Project 4 CUDA A Trous Denoiser

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

• Raymond Yang
  • LinkedIn
  • Tested on:
    • 10/22/2021
    • Windows 10
    • NVIDIA GeForce GTX 1080 Ti.
  • Submitted on: 10/22/2021

Introduction

The objective of this project was to implement A Trous denoiser detailed in this paper. To construct the denoised image, we modulate a noisy path traced image (represented as a color buffer) with a geometry buffer/G buffer (represented as an array of positions and an array of normals). Both elements of the G buffer are cahced on first bounce and only done once per render.

Components of A Trous Denoiser

Scene Colors:

Scene Positions (relative to camera origin):

Scene Normals:

Scene Time to Intersect (unused):

Denoised Output

Performance Analysis

In this section, we compare the runtime costs of A Trous denoising, recorded in microseconds. For Simple Scene, we will use scene/cornell_ceiling_light.txt. For Slightly Less Simple Scene, we will use scene/demo.txt. For the purposes of consistency between runs, the following values will be hardcoded to the window unless otherwise specified:

• Filter Size = 80
• Color Weight = 8.f
• Normal Weight = 1.f
• Position Weight = 1.5f

Simple Scene

Baseline image for reference at 1000 iterations:

Path trace time for 1000 iterations: 33679.50 ms Denoise time for 1000 iterations: 62.97 ms

Path traced image after 10 iterations:

Path traced image after 10 iterations with A Trous denoiser:

Cost of Running Denoiser vs Path Tracer

Trial	1	2	3	4	5
Path Trace Time	326.05	338.88	336.49	345.53	338.01
Denoise Time	62.95	64.02	62.62	64.47	64.56

Lower is better:

Cost of Running Denoiser vs Resolution

Resolution	256x256	512x512	1024x1024
Time (ms)	4.76	20.41	116.61

Lower is better:

Lower is better:

* The increase in time is proportional to the increase in pixelcount for smaller resolutions.

Cost of Running Denoiser vs Filter Size

Filter Size	25	50	100
Time (ms)	45.66	60.41	65.53

Lower is better:

• Note that the increase in time is not substanctial indicating that I am likely calculating the number of iterations required for A Trous denoise incorrectly.

Slightly Less Simple Scene

Baseline image for reference at 1000 iterations:

Path trace time for 1000 iterations: 237650.00 ms Denoise time for 1000 iterations: 86.21 ms

Path traced image after 10 iterations:

Path traced image after 10 iterations with A Trous denoiser:

Cost of Running Denoiser vs Path Tracer

Trial	1	2	3	4	5
Path Trace Time	2431.57	2365.28	2374.86	2480.54	2387.57
Denoise Time	86.39	87.75	87.64	87.86	85.97

Lower is better:

Cost of Running Denoiser vs Resolution

Resolution	256x256	512x512	1024x1024
Time (ms)	6.61	30.41	134.44

Lower is better:

Lower is better:

* The increase in time is proportional to the increase in pixelcount for tested resolutions.

Cost of Running Denoiser vs Filter Size

Filter Size	25	50	100
Time (ms)	64.50	82.70	91.55

Lower is better:

• Note that the increase in time is not substanctial indicating that I am likely calculating the number of iterations required for A Trous denoise incorrectly.

Visual Analysis

In this section, we compare visual effects of A Trous denoising.

Visual Results vs Filter Size

Filter size is the two dimensional area around a pixel. That is, for a filter size of 25 and a pixel located at (i,j) of the image, the surrounding 5x5 grid centered around the pixel are factored into the weighting A Trous denoiser.

• Since I am using a gaussian weighting spread of 5x5, filter sizes below 25 are not considered because the area is smaller than the gaussian spread.
• Filter sizes in [25, 81), are mapped 1:1 to the gaussian spread. Filter sizes in [81, 100] are mapped 2:1 to the gaussian spread. That is, the gaussian weight at (-2, -2) relative to the pixel will actually be the pixel located at (-4, -4) relative to the pixel. Likewise, (1, 1) is mapped to (2, 2).
• NOTE: that this assumes I understand and implemented filter size correction. Given the observations above (runtime for a filter size of 50 is greater than runtime for a filter size of 25), it is very likely I have not implemented filter size correctly.

filterSize = 25:

filterSize = 50:

filterSize = 75:

filterSize = 100:

The only substantive change seems to be between filterSize = 25 and filterSize = 50.

Visual Results vs Material

From left to right, the spheres are of properties reflective, refractive (with index of 1.5), and diffuse. Scene can be found in file /scenes/cornell_ceiling_light_plus.txt.

1000 iterations:

10 iterations:

10 iterations with A Trous diffuse:

With my implementation, A Trous denoiser fails to adequately simulate refractive properties. It is moderately capable of simulating diffuse properties. It is capable of simulating reflective properties.

Visual Results vs Lighting

Small light uses scene: /scenes/cornell.txt Large light uses scene: /scenes/cornell_ceiling_light.txt

Cornell Box with small ceiling light:

Cornell Box with large ceiling light:

Cornell Box with small ceiling light and A Trous denoise:

Cornell Box with large ceiling light and A Trous denoise:

Given poorer lighting conditions, more "holes" are left in the image. As the A Trous denoiser interpolates neighboring colors for each pixel, it emphasizes darker colors leading to poorer results.