User:JeffreyLiu/GSoC2022/Proposal

Project Name
Integrating Many Lights Sampling into Cycles X

Name
Jeffrey Liu

Contact

 * Email: jliu179@illinois.edu
 * Blender Chat: Jebbly
 * GitHub: https://github.com/Jebbly

Synopsis
The Many Lights Sampling project has been a long-standing project over the course of Blender’s development. It is meant to improve Cycles render times in scenes with many lights, so that artists may be more flexible when it comes to creating and lighting environments. Although a lot of work has already been completed in regards to the Many Lights Sampling algorithm, the goal of this project is to merge this work into the main branch after sufficient testing and user feedback.

Benefits
Past GSoC projects have shown that Many Lights Sampling can significantly improve Cycles efficiency for scenes with large numbers of lights. If it is finally merged into official releases, artists will experience drastically decreased render times. Furthermore, the project should also provide a foundation for developers to implement other algorithms that are based on Many Lights Sampling.

Deliverables
The minimum deliverable would be to merge the Many Lights Sampling project into the master branch, and thus official releases. This would include optimizing the algorithm (with specific benchmark numbers), testing the stability on a variety of scenes, and getting user feedback to improve the implementation. End-user documentation should also provide a clear explanation of how any UI options affect the algorithm. Time should be dedicated to making the UI as simple and intuitive as possible.

Project Details
A major part of this project requires merging previous GSoC Many Lights Sampling work to Cycles X. Since Cycles X introduced a rewrite of a large portion of the old Cycles code, much of the original logic must be updated to fit into the current code. This would also include adding support for the Embree implementation.

The next step of the project would be to gather and/or create more scenes to test the stability and performance of the implementation. Furthermore, more feedback on the user interface is needed in order to finalize a set of options and designs.

If time allows, optimizing and stabilizing the algorithm would take priority rather than extending the feature set. Some potential work may include investigating a bottom-up construction of the tree, finding a better method to determine the splitting threshold, or improving any of the current heuristics used. For example, the current splitting cost function doesn’t take into consideration the centroids of the bounding boxes, which may actually have a large impact on the accuracy of resulting orientation bounds. More accurate heuristics should produce a more effective tree.

Project Schedule
Since a lot of work has already been done on the Many Lights Sampling project, the scope of this project will likely be a medium project (~175 hours). I can guarantee approximately 25-30 hours per week to dedicate to this project.


 * May 20 - June 12: Read through the notes from the 2018 and 2022 projects, research any improvements that have been made to the original implementation, and find scenes that can be used to test the stability/performance of the algorithm. Also, read more about light sampling and debug how Cycles’ kernels work in detail.


 * June 13 - June 27: Finish up the initial implementation of the algorithm for the CPU, including support for Embree. Also, create a basic set of options for end-users to interface with, and reach out to developers and users to get feedback on the UI.


 * June 27 - July 11: Apply any user feedback for UI, and begin integration testing + benchmarking scenes. Include time to fix any bugs in the implementation.


 * July 11 - July 25: If the algorithm is stable enough on the CPU, start working on the GPU implementations. Also, catch up on writing any end-user documentation.


 * July 25 - August 8: Complete GPU implementations, including time for testing and benchmarking. Start investigating potential optimizations, including a bottom-up light tree construction and different split heuristics.


 * August 8 - August 22: I applied to be a SIGGRAPH student volunteer, so if I get accepted, I will likely not be available between August 8-11. In the remaining time, test and benchmark the potential optimizations used. Also, implement any requested or suggested changes.


 * August 22 - September 5: I will likely be unavailable for a few days commuting to college. Otherwise, clean up the code so that the work can be merged into main.


 * September 5 - 12: Buffer week for any remaining work. Any extra time can be spent working on polishing the end-user documentation.

Note that this was just the proposed schedule. The actual schedule is likely to change depending on the progress made and any obstacles encountered.

Bio
My name is Jeffrey Liu and I’m currently an undergraduate student at the University of Illinois at Urbana-Champaign. My major is Mathematics + Computer Science, and I’m also planning to pursue a minor in Physics as well.

I have been using Blender since middle school for renders, visualizations, and game development; its open-source software and community have nurtured my passion for computer graphics. As a result, my computer science journey has largely been oriented around graphics. I’ve done learning projects with OpenGL, Vulkan, and CUDA, and I’ve also implemented a variety of rendering techniques in my GitHub projects. Currently, I’ve been reading through PBRT to learn about photorealistic raytracing. Furthermore, I have experience contributing to large-scale C/C++ projects: I participated in Google Code-in 2019 with BRL-CAD where I began integrating the appleseed renderer with BRL-CAD’s geometry.

I’ve already gone through the “Importance Sampling of Many Lights With Adaptive Tree Splitting” paper that covers the original implementation, and I have completed some preliminary work on this patch: https://developer.blender.org/D14664. Outside of this work, I’ve also been stepping through Cycles code to understand how it works, and trying out the testing/benchmarking suite myself.