Abstract:

[Objective] In the deep learning framework, in order to realize large-scale deep learning computing, the heterogeneous OpenCL computing model has become important to improve learning efficiency by making full use of different types of hardware devices and computing resources produced by different manufacturers. Therefore, deep learning frameworks such as MXNet are migrated to the OpenCL computing model to improve their adaptability to large-scale deep learning. In the process of migrating the MXNet deep learning framework, the migration of operations related to multi-dimensional linear data processing, which is common in deep learning computing, is the main issue discussed in this paper. [Methods] By systematically comparing the operating mechanisms of the CUDA computing model and the OpenCL computing model, the logic of multi-dimensional linear data processing in the MXNet deep learning framework compatible with the CUDA computing model is reconstructed based on the OpenCL computing model. [Results] The computing operations related to multi-dimensional linear data processing in the MXNet deep learning computing framework based on the OpenCL computing model for adaptive reconstruction can pass the existing framework tests. [Conclusions] The adaptive reconstruction scheme based on the OpenCL computing model can well solve the common migration problem of multi-dimensional linear data processing related operations when the MXNet deep learning framework is migrated to the OpenCL computing model.

Key words: OpenCL kernel program, MXNet, multidimensional linear data calculation