cuda-pathtrace

cuda-pathtrace is a realtime photorealistic pathtracer implemented in CUDA. It can currently only render diffuse surfaces and scenes containing spheres.

By default, cuda-pathtrace outputs many features which can be used in a denoising algorithm, such as color, surface normals, albedo/texture, depth, along with per-pixel variances for each feature.

cuda-pathtrace's low-sample renders are automatically fed through a deep learning denoising algorithm to provide an interactive realtime photorealistic experience. See the denoise_cnn/ folder for current status of the denoising algorithm experimentation.

Installation

Prerequisites

You will need the following in order to build the application:

• NVIDIA CUDA Toolkit 8.0
• glm
  • Just copy the glm folder to your include path (/usr/include)
• GLFW
  • Clone the repo and cd to the folder
  • Run cmake -DBUILD_SHARED_LIBS=ON .
  • Run make
  • Run sudo make install
  • Copy /usr/local/lib/libglfw.so.3 to your cuda-pathtrace repo, or add /usr/local/lib/ to your LD_LIBRARY_PATH
• Boost (boost_python, boost_program_options, boost_system)
• Python 2.7
• Pytorch for Python 2.7 and CUDA 8.0

Compiling

Type make to build the application (Tested on Ubuntu 16.04).

Usage

To launch cuda-pathtrace in interactive (real-time) mode with denoising enabled, use:

./pathtrace -i -d

To render single frames, cuda-pathtrace accepts the following arguments. The output file .exr will be created with the results of your render.

Options:
  --help                         Print help messages
  -t [ --threads-per-block ] arg Number of threads per block in 2D CUDA
                                 scheduling grid.
  --size arg                     Size of the screen in pixels
  -s [ --samples ] arg           Number of samples per pixel
  --device arg                   Which CUDA device to use for rendering
  -d [ --denoising ]             Use denoising neural network.
  -i [ --interactive ]           Interactive mode - will render single frame
                                 only if not set.
  --nobitmap                     Don't output bitmaps for each channel
  -o [ --output ] arg            Prefix of output file/path
  -x [ --camera-x ] arg          Starting camera position x
  -y [ --camera-y ] arg          Starting camera position y
  -z [ --camera-z ] arg          Starting camera position z
  -c [ --camera-yaw ] arg        Starting camera view yaw
  -p [ --camera-pitch ] arg      Starting camera view pitch

Using the Rendered Features

cuda-pathtrace outputs a multilayered OpenEXR file containing all of the necessary features to train a deep learning denoising algorithm.

In Python

To load the rendered image and features from the EXR file, the python code in denoise_cnn/load_data.py can be used via:

from load_data import load_exr_data

x = load_exr_data("output.exr", preprocess=True)

You will need to install the OpenEXR python bindings. If you are using Windows, I recommened installing from an unofficial precompiled binary - it will make your life 10x easier.

Built With

• C++
• CUDA
• glm - Mathmatics functions for Camera class
• GLFW
• OpenGL
• Pytorch
• Python
• Boost
• tinyexr - Saving OpenEXR files
• stbimage - Saving bitmaps
• Cornell box scene from http://www.kevinbeason.com/smallpt/