運用CUDA在多核心平台上進行平行運算模擬多層不可伸展衣服

Abstract

我們提出平行運算技術去模擬多層不可伸展布料，這模擬在電影、遊戲及衣服設計都有 廣泛應用。要有高質素模擬，就要有高解像度的布料。當布料變得複雜時，由於複雜的 互動，非常難達到互動模擬，而且當這些模擬變得複雜時，新問題就會出現。第一，隨 著三角數量增加，容量處理就要從新考慮；第二，由於有大量的碰撞事件，衣服的穩定 性難以處理，而且中途的計算值可能有溢位；第三，要檢測和處理大量碰撞事件；第四， 當三角形數量增加，它們變得愈小時，這時就要降低時間步長。現時，多核心CPU 和 GPU 的技術都有大進展，它們的計算能力變得更强。一個新的平行架構混合CPU 和 GPU，在未來會佔據市場，我們提出一個整合系統, 在這架構上處理以上四個問題。這 個系統有擴充力去模擬高解像度衣服，我們會用CUDA 去有效地運用GPU。 這個計劃在兩年內完成。 第一年， 我們會建立測試系統，在一個簡單的環境下模擬衣服。我們應用CUDA 發展 平行運算方法去計算衣服運動, 以及檢測和處理碰撞事件。 第二年， 我們會發展新技術去處理多層不可伸展布料, 在複雜的環境下模擬高解像度 衣服, 及利用不對稱方法去調節CPU 及GPU 的工作量。

We propose parallel simulation techniques for multilayered inextensible garments. The simulation of garments is important in many areas including film, game industries and garment design industries. Garments of high resolution are required in order to produce high quality simulation. As the models of garments get more complex, due to the complex interaction between garments themselves and the underlying animated characters, interactive performance is difficult to achieve anymore. There will be new problems explored when simulating high resolution garments. First, as the number of triangles increases, the storage management is an important issue. Second, the stability of the motion of garments is difficult to maintain when there are lots of collision constraints to be considered. Moreover, the intermediate computed value may be overflowed by employing the traditional techniques. Third, there will be many collision events to detect and handle. Forth, the more the number of triangles is, the smaller the size of the triangles become. Decreasing the simulation time step will increase the computation cost. To tackle the four problems, we will develop an integrated simulation system for multilayered inextensible garments on a platform that is mixed with multicore CPUs and manycore GPUs. Our system is scalable to handle garments of high resolution. We will employ CUDA in order to access GPU efficiently and effectively. This project is planned to complete in two years. In the first year, we will build a prototype system that could be scalable to handle garments of low resolution. We will also develop independent parallel techniques using CUDA for detecting collision events and computing the motion of garments on a platform mixed with multicore CPUs and GPUs. In the second year, we will develop new techniques to handle inextensible multilayered garments of high resolution. The simulation of multilayered garments will lead to many collision constraints to compute. Load balancing techniques will be employed to balance work load between CPUs and GPUs. An asymmetric approach will be employed. The multilayered garments will be dressed on animated characters.