unbalanced multiplication and division

Torbjorn Granlund tg at gmplib.org
Thu Oct 29 22:24:47 CET 2009

Paul Zimmermann <Paul.Zimmermann at loria.fr> writes:

  > You should know that our focus have become 64-bit machines since GMP
  > 4.3.  That does not mean we do not optimise  for 32-bit machines,
  > only that they might get less attention.
  on http://www.loria.fr/~zimmerma/{plot1000.ps,plot10000.ps} you will
  find similar curves obtained on a 64-bit Core 2, after "make tune":
  * plot1000.ps shows a similar behaviour (size of the product or dividend
    is 1000 limbs)
  * plot10000.ps shows a similar behaviour for the division, and a decrease
    at about 1100 limbs for the multiplication, and then the time is constant,
    probably due to the use of an FFT of total size 10000+ limbs -> the tuning
    should be improved here.

I think I will not try to understand the exact causes for this right
now, since all code for this is in rapid flux so any analysis will
become obsolete.

When we have all the code in place for the GMP 4.4 release, re-measuring
this would be very useful.


More information about the gmp-devel mailing list