Aside from Nspire 84 not being a thing: I learned much of my initial programming knowledge on a TI-84 Plus Silver Edition, and I think that really taught me to value efficiency. Love seeing other people with similar stories!
I'm a sucker for coding stuff to run on ancient hardware because I started coding when I was a preteen in the 90s, when you had to know how to get the most out of a computer or your game ran like unplayable garbage. Nowadays devs just do whatever and ignore performance and optimization because they just let their end-user's hardware go to waste. EDIT: Indie devs, for the most part, is who I was referring to, but also companies like Bethesda that build layercakes of flashy FX ontop of their ancient engine.
I have a decent pc that can even run Cyberpunk 60fps. But for some unknown reason it freezes randomly even with no load. Like im experiencing 10 second freezes: on desktop; on lockscreen; opening explorer; randomly at any point. So even if the game can run at 10k fps on a calculator im still gonna have ~0.1fps
@@cosy_sweatermight wanna have a look at your power supply, if it's defective it could have random voltage drops on certain lines that would cause GPU issues
@@cosy_sweater i've had a similar issue, it was my motherboard doing silly things to the RAM voltage which sometimes crashed my entire PC, what motherboard do you have?
There is also a much better way to render volumetric lighting. It's called "Froxels". It was first introduced in Assassin's Creed Black Flag and was later adopted by EA, Naughty Dogs, Rockstar and many more! And also it looks absolutely stunning!
@@juliandurchholz its not really state of the art, its close to 10 years old at this point. It ran great back then, and if you cant run foxels, you might as well give up running anything serious. Also it looks great in comparison with the nonsensical method proposed in this video.
@@juliandurchholz"state of the art" technically just means "what the people that know what they're doing use", but the common use is "just invented, best available way to do something".
This works somewhat ok for your specific scene because you expect a buch of sparesly scattered beams in that specific scenario. This will just look weird in a scene not covered with trees, becaue those few hundred beams will not average out into a homogeneus volume, and if you increased their amount to the point where they do, you will be running slower than aproach nº 2, and will be severely bottlenecked by fillrate. You also seem to be spawning the beams from the floor towards the sky and checking for the shadowmap at that position. That means a beam that is being occluded by a tree before it hits the floor gets faded out in its entirety: no beam after the tree NOR BEFORE IT. That too might not look so obviously wrong in that scene, but can look decidedly worse in scenarios with higher depth complexity. Its a clever hack, but with very limited aplicability.
I agree, it's certainly no volumetric light system. Definitely cool, but more a highly specific trick than anything else. Unrelatedly, the CPU readback being involved immediately makes me dislike the technique. It would be slightly less painful if they operated on the opacity data in a compute shader. Kind of does get me thinking about some kind of cached (perhaps voxel based) approach though.
As long as the trees are also receiving shadows, I imagine you could do the same technique at the edges of those shadows. Looking at the window lighting that inspired this technique, I'd expect it to be one of those things like billboarding where it's "good enough" until you really pay attention to it.
Can you please elaborate on your second sentence about a few hundred beams and homogenous volumes? Great point about checking from the floor up. An enhancement could be to check the min and max shadow depth values of each sample, then render the beam between the highest position and the lowest position (using the same opacity formula)
@kabinet0 persistent mapped buffers and fences allow the CPU readback to happen asynchronously. With 2048 beams that's 2048 * 6 floats * 4 bytes = 49kb of data, which isn't enough to cause a bottleneck
@@Vercidium Your aproximation looks somewhat plausible in a scene with very spotty coverage such as bellow trees. A swiss cheese world would also apply hahaha. But try doing that in a wide open landscape. You will still have a bunch of randomly scattered beams instead of a homogeneous solid fog volume. The random beams will make no sense (what is dividing light up into small spot-light-like rays in an open sky?". It will look more like rain than fog in that context. I said "few hundred" out of guesswork. I dont know how many billboards you are actually spawning across the scene. But still, in reasonably consistent and robust implementations of god-rays the same fog that creates these small beams in some scenes, is the fog that creates a dense Z-buffer-fog-like effect in other contexts. Its all recreations of real-world light scatering and shadowing interacting. For good reference of how broad the effect is, take a look at a RDR2 presentation on their atmospherics system they show comparisons to real world and artistic depictions of the same phenomena. AC4:BF also has a description of their implementation in a graphics presentation of theirs, which I believe was the firsr shipping "Froxel" solution. TLoU2 also has an extensive presentation on myriad optimisations and quality improvements they did to it. All that said, what you came up with is still a clever little hack that can work in specific scenes. For a broader game you'll probably want some artist control over where billboards spawn and where they don't, so they are limited to places where they look plausible. If you have a clearing in the forest for example, you'd want to have no beams in the middle of it because then again, they'd lool like weird rain in that context instead of light. Even for a window, your solution might look inproper once you have dinamic objects. Since oclusion is checked from the floor, if a character steps in front of the window, not only will the beams be bocked from reaching the floor, they will misteriously not even hit the character either. A tall enough window and a short enough dynamic object can create a very visible effect of a moving object magically making the sun-rays disapear from the window as long as it passes over the sun-lit part of the floor. PS: I just rewatched the video and remembered your trick is to only draw beams "at the edge of light". Yeah, that reduces the occurance of the "light looking like rain" problem I described. Yet, the reality is that those wide open spaces should also have the volumetric light in them (including the ocean) just instead of scattered beams, it should be a larger volume. Your solution emulates the "feeling" of a subset of a light phonomena, rather than the real thing. Thats why I recomened looking up the RDR2 presentation, they really explain all the different cases where atmospheric scatering can influence a scene.
Very clever. I never got into the technical side of things but when I worked in CryEngine (2007-2015ish). I was the optimization guy in the editor- did most of the optimizations for Mechwarrior: Living Legends environments... Kagoshima (a MW:LL level I designed) took a lot of tricks to get to run reasonably. Making Levels for that game/mod are some of my best memories, and it's still alive today, 15 Years later!
I love people who focus on optimisation because besides the truly impressive technical skills, you only get into optimisation out of two reasons: you cant run your own game/feedback from players, or and thats the reason i respect a lot because they want to make games more accessible. Games are art, more people should be able to enjoy art. Thank you!
Completely agree. Making a game more accessible by optimizing it for potato hardware is awesome. To this day I have never owned a discrete graphics card.
I think the static 2d planes variant of god rays was used in multiple games in 00s. I think Hitman 2 has it and Freedom Fighters, both are from IO Interactive.
Even Crysis 3 uses them, and still looks stunning. But such usage is limited to static scenes only. For dynamic things Crysis 3 uses screen-space Godrays effect. Works pretty well.
It looks alright, but you can tell all the beams are the same size. I wonder if increasing their width with the decrease in opacity would help make the overall scene look better and less uniform. Kind of simulating the beams becoming blurrier.
If you and acerola collaborated on a project, you would somehow manage to optimize something to render at the speed of light in a 360° POV with 4k resolution on every texture and also the game would do my homework and take me out to dinner
@@trued7461 there is probably a middle ground where you gain a little bit of performance and it actually looks the same. But from those images it ain't it chief
They could be made wider. Maybe dynamically. They're pencil thin. Maybe if they're full opacity they're also double width and you scale the same as opacity...
The dynamic beam approach was very interesting. I couldn't help but consider biasing the "dynamic beam" size/ density depending on distance to camera or splaying thebeam's mesh from the threshold of object and sky towards the ground, creating a trapezoid, to emulate penumbra.
This is crazy This channel deserves so much more!!! Keep going with your amazing work! So inspirating for people like me who would love to make video games and also like the software engineering behind it!
Yeah, this channel is so much fun and inspiring for people like us who want to get involved in engine and optimization stuff. Probably not for those who think they know everything already, but great for me. I wish there was more contents and devlogs like this. It's like a light in the darkness.
>Sees video title that I'm interested in. >All I hear is static in my brain with the well put together information because I don't understand. >Nice video.
Hey there, good video! Certainly a decent way to fake light shafts, but I wouldn't say it "looks the same" as raymarched volumetric lighting. True single-scattering will always be the most convincing second to raytracing. Your technique appears more flat and is restricted to a fixed amount of random light shafts, which probably leads to scene-dependent artifacts. What's your hardware for the given benchmarks? Would be nice to have that included somewhere. The "first" approach (screen-space post process) can also be extended to somewhat work for off-screen light sources, to a degree. Are you sure using an intermediate texture with directional blur is superior to epipolar sampling with weight function, as per the traditional GPU Gems 3 solution? I wish your video was a bit longer to cover those details and nuances, go into more depth when comparing other techniques, so as to really give an impression of the tradeoffs.
Another banger video! I always find it very interesting when people find alternative ways of creating same visual effects but in much more optimized way. I also thought about optimizing one effect called SSR (Screen Space Reflections) and came up with a completely different way of doing it, but I haven't tested it yet. Also would be very interesting to see attempts of recreating ray tracing/path tracing with roughly same quality, but a lot more performance, since ray marching is very heavy task for a fragment shader. Would be very interesting to see alternative ways of doing it.
I really appreciate content like this that tells us how to make our code both efficient and simple. I know what it's like to have a low-budget PC and still want to play games on it, and optimizations like this save the day for such gamers. Even if unoptimized code is "enough", it's preventing you as a game dev from implementing lots of interesting features. I wish the software engineering community as a whole would stop pretending resources are unlimited.
I don't think it's fair to say this looks the same as fully integrated volumetrics. Like, sure, in this one particular scenario in a forest where you want lots of small shafts this may work great, but what about other outside scenes where they aren't as closely occluded? What about local lights? What about generally open areas where the vast majority would be fog-filled? Generally this is a lovely technique for those very specific small beams of light, but I don't think this is anything even remotely close to volumetric scattering - hence why there aren't many games that use this particular method. Great video nonetheless, even if I would have hoped for a few more details about placement and transformation of the beams.
I like your content man, very well edited. Just wondering (and sorry if it's a stupid question), but do these optimizing techniques translate to other game engines or are they just specific to the one you're using?
If you want to, you can apply these optimization techniques yourself to other engines, though that means throwing away what the engine itself offers and writing things from scratch, while also building it in such a way that it works with all the other parts of the engine.
This should be possible in any engine that supports shadows and vertex readback. For example Valheim uses a custom render pipeline within Unity, I reckon they could add beams of sunlight like this.
I know this is an older video now, but I'd like to share a method I used 15 years ago in Blitz3D (DirectX7). Calculate the intersection point between Sun and Camera ray, capture an Ortho view from that point towards the sun. Place up to 255 transparent planes along the sun ray, placing them at equal distance step, starting at the point that first intersects the frustum, and ending at the point that last intersects the frustum. This is effectively the 2nd method, but it should be significantly faster than the 2nd method. Downside is that you need to take care of gaps between planes by tilting them towards the player camera.
I think the best move is color your models and textures like if light actually hit and color it like if light didn’t. What’s going on from I am seeing is everything is completely singularly functioned and not programmed to act like how light should hit the environment. The light shouldn’t be faked, but how you color the textures and models can be faked to the point that it looks real. If you let the engine itself do all the work, you won’t get the performances with its default settings, so getting the models and textures early on in development to look like it is real-time before the game engine, can break the limits of the engine and you won’t see the work break itself.
The end result isn't stunning to me visually, however the approach to get there is highly intriguing and gives me all sorts of ideas. These videos are great for getting the indie dev mind thinking.
I always find it fascinating to see experts go ham in thier in their interests and listen in no matter how much i actually understand on the subject at hand 😂
In an old game prototype I made, I had a lot of grass models placed around the world, but they were physical objects and would lag a lot when loading/unloading and area because of them. There is probably a way to use a shader instead of the grass actually being real while still giving me a way to control where the grass is and have the player model interact with it (I'm just not good enough at shaders to make it). That could be a cool video idea, I would love to see you render millions of blades of grass a few thousand times a second lol.
@@Vercidium i think Acerola's approach laks your way of optimising memory for meshes by using only vertex indices. Also you have your own engine and probably don't need to run a compute shader each frame just to have a vertex buffer filled (that's what DrawMeshInstancedIndirect does in unity i think). Would really love to see how you implement grass, probably in future videos
I wonder: what about a voxel map for where light goes? The map would be a 3D greyscale texture (you can tint it later). Initially fill it with all-white voxels, black out voxels where solid geometry exists, then do a directional blur away from the light sources (which would be a single direction for the sun, given it's far enough away that its rays can be considered parallel by the time they reach Earth), and repeat the black-out and blur steps a few times. When rendering, sample the voxel grid along the camera ray to get side-illumination data using a 3D texture sampler. You'd have to experiment to find how low a resolution you can get away with, but it could allow pre-calculating some of the lighting data, and as a bonus, you'd know ahead of time how many steps to take along the ray during rendering, just from the grid resolution and distance. If your swaying trees follow a predictable loop, you could calculate a few frames of voxel lighting data and select from the appropriate one, or even interpolate between them. I have no idea how fast this would be, and it absolutely would be memory-hungry, but by transforming the volumetric lighting to image-processing (just in 3D), it might be faster than doing the real thing. With luck, you could do it at load-time, or chunk-generation, rather than having to keep it on disk. Hell, you could probably do it in a compute shader on the GPU, so the CPU never needs to touch it. Now that I think about it, it sounds like it has a bit in common with Quake's lightmaps, except 3D and only concerned with direct illumination.
I would recommend folks look at the GDC talk for how light beams were handled for the game Abzu. Very cool. In that talk as well is a great summary of how to animate thousand of creatures simply via vertex offsets on the GPU - but that isn't so related to this.
I think it's fair if the output looks comparable to the human eye. The thing that is being optimised is the graphical output, not the completion of a specific process. The fast inverse square Quake is a classic example. If someone replaced the true inverse square with the fast inverse square, I'd certainly consider that an optimisation even though it technically gives a slightly different output and technically isn't the inverse square.
Wouldn't another option be to do a shell texturing approach where you have planes always in front of the screen that are lit by the shadow map lighting you already have in a certain way that looks like volumetric "god rays?"
Idk if this just common knowledge or if this man is just a genius but my god the game space is gonna be so much better from here on out if this guy is working on em. Great video!
It’s amazing that the scene calculations can accidentally be forced to figure out if an infinite set of objects exist in what needs to be rendered for each frame and then doing the calculations on what’s actually in the scene and ray tracing those objects
The visualization of the fragment shader output, with all the colored, positioned fragments, was amazing! Can you explain how you did that? Was it a simulation or is there a tool that helps with this?
Thank you! It's a simulation, I used GL_POINTS to render each pixel, using the colour from the final image. Each pixel was then positioned by sampling the depth buffer in the vertex shader, then re-creating its 3D position from the depth value. I don't fully understand the matrix math, but it's explained well here: stackoverflow.com/questions/32227283/getting-world-position-from-depth-buffer-value
Ok I can add on this... When the day passes less beams should be on screen... Beams should't be statics but move slightly over time... Great concept... But I still can't get dynamic snow done in this way as the sheer amount of vertex points saved is the problem... Im getting closer tho...
The key to creating a simulation that is both accurate and fast is to forget about accuracy, cheat as much as possible, and tell no one. No but for real there will be times when relying on good old RNG gives a result that is only slightly worse than running an expensive simulation. The real key is finding shortcuts and approximations that get you that "good enough" result where no one can honestly tell it isn't 100% accurate without costing too many system resources. Creating an accurate simulation is only half the job, the other half is figuring out the tricks that let you run the simulation in real time.
You should be able to use atomic's within the shader instead of a feedback transform. This will skip the extra step of having to communicate with the cpu.
If it's through leaves, the flickering isn't necessarily wrong - just annoyingly flashy, as flickering leaves are in real life! But crepuscular rays through clouds shouldn't flicker, so this is great for those
Great video. Thats all i have to say tbh... which is why ill ask a off-topic-ish question instead, What did you use to create all the motion graphics like the diagrams and the code boxes?
only problem i see is when the sun's rotation changes quickly, as in 4:53. The rotation pivot should be on the object castig the shadow on the beam, not on the beam's contact with the ground. Outside of that, great approach!