24 Jul
2005
24 Jul
'05
6:46 p.m.
http://plugin.org.uk/libgdither/
Bugfix release. Added patch for 64bit architecture support, thanks to
Andreas Steinmetz, and a bugfix to the shaped dither that increases its
signal to noise ratio. All projects that use this code should update.
The intention is that applications that require this function should
include hte code, rather than adding it as a depenency.
About libgdither:
Libgdither is a library for applying dithering to PCM audio sources.
Dithering is a process to remove the intermodulation distortion from
signals that are to be been reduced in bitdepth.
This code is (c) Steve Harris 2001-2005 and released under the GNU Public
Licence. Details may be found in the file COPYING.
It can do conversions between any combination of:
in out (optionally interleaved)
-------------------------------------------------------------
normalised mono float 8bit unsigned ints
normalised mono double 16bit signed ints
32bit signed ints
normalised float
normalised double
At any bitdepth supported by the input and output formats
25 Jul
25 Jul
3:37 p.m.
Steve Harris <S.W.Harris(a)ecs.soton.ac.uk>uk>:
Me being noisy all the time I still don't know the maths of making
noise. Would you care to explain the rationale in this snippet from
noise.h?
inline static float gdither_noise()
{
static uint32_t rnd = 23232323;
rnd = (rnd * 196314165) + 907633515;
return rnd * 2.3283064365387e-10f;
}
The 64bit fixes have rnd left over as the last uint32_t. Is this on
purpose?
Thanks in advance, Wolfgang
4:53 p.m.
On Mon, Jul 25, 2005 at 07:37:14PM +0200, Wolfgang Woehl wrote:
Me being noisy all the time I still don't know the
maths of making
noise. Would you care to explain the rationale in this snippet from
noise.h?
inline static float gdither_noise()
{
static uint32_t rnd = 23232323;
rnd = (rnd * 196314165) + 907633515;
return rnd * 2.3283064365387e-10f;
}
It's a 'linear congruential sequence generator', one of the simplest
and fastest ways to generate uniform random variables. For a (very
complete) analysis of this algorithm, see D.E. Knuth, "The Art of
Computer Programming", Vol.2. The final multiply reduces the output
to the interval [0,1].
The integer computation done is of the form
x = (x * a + b) mod m
and here m = 2^32. For this m, to get the best results, you need
a mod 8 = 3 or 5 (196314165 mod 8 = 5)
gcd (b, m) = 1 (so 1 would be as good as 907633515)
--
FA
26 Jul
26 Jul
6:46 a.m.
On Mon, Jul 25, 2005 at 08:56:49 +0200, fons adriaensen wrote:
On Mon, Jul 25, 2005 at 07:37:14PM +0200, Wolfgang
Woehl wrote:
Thanks :) Yes, note that its not much good as a random number generator as
its easily predictable, but in this case all thats required is that it
produces noise that is approximatly white (ie. equal energy in each
frequency band), and its efficient. This is the fallback generator, and
the calling application can provide its own at compile time, but it seems
reasonable according to the test processes.
- Steve
Me being noisy all the time I still don't
know the maths of making
noise. Would you care to explain the rationale in this snippet from
noise.h?
inline static float gdither_noise()
{
static uint32_t rnd = 23232323;
rnd = (rnd * 196314165) + 907633515;
return rnd * 2.3283064365387e-10f;
}
It's a 'linear congruential sequence generator', one of the simplest
and fastest ways to generate uniform random variables. For a (very
complete) analysis of this algorithm, see D.E. Knuth, "The Art of
Computer Programming", Vol.2. The final multiply reduces the output
to the interval [0,1]. 
1:51 p.m.
On Tue, Jul 26, 2005 at 09:45:45AM +0100, Steve Harris wrote:
Yes, note that its not much good as a random number
generator as
its easily predictable,
All 'psuedoramdom' generators are predictable, that doesn't make
them less 'good', except for cryptographic applications.
but in this case all thats required is that it
produces noise that is approximatly white (ie. equal energy in each
frequency band), and its efficient.
It's a perfectly good choice for this application. Its only problem
is when you use it to generate random *bits*. Witm m = 2^N, the
lower n bits will form of cycle of lenght 2^n, so e.g. the LS bit is
just altermating 0,1,0,1,0,1 -- not very random.
--
FA
2:25 p.m.
On Tue, Jul 26, 2005 at 05:51:09 +0200, Alfons Adriaensen wrote:
On Tue, Jul 26, 2005 at 09:45:45AM +0100, Steve Harris
wrote:
True, but I'm casting it to a float, so the LS 7-ish bits will be lost
anyway, and the ones above should have sufficiently long period that they
won't be noticable, even if it weren't used for dithering :)
- Steve
Yes, note that its not much good as a random
number generator as
its easily predictable,
All 'psuedoramdom' generators are predictable, that doesn't make
them less 'good', except for cryptographic applications.
but in this case all thats required is that it
produces noise that is approximatly white (ie. equal energy in each
frequency band), and its efficient.
It's a perfectly good choice for this application. Its only problem
is when you use it to generate random *bits*. Witm m = 2^N, the
lower n bits will form of cycle of lenght 2^n, so e.g. the LS bit is
just altermating 0,1,0,1,0,1 -- not very random.
7004
days inactive
7006
days old
linux-audio-dev@lists.linuxaudio.org
5 comments
4 participants
participants (4)
-
Alfons Adriaensen -
fons adriaensen -
Steve Harris -
Wolfgang Woehl