Performance

Measure the function, input size, dtype, and build target used by your application before making performance decisions.

Keep tensor arithmetic on tensors

For elementwise vector calculations, use Chelis tensor operations where they express the computation. A to_list / map / to_tensor round trip moves values through host lists and can cost much more than a native tensor operation in generated C. Some Nautilus functions use lists internally, so measure the complete function as well as the surrounding expression.

Compare like with like

• Check numerical error against a suitable reference before comparing time. A faster result with a different tolerance is a different calculation.
• Keep dtype, input shape, number of threads, and compiler flags visible in the measurement. Most Nautilus modules compute in f32; a direct comparison with an f64 library also compares precision.
• Time the generated C path for a compiled workload. CLI startup and calls from a host-language test harness can dominate short calculations.
• Measure both small and representative large inputs. The fastest choice can change when dispatch overhead stops dominating the work.

Results can vary with the compiler, hardware, operating system, and thread configuration, so measurements from another machine are not a performance promise.