topic IPP's Rounding Off Results is Not Consistent in Intel® Integrated Performance Primitives

IPP's Rounding Off Results is Not Consistent

OKohl — Wed, 08 Feb 2012 11:55:36 GMT

Hello,

I'm puzzled by an apparently inconsistencyin IPP's rounding off results:
In the first example the real result is 2.5 and IPP rounds it off to 2.
In the second example the real result is 93.5 and IPP rounds it off to 94.
Why is the 0.5 result sometimes rounded up and sometimes down?

e.g.

I'm using multiply by constant function with scaling 7 , in two cases:
(I'm giving an example for a single pixel result)

ippiMulC_16s_C1RSfs(Coeficient, ImagePtr, ImageBytesStep, RoiWork, 7);

A) Coeficient = 2, Image Pixel = 160

Result pixel is : 2 x 160 Scaled by 2^7 => 2.5 => IPP rounds the result to 2

B) Coeficient = 44, Image Pixel = 272

Result pixel is :44 x272 Scaled by 2^7 =>93.5 => IPP rounds the result to94

Thanks in advance
Offir

IPP's Rounding Off Results is Not Consistent

igorastakhov — Thu, 09 Feb 2012 08:16:32 GMT

Hi Offir,

IPP provides the same behavior as HW provides by default - and it's mentioned in ippsmanual - all IPP functions by default round FP result to the nearest-even in order to provide more uniform distribution of results. So called "accounting" rounding (when 0.5 is always rounded up) is not uniform as results are always slightly shifted up.

Regards,
Igor

IPP's Rounding Off Results is Not Consistent

OKohl — Thu, 09 Feb 2012 12:15:50 GMT

Many thanks Igor