annulusgames / burstlinq

Extremely fast LINQ aggregation operations implementation optimized by Burst Compiler

Stats

burstlinq's Issues

It's slow on IL2CPP build. isn't it?

[Mono Build] is better than Non BurstCompile

but [IL2cpp Build] is not fast than BurstCompile

TestCode

[BurstCompile(OptimizeFor = OptimizeFor.Performance)]
public unsafe static class BurstLinq
{
    public static int Sum(List<int> source)
    {
        var span = SpanHelper.AsSpan(source);
        fixed (int* ptr = span)
        {
            SumCore(ptr, source.Count, out var result);
            return result;
        }

    }

    [BurstCompile(FloatMode = FloatMode.Fast)]
    internal static void SumCore(int* ptr, [AssumeRange(1, int.MaxValue)] int length, out int result)
    {
        var sum = default(int);
        for (var i = 0; i < length; i++) sum += ptr[i];
        result = sum;
    }
}

public unsafe static class NoBurstLinq
{
    public static int Sum(List<int> source)
    {
        var span = SpanHelper.AsSpan(source);
        fixed (int* ptr = span)
        {
            SumCore(ptr, source.Count, out var result);
            return result;
        }

    }

    internal static void SumCore(int* ptr, [AssumeRange(1, int.MaxValue)] int length, out int result)
    {
        var sum = default(int);
        for (var i = 0; i < length; i++) sum += ptr[i];
        result = sum;
    }
}

public class TestBehaviour : MonoBehaviour
{
    private List<int> ints;
	private int = 1000000;

    void Start()
    {
        ints = Enumerable.Range(0,1000).ToList();
    }

    void Update()
    {
        var localValue = 0;

        Profiler.BeginSample("Linq");
        for (var i = 0; i < length; i++)
        {
            localValue += ints.Sum();
        }
        Profiler.EndSample();
        Profiler.BeginSample("NoBurstLinq");
        for (var i = 0; i < length; i++)
        {
            localValue += NoBurstLinq.Sum(ints);
        }
        Profiler.EndSample();
        Profiler.BeginSample("BurstLinq");
        for (var i = 0; i < length; i++)
        {
            localValue += BurstLinq.Sum(ints);
        }
        Profiler.EndSample();
    }
}

min maxの仕様について

armとx86でsimd minmaxの仕様が違うようで、armはnanを考慮するけど、x86は考慮しないみたいです。
armmax(a, b)=>double.IsNan(b)||a>b?a:b;
x86max(a, b)=>a>b?a:b;
みたいな感じです。
armはarm maxじゃないとsimd効かず遅くなる、x86はarmmaxにすると演算が2つ増えるけどsimdは効くからちょっと遅くなるみたいです。
simdlinqではx86前提なので、x86maxです。
仕様はどうしますか？

error CS1503: Argument 1: cannot convert from 'void*' to 'int*'

Mulitple lines error in following files:
Library\PackageCache\com.annulusgames.burst-linq@bd8729e\Runtime\BurstLinqExtensions.Native.Average.cs
Library\PackageCache\com.annulusgames.burst-linq@bd8729e\Runtime\BurstLinqExtensions.Native.Contains.cs
Library\PackageCache\com.annulusgames.burst-linq@bd8729e\Runtime\BurstLinqExtensions.Native.Max.cs
Library\PackageCache\com.annulusgames.burst-linq@bd8729e\Runtime\BurstLinqExtensions.Native.Min.cs
Library\PackageCache\com.annulusgames.burst-linq@bd8729e\Runtime\BurstLinqExtensions.Native.SequenceEqual.cs

Unity version 2023.2.13.f1
Burst: 1.8.12