Code Transformations for Data Transfer and Storage Exploration Preprocessing in Multimedia Processors

modern processor architectures is a very important and timely problem, as discussed in another article in this issue (“Random-Access Data Storage Components in Customized Architectures,” by L. Nachtergaele, F. Catthoor, and C. Kulkarni). This problem is especially relevant in embedded applications, where cost issues such as memory footprint and power consumption are vital. In this article, we show how source-tosource code transformations play a crucial role in the solution of this problem for multimedia and telecommunications applications. Many of these code transformations can be platform-independent if they are defined carefully, in a well-chosen order.1,2 This very useful property lets us apply them to a given application’s code before we have any knowledge of platform architecture parameters such as memory size, communication scheme, and even processor type. Although the resulting optimized code behaves better on any of the modern platforms, passing it through a platform-dependent stage produces further cost-performance improvements. These optimizations are especially useful when the target is an at least partly customizable memory organization based on embedded DRAM and SRAM. Benini and De Micheli3 provide a good overview of general related work on system-level transformations, focusing especially on reducing power consumption.