While following a discussion on KVR, I thought about adding support for denormals handling in Audio Toolkit
What are denormals?
Denormals or denormal number are numbers that can’t be represented the “usual” way in floating point representation. When this happens, the floating point units can’t be as fast as with the usual representation. These numbers are really low, almost 0, but not exactly 0. So this can often happen in audio processing at the end of the processing of a clip, and sometimes during computation for a handful of values.
In the past, on AMD CPUs, the FPU would even use the denormal process for bigger values than on the Intel CPUs, which lead to poorer performance. This doesn’t happen anymore AFAIK, but if your application is slow, you may want to take a look at a profile and determine if you could have an issue there. Denormals behavior can be detected by an abnormal ratio of floating point operations per cycle (the number is too low).
Flush to zero on different platforms
There are different ways of avoiding denormals. One not so good one is to add background noise to the operations. The issue is what amount (random or constant) and the fact that not all algorithms can handle them.
The better solution is to use the CPU facilities for this. x86 processors have internal flags that can be use to flush denormals to zero. Unfortunately, the API is different on all platforms.
On Windows, the following function is used:
_controlfp_s(&previous_state, _MCW_DN, _DN_FLUSH);
On Linux, as an extension of C99, gcc added this function:
_mm_setcsr(_mm_getcsr() | (_MM_DENORMALS_ZERO_ON));
And finally, OS X has yet a different way of doing things.
ARM platform is yet different. The default compiler has also an API, but if you are compiling with GCC, flush to zero is activated through the command line option -funsafe-math-optimizations.
Using flush to zero
The functions need to set the state before processing anything and then they need to be reused to set the state as it was before. This is to ensure that calling code has the same FPU state. What your functions can handle (arbitrary noise for small values) may not be acceptable for other applications.
The total amount of change in terms of performance may not be impressive. Using the functions to change FPU state means that there is an overhead (that may not be important, but an overhead nonetheless) and that the algorithms will behave slightly differently. So flushing to zero is about compromise.
Flushing denormals to zero may not be mandatory, but having the option to enable it is neat. So this is now available in Audio Toolkit 1.3.2.