Benchmarks¶

Benchmarks measure the v2 primitives that matter: feature maps, solvers, online updates, and ML-ELM fit/forward paths.

Running benchmarks¶

docker compose run --rm dev ./scripts/run_benchmarks.sh

The script builds benchmark targets and writes JSON files to data/benchmarks/latest/.

Output files¶

File	Contents
`bench_feature_maps.json`	Additive, RBF, and ELM-AE transform benchmarks
`bench_solvers.json`	Ridge Cholesky, dual/primal behavior, and RLS update benchmarks
`bench_ml_elm.json`	ML-ELM fit and forward-pass benchmarks
`bench_elm_cuda.json`	Legacy CUDA ELM primitive benchmarks retained for comparison

Required fields¶

Successful benchmark entries include:

name
real_time
cpu_time
iterations
time_unit
custom dataset_size counter
custom device counter such as CPU or GPU:sm_89

GPU benchmark entries may report error_occurred: true on CPU-only hosts. Treat those entries as skipped runtime data, not correctness failures.

Interpreting results¶

Feature-map benchmarks isolate transform cost for additive, RBF, and ELM-AE layers.
Solver benchmarks compare CPU Cholesky paths and RLS updates.
ML-ELM benchmarks measure fit and forward cost, not accuracy.
Use Google Benchmark JSON for downstream badge and table generation.

Example¶

{
  "benchmarks": [
    {
      "name": "BM_AdditiveTransform/1024",
      "iterations": 100,
      "real_time": 12000,
      "cpu_time": 11980,
      "time_unit": "ns",
      "dataset_size": 1024,
      "device": "CPU"
    }
  ]
}