Implementation method for parallel acceleration of ResNet based on general neural network processor

Abstract

The invention discloses an implementation method for parallel acceleration of ResNet based on a general neural network processor. The method comprises the following steps: loading a data set and a weight into a register from a central cache region, carrying out the operation of a convolutional layer, and loading a weight matrix into a matrix multiplication unit from the register; sending the data set matrix to a matrix multiplication unit in a streaming data mode, writing an operation result back to a register, and repeating the process until all data are processed; using a vector compression unit to complete batch standardization operation; completing the operation of linear rectification by using an SIMD operation unit ; completing the operation of a pooling layer by using the SIMD operation unit and a vector compression unit; completing operation of a full connection layer by using a matrix multiplication unit, and writing a result back to the register; and writing the result in the register back to the central cache region. The matrix multiplication unit is adopted to complete the operation of the convolution layer and a full connection layer, the optimal performance and performance power consumption ratio are achieved, and lower energy consumption, shorter model convergence time and higher image recognition speed are achieved.

Description

technical field [0001] The invention relates, in particular, to a method for realizing parallel acceleration of ResNet based on a general neural network processor. Background technique [0002] Residual network "ResNet" is a machine learning model. In the past, when performing deep network training, there was a problem that the difference gradient disappeared. Repeated matrix multiplication may make the difference gradient between the result and the initial content smaller. As the network In-depth, the performance of the model will reach saturation or even decline in the end. "ResNet" solves this problem by establishing a shortcut to alleviate the problem of gradient disappearance, so that our model can learn a constant function to ensure that the performance is at least as before without deteriorating. ResNet makes computer image recognition practical and has been widely and importantly used in life. With the development of deep neural network technology, the number of la...

AI Technical Summary

Problems solved by technology

To implement these operations on the GPU, it needs to be decomposed into basic scalar multiplication and addition, resulting in a large number of intermediate results, frequent reading and writing of memory, and the application effect is not good

Images