ARM64 snippets for C#

ARM64.cs

using System.Runtime.CompilerServices;
using System.Runtime.Intrinsics.Arm;

namespace System.Runtime.Instrinsics;

public static class VectorExtensions
{
    [MethodImpl(MethodImplOptions.AggressiveInlining)]
    internal static int CountMatches<T>(this Vector256<T> mask)
    {
        if (Vector256.IsHardwareAccelerated)
        {
            return BitOperations.PopCount(mask.ExtractMostSignificantBits());
        }

        var (lower, upper) = mask;
        var lowerCount = CountMatches(lower);
        var upperCount = CountMatches(upper);

        return upperCount + lowerCount;
    }

    [MethodImpl(MethodImplOptions.AggressiveInlining)]
    internal static int CountMatches<T>(this Vector128<T> mask)
    {
        if (AdvSimd.Arm64.IsSupported)
        {
            return AdvSimd.Arm64
                .AddAcross(AdvSimd.PopCount(mask.AsByte()))
                .ToScalar() / (8 * Unsafe.SizeOf<T>());
        }

        return BitOperations.PopCount(mask.ExtractMostSignificantBits());
    }

    [MethodImpl(MethodImplOptions.AggressiveInlining)]
    internal static int CountMatches<T>(this Vector64<T> mask)
    {
        return AdvSimd.Arm64
            .AddAcross(AdvSimd.PopCount(mask.AsByte()))
            .ToScalar() / (8 * Unsafe.SizeOf<T>());
    }
}