What happens if you just have the maximum number of voices you think you'd ever need and [switch~] them on and off as needed?
Since you have a large patch, I'd be curious to know how much memory is taken up by having the switched off voices just sitting there.
-Jonathan
That's what I am doing anyway. Memory is not an issue. There is obviously no change in memory by simply switching the voices on or off. After I got most of the control stuff as well as a large number of the [receive] objects out of the way the patch doesn't need that much cpu time at all unless the voices are turned on and playing.
Now that the switched off voices are not hardly doing anything anymore there is no more need to adjust the voice number as it was the case before I got rid of a whole bunch of [receive] objects and cut most of the control messages when the [switch~] object gets turned off. In certain cases I can limit the voice number with different [poly] objects.
One shouldn't underestimate the cpu load when several hundreds of midi controllers per second are modulating hundreds of parameters and the get multiplied by 16 voices constantly because the control messages are still active all of the time.
Ingo
----- Original Message -----
From: Ingo ingo@miamiwave.com To: 'Roman Haefeli' reduzent@gmail.com; 'Ludwig Maes'
Cc: 'Pd List' pd-list@iem.at Sent: Thursday, September 29, 2011 5:33 AM Subject: Re: [PD] Fwd: Variable number of objects?
Actually, I just tried again to simply copy a couple of voices and it
only
took a fraction of a second with a very short dropout - even with the
dsp
on.
I have recently installed Natty instead of Lucid on a new machine. Maybe
it
has something to do with realloc that Mathieu mentioned.
So it looks like dynamic patching of voices isn't "that" much of a problem here anymore. It still takes 7-8 seconds to create 16 voices in my case
with
the dsp off (12 with the dsp on) but that's still faster than restarting
the
patch. I would never have checked again if nobody would have mentioned
it.
I have attached a patch that I used for testing. These voices were
receiving
their input with [receive] so no connections were needed. If you are
using
wired inlets you will have to dynamically add the connections of course.
I added a cut & paste at the end because for some reasons the voices didn't get initialized correctly. Not sure if this is needed for other voice-abstractions.
Ingo
-----Ursprüngliche Nachricht----- Von: pd-list-bounces@iem.at [mailto:pd-list-bounces@iem.at] Im Auftrag
von
Roman Haefeli Gesendet: Donnerstag, 29. September 2011 08:36 An: Ludwig Maes Cc: Pd List Betreff: Re: [PD] Fwd: Variable number of objects?
On Wed, 2011-09-28 at 19:29 +0200, Ludwig Maes wrote:
---------- Forwarded message ---------- From: Ludwig Maes ludwig.maes@gmail.com Date: 28 September 2011 19:29 Subject: Re: [PD] Variable number of objects? To: Ingo ingo@miamiwave.com
I actually meant more in general, also for non-~ signals (i.e. also control rate .pd patches). I referred to polysynth such that people would see more easily what I meant. Are there really no such primitives? That seems like quite a restriction...
How can that take 10 seconds?? I dont see what would cause such a
huge
overhead, i'd expect an increase in computations & memory
though (say
from 10 voices to 11: 10% increase in cpu workload & ram dedicated
to
these voices..., I fail to see what would necessitate a long initialization...)
also, how is it done even with the long delays?
As I understand it, the whole DSP is recompiled whenever it is
changed.
So, if you have a very large patch (and Ingo seems to be an expert in very large patches) and you create another voice, it's easily possible to eat up quite some time.
Also, it's probably much slower the first time, if the voice
abstraction
is built of many other abstractions, which need to be read from disk.
But I guess you are right about the increase in CPU workload _after_
the
DSP graph has been recompiled: A switch from 10 two 11 instances is expected to show only an increase of 10% in CPU usage.
It's been said, that often you can gain quite some time while turning off DSP during dynamic instantiation. But I guess, this makes only a difference when performing several instantiations at the same time.
When
DSP is on, each cycle would cause a complete DSP graph recompilation, whereas when DSP is off, it's only recompiled once (after turning it
on
again).
Roman
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list