Need to avoid LLVM usage MATLAB r2012a doesn't understand PTX provided by new compiler

Guys I got the same problem.
MATLAB doesn’t understand structure of PTX files compiled using LLVM. I know that nvcc switches between old an new compiler depending on specified arch. The question is: How can I avoid LLVM usage during compilation?
BTW: What’s the difference between PTXes? Maybe it’s possible to write parser to transform PTX in MATLAB-friendly form?

Starting with cuda 4.1, nvcc uses LLVM as GPU code compiler for compute capability >= 2.0 by default, whereas Open64 remains the GPU code compiler for compute capability <=2.3. If you want to change this behaviour and force the use of Open64 even for compute capability >=2.0, you can use the compiler switch “-open64”.