Compiler Explorer

Source code

using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;

public static class C {
    internal static Vector128<byte> NarrowNative(Vector128<ushort> lower, Vector128<ushort> upper)
    {
        return Sse2.PackUnsignedSaturate(lower.AsInt16(), upper.AsInt16());
    }
    
    internal static Vector128<byte> Narrow(Vector128<ushort> lower, Vector128<ushort> upper)
    {
        return Vector128.Narrow(lower, upper);
    }
    
    internal static Vector256<byte> NarrowNative(Vector256<ushort> lower, Vector256<ushort> upper)
    {
        const byte control = 0b_11_01_10_00;
        return Avx2.Permute4x64(Avx2.PackUnsignedSaturate(lower.AsInt16(), upper.AsInt16()).AsInt64(), control).AsByte();
    }

internal static Vector256<byte> Narrow(Vector256<ushort> lower, Vector256<ushort> upper)
    {
        return Vector256.Narrow(lower, upper);
    }
    
    internal static Vector512<byte> NarrowNative(Vector512<ushort> lower, Vector512<ushort> upper)
    {
        Vector512<long> control = Vector512.Create(0, 2, 4, 6, 1, 3, 5, 7);
        return Avx512F.PermuteVar8x64(Avx512BW.PackUnsignedSaturate(lower.AsInt16(), upper.AsInt16()).AsInt64(), control).AsByte();
    }
    
    internal static Vector512<byte> Narrow(Vector512<ushort> lower, Vector512<ushort> upper)
    {
        return Vector512.Narrow(lower, upper);
    }
}