[gambit-list] Re: Some tweaks to _num.scm
Marc Feeley
feeley at iro.umontreal.ca
Thu May 18 15:41:59 EDT 2006
Thanks Brad! I've integrated the new _num.scm into the runtime and
all the tests I have pass.
Marc
On 18-May-06, at 3:08 PM, Bradley Lucier wrote:
> Marc:
>
> Here's a _num.scm with some small tweaks to the fft multiplication
> code. Memory allocation is cut by roughly 40% when squaring large
> integers; memory allocation is also cut when the length of one
> integer is > 2^M bits, but the some of the lengths of the two
> integers is <= 2^{2M} bits. And I've defined a new macro for the
> constant inexact-pi/4 to remove a small dependence on the accuracy
> of the atan routine on fft multiplication. (It still depends on
> the accuracy of sin and cos in the interval [0,pi/4), which, on all
> modern systems, is excellent.)
>
> So, we have
>
> Old:
>
> > (time (* 0 (* (expt 3 1000000) (expt 3 100001))))
> (time (* 0 (* (expt 3 1000000) (expt 3 100001))))
> 427 ms real time
> 424 ms cpu time (340 user, 84 system)
> 22 collections accounting for 45 ms real time (14 user, 30 system)
> 43347960 bytes allocated
> no minor faults
> no major faults
>
> New:
>
> > (time (* 0 (* (expt 3 1000000) (expt 3 100001))))
> (time (* 0 (* (expt 3 1000000) (expt 3 100001))))
> 422 ms real time
> 388 ms cpu time (334 user, 54 system)
> 24 collections accounting for 38 ms real time (23 user, 16 system)
> 38710968 bytes allocated
> no minor faults
> no major faults
> 0
>
> Old:
>
> > (time (* 0 (expt 3 1000000)))
> (time (* 0 (expt 3 1000000)))
> 200 ms real time
> 194 ms cpu time (162 user, 32 system)
> 13 collections accounting for 17 ms real time (9 user, 8 system)
> 17846920 bytes allocated
> no minor faults
> no major faults
> 0
>
> New:
>
> > (time (* 0 (expt 3 1000000)))
> (time (* 0 (expt 3 1000000)))
> 172 ms real time
> 167 ms cpu time (150 user, 17 system)
> 10 collections accounting for 12 ms real time (9 user, 3 system)
> 13683504 bytes allocated
> no minor faults
> no major faults
> 0
>
> Brad
>
> <_num.scm>
More information about the Gambit-list
mailing list