Skip to content

fix(build,hip): round-3 batch — HIP vif uninit-err + NVTX libdl + icx ssim fp-model (R3-6/9/10) - #1062

Merged
lusoris merged 1 commit into
masterfrom
fix/round3-build-gpu-batch
Jun 27, 2026
Merged

lusoris merged 1 commit into
masterfrom
fix/round3-build-gpu-batch

Conversation

@lusoris

@lusoris lusoris commented Jun 27, 2026

Copy link
Copy Markdown
Contributor

Round-3 build/GPU error-path fixes (golden-safe). R3-6 HIP integer_vif uninitialized err code; R3-9 NVTX empty libdl target → cc.find_library; R3-10 SSIM AVX2 carve-out missing icx strict-fp. CPU build 1280/1280, libvmaf 3.2.0. R2-6 icpx-motion deferred (needs scoped motion carve-out).

no docs needed: build-system + GPU error-path bug fixes, no user-discoverable surface.
🤖 Generated with Claude Code

…sim fp-model (R3-6/9/10)

R3-6 HIP integer_vif init: declare int err=0 early (fail_stream/fail_submit
returned uninitialized). R3-9 NVTX: cc.find_library('dl') not empty
shared_library('dl'). R3-10 ssim AVX2 carve-out: + _x86_simd_strict_fp_extra
(icx -fp-model=precise; no-op gcc/clang → golden unchanged). R2-6 icpx-motion
deferred (needs scoped motion carve-out). meson libvmaf 3.2.0; CPU 1280/1280.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@lusoris
lusoris merged commit 1187e93 into master Jun 27, 2026
@lusoris
lusoris deleted the fix/round3-build-gpu-batch branch June 27, 2026 19:26
@lusoris lusoris added this to the 1.0.0 — First release milestone Sep 4, 2026
lusoris added a commit that referenced this pull request Sep 16, 2026
22 inline code-scanning findings on #1425, triaged one at a time rather
than dismissed or suppressed in bulk. Five are real:

- `read_luma8()` guarded `shift == 0` after the `bitdepth == 8` case had
  already returned, so the arm was unreachable. Removing it also makes
  visible that `shift` is 2, 4 or 8 there and `shift - 1U` cannot
  underflow.
- `check_filtered_center()` and `check_spatial_mask_first_image()` took
  `VmafPicture` by value and then took its address; both take a pointer.
- `test_vif_lifecycle.c` and `test_vmaf_roi_bounds.c` compared floats
  with `==`. Exact equality is the *intent* — bit-exact frame
  differencing, exactly symmetric saliency — so loosening them to a
  tolerance would hide the defect they exist to catch. They now compare
  bit patterns, the same shape `test_feature_isa_invariance.c` already
  documents for the same reason.

Fifteen are false positives, each checked rather than assumed:

- every `cpp/unused-static-function` hit is called. The `pdjson.c`
  helpers are called directly; `json_begin_error` only through the
  `json_error()` macro; the thread-pool helpers through
  `#define pthread_create observed_create` combined with
  `#include "../src/thread_pool.c"`, an interposition CodeQL cannot
  follow. gcc with `-Wunused-function` reports none of them.
- the `cpp/constant-comparison` hits in `feature_extractor.cpp` and
  `fex_ctx_vector_internal.h` are `SIZE_MAX / sizeof(T)` overflow
  guards. They are provably dead on 64-bit, which is what CodeQL sees,
  and live on i686 — a target `libvmaf-build-matrix.yml` gates. Deleting
  them to quiet a note would be a real 32-bit regression.

The two Semgrep findings are #1062/#1063, already dismissed earlier;
their PR comments are stale, because GitHub applies an in-source
suppression only at alert-creation time.

148/148 tests pass.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant