A computing device, computing method, storage medium and terminal

Abstract

The invention provides a calculation device, a calculation method, a storage medium and a terminal. The calculation device includes a local storage unit; a storage control unit, which is connected to the local storage unit in communication; a multiply-accumulate calculation unit, which is used to execute Various convolution calculations and accumulation calculations; pixel calculation engine unit, the pixel calculation engine unit is used to perform pooling and filtering operations; task control unit, the task control unit is used to configure the calculation mode and input and output addresses of the computing device; among them, the task The control unit is all connected to the storage control unit, the multiplication and accumulation calculation unit and the pixel calculation engine unit to output control information, and the multiplication and accumulation calculation unit and the pixel calculation engine unit are all connected to the storage control unit to realize information transmission. The present invention not only reduces the It not only reduces the cost of visual processing, but also increases the flexibility of computing, while effectively saving data handling and reducing system power consumption.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a computing device, a computing method, a storage medium and a terminal. Background technique [0002] In recent years, deep learning algorithms have been widely used in artificial intelligence applications such as computer vision, especially convolutional neural networks. Compared with traditional computer vision algorithms, deep learning can achieve higher accuracy in application fields such as image classification, object recognition, and speech analysis. Compared with traditional computer vision algorithms that rely on manual design of detection features, deep learning algorithms The training of the model is easier to replicate on a large scale. However, due to the high computational complexity of deep learning algorithms, traditional processors such as CPUs and GPUs cannot perform efficiently, so various deep neural network hardware accelerators have emerge...

AI Technical Summary

Problems solved by technology

Since the performance, cost, and power consumption of general-purpose DSP processors are aimed at a certain type of specific application, they generally cannot achieve the best results. ) uses a dedicated IVE (Intelligent VideoEngine) processor to process traditional computer vision algorithms, and a dedicated NPU (NNIE) to process deep neural network algorithms

[0005] However, in the current solution, when two different computing units are used at the same time, there are large data transfers, resulting in high power consumption.

Images