EgorBot for hamarb123 in #99596

[EgorBot]

Processing dotnet/runtime#99596 (comment) command:

Command

-intel -amd

using BenchmarkDotNet.Attributes;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Runtime.Intrinsics;

public class Benchmarks
{
    private byte[] source = null!;

    [Params(32, 64, 128, 256, 1024, 4096, 16384, 65536)]
    public int N;

    [GlobalSetup]
    public void Setup()
    {
        source = new byte[N];
        new Random(42).NextBytes(source);
    }

    [Benchmark]
    public float TestWork() => Work(ref MemoryMarshal.GetArrayDataReference(source), (uint)source.Length);

    [Benchmark]
    public float TestWorkWithZeroing() => WorkWithZeroing(ref MemoryMarshal.GetArrayDataReference(source), (uint)source.Length);

    //this test should be very latency bound as it has a long dependency chain (and it will even compile first try this time!)

    public static float Work(ref byte b, nuint size)
    {
        Vector256<float> result = Vector256<float>.Zero;
        while (size >= (uint)Vector256<byte>.Count)
        {
            Vector256<byte> v0 = Vector256.LoadUnsafe(ref b);
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)15, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            result += v0.AsSingle();
            b = ref Unsafe.Add(ref b, 1);
            size--;
        }
        return Vector256.Sum(result);
    }

    public static float WorkWithZeroing(ref byte b, nuint size)
    {
        Vector256<float> result = Vector256<float>.Zero;
        while (size >= (uint)Vector256<byte>.Count)
        {
            Vector256<byte> v0 = Vector256.LoadUnsafe(ref b);
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            v0 = (v0.AsSingle() + Vector256.Create(1.0f)).AsByte();
            v0 = Vector256.Shuffle(v0, Vector256.Create((byte)99, 0, 14, 1, 13, 2, 12, 3, 11, 4, 10, 5, 9, 6, 8, 7, 31, 16, 30, 17, 29, 18, 28, 19, 27, 20, 26, 21, 25, 22, 24, 23));
            result += v0.AsSingle();
            b = ref Unsafe.Add(ref b, 1);
            size--;
        }
        return Vector256.Sum(result);
    }
}

(EgorBot will reply in this issue)

[EgorBot]

Benchmark results on linux_azure_genoa

BenchmarkDotNet v0.14.0, Ubuntu 24.04.1 LTS (Noble Numbat)
AMD EPYC 9V74, 1 CPU, 8 logical and 4 physical cores
  Job-FVPUPO : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-JPALZN : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Method	Toolchain	N	Mean	Error	Ratio
TestWork	Main	32	4.028 ns	0.0005 ns	1.00
TestWork	PR	32	3.250 ns	0.0007 ns	0.81
TestWorkWithZeroing	Main	32	5.412 ns	0.0012 ns	1.00
TestWorkWithZeroing	PR	32	3.248 ns	0.0009 ns	0.60
TestWork	Main	64	83.408 ns	0.0115 ns	1.00
TestWork	PR	64	75.637 ns	0.0293 ns	0.91
TestWorkWithZeroing	Main	64	129.789 ns	0.1251 ns	1.00
TestWorkWithZeroing	PR	64	75.297 ns	0.0071 ns	0.58
TestWork	Main	128	242.562 ns	0.0254 ns	1.00
TestWork	PR	128	219.447 ns	0.0186 ns	0.90
TestWorkWithZeroing	Main	128	379.330 ns	0.4133 ns	1.00
TestWorkWithZeroing	PR	128	219.431 ns	0.0263 ns	0.58
TestWork	Main	256	561.700 ns	0.0269 ns	1.00
TestWork	PR	256	508.631 ns	0.0425 ns	0.91
TestWorkWithZeroing	Main	256	878.289 ns	0.7308 ns	1.00
TestWorkWithZeroing	PR	256	508.509 ns	0.0843 ns	0.58
TestWork	Main	1024	2,476.962 ns	0.3858 ns	1.00
TestWork	PR	1024	2,243.843 ns	0.1472 ns	0.91
TestWorkWithZeroing	Main	1024	3,875.836 ns	3.8726 ns	1.00
TestWorkWithZeroing	PR	1024	2,243.851 ns	0.3199 ns	0.58
TestWork	Main	4096	10,137.186 ns	0.8249 ns	1.00
TestWork	PR	4096	9,181.582 ns	0.8262 ns	0.91
TestWorkWithZeroing	Main	4096	15,882.453 ns	20.3580 ns	1.00
TestWorkWithZeroing	PR	4096	9,183.172 ns	2.6816 ns	0.58
TestWork	Main	16384	40,775.922 ns	2.0389 ns	1.00
TestWork	PR	16384	36,933.674 ns	2.0524 ns	0.91
TestWorkWithZeroing	Main	16384	63,865.811 ns	79.2406 ns	1.00
TestWorkWithZeroing	PR	16384	36,932.880 ns	1.8486 ns	0.58
TestWork	Main	65536	163,333.184 ns	11.3526 ns	1.00
TestWork	PR	65536	147,965.859 ns	13.2357 ns	0.91
TestWorkWithZeroing	Main	65536	255,547.157 ns	252.9114 ns	1.00
TestWorkWithZeroing	PR	65536	147,958.248 ns	55.0237 ns	0.58

BDN_Artifacts.zip

cc @hamarb123 (agent_logs.txt). EgorBot manual: link.

[EgorBot]

Benchmark results on linux_azure_cascadelake

BenchmarkDotNet v0.14.0, Ubuntu 24.04.1 LTS (Noble Numbat)
Intel Xeon Platinum 8370C CPU 2.80GHz, 1 CPU, 8 logical and 4 physical cores
  Job-HYSEPG : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-ULQOPH : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Method	Toolchain	N	Mean	Error	Ratio
TestWork	Main	32	4.997 ns	0.0009 ns	1.00
TestWork	PR	32	4.069 ns	0.0012 ns	0.81
TestWorkWithZeroing	Main	32	6.538 ns	0.0021 ns	1.00
TestWorkWithZeroing	PR	32	3.678 ns	0.0008 ns	0.56
TestWork	Main	64	82.628 ns	0.0268 ns	1.00
TestWork	PR	64	72.185 ns	0.0143 ns	0.87
TestWorkWithZeroing	Main	64	139.382 ns	0.0256 ns	1.00
TestWorkWithZeroing	PR	64	71.550 ns	0.0089 ns	0.51
TestWork	Main	128	243.990 ns	0.0723 ns	1.00
TestWork	PR	128	210.388 ns	0.0247 ns	0.86
TestWorkWithZeroing	Main	128	413.672 ns	0.1493 ns	1.00
TestWorkWithZeroing	PR	128	211.923 ns	0.1518 ns	0.51
TestWork	Main	256	569.155 ns	0.0916 ns	1.00
TestWork	PR	256	492.100 ns	0.2155 ns	0.86
TestWorkWithZeroing	Main	256	959.317 ns	0.1196 ns	1.00
TestWorkWithZeroing	PR	256	493.338 ns	0.1968 ns	0.51
TestWork	Main	1024	2,504.693 ns	1.2798 ns	1.00
TestWork	PR	1024	2,156.762 ns	0.3129 ns	0.86
TestWorkWithZeroing	Main	1024	4,220.248 ns	0.3756 ns	1.00
TestWorkWithZeroing	PR	1024	2,157.426 ns	0.2552 ns	0.51
TestWork	Main	4096	10,244.097 ns	3.1170 ns	1.00
TestWork	PR	4096	8,817.136 ns	1.8494 ns	0.86
TestWorkWithZeroing	Main	4096	17,293.932 ns	9.5948 ns	1.00
TestWorkWithZeroing	PR	4096	8,821.353 ns	0.8262 ns	0.51
TestWork	Main	16384	41,268.643 ns	18.7114 ns	1.00
TestWork	PR	16384	35,501.194 ns	16.2738 ns	0.86
TestWorkWithZeroing	Main	16384	69,479.384 ns	12.5593 ns	1.00
TestWorkWithZeroing	PR	16384	35,452.610 ns	5.6369 ns	0.51
TestWork	Main	65536	165,604.372 ns	234.1340 ns	1.00
TestWork	PR	65536	142,501.684 ns	323.1169 ns	0.86
TestWorkWithZeroing	Main	65536	278,419.241 ns	69.8628 ns	1.00
TestWorkWithZeroing	PR	65536	142,335.196 ns	103.7682 ns	0.51

BDN_Artifacts.zip

cc @hamarb123 (agent_logs.txt). EgorBot manual: link.