상호 운용 성능의 함정

Calling Python from Mojo is powerful, but it is not free. Knowing the costs helps you keep your program fast. 🐢

Every crossing into Python carries overhead, since Mojo must talk to the Python interpreter for each call.

A Python call inside a tight loop pays that tax every iteration. Pull such calls outside the loop whenever you can.

Instead of many small calls, send data once and let NumPy process it in bulk, paying the bridge cost a single time.

Moving large data across the bridge often copies it into Python memory. Big copies cost time and double your memory use.

Working on a PythonObject still runs through Python's dynamic machinery, far slower than native Mojo types.

For hot math, convert results into native Mojo types first, then run your fast loops and SIMD on them.

Reserve Python for setup and I/O. Your performance-critical kernels should run entirely in pure Mojo code.

Call import_module a single time and reuse the handle. Re-importing on every call wastes effort for no benefit.

Do not guess where time goes. Measure first, and you may find the bridge, not your math, is the real bottleneck.

Used wisely, interop gives you Python's tools without losing speed: cross the bridge rarely, then compute in Mojo. ⚡

Recall the best way to keep Python interop from slowing a hot loop.

You learned interop has overhead: cross the bridge rarely, batch your data, convert to native types, and keep hot paths in pure Mojo. 🎯