Convolution operation acceleration method for carrying out weight splitting according to bit precision

Abstract

The invention discloses a convolution operation acceleration method for carrying out weight splitting according to bit precision, and the method comprises the following steps: 1), carrying out the traversal and comparison of each bit of each weight W in a convolution kernel, marking non-zero elements ck * k'and q in all bit data, marking input feature map data FIk * k 'corresponding to the weightWk * k' to which the position belongs, transmitting the marked input feature map data FIk * k 'to FIk * k' and q ', and supplementing the unmarked FIk * k' and q 'with' 0 '; 2) for each bit, accumulating all input data classified and given to the current bit; 3) multiplying the accumulation result under each q bit by 2q; 4) accumulating results obtained by shift calculation again to obtain outputfeature map data FO of the current convolution operation; and iteratively repeating the steps 1) to 4) to generate output feature map data of all convolution channels. According to the invention, theconvolution operation gets rid of the dependence on on-chip DSP resources, and the upper limit of the calculation parallelism degree that the convolution operation can reach is improved.

Description

technical field [0001] The invention relates to a convolution operation acceleration method for performing weight splitting according to bit precision. Background technique [0002] Convolutional neural network (CNN) has been widely used in image classification and speech recognition. CNN is built by imitating the biological visual perception mechanism. Visual perception is mainly composed of multi-layer neurons that process data. Mapped to the CNN network structure, it is embodied as a multi-layer convolutional layer connection. According to statistics, volume Product calculation accounts for 99% of the total computation of CNN. In recent years, many applications including autonomous driving have put forward higher real-time requirements for CNN. Therefore, the design of CNN accelerator has become the mainstream research direction in the industry. [0003] High flexibility, low cost, and short design cycle make FPGA-based CNN accelerator designs more suitable for deployme...

AI Technical Summary

Problems solved by technology

[0009] 3. CNN implements the functions of classification and recognition. When deploying CNN accelerators in applications such as automatic driving, peripheral circuits are still required to realize the functions of automatic driving. These circuits have certain requirements for on-chip DSP resources, while existing The convolution scheme of the multiplication is limited by the reduction

Images