ADR-0592: Remove float_vif_score weak HSACO stub now that real HIP kernel ships¶

Status: Accepted
Date: 2026-05-18
Deciders: lusoris, Claude
Tags: hip, build, cleanup

Context¶

ADR-0379 (PR #1025, 2026-03-29) landed a real HIP kernel for float_vif at core/src/feature/hip/float_vif/float_vif_score.hip and registered it in core/src/meson.build under hip_kernel_sources. A later follow-up (PR #1303, "extend hip_hsaco_stubs.c to cover all 13 weak symbols") added a __attribute__((weak)) 1-byte stub for float_vif_score_hsaco to core/src/feature/hip/hip_hsaco_stubs.c so the host link always succeeded — even on configurations where the real HSACO blob did not build.

With enable_hipcc=true builds, the linker now sees two definitions of float_vif_score_hsaco: the strong xxd-embedded blob from the real .hip source and the weak 1-byte stub. The link rule resolves the strong symbol correctly, but the weak/strong array-size mismatch triggers -Wlto-type-mismatch warnings and leaves dead C source in the stubs TU. The user direction was unambiguous: paraphrased, "no stubs anywhere" for HIP HSACO symbols that have a real kernel behind them.

Decision¶

Remove the VMAF_HSACO_WEAK_STUB(float_vif_score_hsaco) entry from core/src/feature/hip/hip_hsaco_stubs.c. The real float_vif_score.hsaco blob produced from core/src/feature/hip/float_vif/float_vif_score.hip becomes the only definition of the symbol, the LTO warning disappears, and the stub TU shrinks by one line. The remaining 11 weak stubs (four ADM kernels, two ssimulacra2 kernels, and five other HIP extractors whose .hip sources are not yet portable) stay in place — each gets the same one-line removal when its kernel becomes real.

Alternatives considered¶

Option	Pros	Cons	Why not chosen
Remove `VMAF_HSACO_WEAK_STUB(float_vif_score_hsaco)` (chosen)	Eliminates the LTO warning, removes dead code, matches user direction.	Linker now hard-requires the real `.hip` to build under `enable_hipcc=true`.	Selected — the failure mode is loud and immediate (link error during dev) rather than silent (init returning `-ENOSYS` at runtime).
Keep the weak stub indefinitely	No behaviour change; defence-in-depth.	Leaves dead code; LTO keeps warning; user directive ignored.	Rejected — user explicitly asked for stub removal.
Suppress the LTO warning with a pragma instead of removing the stub	Single-pragma fix; preserves the fallback.	The fallback is no longer needed (the real `.hip` has been building cleanly since PR #1025).	Rejected — masking dead-code warnings hides drift.

Consequences¶

Positive: One fewer LTO warning, one fewer line of dead C, and the HSACO symbol table is unambiguous (one strong definition only).
Negative: A future configuration that flips enable_hip=true on without enable_hipcc=true will fail to link with an unresolved reference to float_vif_score_hsaco. Acceptable — that combination has never been a supported product surface (the only reason to enable HIP without hipcc is build-system probing, which the host TU's #ifdef HAVE_HIPCC guard already handles).
Neutral / follow-ups: As the remaining 11 weak stubs convert to real .hip kernels in future PRs, each gets the same one-line removal. hip_hsaco_stubs.c will eventually be deleted entirely when the last kernel ports.

References¶

ADR-0379 — original HIP float_vif port (PR #1025).
ADR-0536 — original stubs file design (ADM-only weak stubs).
PR #1303 — extension that introduced the float_vif_score_hsaco weak stub (one of 13 weak symbols added at once).
Source: req (paraphrased: "no stubs anywhere — replace the weak HSACO stub with the real kernel build path"; the user's PR assignment instructed direct removal of the float_vif_score_hsaco row from hip_hsaco_stubs.c after confirming the real kernel builds).