fix(sycl): return from vmaf_sycl_upload_plane only after its copy has completed

[lusoris]

Summary

vmaf_sycl_upload_plane() returned with its host-to-device copy still in flight. Nothing ordered that copy before the frame's compute, which runs on other queues: sycl_apply_input_barriers() waits on last_upload_event, which this function never set. The caller could also free the source while the copy was still reading it; the Windows D3D11 import unmaps its staging texture right after the call. On an Arc A380, every 3840x2160 frame uploaded this way and read at once scored the wrong pixels.

The function now orders the copy behind the slot's previous readers and waits for it before returning. This closes T-SYCL-UPLOAD-PLANE-NO-COMPUTE-FENCE-2026-10-05, which #2075 opened (now on master).

What reproduced (A380, xe)

test_sycl_zero_copy_model_gate, new case test_upload_plane_orders_compute: 8 frames of 3840x2160 that change every frame, uploaded with vmaf_sycl_upload_plane(), read at once by psnr_sycl (enable_chroma=false), compared with the CPU psnr. The test frees its source pictures right after the read, as a caller may.

Build	Frames equal to the CPU (%.17g)
without the fix	0 of 8, on 3 of 3 runs (psnr_y 4.9 to 6.1 where the CPU gives 7.04 to 7.15)
copy event recorded for the compute barrier, no wait	6 of 8 wrong, on 3 of 3 runs
this PR (fence behind the slot's readers, then wait)	8 of 8, on 3 of 3 runs

The test's other cases (576x324) were right before the fix too, because their copies finish first. Their explicit vmaf_sycl_wait_copy_queue() is gone, so every case now runs without a wait of the caller's own.

What changed

• core/src/sycl/common.cpp::vmaf_sycl_upload_plane(): sycl_fence_slot_readers() first (as vmaf_sycl_shared_frame_upload() does, ADR-1369), then last_ev.wait_and_throw() before returning.
• common.cpp, touched, is brought from 4 to 0 sycl-lane clang-tidy findings. Its four getenv() calls (VMAF_SYCL_PROFILE, VMAF_SYCL_TIMING, VMAF_SYCL_IMPORT_DEBUG, VMAF_SYCL_CHECKSUM) now read through the thread-safe snapshot vmaf_gpu_dispatch_env_get() (ADR-0488). The scoped baseline was tightened in the dev container.
• Docs: the libvmaf_sycl.h comment, docs/api/gpu.md and docs/backends/sycl/zero-copy.md state that the call returns after the copy.

Verification (A380; SYCL build with dg2-g11 AOT)

Check	Result
test_sycl_zero_copy_model_gate	5 of 5, three runs
the 66 fast+gpu+sycl tests	66 of 66
test_vmaf_sycl_threads, test_vmaf_feature_backend_sycl (slow)	pass
--suite fast without gpu	328 pass, 0 fail, 1 skipped
scripts/dev/tidy-lane.sh --write --only core/src/sycl/common.cpp sycl (dev container)	exit 0; common.cpp 4 → 0, tightened; the test file stays at 0

Not verified: the Windows D3D11 path, which calls this function (no Windows GPU host).

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: one ordering fix, measured in the state row
• Decision matrix — no alternatives: only-one-way fix (the source is the caller's memory, so the copy must finish before the call returns)
• AGENTS.md invariant note — core/src/sycl/AGENTS.md ("vmaf_sycl_upload_plane() returns after its copy completed")
• Reproducer / smoke-test command — below
• CHANGELOG fragment — changelog.d/fixed/sycl-upload-plane-waits-for-copy.md
• Rebase note — docs/rebase-notes.md, "vmaf_sycl_upload_plane waits for its copy (2026-10-05)"

Reproducer

source /opt/intel/oneapi/setvars.sh
CC=icx CXX=icpx meson setup build-sycl core -Denable_sycl=true -Dsycl_icpx_aot_targets=dg2-g11
ninja -C build-sycl test/test_sycl_zero_copy_model_gate
flock ~/.cache/vmafx-locks/sycl-a380.lock timeout 300 build-sycl/test/test_sycl_zero_copy_model_gate