Introduction

Motivation

In the last decade, graphics processor units have emerged as a cheap and powerful data parallel computational platforms. The architectural innovations, in such processors, have transformed the GPU from fixed-function hardware blocks to programmable units for general purpose computations (GPGPU). The gap between floating-point capability of GPU versus CPU is increasing on a year base and other hardware trends encourages to parallelize the existing serial algorithms.

The evolution of new programming languages, like CUDA- Compute Unified Device Architecture- and OpenCL, gives to the programmer more flexibility in the usage of the GPU processors for general purpose computing. This processing power can be used today to address many of the computational problems, not solvable in adequate time with the previously available resources. The task we are going to focus on is realistic image synthesis of participating media. Most of the light effects, visible in nature, can be described by this type of media. Indeed, those are usually used to describe the property of a generic material. Realistic rendering of those materials requires physical simulations of the light interacting with the media. Those interactions can be thousands or millions based on the type of participating media we are going to simulate. For this reason, simulation of the light can take a long time. Efficiently, using GPU to compute in parallel, those light interactions can improve the time performance to get the final render.

Possible Applications

Realistic image synthesis of solid object interests a wide range of areas. In architecture and product visualization can be used to present a building or a product. In virtual prototyping can be used to predict the appearance of an object before creating it and modifying it in according to that. Realistic Real time rendering is used inside many of the most famous games. Another possible application comes from volume visualization, where realistic rendering plays a central role in 3D shape perception. Most of the available implementation, in this sense, focuses more on performance, rather than perfect light transport simulation. The original idea for this project comes from the 3D printing field. In a recent paper   have used volumetric rendering for accurate prediction of texture colors in 3D printing. This prediction stage is the biggest performance bottleneck of all the pipeline. The reason is that print materials have usually high optical density and simulating the appearance of those materials requires computing thousands of scattering interactions.

This work aims to give a comprehensive view on how to optimize a volumetric path tracing for the GPU architecture, taking in consideration the advantages and disadvantages of different design choices. The solutions found can be used inside a completely GPU based renderer or as a stage of longer pipeline for rendering.

Thesis Structure

The thesis has six chapters:

• the first chapter describes the problem that the thesis aims to solve.

• The second chapter shows the work that has already been done on solving this problem.

• The third chapter describe the theoretical background necessary for the methods that the thesis implements.

• The fourth chapter shows the method implemented and the different behavior of each method.

• The fifth chapter describes some implementation details on the solution adopted and the software created.

• Finally, the sixth chapter shows some results obtained with the GPU methods implemented and a comparison with a standard CPU algorithm.