Re: [ecasound] 64bit float DSProcessing?

From: Sergei Steshenko <sergstesh@email-addr-hidden>
Date: Sun Feb 21 2010 - 00:15:01 EET

--- On Fri, 2/19/10, Klaus Schulz <kls.schlz@email-addr-hidden> wrote:

> From: Klaus Schulz <kls.schlz@email-addr-hidden>
> Subject: Re: [ecasound] 64bit float DSProcessing?
> To: "Kai Vehmanen" <kvehmanen@email-addr-hidden>
> Cc: ecasound-list@email-addr-hidden
> Date: Friday, February 19, 2010, 6:39 AM
> That's a good point. From a theoretical standpoint it
> is more then obvious - isn't it. 64 bit processing
> througout just causes lower errors.  To prove it by
> measurements will be pretty tricky
> I'd guess. A test on several high resolution audio
> system would be feasible I guess. (How about bringing this
> up on one of the annual Linux Audio Conferences?)
> Cheers
> \Klaus

No, it's that simple. For example, DFT for float (32 bits) makes sense
(strictly speaking) for no more than 4K samples - 'sin' and 'cos'
function values become indistinguishable for certain adjacent arguments
(where the functions peak/dip).

Because the function values become indistinguishable, true splitting of
spectrum into frequency bins becomes impossible - mathematically speaking,
orthogonality is lost.

It's easily measurable - the difference between float and double manifests
itself in like several thousands samples (44.1K) in a ~3minutes song. Which
is less than a second, but, still ...



Download Intel&#174; Parallel Studio Eval
Try the new software tools for yourself. Speed compiling, find bugs
proactively, and fine-tune applications for parallel performance.
See why Intel Parallel Studio got high marks during beta.
Ecasound-list mailing list
Received on Sun Feb 21 00:15:07 2010

This archive was generated by hypermail 2.1.8 : Sun Feb 21 2010 - 00:15:08 EET