Dear all:

I am pleased to announce an improved version (still a beta) of my
Mlucas code, v2.7z. See ftp:/209.133.33.182/pub/mayer/README for
pointers to the source code and various binaries.

Changes from v2.7y include:

1) A bug discovered by Brian Beesley related to the printing (but not
the actual computation of) the LL residues has been fixed. This bug
was causing the upper 44 to 48 bits of some Res64s to be shifted one
place in printing. The lower 16 bits were unaffected, and George has
assured me that he'll accept any double-checking results where the
lower 16 bits match, and most of the upper ones also match upon
correction of the shift error. The bug only affects second (or later)
exponents in a multi-exponent run, and only if the run in question
was never restarted, i.e. ran without any interruptions. Note that
the bug could not have prevented you from getting credit for finding
a Mersenne prime, since a string of zeros would be unaffected.

2) The leading space in printed 'M(...' output lines has been removed,
so you no longer have to worry about stripping out leading spaces prior
to sending your results to the PrimeNet server.

3) George's and my e-mail addresses have been updated, so if one of
your runs dicovers the newest Mersenne prime, you'll actually be able
to contact us. :)

4) I've made some improvements in the carry propagation phase of the code,
which should give a small but noticeable speedup on some (especially lower-
end) systems.

5) Thanks to Bill Rea's run-time profiling efforts, the SPARC binary is
now roughly 10% faster at nearly all runlengths.

6) All binaries are now statically linked (e.g. compiled using
the -non_shared option), so there's no more need for mucking about
with Fortran run-time library (RTL) files. Just download, unzip,
give yourself execute permission, and run.

7) The FFT vs. exponent breakpoints have been made slightly more
conservative - Pmax now is about 98% that of Prime95 at each length.
(Several users of v2.7y reported that their jobs quit with fatal
roundoff error messages for exponents near the v2.7y breakpoints).


Savefiles are of course fully compatible with v2.7y. Note that
if for any reason you find v2.7z to run more slowly on your hardware,
another way to work around the Res64 print bug is to simply stop and
restart your current v2.7y job.

On a related front,I've put together an initial version of a GIMPS
source code timings page, whose main purpose is to guide non-x86 users
in choosing the best code for their platforms (or perhaps vice versa :)
The page is at ftp://209.133.33.182/pub/mayer/gimps_timings.html.

Thanks to Brian Beesley, Bill Rea, George, and all the other beta
testers who sent me comments and timings.

Happy hunting,
-Ernst

p.s.: I'm cc'ing this to a short list of known Mlucas users - if you'd
like to be included on or removed from the Mlucas mailing list, or just
to let me know that you're using the code (and roughly how many machines
you're running on), please send me private e-mail.

_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ      -- http://www.tasam.com/~lrwiman/FAQ-mers

Reply via email to