#CUDA-based GPU Pathtracer This project implements an path tracer on GPU using CUDA. It has the following features:
- Diffuse interreflectance
- Supersampling
- Interactive camera
- Depth of field effect
- OBJ model rendering support
Recording of interactive control: http://www.youtube.com/watch?v=noNKCQpq5iA&feature=youtu.be
Result after 2000 iterations:
Result after 120 iterations:
##Implementation Details: This work was extended from my previous project, which is a CUDA-based ray tracer. To make it have a better interactivity, the tracing of an eye ray is broken into multiple kernel invocations, one for each tracing depth. Furthermore, the ray-triangle intersection test is accelerated with the use of Bounding Box.
Two ray-termination schemes are employed: Russian Roulette and predefined depth limit.
Since recursive invocations within CUDA kernels are slow, and patht racing is inherently a recursive operation, this path tracer traces the rays iteratively and builds a stack to mimic the recursive behavior.
To prevent the graphics memory from exhaustion, a predefined tracing depth limit is set.
Within the depth limit, Russian Roulette is used to determine if a ray should terminate or not. Currenlty 50% termination rate is used.
For stream compaction, an index array is constructed, each element stores the index of each pixel, that is, each eye ray.
During the path tracing, the index values of rays deemed terminated are replaced with -1, which will then be remove from the index array when stream compaction is performed. With 50% termination rate at each depht,
the travese time could be greatly reduced, especially on enclosed scenes, where rays bounce around mulitple times. When all rays are terminated or the predefined a
For better sampling efficiency, direct illumination and indirect illumination sampling are separated.
Supersampling is done by sampling each pixel four times.
A 2x2 rotated grid pattern is used to yield better results than the normal grid pattern.
Since this is a path tracer, the sampling results are averaged without special care.
Depth of field effect is realized by randomly offsetting the eye position at the start of each iteration.
##Performance evaluation: **Coupled with Russian Roulette, stream compaction can greatly reduce path traversing time, as shown in the following chart:
Stream compaction ensures that kernel invocations work on valid rays, reducing the chances of branch divergences.
##Third-party code
- GLM object loader from Nate Robins
##Third-party libraries
##Development environment
- Visual Studio 2012 on Windows 7
- How to build
Make sure the project has correct INCLUDE and LIBRARY Pathes of the above libraries.
Make sure the CUDA 5.5 is selected in the [Build Customization] Setting.
Make sure the compute_10/sm_10 compute version is remvoed from the Code Generation setting under the [0CUDA C/C++] setting
Place the needed DLL inside the execution folder.
Place testScene.scene and model teapot.obj in the execution folder.
You are good to go.