3. Arithmetic

In this section, we focus on number representation problems. DSPs generally do not have floating-point units for :

• cost reasons ;

• performance issues: floating-point calculations, even when performed by a specialized unit, are much slower than calculations performed by the arithmetic and logic unit, i.e. on integers;

• questions of utility: most algorithms don't really need the dynamics provided by the floating-point representation; however, we do need to be aware of the few problems it can cause in terms of computational noise and processing non-linearities.

The presentation of representation problems is also an opportunity to introduce some DSP architectures.