- Yu Mao, Robotics Institute, School of Computer Science, Carnegie Mellon University (ymao1@andrew.cmu.edu)
- Rui Zhu, Robotics Institute, School of Computer Science, Carnegie Mellon University (rz1@andrew.cmu.edu)
Also submitted to the afs cloud at
/afs/cs/academic/class/15462-f15-users/ymao1/asst5/
Warning: Before compiling the program, make sure you have:
- CUDA compatible Graphics Card
- CUDA SDK installed
To test Task 1, the BVH build thing, you can type in
./pathtracer ../dae/sky/CBbunny.dae
and hit v for BVH construction (use classic path tracing by default. See parameter -p below).
To test Task 2, you can type in (for example to test the CBsphere (and lambertian) case)
./pathtracer -t 18 -s 16 -p 1 ../dae/sky/CBspheres.dae
and hit r for BDPT rendering.
The -p parameter is a switch for BDPT or classic path tracing (1 is BDPT; 0 for classic path tracing (default)).
Again notice that only AreaLight models (the Cornell Box series) are supported in our little BDPT demo version.
We complemented two tasks according to the proposal:
- Parallel BVH Construction with CUDA Acceleration
- Bidirectional Path Tracing. (AreaLight cases supported)
For the first task, we first implemented the Morton code based method which only runs on CPU. Then we implemented the parallel binary radix tree(BRT) construction algorithm as proposed in Karras, Tero. "Maximizing parallelism in the construction of BVHs, octrees, and k-d trees." for general BRT construction. Thirdly, we implemented the parallel algorithm for BVH construction based on BRT construction algorithm, which augments it by adding a bottom-up parallel bounding box construction phase. We also compared the efficiency of different approaches.
A demo video about our blazing fast BVH construction is available on Youtube (Click the thumbnail below to show):
We are thrilled to see those complex models rendered which would otherwise take hours to render in the past.
Some constructed BVHs and rendered images, as well as analysis of the performance are listed below:
For the second one, we modified the renderer used in Homework 3, which uses traditional path tracing, to one which uses Bidirectional Path Tracing (BDPT). The new renderer is able to render scenes with AreaLight using BDPT with better illumination and variation in shadowed areas. We compared the efficiency of there two methods, as well as the ``photorealisticity" of the rendered images under these two methods. Our BDPT gives better illumination in the soft shaded area, and less variance in the caustic focused light under a glass sphere.
This is a comparison of our renderer and the classic path tracing:
There is also a really cool screen record of our BDPT renderer (Click the thumbnail below to show)
For the last assignment, we are going to choose option F: Advanced Monte Carlo Rendering. We will implement sub-options 1 to 4.
We would improve the BVH construction process by refering to Fast Parallel Construction of High-Quality Bounding Volume Hierarchies. We would implement the method in the previous system for assignment 3.
We would conduct evaluation on the system to compare the efficiency before and after the improvement.
##Task 2: Improve performance of ray tracing
We would refer to Ray Tracing Deformable Scenes using Dynamic Bounding Volume Hierarchies and implement the fast packet-based ray tracing using.
We would conduct evaluation on the system to compare the efficiency before and after the improvement.
We would like to implement bidirectional path tracing to help find the best paths and reduce the cost. Maybe We will consider using parallel computing skills to further improve the efficiency.
We should be able to output the variance of the estimator and the total time cost, and make a comparison with the old routine.
We may refer to the PBRT book for details about the algorithm.
We would use blue noise (maybe adaptive blue noise) and stratified sampling to improve the sampling pattern. We will compare the estimated variance of the estimator and the time consumption.