Image tiling for embedded applications with non-linear constraints

Tiling is a key aspect of the design of embedded image processing applications, due to local memory constraints. To maximize system performance, the designer must select a suitable tile size that balances data transfers and computation. In this work, we present a method for optimal 2D image tile sizing using constraint programming. Unlike previous methods, ours accurately models DMA data transfer times and parallel scheduling overheads with non-linear constraints. Our experiments with a binomial filter demonstrate that we can compute the optimal tiling dimensions that minimize the execution time for different image sizes and internal memory constraints. This technique provides invaluable information for both application developers and system architects that can quickly explore design trade-offs.