Distributed Reconfigurable Hardware for Image Processing Acceleration

Lately, the use of GPUs is dominant in the field of high performance computing systems for computer graphics. However, since there is "not good for everything" solution, GPUs have also some drawbacks that make them not the best choice in certain scenarios: poor performance per watt ratio, difficulty to rewrite code to explode the parallelism and synchronization issues between computing cores, for example. In this work, we present the R-GRID approach based on the grid computing paradigm, with the purpose of integrating heterogenous reconfigurable devices under the umbrella of the distributed object paradigm. With R-GRID the aim is to offer an easy way to non experience hardware developers for building image processing applications using a component model. Deployment, communication, resource sharing, data access and replication of the processing cores is handled in an automatic and transparent manner, so coarse grained parallelism can be exploited effortless in R-GRID, accelerating image processing operations.