Codelets for ToomN1 (for N=2, 3, 4, 6, 8) should be added and here's why. (Also: a significant non-triviality on where cut-off points should be).
bodrato at mail.dm.unipi.it
Thu Apr 5 21:33:15 UTC 2018
Il Mer, 4 Aprile 2018 9:52 am, paul zimmermann ha scritto:
> One possible interpretation is the following: the simple threshold
> mechanism used by GMP might be suboptimal.
We know ;-)
> Since MUL_TOOM8H_THRESHOLD=214 on my computer, you don't need to have a
> large table to cover all Toom variants.
By the way, in the last message I forgot to point to some graphs we
elaborated to show where each multiplication variant is better. They can
be found at https://gmplib.org/devel/, under the title "New Toom
multiplication code for unbalanced operands".
Those graphs are old, they did not yet take into account either Toom6h or
> However, for the unbalanced multiplication, one would need a
> bi-dimensional table, which would be more expensive to store.
Should we really stop the bi-dimensional table at Toom8h? What about the
dispersed border Toom-versus-FFT?
Maybe a more fine-grained table for small sizes... then another table for
larger values.. We shall thinkabout that.
More information about the gmp-discuss