# rav1d-safe threading baseline (2026-03-24)
# CPU: x86_64 (WSL2), checked mode, no unchecked feature
# Configs: 0 = simd_1t (Native AVX2, 1 thread), 1 = scalar_1t (Scalar, 1 thread)
#
# Result: SIMD is 1.47-1.58x faster than scalar single-threaded.
# Multi-threaded scalar would need ~2 tile threads to break even with SIMD 1t.
# Multi-threaded SIMD is blocked by DisjointMut strided guard overlaps.

Image          | SIMD 1t (ms) | Scalar 1t (ms) | SIMD speedup
photo_2k.avif  | 65.3         | 95.8            | 1.47x
photo_4k.avif  | 233.3        | 350.9           | 1.50x
photo_8k.avif  | 1071         | 1697            | 1.58x

# DisjointMut overlap status:
# - f.lf.level (loopfilter level cache): FIXED via stack gather (e2de9f1)
# - Picture plane pixels: BLOCKED by strided_slice_mut creating
#   contiguous guards that span stride*h bytes for an w*h block.
#   Fix requires either strided guard support in DisjointMut or
#   per-row guard acquisition in SIMD functions.
