Vector processor-oriented pooling vectorization implementation method

Abstract

The invention discloses a vector processor-oriented pooling vectorization implementation method, which comprises the following steps: according to AM space parameters and pooling window preset parameters, acquiring transmission block parameters of an input block feature map transmitted into an AM space each time; loading the sub-blocks of the block feature map to the AM space on the basis of DMA operation; according to a preset rule, partitioning the single partitioned feature map sub-block; performing pooling processing on each region of the sub-blocks of the block feature map in sequence according to a preset sequence; exporting an output feature map sub-block obtained after processing from the AM space to a memory based on DMA operation; and repeating the above steps until the pooling results of all the sub-blocks of the block feature map are obtained. The method is clear in logic, safe, effective, reliable and easy and convenient to operate, not only can support non-square feature maps, non-square moving step lengths and non-square pooling windows, but also can support feature map filling, does not need to rearrange the feature maps, and improves the pooling processing efficiency.

Description

technical field [0001] The present invention relates to the technical field of deep learning and convolutional neural network, in particular to a vector processor-oriented pooling and vectorization implementation method. Background technique [0002] Convolutional neural network is the most widely used neural network model in the current deep learning model, and its performance on specific tasks in fields such as image classification has surpassed that of humans. The convolutional neural network model generally consists of convolutional layers, activation layers, pooling layers, and fully connected layers. [0003] The pooling layer is located after the convolutional layer and is used to aggregate statistics on the features extracted by the convolutional layer. In the convolutional neural network, after using the convolution operation to extract the features of the input image, multiple feature maps will be generated. If all the features are used, the amount of calculation ...

AI Technical Summary

Problems solved by technology

[0005] At present, the method mentioned in the patent "CN 108205703 A-multi-input multi-output matrix average pooling and vectorization realization method" has the following problems: (1) before the input feature map is imported into the vector processor core AM space, the input The feature map is rearranged, see S2-S3 on page 2, which greatly affects the processing efficiency; (2) As shown on pages 4, 6, and 7 of the second page, the height and width of the feature map are required to be equal, and the pooling The horizontal movement step is the same as the vertical movement step, and the height and width of the pooling window are equal. It does not support unequal or different situations, that is, it does not support non-square matrix situations; (3) In convolutional neural networks, often through Fill the feature map (Padding) to keep the valid information on the boundary of the feature map. Currently, this document does not support the case of padding.

Images