在可重組系統中使用動態重組排程方式增加3D顯像程式的效能

Abstract

3D顯像程式需要大量的數學運算且資料之間的平行性極高，但是每筆資料處理的時間卻不相同，使用微處理器來執行會使得效能不夠快，設計一個特定的硬體雖然可以得到最大的效能，但是會用上非常多的硬體資源。在某些有硬體資源限制的情形下，有可能會造成效能不彰亦或是根本無法在上面執行。而可重組式運算可以在有限的硬體資源下進行大規模的運算，更能使硬體結構直接切合運算的需求，達到高效能的目的。因此在本篇論文中，探討3D顯像程式如何在可重組式系統上執行，透過 1. 程式分割：不破壞最大平行性的情形下，將程式分割成幾個片段 2. 工作排程：利用資料平行性盡量重用硬體和並且減少重組的次數進而降低總執行時間 透過以上兩點，我希望可以找到一個讓3D顯像程式在可重組式系統上執行最適合的方式。

3D rendering applications have lots of arithmetic operations and high parallelism but execution times of each data are not the equal. General purpose computing is not fast enough for 3D rendering, Application specific computing (ASC) supports customization of applications in the form of hardware. Due to customization of hardware, this approach offers maximum performance for executing applications, but it will also cost a lot of hardware resource. In some situation with hardware constraint, it may not get high performance or even can not execute by using ASC. Reconfigurable computing can execute lots of computation with limited hardware resource and customize hardware circuit to fit the application needs for high performance. In this thesis, I discuss how 3D rendering application can be mapped into reconfigurable system, by: 1. Analysis of rendering process: analyze operations, parallelism and computation flow 2. DFG scheduling: reuse hardware and decrease reconfigurable frequency using data parallelism to minimize total execution time Through this, I propose a method to execute 3D rendering application on reconfigurable system.