Doubles versus floats

[Michael Gogins-2]

I have completed some basic tests of Csound performance and audio precision,
comparing the float configuration of Csound 5.11.1 with the double
configuration, both built with MinGW 4.40/MSys and running on Windows XP. I
used the examples/python/CsoundPerformanceTest.py script which I have
checked into Csound CVS.

I performed double-blind ABX listening tests. Over the course of about half
an hour, I was able to discriminate two short sections of the test
soundfile, rendered at 96000 Hz to a stereo float sample soundfile at ksmps
= 1, with 0.08% probability of error. I am reasonably confident that I could
have found another spot or two in the soundfile where differences stood out.
These were either heavily filtered sounds, or physical models involving
delay lines. It is rather tedious to identify these audibly different
segments, which are at most a few seconds long, but once identified, one can
hear the difference just about every time. I would think that in a very few
cases the differences could become artistically significant -- but only in a
very few cases.

I perfomed performance tests for master-quality rendering (same as for the
listening test), CD-quality rendering at 44100 Hz stereo to 16 bit soundfile
at ksmps = 100, and real-time audio. I took 8 samples each of master and CD,
and 6 samples each of audio rendering. A two-tailed t-test showed that there
is a difference in performance between float and double configurations at
the P=0.001 confidence level for master quality, P=0.018 confidence level
for CD quality, and P=0.049 for real-time performance. We can be reasonably
confident then that there is a real advantage in performance of the float
configuration over the double configuration. The difference appears to be on
the order of 10% to 15% faster for the float configuration.

These results are consistent with earlier, simpler tests.

At this point, I have reached no conclusion about which configuration to
package and release.

Regards,
Mike



------------------------------------------------------------------------------
Come build with us! The BlackBerry® Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9-12, 2009. Register now!
http://p.sf.net/sfu/devconf
_______________________________________________
Csound-devel mailing list
Csound-devel@...
https://lists.sourceforge.net/lists/listinfo/csound-devel

[jpff-2]

> I have completed some basic tests of Csound performance and audio
> precision,
> comparing the float configuration of Csound 5.11.1 with the double
> configuration, both built with MinGW 4.40/MSys and running on Windows XP.
> I
> used the examples/python/CsoundPerformanceTest.py script which I have
> checked into Csound CVS.
>
.....

I take it this is on 32bit Windows.  I am suspecting that on a 64bit
machine one might see different results.  I suppose I ought to do th
eexperiments


------------------------------------------------------------------------------
Come build with us! The BlackBerry® Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9-12, 2009. Register now!
http://p.sf.net/sfu/devconf
_______________________________________________
Csound-devel mailing list
Csound-devel@...
https://lists.sourceforge.net/lists/listinfo/csound-devel