Application Source Codes Profiling for ASIP Memory Subsystem Design

Abstract

Application Specific Instruction-set Processors (ASIPs) are a realistic solution for domain-specific applications. To reach optimal system-level performance, memory subsystem design is considered in the pre-architecture design stage, which narrows down the huge design space to applications of a specific domain. Source code profiling approach aims to understand the characteristics of applications to guide ASIP design. The memory profiler proposed in the paper uses dynamic profiling technique to generate memory traces, and the live intervals of memory objects are computed by load-store information. Then memory requirement diagram is plotted according to instruction counts. The minimum memory requirement of the application is acquired from the diagram and guides the design of memory subsystem. The profiler is tested using a computing kernel, and the memory subsystem design suggestions are given according to the profiling results.