Abstract:

Dataflow programming language simplifies the domain programming and offers an attractive way to express the parallelism of mission computing and data communication on task level and data level. For the problems such as too much data parallelism, task parallelism and pipeline parallelism in multiCPU and multiGPU architectures, we propose an efficient data flow compilation framework. The framework takes the synchronous data flow graph as the beginning input, and uses many partition methods to distribute the tasks to multiCPU and multiGPU. According to the parallelism of tasks and communication, the tasks classification method assigns the tasks to GPU or CPU. We propose a GPU task horizontal splitting method to divide the tasks distributed to GPU into many blocks, and one GPU executes one block. The GPU task horizontal splitting method avoids the communication between GPU and GPU. The CPU dispersed task balancing partition method chooses appropriate CPU cores and balances the tasks distributed to these CPU cores. The method satisfies load balancing and raises the utilization rate of CPU cores. We choose a multiCPU and multiGPU heterogeneous architecture as the experiment platform and the common algorithms in media processing applications as benchmarks. Our experiments verify the effectiveness of the proposed methods.

Key words: heterogeneous architecture, dataflow program, task partition, storage optimization