Affiliations: IMEC-DESICS, Kapeldreef 75, B3001 Leuven-Heverlee,
Belgium | E.E. Department of University of Patras, Greece | E.E. Department of K.U. Leuven, Belgium
Abstract: Powerful multimedia applications are running more and more on very
compact and resource-scarce, portable systems. As a consequence, system design
optimization, its associated time-to-market constraints and the required
automated tool support are becoming increasingly important issues, especially
in situations where product derivatives and extensions introduce unforeseen and
possibly dramatic constraints in the system optimization process. Nevertheless,
the system designer remains an irreplaceable cornerstone for steering the whole
system optimization process. This paper presents the relationship of aforementioned aspects in
the context of optimizing data access to memory, which is the dominant factor
determining the system-on-a-chip area, data throughput and power consumption.
The case study of a 1D and 2D forward and inverse wavelet transform,
interacting with surrounding system modules imposed by current multimedia
compression standards, leads the reader through the peculiar technical
counter-measures and script-based optimization steps to be followed for
reaching a satisfactory global optimization. In particular, the data
dependencies between the different functional modules are shown to be crucial
in the memory optimization process and lead to non-trivial/counter-intuitive
decision takings that can increase the energy consumption gains compared to
more commonly-accepted, though suboptimal approaches. An example is the
counter-intuitive observation that though JPEG2000 uses independently entropy
coded blocks over its wavelet subbands, it requires more memory because of
"hidden" data dependencies, than its zero-tree based MPEG-4 counterpart, whose
intricate entropy coding spreads over all the subbands. Hence, to achieve an
overall optimal implementation with good trade-offs between efficiency and
cost, it is strongly suggested that algorithmic and implementation designers
should co-operate in early stages of the multimedia systems design, facilitated
by high-level memory cost estimation analyses.