User-Centric Architecture Design for Computer Graphics

Over the past decades, the vast improvement in graphics technology makes graphical applications able to deliver photo-realistic scenes and provide immersive user experience. The recent studies show that the performance of 3D rendering in a Graphics Processing Unit (GPU) that converts a 3D vector stream into a 2D frame with 3D image effects significantly impact users' experience with modern computer graphics.

Aiming to improve rendering performance, I observe that the off-chip memory bandwidth becomes a critical obstacle due to the high texture throughput in 3D rendering. In this study, I first introduce two architectural approaches to reduce the texture throughput by enabling Processing-In-Memory (PIM) and perception-oriented approximation in GPU for efficient 3D rendering, respectively. In the first work, I enable emerging 3D stacked technology for 3D rendering to speed up the texture filtering process. In the second work, I propose a perception-oriented hardware approximation model by leverage the inner-relationship between anisotropic and isotropic filtering. The experiment results show the two architecture designs improve the overall rendering performance by 43% and 17%, respectively.

On the other hand, Virtual Reality (VR) which was invented back in the 1990's is now experiencing a rapid growth. Compared with conventional graphics applications, VR promises a fully immersive experience, which poses an even more critical requirement for graphics hardware. In the high interactive VR world, the image display latency highly degrades this illusion of stability and has been identified as a major challenge of existing VR systems. In order to alleviate this negative effect, asynchronous time warp (ATW) has been proposed by VR vendors to map the rendered stereoscopic frame in the correct position using the latest head-motion information. In this dissertation, we propose a preemption-free Processing-In-Memory based ATW design which asynchronously executes ATW within a 3D-stacked memory, without interrupting the rendering tasks on the host GPU.