Floating-point operations (9) in VDSP (bf561): Comparison of long double and float

By default, VDSP that the two types are the same, so we compare long double and float.

1.1 Type representation

The following are two types of data representations:

TD style= "BORDER:0;PADDING:0;" > exponential number of digits

 	float	long double
number of bytes	8
symbol bit	1	1
11
number of digits	52
minimum	1.1754943508222875e-38f
Max	3.4028234663852886e+ 38F	resolution	1.1920928955078125e-07f

1.2 Operational efficiency under normal conditions

The following table gives the efficiency comparison of two types of operations, select High Performance

	Float	Long double
Add and subtract operation	132 Cycle	166 Cycle
Multiplication operations	Cycle	179 cycle
Division operations	Cycle	1506 Cycle

Khan, this long double is too harsh for division!

The following table gives an efficiency comparison of two types of operations, choosing strict IEEE:

	Float	Long double
Add and subtract operation	310 Cycle	381 Cycle
Multiplication operations	274 cycle	585 Cycle
Division operations	497 cycle	1063 cycle

There's a value here that seems odd, and after choosing the strict IEEE option, long double makes division faster!

This explains the two different options in VDSP's documentation:

THE-FAST-FP (fast floating point) switch directs the compiler to link with the high-speed floating-point emulation librar Y. This library relaxes some of the IEEE floating-point Standard's rules for checking inputs against Not-a-number (NaN) A D denormalized numbers to improve performance. This switch was enabled by default.

THE-IEEE-FP (slower floating point) switch directs the compiler to link with the fully-compliant floating-point Library. This library obeys all of the IEEE floating-point Standard's rules, and incurs a performance penalty when compared with th E default floating-point emulation library.

're told all computations should be slower after using the strict IEEE option, but a long double floating-point division seems to be an exception!