Error evaluation method for multiple approximation technologies based on probabilistic graph model in special accelerator

Abstract

The invention provides anerror evaluation method for multiple approximation technologies based on probabilistic graph model in special accelerator, wherein the approximation technology combination of any part is rapidly assessed and selected through the probability graph model, and on the premise that the output quality is guaranteed, the power consumption is reduced to the maximum extent. The method at least comprises the steps: converting a data flow diagram of a specified application into a Bayesian network of a probabilistic graph model according to a predefined mapping rule, the network comprising approximate configuration nodes in two states and input nodes in three states; then establishing an efficient and accurate dependency subgraph based on structure learning, and determining a conditional probability table of each node based on node parameter learning according to configuration of different approximation technologies; and according to a variable elimination inference algorithm, based on a Bayesian network and a conditional probability table, accurately solving a required marginal probability, and outputting an error distribution condition.

Description

technical field [0001] The invention relates to the technical field of a rapid evaluation method for multiple approximation techniques in a dedicated accelerator and the establishment of a probability graph model, in particular to an error assessment method for multiple approximation techniques based on a probability graph model in a dedicated accelerator. Background technique [0002] The power consumption problem of energy drives the emergence of new approximation technologies. By reducing the output quality of image processing, data mining, pattern recognition and other applications, approximate calculation, storage and communication can be greatly improved in terms of power consumption and performance. , in order to design for error-resistant application accelerators. This approximation technique can be applied to different components of a given application accelerator, but at different levels with different quality / power (or performance) tradeoffs. [0003] To maximize...

AI Technical Summary

Problems solved by technology

Although several probabilistic approaches to approximate adders or exact scaling have recently been proposed for error estimation, the hybrid use of multiple approximation techniques can better achieve an accurate quality / power trade-off, but has not been considered yet

Images