[issue2819] Full precision summation
report at bugs.python.org
Thu Aug 14 13:10:15 CEST 2008
Mark Dickinson <dickinsm at gmail.com> added the comment:
Here's a patch, in final form, that replaces fsum with an lsum-based
implementation. In brief, the accumulated sum-so-far is represented in
huge_integer * 2**(smallest_possible_exponent)
and the huge_integer is stored in base 2**30, with a signed-digit representation (digits
in the range [-2**29, 2**29).
What are the chances of getting this in before next week's beta?
I did toy with a base 2**52 version, with digits stored as doubles. It's attractive for
a couple of reasons: (1) each 53-bit double straddles exactly two digits, which makes
the inner loop more predictable and removes some branches, and (2) one can make some
optimizations (e.g. being sloppy about transferring single-bit carries to the next digit
up) based on the assumption that the input is unlikely to have more than 2**51 summands. The result was slightly faster on OS X, and slower on Linux; the final rounding code
also became a little more complicated (as a result of not being able to do bit
operations on a double easily), and making sure that things work for non IEEE doubles is
a bit of a pain. So in the end I abandoned this approach.
Added file: http://bugs.python.org/file11108/fsum11.patch
Python tracker <report at bugs.python.org>
More information about the Python-bugs-list