Parallelized Ray Tracing In One Weekend

The parallelized version for the book, GitHub v3.2.3, Ray Tracing in One Weekend by Peter Shirley.

TL;DR

Optimize and Benchmark Ray Tracing code with CUDA, OpenMP.

Reports

• Final Presentation
• Final Report
• Project Proposal

Samples

Sample 10 times per pixel

Sample 500 times per pixel

Benchmark results

OpenMP

CUDA

Map size (Number of objects)

Environment

Hardware

• CPU: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz (10 cores 20 threads)
• GPU: NVIDIA RTX 2080Ti 12GB
• RAM: DDR4 2666MHz 128GB

Software

• Host OS: Ubuntu 20.04
• Container: Docker engine 24.0.6 with rootless mode
• Guest OS: Ubuntu 22.04
• Linux kernel: 5.15.0-87-generic
• OpenMP: 4.5
• CUDA: 12.3
• GCC: 11.4.0

How to build?

export CMAKE_BUILD_TYPE="Release"

# build all implementations with the best settings
CUDAFLAGS="-DCUDA_BLOCK_SIZE=8" cmake -B build . && cmake --build build

How to run?

# run serial implementation
./build/serial/InOneWeekend > test.ppm

# run OpenMP implementation
OMP_NUM_THREADS=22 ./build/openmp/OMPInOneWeekend > test.ppm

# run CUDA implementation
./build/cuda/CUDAInOneWeekend > test.ppm

Benchmark

Notice

These scripts are for reference only. If you want to reproduce the result, the OpenMP implementation is much more tricky. You have to change the scheduling method by yourself in openmp/src/main.cc.

Before running

We provide the devcontainer to quickly set up the enviroment (for rootless docker only). If you want to set up by yourself, you can follow the Dockerfine in .devcontainer folder to install.

OpenMP

./scripts/benchmark_omp.sh

CUDA

./scripts/benchmark_omp.sh

Map size

./scripts/benchmark_map_size.sh