Datapath and memory co-optimization for FPGA-based computation

Abstract

With the large resource densities available on modern FPGAs it is often the available memory bandwidth that limits the parallelism (and therefore performance) that can be achieved. For this reason the focus of this thesis is the development of an integrated scheduling and memory optimisation methodology to allow high levels of parallelism to be exploited in FPGA based designs. A manual translation from C to hardware is first investigated as a case study, exposing a number of potential optimisation techniques that have not been exploited in existing work. An existing outer loop pipelining approach, originally developed for VLIW processors, is extended and adapted for application to FPGAs. The outer loop pipelining methodology is first developed to use a fixed memory subsystem design and then extended to automate the optimisation of the memory subsystem. This approach allocates arrays to physical memories and selects the set of data reuse structures to implement to match the available and required memory bandwidths as the pipelining search progresses. The final extension to this work is to include the partitioning of data from a single array across multiple physical memories, increasing the number of memory ports through which data my be accessed. The facility for loop unrolling is also added to increase the potential for parallelism and exploit the additional bandwidth that partitioning can provide. We describe our approach based on formal methodologies and present the results achieved when these methods are applied to a number of benchmarks. These results show the advantages of both extending pipelining to levels above the innermost loop and the co-optimisation of the datapath and memory subsystem.