With a test image set of 2800x1400 bytes on a Intel Core i7 5960X this improves runtime of MODE_HH with about 10%. (this particular replaced code segment is approx 3 times faster than the non-SSE2 variant). I was able to reduce runtime by 130 ms by this simple fix.

The second part of the SSE2 optimized part could probably be optimized further by using shift SSE2 operations, but I imagine this would improve performance 10-20 ms at best.