Parallax Ray Marching [Feature Request]

[capslpop]

Hello, so I am a voxel developer that really loves to experiment with voxel parallax ray marching. I believe that this could have really big performance gains for a LOD mod like this. I would love to contribute and test this out.

Parallax ray marching is a technology that teardown uses to render its voxels. Its basically a hybrid approach to rasterization. Where you render simple box privatives that are ray marched.

I would love to really compare the difference between meshed graphics rendering and the parallax ray marched alternative. I am new here and would love some insight into how to integrate this into the existing architecture. Mainly the chunks need to be rendered as aabb boxes for 8x8x8 chunks. And the chunks should be sorted front to back for early-z optimizations. Any suggests would be helpful!

Thanks!

[FluffyPuppyKasey]

How, exactly, would ray marching improve performance? You've given us absolutely zero useful information beyond "It'll improve performance"

From what I know about ray marching, that'd be far more demanding and is very likely out of scope for a Minecraft mod designed to let the player see further

[capslpop]

So the implementation looks like this:

First an AABB representation of the chunk is assembled in the vertex shader. Then in the fragment shader the chunk is ray marched. If there is a hit then you grab the correct texture and display that to the screen. If there is no hit you call discard. That's it. No need to store meshes or update them on the CPU.

Why this is likely faster? Because you don't have to process as many triangles. Since your only drawing an AABB for each chunk then marching through it. Your basically trading vertex shading for fragment shading time. And with LODs you get a lot of tiny triangles that are performance intensive.

Optimizations: Storing the chunk as a bit mask can help a LOT here since you don't have to load every block state at once (for transparent blocks you can continue the march after texture sampling the block location when hit). Also, generally sorting the chunks from front to back each frame can be a nice optimization (though this is not strictly needed because using discard alone doesn't fully disable early-z optimizations).

This paper shows that using gl_points and intersecting an AABB can be faster that meshing when rendering voxels. However, implementing this using chunks instead will likely out compete any cutting edge meshing algorithms for voxels.
https://research.nvidia.com/labs/rtr/publication/majercik2018raybox/?utm_source=chatgpt.com

Why care about performance so much? I have been curious about this for a while. And I would love to directly compare meshed performance vs ray marched performance in a real world example.

I would love to try this out and get your input into how I might integrate this into voxy. I am very new here. Is there any documentation for voxy that I can read?