Is there a more optimized floating point software library available?
Yes. Since 2004, gumstix has switched from using the kernel floating point emulation in linux to using the softfloat GCC library — the major difference is that every floating point operation no longer results in a switch to kernel code form userspace code with the overhead inherent in that. The result is about a 3x speedup in FP performance relative to the older in-kernel code (We haven’t actually done any significant benchmarking — Craig has qualitatively noticed a speedup, and stuff on the web says 3x is typical — it could be that has been improved upon over time). The benchmarks are done using softfloat.
