Li Benchmarks

ml_conv2d_forward redvalidity pass

Tier 1 · wall_time · algo registry stub · variant algo_registry · OS linux

Validity gate

passpass — perf claims allowed when ratio is green

Source: latest.csv:perf_present

Wall-clock row exists only — no explicit correctness signal yet.

Ratio vs best competitor: 0.750 (75.0% of cpp speed) (best in series: cpp — Li is never labeled best)
Best competitor value: 0.0027 s

Numeric oracle

No analytical verify rows in ingest yet. Run tier-1 with --verify so latest.csv exports verify_ulps / verify_within_1ulp.

Problem size: algo registry stub
Category: micro
Pillar: numerics
Package: li-math
Li / catalog oracle (mean ± σ): 0.0036 s / 0.0027 s (cpp)
Ratio vs catalog oracle: 1.3333×
Best competitor: cpp (0.0027 s)
Li relative speed vs SOTA: 0.750 (1.0 = cpp speed)
Validity: validity pass
Threshold: 1.2×
Source: li-math/benchmarks/ml/ml_conv2d_forward

PH ids: PH-5b

Performance vs best competitor

Compare oracle: cpp

Relative speed vs best competitor (cpp) — SOTA = 1.0, higher is better. Absolute s values are in the table below.Validity gate not passed — bars are informational only.

cpp (release)SOTA
1.000
rust (release)
1.000
julia (release)
0.818
li (release)
0.750

Absolute measurements

Language comparison
Lang	Mean ± σ	Runs	Unit	Variant	OS
li	0.0036	—	s	release	linux
cpp	0.0027	—	s	release	linux
rust	0.0027	—	s	release	linux
julia	0.0033	—	s	release	linux
Metric: wall_time (value = mean of timed runs)

Host OS

Measurement hosts by language
Scope	OS
Row aggregate	linux
li (release)	linux
cpp (release)	linux
rust (release)	linux
julia (release)	linux

← Overview