FilterCmp_Avx_Next.cs

public unsafe static int FilterCmp_Avx_Next(Span<long> items)
{
    var len = items.Length;
    if (len <= 0)
        return 0;

    ref var permuteStart = ref Unsafe.AsRef(PermuteTable[0]);
    int outputIdx = 0;
    int i = 0;
    ref var output = ref items[i];
    if (i + Vector256<long>.Count <= len)
    {
        var next = Vector256.LoadUnsafe(ref Unsafe.Add(ref output, i));
        len -= Vector256<long>.Count;
        for (; i + Vector256<long>.Count <= len; i += Vector256<long>.Count)
        {
            var v = next;
            next = Vector256.LoadUnsafe(ref Unsafe.Add(ref output, i + Vector256<long>.Count));
            var bits = v.ExtractMostSignificantBits();
            if (bits == 0) // do we have _any_ negatives here?
            {
                v.StoreUnsafe(ref Unsafe.Add(ref output, outputIdx));
                outputIdx += Vector256<long>.Count;
                continue;
            }

            // complex case, we have to deal with some negatives
            var permute = Vector256.LoadUnsafe(ref Unsafe.Add(ref permuteStart, (int)bits * (sizeof(int) * 8))).AsInt32();
            var m = Avx2.PermuteVar8x32(v.AsInt32(), permute).AsInt64();
            m.StoreUnsafe(ref Unsafe.Add(ref output, outputIdx));
            outputIdx += 4 - BitOperations.PopCount(bits);
        }
    }

    // remainder, do that in a scalar fashion
    for (; i < len; i++)
    {
        ref var cur = ref Unsafe.Add(ref output, i);
        if (cur < 0)
            continue;
        Unsafe.Add(ref output, outputIdx++) = cur;

    }
    return outputIdx;
}