search

jpcima / adljack

OPL3/OPN2 synthesizer using ADLMIDI and OPNMIDI, for Jack and cross-platform audio
Boost Software License 1.0
41 stars 3 forks source link

MIDI latency work #9

Open jpcima opened 5 years ago

jpcima commented 5 years ago

Feature request by @Wohlstand

Process in segmented buffers for latency reduction. A work was made in the timestamps branch.

Problem RtMidi sends wrong timestamps! :imp: The problem is present in different APIs. (ALSA, WinMM) Certainly, the timestamping code is not very robust and needs rework..

At the first event with Wine and DirectSound, RtMidi can emit a timestamp 4292560.577000 seconds.

In ALSA, there is seemingly a wrong calculation in a difference of timespec values... I converted the computation to using floating point, since after all timeStamp is typed double... This particular problem seemed resolved, but latency problems were still present.. rtmidi-alsa.diff.gz

jpcima commented 5 years ago

I've made workarounds, now it's possible for winMM to work. In the latency range which used by the standalone Windows software, adljack seems to work now better. But at higher latencies in 100ths ms, it becomes more and more crap as value goes up.

I've raised MIDI buffer size as necessary, because now it must have capacity for accumulation of MIDI events for a longer duration. For now, MIDI input does not act in RT way, waiting to have available buffer space if it's full. No problem for ALSA MIDI, but in case of Jack, it can create xrun.

No work was pushed to the master branch, while ALSA remains broken and needs identifying the precise problem.

Wohlstand commented 5 years ago

I saw you have took some fixes into RtMidi library. Does this helps you to solve ALSA MIDI trouble with time stamps?

jpcima commented 5 years ago

At least, the timestamp don't emit crazy values at this point, but still some trouble remain. I've not merged yet, into this adljack branch, the RtMidi patches I got accepted upstream.

The last time tried, ALSA had still trouble despite patching, the midi event delays were not all well spaced at arrival. I think, I have demonstrated that RtMidi has a really weak point on timestamping and needs more fixing.

However, the winMM situation is well improved, in the moderate buffer size values, such as upper bounds savihost has. Still, in the very high range (think 0.5s and up), the MIDI gets increasingly desynced, meaning still that something is wrong with damn timestamping.

At one point, I should compare RtMidi vs Juce code in ADLplug, because that other manages to gets the timing all perfect without any troubles.

jpcima commented 5 years ago

@Wohlstand so, right now I merged in the timestamps code. It's not perfect but I expect it to help the latency problem. Please give a try on your side.

luzpaz commented 7 months ago

CC @Wohlstand

Wohlstand commented 7 months ago

At some moment, I tweaked the code to resolve another problem that I found by myself that latency would get unexpectedly bloat and lead a crash later.

Wohlstand commented 5 months ago

I forgot to say, a while ago I adjusted some internal delays, and problem has gone.