[Openmcldevel] Fun with Measurements
Eric Marsden
eric.marsden at free.fr
Mon Oct 30 19:42:19 UTC 2006
>>>>> "bf" == Brent Fulgham <bfulg at pacbell.net> writes:
,
 ;; this can be 1e6 on most compilers, but for COMPUTEPIDECIMAL on
 ;; OpenMCL we lose lotsa precision
 (defun fuzzyeql (a b)
 (< (abs (/ ( a b) b)) 1e4))

 (defun runpidecimal/small ()
 (assert (fuzzyeql pi (/ (computepidecimal 200) (expt 10 198)))))

 (defun runpidecimal/big ()
 (assert (fuzzyeql pi (/ (computepidecimal 1000) (expt 10 998)))))
`
bf> The comment regarding OpenMCL is not mine, this was in the benchmark
bf> sources. Does anyone know why this comment is being made?
I wrote this comment. My (also somewhat fuzzy) recollection is that
OpenMCL on 32bit MacOS X was getting much less precision than the
other implementations I was testing. This may be due to lower
precision in the PowerPC FPU (64bit rather than 80bit on x86 given
various assumptions). I just checked that OpenMCL on AMD64 works
with a fuzziness of 1e6.

Eric Marsden
More information about the Openmcldevel
mailing list