1013

Date | Nov 07, 2014 |
---|---|

Speaker | Hiroshi Fujiwara |

Dept. | Kyoto University, Graduate School of Informatics |

Room | 선택 |

Time | 11:00-12:00 |

강의실:

We discuss an efficient parallel computation of the stationary radiative transport equation (RTE) in the three dimensions by using GPU (Graphics rocessing Unit). RTE is an integro-differential equation and it appears as a mathematical model of light propagation in human bodies. Its discretization is essentially a five dimentional large scale problem, since light intensity in RTE depends on the position and the velocity.

We employ two techniques for fast computations : GPU to accelerate calculation of the integral term in RTE, and software pipelining with OpenMP to hide transfer delays between the host memory and GPU device memory. The finite difference also appears in discretization of RTE. Its memory access pattern is quite different from that of numerical integration, and CPU processes it simultaneously by software pipelining techniques in the proposed algorithm. In the presentation, we show the current hybrid parallel implementation on a PC cluster with GPUs,

and discuss efficiency of the propolsed parallel computing.

This talk is based on a project with Prof. Y. Iso and Human Brain Research Center in Kyoto University.

We discuss an efficient parallel computation of the stationary radiative transport equation (RTE) in the three dimensions by using GPU (Graphics rocessing Unit). RTE is an integro-differential equation and it appears as a mathematical model of light propagation in human bodies. Its discretization is essentially a five dimentional large scale problem, since light intensity in RTE depends on the position and the velocity.

We employ two techniques for fast computations : GPU to accelerate calculation of the integral term in RTE, and software pipelining with OpenMP to hide transfer delays between the host memory and GPU device memory. The finite difference also appears in discretization of RTE. Its memory access pattern is quite different from that of numerical integration, and CPU processes it simultaneously by software pipelining techniques in the proposed algorithm. In the presentation, we show the current hybrid parallel implementation on a PC cluster with GPUs,

and discuss efficiency of the propolsed parallel computing.

This talk is based on a project with Prof. Y. Iso and Human Brain Research Center in Kyoto University.

TEL 02-880-5857,6530,6531 / FAX 02-887-4694