LoongArch64 FP16 hardware support

[heiher]

LoongArch is a RISC instruction set architecture and currently a Tier-2 (with host-tools) target ¹ in the Rust upstream community.

This patch introduces FP16 conversion functions based on the LoongArch SIMD extension to improve performance.

Benchmarks:

HalfFloatSliceExt::convert_from_f32_slice/constants
                        time:   [10.816 ns 10.823 ns 10.831 ns]
                        change: [-63.769% -63.728% -63.693%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_from_f32_slice/large
                        time:   [137.68 ns 137.77 ns 137.88 ns]
                        change: [-94.847% -94.841% -94.834%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_from_f64_slice/constants
                        time:   [12.656 ns 12.669 ns 12.684 ns]
                        change: [-78.455% -78.418% -78.367%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_from_f64_slice/large
                        time:   [544.15 ns 544.49 ns 544.91 ns]
                        change: [-89.799% -89.791% -89.781%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_to_f32_slice/constants
                        time:   [6.0412 ns 6.0442 ns 6.0482 ns]
                        change: [-74.100% -74.068% -74.042%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_to_f32_slice/large
                        time:   [512.78 ns 513.08 ns 513.45 ns]
                        change: [-77.628% -77.526% -77.422%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_to_f64_slice/constants
                        time:   [10.779 ns 10.784 ns 10.792 ns]
                        change: [-49.028% -48.922% -48.813%] (p = 0.00 < 0.05)
                        Performance has improved.

HalfFloatSliceExt::convert_to_f64_slice/large
                        time:   [923.19 ns 923.77 ns 924.50 ns]
                        change: [-80.876% -80.862% -80.849%] (p = 0.00 < 0.05)
                        Performance has improved.

Footnotes

1. https://doc.rust-lang.org/stable/rustc/platform-support/loongarch-linux.html ↩

[Amanieu]

@heiher This causes the crate to fail to build on stable. The usage of nightly features should be gated behind a Cargo feature.

[heiher]

@heiher This causes the crate to fail to build on stable. The usage of nightly features should be gated behind a Cargo feature.

Thanks! A new nightly feature has been added to gate the usage of the nightly-only features: #136