A convolution operation device and method thereof

Abstract

The invention provides a convolution operation structure and method thereof. The convolution operation structure includes a conversion unit structure, and the conversion unit structure includes: a first register group, a second register group, a multiplier, an adder, and an intermediate result register group, wherein the first register group and the The second set of registers is controlled to perform matrix multiplication by the multiplier and the adder on the first matrix and the second matrix respectively stored by them during the first round of calculation, and will pass through the matrix multiplication The obtained first result matrix is ​​stored in the intermediate result register group; and the second register group and the intermediate result register group are controlled to, during the second round of calculation, perform the third The matrix and the first result matrix perform matrix multiplication through the multiplier and the adder, and store a second result matrix obtained through the matrix multiplication into the first register group.

Description

technical field [0001] The invention relates to the improvement of the neural network processor including the convolution operation unit, in particular to the improvement of the convolution operation structure capable of realizing the winograd convolution operation. Background technique [0002] Deep learning technology has developed rapidly in recent years, and it has been widely used in solving advanced abstract cognitive problems, such as image recognition, speech recognition, natural language understanding, weather prediction, gene expression, content recommendation and intelligent robots. And has excellent performance, so it has become a research hotspot in academia and industry. Deep neural network is one of the perception models with the highest level of development in the field of artificial intelligence. This type of network simulates the neural connection structure of the human brain by building a model, and describes the data features layered through multiple tran...

AI Technical Summary

Problems solved by technology

However, the existing winograd convolution operation unit is used to realize the above-mentioned matrix conversion hardware structure and the matching processing operations are very complicated, resulting in relatively poor circuit area and utility of the winograd convolution operation unit

Moreover, since the convolution operation of winograd is not ideal for the calculation effect of certain network sizes (for example, the acceleration effect of the convolution operation with a convolution kernel size smaller than 3×3 is not obvious, and the calculation speed is even lower than the traditional convolution operation Unit), so many neural network processors equipped with winograd convolution operation unit will also set up traditional convolution operation unit to deal with the calculation requirements for different network sizes

In these neural network processors, there is no computing unit that can be multiplexed for traditional convolution operations and winograd convolution operations, which makes the neural network processors need to be set up for traditional convolution operations and winograd convolution operations. corresponding circuit

Images