Resources
Highlights
Summary
In this group project, we studied the performance of a parallel ray tracing application, from CPU-based implementations to CUDA-accelerated ones.
Starting with a naïve CPU renderer, we progressively optimized the application using the right compiler flags, bounding volume hierarchies (BVH) and smarter thread scheduling strategies.
Then, we ported the C code to CUDA, a task that required the conversion of recursive functions and a completely different thread organization.
Finally, we optimized the CUDA code by introducing a tile-based thread scheduling and better data alignment, culminating in a renderer capable of achieving 65 FPS at 1080p, roughly 28 times faster than the best CPU implementation.
Pictures
An example of rendered scene
A visualization of the ray-triangle intersection process
Speedup of different compiler flags with respect to GCC’s default (CPU’s naïve version)
Speedup of every improved version we developed with respect to the first one