Last I played with FFT convolution it didn't end up being a win for any sane filter size, even compared to brute-force convolution with arbitrary non-separable kernels. Admittedly I was using off-the-shelf tools in Nuke, but I would imagine both their FFT and Convolve nodes are reasonably optimized.

There's an analysis of the 1D case here with some interesting cautions at the end: http://www.engineeringproductivitytools.com/stuff/T0001/PT15...