Microarchitectural-level simulator for parallel tile rendering on mobile GPUs
Mobile devices have led the boom in the technological segment in the recent years. They have witnessed a tremendous improvement in screen resolution and high-quality graphics because of the growing demand for playing games and other animated graphics applications. However, the demand for rendering m...
| Author: | |
|---|---|
| Format: | master thesis |
| Publication Date: | 2022 |
| Country: | España |
| Institution: | Universitat Politècnica de Catalunya (UPC) |
| Repository: | UPCommons. Portal del coneixement obert de la UPC |
| Language: | English |
| OAI Identifier: | oai:upcommons.upc.edu:2117/376125 |
| Online Access: | https://hdl.handle.net/2117/376125 |
| Access Level: | Open access |
| Keyword: | Graphics processing units Computer graphics Energy consumption GPU mòbil Graphics Pipeline Tile-Based Rendering eficiència energètica energy efficiency mobile GPU Infografia Energia--Consum Àrees temàtiques de la UPC::Informàtica::Infografia |
| Summary: | Mobile devices have led the boom in the technological segment in the recent years. They have witnessed a tremendous improvement in screen resolution and high-quality graphics because of the growing demand for playing games and other animated graphics applications. However, the demand for rendering more realistic scenes brings with it a significant increase in computation and memory bandwidth. This inevitably translates to an increase in energy consumption. Since GPUs are battery operated, energy-efficiency is an important design factor as it dictates their autonomy. In this work, we present a novel technique which we term Parallel Tile Rendering (PTR), which aims to exploit new sources of parallelism in a GPU. Under PTR, we rasterize multiple tiles in parallel using two different rasterization lanes, called Raster Units, in architectures for mobile GPUs. In this way, we dramatically reduce the required cycles for rasterization, which has been seen to be the most time-demanding process when rendering images. Experimental results show that PTR can achieve an average speedup of 83% for a wide range of different benchmarks, each of them with different characteristics. In fact, it is much more effective than having the same amount of computing resources but in a single Raster Unit, with an increase in performance of 8.3% on average. Moreover, PTR provides significant energy savings with an average decrease of 9.86%. |
|---|