Method for multiplexing IO units in stream processor

Abstract

The invention discloses a method for multiplexing IO units in a stream processor, which aims to solve the technical problems of overcoming the limitation that one IO unit can only execute one stream in the design of the stream processor, achieving the extended support for a large amount of input and output streams by a small amount of IO units, and reducing the hardware expense. The technical scheme is that the method comprises the following steps: modifying a compiler, adding an IO unit multiplexing processing module in the compiler, distributing a stream identifier to each stream, and scheduling IO operations to the IO units appointed by the stream identifier for execution; improving a hardware structure of an operational cluster in the stream processor, and adding K multipath selectorsbetween the IO units and a stream buffer; and decoding the stream identifier by the IO units, using the decoding result as an arbitration signal, and arbitrating the visit to the streams by each multipath selector according to the arbitration signal so as to achieve the input and the output of the streams. By using the method, the number of the input and output streams is not limited by the number of the IO units any more, the number of the IO units is greatly reduced, and the hardware expense is small.

Description

technical field [0001] The invention relates to a design method of a flow processor, in particular to an IO unit multiplexing method in the design of the flow processor. Background technique [0002] The stream processor is a typical representative of a new generation of intensive computing-oriented high-performance microprocessors, specifically for stream applications. Streaming applications are computationally intensive: Compared with traditional desktop applications, streaming applications perform a large number of arithmetic operations on each data fetched from memory. [0003] The stream processor is composed of a scalar core, a DRAM controller, a storage controller, a stream controller, a microcontroller, a stream register file SRF (Stream Register File), and multiple computing clusters, and all components are connected through an on-chip general bus. The main idea of ​​stream processing is to divide the application program into computing core functions (kernels), gen...

AI Technical Summary

Problems solved by technology

A large number of IO units will make the size of the selector in the interconnection switch in the cluster become very large, and the number of selectors will increase as the number of IO units increases, which will lead to a sharp increase in design area and delay

In addition, each IO unit requires a dedicated decoding unit and execution unit, and a large number of IO units will also consume a certain amount of hardware logic

[0008] (3) Most importantly, each IO unit can only execute one stream, which also limits the number of streams supported by the processor. Once the data streams that need to be transmitted at the same time exceed the number of IO units, the system will no longer be able to provide support, requiring programmers Adjust the program to bring inconvenience to the programmer

Images