Round-half-even or Round-toward-nearest

I just finished the reading of An introduction to different rounding
algorithms,  http://www.eetimes.com/document.asp?doc_id=1274485.   I must
admit I never pay much attention on the trick of the quantization. The
approach of  Round-toward-nearest is always used in my design. Did anyone
have  practical experiences to show the DSP filter performance improved or
degraded by using Round-half-even instead of Round-toward- nearest?

---------------------------------------
Posted through http://www.DSPRelated.com

>I just finished the reading of An introduction to different rounding
>algorithms,  http://www.eetimes.com/document.asp?doc_id74485.   I must
>admit I never pay much attention on the trick of the quantization. The
>approach of  Round-toward-nearest is always used in my design. Did
anyone
>have  practical experiences to show the DSP filter performance improved
or
>degraded by using Round-half-even instead of Round-toward- nearest?
>
>---------------------------------------
>Posted through http://www.DSPRelated.com

Various rounding methods help avoid dc bias. With direct truncation dc is
very likely as we always round towards minus inf. With classic rounding to
nearest integer we don't expect much dc except if signal has dc offset. dc
then occurs due to midpoint rounding bias. This also means that such dc
bias gets high if we are rounding one or so LSB and it also explains build
up of error in accumulators. For this end rounding to nearest even is used
to avoid dc bias as the midpoint rounding is alternated.

Kaz
 
---------------------------------------
Posted through http://www.DSPRelated.com