New thresholds in table

Torbjorn Granlund tg at
Mon Nov 14 11:09:39 CET 2011

nisse at (Niels Möller) writes:

  I have tweaked the tuneup parameters a bit. I removed min_size for
  HGCD_APPR_THRESHOLD (was 30, now uses the default of 10). This threshold
  is definitely lower than I had expected. On, it's 10 on one machine
  (itanium2), and 11 or 12 on several 64-bit machines (x86_64 and power).
I suppose 10 is still too high then.  Or doesn't the code support lower

The spreading of measured values for k10/k8 makes me suspect that
something might not be right.

All k10 and  k8 machines using 64-bit limbs get values less then 30,
but one is far off that:          171

These machines have two nehalem generation CPUs:   169          12

(There is a slight difference in integer division performance, but I
doubt you're using integer division.)

  I also increased max_size and step_factor for HGCD_REDUCE_THRESHOLD.
  That might have broken tuneup a bit, I see results are missing for a
  large number of machines for last night. Typical values seem to be a few
  thousand limbs.
Missing values are not alarming, (many) red ones are.  Things look good

Most of the GMP testing is nowadays done on emulators or under
virtualisers; running tuneup there makes limited sense.


More information about the gmp-devel mailing list