Atom made the changes public in the new release 0.26, the performance increase is absolutly amazing!
preview note: i managed to utilize the BFI_INT instruction on AMD GPU's 5xxx and 6xxx on oclHashcat v0.26 which gave a nice performance boost. afaik oclHashcat v0.26 is the first BFI_INT hacked OpenCL based program. here is a preview from a 4 x hd5970 rig:
i think this hack also make oclHashcat the worlds fastest SHA1 single hash cracker.
To give a little descriptions, BFI_INT is an opcode of the GPU that is not available in the compiler. Atom made a permutable kernel that builds and patches itself at the runtime to be able to use the opcode!