Keep the client schedule floor at 2 ms while accepting shorter guest
frame intervals for cadence tracking. This lets sources above 500 Hz
trigger immediate final-frame delivery when their cadence breaks.
Signal the shared D3D11 source only after a retained candidate and its
command slots have been acquired. Superseded source frames no longer queue
a fence that no Direct3D 12 work will consume.
Signal a per-slot event after callbacks finish and the command slot has
returned to the free state. This removes the one millisecond polling delay
when candidate state becomes visible just before its slot is reusable.
Drop an acquired frame immediately when both retained candidate slots are
unavailable. Waiting for up to 2 ms held the IddCx surface and
backpressured sources whose frame period was shorter than the wait.
Derive the source rate from unique frames observed over the actual log
interval. The old EWMA discarded intervals below 2 ms, capping the
reported rate at 500 Hz and understating high-rate sources.
Allow the timing owner to keep one delivery pending on each independent
owner queue instead of waiting for the client to release every frame.
Count shared fallback deliveries against the same two-frame limit and
match republished frames across all active lanes. This hides the client
release round trip without allowing a third delivery for the owner.
Store retained frame candidates in the final linear transport layout.
Prepare texture output in local DEFAULT buffers before cadence holding.
Then publish full frames with a single buffer copy.
Keep partial damage on matching pitched layouts and leave effect-owned
buffer copies unchanged. This moves detiling out of the
deadline-critical IVSHMEM or readback transfer.
Ignore allocation alignment when matching retained candidate resources.
D3D12 may resolve automatic alignment requests in GetDesc.
Comparing that value recreated the candidate for every benchmark sample.
Measure native and packed over matching wall-clock intervals.
Use preparation and publication intervals, excluding the cadence hold.
This avoids mixing GPU timestamps with fallback timing. The CPU IVSHMEM
copy is included only on indirect systems.
Track damage received after each retained candidate snapshot. When a
candidate is submitted successfully, replace the global pending set
with its trailing damage instead of requiring the source generation
to remain unchanged.
This lets published full-frame damage retire while preserving newer
partial or full updates.
Run effect timing samples only when the retained candidate pipeline is idle
and force each completed sample to publish immediately. This prevents work
from superseded candidates from being charged to a surviving sample.
Include the indirect CPU copy to IVSHMEM in the measured copy stage while
preserving the wall-clock fallback when GPU timestamps are unavailable.
Treat a failed desktop texture bind as an incomplete render so full target
restoration remains latched until a frame can actually be drawn.
Carry retired damage into the pending record when the drawable texture does
not contain it, closing the publication gap between texture slots and their
matching damage metadata.
Prepare each acquired content frame into one of two private GPU
candidates so the IddCx frame can be released without copying it to
IVSHMEM immediately.
Wake a high-resolution publisher at the client deadline and send the
newest completed candidate. Preserve accumulated damage across
superseded candidates, include the scheduling hold in end-to-end timing,
and keep transport work on the existing physical copy queue.
Move splash transitions onto the render thread and derive opacity from
elapsed monotonic time. Pace active animation at a display-compatible
minimum of 60 Hz while allowing incoming frames to satisfy that demand.
Reset the splash cleanly across reconnects and latch state changes with
the full-frame restoration decision.
Only report the client queue lead when JIT rendering provides a valid
measurement. Clamp linear phase errors instead of wrapping them into the
opposite direction at half of a refresh period.
Timestamp OBS frames after the shared framebuffer becomes ready.
Always track Wayland surface outputs alongside fractional scaling and use
RandR modes for X11 cadence. Remove render-derived feedback so a
throttled client cannot teach the producer to remain throttled.
Release scheduling after nominal output timing remains unavailable.
Use the current wl_output mode for the requested producer cadence.
Presentation feedback can follow a slower VRR cadence, feeding that low
update rate back into the IDD. Select the fastest nominal mode when a
window spans multiple outputs.
Mark the active fade as requiring a fully restored render target.
This also covers guest-driven renders between animation ticks so they
cannot blend over stale splash content.
Fully invalidate the render target when the full-screen splash changes.
This restores the desktop before alpha blending and prevents stale fade
states from cycling between swapchain buffers when guest updates are
sparse.
Discard the learned guest arrival period when a frame follows a gap
much larger than expected. Force that frame through and relearn the
arrival rate so sparse updates are not held for a frame that never
arrives.
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.
