Performance

The image below shows the approximate performance of:

• fpm (the Q16.16 type)
• libfixmath (fix16), an alternative C library (see notes).
• Compositional Numeric Library (CNL), an experimental library that is being developed in C++’s SG14.
• native FPU operations (float and double).

The graph below plots the average number of nanoseconds a single operation takes: lower is better. The data was collected with Google Benchmark on an Intel Core i7-5820K, 3.3GHz.

The results show the following:

• Compared to libfixmath, the performance of fpm is at least as good, except for exp, where it’s considerably slower.
• Compared to CNL, fpm only matches the performance for sqrt. However, CNL does not support the majority of benchmarked functions.
• Compared to native single-precision floating-point operations, fpm is slower by up to an order of magnitude for most functions, except for the basic add, sub and several power or trigonometry functions, where it is faster.

Notes

For a fair comparison, libfixmath was compiled with FIXMATH_NO_CACHE. It should also have been compiled with FIXMATH_NO_OVERFLOW (since fpm does not detect overflow), but libfixmath failed to compile with that option.