Performance Regression in Rust Module with PyO3 in Python 3.13.2

[Smart-SangGe]

I've recently developed a computational module in Rust using PyO3 and observed significant performance differences between Python 3.12.9 and 3.13.2. I'm sharing my findings to understand if others have experienced similar issues or if there's something wrong with my implementation.

Environment
Python 3.12.9: [GCC 14.2.0] (February 5, 2025)
Python 3.13.2: [GCC 14.2.0] (February 5, 2025)

Benchmarking Results
Python 3.12.9
rust multiply execution time: 0.000855 seconds
ecc_py multiply execution time: 0.001347 seconds
rust add execution time: 0.000415 seconds
ecc_py add execution time: 0.000520 seconds

Python 3.13.2
rust multiply execution time: 0.009020 seconds
ecc_py multiply execution time: 0.001272 seconds
rust add execution time: 0.001548 seconds
ecc_py add execution time: 0.000463 seconds

Observations
In Python 3.12.9, the Rust implementation is faster for both operations:

For multiplication: Rust is ~1.6x faster than the Python implementation
For addition: Rust is ~1.3x faster than the Python implementation

However, in Python 3.13.2, this advantage is completely lost:

For multiplication: Rust is ~7x slower than the Python implementation
For addition: Rust is ~3.3x slower than the Python implementation

This seems to indicate a significant regression in PyO3 performance with Python 3.13.2. Has anyone else encountered similar issues? Any suggestions for optimizing the Rust/PyO3 code for Python 3.13.2?

[birkenfeld]

Important data points we'd need to know: what PyO3 version(s) are you using, what Rust compiler version, and are you using the freethreaded build of Python 3.13?

[Smart-SangGe]

I'm using
pyo3 0.23.5
rustc 1.85.0 (4d91de4e4 2025-02-17)
python3.12/unstable,kali-rolling,now 3.12.9-1 amd64
python3.13/unstable,kali-rolling,now 3.13.2-1 amd64 (seems not the no-gil version)

[ngoldbaum]

It'd probably be useful to try profiling both, with e.g. cargo flamegraph to identify why the newer version is slower.

[ngoldbaum]

It might also be that CPython is faster in 3.13...

Also if you can share a reproducible example that demonstrates the issue you're having that would be a great bug report.

[Smart-SangGe]

I've learned that the rust-analyzer hints were showing pyo3 version 23.5, but I was actually compiling with 0.23.2 all along. After running cargo update to upgrade to version 23.5, the performance issues I was experiencing have been resolved. I had been confused about the version discrepancy for a while, not realizing that dependencies specified with wildcards in Cargo.toml don't automatically update without an explicit cargo update command.