Sparse convolutional neural network accelerator and calculation method

Abstract

The invention discloses a sparse convolutional neural network accelerator and a calculation method. The accelerator comprises an accelerator body, the convolution calculation module is used for carrying out multiply-add processing on the input feature map and generating an intermediate result, the convolution calculation module is composed of nine calculation units, each calculation unit is composed of one or more multiply-add devices, the multiply-add devices in the same calculation unit have the same activation input, and the nine calculation units are provided with an additional activationinput; a non-linear and pooling module which is used for carrying out non-linear calculation and pooling calculation on the intermediate result to generate an output feature map; and a full connectionlayer calculation module which is used for carrying out full connection calculation on the output feature map to generate a final result. Different working modes are obtained through combination fora plurality of convolution calculation modules, so that the activation sparsity can be effectively utilized to accelerate the convolution neural network calculation, and meanwhile, relatively low additional resource overhead and relatively low load unbalance are generated.

Description

technical field [0001] The present disclosure relates to the field of deep learning, in particular, to a sparse convolutional neural network accelerator and a calculation method. Background technique [0002] In recent years, due to the acquisition of massive data in the era of big data and the significant improvement of computer performance, deep learning algorithms represented by convolutional neural networks have shown great advantages in many fields. However, like the typical classification network VGG-16, it requires 15.5G multiplication and addition operations and 138M parameters. The huge amount of calculations and parameters make the practical application of convolutional neural networks difficult. Experiments show that convolutional neural networks have inherent sparsity, and effective use of sparsity can greatly improve computing performance by reducing the amount of calculations. However, the existing design of convolutional neural network accelerators ignores the...

AI Technical Summary

Problems solved by technology

However, the use of activation sparsity requires additional logic to judge or process activation inputs, which will increase additional resource overhead; in addition, different activation input sparsities are different, and queue lengths are different. In parallel computing, waiting for the longest queue will cause Reduce parallel efficiency; and, there is an upper limit to the way to share activation input and expand output feature maps using tiling. Excessive expansion and output feature map directions will lead to reduced parallel efficiency

Images