CUDA Denoiser For CUDA Path Tracer

University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 4

• Shubham Sharma
  • LinkedIn, personal website.
• Tested on: Windows 10, i7-9750H @ 2.26GHz, 16GB, GTX 1660ti 6GB (Personal Computer). *GPU Compute Capability: 7.5

Project Overview

In this project, I implemented a pathtracing denoiser based on "Edge-Avoiding A-Trous Wavelet Transform for fast Global Illumination Filtering" by Dammertz, Sewtz, Hanika, and Lensch. The Denoiser uses geometry buffers (G-buffers) which is convoluted with the original Ray Traced output to provide a smooth result and therby acting as smoothing filter.

Algorithm Overview

• The input is a path-traced image,

• Normal buffer

• Position buffer

The algorithm uses Gaussian filter with increasing step width for multiple iterations to get a denoised result.

Performance Analysis

The following figures show the renders with different denoising iterations.

All the Images Use The Following Parameters

• Filter Size = 25
• Color Weight = 8.1000
• Normal Weight = 0.350
• Position Weight = 0.700

Reference Image (1000 Iterations)

10 Iterations

• Reference

• Denoised

2 Iterations

• Reference

• Denoised

Performance Numbers

Cost Denoiser vs Path Trace

• Filter Size = 80
• Color Weight = 8.1000
• Normal Weight = 0.350
• Position Weight = 0.700

Denoiser vs Resolution

• Filter Size = 80
• Color Weight = 8.1000
• Normal Weight = 0.350
• Position Weight = 0.700

This demonstrates that the average time increases linearly with the number of pixels in the image.

Denoiser vs Filter Size

The À-Trous filter scales in power of 2 which aligns with the demonstration of this graph curve which behaves as a logarithmic curve.

Notable Findings

• How visual results vary with filter size.

• • In my Findings the visual results do not change linearly with filter size. As the filter size increase the difference in results become less percieveable

• how do results compare across different scenes - for example, between cornell.txt and cornell_ceiling_light.txt. Does one scene produce better denoised results? Why or why not?

• • A well Lighted scene will have

Optimisations

G-Buffer optimization

• Optimising per-pixel positions and normals which are stored as glm::vec3s results in smaller storage buffers size and access times. You can see the performance improvements from charts and numebrs below

• • positions are now reconstructed based on pixel coordinates using an inverted projection matrix.
• • Normals are compacted from glm::vec3 to glm::vec2 saving 4byte per float value stored.

• Optimised Charts

Trial # 1 2 3 4
Path Tracer 80.135ms 82.127ms 84.653ms 85.317ms
Denoiser 13.0652ms 13.157ms 13.265ms 13.59ms

• UnOptimised Charts

Trial # 1 2 3 4
Path Tracer 87.6103ms 89.3156ms 90.134ms 90.9634ms
Denoiser 9.5657ms 9.7115ms 9.7557ms 9.7799ms

Comparing A-trous and Gaussian filtering

• A-trous

• Simple Gussian Blur

As we can see the A-tours filter does a much better job in preserving the edges which almost seem too fade in simple Gaussian Blur.

Filter Size	A Trous	Simple Gaussian
25	8.5991	1.3322
80	13.590	3.1142