http://swrc.ontoware.org/ontology#InProceedings
SAT Based Formulation of Automatic Generation of Parallel Computing from Specification
en
回路設計・評価
University of Tokyo
University of Tokyo
University of Tokyo
University of Tokyo
Gao Ruitao
Amir Masoud Gharehbaghi
Tomohiro Maruoka
Masahiro Fujita
In recent years, methods using deep learning have been widely used in various fields. And It is known that a large portion of computation time in deep neural network is taken by matrix multiplication. There is close connection between neural network and matrix multiplication. In this paper, parallel computing solution for matrix-vector multiplication on certain ring-connected cores is automatically generated. The basic method is to formulate matrix-vector multiplication on ring-connected architecture as a SAT problem and use SAT solver to get the mapping solution. According to the experiment results, parallel computing solution of 16x16 matrix can be generated in short time. Moreover, solutions for sparse matrix multiplication can be generated.
In recent years, methods using deep learning have been widely used in various fields. And It is known that a large portion of computation time in deep neural network is taken by matrix multiplication. There is close connection between neural network and matrix multiplication. In this paper, parallel computing solution for matrix-vector multiplication on certain ring-connected cores is automatically generated. The basic method is to formulate matrix-vector multiplication on ring-connected architecture as a SAT problem and use SAT solver to get the mapping solution. According to the experiment results, parallel computing solution of 16x16 matrix can be generated in short time. Moreover, solutions for sparse matrix multiplication can be generated.
DAシンポジウム2019論文集
2019
119-124
2019-08-21