Docs 1.10.0 · 1 min read

Maximum performance on 1.10.0 — what to do, and why

Apsi 1.10.0 runs every program on its bytecode engine (Apsi compiled to Python bytecode). It has no native compiler yet — that is the headline of 2.0.0. The habits below are what make 1.10.0 fast, and why each one matters.

1. Let tensors do the loops

Do: sum(x * y) instead of a for loop adding x[i] * y[i]; A @ B instead of three nested loops.

Why: one tensor operation runs as optimised machine code inside NumPy for the whole array; a loop runs the bytecode engine once per element.

Difference: 50–500×. On the benchmarks, which are written as explicit loops on purpose, 1.10.0 is hundreds of times slower than C — that is the cost of per-element loops.

2. Use the built-ins

Do: sorted, argmax, cumsum, linalg.solve, evolve, nn.adam … rather than writing them yourself.

Why: the built-ins are vectorised and tested; your own loop version runs element by element.

3. Keep units on, they are cheap

Why: units are checked by the checker before running and validated once inside evolve, which then integrates on plain numbers. Leaving them out saves little and loses the safety.

4. Choose number formats deliberately

f32 / bf16 halve memory traffic for large AI models (dtype=f32, @precision(f32)).

5. Measure

apsi bench app.apsi

6. Need loops at C speed?

Upgrade to 2.0.0 and put @native on the function — the same loop-heavy code runs about 500× faster on the benchmarks. Programs written for 1.10.0 run on 2.0.0 unchanged.

Found a mistake? Report it · Markdown source