* Compile the mmx assembler for scaling (I will post the cmake solution I prepared with lots of help from Alexander to the build system ml). On an mmx-enabled system, scaling dropped from 30-100 ms to 6-30ms.
* Smaller checks apparently paint lot faster