I have such requirements to the hash function in my project:
- operating on 32-bit data,
- not mathematically breakable (built on existing hashes known to be secure),
- free of branching,
- as much as possible, free of repetitive rounds and rotations,
- long enough to require a huge number of gates on the ASIC.
The idea is to make efficient hash focused on GPU units, and not prospective for ASICs implementation.
(Skein-256 almost seems to fit these requirements?)