19 Feb
2007
19 Feb
'07
1:55 p.m.
Stefano D'Angelo wrote:
Actually, *
That, per se, is no optimization, but moving stuff around can help
making the other rules apply.
Like, if you have:
eq -> LTIfilter -> eq ->
you can first swap the first two:
LTIfilter -> eq -> eq ->
and then reduce the second and third:
LTIfilter -> sum_of_eqs ->
Ok. Filter chains without outputs nor side-effects. (Like optimizing
away pure functions)
What's "naspro"?
2007/2/19, Camilo Polyméris
<cpolymeris(a)gmx.net>et>:
I meant Jeff's idea: the simplification of
parallel filters. He
mentioned the SynthEdit API using it.
Stefano D'Angelo wrote:
Yes, I agree I think if one comes up with a couple of "rules" like that,
it could be possible to design a system which automatically simplifies
processing networks.
To recap:
* If two parallel filters have equal control inputs and their outputs
are summed, replace with one filter and feed with summed inputs.
Maybe too specific... maybe also plugins with different control inputs
can be merged, I must see this. 2007/2/19, Camilo Polyméris
<cpolymeris(a)gmx.net>et>:
> Jeff McClintock wrote:
> > >>>I actually don't know how many plugins are LTI, but, for
example, a
> > >>>lot of delays, reverbs,
choruses, eq. filters, compressors,
> modulators
> > >>>and "sound mixers" should be, and that's quite enough
after all.
> >
> > Yeah, It's a good optimization. The SynthEdit plugin API supports
> > inputs being flagged as 'linear', if several such plugins are
used in
> > parallel they are automatically
collapsed into a single instance
which
> > is fed the summed signals of the
original plugins. Plugin are
> > collapsed only when their control inputs are the same.
> >
> > BEFORE optimation:
> >
> > [plugin]-->[delay1]------>
> > [plugin]-->[delay2]-/
> >
> > AFTER:
> >
> > [plugin]--->[delay1]--->
> > [plugin]-/
> >
> > e.g. two parallel 100ms delays are combined. Two different length
> > delays aren't.
> >
> > This is most useful in synth patches where each voice is an
> > identical parallel sub-patch.
> >
> >
> > Jeff McClintock
> >
> How often are more than one plugin with the same control inputs
used in
paralel?
I was rather thinking of colapsing (or swapping) plugins in
series. They'd have to be linear and time invariant, of course.
Or maybe plugins could 'know' how to colapse themselves, sort of like
overriding Plugin::operator+(const Plugin&), to use a C++ metaphor.
Well, stereo sounds passing through mono plugins is one case.
However as Jeff describes this optimization, it is applicable when
output signals are summed, and I don't know how often it happens.
Anyway it is another idea to optimize processing for linear plugins,
definitively not something to discard.
This makes me think that some common basic "pieces" like mixers and
delay filters can have special properties which involve even more
aggressive optimization. Maybe it's worth considering how this special
blocks could be developed and used.
Stefano
* If two serial filters are LTI, they impulse
response can be added to
one filter.
Added = multiplied :-) * If two serial filters are LTI, and their
impulse response is unknown,
they can be swapped.
Yes, but why? * Filter "classes" could know how to
merge to instances into one. Those
instances may even cancel each other out.
Yes
* Remove filter chains which have no outputs.
Absolutely not: what about a GUI oscillator?
etc... With a
little thinking and some formal work, one could come up
with more ideas like those.
I think too that this is an interesting path to follow: NASPRO (the
wrapper) will absolutely go this way, just after wrapping LADSPA,
DSSI, LV2 (without extensions) and similar.
When LV2 extensions will be implemented then work on these stuff will
begin.
Software like
puredata and jMax (which use such "common basic pieces" in
many diferent configurations) could benefit from such a system. I looked
at their websites, but could not find any references to similar ideas.
Good, another use for it :-)
Stefano