There is some sort of memory corruption going on and these plugins
usually segfault on the audio thread. I'm clueless as to what could be
causing this (I wouldn't be surprised if this is caused by an
interaction between the VST3 SDK and Wine's Windows.h implementation),
but it's probably best to disable loading 32-bit VST3 plugins completely
until this has been fixed.
Or technically, two, since the group bridge also does the same loop. We
no longer need special handling for VST2 and VST3 plugins, so we can
simplify things a bit here.
Even though the interface ID is passed as an FIDString and there is a
function to compare FIDStrings obviously doesn't mean that you're
supposed to use them! Duh.
`std::char_traits<wchar_t>::length()` would return the wrong length for
a wide string under Wine, and thus also for `String128`. This would
cause parts of the strings to get cut off. As a workaround we'll just
serialize the entire container including all null bytes at the end.
Since the restart will always be called from another thread, and when a
plugin asks for a restart during initialization this can quickly lead to
issues.
This is in some cases needed to get decent performance in REAPER, as
REAPER seems to query this information (which cannot change without the
plugin requesting a restart) four times per second.
This can absolutely destroy performance when using VST3 plugins. REAPER
also does something similar with bus information from the audio
thread (for which we had to build in a caching layer), but with these
plugin bridges a VST3 plugin will quickly use ten times as much CPU time
as it should be using.
In commit 9788f21e0e we changed the input
focus grabbing to only grab focus once per event cycle, because
otherwise Bitwig would fight over who was getting focus. This is a bit
hacky, and with this approach you could in theory still have an
unnecessary focus grab 60 times per second (i.e. the default event
polling rate). Now we instead check the current focus target, and only
request focus when it's actually necessary.
It sort of goes against yabridge's principles to not do these
unnecessary `audioMasterGetTime()` calls if the plugin does that, but it
also hurts the user experience to not have this as a default.
This works around an issue with REAPER. During every processing cycle
REAPER would query how many input and output busses we have, and it
would then enumerate over all of those busses. This meant that if a VST3
plugin has 32 output busses, then REAPER will do 34 extra function calls
before processing audio.
We didn't initialize the field, and we also didn't copy the updated
value back (since everything else is a pointer to the original data).
This fixes audio channels in REAPER randomly being silence, as this
field would otherwise be uninitialized.
Thanks a lot to @kytdkut for finding this issue!
Ever since 0bed2b7bc0 REAPER will randomly
not play back one or more audio channels for plugins that support
IPlugViewContentScaleSuport. If you return `kNotImplemented` then this
bug doesn't occur. REAPER should definitely fix this soon. With
`kResultFalse` this issue still occurs, hence this change so you can use
`vst3_no_scale` to work around this REAPER bug.
I haven't seen this cause issues myself, but it would not surprise me if
a plugin that produces a constant stream of FIXMEs would steal resources
from audio processing if the threads that are relaying those messages
are set to SCHED_FIFO.
FabFilter plugins will exchange messages that have to be handled from
the GUI thread, or they'll get stuck waiting on a synchronisation
object. This probably hurts GUI performance significantly but luckily it
only affects Ardour.
We'll periodically copy the scheduling priorities from the host's audio
threads to the Wine plugin host's audio threads. The overhead of doing
this is about 1 microsecond on my system, so doing this every cycle
really adds up. But getting the Unix epoch time and comparing some
timestamps has a neglegible overhead, so this should give you the best
of both worlds.
Next we'll do the same thing for VST3 plugins.
As suggested by @jhernberg
Robbert van der Helm