WIP: Cycles octahedral normals (and tangents) #121704

Alex Fuller · 2024-05-12T05:10:37+02:00

Alex Fuller commented

2024-05-12 05:10:37 +02:00

First-time contributor

While experimenting with corner normals, an issue coming up is the extra memory overhead that would incur if there was a hypothetical swap from vertex normals as the default. One of the known effective compression techniques is octahedral normals:

https://knarkowicz.wordpress.com/2014/04/16/octahedron-normal-vector-encoding/
https://twitter.com/Stubbesaurus/status/937994790553227264
https://jcgt.org/published/0003/02/01/paper.pdf
https://www.shadertoy.com/view/llfcRl

I went with compressing/quantising to a single uint, it seems to have enough precision and is 1/3rd the memory usage over float3.

Another concern is any performance overheads that would incur from this approach, so there is some benchmark data. Octahedral is just the normals, and octahedral_tangents combines both normals and tangents (as well as compressing tangent_sign to a char/8-bit value that gets a cast to float).

I've only just tested on CPU and OptiX (Nvidia RTX 2070) but the results seem like they do add overhead, and junkshop has the best memory savings probably due to using more normal maps than the other tests.

Renders seem to 'match' enough, the quantisation probably causes some samples to differ very slightly.

While experimenting with corner normals, an issue coming up is the extra memory overhead that would incur if there was a hypothetical swap from vertex normals as the default. One of the known effective compression techniques is octahedral normals: https://knarkowicz.wordpress.com/2014/04/16/octahedron-normal-vector-encoding/ https://twitter.com/Stubbesaurus/status/937994790553227264 https://jcgt.org/published/0003/02/01/paper.pdf https://www.shadertoy.com/view/llfcRl I went with compressing/quantising to a single uint, it seems to have enough precision and is 1/3rd the memory usage over float3. Another concern is any performance overheads that would incur from this approach, so there is some benchmark data. Octahedral is just the normals, and octahedral_tangents combines both normals and tangents (as well as compressing tangent_sign to a char/8-bit value that gets a cast to float). I've only just tested on CPU and OptiX (Nvidia RTX 2070) but the results seem like they do add overhead, and junkshop has the best memory savings probably due to using more normal maps than the other tests. Renders seem to 'match' enough, the quantisation probably causes some samples to differ very slightly.

cycles_junkshop_main.png

15 MiB

results.html

25 KiB

cycles_junkshop_octahedral.png

15 MiB

cycles_junkshop_octahedral_tangents.png

15 MiB

👍 4 👀 2

Alex Fuller added 1 commit 2024-05-12 05:10:48 +02:00

b9e876232b Cycles: Octahedral normals

Alex Fuller force-pushed feature/octahedral_tangents from ba15d4fde0 to 3cd950f8b1

2024-05-14 15:21:47 +02:00

Compare

Alex Fuller commented

2024-05-14 15:29:49 +02:00

First-time contributor

I made a modification in this latest version. I've updated the functions to be templated so you can specify the precision of the octahedral quantisation. I've made this so I could compress tangents into 30 bits instead of 32 bits so I can also fit tangent_sign in there, as it only needs to store 1.0f or -1.0f (with the default as 1.0f) which also simplifies the code quite a bit across the board. I needed to make it compatible with OSL but I've not tested that codepath yet.

Memory should go down a little bit more and visually I can't see a difference - but I'd appreciate more eyes on the signage code, or a unit test that can verify that the flag is working as intended - looks like 97.13mb less in junkshop from the previous commit (or 336.81mb less from main branch).

I made a modification in this latest version. I've updated the functions to be templated so you can specify the precision of the octahedral quantisation. I've made this so I could compress tangents into 30 bits instead of 32 bits so I can also fit tangent_sign in there, as it only needs to store 1.0f or -1.0f (with the default as 1.0f) which also simplifies the code quite a bit across the board. I needed to make it compatible with OSL but I've not tested that codepath yet. Memory should go down a little bit more and visually I can't see a difference - but I'd appreciate more eyes on the signage code, or a unit test that can verify that the flag is working as intended - looks like 97.13mb less in junkshop from the previous commit (or 336.81mb less from main branch).

results.html

25 KiB

cycles_junkshop_octahedral_tangents.png

15 MiB

🚀 1

Alex Fuller force-pushed feature/octahedral_tangents from 3cd950f8b1 to 680ad333ab

2024-05-14 15:40:59 +02:00

Compare

Alex Fuller commented

2024-05-14 15:41:55 +02:00

First-time contributor

...and one minor tweak so that non-default tangents are correctly tagged for encoding/decoding.

Alex Fuller force-pushed feature/octahedral_tangents from 680ad333ab to f12776cdb2

2024-05-14 15:52:30 +02:00

Compare

Alex Fuller force-pushed feature/octahedral_tangents from f12776cdb2 to a627942c02

2024-05-14 16:27:14 +02:00

Compare

Brecht Van Lommel requested review from Brecht Van Lommel 2024-05-21 19:46:00 +02:00

This pull request is marked as a work in progress.

This branch is out-of-date with the base branch

View command line instructions.

Checkout

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

git fetch -u feature/octahedral_tangents:boberfly-feature/octahedral_tangents

git checkout boberfly-feature/octahedral_tangents

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 octahedral normals (and tangents) #121704

Checkout