以圖型應用為主的用戶端多核心嵌入式系統---總計畫 (I)

Abstract

未來的行動裝置上，決勝點將在於是否有高品質的三維圖型輸出，高效能、低耗電和適應執行環境的能力。而未來高階的行動裝置系統架構上將有數個同質多核心的CPU，搭配GPU，同時CPU本身會支援一些SIMD指令集。所以在三維圖型的處理上可以有很多不同的做法。以混合CPU/GPU來進行進階全域照明的繪圖法(結合光線追蹤與傳統rendering的優點)，不但提高圖形品質而且較容易依系統執行時的效能，網路頻寬，耗電，供電情況作圖形輸出品質和耗電的調整。在這方法上，以光線追蹤為基礎的計算可利用同質多核心的MIMD特性，而需要預先計算rasterization則交由GPU處理，或是為省電考量, 甚至可以交由CPU利用SIMD指令再配上MIMD來處理。本計劃以高品質，高適應性的三維繪圖應用為基礎, 透過垂直整合將a)創新的智慧性繪圖方法，b) 先進的耗電量測、監管、控制和優化方法，c)支援同質/異質多核心的靜態與動態編譯方法, 以及d) 可程式化的三維圖型處理器, 建立在一用戶端多核心嵌入式系統上。我們也提供高效率的系統模擬，來驗證評估上述四者垂直整合相互呼應的成效。此研究成果希望能配合台灣心(Taiwan core)計畫開發, 為國內未來嵌入式系統的發展鋪路.

For a market successful mobile device, the determining factor would be if the product has high quality 3-D graphics, crisp response, great power efficiency and highly adaptable. Future high-end mobile devices are likely to have multiple CPU cores, accompanied with a GPU, and the CPU also supports a SIMD instruction set. On such systems, 3-D graphics processing could be very versatile. We believe a hybrid CPU/GPU rendering method which combines the advantage of ray tracing and traditional rendering, can improve the quality of graphics and be adaptive to the actual working conditions of the device in terms of power consumption, power supply, available network bandwidth, and the resolution of the display. In the hybrid method, the ray tracing part can take advantage of the homogeneous multi-cores, and the pre-computation of rasterization can be dispatched to a GPU. Or to save power and energy, the pre-computation may be carried out by SIMD instructions on the CPU. Based on this graphics model, we collaboratively research on a) novel algorithms for hybrid CPU/GPU rendering, b) power monitoring, management, and adaptive optimizations, c) static and dynamic compiling techniques for homogeneous/heterogeneous multi-core systems and d) a programmable and power efficient 3-D graphics processor. We also provide a highly efficient emulation system to verify our developed applications and system software, and evaluate the effectiveness of this application-architecture co-exploration. This research will work with the design and development of Taiwan core, and wish to set a guiding line for future embedded systems developed in Taiwan.