System on Chip Seminar
Global data storage and transfer exploration in (parallel) processor architectures for multi-media applications.
Francky Catthoor, IMEC
Many current applications are data-dominated. This happens in embedded multi-media applications like image, video, speech and audio processing but also in broadband wired or wireless telecom terminals and even modern network protocols. For these application domains, one of the most crucial problems in the design trajectory is related to data storage access and global data communication for complex data types like multi-dimensional arrays or sets of labeled data.
A comprehensive design exploration script has been developed to deal with this in the past for customized processor architectures. Models and automatable optimisation techniques have been investigated and developed to support the memory organisation tasks. They have a large impact on both (board) cost and power consumption. Practical examples have demonstrated potential savings in the range of an order of magnitude rather than a few dozen %. This relies especially on system-level transformations of the initial algorithmic specification. In the last 5 years, also extensions towards data storage and transfer related issues in (partly) predefined signal and video processor architectures have been studied. The focus of this talk will be mainly on these issues, including global data-flow and control-flow (mainly loop) transformations, task versus data parallelism exploitation and agressive cache oriented code optimisations. Experiments on real-life multi-media applications show a speed-up of up to a factor 3 and a typical reduction in memory bus load (on-chip and off-chip) with an order of magnitude. The result is also a significant energy consumption reduction.