Ray Tracer

A small path tracer implemented in C++ with both CPU and CUDA renderers. Includes diffuse, metal, and dielectric materials; recursive ray scattering; antialiasing; gamma correction; and a depth-of-field camera.

The CUDA version parallelizes rendering across pixels so each GPU thread traces and samples one pixel independently. This makes it a good project for comparing a straightforward CPU renderer with a GPU implementation of the same ray-tracing logic.

Features

• CPU renderer in main.cpp
• CUDA renderer in main.cu
• Randomized scene generation with hundreds of spheres
• Lambertian, metal, and dielectric materials
• Camera aperture and focus distance for depth of field
• Per-pixel random sampling with CURAND in the CUDA path
• PPM image output
• Makefile targets for building and profiling

Implementation Details

The CPU and CUDA implementations share the same basic rendering model: rays are generated from a camera, tested against a list of hittable objects, scattered through materials up to a maximum depth, and accumulated into a final pixel color.

The CUDA renderer uses:

• 16x16 thread blocks
• One thread per pixel
• A 1200x608 output resolution
• 100 samples per pixel
• Up to 50 ray bounces per sample
• cudaMallocManaged for the framebuffer
• Device allocations for the world, camera, object list, and CURAND states

The scene and camera are created on the GPU, then the render kernel fills the framebuffer in parallel. Each pixel gets its own CURAND state so antialiasing and material scattering can use independent random samples.

Requirements

• NVIDIA GPU with CUDA support
• CUDA Toolkit with nvcc, nsys, and ncu
• C++ compiler, configured as g++ in the Makefile
• make

The Makefile assumes CUDA is installed at /usr/local/cuda unless CUDA_PATH is set.

Building

Build the CPU renderer:

make cpuart

Build the CUDA renderer:

make cudart

On Windows, the executables are built as cpuart.exe and cudart.exe. On Linux/macOS-style environments, they are built as cpuart and cudart.

Running

Run the CPU renderer:

./cpuart

Run the CUDA renderer:

./cudart

Both renderers write a PPM image to:

render.ppm

The CUDA renderer also prints the render time to standard error.

Profiling

Build and profile the CUDA renderer with NVIDIA Nsight Systems:

make profile_basic

Collect a compute profile with NVIDIA Nsight Compute:

make profile_compute

Collect selected occupancy and instruction metrics:

make profile_metrics

Profiling reports are written to the reports/ directory. You can set a custom report tag with:

make profile_basic TAG=my_run

Cleaning

Remove generated executables and output files:

make clean