In this project, I have implemented a CUDA based raytracer utilizing the base code from the course CIS565 at University of Pennsylvania. In the base code, I was provided with functions such as file I/O, OpenGL setup, base CUDA kernel set up, etc. The following are the features that I implemented.
- Raycasting from a camera into a scene through a pixel grid
- based on the camera configuration, a ray is cast through each pixel and intersected with each object in the scene.
- Phong lighting for one point light source
- Given an arbitrary number of light sources, Phong shading model is applied to each surface attenuated with the distance away from the light source.
- Diffuse lambertian surfaces
- Raytraced shadows
- Each light source will be used to cast shadows.
- Cube intersection testing
- Ray-Cube intersection with slabs
- Sphere surface point sampling
- Randomly sampling a point on a sphere utilizing thrust.
- Supersampled antialiasing
- Comparision:
- Comparision:
- Specular highlights
- Simple mirror reflection
- Rays are reflected from reflective surfaces and samples the color of the object that is intersected.
- Soft shadows.
Work in Progress:
Video: http://youtu.be/Tnpnndr-x28
Over the course of the project, I have attempted several experiments in order to speed up the run time of my program.
-
Dividing large kernels into smaller kernels.
- While most of the core ray tracing algorithm is still in one kernel, I have taken out some of the more minor calculations and launched as separate kernels. Unfortunately, I did not notice a major speed up as a result of this process.
-
Changing the number of threads per block.
- Since kernels issue instructs in warps with 32 threads, I decided to change the number of threads per block to 32. There was a slight improvement in run time. With 8 threads per block, I was getting around 4 frames per second and with 32 threads per block, it went up to around 5.45 frames per second.
Ideas that I did not get to try.
- Have each of the super sampling rays be separate threads as well.
- Accumulate shadow ray colors across multiple iterations instead of casting multiple shadow rays per iteration.