The Research On Memory Access Mechanism Of Stream Architecture

Stream architectures are outstanding in emerging high-performance computer architectures. They are used to accelerate the media applications and scientific computing which share several important characteristics: compute intensity, parallelism, and locality. The basic idea of stream architecture is to express applications be a collection of data streams passing through a series of computation kernels. Each stream is a sequence of homogeneous data records and each kernel is a loop body that is applied to each record of the input stream(s). By decoupling memory access from computing, stream architectures can not only express media applications and scientific computing effectively, but also make good use of the advances of VLSI.The Stanford Imagine Stream Processor is one of the typical stream architectures. Imagine takes two-level model in programming language, compiler and hardware. The two-level model decouples memory access from computing and exploits all kinds of parallelism and locality thoroughly. So we focus on Imagine and study her execution mechanism.This paper will study the memory access mechanism of stream architecture.First, we define the concept of stream processing and introduce the software and hardware structure of Imagine stream architecture briefly.Second, comparing with traditional micro-processor, we discuss the memory access mechanism of stream architecture from programming model, bandwidth hierarchy and compiler optimization respectively.Third, we study the structure of streaming memory system which is based on data-level parallelism and improve it to optimize its DRAM bandwidth further.Finally, we implement "chaining" stream in stream architecture to exploit parallelism between computing and memory access further. Simulation results show that "chaining" stream can accelerate some stream applications.