Up-sampling reverse block mapping method

Abstract

The invention discloses an up-sampling reverse block mapping method. The method comprises the following steps: S1, reading input feature map data and storing the input feature map data in a shift cache region; S2, finding out pixel points in the blocks of the output feature map, and mapping the pixel points to the positions of the most adjacent four pixel points in the input feature map; S3, calculating to obtain the pixel value of the output feature map by adopting an assembly line mode: S31, in the vertical direction, carrying out one-time multiplication on four most adjacent pixel points and the parameters in the column direction, and adding four obtained intermediate values to obtain two intermediate values, and S32, multiplying the two intermediate values with the parameters in the row direction to obtain two intermediate values, and adding the two intermediate values once to obtain up-sampled pixel points of the output feature map; S4, after the data of one block is processed, returning to the S1, and processing the next block; and S5, after the input feature map is processed, continuing to process the next feature map according to the register instruction.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to an upsampling reverse block mapping method. Background technique [0002] The emergence of deep learning algorithms has made a breakthrough in the application of artificial intelligence technology. In the early days, deep learning algorithms were mainly run on servers with high-performance GPUs. With the in-depth application of deep learning, people found that although this method is simple and efficient, it also has problems such as high power consumption and large volume. Aiming at these problems, neural network accelerators can effectively solve them. As an important part of the implementation of neural network accelerators, upsampling has received higher attention and development. [0003] The main function of upsampling is to use existing pixels to construct new pixels. It is often used in neural network accelerators to enhance feature maps and restore feature map ...

AI Technical Summary

Problems solved by technology

The disadvantage of this is that data multiplexing cannot be achieved, because pixels with different output feature maps can be mapped to four adjacent pixels with the same input feature map. If it is a point-to-point mapping, repeated reading will inevitably occur. This results in a waste of time and resources

Images