01 / WITHOUT SEMANTICCOMPUTE
Inspect the output
The candidate is plausible at normal scale. A few isolated bad channels and rounding differences are easy to miss, especially in a large texture or a long render sequence.
Candidate output, visually plausible
normal image viewone-ULP driftlarge numeric divergencespecial-value failure
02 / WITH SEMANTICCOMPUTE
Measure and localise
Apply a stated tolerance, count every mismatch, locate the first and worst pixel/channel, and distinguish small ULP drift from a NaN or a genuine algorithmic divergence.
03 / AFTER THE FIX
Re-verify the result
The package does not edit a shader. It makes the failure inspectable, supplies a likely remedy, and checks the developer's corrected output against the same trusted reference.
What the numbers differentiate
Switch between exact comparison and ULP(1). The tolerance policy matters: a one-step Float32 rounding difference may be acceptable in a stated contract; a NaN or a large numerical change is not hidden by that choice.
| Classified cause | Channels | Under exact | Under ULP(1) | How it reflects in an image |
|---|
Corrected candidate after the prescribed changes
same reference, re-verifiedRe-verification
0 diverging
The corrected candidate is bit-identical to the fixture reference. In a real workflow, the same report records the backend, precision, texture format, dimensions, tolerance and run status.
This page is a product visualisation, generated from a deterministic illustrative texture fixture in the browser. It demonstrates the reporting model behind sc_check_metal_texture_parity; it does not claim a Metal kernel ran, a GPU benchmark was performed, or a customer image was verified. Replace the fixture with a real reference/candidate pair when publishing run evidence.