Heterogeneous system, acceleration method and device thereof and readable storage medium

Abstract

The invention discloses a heterogeneous system, and the system comprises: a preloading module which is used for the preloading of all heterogeneous operators in a heterogeneous device; the selection module which is used for determining an optimal execution scheme according to the theoretical consumption of all the heterogeneous operators; and the burning module which is used for loading and burning the heterogeneous operator corresponding to the optimal execution scheme. According to the method, the preloading module is used for providing an implementation basis for calling of the heterogeneous operator, loading and burning can be completed directly by sharing the memory space, secondary copying of the memory is not needed, heterogeneous acceleration is completed on the premise that an original program is not modified, the system can be applied to the field of acceleration of traditional software, and more expansion possibilities are provided for the traditional software. Correspondingly, the invention further discloses an acceleration method and device of the heterogeneous system and a readable storage medium.

Description

technical field [0001] The invention relates to the field of heterogeneous computing, in particular to a heterogeneous system, its acceleration method, device and readable storage medium. Background technique [0002] At present, heterogeneous systems have their own strengths, greatly improving the overall computing power of the system. The mainstream heterogeneous systems include "CPU+GPU", "CPU+FPGA" and "CPU+ASIC". These systems fully reflect the characteristics of heterogeneous hardware at the hardware design level. For example, compared with CPU (central processing unit, central processing unit), GPU (Graphics Processing Unit, graphics processing unit) has higher parallelism and higher stand-alone Calculate the peak value; FPGA (Field Programmable Gate Array, field programmable logic gate array) has higher performance per watt, etc.; ASIC (Application Specific Integrated Circuit, application specific integrated circuit) has dedicated functions, and each time unit can fo...

AI Technical Summary

Problems solved by technology

[0003] The three mainstream architectures have their own complex development environments and processes

For example, development frameworks such as CUDA (Compute Unified Device Architecture, Unified Computing Device Architecture), OpenCL, etc. These frameworks have the characteristics of complex environment, long development cycle, and high threshold. It is difficult for ordinary developers to make full use of heterogeneous performance and achieve ideal results.

Because of this, these mainstream architectures generally focus on artificial intelligence, data analysis, scientific computing, video image processing, video transcoding and other software development complex, long development cycle, newly developed mainstream emerging scenarios, which cannot be considered for traditional applications

Traditional applications need to modify the source code and use the corresponding development framework for complex secondary development. The development cycle is long and the threshold is high. After the development is completed, there is no load option, resulting in no good applications and breakthroughs in the field of accelerated traditional software.

Images