Dataaow-driven Memory Allocation for Multi-dimensional Signal Processing Systems Category 3: High-level Synthesis, Memory System Synthesis Dataaow-driven Memory Allocation for Multi-dimensional Signal Processing Systems

All appropriate organizational approvals for the publication of this paper have been obtained. If accepted, the author(s) will prepare the nal manuscript in time for inclusion in the Conference Proceedings and will present the paper at the Conference. Abstract Memory cost is responsible for a large amount of the chip and/or board area of customized video and image processing systems. Therefore, design support to determine the storage organization for the multi-dimensional (M-D) signals is crucial to reduce the system architecture design time, while maintaining a suuciently high design quality in terms of area and/or power. In this paper, a novel background memory allocation and assignment technique is presented. It is intended for a behavioural algorithm speciication, where the procedural ordering (especially the loop ordering) of the memory related operations is not yet fully xed. Therefore , we have introduced a optimisation method driven by data-ow analysis, instead of the more restricted scheduling-based exploration in current approaches { which start from a procedurally interpreted speciication in terms of loops. Our novel allocation approach can process non-procedural functional descriptions, containing conditions (data-dependent or not), delays, M-D signals with complex aane indices, nests of loops having as loop boundaries { aane functions of the surrounding loop iterators. The allocation/assignment technique yields one (or optionally several) distributed (multi-port) memory architecture(s) with fully-determined characteristics, complying with a given clock cycle budget allocated for read/write operations. Our allocation approach currently employs the estimated silicon area occupied by a multi-port memory connguration as a steering cost. This estimation is provided by an area model for on-chip memories. Moreover, our approach can accurately deal with large multi-dimensional indexed signals: the allocation system incorporates a polyhedral data-ow analysis, allowing to operate directly with groups of signals analytically determined, rather than with individual signals. These claims are substantiated on several realistic test vehicles.