This section describes several programming guidelines that can help you improve the performance of a floating-point applications:
Avoid exceeding representable ranges during computation; handling these cases can have a performance impact.
Use a single precision type (for example, float) unless the extra precision obtained through double or long double is required. Greater precision types increase memory size and bandwidth requirements.
Reduce the impact of denormal exceptions for all supported architectures.