Disable fast-math for the timing graph translation unit because its
bucket arrays use NaN values to break plot lines across empty intervals.
Keep fast-math enabled for remaining client sources while allowing
Clang's strict floating-point diagnostics to remain fatal.
Treat every completed swap as satisfying the interactive overlay
refresh requirement. Only force a JIT overlay render after the 60 Hz
deadline, and rearm the timer-based deadline after frame-driven swaps.
This prevents interactive overlays from adding redundant swaps when
incoming frames already keep the client at or above the required rate.
Unbind the streamed desktop pixel upload buffer before ImGui processes
CPU-backed font atlas updates.
This prevents atlas pointers from being interpreted as PBO offsets when
DMA is disabled, which caused newly requested glyphs to be missing.
Expand ImGui window damage to include its border and round both logical
window bounds and framebuffer-scaled bounds outward.
This prevents a one-pixel strip of stale desktop content when an
overlay window is resized at fractional positions or display scales.
Correlate received frames with the texture update actually consumed by
the renderer and retain damage until its token becomes renderable.
Split client latency into dispatch, import, queue, preparation, setup,
effects, desktop, composition, and swap stages in FRAME LATENCY.
Exclude diagnostic overlay work from composition and preserve client
wait time when producer timing is unavailable.
Publish joined samples through a bounded token queue, avoid sleeping
while producer timing is finalized, and correct Wayland photon units.
Use calibrated copy-queue timestamps to separate source and effect
waits from the actual framebuffer copy.
Exclude producer readiness waits from client import timing so the
same interval is not counted in both Copy and Import.
libdecor_decorate() already creates an xdg_surface and xdg_toplevel
internally for the frame. Calling xdg_surface_get_toplevel() again on
that same xdg_surface just to fetch the toplevel for icon handling
assigns the xdg_toplevel role to the underlying wl_surface a second
time, which is a protocol violation.
On GNOME Shell 50 (mutter) this is no longer tolerated: mutter
terminates the Wayland connection ("WL: error in client
communication"), and since libdecor_shellInit() waits for the initial
configure with a blocking, error-less libdecor_dispatch(-1) loop, the
client hangs forever with no window and no diagnostic output. This
made -DENABLE_LIBDECOR=ON (the default when libdecor-0 is present)
appear completely broken on current GNOME/Wayland.
Use libdecor_frame_get_xdg_toplevel() instead, which returns the
xdg_toplevel libdecor already owns, without requesting a second role
assignment.
Require libdecor-0 >= 0.1.1, which is available in Debian oldstable,
so the new accessor is guaranteed by the build dependency.
Add an optional EGL framebuffer capture path for the test transport.
Capture the fully composed frame before presentation and retain its
surface format and HDR state for validation.
Add GoogleTest integration tests which run the production client
under a headless Weston compositor and compare every output pixel
against an independent reference implementation.
The matrix covers BGRA, RGBA, BGR32, RGB24, PQ RGBA10 and scRGB
RGBA16F with full, moving, overlapping, maximum-count, invalid, null
and zero-area damage rectangles. Validate HDR-to-SDR conversion and
native HDR output metadata when supported by the compositor.
Keep tests opt-in through ENABLE_RENDER_TESTS and document how to run
the software and native HDR tiers.
NVIDIA's Wayland EGL implementation sets explicit-sync acquire and
release points while presenting a frame. The requests emitted by
eglSwapBuffers and its final surface commit must be treated as one
logical transaction.
The pointer-warp path can commit the main surface from the input
thread while the render thread is inside eglSwapBuffers. If this
occurs between the explicit-sync requests, the cursor commit applies
incomplete pending state and the compositor disconnects the client
with:
explicit sync is used, but no release point is set
Add a surface lock and hold it across EGL presentation. Use the same
lock for direct commits, frame callback registration, and
pointer-constraint updates, replacing the narrower confinement lock.
This prevents another thread from splitting the EGL transaction and
removes the need for __NV_DISABLE_EXPLICIT_SYNC=1.
Move shared-memory ownership, LGMP session handling, queue access, and
DMA setup behind a transport interface. The LGMP backend preserves the
existing zero-copy frame and DMA paths while owning its lgmp:* options.
Expose the initialized EGL context through a versioned renderer interop
record for future accelerated decode backends. Add an LGMP-independent,
deterministic test transport for graphics-pipeline validation.
Model playback latency from the device period, arrival jitter, source
packet phase, and resampler delay. Treat audio:latencyOffset as an
explicit addition to this minimum and align the first device pull to
the next packet deadline. This starts playback near its steady-state
target without unnecessary prefill or startup underruns. Use a
512-frame default period as a practical low-latency baseline.
Replace the startup clock hold with a one-sided proportional
acquisition controller, then hand off to source/device rate
feed-forward and a slow phase loop. Calibrate the logical device
timeline at handoff, discard correction that opposes the current
error, and slew-limit rate changes. This prevents startup drain,
integral wind-up, overshoot, and long convergence while preserving
clock-drift compensation.
Allow audio backends to expose a real-time resampler and use
PipeWire's adaptive resampler when version 1.4 or newer supports it.
Retain libsamplerate as the fallback and add audio:resampler to select
the implementation. Wait for PipeWire stream setup to complete and
propagate rate-control failures cleanly.
Track PipeWire input-consumption and output-equivalent clocks
separately. The input clock measures ring latency while the output
clock drives feed-forward using the ratio that governed each request.
This removes delayed self-feedback that made adaptive resampling
oscillate between the correction limits.
Reduce audio:debug output to useful latency, clock, jitter, and xrun
values, and scale the playback graph from the startup estimate. Update
the option names and documentation for the new latency model.
Avoid querying playback timing on every graph cycle. Refresh backend
latency only when audio diagnostics request it, and defer callback
errors for reporting from non-realtime contexts.
Move microphone delivery to a bounded SPSC queue drained by a sender
thread. This keeps PureSpice locking and socket writes out of the
PipeWire capture callback while bounding overload behavior.
Validate and recycle capture buffers before enqueueing their contents.
These changes improve deadline margin when using small period sizes.
Bound playback writes and synchronization slews to the physical ring
storage, preventing overwritten samples from being treated as valid PCM.
Trigger clock resynchronization when output must be dropped.
Wait for enough startup audio to cover backend demand and a complete
source packet. Generate silence if playback begins early instead of
rewinding the reader into stale ring storage.
Honor PipeWire playback frame requests and fully initialize empty chunks.
This keeps playback reliable when using small device period sizes.
