Skip to content

Latest commit

 

History

History
 
 

monte_carlo_ray_tracing

README.md

Graphcore

Monte Carlo Ray Tracing

Example output image

A proof of concept Monte Carlo ray tracing application.

This uses the third party repository light which in turn is based on Smallpaint by Károly Zsolnai-Fehér, both of which are released under the MIT license.

File structure

  • src/ C++ files for the Poplar application.
  • src/codelets/ IPU path tracing compute kernels.
  • src/exr/ C++ code for EXR post processing tools.
  • light/ Submodule that fetches the custom version of Smallpaint.
  • scripts/ Utility scripts.
  • CMakeLists.txt CMake build description.
  • README.md This file.

How to use this demo

  1. Prepare the environment.

Install the Poplar SDK following the instructions in the Getting Started guide for your IPU system. Make sure to source the enable.sh script for poplar.

  1. Install the apt dependencies for Ubuntu 18.04 (requires admin privileges):
sudo apt install $(< required_apt_packages.txt)
  1. Build and run the application:
git submodule update --init --recursive
mkdir build
cd build
cmake ../ -G Ninja
ninja
./ipu_trace --outfile image.png -w 960 -h 1080 --tile-width 40 --tile-height 18 --samples 1000 --samples-per-step 100 --ipus 1

This will render an image to a final sample count of 1000 paths per pixel (preview images are output at 100 samples per pixel intervals). For a full description of options: ./ipu_trace --help

The output image format is determined by the extension given to the --outfile option (support for different formats comes directly from OpenCV). In addition to the chosen output format light always saves the raw high dynamic range (HDR) output in an EXR file (e.g. image.png.exr). You can use pfs tools to manipulate and view the EXR file (e.g. to apply tone mapping: pfsin image.png.exr | pfstmo_reinhard05 | pfsout tone_mapped.png).

Running the tests

In a Python3 environment:

pip install -r tests/requirements.txt
pytest ./tests

NOTE: The test relies on Ubuntu's apt installed convert program and will not work with other versions such as the one bundled with oneAPI (see below).

Distributed rendering

The script 'scripts/distributed_render.sh' shows how to utilise many IPUs to render a high quality image. This script requires the following external dependency for the final de-noising step:

Minimal instructions for a local install of Intel's renderkit (Open Image Denoiser)

Licensing details for oneAPI can be found here: oneAPI licensing. Download installer script from: https://software.intel.com/content/www/us/en/develop/tools/oneapi/rendering-toolkit/download.html

  • Select Linux, Web & Local, Installer type online
  • Download 17MB, 733MB required install.
  • Continue as guest:
chmod +x l_RenderKit_p_2021.2.0.739.sh
sh ./l_RenderKit_p_2021.2.0.739.sh -s -a --silent --eula accept --install-dir ~/workspace/intel

Then activate it:

source ~/workspace/intel/setvars.sh

Comprehensive installation instructions can be found here: oneAPI installation guide

Running the distributed render script

Because the simple scenes used here fit in the in-processor memory of a single IPU device distribution is trivial. The script performs the following steps:

  • Compile a path tracer that uses two IPUs using the --compile-only option.
  • Launch the pre-compiled graph over 8 x 2 IPU devices (16 IPUs total).
  • Average the results on the host (jobs are launched with different seeds so the images they produce are independent samples).
  • Tone map and de-noise the combined image on the host using third party tools.

Once you have installed all dependencies then you can run the script as follows:

source <path-to-renderkit>/intel/setvars.sh
bash scripts/distributed_render.sh final.png