This repo contains the course project for CMU 15-618: Parallel Computer Architecture and Programming Fall 2023. We adatped a CPU path tracer from Yocto/GL to CUDA and extended it with the wavefront design and the wide BVH from the following papers:
- Megakernels Considered Harmful: Wavefront Path Tracing on GPUs
- Efficient Incoherent Ray Traversal on GPUs Through Compressed Wide BVHs
We compare the performance of our wavefront path tracer, the original megakernel version in CUDA (without using OptiX), and the multithreaded CPU version. Our implementation achieves a speedup of 1.02x - 1.79x comparing to the megakernel CUDA version, and 1.49x - 8.20x comparing to running on a 16-core CPU.
First, check the original repo for requirements. By default, cmake .. will generate configuration for the wavefront version. To generate the megakernel version without OptiX, run
cmake -DWAVEFRONT=OFF ..To generate the original CUDA version using OptiX, run
cmake -DWAVEFRONT=OFF -DCUSTOM_CUDA=OFF ..The target names are wavefront_trace, cuda_trace, and ycutrace, respectively.
To use the wide BVH, specify --wbvh in the command line. To use the three-stage (logic, material, ray cast) wavefront pipeline instead of the two-stage one (logic, ray cast), add --matstage to the command for wavefront_trace.