bitshifter / glam-rs / 36314158559

99%

push

github

perf(mat2): speed up NEON Mat2 determinant and inverse (#878)

Optimises the NEON implementations of `Mat2::determinant` and
`Mat2::inverse` (and therefore `Affine2::inverse`) discussed in #830.

- `determinant`: use the scalar body, which reaches parity with
`scalar-math`.
- `inverse_checked`: compute the determinant with lane multiplies and
use a scalar reciprocal instead of a 4-wide vector divide.

### Benchmarks (Raspberry Pi 4, Cortex-A72)

Criterion, 3 interleaved runs per config, medians. `speedup =
scalar_median / neon_median`; `1.00` is parity.

| bench | before | after |
| --- | --- | --- |
| mat2 determinant/16 | 0.36 | 1.00 |
| mat2 determinant/1024 | 0.30 | 1.00 |
| mat2 inverse/16 | 0.44 | 0.65 |
| mat2 inverse/1024 | 0.43 | 0.65 |
| affine2 inverse/16 | 0.73 | 0.96 |
| affine2 inverse/1024 | 0.72 | 0.92 |

NEON-vs-scalar geomean over these benches: 0.47 -> 0.85.

Refs #830.

9 of 9 new or added lines in 1 file covered. (100.0%)

93360 of 94290 relevant lines covered (99.01%)

123338.85 hits per line