EasyNUSC: An Efficient Heterogeneous Computing Framework for Non-uniform Sampling Two-Dimensional Convolution Applications.

Non-uniform sampling two-dimensional convolution (NUSC for short) is a practical method in the field of 2D space image processing. NUSC maps sampling data of non-uniform distribution to a regular output grid through convolution. The growth rate of such data volume continues to increase, and the computational performance of NUSC is one of the key issues to be solved. Heterogeneous computing platforms provide advanced computing capabilities for accelerating NUSC performance. But heterogeneous programming and performance tuning are complex. A simple and efficient dedicated programming model and corresponding runtime framework can effectively solve such a problem. This paper proposes a parallel programming model and framework for the development of NUSC applications in heterogeneous computing environments, named EasyNUSC. When developing NUSC applications, EasyNUSC can automatically parallelize NUSC applications and perform tedious work. Developers no longer need to pay attention to the details of algorithm parallelization and task scheduling. In terms of performance optimization, this paper proposes a series of strategies in vectorization, memory access, and data reuse. The experimental data shows that EasyNUSC achieves up to 339 times the performance of a serial program within a single node, while providing excellent scalability.