Flow program multi-granularity division and scheduling method based on GPU/CPU hybrid architecture

Abstract

The present invention discloses a flow program multi-granularity division and scheduling method based on a GPU / CPU hybrid architecture. The method comprises: according to computation features of each task of a data flow program and data traffic among tasks, distributing each task to a suitable computing platform; splitting GPU end tasks to each GPU in a balanced manner according to parallelism of the GPU end tasks, so as to avoid that great communication overheads among GPUs affect overall execution performance of an application; by selecting suitable CPU cores, distributing CPU end tasks among the CPU cores in a balanced manner, so as to ensure load balance and improve the utilization rate of each CPU core; using a plurality of data storage structures and a plurality of methods of access types to improve memory access efficiency; and generating a target template class and reducing the number of target nodes to reduce the redundancy of object codes. According to the invention, the flow program multi-granularity division and scheduling method based on a GPU / CPU hybrid architecture achieves load balance on a CPU platform and a GPU platform separately, reduces communication overheads, improves data access efficiency, reduces the redundancy of object codes, and improves execution performance of a data flow program.

Description

technical field [0001] The invention belongs to the technical field of a system platform based on a GPU / CPU hybrid architecture, and more specifically relates to a multi-granularity division and scheduling method for stream programs based on a GPU / CPU hybrid architecture. Background technique [0002] The emergence of CPU multi-core systems has greatly improved the execution efficiency of applications in related fields, especially in computing-intensive application fields such as digital media processing and scientific computing. The continuous upgrading of the graphics processing unit (Graphic Processing Unit, hereinafter referred to as GPU) system structure has promoted the development of GPU computing, so that it can not only efficiently perform computer graphics calculations, but also has great advantages in parallel computing of large-scale data. The advantages. Compared with CPU multi-core systems, GPU has the characteristics of multiple core processing units, fine th...

AI Technical Summary

Problems solved by technology

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a stream program multi-granularity division and scheduling method based on the GPU / CPU hybrid architecture, aiming at the massive data parallelism, task parallelism and pipeline parallelism existing under the GPU / CPU hybrid architecture and other issues, using the synchronous data flow graph as an intermediate representation, a multi-granularity division and scheduling method for data flow programs based on GPU / CPU hybrid architecture is designed and implemented. This method effectively improves the performance of data flow programs under GPU / CPU hybrid architecture. execution performance

Images