bitshifter / glam-rs / 35218159043

99%

push

github

perf(simd): broadcast dot product without an extra shuffle (#833)

The SSE2 broadcast dot product can save a shuffle by computing the
result directly in all four lanes.

## Solution

`dot4_into_m128` now sums the products directly in all four lanes,
keeping the `(x + z) + (y + w)` grouping. Tests and benchmarks are
included.

## Benchmarks

On my Xeon E5-2696 v4:

| Measurement | Before | After | Change |
|---|---:|---:|---:|
| 256-dot-product batch | 301.673 ns | 267.888 ns | -11.2% |
| Throughput | 848.60 M/s | 955.62 M/s | +12.6% |
| Quaternion lerp | 6.491 ns | 6.095 ns | -6.1% |
| Cycles/dot | 3.772 | 3.335 | -11.6% |
| Instructions/dot | 12.104 | 11.104 | -8.3% |
| Allocations | 0 | 0 | Unchanged |

Timing: Criterion on Windows. Counters: five alternating, core-pinned
runs under WSL. Normalization timing stayed within the noise threshold
(+0.7%).

## Testing and linting

- Three million comparisons passed: non-NaN outputs matched bit for bit;
NaNs matched by classification.
- Regression tests cover cancellation, signed zero, subnormals,
overflow, infinities, and NaNs.
- All nine feature configurations, MSRV, and portable SIMD checks
passed.
- Glam-specific Clippy and formatting passed.
- The existing Windows CI-tool lint remains.

🤖 AI disclosure: I used Codex with GPT-6 Astra at xhigh reasoning effort
to help identify the optimization and construct the benchmark. I have
reviewed and understand the resulting implementation.

8 of 8 new or added lines in 2 files covered. (100.0%)

93262 of 94149 relevant lines covered (99.06%)

123554.19 hits per line