Multi-sample multi-channel convolutional neural network Same convolution vectorization implementation method

Abstract

The invention discloses a multi-sample multi-channel convolutional neural network Same convolution vectorization implementation method, which comprises the steps of 1, storing input feature data set data according to a sample dimension priority mode, and storing data of convolution kernels according to a number dimension priority mode of the convolution kernels; 2, dividing a data matrix of the input feature data set into a plurality of matrix blocks according to columns; step 3, transmitting the convolution kernel data matrix to the SM of each kernel each time, transmitting a sub-matrix formed by row extraction from the input feature data matrix to the AM of each kernel, executing vectorization matrix multiplication calculation and parallelization matrix multiplication calculation, and executing zero supplement in the calculation; 4, storing an output characteristic matrix calculation result in an off-chip memory; and step 5, repeating the steps 3 to 4 until all calculations are completed. According to the invention, Same convolution vectorization can be realized, and the method has the advantages of simple implementation operation, high execution efficiency and precision, small bandwidth requirement and the like.

Description

technical field [0001] The invention relates to the technical field of vector processors, in particular to a multi-sample multi-channel convolutional neural network Same convolution vectorization implementation method. Background technique [0002] In recent years, deep learning models based on deep convolutional neural networks have made remarkable achievements in image recognition and classification, target detection, video analysis, etc. The rapid development of related technologies such as data processing and processors. Convolutional Neural Networks (CNN) is a type of Feedforward Neural Networks (Feedforward Neural Networks) that includes convolution calculations and has a deep structure. It is one of the representative algorithms for deep learning. The input layer of the convolutional neural network can process multi-dimensional data. Since the convolutional neural network is the most widely used in the field of computer vision, when designing the convolutional neural...

AI Technical Summary

Problems solved by technology

[0005] (1) Requires at least double the memory overhead;

[0006] (2) The storage locations of 0 elements are discontinuous, resulting in high operation overhead for supplementing 0 elements;

[0007] (3) Copying the original image data takes a lot of time and overhead

[0009] According to the architectural characteristics of vector processors, there are currently various vectorization implementation methods for convolution calculations. For example, a vectorization method for convolutional neural network operations of vector processors disclosed in Chinese patent application 201810687639.X, and patent application 201810689646.3. A GPDSP-oriented convolutional neural network multi-core parallel computing method, patent application 201710201589.5 discloses a vectorized implementation method of two-dimensional matrix convolution for vector processors, etc., such schemes use weight data loading In the vector array memory AM, the input image feature data is loaded into the scalar storage SM of the vector array memory to complete the convolution calculation, and most of them use the method of reordering the data according to the order of the third dimension. The vectorized implementation method of Same convolution in the product neural network, and the following problems will exist when the above-mentioned traditional scheme is applied to the vectorized implementation of Same convolution:

[0010] 1. The weight data cannot be effectively shared, which will waste storage bandwidth and fail to fully utilize the computing efficiency of the vector processor

[0011] 2. Since the size of the third dimension is uncertain and does not match the number of processing units of the vector processor, and the size of the third dimension of different convolutional neural network models and different convolutional layers is different, the loading data efficiency of the above-mentioned various schemes will be large. Affected and not universal

Images