James comment:

http://www.ais.uni-bonn.de/~schulz/
I think I missed it when it came out, but Hannes schulz compared theano's conv to Alex krizhevsky's and it looks like Alex's code is pretty fast. I did 70% of the job of porting it to theano ops in the theanolinear project, but if anyone wants to see 5x improvements in some cases they might want to follow up I'm that.

https://github.com/jaberg/TheanoLinear/tree/master/theano_linear/unshared_conv

What is the updated version for small patch? (The benchmarking on this
page was done quite a while ago, and it didn't look especially
thorough, but I've heard from a few sources that Alex's code is pretty
darned fast so we should compare it.)