Automatic Data Path Generation from C code for Custom Processors

Abstract

The stringent performance constraints and short time to market of modern digital systems require automatic methods for design of high performance applicationspecific architectures. This paper presents a novel algorithm for automatic generation of custom pipelined data path for a given application from its C code. The data path optimization targets both resource utilization and performance. The input to this architecture generator includes application C code, operation execution frequencies obtained by the profile run and a component library consisting of functional units, busses, multiplexers etc. The output is data path specified as a net-list of resource instances and their connections. The algorithm starts with an architecture that supports maximum parallelism for implementation of the input C code and iteratively refines it until an efficient resource utilization is obtained while maintaining the performance constraint. This paper also presents an algorithm to choose the priority of application basic blocks for optimization. Our experimental results show that automatically generated data paths satisfy given performance criteria and can be obtained in a matter of minutes leading to significant productivity gains.