以OpenCL實現蒙地卡羅光線追蹤之加速與探討

Abstract

隨著硬體科技越來越進步，圖形處理器從固定的pipeline架構，到可以廣泛應用的通用型圖形處理器程式設計(GPGPU programming)越來越成熟，有許多開發平台都包含光線追蹤的功能，例如：DirectX Ray Tracing、OptiX、Embree 等等，利用平行化的優勢，解決光線追蹤需要的龐大計算量。 開發者在平行程式編寫上有 CUDA、OpenCL等便於平行化開發的框架，自由的開發環境使GPU kernel有多種編寫方式，在如何設計並優化GPU kernel上有許多研究，包括提升硬體利用率的方法或是同質性的計算流程等有利平行化的設計方式。 本研究以OpenCL為開發平台，探討基於物理渲染(Physically Based Rendering)的蒙地卡羅路徑追蹤法(Monte Carlo Path Tracing)具有的計算特性，並分析如何利用這些特性進一步提升平行化效率，同時考慮花費成本在追蹤的各個階段帶來的影響。With the hardware progressing, graphics processors have developed from the fixed pipeline architecture into the widely-used GPGPU programming. There are many de-velopment tools including the Ray Tracing, such as DirectX Ray Tracing, OptiX, Em-bree, … etc. These tools deal with the huge amount of calculation required for Ray Tracing with the advantage of parallelization. Developers have many ways to GPU kernels programming because of the flexible software released for facilitating parallel development, such as CUDA, OpenCL and other APIs. There are lots of studies on the designing and optimizing GPU kernels, in-cluding improving hardware utilization or homogeneity and other favorable parallel de-sign methods. This research, through OpenCL as the development tools, aimed to explore the computational features of the Physically-Based Monte Carlo Path Tracing, and to ana-lyze how to further improve the efficiency of parallelization. Meanwhile, the influence of the cost for each phase are also discussed in the research.