FOR loops that count down are evaluated incorrectly on Evergreen GPUs.
The error is decribed here:
http://forums.amd.com/devforum/messageview.cfm?catid=390&threadid=140905
After getting the newest SDK release I found that the GPU compler still makes this basic translation error.
It is really hard justifying using OpenCL on AMD GPUs for any serious project under these circumstances.