抄録
An implementation of particle filter algorithm on Cell Broadband Engine in parallel way have been proposed. Major problem for the implementation is resampling procedure of the algorithm, which is difficult to parallelize. Another problem is Local Store (LS) of computational core,which is called SPU, being limited to 256KB.
Further subjects are effective use of Single Instruction Multiple Data (SIMD) operations,use of dual operation issuing property,
and double buffering in data transfer. We have tried to implement the algorithm following the policy below;1) implement full algorithm without any simplification due to the parallel implementation,2) hold generality of particle filter algorithm,3) store particles separately in LS of SPUs,4) all particles are transferred to Main Storage (MS) after resampling step,and 5) SIMD operations are effectively used in each part of the algorithm. We have used Mailbox facility for synchronization of SPUs and main CPU,which is called PPU. Direct Memory Access transfer has been employed for data transfer between LS and MS. Performance improvement of the parallel implementation has been examined through some experiments of performance evaluation.
