Ray Marching, and making 3D Worlds with Math 

Even if I ever understood before, this might be the time it actually sticks :) for some reason I always mixed it up with tracing, or maybe something to do with section searches

​@@johanngambolputty5351 well technically you're still tracing rays, right? certainly it's ray casting, the term they used in 3d games back in the 90s to collision check your projectile trajectories and the such, which is just 1 ray. [ray tracing] is a routine that casts a ray for every angle your "caster" can "see". in 3d rendering often a camera, so the camera settings dictate the position and the angles the caster limits itself to, the pixels it traces are projected into angles depending on FOV/perspective values from the camera, they ARE the set/array of angles to process basically. it's still raycasting too, raycasting a lot in fact, raycasting is still the integral part that does the work. (a 1080p image is about 2 million pixels, so ~2 million rays if just 1 ray for every pixel, modern path tracing do many many more per pixel) there's also bounces for light ofcourse and the such, but for the sake of simplicity let's keep it at 1 bounce. [path tracing] is a more complex raytracing that can also do reverse tracing for caustics and the such, has way more rays for more modern PBR rendering methods, z distance information feedback for post processing and the such, essentially raytracing on crack. [ray marching] i consider it to be a variant/derivate of ray tracing, i'm not sure how disputed that is, but i'd put it under that moniker. but it is smart/dynamic, rather than checking everything, it makes bigger jumps, and you can reduce the amount of cycles by a great deal just working with minimum lenghts set on a higher value , but very small object intersections aren't "seen" because the ray jumps over them, this disadvantage is exploited as an advantage in the demos in this video by interpolating and blending the missing parts together the "looking over it" principle is also used in metaballs, but metaballs calculate a lot of shit on the way, this method actually drops data and fills in the gaps, it has a bit less control, but you could put extra data in vectors and use those with proximity for all kinds of effects too i don't think the vector lengths (jump distances) strictly have to be dynamic for it to be ray marching, they could be static and just have it jump the same distance every time as a crude version until it hits something it's not ray marching if you only shoot 1 ray either, so it's just grid-snapped ray casting strictly taken, but the distributed is implied although there may be differentiation in terminology these days perhaps, but that was also understood as ray marching back in the day ray marching is a de facto collision detection algorithm, as is ray tracing, because they're all raycasting and the rays are looking for something to bounce on, that precisely is what collision detection is, finding an intersection. as with many code, raymarching differs a lot between CPU based (serial functional programming) and GPU based (parallel logical programming), they need a quite different approach. the crude version of ray marching basically is raytracing, but the ray casting it does, is snapped to a grid with the grid's rotation relative to the individual ray's angle so it always traces over a striaght line (which is exactly what "local coordinates" are) short story long: i consider raymarching a variant of raytracing, i'd define them as: "distributed raycasting algorithms" when in doubt, use the term ray casting, that's the underlying principle for all of them, including ray casting itself obviously, so you can never be wrong then ;) even simple pixel-based detection is essentially just crude ray casting with the vectors being snapped to integer values, using global coordinates, and dropping all the benefits of unit vectors, when you think about it, right? lol i just had an epiphany, never even thought of it about like that myself even.

@@dutchdykefinger True, its a clever sort of tracing that you can do if you use distance fields instead of polygons, especially since the step size is not just dynamic but optimal (largest while guaranteed not to hit anything). I'm not familiar with this stuff, so I just imagined constant step sizes for tracing (against polygon primitives)... though I'm sure it can get way more clever

Ooooh sounds cool! Do you have any demos or code bases you can share?

@@keyb not yet unfortunately. The first year was mainly reading papers, understanding the heaps of theory and finding a sensible way of actually making it performant. There are bunch of challenges i wanted / had to solve first: - how do i make dynamic scenes with SDFs instead of scenes hardcoded into a shader? - how do i avoid as many branches as possible? - how do i completely avoid the default render pipeline? - how would instancing work? - how can i drastically improve performance? - etc. And the second year was mostly occupied by learning vulkan tbh. But i‘ve been making some major strides recently and i‘m hoping to finish a 2D version before end of the year. That will allow me to make some tools to drastically accelerate my dev process and it will all snowball from there. End-goal is a full 3d engine for real time rendering :D I can bookmark this video and let you know in this thread when something showable is abailable if you want? :)

Super cool! Got a demo I can watch of it in action?

@@simondev758 not yet unfortunately :D See comment above yours hahaha I‘d be happy to send you a demo as soon as something showable is ready, hopefully near the end of this year :)

​@@Chribit Yes please! It sounds like a ton of work (especially in the sense of optimizations) so you still probably have a while to go. Whenever you do get a working model please let me know!

yeah, the naive ray marching function steps a point towards the target and returns when it ends up inside an entity/object

Do you mean sphere tracing is just ray marching with SDFs?

​@@diodin8587basically, if you open up the wikipedia page of ray marching there's a section on it.

I believe they used a 2D version for text on signs in Team Fortress 2. I vaguely recall reading about it back in the 2ks.

Awesome!

Yeah, I think if I were to redo it, I'd add some animations around computing things like the normal/shadows in the same way that I did for the raymarching. Next time!

Agreed, the performance cost is definitely real. I was mostly thinking in terms of implementation cost for a lot of these effects, but yeah the ray marching loop can't early out as easily.

Let me know if you do!

Hah

Me too!

@@simondev758 just found this ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-ARlbxXxB1UQ.html

Dreams on PS4/PS5 uses this technique extensively.

There's worlds in VRChat featuring Raymarching shaders that you can experience in VR. It's quite fascinating. One of the worlds was called "Treehouse In The Shade".

@@KillFrenzy96 thank you so much!

I think it could, but that'd be a major challenge. Maybe for a day when I'm super bored heh

Definitely planning on playing with it

Thanks, I think I need to find a better recording location too heh

@@simondev758 Yeah, that could work as well 😅

All done in shader, pretty much everything you see is based on what I explain in this video and the next (lerp).

100% done in shaders yes

yes, they are. Inigo Quilez has excellent tutorials about it. You can see hundreds, if not thousands of samples of SDFs on ShaderToy. The platform was created by him partially as well.

@@simondev758 think you just sold me on the course. Awesome.

@@AntonioNoack thanks! I will check it out

@@brennonwilliams9181 Heh awesome. And yes, definitely check out Inigo's site, it's in the description, amazing resource for ray marching & sdf's.

Most implementations set a max # of iterations, usually about 100 or so get's you a good result, a lot less depending on the scene.

@@simondev758 at each point/iteration, we have a different set of scanning angles around the point, right? So how about the scanning resolution per sample point usually?

I probably forgot to upload it, lemme go find the code and do that.

@@simondev758Not to come off as rude, but did you upload it yet? I also couldn't find it and I'd be super interested...

@@simondev758 Just casually dropping by to remind you to upload this code. Cheers :)

These were done in the fragment shaders.