Results measured on Haswell.

For convolve_2d
Width Height Improvement
    4      4        8.3%
    8      8        5.6%
   64     64         13%
    4     16         13%
   32      8          9%

For convolve_x
Width Height Improvement
    4      4          0%
    8      8          0%
   64     64         21%
    4     16          0%
   32      8         29%

Change-Id: Ic13215d61f48400b3e31d5ffea3af11c0b3f0ed7