a much better way to do your multithreaded processing is to have N worker threads do interleaved scanlines. there's plenty of parallelism to go around without resorting to splitting threaded work at the subpixel (!!) level; futhermore, unless there are extremely regular correlations in the image (causing different execution times) it should result in good (even) workload distribution. the threads do not share data at all and write to independent memory regions...

why on earth did you split it at the aa level?! this forces a square number of threads unless you get funky with precomputed antialiasing patterns (having toroidal symmetry), and since cpu cores come in powers of two usually that's 1, 4, 16, 64, etc., and each thread's results must then be gathered and combined, written again, ...

edit: i see the threading isn't relaeted to antialiasing at all, but the final image dimensions. interleaved threading will provide much better load balancing than blocked will, and when you add antialiasing into the mix (long render times...) it'll make a significant difference.