As Will pointed out, my current approach to improving the midi latency issues
collapses when things get busy. I think I've actually known it for a while,
but I've somehow been willing it not to be true, which I think derives from
a healthy dose of "bear of very little brain" syndrome.
The way I've been going about it brings all the noteon object creations, and
their significant new memory allocations, _into_ the realtime path. When things
get even moderately busy, it fails miserably. Somehow I think I've been kidding
my self for a while that this wasn't happening. It is.
The first thing that comes to mind for solving it is some kind of pool
of prepared notes, recycling used ones back into the pool once they're done.
That's going to be some seriously complex code, quite likely beyond my
capabilities.
Meanwhile I think the best I can offer is a slight advance on 0.58.1,
incorporating the new midi handling code but leaving the interleaving of midi
events and audio generation as it's always been. In practical terms all that
can really offer is arguably "nicer" code but no real performance improvements.
And of course, there's still the segfault issues to resolve. Such is life.
cheers, Cal
