[PD] Multiprocessing in Pd (was DSP and Gem in the same instance of Pd)

Fri Apr 1 07:09:28 CEST 2016

Pd's constraints make the automatic allocation unlikely.

Read this:
http://puredata.info/community/conventions/convention09/puckette.pdf

'Since at least 1990, users and critics of Max/FTS have observed that it
would be desirable for objects to be automatically allocated to processors
in a way that would minimize the bandwidth of interconnections between the
objects. This would free the user from the cumbersome task of understanding
the actual flow of data between objects in the patch; the software would
automatically assess that. This didn’t prove practical, for two reasons.
First, as has long been well known, one can’t compute the quantity of data
that will flow between any given pair of objects in a patch (at least, not
if the patching language is able to solve arbitrary computing problems).
Predicting how much data will flow where is hopeless. The second problem is
that nobody has been able to make an expressive patching language that
doesn’t depend on objects sharing data. In Max/FTS (and in Pd as well) this
takes the form of “named” objects such as arrays. Any automatic
distribution of patches that allows accessing arrays would have to place
every object that accesses any particular array on the same processor, or
else use some kind of locking mechanism that would be unlikely to work in
real time. Also, any situation in which there is of recombination of
message fanout would require that both message paths be synchronized, i.e.,
that both message paths go through the same itinerary of processors or be
otherwise delay-equalized. In combination, these constraints would require
that, for complete transparency, almost any interesting patch would have to
reside on a single processor. It appears to be an inescapable fact that
multiprocessing has non-hideable effects on the execution of “patches” and
can’t effectively be carried out without the user’s active participation.'

It might be easier if Pd used a system of buses for routing rather than
arbitrarily drawn patch connections, or if a graphical patching environment
had a good way of implementing something like SuperCollider's Synth, which
works with a flexible node order but has relatively limited means of input
and output. Here's what I see as the bare minimum of what we'd need to
address to make your wish a reality:

1) Unit generators instantiated in Pd have to exist somewhere in a running
patch to output. This is in distinction to SC3 and csound, where instances
of Synths or instrument templates are instantiated and destroyed. In the
latter two, the order of creation and destruction of instruments (and in
csound, the order they're defined in the orchestra) matters a lot in the
DSP graph, which makes it more predictable. SC3 also has user methods for
ordering nodes.

2) Connections in Pd are flexible and arbitrarily complex, which makes the
DSP graph a lot more ramified than a mixer model with buses, inserts, aux
sends, etc., and therefore much less predictable in the abstract.

3) Pd is deterministic, which means that (as noted in the quote above), any
memory sharing across threads would need to involve locks, which can be
killer in real-time, not to mention difficult to scale and guarantee thread
safety. [pd~] communicates via FIFO because it needs to be able to keep
messages and audio in sync by block.

I'm sure there's more.

On Thu, Mar 31, 2016 at 11:07 PM, William Huston <williamahuston at gmail.com>
wrote:

> Slightly OT-- but related:
> I'm going to throw out a wish-list item
> which is probably impossible or very hard to implement,
>
> Find a way for the DSP Graph compiler to naturally
> break up the task into small chunks, which all use shared
> memory in a thread-safe way, so that the PD job automagically
> spreads itself out over available cores without any special
> work by the programmer.
>
> Now that I've got my big, fast, lots of ram, 6-core AMD
> box running again, I notice that I can run MUCH larger
> graphs there than on my Raspberry Pi.
>
> But I think it's just a raw CPU speed, or Cache Speed, or
> speed to RAM which matters, and not the number of cores.
>
> I notice that on my Pi there seems to be two processes, one for
> the GUI and one for the DSP. 2 cores are wasted unless I use
> the [pd~] object, and I have to basically guess how to split up the job.
>
> I know, I'm dreaming here....
>
>
>
>
>
>
> On Thu, Mar 31, 2016 at 10:57 PM, Lucas Cordiviola <lucarda27 at hotmail.com>
> wrote:
>
>> And more,
>>
>> A single thread calculation divided between 2 cores in its 1 core time is
>> more stable.
>>
>> 1 0
>> 0 1
>> 1 0
>> 0 1
>>
>> transistors have more idle time.
>>
>> Mensaje telepatico asistido por maquinas.
>>
>> ------------------------------
>> From: brbrofsvl at gmail.com
>> Date: Thu, 31 Mar 2016 22:45:57 -0400
>> Subject: Re: [PD] DSP and Gem in the same instance of Pd
>> To: jancsika at yahoo.com
>> CC: lucarda27 at hotmail.com; pd-list at lists.iem.at
>>
>>
>> Right, so the point of [pd~] is that the OS can now throw whatever is
>> going on in the subprocess onto another core. The idea from what I've heard
>> for Gem is that you can leave the DSP off in the [pd~] instance, run Gem
>> from there (on another core, possibly). Then if together they would have
>> maxed out one core they could split the work among two and proceed in time.
>>
>> But if the problem is that Gem has to wait for something to happen
>> elsewhere before it can proceed, it won't help. Kind of in the same way
>> that running an infinite [until] loop on the subprocess will halt the main
>> process, too.
>>
>> On Thu, Mar 31, 2016 at 9:24 PM, Jonathan Wilkes via Pd-list <
>> pd-list at lists.iem.at> wrote:
>>
>> But [cpu_hungry_hippo~] needs input from [pd~] in order to
>> compute its output.  So [pd~] must send output before [cpu_hungry_hippo~]
>> can execute its perform routine.
>>
>>
>> On Thursday, March 31, 2016 9:17 PM, Lucas Cordiviola <
>> lucarda27 at hotmail.com> wrote:
>>
>>
>> Isn`t
>>
>> [pd~] <-- some dsp stuff going on in here
>>
>> To take advantage of multi-core CPUs?
>>
>> Mensaje telepatico asistido por maquinas.
>>
>> ------------------------------
>> Date: Fri, 1 Apr 2016 00:37:26 +0000
>> To: brbrofsvl at gmail.com; reduzent at gmail.com
>> CC: pd-list at lists.iem.at
>> Subject: Re: [PD] DSP and Gem in the same instance of Pd
>> From: pd-list at lists.iem.at
>>
>> I'm not sure I understand [pd~].  Consider:
>>
>> [foo~]
>> |
>> [pd~] <-- some dsp stuff going on in here
>> |
>> [cpu_hungy_hippo~]
>>
>> How does [pd~] help me in this case, as opposed to just putting the
>> "dsp stuff" directly in the patch?
>>
>> And in general, how is the super-process able to anything
>> other than block when waiting for output from [pd~]?
>>
>> -Jonathan
>>
>>
>>
>> On Thursday, March 31, 2016 5:17 PM, Matt Barber <brbrofsvl at gmail.com>
>> wrote:
>>
>>
>> One other thing that's helped in an emergency is increasing Pd's audio
>> buffer in the preferences.
>>
>> One thing I've heard of but never tried is running Gem from a slave
>> instance in [pd~]. I don't know enough about it to know whether this could
>> work or why; it might just be a rain dance.
>>
>> On Thu, Mar 31, 2016 at 7:16 AM, Roman Haefeli <reduzent at gmail.com>
>> wrote:
>>
>> On Thu, 2016-03-31 at 11:35 +0200, cyrille henry wrote:
>> >
>> > Le 31/03/2016 11:19, Roman Haefeli a écrit :
>> > >
>> > > BTW: Why does the graphics rendering|clock have precedence over the
>> > > audio rendering (at least, it seems to be like that in Pure
>> Data/Gem)? I
>> > > guess most softwares do it the other way around, since clicks are much
>> > > more noticeable than a frame being a few milliseconds late.
>> >
>> > Gem have no precedence over audio : they both have the same priority.
>> > when having priorities on audio, the openGL rendering did not have
>> > fixed frame rate, and it's not possible any-more to have smooth hight
>> > speed movement.
>> >
>> > So, i like the way it is, even if it cause implementation problem.
>>
>> Oh, now since I understand, I like the way it is, too ;-)
>>
>> > one possible explanation of your problem is that you are rendering a
>> > 60 fps, and that openGL is sync on the 60fps screen.
>> > You can have jitter between the 2 different 60fps clock. If Gem is
>> > waiting for the screen, then everything (including audio) is on pause.
>>
>> That is exactly what I was doing.
>>
>> > if this is the cause of your problem, then reduce Gem fps to 59, or
>> > remove openGL syncro (sync to vblank).
>>
>> This is exactly what helped (reducing fps to 59). Thanks for your sharp
>> thinking.
>>
>> Roman
>>
>>
>> _______________________________________________
>> Pd-list at lists.iem.at mailing list
>> UNSUBSCRIBE and account-management ->
>> http://lists.puredata.info/listinfo/pd-list
>>
>>
>>
>> _______________________________________________
>> Pd-list at lists.iem.at mailing list
>> UNSUBSCRIBE and account-management ->
>> http://lists.puredata.info/listinfo/pd-list
>>
>>
>>
>> _______________________________________________ Pd-list at lists.iem.at
>> mailing list UNSUBSCRIBE and account-management ->
>> http://lists.puredata.info/listinfo/pd-list
>>
>>
>>
>> _______________________________________________
>> Pd-list at lists.iem.at mailing list
>> UNSUBSCRIBE and account-management ->
>> http://lists.puredata.info/listinfo/pd-list
>>
>>
>>
>> _______________________________________________
>> Pd-list at lists.iem.at mailing list
>> UNSUBSCRIBE and account-management ->
>> http://lists.puredata.info/listinfo/pd-list
>>
>>
>
>
> --
> --
> May you, and all beings
> be happy and free from suffering :)
> -- ancient Buddhist Prayer (Metta)
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.puredata.info/pipermail/pd-list/attachments/20160401/71e868d8/attachment-0001.html>