Signal Processing and Speech Communication Laboratory
homeevents › Guest Lecture: Thibault Hilaire

Guest Lecture: Thibault Hilaire

Start date/time
Fri Jan 27 13:00:00 2017
End date/time
Fri Jan 27 13:00:00 2017
Seminar room IDEG 134 (Inffeldgasse 16c)

**Prof. Thibault Hilaire **(Université Pierre et Marie Curie (Maître precision de Conférences), Laboratoire d’Informatique de Paris), will give a guest lecture with the title

From controller to code: an overview of the finite-precision implementation flow

Transforming an algorithm from mathematical notation to fixed-point code running on some CPU or custom hardware is a complicated and error-prone task that requires large skills (numerical analysis, computer architecture, computer arithmetic, etc.), especially when we want guaranty on the numerical quality of the code. This task is typically performed using simulation tools (like Matlab/simulink) that help developers/designers to evaluate (using simulation) the impact of the finite precision computations. But this approach does not provide any guarantee on the numerical quality of the implementation, and thus no mathematical bound on the output error. Moreover, applied to signal processing or control algorithms, this approach cannot deal with implementation possibilities, like the reorganization of the algorithm (several structures exists for a given linear filter), the reorganization of the computation reorder, the bit-width optimization, etc. We will present a complete flow, applied to linear signal processing/control algorithm (but not restricted to) that provides fixed-point (or floating-point) implementations from mathematical specification. with a reliable bound on the output error. This is based on basic bricks like a reliable evaluation of the magnitude of each variable (in order to define their fixed-point format and guaranty that no overflow will occur), an error analysis to bound the output error (difference between the exact algorithm and the implemented one), a bit-width optimization (smallest bitwidth that still guaranty the error to be bounded by a given epsilon), and fixed-point code generation (C code with integers, or VHDL operator).