Buffered async video lags behind

Operating System Info

Ubuntu 24.04

Other OS

No response

OBS Studio Version

31.1.2

OBS Studio Version (Other)

No response

OBS Studio Log URL

https://obsproject.com/logs/MLjBf83CxDTC3U8y

OBS Studio Crash Log URL

No response

Expected Behavior

The square for the Sync Test (Async Video/Audio Source) should alternative from black to white every second. There should be a tone playing when the square is white.

Current Behavior

In the original version of the (test) source: The tone plays while the audio is black, and is silent while it is white.

In the modified version (using absolute timestamps): there is a 1 second delay before the square starts alternating colour.

Steps to Reproduce

1. Build OBS with the test sources.
2. Add a Sync Test (Async Video/Audio Source) to the scene.
3. Start/End recording then listen to result.

I also modified the test source to use an absolute timestamp (i.e. one that doesn't start at zero) by making the following changes:

--- a/test/test-input/sync-async-source.c
+++ b/test/test-input/sync-async-source.c
@@ -60,7 +60,7 @@ static void *video_thread(void *data)
        uint64_t cur_time = os_gettime_ns();
        bool whitelist = false;
        double cos_val = 0.0;
-       uint64_t start_time = cur_time;
+       // uint64_t start_time = cur_time;
 
        struct obs_source_frame frame = {
                .data = {[0] = (uint8_t *)pixels},
@@ -80,8 +80,8 @@ static void *video_thread(void *data)
        while (os_event_try(ast->stop_signal) == EAGAIN) {
                fill_texture(pixels, whitelist ? 0xFFFFFFFF : 0xFF000000);
 
-               frame.timestamp = cur_time - start_time;
-               audio.timestamp = cur_time - start_time;
+               frame.timestamp = cur_time; // - start_time;
+               audio.timestamp = cur_time; //- start_time;
 
                if (whitelist) {
                        for (size_t i = 0; i < sample_rate; i++) {

I then performed the same test.

Anything else we should know?

There are two relatively obvious causes of these bugs within obs-source.c: Firstly, the following test incorrectly short-circuits the frame-readying procedure if the first timestamp was zero:

https://github.com/obsproject/obs-studio/blob/7778070cbd8e4689d91d90068091ced467c5fdef/libobs/obs-source.c#L4072

The second issue is that changes in the system (video) time are only accounted for when the queue is non-empty. This effectively guarantees that async video is one frame behind at all times; see the code surrounding:

https://github.com/obsproject/obs-studio/blob/7778070cbd8e4689d91d90068091ced467c5fdef/libobs/obs-source.c#L4016).

The potential for incorrect accumulation for last_frame_ts is almost certainly related to other issues with buffered async video observed here: https://obsproject.com/forum/threads/video-gradually-falling-behind-audio-where-can-i-start-investigating.187021/

Start with the referenced sections of libobs/obs-source.c around lines 4016 and 4072, then reproduce the behavior using test/test-input/sync-async-source.c and the linked OBS log. Compare zero-based and absolute timestamps while tracing buffered async video timing. Done means the Sync Test alternates on time with its audio tone and no persistent one-frame or one-second lag remains.