Neural network approximate multiplier implementation method and device based on preprocessing

Abstract

The invention provides a neural network approximate multiplier implementation method and device based on preprocessing. The method comprises the following steps: before the unknown operand is input, determining a first result of truncation of a known operand after a target digit is reserved and determining a first displacement corresponding to the truncation; after the unknown operand is input, determining a second result of truncation of the unknown operand after the target digit is reserved and determining a second displacement corresponding to the truncation; multiplying the first result and the second result to obtain a partial product; and performing displacement on the partial product according to the first displacement amount and the second displacement amount to obtain an approximation result. According to the method, known operands are preprocessed and stored in advance, and a plurality of operands are prevented from being processed at the same time, so the energy consumption of a multiplier unit in a processing step can be reduced, the design area of the multiplier unit is reduced, energy consumption required by the whole neural network during calculation is further reduced, and time for outputting a result by the neural network is reduced.

Description

technical field [0001] The invention relates to the field of computers, in particular to a method and device for realizing a neural network approximate multiplier based on preprocessing. Background technique [0002] In neural networks, convolution operations are widely used. Therefore, the number of multiply-accumulate operations is extremely large, and the multiply-accumulate unit is the computing unit that requires the most space and power consumption in the entire neural network, and is the main source of computing resource requirements. Therefore, the multiplier, as one of the most important arithmetic modules in the processor, has a decisive impact on the performance and energy efficiency of the hardware unit of the computational neural network. Although errors are undesirable in an essential sense, due to the iterative nature of neural networks, they have a certain tolerance for calculation errors, and this error recovery capability can be used to save energy. There...

AI Technical Summary

Problems solved by technology

[0005] However, the current design based on these directions is only an approximate multiplier design from the perspective of saving energy consumption of a single multiplier

Therefore, current approximate multipliers for neural networks are still energy-intensive and large in area

Images