OPENVX STANDARD TENSOR IMPLEMENTATION FOR ARTIFICIAL NEURAL NETWORKS ON SPECIALIZED PLATFORMS

Abstract

NNE extension for OpenVX 1.2 standard adds functions that implement layers of artificial neural networks. Each layer receives a tensor object as input, processes it and passes the result tensor to the next layer. Implementation recommendations and hardware optimization of tensor processing functions were proposed. The types of data supported by the OpenVX standard for creating of tensors are given, and the operations for processing them are described. The mathematical and software models of tensors are described in detail, which allow performing element-by-element processing. A new program model of the tensor is proposed, which allows to reduce the implementation of functions for processing tensors to the re-invocation of functions for image processing of the OpenVX standard. There are a detail description of functions algorithms from NNE extension based on the proposal method in the paper. Advantages and disadvantages of considered approach function implementation were shown.