I've implemented a method for parsing an unsigned integer string of length <= 8 using SIMD intrinsics available in .NET as follows:

public unsafe static uint ParseUint(string text)
{
  fixed (char* c = text)
  {
    var parsed = Sse3.LoadDquVector128((byte*) c);
    var shift = (8 - text.Length) * 2;
    var shifted = Sse2.ShiftLeftLogical128BitLane(parsed, 
      (byte) (shift));

    Vector128<byte> digit0 = Vector128.Create((byte) '0');
    var reduced = Sse2.SubtractSaturate(shifted, digit0);

    var shortMult = Vector128.Create(10, 1, 10, 1, 10, 1, 10, 1);
    var collapsed2 = Sse2.MultiplyAddAdjacent(reduced.As<byte, short>(), shortMult);

    var repack = Sse41.PackUnsignedSaturate(collapsed2, collapsed2);
    var intMult = Vector128.Create((short)0, 0, 0, 0, 100, 1, 100, 1);
    var collapsed3 = Sse2.MultiplyAddAdjacent(repack.As<ushort,short>(), intMult);

    var e1 = collapsed3.GetElement(2);
    var e2 = collapsed3.GetElement(3);
    return (uint) (e1 * 10000 + e2);
  }
}

Sadly, a comparison with a baseline uint.Parse() gives the following, rather unimpressive, result:

What are some of the ways the above code can be improved? My particular areas of concern are:

• The way a bit shift of the SIMD register happens with a calculation involving text.Length
• ~~The unpacking of UTF-16 data using a MultiplyAddAdjacent involving a vector of 0s and 1~~
• The way elements are extracted using GetElement() -- maybe there's some ToScalar() call that can happen somwehere?