Optimize Enumerable Min/Max final reduction with shuffles

[EgorBo]

Note

This PR was filed by AI (GitHub Copilot CLI).

Inspired by a similar helper in Tensors. I guess we don't want to consume Tensors in Linq due to IL size concerns. It also doesn't look beneficial to share this function via source file due to different generics being used. Maybe Vector128 will get this as public API at some point.

Benchmark

[Params(16, 64)]
public int Length;

private byte[]   _bytes;
private sbyte[]  _sbytes;
private short[]  _shorts;
private ushort[] _ushorts;
private int[]    _ints;
private uint[]   _uints;
private long[]   _longs;
private ulong[]  _ulongs;

[Benchmark] public byte   MaxByte()   => _bytes.Max();
[Benchmark] public sbyte  MaxSByte()  => _sbytes.Max();
[Benchmark] public short  MaxShort()  => _shorts.Max();
[Benchmark] public ushort MaxUShort() => _ushorts.Max();
[Benchmark] public int    MaxInt()    => _ints.Max();
[Benchmark] public uint   MaxUInt()   => _uints.Max();
[Benchmark] public long   MaxLong()   => _longs.Max();
[Benchmark] public ulong  MaxULong()  => _ulongs.Max();

Full benchmark request and raw output: EgorBot/Benchmarks#198

Results

Speed-up = main time / PR time (higher is better). Numbers below are from EgorBot.

ARM64 — Neoverse-N2 (ubuntu24_azure_cobalt100)

Method	Length	main	PR	Speed-up
MaxByte	16	51.54 ns	1.85 ns	27.85×
MaxSByte	16	49.58 ns	1.90 ns	26.16×
MaxShort	16	23.66 ns	0.98 ns	24.16×
MaxUShort	16	21.82 ns	1.22 ns	17.93×
MaxInt	16	1.72 ns	1.84 ns	0.93×
MaxUInt	16	1.93 ns	1.74 ns	1.11×
MaxLong	16	3.77 ns	3.91 ns	0.96×
MaxULong	16	3.81 ns	3.93 ns	0.97×
MaxByte	64	49.64 ns	2.03 ns	24.51×
MaxSByte	64	48.76 ns	2.04 ns	23.95×
MaxShort	64	22.35 ns	3.53 ns	6.33×
MaxUShort	64	22.33 ns	3.36 ns	6.64×
MaxInt	64	6.33 ns	6.21 ns	1.02×
MaxUInt	64	6.17 ns	6.51 ns	0.95×
MaxLong	64	23.55 ns	23.51 ns	1.00×
MaxULong	64	23.36 ns	23.44 ns	1.00×

Intel — Emerald Rapids / AVX-512 (ubuntu24_azure_emeraldrapids)

Method	Length	main	PR	Speed-up
MaxByte	16	8.88 ns	0.93 ns	9.54×
MaxSByte	16	9.86 ns	0.89 ns	11.10×
MaxShort	16	4.42 ns	1.55 ns	2.87×
MaxUShort	16	4.45 ns	1.02 ns	4.38×
MaxInt	16	1.66 ns	0.88 ns	1.89×
MaxUInt	16	1.81 ns	1.50 ns	1.20×
MaxLong	16	0.96 ns	0.89 ns	1.08×
MaxULong	16	1.00 ns	0.94 ns	1.06×
MaxByte	64	16.31 ns	1.44 ns	11.39×
MaxSByte	64	10.03 ns	1.59 ns	6.33×
MaxShort	64	6.52 ns	1.37 ns	4.77×
MaxUShort	64	6.54 ns	1.54 ns	4.26×
MaxInt	64	2.44 ns	1.68 ns	1.47×
MaxUInt	64	2.84 ns	1.94 ns	1.47×
MaxLong	64	2.72 ns	3.17 ns	0.86×
MaxULong	64	3.17 ns	2.97 ns	1.07×

AMD — EPYC 9V45 (Turin) / AVX-512 (ubuntu24_azure_turin)

Method	Length	main	PR	Speed-up
MaxByte	16	4.41 ns	0.70 ns	6.27×
MaxSByte	16	4.96 ns	0.83 ns	5.96×
MaxShort	16	3.01 ns	0.87 ns	3.45×
MaxUShort	16	2.09 ns	0.89 ns	2.35×
MaxInt	16	1.48 ns	0.98 ns	1.52×
MaxUInt	16	1.64 ns	0.98 ns	1.68×
MaxLong	16	0.94 ns	0.91 ns	1.03×
MaxULong	16	1.05 ns	0.91 ns	1.16×
MaxByte	64	5.86 ns	1.27 ns	4.61×
MaxSByte	64	7.57 ns	1.28 ns	5.93×
MaxShort	64	2.78 ns	1.39 ns	2.01×
MaxUShort	64	2.96 ns	1.13 ns	2.61×
MaxInt	64	2.27 ns	1.61 ns	1.41×
MaxUInt	64	1.97 ns	1.62 ns	1.22×
MaxLong	64	3.02 ns	3.00 ns	1.01×
MaxULong	64	2.87 ns	2.10 ns	1.37×

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-linq
See info in area-owners.md if you want to be subscribed.

[EgorBo]

@MihuBot

[EgorBo]

Note

Benchmark generated by AI (GitHub Copilot CLI).

@EgorBot -linux_amd -linux_intel -linux_arm64

using System.Linq;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(Bench).Assembly).Run(args);

public class Bench
{
    [Params(16, 64)]
    public int Length;

    private byte[]   _bytes   = default!;
    private sbyte[]  _sbytes  = default!;
    private short[]  _shorts  = default!;
    private ushort[] _ushorts = default!;
    private int[]    _ints    = default!;
    private uint[]   _uints   = default!;
    private long[]   _longs   = default!;
    private ulong[]  _ulongs  = default!;

    [GlobalSetup]
    public void Setup()
    {
        var rng = new Random(42);

        _bytes = new byte[Length];
        rng.NextBytes(_bytes);

        _sbytes = new sbyte[Length];
        for (int i = 0; i < Length; i++) _sbytes[i] = (sbyte)rng.Next(sbyte.MinValue, sbyte.MaxValue + 1);

        _shorts = new short[Length];
        for (int i = 0; i < Length; i++) _shorts[i] = (short)rng.Next(short.MinValue, short.MaxValue + 1);

        _ushorts = new ushort[Length];
        for (int i = 0; i < Length; i++) _ushorts[i] = (ushort)rng.Next(0, ushort.MaxValue + 1);

        _ints = new int[Length];
        for (int i = 0; i < Length; i++) _ints[i] = rng.Next();

        _uints = new uint[Length];
        for (int i = 0; i < Length; i++) _uints[i] = (uint)rng.Next();

        _longs = new long[Length];
        for (int i = 0; i < Length; i++) _longs[i] = rng.NextInt64();

        _ulongs = new ulong[Length];
        for (int i = 0; i < Length; i++) _ulongs[i] = (ulong)rng.NextInt64();
    }

    [Benchmark] public byte   MaxByte()   => _bytes.Max();
    [Benchmark] public sbyte  MaxSByte()  => _sbytes.Max();
    [Benchmark] public short  MaxShort()  => _shorts.Max();
    [Benchmark] public ushort MaxUShort() => _ushorts.Max();
    [Benchmark] public int    MaxInt()    => _ints.Max();
    [Benchmark] public uint   MaxUInt()   => _uints.Max();
    [Benchmark] public long   MaxLong()   => _longs.Max();
    [Benchmark] public ulong  MaxULong()  => _ulongs.Max();
}

[Copilot]

Pull request overview

This PR changes the final reduction step in System.Linq’s vectorized integer Min/Max implementation to reduce a Vector128<T> accumulator down to a single scalar using shuffle-based horizontal reduction instead of a scalar per-lane loop.

Changes:

• Replaced the per-element for loop reduction of best128 with a helper (HorizontalMinMax) that performs log2(lane-count) reductions using Vector128.Shuffle.
• Added HorizontalMinMax<T, TMinMax> to centralize the reduction logic for different element sizes (byte/short/int/long lane counts).

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.