Real-device WebGPU crash in directional shadow receiver path after Three shadow render completion

Real-device WebGPU crash in directional shadow receiver path after Three shadow render completion

Summary

We reduced a WebGPU crash in our React Native / Expo app to a small repro that fails on both iOS simulator and a real iPhone, but only when a directional-light shadow receiver path is active.

What is already proven:

• raw WebGPU passes on simulator and real device
• raw Three WebGPU passes on simulator and real device
• patched R3F WebGPU passes on simulator and real device
• the reduced shadow receiver repro fails on simulator and real device
• three completes the shadow render path on the JS side
• the only local patch that materially changed the failure boundary was in react-native-wgpu
• no safe local workaround is proven

Current best attribution:

• this does not look like a generic WebGPU, generic Three, or generic R3F failure
• this looks like a native/backend WebGPU failure on the react-native-wgpu / Dawn / Metal path during the shadow-pass submission lifecycle
• our best current stage classification is: during_shadow_pass_submission_before_native_present_completion

Environment

• react-native-wgpu: 0.5.8
• react-native: 0.81.5
• expo: ~54.0.0
• react: 19.1.0
• three: 0.183.2
• @react-three/fiber: 9.5.0
• real device used for confirmation: iPhone 16 Pro Max

Passing controls

These all pass on simulator and real device:

• raw WebGPU
• raw Three WebGPU
• patched R3F WebGPU

Failing reduced repro

The failing repro is a reduced scene with a directional light and a shadow receiver.

Final narrowed failing condition:

• receiveShadow = true
• directionalShadow = true

This still fails across:

• materials:
  • MeshStandardMaterial
  • MeshLambertMaterial
  • MeshBasicMaterial
• geometries:
  • circle
  • plane
  • ring

So this is not just “standard material on a circle”.

Observed behavior

On a real iPhone:

• the reduced repro reaches route-ready
• the process is gone by the end of the probe window
• deeper native trace capture could not be flushed before failure from this environment

On simulator:

• the same reduced repro also fails

Important local findings

1. three is not the first failing layer anymore

Our engine-level source-debugging narrowed the boundary further:

• three reaches/completes the shadow render path on the JS/engine side
• the remaining blocker appears after that, in the native/backend path

2. A react-native-wgpu patch was the only patch that materially changed the failure boundary

We tested several local patch candidates.

Only one materially changed behavior:

• a native present/backend guard in react-native-wgpu

That candidate:

• made the reduced repro pass on simulator
• made broader Doorway validation scenes pass on simulator and real device
• but the reduced shadow repro still failed on real device in the clean confirmation rerun

This is why react-native-wgpu is now our primary issue target.

3. Warm-up / delayed shadow enable did not solve it

We also tested delayed shadow enable / warm-up gating.

Result:

• no usable effect on real device
• the reduced repro still failed before the gate meaningfully activated

Why we think this belongs here

At this point the evidence suggests:

• the reduced repro is valid
• generic WebGPU and generic Three/R3F are fine
• the failure is in the native/backend path specific to the shadow-pass lifecycle on this stack
• the only patch that moved the boundary at all was in react-native-wgpu

That makes react-native-wgpu the primary owner to inspect next.

Primary ask

Can you advise where to inspect next in the native/backend path for this reduced real-device shadow-pass failure?

The most relevant current question is:

• why does the reduced directional-shadow receiver repro still die on real device after three shadow render completion, even though a native/backend guard is enough to make broader validation surfaces pass?

If there is a specific Dawn / Metal / react-native-wgpu texture, render-pass, queue-submit, or present-adjacent path that matches this pattern, that is the highest-value direction for us.

If helpful, I can provide:

• the reduced repro surface
• machine-readable simulator vs real-device comparison outputs
• the exact local patch candidates we tested

Start with the reduced directional-shadow receiver repro and the react-native-wgpu native present/backend guard, comparing simulator and iPhone behavior after Three completes shadow rendering. Trace the Dawn/Metal texture, render-pass, queue-submit, and present-adjacent stages named in the report. Done means identifying the native failure boundary and demonstrating a fix against the reduced repro on both simulator and real device.