I'm not saying sample-accurate, because it turns out, Fluidsynth has an internal buffer size and thus timing granularity of 64 frames. So 64 frames is the max. accuracy attainable for SF2. But it's better than nothing. Big thanks to David Henningsson of the Fluidsynth dev team, who very helpfully answered questions. A great guy.
In addition, there are some fixes to earlier commits here, which I ran into while working on the SF2 timing.
Recent builds of MinGW64 runtime include the winpthreads library which
is not linked statically anymore and thus doesn't required non-portable
(de-)initialization function calls.
- Remove the redundant hasSampleExactData() function. Instead, signal lack of s.ex.data by returning a NULL in valueBuffer()
- Cache s.ex.buffers and only update them once per period
- Make valueBuffer() in AutomatableModel threadsafe so that it can be used for NPH's sharing the same model
- Add sample-exactness to instrumenttrack's vol & pan knobs
Change in handling of frameoffset for multistreamed instruments and sampletracks.
- Instead of holding the offset for the lifetime of the playhandle, negate the offset in the first period
- Multistream-instruments require some small changes: they have to now check for the offset and accordingly leave empty space in the start of the period (already done in this commit)
- There are possibly optimizations that can be done later
- This change is necessary so that we can have sample-exact models, and sample-exact vol/pan knobs for all instruments. Earlier multistream instruments were always rendering some frames ahead-of-time, so applying sample-exact data for them would have been impossible, since we don't have the future-values yet...
Changing the velocity after noteon doesn't really seem to work on fluidsynth (may be dependent on soundfont file) - seems like the panning changes when velocity is changed.
So I disabled that, after which everything works fine.
Changing the velocity after noteon doesn't really seem to work on fluidsynth (may be dependent on soundfont file) - seems like the panning changes when velocity is changed.
So I disabled that, after which everything works fine.
Issue: Currently, we use threads to process all PlayHandles, so there's no guarantee of the order they are processed in. This causes timing inaccuracy and jitter: notes of instruments that use both NPH's and IPH's can get randomly delayed by one entire period.
The issue is solved thusly:
- When processing an IPH, we check if the instrument is midi-based. If yes, we just process it normally (no NPH's to worry about).
- If it's not, then it also uses NPH's, so we'll have the IPH wait until all NPH's belonging to same instrument have been processed. There's some similar code in the new FX mixer, I pretty much just copied how we do it there.
