lesavka/client/src/sync_probe/capture/video_packets.rs

464 lines
18 KiB
Rust

use super::*;
pub(super) struct MjpegProbeFrames {
dark: Vec<u8>,
regular_pulse: Vec<u8>,
marker_pulse: Vec<u8>,
coded_pulses: BTreeMap<u32, Vec<u8>>,
}
pub(super) enum VideoPacketSource {
Mjpeg(MjpegProbeFrames),
Pipeline {
pipeline: gst::Pipeline,
src: gst_app::AppSrc,
sink: gst_app::AppSink,
first_sample_pts: Option<Duration>,
},
}
#[derive(Clone, Copy)]
pub(super) enum ProbeFrameKind {
Dark,
RegularPulse,
MarkerPulse,
Coded(u32),
}
pub(super) struct EncodedVideoData {
pub data: Vec<u8>,
pub pts: Duration,
}
/// Build the encoded video source used by the synthetic client transport probe.
///
/// Inputs: negotiated camera profile and pulse schedule.
/// Outputs: either pre-encoded MJPEG signature frames or a live H.264 encoder.
/// Why: MJPEG encoding 1080p frames in the hot loop caused transport jitter, so
/// deterministic signature frames are encoded once before capture starts.
pub(super) fn build_video_packet_source(
camera: CameraConfig,
schedule: &PulseSchedule,
) -> Result<VideoPacketSource> {
match camera.codec {
CameraCodec::Mjpeg => {
let coded_pulses = schedule
.event_width_codes()
.iter()
.copied()
.map(|code| {
encode_mjpeg_probe_frame(
camera,
&build_coded_probe_frame(
camera.width as usize,
camera.height as usize,
code,
),
)
.map(|frame| (code, frame))
})
.collect::<Result<BTreeMap<_, _>>>()?;
Ok(VideoPacketSource::Mjpeg(MjpegProbeFrames {
dark: encode_mjpeg_probe_frame(
camera,
&build_dark_probe_frame(camera.width as usize, camera.height as usize),
)?,
regular_pulse: encode_mjpeg_probe_frame(
camera,
&build_regular_probe_frame(camera.width as usize, camera.height as usize),
)?,
marker_pulse: encode_mjpeg_probe_frame(
camera,
&build_marker_probe_frame(camera.width as usize, camera.height as usize),
)?,
coded_pulses,
}))
}
CameraCodec::H264 | CameraCodec::Hevc => {
let pipeline = build_encoded_pipeline(camera)?;
let src = pipeline
.by_name("sync_probe_video_src")
.context("missing sync probe video appsrc")?
.downcast::<gst_app::AppSrc>()
.expect("video appsrc");
let sink = pipeline
.by_name("sync_probe_video_sink")
.context("missing sync probe video appsink")?
.downcast::<gst_app::AppSink>()
.expect("video appsink");
pipeline
.set_state(gst::State::Playing)
.context("starting sync probe pipeline")?;
Ok(VideoPacketSource::Pipeline {
pipeline,
src,
sink,
first_sample_pts: None,
})
}
}
}
/// Stop any live encoder held by a video packet source.
///
/// Inputs: packet source being consumed by the video thread.
/// Outputs: GStreamer shutdown side effects only.
/// Why: the probe should release local encoder resources even when the RCT
/// capture or upstream transport test exits early.
pub(super) fn stop_video_packet_source(packet_source: VideoPacketSource) {
if let VideoPacketSource::Pipeline { pipeline, src, .. } = packet_source {
let _ = src.end_of_stream();
let _ = pipeline.set_state(gst::State::Null);
}
}
/// Encode one RGB probe frame as MJPEG.
///
/// Inputs: camera profile and raw RGB frame bytes.
/// Outputs: JPEG packet payload bytes.
/// Why: pre-encoding still frames keeps the client transport test focused on
/// bundled network timing rather than local software JPEG throughput.
fn encode_mjpeg_probe_frame(camera: CameraConfig, frame: &[u8]) -> Result<Vec<u8>> {
let video_caps = format!(
"video/x-raw,format=RGB,width={},height={},framerate={}/1",
camera.width,
camera.height,
camera.fps.max(1)
);
let desc = format!(
"appsrc name=sync_probe_still_src is-live=false format=time do-timestamp=false caps={video_caps} ! \
videoconvert ! jpegenc quality=90 ! image/jpeg,parsed=true,width={},height={},framerate={}/1 ! \
appsink name=sync_probe_still_sink emit-signals=false sync=false max-buffers=1 drop=false",
camera.width,
camera.height,
camera.fps.max(1),
);
let pipeline = gst::parse::launch(&desc)
.with_context(|| format!("building still MJPEG encoder: {desc}"))?
.downcast::<gst::Pipeline>()
.map_err(|_| anyhow::anyhow!("still MJPEG encoder did not build a pipeline"))?;
let src = pipeline
.by_name("sync_probe_still_src")
.context("missing still MJPEG appsrc")?
.downcast::<gst_app::AppSrc>()
.expect("still appsrc");
let sink = pipeline
.by_name("sync_probe_still_sink")
.context("missing still MJPEG appsink")?
.downcast::<gst_app::AppSink>()
.expect("still appsink");
pipeline
.set_state(gst::State::Playing)
.context("starting still MJPEG encoder")?;
let frame_step = Duration::from_nanos(1_000_000_000u64 / u64::from(camera.fps.max(1)));
let mut buffer = gst::Buffer::from_slice(frame.to_vec());
if let Some(meta) = buffer.get_mut() {
meta.set_pts(Some(gst::ClockTime::ZERO));
meta.set_dts(Some(gst::ClockTime::ZERO));
meta.set_duration(Some(gst::ClockTime::from_nseconds(
frame_step.as_nanos() as u64
)));
}
src.push_buffer(buffer)
.map_err(|err| anyhow::anyhow!("pushing still MJPEG frame failed: {err:?}"))?;
let _ = src.end_of_stream();
let sample = sink
.try_pull_sample(gst::ClockTime::from_seconds(2))
.context("still MJPEG encoder produced no sample")?;
let data = sample
.buffer()
.context("still MJPEG sample had no buffer")?
.map_readable()
.context("mapping still MJPEG sample")?
.as_slice()
.to_vec();
let _ = pipeline.set_state(gst::State::Null);
Ok(data)
}
/// Build the live encoder pipeline for non-MJPEG negotiated profiles.
///
/// Inputs: camera profile.
/// Outputs: an appsrc-to-appsink GStreamer pipeline.
/// Why: inter-frame codecs cannot reuse still JPEG packets, but they still need
/// the same RGB signature frames so analyzer identity remains comparable.
fn build_encoded_pipeline(camera: CameraConfig) -> Result<gst::Pipeline> {
let video_caps = format!(
"video/x-raw,format=RGB,width={},height={},framerate={}/1",
camera.width,
camera.height,
camera.fps.max(1)
);
let (encoder, parse_chain) = match camera.codec {
CameraCodec::H264 => (
pick_h264_encoder(camera.fps.max(1))?,
"h264parse config-interval=-1 ! video/x-h264,stream-format=byte-stream,alignment=au",
),
CameraCodec::Hevc => (
pick_hevc_encoder(camera.fps.max(1))?,
"h265parse config-interval=-1 ! video/x-h265,stream-format=byte-stream,alignment=au",
),
CameraCodec::Mjpeg => unreachable!("MJPEG uses pre-encoded still frames"),
};
let video_branch = format!(
"appsrc name=sync_probe_video_src is-live=true format=time do-timestamp=false caps={video_caps} ! \
queue max-size-buffers=4 leaky=downstream ! videoconvert ! \
{encoder} ! {parse_chain} ! \
appsink name=sync_probe_video_sink emit-signals=false sync=false max-buffers=4 drop=true",
);
gst::parse::launch(&video_branch)
.with_context(|| format!("building sync probe pipeline: {video_branch}"))?
.downcast::<gst::Pipeline>()
.map_err(|_| anyhow::anyhow!("sync probe description did not build a pipeline"))
}
/// Choose an available low-latency H.264 encoder.
///
/// Inputs: target frame rate, used for GOP sizing where the encoder supports it.
/// Outputs: a GStreamer encoder element description.
/// Why: this probe should run on different developer hosts without hardcoding a
/// single hardware encoder, while still preferring low-latency behavior.
fn pick_h264_encoder(fps: u32) -> Result<String> {
if gst::ElementFactory::find("x264enc").is_some() {
return Ok(format!(
"x264enc tune=zerolatency speed-preset=ultrafast bitrate=2500 key-int-max={}",
fps.max(1)
));
}
if gst::ElementFactory::find("openh264enc").is_some() {
return Ok("openh264enc bitrate=2500000".to_string());
}
if gst::ElementFactory::find("v4l2h264enc").is_some() {
return Ok("v4l2h264enc".to_string());
}
bail!("no usable H.264 encoder found for sync probe")
}
/// Choose an available low-latency HEVC encoder.
///
/// Inputs: target frame rate, used for GOP sizing where the encoder supports it.
/// Outputs: a GStreamer encoder element description.
/// Why: the client-to-server probe should exercise the same HEVC transport
/// shape as real webcam uplink without requiring a specific GPU encoder.
fn pick_hevc_encoder(fps: u32) -> Result<String> {
if gst::ElementFactory::find("x265enc").is_some() {
let keyframe_interval = low_latency_hevc_keyframe_interval(fps);
return Ok(format!(
"x265enc tune=zerolatency speed-preset=ultrafast bitrate=2500 key-int-max={}",
keyframe_interval
));
}
for encoder in ["nvh265enc", "vah265enc", "vaapih265enc", "v4l2h265enc"] {
if gst::ElementFactory::find(encoder).is_some() {
return Ok(encoder.to_string());
}
}
bail!("no usable HEVC encoder found for sync probe")
}
/// Match the real webcam HEVC keyframe cadence in synthetic transport probes.
///
/// Inputs: target frame rate. Output: low-latency keyframe interval in frames.
/// Why: the client-to-RCT probe should stress the same inter-frame shape as
/// real webcam uplink; a one-second GOP made coded flashes less representative
/// than Lesavka's default live-call HEVC pipeline.
fn low_latency_hevc_keyframe_interval(fps: u32) -> u32 {
fps.clamp(1, 5)
}
/// Select the visual signature for a video timestamp.
///
/// Inputs: deterministic pulse schedule and current video PTS.
/// Outputs: frame kind used by packet encoding.
/// Why: the frame decision must be shared by MJPEG and H.264 so both codecs
/// carry the same event identity to the RCT analyzer.
pub(super) fn probe_frame_kind(schedule: &PulseSchedule, pts: Duration) -> ProbeFrameKind {
if !schedule.flash_active(pts) {
return ProbeFrameKind::Dark;
}
if let Some(code) = schedule.event_code(pts) {
return ProbeFrameKind::Coded(code);
}
if schedule.pulse_is_marker(pts) {
ProbeFrameKind::MarkerPulse
} else {
ProbeFrameKind::RegularPulse
}
}
/// Produce an encoded video payload for a probe frame.
///
/// Inputs: packet source, frame kind, raw RGB frame, and timing metadata.
/// Outputs: encoded video bytes, or `None` when the live encoder is drained.
/// Why: bundled transport tests need fresh video packets paced from local PTS,
/// but MJPEG and H.264 require different packet-production paths.
pub(super) fn video_packet_data(
packet_source: &mut VideoPacketSource,
frame_kind: ProbeFrameKind,
raw_frame: &[u8],
pts: Duration,
frame_step: Duration,
) -> Option<EncodedVideoData> {
match packet_source {
VideoPacketSource::Mjpeg(frames) => match frame_kind {
ProbeFrameKind::Dark => Some(EncodedVideoData {
data: frames.dark.clone(),
pts,
}),
ProbeFrameKind::RegularPulse => Some(EncodedVideoData {
data: frames.regular_pulse.clone(),
pts,
}),
ProbeFrameKind::MarkerPulse => Some(EncodedVideoData {
data: frames.marker_pulse.clone(),
pts,
}),
ProbeFrameKind::Coded(code) => frames
.coded_pulses
.get(&code)
.cloned()
.map(|data| EncodedVideoData { data, pts }),
},
VideoPacketSource::Pipeline {
src,
sink,
first_sample_pts,
..
} => {
let mut buffer = gst::Buffer::from_slice(raw_frame.to_vec());
if let Some(meta) = buffer.get_mut() {
let pts_time = gst::ClockTime::from_nseconds(pts.as_nanos() as u64);
meta.set_pts(Some(pts_time));
meta.set_dts(Some(pts_time));
meta.set_duration(Some(gst::ClockTime::from_nseconds(
frame_step.as_nanos() as u64
)));
}
if src.push_buffer(buffer).is_err() {
return None;
}
freshest_probe_video_sample(sink).and_then(|sample| {
let buffer = sample.buffer()?;
let sample_pts =
normalized_sample_pts_duration(buffer, first_sample_pts).unwrap_or(pts);
let map = buffer.map_readable().ok()?;
Some(EncodedVideoData {
data: map.as_slice().to_vec(),
pts: sample_pts,
})
})
}
}
}
/// Read the encoder output timestamp for one sample buffer.
///
/// Inputs: encoded sample buffer. Output: packet PTS as a `Duration` when the
/// encoder preserved it. Why: inter-frame encoders may return an older access
/// unit than the frame just pushed, so transport packets must use the actual
/// output PTS instead of the current input-loop PTS.
fn sample_pts_duration(buffer: &gst::BufferRef) -> Option<Duration> {
buffer.pts().map(|pts| Duration::from_nanos(pts.nseconds()))
}
/// Rebase encoder output timestamps onto the probe's zero-based timeline.
///
/// Inputs: encoded sample buffer and mutable first-sample timestamp.
/// Output: normalized sample PTS.
/// Why: some GStreamer encoders emit a segment-offset PTS while still preserving
/// correct sample-to-sample cadence, so the probe keeps the cadence and drops
/// the absolute segment origin before bundling media for transport.
fn normalized_sample_pts_duration(
buffer: &gst::BufferRef,
first_sample_pts: &mut Option<Duration>,
) -> Option<Duration> {
let sample_pts = sample_pts_duration(buffer)?;
let first = first_sample_pts.get_or_insert(sample_pts);
Some(sample_pts.saturating_sub(*first))
}
/// Drain a live appsink and return the newest encoded sample.
///
/// Inputs: GStreamer appsink for the H.264 probe pipeline.
/// Outputs: most recent sample if one was produced.
/// Why: the transport probe should prefer freshness over preserving an encoder
/// backlog that would make client-origin media look older than it really is.
fn freshest_probe_video_sample(sink: &gst_app::AppSink) -> Option<gst::Sample> {
let mut newest = sink.try_pull_sample(gst::ClockTime::from_mseconds(250));
while let Some(sample) = sink.try_pull_sample(gst::ClockTime::ZERO) {
newest = Some(sample);
}
newest
}
#[cfg(test)]
mod tests {
use gstreamer as gst;
/// Verifies synthetic HEVC probes use the same short GOP shape as live
/// camera transport.
///
/// Input: representative target frame rates. Output: bounded keyframe
/// interval. Why: coded flash recovery should fail for real transport
/// reasons, not because the probe used an easier one-second GOP.
#[test]
fn low_latency_hevc_keyframe_interval_matches_live_camera_default() {
assert_eq!(super::low_latency_hevc_keyframe_interval(0), 1);
assert_eq!(super::low_latency_hevc_keyframe_interval(1), 1);
assert_eq!(super::low_latency_hevc_keyframe_interval(5), 5);
assert_eq!(super::low_latency_hevc_keyframe_interval(20), 5);
assert_eq!(super::low_latency_hevc_keyframe_interval(30), 5);
}
/// Verifies encoded packet timestamps come from the encoder output sample.
///
/// Input: one encoded GStreamer buffer with explicit PTS. Output: matching
/// `Duration`. Why: HEVC encoders may return a delayed access unit, so the
/// bundle must carry the timestamp of what actually left the encoder.
#[test]
fn sample_pts_duration_uses_encoder_output_pts() {
gst::init().expect("gst init");
let mut buffer = gst::Buffer::with_size(4).expect("buffer");
{
let meta = buffer.get_mut().expect("mutable buffer");
meta.set_pts(Some(gst::ClockTime::from_mseconds(123)));
}
assert_eq!(
super::sample_pts_duration(buffer.as_ref()),
Some(std::time::Duration::from_millis(123))
);
}
/// Verifies encoder segment origins are removed while cadence is retained.
///
/// Input: two encoded buffers whose PTS starts far from zero. Output:
/// zero-based probe timestamps. Why: the server analyzer compares client
/// media against the synthetic probe timeline, not GStreamer segment
/// wall-clock origins.
#[test]
fn normalized_sample_pts_duration_preserves_cadence_without_segment_origin() {
gst::init().expect("gst init");
let mut first_sample_pts = None;
let mut first = gst::Buffer::with_size(4).expect("first");
first
.get_mut()
.expect("first mutable")
.set_pts(Some(gst::ClockTime::from_seconds(3_600)));
assert_eq!(
super::normalized_sample_pts_duration(first.as_ref(), &mut first_sample_pts),
Some(std::time::Duration::ZERO)
);
let mut second = gst::Buffer::with_size(4).expect("second");
second.get_mut().expect("second mutable").set_pts(Some(
gst::ClockTime::from_seconds(3_600) + gst::ClockTime::from_mseconds(33),
));
assert_eq!(
super::normalized_sample_pts_duration(second.as_ref(), &mut first_sample_pts),
Some(std::time::Duration::from_millis(33))
);
}
}