Omniverse Realtime Viewer

<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -->

<!-- SPDX-License-Identifier: Apache-2.0 -->

Spatial UI Performance And Telemetry

Render FPS is the primary budget. Separate state-changing work from pose-only work and keep both away from blocking XR-frame operations.

Retained And Asynchronous Work

• Build hierarchy/property models only when stage generation or requested data changes.

• Run Yoga and Skia off the XR thread and publish the newest complete immutable revision through a bounded latest-wins handoff.

• Upload dense UI pixels only for a new document revision.

• Pre-author analytic controller, ray, tooltip, gizmo, and chrome geometry; update bounded pose/vertex data instead of creating prims.

• Extract stage meshes, bounds, and selectable mappings at generation boundaries. Keep USD traversal and pxr out of the frame loop.

• Use one immutable input snapshot and one bounded transform publication batch.

• Cache stable tracking-to-world and model metadata for the input lifetime; invalidate them on their explicit lifecycle boundaries.

Do not deep-copy status trees, serialize proof artifacts, poll Fabric, launch an idle ray query, or log unchanged pose dictionaries every frame. Explicit audits are finite and armed; they stop automatically after consuming their bounded evidence.

A persistent raster subprocess is still synchronous if the XR owner writes to its pipe and waits for the reply. Hover, capture and selection changes must enqueue bounded latest-wins revisions; keep showing the previous complete raster and its matching hit map until the new pair is ready. A response timeout does not bound a blocking request-pipe write. Test interaction bursts and a stalled worker, not only unchanged-panel steady state. Keep at most one render in flight, one replaceable pending request, and one complete result. Bind subprocess readers to their own process generation; discard late output after reset. Give both request writes and responses a worker-side deadline and interrupt a stuck child on shutdown. Retry only a bounded number of times while retaining the last complete image/hit-map pair.

Account for bytes per revision: 2048x1024 RGBA8 is 8 MiB, four times 1024x512. Raising raster resolution does not justify uploading it for pose changes. Coalesce obsolete revisions and measure raster/IPC/upload latency separately. Do not scan a full RGBA canvas for routine diagnostic coverage on the XR owner. Compute bounded diagnostics in the raster worker or arm a finite audit. Track geometry upload bytes separately from texture bytes; a narrowly defined transfer counter is not necessarily total host-to-GPU traffic.

Required Counters

• Raster: model generations, render revisions, texture uploads, unchanged revision upload attempts, renderer queue drops and latency.

• Primary path: injected/omitted frames, per-eye draws and scissors, full-eye copies/bytes, transfers/bytes, same-frame waits, slot exhaustion.

• Timing: bounded CPU compose and queue-submit windows; GPU timestamps read only after an existing frame-slot fence has signaled.

• Whole owner loop: include input, post-submit UI work, diagnostics and capture contention; a timer stopped at scene submission hides later work.

• Geometry: per-eye accepted, clipped, composed, and draw-included surfaces, bounds, index spans, and rejection reasons.

• Input: epoch, sample/display time, owner/focus/activity/validity, explicit reference space, ray hit/UV/distance, capture phase, and cancellation reason.

• Identity: native integration build/hash, manifest or hook selection, runtime versions, and process start time.

Unavailable measurements remain unknown, not passed.

Connected-session measurement and targeted optimization

Bound connection waiting separately from the capture window. Exclude disconnected pacing and shader warmup from steady headset FPS. Report frame interval median/tails plus gesture-active versus idle windows; correlation is not causation. Preserve scene/view, resolution, material policy, foveation and mirror demand for A/B runs. Stop a one-shot collector after its requested session, disconnect, deadline or error; do not silently rearm it.

Timing lanes can nest: SDK wall time includes runtime waits and child begin/bind/submit lanes; controller time can contain gesture and tool work. Never sum overlapping lanes. Native compositor CPU timing can exclude earlier mesh construction. Report coordinate conversion, mesh packing/construction, mirror capture/copy/publication and post-submit owner work explicitly; unknown GPU or transport latency stays unknown.

Before rewriting a renderer, measure bounded CPU work: batch 26/52-joint conversion against one active anchor snapshot, pack analytic strokes into reusable contiguous buffers, preserve native mesh capacity, and avoid creating unused frame-result wrappers. Keep a scalar reference for equivalence checks, one immutable input snapshot, and a documented fallback for unavailable batch capabilities. Do not cache an anchor across a real change or drop validity flags. Disable unused fixture/physics work rather than merely hiding its visuals. One-time hide/reset publication is acknowledged only after successful owner publication; include activation generation so stale async results cannot revive an earlier fixture. CPU microbench improvements are not headset FPS evidence.

Proof Boundaries

Track these separately:

1. Application state was accepted.

2. A finite per-eye mesh was composed.

3. The submitted draw and scissor included its indices and bounds.

4. The depth/alpha contract was valid.

5. Measured fragments survived scissor/depth and changed the primary output.

6. Correlated decoded or headset pixels showed the surface.

A later boundary cannot be inferred from an earlier one. If bounded fragment instrumentation is needed, consume results only after an already-signaled production fence; do not add a second draw, full-eye readback, or same-frame wait.

Steady-State Gate

Across a settled interval with an unchanged panel raster, require zero unchanged-revision texture uploads, primary full-eye copies/transfers, same-frame fence waits, steady Fabric reads, and unbounded diagnostics. Require an explicit draw/scissor budget per eye and bounded CPU/GPU timing samples. Batch compatible surfaces, allowing measured foreground/depth-policy splits; count actual draw commands separately from render passes and scissor changes. Pose updates may upload bounded geometry even when texture uploads are zero. Report texture, geometry and total host-upload bytes separately; do not redefine zero-copy to mean no full-eye copy while hiding these smaller transfers.

Compare matched scene/view, resolution, foveation, mirror demand, material policy, build and warmed shader state. Report steady and interaction-time frame intervals separately. A short startup GPU timing window cannot explain an unrelated end-to-end FPS difference. Isolate AR material/refraction/cutout options individually while rechecking correct transparency and edge quality.

Count native eye capture independently from mirror publication. Disarming a one-shot reader or hiding a preview does not disable create-time native capture. End diagnostic-only capture through the documented context lifecycle; do not rebuild a context merely because a shared mirror still needs capture. For create-time-only capture, explicitly opt the diagnostic into a context refresh afterward (which may require client reconnect), or require an existing capture consumer. Coalesce cleanup with a pending lifecycle change. Cancel old workers before detach and reject their completion against replacement contexts. Attempt/time budgets bound application retries, not an unresponsive GPU driver.

Exercise retirement/disconnect/context refresh with GPU work in flight. Fence polling in the steady loop does not prove bounded teardown. Define timeout and device-loss handling without freeing resources the GPU still owns.