mail.dm.unipi.it writes: Remember that toom42 simply is unbalanced Toom-3 (i.e. same cost as toom33), toom63 is unbalanced Toom-4'n'half and mpn_mulmod_bnm1 can be used for unbalanced FFT. With this point of view, it should be easier to understand that, in the FFT range, mulmod is the

Re: Cancellation with hgcd / unbalanced mulmod_bnm1

2011-11-12 Thread Niels Möller
Torbjorn Granlund t...@gmplib.org writes: Perhaps one should look into a specialised toom-like primitive that directly takes the cancellation into account? This should not be hard to investigate. Like, extending toom_mullo beyond tomm2 and Mulders' trick? Or do you think one can do something

Re: HGCD_REDUCE_THRESHOLD

2011-11-12 Thread Niels Möller
Torbjorn Granlund t...@gmplib.org writes: I tried adding param.max_size = 1 to tune_hgcd_reduce, and on shell.gmplib.org I now get 3681. Measurements become very slow, one might want to set a coarser step factor too. I tried some values, amd I think max_size = 1 (or something a bit