WIP: Cycles: implement faster GGX VNDF sampling #109757

Weizhen Huang · 2023-07-05T19:57:47+02:00

Weizhen Huang commented

2023-07-05 19:57:47 +02:00

from paper Sampling the GGX Distribution of Visible Normals at HPG 2023

The function itself uses less cross(), inversesqrt() and sqrtf(), is claimed to be 2.5x faster than the original function and take 3~7% less time in a simple test scene.
However, it doesn't seem to bring any visible speedup in Cycles.
Besides, the original implementation benefits from concentric disk sampling, whereas the new version does not. We could use sample_uniform_cone_concentric() to sample the spherical cap, but that doesn't have low distortion in the lower hemisphere.
Below is a visualization of mapping regular grid points to half vectors using different methods.

old	new	new with concentric

It is not clear if we want to use the new sampling method or not, but still create a pull request for the record.

from paper [Sampling the GGX Distribution of Visible Normals](https://diglib.eg.org/bitstream/handle/10.1111/cgf14867/v42i8_03_14867.pdf) at HPG 2023 The function itself uses less `cross()`, `inversesqrt()` and `sqrtf()`, is claimed to be 2.5x faster than the original function and take 3~7% less time in a simple test scene. However, it doesn't seem to bring any visible speedup in Cycles. Besides, the original implementation benefits from concentric disk sampling, whereas the new version does not. We could use `sample_uniform_cone_concentric()` to sample the spherical cap, but that doesn't have low distortion in the lower hemisphere. Below is a visualization of mapping regular grid points to half vectors using different methods. |old|new|new with concentric| |--|--|--| |![before.png](/attachments/ae94db55-4e27-47f8-9100-ba287456fcb2)|![after.png](/attachments/83712c36-61ed-4525-badc-4571f550b597)|![after_concentric.png](/attachments/7342f416-43b2-4ed7-8822-2e59451510c7)| It is not clear if we want to use the new sampling method or not, but still create a pull request for the record.

after.png

80 KiB

after_concentric.png

87 KiB

before.png

85 KiB

Weizhen Huang added the

Module

Render & Cycles

label 2023-07-05 19:57:47 +02:00

Weizhen Huang added 1 commit 2023-07-05 19:57:59 +02:00

Cycles: implement faster GGX VNDF sampling 0cb8451176

from paper "Sampling the GGX Distribution of Visible Normals"

Weizhen Huang changed title from ~~Cycles: implement faster GGX VNDF sampling~~ to WIP: Cycles: implement faster GGX VNDF sampling

2023-07-06 13:04:25 +02:00

Weizhen Huang commented

2023-12-23 20:49:23 +01:00

There is a new paper at SIGGRAPH Asia 2023
Bounded VNDF Sampling for Smith–GGX Reflections which is based on the above-mentioned paper, but raises the lower bound so that less samples get rejected after reflection.
Note that their method can only be used for GGX reflection, and makes the pdf computation more complicated.

There is a new paper at SIGGRAPH Asia 2023 [Bounded VNDF Sampling for Smith–GGX Reflections](https://gpuopen.com/download/publications/Bounded_VNDF_Sampling_for_Smith-GGX_Reflections.pdf) which is based on the above-mentioned paper, but raises the lower bound so that less samples get rejected after reflection. **Note that their method can only be used for GGX reflection, and makes the pdf computation more complicated.**

Weizhen Huang referenced this pull request

2024-01-04 09:53:20 +01:00

EEVEE-Next: implement more efficient and less noisy GGX VNDF sampling #116773

Weizhen Huang commented

2024-01-04 10:13:31 +01:00

About my concerns regarding the lack of concentric mapping, the author of the paper, Yusuke Tokuyoshi, suggested to look at the reflection vectors instead of the half-vectors, as their method can offer lower distortions when the roughness is high.
I'm not convinced by their reply, as the result for lower roughness might still be worse. However, considering the benefit of less rejected samples, it might still worth just trying.

About my concerns regarding the lack of concentric mapping, the author of the paper, Yusuke Tokuyoshi, suggested to look at the reflection vectors instead of the half-vectors, as their method *can* offer lower distortions when the roughness is high. I'm not convinced by their reply, as the result for lower roughness might still be worse. However, considering the benefit of less rejected samples, it might still worth just trying.

❤️ 2

This pull request has changes conflicting with the target branch.

intern/cycles/kernel/closure/bsdf_microfacet.h

View command line instructions.

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u faster_ggx_vndf:weizhen-faster_ggx_vndf

git checkout weizhen-faster_ggx_vndf

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

WIP: Cycles: implement faster GGX VNDF sampling #109757

Checkout