To do
Schnell_pi is far from perfect. It is version 1.0
with only the very basic computation possible. Here
are some plans for improving the programs:
- Additional tweaks to reduce the memory read/writes.
I expect only small gains here. I may improve the speed
by 10% for 1M digits, not much more;
- Decrease the memory needed. I know how to reduce
it from 10.375*(number of digits) to 6.25*(number of digits).
This is not very difficult, but boring. At the same time,
this will increase the speed for large computations
(above 8 M digits);
- Write a multi-threaded version to be used on two-processor
PC. This should be not too difficult. However, the speedup
will not be a factor 2, although the FFT is easy to
parallelize over 2 processors. The memory bus contention
should limit the speedup around 1.5.
On a longer time scale, I plan to develop a version
with intermediate disk read/writes - à la PiFast -
just to see who is the fastest between AGM and Chudnovsky.
dominique@delande.nom.fr
Last modified: Fri Aug 10 18:14:36 CEST 2001