[Openmcl-devel] Fun with Measurements
bfulg at pacbell.net
Wed Nov 1 05:51:29 UTC 2006
On Oct 29, 2006, at 5:54 PM, Gary Byers wrote:
> I was going to say that there wasn't likely to be a single smoking
> gun here; there
> may not be, but this may be pretty close to it.
> - the ppc32 version sometimes goes to great lengths to avoid
> calculating both
> the quotient and remainder in things like TRUNCATE when it can
> tell that only
> one of them will be used; the 64-bit versions tend to compute
> both results
> and one of them is often ignored.
> I did a 64-bit version of CCL::%BIGNUM-BIGNUM-GCD that does
> destructive shifts
> on copies of its arguments; the GCD call above went from 15
> milliseconds to
> less than 1, and the (COMPUTE-PI-DECIMAL 200) test went from about
> 1.5 seconds
> to about .12 seconds, so (at least in this case) there seems to
> have been a
> smoking gun.
I reran the benchmarks and posted them to the Wiki page. What a huge
Test Original 64 bit Revised 64-bit
PI-DECIMAL/SMALL 10.49 times 0.47 times.
PI-DECIMAL/BIG 8.13 0.71%
PI-RATIOS 7.79 2.60
Very nice work!
More information about the Openmcl-devel