Shading Language Compiler Backend for a Mobile GPU: Case Study on Register Allocation and Code Scheduling

• Main Authors: Jian-Hao Su, 蘇建豪
• Other Authors: Shin-Wei Liao
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/13043048243897298472

碩士 === 國立臺灣大學 === 資訊網路與多媒體研究所 === 97 === 3D computer graphic has become more and more important today, applications like games, virtual reality and so on prevail the way for 3D computer graphic. Accelerator for 3D Graphic has become a necessary component in computer and mobile device. GPU is the new name of graphic hardware since it has taken nearly all workload of 3D graphic processing form CPU. As the evolution of GPU, programmability is now provided for more complex computation. Shading language is the key to release the power of GPU programmability. It provides the high hardware abstraction and more flexibility programmability rather than programming GPU in low-level hardware related assembly. Shading language compiler translates shader codes to low level instructions. We have developed an OpenGL ES shading language compiler for a mobile GPU developed by Media IC & System and DSP IC Lab at NTU. In this thesis, we describe the compiler we have done and study the backend of shading compiler for the mobile GPU hardware. Optimization methods are proposed to improve the performance of current hardware. Code scheduling in compiler detects data forwarding and generates code to use hardware forwarding path. Register allocation supports packing to use remain part of vector type in shading language. In current experiment, at most 28% register usage can therefore shrink.