Method, device, computing device, and computer storage medium for allocating display memory

Abstract

The invention discloses a method, a device, a computing device and a computer storage medium for allocating a display memory, wherein the method comprises the following steps of analyzing the data flow paths of one or more depth learning algorithm models loaded in the device to obtain an analysis result; according to the parsing result, obtaining the first number of display memory blocks needed inthe data flow process of each depth learning algorithm model; determining an allocation rule of a first number of display blocks in a data flow process of each depth learning algorithm model; allocating display memory blocks to the depth learning algorithm model according to allocation rules corresponding to each depth learning algorithm model. In accordance with that invention, according to theanalysis results of the data flow path, the number of the display memory blocks and the allocation rules of the number of the display memory blocks needed by the depth learning algorithm model are determined, so that the display memory in the device can play the role of cache to the maximum extent, and the project implementation cost of the depth learning algorithm model can be reduced by improving the utilization rate of the display memory.

Description

technical field [0001] The invention relates to the technical field of computers, in particular to a video memory allocation method, device, computing equipment and computer storage medium. Background technique [0002] With the rapid development of hardware technology, the mainstream hardware devices currently on the market can already meet the basic operation of deep learning algorithms, but there are many problems hidden in actual industrial production. One of the more typical problems is as follows: most open source deep learning frameworks are mainly developed for academic research, but the memory utilization rate of graphics cards has not been very high, and general graphics card devices cannot afford to feature ultra-large-scale deep neural networks such as ResNet Extract the memory requirements of deep learning algorithms for networks. Moreover, in actual industrial use, if you use an open source deep learning framework, such as Caffe, a general deep learning framew...

AI Technical Summary

Problems solved by technology

One of the more typical problems is as follows: most open source deep learning frameworks are mainly developed for academic research, but the memory utilization rate of graphics cards has not been very high, and general graphics card devices cannot afford to feature ultra-large-scale deep neural networks such as ResNet Memory Requirements of Deep Learning Algorithms for Extractive Networks

Moreover, in actual industrial use, if you use an open source deep learning framework, such as Caffe, which is a general deep learning framework, its hardware cost is very expensive, which will cause a graphics card to be unable to load multiple deep learning algorithm models.

[0003] Therefore, a solution suitable for industrial production is needed to overcome the problems of low memory utilization and high implementation cost of existing deep learning algorithm models during operation

Images