Scalable parallel data loading device and method

Abstract

The invention provides a scalable parallel data loading device and method, which are used for parallel loading acceleration of the data tensor, can be conveniently scaled and expanded, provide data input for parallel execution units of any scale, and have the characteristic of high data bandwidth. The device comprises a parallel input register array, the design size and the calculation unit array size meet a certain rule, and hardware customization design is carried out based on the rule. The method comprises an input data transformation method for determining a sequence of data after dimension reduction; and the device also includes a parallel input data access engine which performs parallel access on the data in the parallel input register array, has a specific control algorithm and a circuit structure, and is optimized for facilitating chip implementation. A set of hardware data processing method is designed and formed, the hardware data processing method comprises a transformation algorithm and an addressing rule, the utilization rate of local space information of input data is improved, high-bandwidth data input is provided for a parallel acceleration computing unit, and the number of times of access to a main memory is reduced.

Description

technical field [0001] The invention belongs to the field of computer hardware, artificial neural network algorithm deployment hardware acceleration, tensor computing hardware, and the field of digital integrated circuit design, and specifically relates to a key processing device for input data of a deep convolutional neural network hardware acceleration chip, and its design method. Background technique [0002] The deep convolutional neural network algorithm is composed of multiple layers of specific neuron algorithm layers and hidden layers, mainly including convolutional layers, and the main operator is convolution calculation of matrix or vector. The main characteristics of this computing task are the large amount of input data, the coupling of the input data with spatial feature information, and the data calculated by each convolution often overlaps with the already calculated data, and the input data is often data in tensor format. The required calculation data is ext...

AI Technical Summary

Problems solved by technology

In the inventions described in patent documents 1 and 3, due to the different sizes of different neural network algorithm layers and different degrees of data reuse, accelerator resources are wasted, so that it is necessary to cooperate with other heterogeneous processors to help solve data-related problems; patent The storage method described in 3 needs to back up more data, resulting in too large a Buffer size; the method of patent 2 adopts the idea of ​​reconfigurable computing, although it pays great attention to saving resources and waste, its data segmentation and arrangement methods are very complicated; the method of patent 4 The invention is too coupled with the design of the central processing unit, and the design and implementation complexity is too high

Images