Repository navigation
docs: record the fork's check of the upstream defects verified on 6ec23e8f2 - #1761
Merged
Merged
Conversation
…it-identical (#1734) * fix(cuda): compute float_adm in the CPU's arithmetic so the twin is bit-identical float_adm_cuda matched the CPU extractor on 144 of 791 measured scores and was up to 1.3e-5 from it (adm_scale0 at 3840x2160). An exact twin was built first and the old constructs were put back one at a time; all of them together give the old twin's output bit for bit. Sizes alone: - the angle test's threshold as cos^2 * (|o|^2 * |t|^2), where adm_angle_flag_s() evaluates (cos^2 * |o|^2) * |t|^2 (1.3e-5); - each row reduced per tile and the rows added in fp64, where adm_csf_den_scale_s() and adm_cm_s() add a row into one float and the rows into another (4.4e-7); - CSF weights from a host copy of dwt_quant_step() with fp32 intermediates, four of eight default weights 1 to 3 ulp off (2.1e-7); - t / (o + eps), where the x86 CPU multiplies by rcp_s(), a Newton step on the processor's RCPSS estimate (1.3e-7); - the masking threshold's centre taps after its neighbours, where adm_cm_thresh3x3_s() adds the centre fifth in one sum per band (9.4e-8); - fp32 1/15 and 1/30 where the CPU's are double literals (7.2e-8, 1.5e-10), and an fp32 gain limit (1.0e-7 at 1.2); - the frame sums floored at 1e-2 * area / 1080p where compute_adm() uses 1e-10: adm2 = 1 instead of the CPU's 0 on a flat 16-bit frame with one raised sample and adm_noise_weight=0. ADR-1420: - cuda/float_adm/float_adm_device.h holds the decouple, the CSF, the threshold and the reduction terms of adm_tools.c operation for operation, with explicit round-to-nearest intrinsics on the device; the host compiles it for a test. - RCPSS is specified by an error bound, so the reference's quotient belongs to the host processor. adm_reciprocal_model_probe() fills a 4096-entry table from the instruction and proves the model against it (9.4 ms at extractor start); the decouple kernel evaluates it in integer arithmetic. A CPU build that divides (MSVC, ARM) makes the twin divide. - float_adm_terms stores nine terms per sample of the reduced region, float_adm_row_sums adds each row in one thread, the host adds the rows. - adm_tools.c exports its reduced region, CSF weights, angle constant and reciprocal estimate (adm_float_reference.h), and four of its reductions call the exported adm_pool_bands_s() instead of repeating it. CPU scores are unchanged (4 752 outputs). - EXACT_TWINS lists float_adm / cuda: the gate cell is an equality. Measured on an RTX 4090 at --precision max against master 5c8b9e9, identical scores before and after: Netflix 576x324 8-bit 66/336 and 336/336, 10-, 12- and 16-bit 2/21 and 21/21, checkerboard 1 px 2/21 and 21/21, 10 px 4/21 and 21/21, BBB 3840x2160 66/350 and 350/350; 658 and 2034 of 2034 with debug=true, also with clang's CUDA driver and with non-default gain limit, bypass, noise weight and viewing geometry. Not identical: adm_p_norm other than 1 or 3, within 1.1e-7 (device powf). A run of the twin alone takes 1.87 and 1.98 ms per 4K frame; its kernels take 0.76 and 1.11 ms (T-CUDA-FLOAT-ADM-EXACT-THROUGHPUT-2026-10-01). Found on the way and recorded: the SYCL, HIP and Metal twins carry the same arithmetic and floor (T-GPU-FLOAT-ADM-CPU-ARITHMETIC-2026-10-01, T-GPU-FLOAT-ADM-FRAME-SUM-FLOOR-2026-10-01); the CPU's float ADM depends on the processor through RCPSS (T-FLOAT-ADM-RECIPROCAL-ESTIMATE-HOST-DEPENDENT-2026-10-01), reads outside its bands below 17x17 (T-FLOAT-ADM-TINY-FRAME-BAND-READS-2026-10-01) and files its debug ratio under an unsuffixed key (T-FLOAT-ADM-DEBUG-KEY-UNSUFFIXED-2026-10-01). Tests: test_cuda_float_adm_parity asserts equality over 15 cases, each of which fails on the old twin; test_float_adm_device_math compares the header with the CPU routines and the reciprocal model with the instruction on the host; test_cuda_float_adm_exact_contract.py pins the design. * docs: regenerate the indexes and the citation map after rebasing
* chore(deps): Update dependency openai to >=3.22.1 * chore(deps): regenerate dev-llm lock for openai >=3.22.1 * docs: regenerate the indexes and the citation map after rebasing
…nd HISS standard (ADR-1142) (#1767) * test(core): refactor test_barten_csf for standards and HISS compliance * test(core): refactor test_psnr_hvs_simd for HISS modularity * test(core): refactor test_propagate_metadata for standards compliance * test(core): refactor test_thread_pool for standards compliance * test(core): refactor test_context for standards compliance * test(core): refactor test_predict for standards compliance * test(core): refactor test_ciede for standards compliance * test(core): refactor test_pic_preallocation for standards compliance * test(core): refactor test_locale_handling for branch and locale hygiene Add SPDX license identifier, wrap test TU in ADR-0141 and ADR-1138 NOLINT brackets, extract output verification and tmpfile helpers to satisfy branch complexity limits, and replace rewind with checked fseek. * test(core): refactor test_dict for anonymous namespaces and branch hygiene Add SPDX license identifier, replace NULL with nullptr, move test functions into discrete anonymous namespaces to stay within HISS-04 60-line block budgets, and simplify assertions into modular helpers to satisfy branch complexity limits. * test(core): flatten inner 8x8 block loop in test_psnr_hvs_simd Flatten nested i and j block loops into a single k loop to remain within the nesting depth threshold 4 without altering floating-point accumulation. * chore(ci): tighten CPU tidy and HISS debt baselines for batch A Record updated debt ratchets for 10 core test TUs in tidy-baseline-cpu.json and decrement active infractions by 2 in .standards-baseline.json and README.md. Add changelog fragment for std-core-tests-a and update CHANGELOG.md. * docs: regenerate the indexes and the citation map after rebasing
…s bit-identical (#1759) * fix(cuda): return ssimulacra2's sums in the CPU's order so the twin is bit-identical ssimulacra2_cuda computed the CPU's planes and the CPU's fp64 per-pixel terms and added the terms in a fixed tree. ssim_map() and edge_diff_map() add each of a channel's six terms pixel after pixel into one double; every add rounds, so the order changes the last digits. 8 of 113 measured frames matched and the rest were up to 7.3e-11 away. ADR-1433: the device forms the sums of those loops. - feature/ordered_sum.h: while a running sum of non-negative doubles stays in one binade, adding a term adds an integer. A run of terms carries two integer increments (even and odd start, because a tie rounds to even), runs compose in order, and a walk adds each chunk's increment after checking it against the exact sum; a chunk that crosses a binade is added term by term. Compiled into the kernels and into a host test. - ssimulacra2_device.cu: chunk_sums, chunk_plan, chunk_units and ordered_totals replace combine_partials and combine_final. A chunk is 1024 pixels in raster order; 9 to 27 of 8100 chunks per sum take the term-by-term path on a 4K frame. The readback stays 864 bytes. - The gate lists ssimulacra2 / cuda as an exact twin. Measured on an RTX 4090 at --precision max against master 5c8b9e9, identical frames and largest difference, before and after: Netflix 576x324 8-bit 8/48, 1.3e-13 and 48/48; 10-, 12- and 16-bit 0/3 and 3/3; checkerboard 1 px 0/3, 3.4e-13 and 3/3; checkerboard 10 px 0/3, 7.3e-11 and 3/3; BBB 3840x2160 0/50, 1.5e-12 and 50/50 (200 frames through the gate: 0). It costs time: a run of the twin alone takes 15.6 ms per 3840x2160 frame instead of 7.8 ms and 1.7 ms instead of 0.4 ms at 576x324; the CPU extractor takes 126 ms per 4K frame on sixteen threads (T-CUDA-SSIMULACRA2-EXACT-THROUGHPUT-2026-10-01). The SYCL and HIP twins keep their tree of fp32 pairs (T-GPU-SSIMULACRA2-SUM-ORDER-2026-10-01). * docs: regenerate the indexes and the citation map after rebasing
…23e8f2 (#1761) * docs: record the fork's check of the upstream defects verified on 6ec23e8f2 Fifteen defects reproduced on Netflix master were run against the fork: three reproduced and are fixed (#1305, #1420 as a hang, #1613), two are documented (#910, #755 and #1180), ten are not affected. The dated section in known-upstream-bugs.md and the Confirmed not-affected rows of the state ledger carry the evidence; Netflix 8e7a1ac4e (revert of #1476) needs nothing from the fork. * docs: regenerate the indexes and the citation map after rebasing
lusoris
force-pushed
the
docs/known-upstream-bugs-2026-10-01
branch
from
October 1, 2026 20:57
29cf75f to
4a0e1d8
Compare
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.
Summary
Records the result of checking the fork against the fifteen upstream defects verified on Netflix master
6ec23e8f2: a dated section indocs/development/known-upstream-bugs.md(upstream issue, defect, what the fork does and the evidence, or the fork PR that fixed it) and the not-affected results indocs/state.md("Confirmed not-affected"). No code changes.Outcome: three reproduced and are fixed (#1305 vif accumulator stream by #1750, #1420 as a hang in
vmaf_close()by #1752, #1613 chroma download by #1754), two are documented (#910, #755 and #1180 by #1753), ten are not affected. Also records that Netflix master8e7a1ac4e(revert of #1476) needs nothing from the fork.Type
docs— documentation onlyDeep-dive deliverables (ADR-0108)
AGENTS.mdinvariant note — no rebase-sensitive invariants: documentation only.Reproducer
~/.cache/vmafx-upstream-rebase/evidence/<issue>/holds the upstream reproducers; each row says what was run.Notes