Research-0434: VIF log2 LUT Shrink and Gaussian Filter Cache¶

Summary¶

Investigation into two low-risk performance improvements to the integer and float VIF paths, targeting the hot vif_statistic_avx512 function and per-frame filter allocation.

LUT Index Analysis (Win #1)¶

The log2_32 normalization uses __builtin_clz to compute k = 16 - clz(temp), then shifts temp >>= k. After this:

• k is chosen so that bit 15 of the result is 1 and bit 16 is 0.
• Result is always in [32768..65535] = [0x8000..0xFFFF].
• The always-set bit 15 is redundant as an index discriminator.
• Stripping it via & 0x7FFF yields an index in [0..32767] — a 32768-entry table.

Initial analysis incorrectly claimed a 16384-entry table (& 0x3FFF) was sufficient. A C verification test (/tmp/test_lut2.c) was used to confirm the correct mask is 0x7FFF before landing. The 16384-entry version produced values off by ~1.0-1.1 VMAF points (detected by the Netflix golden gate tests immediately).

The corrected 32768-entry table (64 KB) halves the original 128 KB footprint. It does not fit in L1D (32 KB on AMD Zen) but reduces L2 occupancy and TLB entry count for the three _mm512_i32gather_epi64 sites in vif_statistic_avx512.

Filter Cache Analysis (Win #3)¶

vif_get_filter(out, scale, kernelscale) calls get_1d_gaussian_kernel(out, n, n/5.0f) which computes a 1D Gaussian by calling expf for each coefficient. With scale from 0–3, kernelscale = 1.0 (default), the four filter widths are:

• scale 0: 17 taps
• scale 1: 9 taps
• scale 2: 5 taps
• scale 3: 3 taps

Across 4 scales × per-frame call = 34 expf evaluations per frame, fully avoidable since vif_kernelscale is read-only after VifState.init().

Benchmark Results¶

Platform: AMD Zen (host), AVX-512, BBB 1080p YUV (300 frames), vmaf_bench tool.

Note: the projections from /tmp/perf_findings.md (+8–12% total speedup for integer VIF) assumed L1 fit; since the corrected table is 64 KB (L2, not L1), the measured improvement is smaller (~4%). The gain is real and consistent across 5 runs.

References¶

• integer_vif.h: LUT definition, log2_32, log2_64
• integer_vif.c: log_generate
• vif_avx512.c: three gather sites
• vif.c: compute_vif loop
• float_vif.c: VifState, init, extract