About this blog
The journey towards a path tracer
Entries in this blog
A lot of progress in this new entry! We have implemented the octree, which makes our renderer scale well with a large number of objects in our scenes. We have finished our block editor, which allows us to create, save and load worlds to render. And finally, we added some form of offline rendering.
Off-line renderer - Be sure to watch in 720p!
I've implemented a system in which you give a set of (point, direction) pairs, which the camera will be at at some point. You also specify a list of times, which indicate the time it takes for the camera to go from one pair to the next. If you now specify the number of frames per second, the system will interpolate all camera positions and directions for all frames.
Using this system, I set the camera position and direction, let it run a given number of samples per pixel, save the resulting image as a .jpeg file, and then move on to the next position and direction. In the end, I use some other program (MonkeyJam) to paste all these images together into a movie file.
The YouTube movie you see above is rendered at 1280x720, with 200 samples per pixel, at 30 frames per second.
We can move around, much like a ghost cam in some FPS. You use ASWD to move around, Spacebar and CTRL to ascend and descend, and the mouse to look around. Left mouse-click adds an object, while right mouse-click removes one.
Using other keys on the keyboard, you can choose which color the next object will have, its material type, its albedo (or brightness for a light), and its shape. You can also increase and decrease the brightness of the "skylight".
My friend did all the work on the octree. Since CUDA doesn't support recursion very well, we had to do all operations on this tree stackless.
We are happy with the results, though we did expect more speed gain from the switch to CUDA. Since we are not using textures or meshes, all we have in GPU memory is our octree, which is rather small. The time spent on memory access was rather insignificant compared to time spent on performing calculations. This made a lot of memory optimizations, which we learned about in class, not useful.
We think that the minor gains are to blame on the very branching nature of our kernels. Running one instruction on a whole lot of different data is fast, but if the instruction that we're at with the calculations for one pixel is different from another, we won't benefit from this. So whenever one ray hits a different type of material than another ray does, a different piece of code is run to sample the new direction. We think that the path tracer would be a whole lot faster if we could figure out ways to reduce the branching.
Two of the subjects I'm following at the university at the moment are Additional Component Computer Graphics and Algorithms for Massive Data. The first pretty much lets you propose a graphics related project while the second focuses on processing large amounts of data or computations efficiently. So, of course, my proposal was to implement a path tracer with CUDA. This proposal was accepted for both subjects! So now I'll be working on something I would be anyway, but now with a friend and I get credit for two courses while I'm at it. Yay!
With the knowledge gained from implementing a simple path tracer, surfing the internet and watching lectures about global illumination on YouTube (http://www.youtube.com/playlist?list=PLslgisHe5tBPckSYyKoU3jEA4bqiFmNBJ), I decided to start over with an empty project and set up a clean code structure.
Working indirect illumination. Taking things, such as the BRDF of the material and probability of sampling a given ray, into account
Cosine-weighted sampling for diffuse materials
This means that I'm pretty much just as far as I was before, but now with cosine-weighted sampling.
I'm currently trying to implement a direct lighting approach. At every ray-object intersection point p, I sample a random point on each of the lights and calculate their contribution to the lighting of point p. Given that I have not implemented any kind of fall-off of a light's intensity over distance, an image generated with just direct lighting appears far too bright. As such, the combination of direct and indirect lighting at every intersection is heavily dictated by the (incorrect) direct lighting. This can clearly be seen in the gallery album included with this post. Or, at least, I hope so. Since I've never uploaded one before.
However, I'm wondering if implementing fall-off of the light's intensity will "fix" this. Since it will make the indirect lighting image even more dark, making it contribute even less compared to the direct lighting image. But I guess we will see how it all works out.
An optimization to reduce noise might be to sample the light's shape projected onto the unit hemisphere around the normal, instead of sampling just some random point on the light's shape. Since for a point p, a large fraction of the points on, for example, a sphere are occluded by the sphere itself. This would lead to a whole lot less samples that lead to no contribution.
After direct lighting is fixed, the next thing on the agenda is... CUDA! In the next entry, I'm hoping to show off some slick videos of how awesomely fast and interactive our path tracer will be. Stay tuned!
Depth of field
My friend had added depth of field to the path tracer, which shows some nice results.
This effect was achieved by picking a focal point on the focal plane for every pixel, and then jittering our camera rays to go through this focal point.
After some significant revisions on our two reports for the two subjects for which we did this project, we are finally finished. I feel like I've learned an awful lot more about rendering, mostly due to looking at it from a different angle than the approach I'm used to (rasterization). Working on this project has been a joy for me and I'm happy with the results.
Having finished the report does not mean that we're finished with this project. We do intend to find some time to add more features. But, in reality, time might be sparse. My interests have shifted to learning how to implement these kind of effects (AA, DoF, GI) in a rasterization setting.
I hope people enjoyed having a look at this series of blog posts. Maybe there'll be more. Thanks for reading!
Yes! we've made great progress on my project's conversion to CUDA! Without any optimizations, it's already 6 times as fast as the CPU version. It's not as much as I'd hoped, though. But, hell, it's real-time(-ish) now!
No pretty pictures this time... Instead, YouTube video!
Getting the initial idea of keeping host and device well separated isn't all that hard. However, actually putting it into practice turned out to cause quite some headaches. Debugging a program with hundreds of threads running in parallel is hard, I tell ya. But now, all of the basic functionality seems to be in place.
Octree - way more objects in the scene!
Optimization - 6 times as fast as a single CPU core should be far from the limit!
Block editor - we really need to be able to place/remove objects for our assignment...
This is a journal about how a young student decided to start implementing a path tracer.
I'm Arjan, a 23 year old CSE master student from the Netherlands. Having worked for Triumph Studios [size=2](the creators of the Overlord and Age of Wonders series) and starting a one man software development company [size=2](without many projects, I can tell ya ), I do have at least some work field experience under my belt. So to spice up my humble portfolio, I decided to implement a path tracer.
Why? Because this whole physically based rendering thing seems awesome! No more cheating and hacking to approximate real world visual effects. Instead, let's try to simulate the behaviour of light as closely as possible.
So I got my hands on a copy of Physically Based Rendering: From Theory to Implementation. However, after the first 2 or 3 chapters, I felt like I was just copying how someone else had written his path tracer. I'd much rather be explained the basic concepts, implement those, see pretty pictures and then just improve from there on. This is why I put down the book and fired up the internet. I try to get familiar with the basic concepts and implement them. Whenever I'm stuck, I ask the lovely GameDev'ers around here.
Now, it's image time! Every body loves pretty images! Or, well.. my first attempt wasn't all that pretty:
Turns out I messed up my intersection code. You can imagine my joy when I fixed it and came to this:
Now, a couple days later, I had found my time to implement a mirror material and anti-aliasing. Just look at how pretty those spheres are:
Well, that's all for now! In the future, I plan to implement more features and keep the world updated through this journal. Some of those features would be:
An extra direct lighting sample instead of just one random sample in a hemisphere
More primitives, so far I only have triangles and spheres.
Camera with a lens instead of a pinhole camera
Some tree structure to speed up the intersection check
Support for participating media
Run it on the GPU, with those fancy CUDA core thingies