CUDA Ray Tracing in One Weekend

Notes

Bugs

• dielectric does not seem to be refracting/reflecting correctly; changing precision does not work, so it must be some implementation issue...somewhere...

Float vs. Double

Runtime ~5s for float vs. ~35s for double.

ppm_diff

ppmdiff takes three .ppm files as arguments, e.g., input1.ppm, input2.ppm, and output_diff.ppm. It calculates the differences between the pixels in input1.ppm and input2.ppm and writes it to output_diff.ppm. Minor differences between the images thus result in _darker areas in the output image, while larger differences result in brighter areas in the output image.

Usage: ./ppm_diff base_image.ppm ./src/DoubleCUDAInOneWeekend/cuda_image.ppm diff_image.ppm

Refactoring Motivation

A great reference and starting point was Roger Allen's Accelerated Ray Tracing in One Weekend in CUDA. The code certainly still works, but some thing don't fully align with the most recent Ray Tracing in One Weekend book. That was one reason to refactor, aside from just better understanding every part of the code myself.

A more technical reason, however, was that Roger Allen had chosen to create the world using one thread. I felt that a more intuitive design would be creating the world on the host, and transferring it to the device. As of writing (April 17, 2025), simply adding __device__ qualifiers where appropriate to the base version and transferring host-created hittable objects (its children, really) to the device would not work, as "It is not allowed to pass as an argument to a global function an object of a class with virtual functions.".

As a result,I ended up refactoring to a more C-like style, where I use a function-based approach and structs, ridding much of the polymorphism, inheritance, abstract classes, etc. concepts of C++. Well, at least for the hittables (hittable.h).

Relevant Sections from the Original README

Optimized Builds

CMake supports Release and Debug configurations. These require slightly different invocations across Windows (MSVC) and Linux/macOS (using GCC or Clang). The following instructions will place optimized binaries under build/Release and debug binaries (unoptimized and containing debug symbols) under build/Debug:

On Windows:

cmake -B build
cmake --build build --config Release  # Create release binaries in `build\Release`
cmake --build build --config Debug    # Create debug binaries in `build\Debug`

On Linux / macOS:

# Configure and build release binaries under `build/Release`
cmake -B build/Release -DCMAKE_BUILD_TYPE=Release
cmake --build build/Release

# Configure and build debug binaries under `build/Debug`
cmake -B build/Debug -DCMAKE_BUILD_TYPE=Debug
cmake --build build/Debug

We recommend building and running the Release version (especially before the final render) for the fastest results, unless you need the extra debug information provided by the (default) debug build.

CMake GUI on Windows

You may choose to use the CMake GUI when building on windows.

1. Open CMake GUI on Windows
2. For "Where is the source code:", set to location of the copied directory. For example, C:\Users\Peter\raytracing.github.io.
3. Add the folder "build" within the location of the copied directory. For example, C:\Users\Peter\raytracing.github.io\build.
4. For "Where to build the binaries", set this to the newly-created "build" directory.
5. Click "Configure".
6. For "Specify the generator for this project", set this to your version of Visual Studio.
7. Click "Done".
8. Click "Configure" again.
9. Click "Generate".
10. In File Explorer, navigate to build directory and double click the newly-created .sln project.
11. Build in Visual Studio.

If the project is succesfully cloned and built, you can then use the native terminal of your operating system to simply print the image to file.

Running The Programs

You can run the programs by executing the binaries placed in the build directory:

build\Debug\inOneWeekend > image.ppm

or, run the optimized version (if you compiled with the release configuration):

build\Release\inOneWeekend > image.ppm

The generated PPM file can be viewed directly as a regular computer image, if your operating system supports this image type. If your system doesn't handle PPM files, then you should be able to find PPM file viewers online. We like ImageMagick.