nvidia_logoWladimir Jasper van der Laan released a few week ago a disassembler for CUDA G80/G90 CUBIN (CUDA true hardware binary NVidia never described) and today he just released a first version of an assembler for this format. With these two tools, we now have an entire toolchain allowing very interesting optimization on the code produced by CUDA.

http://www.cs.rug.nl/~wladimir/decuda/