Re: [fluid-dev] Thread safety long-term thoughts

[josh]

Quoting David Henningsson <address@hidden>:
> Btw; with the sample timers we also have the entire midi file player
> (with its fopen calls etc) inside the audio thread. I'm guilty of that,
> and it should be fixed by inserting a sequencer between the player and
> the audio thread, at least in real-time use cases. Perhaps something
> for 1.1.2.


This scenario is less critical IMO, since the audio buffers could be  
increased without issue (so what if the MIDI file plays 100ms later).   
Good to know though.  We may want FluidSynth to be smarter about  
selecting default buffer sizes in the future, depending on the use case.


> > I'm sure the biggest time consumer in this regard is the note-on processing.
>
> Assuming the note-on call makes no system calls, and the soundfont
> isn't paged out by the OS, the note-on event should complete within a
> good fixed time, or is there something else bugging us?


What bugs me about it is that the current note-on processing isn't  
very efficient and is affected by the complexity of the underlying  
instruments.  I'm not sure how much of an issue it really is, but if a  
lot of note-on events occur in a short period of time, it could lead  
to xruns.  Prior versions of FluidSynth processed the note-on event in  
whatever thread was sending the note-on, which was usually the MIDI  
thread which was usually running lower priority, so note-on processing  
could not starve the audio thread.

It would be good to improve the efficiency of the internal note-on  
event handlers, as I described previously, but it may also make sense  
to process these outside of the synthesis thread in the multi-thread  
case, especially since the note-on handler could be supplied by  
another application.


> > Here is an example of what I was talking about, in regards to the   
> > state machine depending heavily on the existing voices.  Say for   
> > example the pitch bend controller changes on a channel.  All voices  
> >  are then scanned and those which are active on the given channel   
> > are modulated in respect to the pitch bend controller.  The active   
> > voices and their parameters are private to the synthesis thread.  I  
> >  don't YET see what additional processing could be moved outside of  
> >  the synth context for most events, from what it is right now.    
> > Though I could see perhaps grouping controller changes, so only one  
> >  update/calculation occurs.
>
> Perhaps there is not that much to gain performance-wise from moving
> things out of audio thread context then. It would just feel less messy,
> I guess, if we splitted the synth object in two parts, one with strict
> real-time and one without.


I kind of felt like this sort of separation occurred with the thread  
safety changes, at least I have a pretty good view of them being  
separate, having delved into what was previously much messier.  I  
think there are improvements that can be made, but I think its a lot  
better off than before.  A lot of the calculations we are talking  
about, in response to events, are pretty minimal.  All of the ones  
that require mutex locking or other OS calls are handled in  
non-realtime.  One exception which should be fixed is in regards to  
SYSEX MIDI tuning events and Jack MIDI.


> And instead of creating shadow variables, all variables should belong
> the state machine, unless explicitly needed by the voices directly.

For clarification:  Shadow variables were provided in only 2 cases for  
the purpose of returning the most recent value assigned when querying  
their values.  The shadow values are used when querying the value  
only.  The 2 cases are presets, since atomic operations can't be used  
in that case (might be possible if there was preset reference  
counting) and polyphony, since polyphony is accessed fairly often in  
the synthesis thread and I thought it was better to shadow the value  
than have to atomically read it every time and also deal with issues  
that might arise if the value is changed while looping over voices.

> In the case above, the current pitch bend controller current value
> belongs to the state machine. The new value is sent to the audio
> thread. Btw, if someone tries to read the current pitch bend controller
> value just after having set it, will it work?

Yes it will work.  As far as I know, all values which can be queried  
should now work, as far as appearing immediately.  For all parameters,  
except presets and polyphony, the value is set and accessed atomically  
by all threads.  When an event occurs, it assigns the new value to the  
parameter and sends an update event to the synthesis thread.  The  
update event does not contain the value, but just tells the synth  
thread to update the parameter based on the latest value.  If an event  
is set from within the synth thread, its pretty much the same, except  
no queuing of the update occurs.

> > The way it is now, is much closer to the synth thread being a voice  
> >  renderer than it was with 1.0.9.  It seems like its about   
> > identifying additional stuff to move outside of it.  Note on events  
> >  are the only thing I can think of at the moment, that could use   
> > some improvement.  Can you think of any others?
>
> You probably have a better overview over the time and locking needed by
> various events than I have, but I'm thinking about presets, tuning,
> soundfont loading etc, but I guess they are already moved out of
> synthesis context.


Yes, they are already outside of synth context.  The stuff that is  
still inside synth context is within the _LOCAL function variants.


> > > The midi thread and shell threads will do the state machine work, when
> > > they make calls against the fluid_synth object.
> >
> > We still need to do something about the Jack MIDI case, like   
> > queuing the events back to another thread (return queue for   
> > example).  Like what is now being done with program changes.
>
> Something like that, let the non-realtime stuff complete when it's
> done, yet we must then queue all simple events only if there is
> non-realtime processing in progress, to prevent reordering... (Sigh.)


Re-ordering is indeed one to watch out for.  I hope there aren't any  
more cases like that currently.  Program changes are one case where  
there is some expectation that a device might take 10s of milliseconds  
or so, to load up a patch, so hopefully we can kind of assume that  
note-ons wont occur for a little time after a program change.  If its  
all happening from outside of synth context, then it will all be  
queued, without re-ordering issues, its only the Jack MIDI case that  
might be a problem.


> Perhaps an option to let the libfluidsynth user call a function
> periodically if he wants to skip the additional thread.


Well it isn't just for garbage collection now.  Its also being used to  
handle program changes, which should happen ASAP.


> Could we do something in create_audio_driver/delete_audio_driver to
> register with the synth, and say that it is now operating with an audio
> driver, and that means real-time operation? And if that is not the
> case, handle events directly. Likewise, we could make the MIDI drivers
> (and shell thread) register that we now have additional threads
> referencing the state machine, so the state machine must be
> multi-threaded.


Yeah, I think something like that could be good, rather than trying to  
auto detect it.  A simple API function like:
void fluid_synth_multi_thread_enable(fluid_synth_t *synth, int enable);

And/Or, like I mentioned before, we could try to improve the  
multi-thread use auto-detection.  I haven't fully looked into the  
scenario, but if we made all the public API event functions  
automatically enable multi-threading and then provide an alternative  
to fluid_synth_handle_midi_event, say  
fluid_synth_handle_midi_event_noqueue() that might suffice and  
somewhat simplify things.


> (Note: potential screwup with libfluidsynth users creating their own
> audio drivers, although I assume it was screwed up in 1.0.9 the same
> way as well.)


In regards to thread safety you mean?  As things are now, it just  
assumes it is multi-threaded, which means slightly less efficiency.


> // David


Josh