High Throughput Hardware Acceleration for Image Generation using HLS
A Bhanu Prasad, Kuruvilla Varghese · 2023
New machine learning techniques have been developed and used to generate new data in recent years. Genera-tive adversarial networks is a neural network model that was developed by Ian Goodfellow in 2014. GAN consists of two neural networks, a generator, and a discriminator, that compete to improve the generator's ability to produce realistic outputs. The generator network tries to generate new data, whereas the discriminator network tries to detect whether the data is real or fake. Both networks compete with each other intending to generate new and unique data. In image processing, convolutional neural networks, generative adversarial networks, and numerous other advanced neural networks require numerous multiply and accumulate units, which necessitates increased processing power. In our work, we implement the conditional generative adversarial network, which can generate new images from the class label input in the hardware. As a case study, we implement the Fashion-MNIST dataset. We implement the generator network and random number generator using High Level Synthesis on the ALVEO U250 data center acceleration card. This work achieves a throughput of 252 mega frames per second. Our work presents the CPU-FPGA system utilizing the Xilinx Vitis tool flow, including end-to-end integration and hardware-specific kernel optimizations.