fix(hip): stage the luma plane onto the device before the ADM DWT2 launch

[lusoris]

Summary

integer_adm_hip faulted the GPU on the first frame and killed the process:

Memory access fault by GPU node-1 on address 0x556905b5a000.
Reason: Page not present or supervisor privilege.

Tracked as T-HIP-INTEGER-ADM-GPU-PAGE-FAULT-2026-09-05; reproduced here on a gfx1030.

Running the test under AMD_SERIALIZE_KERNEL=3 HIP_LAUNCH_BLOCKING=1 AMD_LOG_LEVEL=3 named the offender as the very first kernel launched, adm_dwt2_8_vert_hori_kernel_4_16_32768_128_8_uint8_t, and put the faulting address in the host heap range — the tell that a host pointer reached a device kernel.

It did. extract_fex_hip passed ref_pic->data[0] and dis_pic->data[0] straight into dwt2_8_device_hip / dwt2_16_device_hip, but the HIP backend is host-pic (ADR-0530), so those are HOST pointers. The CUDA twin needs no staging because vmaf_cuda_picture_* hands it device memory already — the HIP port inherited the CUDA call shape without the thing that made it valid. integer_psnr_hip already stages correctly via its ref_in / dis_in buffers, which is exactly why test_hip_psnr_parity passed while the ADM test cored.

Fixed by allocating a per-side device staging buffer for the scale-0 luma plane in init_fex_hip and copying the host plane across with hipMemcpy2DAsync before the DWT2 launch. The staged rows are tightly packed, so the element stride handed to the kernel becomes w rather than the picture's stride.

on gfx1030	before	after
test_hip_adm_parity	GPU coredump (process killed)	3/3 passing
HIP parity suite	16 pass / 2 fail	17 pass / 1 fail
vmaf --backend hip --feature adm_hip	process killed	integer_adm2 = 0.962084

The one remaining HIP failure is test_hip_ssim_parity (4.53e-03), the separately-tracked T-GAP-HIP-INTEGER-SSIM-FLOAT-KERNEL-DEFERRED-2026-09-02 — an 11-tap float Gaussian where the CPU uses a 9-tap int64 kernel, formally deferred under ADR-0564. Untouched here.

Also in this PR: HIP large-fixture parity variants (ADR-1206)

With the ADM extractor able to run, the 960x540 sweep that found the SYCL bug (#1369) could finally run on HIP too. Result on gfx1030: 16 pass, 0 fail, 1 skip — no new HIP defects, which is a useful negative result.

Two tests are handled explicitly rather than silently:

• float_ssim_hip is a documented v1 scale=1-only extractor; its large variant records that contract as a skip, and still fails loudly if the twin ever stops refusing.
• test_hip_ssim_parity is excluded from the list: its divergence is the ADR-0564-deferred 11-tap float Gaussian, so a second fixture would only re-report a deferred gap at another resolution.

Metal is now the only family without large-fixture variants — no Apple hardware here to verify against.

Type

• fix — bug fix
• sycl / cuda / simd — backend-specific (HIP)

Checklist

• Commits follow Conventional Commits.
• make format && make lint is green locally (pre-commit run --files clean; assertion-density, check-copyright, check-state-md-rows pass).
• Unit tests pass: meson test -C build — HIP parity suite 17/1 on a gfx1030; the failure is the pre-existing deferred ssim gap.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2 — equivalent done directly: test_hip_adm_parity compares CPU vs HIP at places=4 and now passes, where it previously could not run at all.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap — only the HIP twin was broken; CUDA/SYCL/Metal receive device pictures or already stage.
• If I added a new .c / .h — no new source files.
• If this is a breaking change — not breaking. The extractor previously crashed; it now returns scores.
• If this PR adds an ADR, the row lives in docs/adr/_index_fragments/ — index in sync.

Bug-status hygiene (ADR-0165)

• docs/state.md updated — T-HIP-INTEGER-ADM-GPU-PAGE-FAULT-2026-09-05 moved to Recently closed; it also closes the second half of T-GAP-HIP-INTEGER-ADM-PICTURE-STAGING-DEFERRED-2026-09-02.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• If I believe a golden value must change — none does. The golden gate is CPU-only; this is a HIP-only change.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: the diagnosis is one serialized-kernel run that named the faulting kernel plus one read of the call site, both recorded in ADR-1211's Context.
• Decision matrix — ## Alternatives considered in ADR-1211, including making the whole HIP backend device-pic and disabling the extractor with .flags = 0, and why both were rejected.
• AGENTS.md invariant note — added to core/src/hip/AGENTS.md.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/hip-integer-adm-picture-staging.md.
• Rebase note — docs/rebase-notes.md, "ADR-1211 — HIP is host-pic; device kernels need staged input".

Reproducer

Needs a ROCm device. Note -Denable_hipcc=true — without it the HIP picture allocator is an -ENOSYS stub and every HIP extractor fails with problem reading pictures for an unrelated reason.

meson setup build core -Denable_hip=true -Denable_hipcc=true -Denable_cuda=false -Denable_sycl=false
ninja -C build
meson test -C build test_hip_adm_parity

On master the process dies with Memory access fault by GPU node-1; on this branch the test is 3/3.

Known follow-ups

• The staging copy is followed by a hipStreamSynchronize for correctness on this first cut; overlapping it with the existing ref/dis events is a performance follow-up.
• If more HIP extractors come to need device input, making the HIP backend device-pic (as CUDA is) becomes the better shape than per-extractor staging — see the ADR's alternatives table.

🤖 Generated with Claude Code