Taking into account the application domain - speech codec, and G.729 and G.723.1 specifically:
- is it recommended to enable flush-to-zero and denormals-are-zero modes
- should I expect any performance improvements
- should I expect incorrect operation (maybe in some corner cases)?
G.723.1 is a fixed-point codec. So you may notenable FZM and can't expect any benefits if it enabled.
The same true for fixed-point version of the G.729 codec.
As for floating-point version of the G.729, you can expect denormals during encoding. Their appearance depend on input data (for example large enough zero block after non-zero data).
The default codec operation mode is precise. To build codec with fast mode enabled you should add define CLIPPING_DENORMAL_MODE to the USC library makefile.
Igor S. Belyakov