r4

• The parameter value is now scaled in a saner way.
• Speed up gaussian blur's code and add SIMD optimization.

r3

• Add "gaussian blur only" mode.
• Support 32-bit float input.

r2

• Skip direction calculation for mode=1.
• Add op parameter.

r1

Add AVX2 auto-dispatched code path in ICC

It speeds up a bit on AVX2-capable CPU.