高级检索

面向3D生成对抗网络的频域加速器设计

Frequency-Domain Acceleration for 3D Generative Adversarial Networks

  • 摘要: 三维生成对抗网络(3D generative adversarial networks, 3D GANs)广泛应用于模型预测、对象生成等领域. 针对当前 3D GANs 加速器存在计算开销高、访存能耗高等问题, 提出一种基于快速傅里叶变换(FFTs)的 3D GANs 频域加速器设计(FAG). 首先针对 3D GANs 加速器计算开销高的问题, 提出支持频域 3D GANs 计算的硬件架构, 利用频域低计算复杂度特性和频域反卷积中的零值模式以减少计算开销; 其次针对 3D GANs 加速器访存能耗高的问题, 提出支持 3D 卷积和反卷积计算的频域映射调度流, 利用频域共轭对称特性和反卷积中的零值模式减少访存开销. 基于ModelNet 数据集和 3 个 3D GANs 模型(3DGAN, 3D-IWGAN 和 3D-PhysNet)的评估结果表明, 与具有代表性的频域加速器相比, FAG 可使性能提高 76%, 能效提高 141%; 与具有代表性的空间域加速器相比, FAG 的性能提高 6 倍, 能效提高 46 倍.

     

    Abstract: 3D generative adversarial networks (3D GANs) are widely utilized in model prediction and object generation. To address the challenges of massive computation and significant energy consumption in accelerating 3D GANs, a novel fast Fourier transform based frequency-domain accelerator (called FAG) is proposed. Firstly, FAG provides a frequency-domain hardware architecture, which utilizes the compact computation complexity and the zero repeat pattern in deconvolution, to reduce the computation overhead in 3D GANs accelerators. Secondly, FAG exploits the frequency-domain characteristic of Hermitian symmetry and the zero repeat pattern of deconvolution to significantly reduce data movements, and it utilizes the repeat pattern in deconvolution to significantly reduce data movements. Comprehensive evaluations based on the ModelNet dataset and three 3D GAN models (3DGAN, 3D-IWGAN and 3D-PhysNet) show that, the performance and the energy efficiency can be improved by 76% and 141%, respectively, compared with the frequency-domain baseline; FAG achieves 6×higher performance and 46×better energy efficiency compared with the spatial baseline.

     

/

返回文章
返回