dot 2M × 20
Order-free sum; approximate floating point; contiguous f64 vectors; whole process.
BENCHMARK OBSERVATORY
Numbers without source are not results.
These are median whole-process times in milliseconds. Initialization and output are included. Answers are checked before timings are accepted. Missing runtimes are not estimated.
- Snapshot
- 2026-07-23
- Machine
- Apple arm64 · macOS 26.6
- Runs
- 7 per command · median
- Source state
- 301bfee · dirty, recorded
01 / RESULTS
Lower is better.
| Runtime | dot 2M × 20 | slerp 2M |
|---|---|---|
| lulang AOT | 64.053 ms | 12.738 ms |
| lulang JIT | 56.535 ms | 36.039 ms |
| lulang selfhost | 13.677 ms | 12.096 ms |
| C++ -O3 | 27.815 ms | 14.259 ms |
| C++ fast | 11.902 ms | 12.669 ms |
| Rust | 27.089 ms | 14.583 ms |
| Julia | — | — |
| NumPy | 63.199 ms | 135.866 ms |
| JavaScript | 44.622 ms | 35.352 ms |
02 / SOURCE AND ASSUMPTIONS
The comparison is inspectable.
03 / ABLATIONS
Optimizations can be switched off.
The same source can be rerun with individual transformations disabled. Layout is explicit in the report: contiguous vectors for dot, value quaternions for slerp, and SoA for record-array kernels.
Read the experiments ↗- LU_MATH
- inline | call
- LU_IFCONV
- on | off
- LU_LICM
- on | off
- LU_SIMD
- on | off
- LU_LAYOUT
- soa | aos
05 / REPRODUCE
Run the same measurement.
python3 benchmarks/run_observatory.py --runs 7 --bootstrapThe runner builds the host and three-stage self-hosted compilers, checks numerical answers, records tool versions, and writes the table.