Repository navigation
fix(feature-cpu): CIEDE 4:2:2 chroma swap (OOB) + cambi init leak + integer_ssim 16bpc overflow (bug-hunt r1+r2) - #1050
Merged
Merged
Conversation
lusoris
force-pushed
the
fix/bughunt-feature-cpu
branch
2 times, most recently
from
June 27, 2026 14:45
c69a11d to
dfa581c
Compare
lusoris
changed the title
lusoris
marked this pull request as ready for review
June 27, 2026 14:54
lusoris
enabled auto-merge (squash)
June 27, 2026 14:54
…nteger_ssim 16bpc overflow Bug-hunt feature-cpu batch (rounds 1+2): - CIEDE2000 4:2:2 chroma-upsample ss_hor/ss_ver flag swap (heap OOB read + wrong scores on YUV422P); fix + load-bearing 422 regression tests. - cambi init() partial-failure leak: route error paths through null-tolerant close_cambi(). - integer_ssim 16-bpc samplemax*samplemax signed-int overflow (round-2 R2-1): 65535*65535 > INT_MAX wraps to -131071 in int, corrupting the SSIM c1/c2 stability constants for 16-bit input and diverging from the CUDA/HIP/SYCL twins (which use int64_t/double). Hoist const double sm = (double)samplemax. 8/10/12-bpc bit-unchanged (product fits int). Adds load-bearing test_ssim_16bit_distorted_in_range (0.999992 with fix vs 1.185 out-of-[0,1] with the overflow). Round-2 R2-8 (cambi histogram alloc int-overflow) was already fixed in tree (c_values_histograms already casts (size_t)alloc_w) -- skipped. Golden-safe: 8-bit golden SSIM/CIEDE unchanged; no Netflix assertions touched. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
lusoris
force-pushed
the
fix/bughunt-feature-cpu
branch
from
June 27, 2026 14:55
dfa581c to
15ea919
Compare
4 of 6 tasks
9 of 12 tasks
lusoris
added a commit
that referenced
this pull request
Oct 2, 2026
…each with its size and the upstream change that ends it (ADR-1479 to ADR-1486) The reference for code inherited from Netflix/vmaf is Netflix's source; a difference needs an ADR. The upstream parity audit of 2026-10-02 found deliberate differences that had none of their own, or whose ADR (ADR-1033) names neither upstream's behaviour nor the size: - ADR-1479 ciede on 4:2:2: chroma flags (fork PR #1050); 0.153 on 48 of 48 frames; Netflix/vmaf#1611. - ADR-1480 speed_temporal buffers at speed_prescale above 1 (#1643); up to 195, upstream segfaults on two fixtures; Netflix/vmaf#1627. - ADR-1481 a failing extractor fails the run (#871); status only, 78 probe runs where upstream is silent and 88 where it crashes. - ADR-1482 integer adm on frames of 17 to 32 pixels (#1473, #1507); scale 3 up to 0.23; Netflix/vmaf#1599, #1600. - ADR-1483 odd-sized chroma planes round up (4f08d32); psnr_cb / cr up to 0.684 / 0.826 dB, ciede 0.198. - ADR-1484 float_ms_ssim magnitude before pow() (#641, ADR-1033 item 2); NaN upstream on the 10 px checkerboard; Netflix/vmaf#1665. - ADR-1485 apsnr of a plane without error (#641, item 1); 114 against 60 dB; Netflix/vmaf#1666. - ADR-1486 float_motion scale-1 stride (#641, item 9); up to 25.1; Netflix/vmaf#1667. Each ADR gives upstream's file and line at Netflix 9e48141b, the fork's lines, the reason found in the fork's pull request, commit or code, and the measured size from the audit. Documentation only.
15 of 26 tasks
lusoris
added a commit
that referenced
this pull request
Oct 2, 2026
… and mirror them in the CUDA, SYCL and HIP twins (ADR-1476) (#1892) * fix(ciede): form ciede2000()'s two products in float as upstream does and mirror them in the CUDA, SYCL and HIP twins (ADR-1476) Netflix's ciede2000() writes sqrt(c_prime_1 * c_prime_2) and + r_sub_t * chroma * hue with float operands, so both products are rounded to float before the double expression widens them (libvmaf/src/feature/ciede.c:224-225, :235-236 at 9e48141b). PR #552, a CodeQL sweep, cast the first operand of each to double, which keeps the products exact and changes the score. The upstream parity audit of 2026-10-02 found it (U4); the golden gate does not see it. ciede.c carries upstream's expressions again; the squares stay products (ADR-1467). cuda/integer_ciede/ciede_device.h and ciede_ff_math.h (SYCL and HIP) form the same float products in place of the exact ones. Against Netflix master (C API, %.17g, 327 frames from 8x8 to 3840x2160, GCC 16.2.1, glibc 2.44), scalar and default dispatch: 153 frames identical, 7 before. The rest: 119 frames up to 2.2e-11 are ADR-1467's product where upstream calls powf(x, 2) (272 of 327 with that call put back), 48 are 4:2:2 input (chroma flag fix, PR #1050), 7 are odd sizes (chroma planes rounded up). The CPU score moves by up to 1.3e-9 on frames of 160x90 and larger. Twins, 153 frames at --precision max, largest difference from the CPU before and after: CUDA (RTX 4090) 8.4e-13, 2.1e-12; SYCL (Arc A380) 7.3e-13, 2.1e-12; HIP (gfx1036) 7.3e-13, 2.1e-12; bound 1e-9. test_ciede_upstream_products replays the CUDA header's ciede_delta_e() with float and with widened products over 20 000 colour pairs (193 tell them apart); test_ciede_device_math holds the header to the CPU extractor bit for bit. The SYCL and CUDA source contracts pin upstream's statements. Netflix golden gate: 271 passed, 12 skipped on x86-64 and aarch64. * docs: regenerate the indexes and the citation map after rebasing
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
Confirmed, adversarially-verified fixes from the codebase bug-hunt sweep. See commit body for per-finding detail + verification. Golden-safe: no Netflix assertAlmostEqual modified.
Reproducer / smoke
See commit body — each fix lists its build + test evidence (golden gate / parity / regression tests run locally).
Deliverables (ADR-0108)
core/src/feature/AGENTS.md.changelog.d/fixed/bughunt-feature-cpu-ciede-422-cambi-leak.md.docs/rebase-notes.mdentry.T-BUGHUNT-FEATURE-CPU-2026-06-27(Recently closed).I did not modify any Netflix golden assertAlmostEqual value.
no docs needed: bug fixes to existing feature extractors (CIEDE2000 4:2:2 chroma swap, cambi init-OOM leak, integer_ssim 16-bpc samplemax² overflow) — correctness only, no new or changed user-discoverable surface or output schema; ADR-0100 / CLAUDE §12 r10 bug-fix exclusion.