I use a similar conditional write paradigm on the GPU and it's usually easiest to do an unconditional write and update the address using a branchless conditional, assuming you are using a system with strict write ordering. Usually the unnecessary writes won't make it out of L1 cache.