Markov Monte Carlo Algorithm acceleration method based on FPGA (field-programmable gate array) heterogeneous platform

Abstract

The invention discloses a Markov Monte Carlo algorithm acceleration method and system based on an FPGA (field-programmable gate array) heterogeneous platform. The method comprises: allowing a host to finish initialization, and transmitting generated random number data to a DDR (double data rate) memory of an FPGA terminal through a PCI-e (peripheral component interconnect express) interface, wherein the initialization includes constructing a probability space and generating random numbers; allowing the FPGA terminal to read the random number data in the DDR memory, and using a constructed Markov chain to perform concurrent pipeline operation on the random number data, and acquiring and transmitting corresponding result data to the host. The random number data in the DDR memory are read via the FPGA terminal, and the constructed Markov chain is used to perform concurrent pipeline operation on the random number data; it is possible to use the constructed Markov chain in an FPGA board to finish a core computing process of the Markov Monte Carlo method in a concurrent and pipeline manner, the cost is reduced, and concurrency efficiency is improved.

Description

technical field [0001] The invention relates to the technical field of computer applications, in particular to a Markov Monte Carlo algorithm acceleration method and system based on an FPGA heterogeneous platform. Background technique [0002] With the development of computing technology, there are more and more application scenarios of Monte Carlo algorithm, such as molecular dynamics, quantum computing, computational chemistry, etc. There are many ways to implement Monte Carlo algorithm, among which the method based on importance sampling (Metropolis sampling) Markov Chain Monte Carlo (MCMC) is the most widely used method. The basic theory of MCMC is the Markov process. The integral function is as similar as possible. In order to sample on this distribution, it is necessary to start from any state, continuously perform state transitions, and finally converge to a stable distribution. After stabilization, weighted average can be done. As the complexity of computing scenar...

AI Technical Summary

Problems solved by technology

The parallelism of Markov chains, the special structure of MCMC is especially suitable for parallel computing. Each thread (or process) can be used to calculate a Markov chain, and there is no need for communication between each Markov chain, so the parallel efficiency is relatively high. High, the problem with this method is that because each Markov chain needs to be preheated (the first sample needs to be removed), it will cause a waste of computing resources, and the degree of parallelism is proportional to the number of cores. This method relies on large Scale computing technology (especially cluster), so the cost is high; the parallel of numerical calculation only generates a Markav chain, and the data calculation part is parallel. This method has a small warm-up cost, but requires communication, so the parallel efficiency is low

Moreover, the degree of parallelism also depends on the number of cores, which is also based on clusters, so the cost is also high

Images