Conclusion

In this work we have implemented and analyzed different methods to optimize a volumetric path tracer. We have seen that, especially on GPU, the type of path tracing algorithm that performs best is tightly coupled with the type of scene we want to render. If the scene is simple, methods that are simple are performing better. If the scene is complex, it is better to use a different strategy and concentrate on utilization and data locality. We have proposed new variants of existing path regeneration algorithms based on different synchronization levels. Moreover, we have implemented a new sorting strategy which builds upon the idea of maintaining the ray coherence of the path tracer not only for the primary generated paths but also during scattering. This new method is performing as good as the best streaming algorithm but we believe that, using bigger and different datasets, it can lead to further improvements in time performance.

Future Work

There are many possible future directions to take based on the results showed in this work. The first step is the use of out-of-core textures and CUDA managed memory. Indeed, our work is currently limited by the available memory in the GPU. We have implemented a zero copy texture which however is poor in performance compared to the normal version. The usage of more memory for testing will permit to test bigger scenes where we believe the sorting algorithms can perform better. Moreover, the implementation can be generalized to handle any type of scenes and geometry using the state of the art intersection framework provided by . After that the algorithms should be tested on different type of hardware which can lead to more details about the different use of the cache between different GPU architectures. Finally, CUDA is not the only language which permits GPGPU computing and for this reason the implementation should be ported and tested also with other GPGPU programming languages like OpenCL.