Core debugging skill. Runs the same (ref, dist) through every enabled backend (cpu, cuda, sycl, vulkan), reports per-frame per-feature ULP divergence, flags anything beyond configurable tolerance. Use after any SIMD / GPU change. Mirrors the T6-8 GPU-parity CI gate (ADR-0214) locally.
Installs into .claude/skills of the current project.
Are you the author of Cross Backend Diff?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/vmafx-cross-backend-diff)
---
name: cross-backend-diff
description: Core debugging skill. Runs the same (ref, dist) through every enabled backend (cpu, cuda, sycl, vulkan), reports per-frame per-feature ULP divergence, flags anything beyond configurable tolerance. Use after any SIMD / GPU change. Mirrors the T6-8 GPU-parity CI gate (ADR-0214) locally.
---
# /cross-backend-diff
## Invocation
```text
/cross-backend-diff --ref=PATH --dist=PATH --width=W --height=H --pixfmt=420p
[--bitdepth=8] [--frames=48] [--tolerance-ulp=2]
[--backends=cpu,cuda,sycl,vulkan]
```
## Steps
1. Build every backend in `--backends` (reuse cached builds where possible).
2. Run `build/tools/vmaf` per backend with `--json --precision=17` for every
feature (default: adm, vif, motion, psnr, ssim).
3. Parse JSON into `(backend, frame, feature) -> score` tables.
4. For every pair of backends, compute:
- Absolute max diff
- ULP distance (double bits XOR)
5. Report table:
```text
Feature Pair MaxAbsDiff MaxULP WorstFrame Verdict
adm cpu vs cuda 3.2e-16 1 17 OK
vif cpu vs sycl 4.1e-15 18 23 FAIL(>2)
...
```
6. Exit 0 if all ULPs ≤ `--tolerance-ulp`, else 1.
## Notes
- ULP > tolerance = blocker. Fix reduction to use double accumulation
(see `simd-reviewer`) or open CODEOWNERS-approved exception.
- Default test clip: Netflix normal pair (`src01_hrc00_576x324.yuv` ↔
`src01_hrc01_576x324.yuv`). Checkerboard variants: use two
`checkerboard_1920_1080_10_3_*_0.yuv` pairs.
- Local-dev mirror of **T6-8 GPU-parity CI gate**
([ADR-0214](../../../docs/adr/0214-gpu-parity-ci-gate.md)): gate enforces
same per-feature cross-device variance budget on every PR. Run before push
to catch parity break locally, not burn CI cycle.
- GPUs NOT bit-identical to CPU as class invariant: close, but Netflix golden
gate CPU-only by design. Don't claim CUDA/SYCL/Vulkan parity stricter than
per-feature variance budget in ADR-0214.
