As many speech codecs have to meet the bit-exact requirement, Intel IPP functions described in this chapter use rounding modes that are different from the default rounding mode used in general signal processing functions. For most signal processing functions, the default rounding mode can be described as “nearest even”, so that the fixed point number x=N + α, 0 ≤ α < 0, where N is an integer number, is rounded as given by:
For example, 1.5 will be rounded to 2 and 2.5 to 2.
For speech coding functions in this chapter, there are two different rounding modes.
The default rounding mode is “clipping”, so that the fractional part of the fixed point number is cut off and the result of rounding is always less than the initial value. Specifically, the fixed point number x=N + α, 0 ≤ α < 0, where N is an integer number, is always rounded to N. For example, -1.3 will be rounded to -2 and 1.7 to 1. No special suffix is added to the names of functions that use this default mode.
Another rounding mode is “nearest right”, in which the fixed point number x=N + α, 0 ≤ α < 0, where N is an integer number, is rounded as follows:
For example, 1.5 will be rounded to 2 (same as for “nearest even” mode) but -1.5 to -1 (different from “nearest even” mode where -1.5 is rounded to -2).
The suffix “NR” is added to names of functions that use the “nearest right” rounding mode.