Oh my... I'd have to look at my code. You could do that as well 
I was driving the 8x8 matrix with a screen refresh of about 50Hz to 60Hz with 15bit colors (5bit per color). Doing a back of the hand calculation gives 60x8(rows)x32(intensity steps) = 15.4kHz for the row switching. I don't have a scope at hand, so I can't measure it on the real thing ;-(
Anyhow, the turn-off delay for the UDN2981A is something about 5µs, and it was too long for that purpose. The source driver of the R-Duino has a turn-off delay of 1µs, and I haven't seen any ghosting on published images.