bitshifter / glam-rs / 35342051359

99%

Build Type

push

github

Committed by web-flow

Commit Message

perf: use fma whenever the target supports it (#853)

# Objective

- Fixes #736.

The `fast-math` feature only gated one thing: FMA in the SSE2 back-end,
and only when the `fma` target feature was also enabled. Meanwhile the
NEON back-end has always used fused multiply-add unconditionally, and
the explicit `mul_add` API already uses FMA when the target supports it.
This was inconsistent.

## Solution

- FMA is now used in the SSE2 implementation whenever the target
supports it, instead of being gated by `fast-math`. This makes SSE2
consistent with NEON, which has always used fused multiply-add. FMA has
a single rounding step so the affected operations are at least as
accurate as before.
- The `fast-math` feature is deprecated and is now a no-op. It is kept
so existing manifests continue to build, and will be removed in a future
breaking release.
- NEON `vmla`/`vmls` intrinsics are switched to the explicit
`vfma`/`vfms` forms.
- Additional internal multiply-adds are fused: `Vec3A::cross`,
`Quat::slerp` and `Quat::mul_quat`.
- Added a CI job that tests with `-C target-feature=+fma`.

Measured on x86_64 with `cargo bench --bench quat --bench vec3a --
--quick` (`fma` vs baseline without FMA):

| bench | 16 | 1024 |
|---|---|---|
| `vec3a cross` | −3.0% | −1.8% |
| `vec3a slerp` | −10.6% | −11.2% |
| `quat slerp` | −10.7% | −10.9% |
| `quat mul quat` | −15.3% | −25.3% |

Coverage Stats

14 of 15 new or added lines in 6 files covered. (93.33%)

36 existing lines in 5 files now uncovered.

93232 of 94143 relevant lines covered (99.03%)

123530.13 hits per line