Matrix3 multiplication is 4x slower in 0.18

[KeyboardDanni]

AMD Ryzen 7 3700X Windows 10 19044.1566 rustc 1.60.0-nightly (1e12aef3f 2022-02-13)

The other day I was testing sprite performance in my engine and noticed it was a lot lower than usual. Investigated the cause and found it was when I updated cgmath from 0.17 to 0.18.

With 0.17, my benchmark that does 100k translations, rotations, and scales on a Matrix3:

test bench_transform_matrix ... bench: 670,110 ns/iter (+/- 13,076)

With 0.18:

test bench_transform_matrix ... bench: 2,755,590 ns/iter (+/- 10,904)

The bench itself Code for the transform wrapper

Since the transform happens for each sprite, the performance difference adds up quickly.

Using default features with -O2/-O3. Compiler target is x86_64-pc-windows-msvc.

[kvark]

Wow that's quite concerning! Thank you for bringing this up.

[peppidesu]

comparing the benchmarks provided by the crate shows a similar (3x) increase:

cargo bench --feature rand _bench_matrix3_mul_m    (0.18.0)
>> test _bench_matrix3_mul_m       ... bench:          13 ns/iter (+/- 0)
cargo bench _bench_matrix3_mul_m                   (0.17.0)
>> test _bench_matrix3_mul_m       ... bench:          4 ns/iter (+/- 0)

There was no significant increase for any of the other benchmarks regarding Matrix3

[peppidesu]

It seems the problem is already fixed on the master branch, getting 4ns results there.

[peppidesu]

Found the issue:

https://github.com/rustgd/cgmath/blob/637c566cc2141203d8d99c03e7ab770796c44f5f/src/vector.rs#L311-L337

This is on v0.18.0. The addition of default_fn! prevents #[inline] from working, which is causing the slowdown. This was fixed in #548 by moving the #[inline] into the macro.

[kvark]

Great, thank you for investigation!