fix(denoise): meta=1 residual accumulators overflow uint32 at 4K/8K#42
Merged
Conversation
The opt-in meta=1 residual-energy accumulators (abs_sum/sq_sum, uint32) summed |in-out|*GS with GS=1e6 over ~W*H/16 pixels per slice. At 4K a slice is 518K pixels, so a realistic residual of 0.02 sums to 1.04e10 -- 2.4x over UINT32_MAX -- and wraps silently, corrupting residual_energy/sigma_estimate/psnr in the PEL_SEC_DENOISE telemetry (overflow threshold: residual >=0.008 at 4K, >=0.002 at 8K). Lowered GS to 1e3 (worst-case 8K sum 2.07e9 < UINT32_MAX) and mirrored the host divisor. residual_energy (the mean) stays accurate; sigma_estimate is now coarse-but-bounded (a precise sigma would need a 64-bit atomic accumulator, noted as a follow-up). Opt-in telemetry only -- no effect on denoised pixel output. Found by an adversarial shader bug-hunt; the same hunt's false positives (a var/varr naming nit, an intentional lockstep barrier) were adversarially refuted and not changed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Fixes a confirmed defect from an adversarial shader bug-hunt: the
vf_pelorus_denoise_vulkanopt-inmeta=1residual-energy accumulators (abs_sum/sq_sum,uint32) silently overflow at 4K/8K.The bug
The GLSL emits
atomicAdd(abs_sum[slice], uint(|in-out| * GS))withGS=1e6, andslice = wg & 15spreads the whole frame across only 16 global accumulators → ~W·H/16pixels/slice (518K at 4K, 2.07M at 8K). A realistic denoise residual of 0.02 at 4K sums to 518400·0.02·1e6 = 1.04e10, 2.4× overUINT32_MAX→ wraps mod 2³² →residual_energy/sigma_estimate/psnrin thePEL_SEC_DENOISEtelemetry are garbage. Overflow threshold: residual ≥0.008 (4K), ≥0.002 (8K) — i.e. virtually any real denoise at ≥4K.The fix
Lowered
GSto 1e3 (worst-case 8K sum 2.07e9 < UINT32_MAX, ~2× margin) and mirrored the host divisor.residual_energy(the mean) stays accurate;sigma_estimateis now a coarse-but-bounded estimate (a precise sigma would need a 64-bit atomic accumulator — noted as a follow-up). Opt-in telemetry only — zero effect on denoised pixel output, and no.complockstep (the accumulator is a filter-only host-readback path).Provenance
Found by a 5-finder → adversarial-verify bug-hunt over the session's new shared-memory GLSL. The hunt's false positives were correctly refuted and not changed: a
var/varrlocal-name divergence (cosmetic, identical SPIR-V) and thepel_load_tileleading barrier (intentional — load-bearing in the filter's multi-plane loop, mirrored in the standalone for CI parity).Deliverables
Regenerated patch 0003 + changelog fragment. No ABI/surface change (same
PEL_SEC_DENOISEfields, just correct values). Pure bugfix — no ADR.🤖 Generated with Claude Code