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(64, 128, 256, 1024, 4096, 16384, 65536, 1 << 20)]
    public int N;

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

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

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

    //test to see if the new codegen using shuf over perm is meaningfully faster
    //this should be more pipelineable than the other test I ran

    public static byte Work(ref byte b, nuint size)
    {
        Vector256<byte> result = Vector256<byte>.Zero;
        while (size >= (uint)Vector256<byte>.Count + 7u)
        {
            Vector256<byte> v0 = Vector256.LoadUnsafe(ref b);
            Vector256<byte> v1 = Vector256.LoadUnsafe(ref b, 1);
            Vector256<byte> v2 = Vector256.LoadUnsafe(ref b, 2);
            Vector256<byte> v3 = Vector256.LoadUnsafe(ref b, 3);
            Vector256<byte> v4 = Vector256.LoadUnsafe(ref b, 4);
            Vector256<byte> v5 = Vector256.LoadUnsafe(ref b, 5);
            Vector256<byte> v6 = Vector256.LoadUnsafe(ref b, 6);
            Vector256<byte> v7 = Vector256.LoadUnsafe(ref b, 7);
            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));
            v1 = Vector256.Shuffle(v1, 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));
            v2 = Vector256.Shuffle(v2, 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));
            v3 = Vector256.Shuffle(v3, 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));
            v4 = Vector256.Shuffle(v4, 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));
            v5 = Vector256.Shuffle(v5, 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));
            v6 = Vector256.Shuffle(v6, 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));
            v7 = Vector256.Shuffle(v7, 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 ^ v1) - (v2 + v3)) ^ ((v4 + ~v5) + (v6 + v7));
            b = ref Unsafe.Add(ref b, 8);
            size -= 8;
        }
        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));
            result += v0;
            b = ref Unsafe.Add(ref b, 1);
            size--;
        }
        return Vector256.Sum(result);
    }

    public static byte WorkWithZeroing(ref byte b, nuint size)
    {
        Vector256<byte> result = Vector256<byte>.Zero;
        while (size >= (uint)Vector256<byte>.Count + 7u)
        {
            Vector256<byte> v0 = Vector256.LoadUnsafe(ref b);
            Vector256<byte> v1 = Vector256.LoadUnsafe(ref b, 1);
            Vector256<byte> v2 = Vector256.LoadUnsafe(ref b, 2);
            Vector256<byte> v3 = Vector256.LoadUnsafe(ref b, 3);
            Vector256<byte> v4 = Vector256.LoadUnsafe(ref b, 4);
            Vector256<byte> v5 = Vector256.LoadUnsafe(ref b, 5);
            Vector256<byte> v6 = Vector256.LoadUnsafe(ref b, 6);
            Vector256<byte> v7 = Vector256.LoadUnsafe(ref b, 7);
            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));
            v1 = Vector256.Shuffle(v1, 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));
            v2 = Vector256.Shuffle(v2, 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));
            v3 = Vector256.Shuffle(v3, 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));
            v4 = Vector256.Shuffle(v4, 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));
            v5 = Vector256.Shuffle(v5, 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));
            v6 = Vector256.Shuffle(v6, 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));
            v7 = Vector256.Shuffle(v7, 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 ^ v1) - (v2 + v3)) ^ ((v4 + ~v5) + (v6 + v7));
            b = ref Unsafe.Add(ref b, 8);
            size -= 8;
        }
        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));
            result += v0;
            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-JRMVGK : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-BNRGWM : .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	64	8.060 ns	0.0717 ns	1.00
TestWork	PR	64	8.344 ns	0.0940 ns	1.04
TestWorkWithZeroing	Main	64	9.771 ns	0.0351 ns	1.00
TestWorkWithZeroing	PR	64	8.341 ns	0.1607 ns	0.85
TestWork	Main	128	21.339 ns	0.3073 ns	1.00
TestWork	PR	128	23.009 ns	0.2659 ns	1.08
TestWorkWithZeroing	Main	128	27.356 ns	0.5715 ns	1.00
TestWorkWithZeroing	PR	128	23.892 ns	0.5034 ns	0.87
TestWork	Main	256	49.933 ns	0.9419 ns	1.00
TestWork	PR	256	50.601 ns	1.0426 ns	1.01
TestWorkWithZeroing	Main	256	58.178 ns	1.1913 ns	1.00
TestWorkWithZeroing	PR	256	51.124 ns	1.0582 ns	0.88
TestWork	Main	1024	221.675 ns	3.4949 ns	1.00
TestWork	PR	1024	219.388 ns	3.9195 ns	0.99
TestWorkWithZeroing	Main	1024	252.837 ns	2.5360 ns	1.00
TestWorkWithZeroing	PR	1024	217.202 ns	3.7242 ns	0.86
TestWork	Main	4096	920.566 ns	14.9785 ns	1.00
TestWork	PR	4096	890.564 ns	11.4375 ns	0.97
TestWorkWithZeroing	Main	4096	1,038.222 ns	5.8478 ns	1.00
TestWorkWithZeroing	PR	4096	895.899 ns	9.8948 ns	0.86
TestWork	Main	16384	3,695.416 ns	8.3498 ns	1.00
TestWork	PR	16384	3,572.049 ns	5.1600 ns	0.97
TestWorkWithZeroing	Main	16384	4,174.213 ns	5.9289 ns	1.00
TestWorkWithZeroing	PR	16384	3,578.050 ns	6.2754 ns	0.86
TestWork	Main	65536	14,818.742 ns	15.8500 ns	1.00
TestWork	PR	65536	14,359.367 ns	16.1748 ns	0.97
TestWorkWithZeroing	Main	65536	16,721.886 ns	21.0476 ns	1.00
TestWorkWithZeroing	PR	65536	14,335.655 ns	10.1123 ns	0.86
TestWork	Main	1048576	236,854.538 ns	262.3565 ns	1.00
TestWork	PR	1048576	229,811.150 ns	249.4602 ns	0.97
TestWorkWithZeroing	Main	1048576	267,728.676 ns	140.5791 ns	1.00
TestWorkWithZeroing	PR	1048576	229,378.676 ns	220.4589 ns	0.86

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-POPUCL : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-WUZLBP : .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	64	10.477 ns	0.0021 ns	1.00
TestWork	PR	64	10.255 ns	0.0049 ns	0.98
TestWorkWithZeroing	Main	64	12.716 ns	0.0047 ns	1.00
TestWorkWithZeroing	PR	64	9.887 ns	0.0300 ns	0.78
TestWork	Main	128	28.710 ns	0.0111 ns	1.00
TestWork	PR	128	25.435 ns	0.1354 ns	0.89
TestWorkWithZeroing	Main	128	33.987 ns	0.0108 ns	1.00
TestWorkWithZeroing	PR	128	26.736 ns	0.0253 ns	0.79
TestWork	Main	256	64.977 ns	0.0135 ns	1.00
TestWork	PR	256	56.324 ns	0.0955 ns	0.87
TestWorkWithZeroing	Main	256	76.767 ns	0.0187 ns	1.00
TestWorkWithZeroing	PR	256	57.097 ns	0.1499 ns	0.74
TestWork	Main	1024	285.519 ns	0.0380 ns	1.00
TestWork	PR	1024	238.199 ns	0.5961 ns	0.83
TestWorkWithZeroing	Main	1024	332.434 ns	0.0453 ns	1.00
TestWorkWithZeroing	PR	1024	241.675 ns	2.0290 ns	0.73
TestWork	Main	4096	1,187.669 ns	0.0767 ns	1.00
TestWork	PR	4096	1,026.536 ns	20.0993 ns	0.86
TestWorkWithZeroing	Main	4096	1,376.965 ns	0.5258 ns	1.00
TestWorkWithZeroing	PR	4096	1,024.971 ns	20.4914 ns	0.74
TestWork	Main	16384	4,758.117 ns	0.5854 ns	1.00
TestWork	PR	16384	4,179.029 ns	2.6244 ns	0.88
TestWorkWithZeroing	Main	16384	5,519.795 ns	1.7770 ns	1.00
TestWorkWithZeroing	PR	16384	4,177.967 ns	2.0834 ns	0.76
TestWork	Main	65536	19,054.506 ns	1.2671 ns	1.00
TestWork	PR	65536	16,550.536 ns	13.1765 ns	0.87
TestWorkWithZeroing	Main	65536	22,109.951 ns	9.0549 ns	1.00
TestWorkWithZeroing	PR	65536	16,448.069 ns	15.1416 ns	0.74
TestWork	Main	1048576	305,015.070 ns	61.3536 ns	1.00
TestWork	PR	1048576	252,464.712 ns	75.2583 ns	0.83
TestWorkWithZeroing	Main	1048576	354,147.083 ns	175.3066 ns	1.00
TestWorkWithZeroing	PR	1048576	252,433.582 ns	124.9834 ns	0.71

BDN_Artifacts.zip

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