Fr0stbyte Shares His Gamedev Man-Crushes

[fr0stbyte124]

As cool as that would be, I think both guys have enough on their plate. And don't get me wrong, what we have in place for Futurecraft will be fine. I just wish I could figure out what the heck the Atomontage engine is doing. Hoping the dev will publish that paper at some point or otherwise explain it in more detail. A couple of his responses make me sort of wonder if it is terribly different from our current approach. Man, having a demo would help so much.

[Avenger_7]

Why not just explain what you're doing and ask if he can share some "insight"?

[fr0stbyte124]

Already did. I'll let you know if he responds.

[fr0stbyte124]

The mystery deepens. Apparently, 2 years ago the engine was running single-threaded and had no programmable shaders. This completely overturns all of my assumptions. Without programmable shaders, you can't do GPU ray tracing, meaning if polygons are not involved (and they're definitely not), the GPU can't be doing a heck of a lot, and that the CPU is the thing raytracing. This also opens up opportunities which aren't available on a GPU. For all its raw power, a shader thread cannot chose where it will output to. Vertices can, but not efficiently enough for a voxel engine. For a large part of the traversal, most of those traced rays are producing the exact same output, so while the CPU is not as powerful as the GPU, it may be able to work smarter. Additionally, there are more advanced compression methods available which I had falsely assumed were done on the GPU.

For instance, the highly optimized CPU voxel engine, Voxlap, famously used in Ace of Spades, stored geometry to a 2D map with variable-length height data which was run-length encoded (normally you have a value followed by a count of how many times that value is repeated, followed by a new value).

So now if we assume Atomontage is doing CPU raycasting, but not per pixel, he would be doing some sort of tree traversial to work out which surfaces are most likely visible in each chunk and rasterizing them, only I still don't think this is the solution. That alone shouldn't be fast enough, nor does it begin to delve into the compression used. There's also a curious artifact which happens from time to time where the memory is cleared and reloaded, and during that time, the scene is temporarily at a lower resolution. This means there is either some decompression, or the data is not fully traversed each frame. This is getting really hard to figure out. So many more questions than answers.

[Ivan2006]

And you are trying to get the stuff they use working for FC if I got that right?

[fr0stbyte124]

And you are trying to get the stuff they use working for FC if I got that right?

We'd never be able to pull it off. I'm just looking for new ideas. We can't do much more than we are already doing with conventional strategies without alienating the weak hardware users. But Atomontage pulls it off. Massive voxel maps and weak hardware, and I'm dying to know how. Whatever he's doing is something I've never seen before, and I get around. It's very likely something completely new.

[Ivan2006]

don´t die to know it.
You are very much needed here.

[fr0stbyte124]

Man, I'm starting to think he's not going to respond to me. And if the pattern continues, it's going to be another year before any new info appears on the engine.
Not gonna lie, it sort of feels like I just got dumped. So BRB, gonna eat this whole tub of ice cream and tell myself I'm still pretty.

[Keon]

It's okay.

[Ivan2006]

It's okay.

You are being mean.
Stop trying to drive fr0st to suicide.
I hate that kind of thing.

[fr0stbyte124]

Who's the babe?

[hyperlite]

Get back to work! Like, your real job.