Split accumulator for convolutional neural network accelerator

Abstract

Disclosed embodiments relate to a split accumulator for a convolutional neural network accelerator, comprising: arranging original weights in a computation sequence and aligning by bit to obtain a weight matrix, removing slack bits in the weight matrix, allowing essential bits in each column of the weight matrix to fill the vacancies according to the computation sequence to obtain an intermediate matrix, removing null rows in the intermediate matrix, obtain a kneading matrix, wherein each row of the kneading matrix serves as a kneading weight; obtaining positional information of the activation corresponding to each bit of the kneading weight; divides the kneading weight by bit into multiple weight segments, processing summation of the weight segments and the corresponding activations according to the positional information, and sending a processing result to an adder tree to obtain an output feature map by means of executing shift-and-add on the processing result.

Description

CROSS REFERENCE TO RELATED APPLICATIONS[0001]This application is a national application of PCT / CN2019 / 087769, filed on May 21, 2019. The contents of PCT / CN2019 / 087769 are all hereby incorporated by reference.BACKGROUND OF THE INVENTION1. Field of the Invention[0002]The invention relates to the field of neural network computation, and particularly to a split accumulator for a convolutional neural network accelerator.2. Related Art[0003]A deep convolutional neural network has achieved a significant progress in application of machine learning, for example, real-time image recognition, detection and natural language processing. In order to improve accuracy, architecture of the advanced deep convolutional neural network (DCNN) owns a complex connection and massive neurons and synapses to satisfy requirement for high accurate and complex tasks. In a convolution operation, weights are multiplied by corresponding activations, and finally, products are added up to perform a summation. That i...

AI Technical Summary

Benefits of technology

The invention is a split accumulator for a convolutional neural network accelerator that uses weight kneading technique for the acceleration. This technique improves the efficiency of modern DCNN models by reducing the number of weights without loss of accuracy. The patent also describes the architecture of a Tetris accelerator that further enhances the acceleration of convolution computation. The technical effects of the invention are increased efficiency and speed in the acceleration of convolutional neural networks.

Problems solved by technology

Different weights or activations may produce different latency time in the process of acceleration, so unexpectable cycles are produced.

Design of hardware necessarily covers the worse case, and only if the cycle of the worse case functions as a cycle of the accelerator, a processing cycle is increased, and a frequency of the accelerator is reduced while also adding complexity of the design.

Moreover, accuracies desired by different DCNN models are different, and even different layers of the same model have different requirements for accuracy, so the multiplier designed for the convolutional neural network accelerator must cover the worse case.

A main component of the typical DCNN accelerator is a multiply-adder, and the main problem of the multiply-adder is to perform invalid operation.

Secondly, an intermediate result of the multiplication and addition operation is useless.

However, such will necessarily sacrifice accuracy of the result, and in particular, in large data sets, accuracies of these solutions are quite seriously damaged.

However, it is impossible to avoid computing the zero bits.

However, it is quite difficult to realize this object, because it is necessary to modify the current MAC computing mode, and reconstruct hardware architecture to support a new computing mode.

Images