SYCL Zero-Copy NaN Diagnosis — Phase 1 Research Digest¶

Diagnostic branch: fix/libvmaf-sycl-zerocopy-nan Verdict commit: see git log --oneline -- docs/research/sycl-zero-copy-nan-diagnosis.md Date: 2026-06-30 Phase: 1 (Diagnose) — Plan 01-02 (DIAG-02) ADR-0108 role: Phase 2 / Phase 4 research-digest stub (citable by commit hash from the fix PR)

Summary¶

This digest records the decisive measurement that selects the Phase 2 fix direction for the libvmaf_sycl zero-copy NaN bug in VMAFx. The bug causes the libvmaf_sycl FFmpeg filter (QSV/VAAPI zero-copy path) to return VMAF score: nan on matched QSV-decoded 4K HDR input while the equivalent libvmaf host-upload filter (same machine, same SYCL device) produces correct scores at ~119 fps.

Verdict: hyp 1 confirmed — buffer content wrong in the VA-import path → Phase 2 = FIX-01 (double-buffer the VA-import path; converge onto the shared_frame_upload slot-toggling pattern from common.cpp:597-598).

Test Setup¶

Machine: NAS server, Intel Arc A380 (SYCL device 0), Fedora, Kernel 6.19 Container: vmafx built from Containerfile.vmafx (oneAPI basekit + ffmpeg n8.1.1 + VMAFx) Probe code: vmaf_sycl_checksum_y_slot (commits b9c0088d + df8a382e on this branch) injected via probe-diag.patch build-context patch (not pushed to remote).

Input pair (Pair 1 — 4K HDR, PRIMARY):

Both decoded with QSV (hevc_qsv / av1_qsv). Formats match — no conversion needed.

Oracle: libvmaf filter with sycl_device=0 + VMAF_SYCL_NO_GRAPH=1 (host-upload path, proven working at ~119fps). Input converted via hwdownload,format=p010le before the filter. Reproducer: libvmaf_sycl filter (zero-copy VA-import path, broken). Env gate: VMAF_SYCL_CHECKSUM=1 for both runs. Frame cap: 100 frames (NaN signature appears within first 10-15 frames).

Key Measurements¶

Aggregate VMAF¶

Zero-cost gate: Reproducer without VMAF_SYCL_CHECKSUM also gives NaN — probe does not perturb the result.

Slot Behavior¶

The D2H probe reads state->cur_compute and logs it as slot=N for each frame.

The host-upload path correctly toggles cur_compute via shared_frame_upload (common.cpp:597-598). The VA-import path calls vmaf_sycl_advance_frame which increments only frame_counter, not cur_compute — cur_compute stays 0 forever.

Per-Frame Checksum Diff¶

80 frames compared (oracle: 100 frames; sycl: 80 frames before VA surface error).

Sample rows:

Frame  Host-slot  Host-ref-CRC  Host-dis-CRC  Sycl-slot  Sycl-ref-CRC  Sycl-dis-CRC  Match
    0          0    0x09271dc5    0x09271dc5          0    0xa3471dc5    0xa3471dc5    DIFF/DIFF
    8          0    0x09271dc5    0x09271dc5          0    0xbcf83fca    0xa3471dc5    DIFF/DIFF
    9          1    0x09271dc5    0x09271dc5          0    0x6800c257    0xa3471dc5    DIFF/DIFF
   10          0    0xcefe8929    0x28e13197          0    0xb92ea64f    0x6800c257    DIFF/DIFF
   11          1    0x045103d1    0xda0a893b          0    0x1d955aa7    0xbcf83fca    DIFF/DIFF
   12          0    0xd8482266    0x02a49da4          0    0x92638727    0xb92ea64f    DIFF/DIFF
   79          1    0xcd3a0f58    0x4bd3c4fb          0    0x4f204a63    0x0cf86c66    DIFF/DIFF

Secondary observation: sycl_dis[N] frequently matches sycl_ref[N-k] for small k (1–6), with irregular offsets. Example: sycl_dis[10]=0x6800c257 = sycl_ref[9], sycl_dis[11]=0xbcf83fca = sycl_ref[8], sycl_dis[12]=0xb92ea64f = sycl_ref[10]. This pattern (dis buffer reading stale ref data from earlier frames) is consistent with either (a) the dis VA-import writing to shared_ref_buf[0] instead of shared_dis_buf[0] due to a pointer confusion, or (b) a race where the single shared slot 0 is read during an in-flight import. Phase 2 should instrument vmaf_sycl_get_shared_dis to confirm the pointer target.

Per-Frame VMAF¶

Sycl VMAF=100.000000 (ref==dist, impossible) from frame 1 onwards. Sycl integer_motion mean 4157.5, max 4903.3 (vs oracle normal range 0–48). Both signatures match HANDOFF.md predictions.

Frame-0 Caveat¶

Frame-0 checksums DIFFER between oracle and sycl (0x09271dc5 vs 0xa3471dc5) because the two code paths allocate USM buffers with different initial states. However, frame-0 VMAF matches (97.428027 on both paths) because frame 0 has no temporal predecessor — integer_motion[0]=0 and adm[0]=vif[0]=1.0 from initialization. Frame-0 VMAF agreement is NOT evidence of buffer correctness. Hyp-2 determination uses frames >= 1.

Verdict¶

hyp 1: buffer content wrong (slot/overwrite race) → Phase 2 = FIX-01¶

ANY frame having crc_host != crc_sycl for ref or dis selects hyp 1. We have 80/80 such frames — an unambiguous result.

The sycl path puts the WRONG bytes into the compute slots. With cur_compute permanently 0, both imports (ref and dis) and all compute operations share a single slot-0 buffer, creating races where import for frame N+1 overwrites the slot that compute is reading for frame N. Additionally, the dis buffer pattern suggests the dis import may be targeting the wrong USM buffer pointer entirely.

Fix direction (Phase 2 — FIX-01):

1. Make vmaf_read_pictures_sycl write to the UPLOAD slot (not compute slot), then advance cur_compute — mirroring shared_frame_upload (common.cpp:597-598).
2. Verify that vmaf_sycl_import_va_surface for is_ref=0 calls vmaf_sycl_get_shared_dis (not vmaf_sycl_get_shared_ref) to target the correct USM buffer.
3. Ensure the slot-1 combined command graph (recorded at common.cpp:834-835) is replayed for frame N while frame N+1 is imported to slot 0 — proper double-buffer separation.
4. Re-run this probe after the fix: verify checksums match oracle for all frames.

Research Predictions vs Observations¶

FFmpeg-Patch Parity¶

vmaf_sycl_checksum_y_slot is internal to core/src/sycl/common.h — not in any public header or consumed by ffmpeg-patches/0005-*. No patch update required (CLAUDE.md rule 14).

Files Referenced¶

The salient evidence rows are embedded above. The CSV files are scratch-space diagnostic artifacts; the embedded rows in this document are the durable record.