Research-2127: CPU options on the SYCL PSNR, SSIM and float-motion twins — 2026-09-29¶

• Status: Active
• Workstream: RC3 SYCL, T-BUG048-GPU-OPTION-PARITY-REMAINDER-2026-09-26 (SYCL part); decision in ADR-1365
• Last updated: 2026-09-29

Question¶

The SYCL twins psnr_sycl, integer_ssim_sycl, float_ssim_sycl and float_motion_sycl lacked CPU options (min_sse, enable_mse, reduced_hbd_peak, enable_apsnr; enable_db, clip_db, enable_lcs; motion_max_val). Where does each option act, can it be applied from what the device already reduces, and how close to --backend cpu does each get on real hardware?

Sources¶

• CPU references: core/src/feature/integer_psnr.c, integer_ssim.c, float_ssim.c + ssim.c + iqa/ssim_tools.c, float_motion.c.
• Twins: core/src/feature/sycl/integer_psnr_sycl.cpp, integer_ssim_sycl.cpp, float_motion_sycl.cpp; the Metal float_ssim_metal.mm L/C/S reduction for comparison.
• Toolchain: oneAPI DPC++ 2026.1 (icpx), AOT for bmg-g21,adl-s, vmaf-dev-mcp:ocloc.
• Host: i9-12900K; Arc B580 (Level Zero device 0) and UHD 770 (device 1) through WSL2 /dev/dxg.

Findings¶

Where each option acts¶

So only enable_lcs needs device work: a second vertical-pass kernel that computes L, C and S from the moments it already has and reduces three more per-work-group partials. Everything else is host arithmetic on values the device already hands back once per frame.

Identical frames and the dB form¶

With enable_db the CPU reports +inf for a perfect score, or the clip_db ceiling (ADR-1221). Both CPU paths give exactly 1 for identical frames: integer_ssim.c computes in exact-integer doubles, and iqa_ssim sums identical l, c, s = 1 terms. The fp32 twins did not. In integer_ssim_sycl, the denominator's second factor x2*w - mx^2 + y2*w - my^2 + c2 rounds differently from the numerator's 2*(xy*w - mxmy) + c2, and icpx contracts a*b + c inside one expression, so a numerator and denominator that are mathematically equal differ in the last bits. With a residue of 1e-7 planted in the ratio, identical 161x91 frames score 69.2 dB on the device against the CPU's 93 dB ceiling. Holding each product in a named temporary, summing the two variances as a pair and returning 1 on numerator == denominator makes the two sides the same operations mirrored; identical windows then score exactly 1. The default linear scores move by at most 1.05e-8 against the pre-change twin on both devices (2.2e-8 on identical frames, which now reach 1).

Two float_motion_sycl divergences found on the way¶

• motion_force_zero was in the option table but only flush() read it; collect() emitted real motion scores. The CUDA, HIP and Metal twins honour it.
• The debug VMAF_feature_motion_score was emitted without motion_fps_weight; the CPU emits motion_clip(score). The CUDA twin has the same gap (not changed here).

Parity per option¶

CPU vs SYCL twin, per frame, --precision max, worst absolute difference over both devices (B580 and UHD 770 agree to the digits shown). Fixtures: the Netflix src01 576x324 pair (48 frames, 8-bit 4:2:0); 853x480 4:4:4 8-bit (12 frames) and 10-bit (8 frames), resampled from that pair (the CLI refuses odd widths for 4:2:0); 576x324 4:2:0 10-bit (12 frames, << 2 plus LCG dither); an identical 853x480 pair (4 frames). float_ssim runs with scale=1 because 853x480 auto-resolves to 2, which the twin does not implement.

Every PSNR value, mse_* and apsnr_* aggregate included, is bit-identical. The dB rows are the linear differences magnified by 4.34 / (1 - ssim); the parity-gate tolerance (5e-5) is a linear-domain figure. The baseline-vs-new comparison of the default outputs is bit-identical for psnr_sycl and float_motion_sycl on both devices.

The CPU psnr extractor, now calling psnr_score.h, is bit-identical to the previous binary for 8 option sets on the four real fixtures, and the non-slow Netflix golden gate passes on a gcc build (271 passed, 12 skipped).

The test fails when the implementation is wrong¶

test_sycl_twin_option_parity was run against planted regressions, one at a time, on the B580: the peak ignoring reduced_hbd_peak (10-bit psnr_y off by 2.6e-2), mse_* emitted for luma only, the motion cap removed (motion2_mmxv_39.9 51.0 instead of 39.9), the SSIM ratio scaled by 0.9999999 (identical frames 69.2 dB instead of the 93 dB ceiling) and the structure term using the unclamped covariance (float_ssim_s off by 1.5e-2). Each failed the test; the unmodified build passes on both devices. Dropping the release of the L/C/S partials fails test_sycl_init_unwind.

Open questions¶

• The CUDA, HIP and Metal twins still lack these options; integer_ssim_hip has no enable_db / clip_db and float_ssim_hip no enable_lcs, beyond what the state row listed. They cannot be tested on this host.
• float_ssim_cuda declares an enable_chroma option the CPU float_ssim does not have.