high performance computing on graphics processing units: hgpu.org

CPU has insufficient resources to satisfy the efficient computation of the Convolution Neural Network (CNN), especially for embedded applications. Therefore, heterogeneous computing platforms are widely used to accelerate CNN tasks, such as GPU, FPGA and ASIC. Among these, FPGA can accelerate the computation by mapping the algorithm to the parallel hardware instead of CPU, which cannot fully exploit the parallelism. By fully using the parallelism of the Neural Network’s structure, FPGA can reduce the computing costs and increase the computing speed. However, the development of FPGA requires designers’ great design skills. As a heterogeneous development platform, OpenCL has some advantages such as high abstraction level, short development cycle and strong portability, which can make up for the lack of unskilled designers. This paper uses Xilinx SDAccel to realize the parallel acceleration of CNN task, and it also proposes an optimizing strategy of single convolutional layer to accelerate CNN. Simulation results show that the calculation speed could be improved by adopting the proposed optimizing strategy. Compared with the baseline design, the strategy of single convolutional layer could increase the computing speed 14 times. Performance of the whole CNN task could be improved 2 times than before, and the speed of image classification could attain more than 48 fps.