Heterogeneous format storage method based on cpu and gpu heterogeneous platforms

Abstract

The invention relates to a heterogeneous format storage method based on a CPU (central processing unit) and GPU (graphics processing unit) heterogeneous platform. The method includes the steps: firstly, detecting scales of computational data of a GPU and the communication bandwidth between a CPU and the GPU; secondly, computing number of segments of segmenting transmission according to results detected by step 1; thirdly, executing data format conversion for data of a first segment; fourthly, uploading completely converted data to the GPU, executing computation by the GPU, and executing format conversion for data of next segment; fifthly, judging whether currently computed data are a final data segment or not, finishing algorithm execution if the currently computed data are the final data segment, and executing step 4 if the currently computed data are not the final data segment. Heterogeneous format storage methods are used on different types of processors, the method mainly researches a hidden method of overhead caused by format conversion needed when the CPU executes the GPU, so that higher computation throughput is obtained by an algorithm in an overall node, and the method has wide practical values and application prospects in computational intensive applications such as high-performance computation.

Description

technical field [0001] The invention relates to a heterogeneous platform composed of a CPU and a GPU. The architectures of the CPU and the GPU are very different. In order to give full play to the data processing capabilities of the two processors, the same data adopts different formats on the CPU and the GPU. Storage, specifically related to a heterogeneous platform composed of CPU and GPU, the same data is stored in a heterogeneous format on the CPU side and the GPU side, when the CPU performs format conversion on the GPU side data, the additional overhead caused by the conversion process is hidden The realization form belongs to the field of computer system structure. [0002] technical background [0003] At present, many high-performance computer computing nodes are composed of processors with two different architectures: multi-core CPU and GPU. Because the architectures of these two processors are quite different, the parallel algorithm of the same data structure data i...

AI Technical Summary

Problems solved by technology

However, due to the difference between the two storage formats of the same data, using both formats of data on a heterogeneous platform composed of CPU and GPU will bring additional overhead of format conversion, so research effectively hides the additional overhead of format conversion The method to ensure the multiplication of the transposed sparse matrix by the product of the sparse matrix and the dense vector is of great significance to obtain high computing throughput on the heterogeneous platform composed of CPU and GPU.

Images