What is the canonical audio sample data type in iOS 5

In the iOS 5.0 documentation it is stated that the canonical audio data type is 16 bit signed int (link):

The canonical audio data sample type for input and output.

typedef SInt16 AudioSampleType;

Discussion

The canonical audio sample type for input and output in iPhone OS is linear PCM with 16-bit integer samples.

However if I right-click "jump to definition" on AudioSampleType I see the following definition, in CoreAudioTypes.h:

#if !CA_PREFER_FIXED_POINT
typedef Float32     AudioSampleType;
typedef Float32     AudioUnitSampleType;
#else
typedef SInt16      AudioSampleType;
typedef SInt32      AudioUnitSampleType;
#define kAudioUnitSampleFractionBits 24
#endif

and again when jump-to-def for CA_PREFER_FIXED_POINT I see:

#if !defined(CA_PREFER_FIXED_POINT)
    #if TARGET_OS_IPHONE
        #if (TARGET_CPU_X86 || TARGET_CPU_X86_64 || TARGET_CPU_PPC || TARGET_CPU_PPC64) && !TARGET_IPHONE_SIMULATOR
            #define CA_PREFER_FIXED_POINT 0
        #else
            #define CA_PREFER_FIXED_POINT 1
        #endif
    #else
        #define CA_PREFER_FIXED_POINT 0
    #endif
#endif

Checking in my code at run-time, I see that CA_PREFER_FIXED_POINT is defined to be 1, both on the simulator and on my iPod.

So, my questions:

What is the canonical type? Is it always SInt16 on the device?
Under what case does the 3rd line above evaluates to 'true'? I mean, which device runs iPhone OS and use one of listed CPU's?
Is there a use case where I should re-define CA_PREFER_FIXED_POINT to 0 (when programming for iPhone)?

Solution

Read the contents of the link, and this line in your headers again:

#define kAudioUnitSampleFractionBits 24

The canonical type for audio input and output is equivalent to SInt16.

The canonical type for other audio processing, such as the new iOS 5 filter Audio Units, is 8.24 signed fixed-point.

If you do your own DSP code for near real-time iOS audio processing, benchmark it with the different types, as on some of the newest ARM cores, sequences of 32-bit floats are often faster than using either of the above canonical types, and coded in NEON asm code even faster.