Conversation
|
I'm getting reports that the studio might feel sluggish with this patch. I don't feel it but perhaps this is related to computer speed. One plausible explanation is that now with the audio synthesis done in the callback, maybe SDL only has one thread for doing event and audio handling? And if we are doing lot of stuff in the audio callback (the whole sound synthesis), then there might not be enough time to handle e.g. key events. This is just a theory; please test and comment. |
Previously, the sound registers were used to synthesize fixed amount of sound after every TIC() and this was pushed to SDL. The problem with this approach was crackling if TIC() takes too long to reach 60Hz. In this fix, the sound registers are instead pushed to a ring buffer and the audio is synthesized in SDL callback to render audio. If the ring buffer does not have enough values, the sound synthesis uses the last good values from the ring buffer, so the music slows down yes, but at least it does not crackle. The corollary to this is that calling tic_core_tick_end(tic) is not enough during the Studio sound export; one should also call tic_core_synthesize_sound which does the job of pulling the sound registers from the tail of the ring buffer and doing the sound synthesis. The ring buffer is kept short on purpose, because if there slight desync between the TIC() frequency and sound frequency, the TIC() frequency might run faster and delay might build up. So, the ring buffer length sets a maximum on the delay that this technique can create.
|
I reduced slightly the callback frequency; SDL_AudioSpec.samples is now 1024. I also made the ring buffer have a length of 6 items. Furthermore, I cleaned some unnecessary whitespace. |
|
Hi, thank you for the PR, but unfortunately I'm in a hospital now and can't review/merge it. Sorry. |
|
No problem! Get well soon! |
nesbox
left a comment
nesbox
left a comment
There was a problem hiding this comment.
Just tested your version and it works perfectly, thank you very much for fixing this annoying issue.
One thing, we also have to update the player.c and other implementations, like libretro/n3ds/sokol/...
2 participants
Previously, the sound registers were used to synthesize fixed amount of sound after every TIC() and this was pushed to SDL. The problem with this approach was crackling if TIC() takes too long to reach 60Hz. In this fix, the sound registers are instead pushed to a ring buffer and the audio is synthesized in SDL callback, to render audio. If the ring buffer does not have enough values, the sound synthesis uses the last good values from the ring buffer, so the music slows down yes, but at least it does not crackle.
The corollary to this is that calling tic_core_tick_end(tic) is not enough during the Studio sound export; one should also call tic_core_synthesize_sound which does the job of pulling the sound registers from the tail of the ring buffer and doing the sound synthesis.
The ring buffer is kept short on purpose, because if there slight desync between the TIC() frequency and sound frequency, the TIC() frequency might run faster and delay might build up. So, the ring buffer length sets a maximum on the delay that this technique can create.
Just to note that you might want to experiment with TIC_SOUND_RINGBUF_LEN and SDL_AudioSpec.samples to get good amount of buffering, without significant delays, and not too much overhead. SDL_AudioSpec.samples controls how many times a second the callback is called and presumably there's a small overhead with each call.
This fix also btw probably fixes the complaints about the sound getting desynchronized with the video; the short ring buffer ensures that the audio can never be too much out of sync with the video.