basic/

basic.rs

#[cfg(all(feature = "x86-v4", target_arch = "x86_64"))]
mod x86 {
	use aligned_vec::avec;
	use diol::prelude::*;

	use core::arch::x86_64::*;
	use core::mem::transmute;
	use pulp::Simd;
	use pulp::x86::V4;

	fn sum_scalar(bencher: Bencher, len: usize) {
		let v = &*avec![0.0f64; len];

		bencher.bench(|| black_box(v.iter().sum::<f64>()));
	}

	#[target_feature(enable = "avx512f")]
	#[allow(clippy::missing_transmute_annotations)]
	unsafe fn sum_stdarch_imp(v: &[f64]) -> f64 {
		let mut acc0 = _mm512_set1_pd(0.0);
		let mut acc1 = _mm512_set1_pd(0.0);
		let mut acc2 = _mm512_set1_pd(0.0);
		let mut acc3 = _mm512_set1_pd(0.0);

		// 512 = 64 * 8
		let (head, tail) = pulp::as_arrays::<8, _>(v);
		let (head4, head1) = pulp::as_arrays::<4, _>(head);

		for [x0, x1, x2, x3] in head4 {
			let x0 = transmute(*x0);
			let x1 = transmute(*x1);
			let x2 = transmute(*x2);
			let x3 = transmute(*x3);

			acc0 = _mm512_add_pd(acc0, x0);
			acc1 = _mm512_add_pd(acc1, x1);
			acc2 = _mm512_add_pd(acc2, x2);
			acc3 = _mm512_add_pd(acc3, x3);
		}

		for x0 in head1 {
			let x0 = pulp::cast(*x0);
			acc0 = _mm512_add_pd(acc0, x0);
		}

		acc0 = _mm512_add_pd(acc0, acc1);
		acc2 = _mm512_add_pd(acc2, acc3);
		acc0 = _mm512_add_pd(acc0, acc2);

		let acc: [__m256d; 2] = pulp::cast(acc0);
		let acc = _mm256_add_pd(acc[0], acc[1]);

		let acc: [__m128d; 2] = pulp::cast(acc);
		let acc = _mm_add_pd(acc[0], acc[1]);

		let acc: [f64; 2] = pulp::cast(acc);
		let mut acc = acc[0] + acc[1];

		for x0 in tail {
			acc += *x0;
		}
		acc
	}

	fn sum_stdarch(bencher: Bencher, len: usize) {
		let v = &*avec![0.0f64; len];

		bencher.bench(|| unsafe { black_box(sum_stdarch_imp(v)) });
	}

	fn sum_pulp(bencher: Bencher, len: usize) {
		if let Some(simd) = V4::try_new() {
			let v = &*avec![0.0f64; len];

			bencher.bench(|| {
				struct Imp<'a> {
					simd: V4,
					v: &'a [f64],
				}

				impl pulp::NullaryFnOnce for Imp<'_> {
					type Output = f64;

					#[inline(always)]
					fn call(self) -> Self::Output {
						let Self { simd, v } = self;

						let (head, tail) = pulp::as_arrays::<8, _>(v);

						let mut acc0 = simd.splat_f64x8(0.0);
						let mut acc1 = simd.splat_f64x8(0.0);
						let mut acc2 = simd.splat_f64x8(0.0);
						let mut acc3 = simd.splat_f64x8(0.0);

						let (head4, head1) = pulp::as_arrays::<4, _>(head);

						for [x0, x1, x2, x3] in head4 {
							let x0 = pulp::cast(*x0);
							let x1 = pulp::cast(*x1);
							let x2 = pulp::cast(*x2);
							let x3 = pulp::cast(*x3);

							acc0 = pulp::cast(
								simd.avx512f._mm512_add_pd(pulp::cast(acc0), pulp::cast(x0)),
							);

							acc0 = simd.add_f64x8(acc0, x0);
							acc1 = simd.add_f64x8(acc1, x1);
							acc2 = simd.add_f64x8(acc2, x2);
							acc3 = simd.add_f64x8(acc3, x3);
						}
						for x0 in head1 {
							let x0 = pulp::cast(*x0);
							acc0 = simd.add_f64x8(acc0, x0);
						}

						acc0 = simd.add_f64x8(acc0, acc1);
						acc2 = simd.add_f64x8(acc2, acc3);
						acc0 = simd.add_f64x8(acc0, acc2);

						let tail = simd.partial_load_f64s(tail);

						simd.reduce_sum_f64s(simd.add_f64x8(acc0, tail))
					}
				}

				simd.vectorize(Imp { simd, v })
			});
		}
	}

	fn sum_pulp_dispatch(bencher: Bencher, len: usize) {
		let v = &*avec![0.0f64; len];

		bencher.bench(|| {
			struct Imp<'a> {
				v: &'a [f64],
			}

			impl pulp::WithSimd for Imp<'_> {
				type Output = f64;

				#[inline(always)]
				fn with_simd<S: Simd>(self, simd: S) -> Self::Output {
					let Self { v } = self;

					let (head, tail) = S::as_simd_f64s(v);

					let mut acc0 = simd.splat_f64s(0.0);
					let mut acc1 = simd.splat_f64s(0.0);
					let mut acc2 = simd.splat_f64s(0.0);
					let mut acc3 = simd.splat_f64s(0.0);

					let (head4, head1) = pulp::as_arrays::<4, _>(head);

					for &[x0, x1, x2, x3] in head4 {
						acc0 = simd.add_f64s(acc0, x0);
						acc1 = simd.add_f64s(acc1, x1);
						acc2 = simd.add_f64s(acc2, x2);
						acc3 = simd.add_f64s(acc3, x3);
					}

					for &x0 in head1 {
						acc0 = simd.add_f64s(acc0, x0);
					}

					acc0 = simd.add_f64s(acc0, acc1);
					acc2 = simd.add_f64s(acc2, acc3);
					acc0 = simd.add_f64s(acc0, acc2);

					simd.reduce_sum_f64s(simd.add_f64s(acc0, simd.partial_load_f64s(tail)))
				}
			}

			pulp::Arch::new().dispatch(Imp { v })
		});
	}

	pub fn main() -> std::io::Result<()> {
		let mut bench = Bench::new(BenchConfig::from_args()?);

		bench.register_many(
			list![sum_scalar, sum_stdarch, sum_pulp, sum_pulp_dispatch],
			[64, 256, 1024],
		);
		bench.run()?;
		Ok(())
	}
}

fn main() -> std::io::Result<()> {
	#[cfg(all(feature = "x86-v4", target_arch = "x86_64"))]
	x86::main()?;
	Ok(())
}