Saturday, October 28, 2017

Trees on Fire

As promised at the end of my previous post, I've implemented fire spreading in trees in 3DWorld. I'm happy with the way it turned out, both the visual quality and simulation accuracy.

I chose to use a more physically correct fire propagation algorithm for trees rather than a 3D voxel/volume based approach as described here. Fire is fueled and spread by individual tree branch cylinder segments, and nearby leaves are burned black. The tree trunk, roots, and larger branches can all be on fire. A branch section has the following fire-related properties:
  • Hitpoints: how much heat damage this segment can take before catching on fire; scales as a function of branch radius so that thin branches are easier to burn (kindling)
  • Fuel: how much fuel is available to the fire from this branch segment; scales as volume
  • Burn rate: rate at which fuel is consumed; also controls the fire size/intensity; scales as surface area, since the branch surface is what's actually burning
As far as I can tell, the scaling is physically accurate to first order. Small, thin branches will burn easily and quickly. Trunks take a while to catch fire, but burn with high intensity for a long time.

Tree branches are each composed of several cylindrical sections that have their end point vertices merged to produce a closed 3D volume. One or more leaves are placed along the thinner branch segments. Each burning branch segment has a fire randomly placed somewhere on its upper surface, with wider branches having larger fires. Every frame, the active fires burn nearby branch segments and leaves that are within their damage radius. Segments take damage until their hitpoints are reduced to zero, at which point they catch fire. Leaves are burned to a black color and eventually are randomly either consumed, or fall off the tree and slowly drift to the ground. The bark color is gradually and uniformly transitioned from brown to black as leaves are burned off.

Below are some screenshots of trees in various stages of fire.

Simulation and rendering of fire spreading to three trees using a realistic fuel/volume and surface area model. Oh, look, I finally added a real health bar at the top left of the screen!

Fires are drawn as animated billboards using a fire texture atlas, the same rendering method used to draw grass fires in the previous post. Additive blending is used to provide emissive color and make the branches behind the fires appear to be red hot. Fires also produce occasional puffs of dark gray smoke that rise into the sky. When the player is close to a burning tree, a looping fire sound is played and a shimmering heat screen-space postprocessing effect is enabled.

Closeup of a burning tree with glowing red hot branches, blackened branches,burnt leaves, and falling leaves.

Grass fires and tree fires have been integrated together so that they can start each other. Burning tree roots will set the grass on fire, and burning grass will set the lower tree trunk on fire. In addition, fire can spread between the small branches of nearby trees, quickly setting dense clumps of trees ablaze.

I've put a lot of work into making tree fire spreading efficient so that it scales to hundreds of trees. The next three screenshots show a forest scene with 140 trees chosen from 100 unique procedurally generated models. Each tree has an average of 800 branch cylinders and between 4,000 and 16,000 leaves; bushes on the ground have fewer leaves. That's over 100K total leaf cylinders and around 1.4M leaves that can be damaged. I get a framerate of around 35 FPS (45 FPS after the latest optimizations) when most of these are burning.

Starting a forest fire by firing rockets at trees. The trees have begun to burn, as well as the grass around their trunks.

One key optimization is to only allow each individual fire to spread once every several frames. I chose to call the update (spread) logic for any given fire at most once every 16 frames (leaves) / 4 frames (branches). Each frame, a sliding window of leaves and branches is processed so that they all get updated at the same average rate.

Another optimization applies to the later stages of tree burning. Just like in the real world, fire spreads quickly through trees, but takes some time to finally die out. Most of the burning period of a 3DWorld tree is spent after all branches have been ignited and most of the leaves have been burned off. In this phase, the fires have nothing else to damage. Each fire tracks the last frame in which it applied damage. Fires that haven't applied damage in the previous frame are put into sleep for the next 8 frames, saving CPU cycles.

I had to make the smoke effect very sparse because smoke particles add a lot of GPU pixel fill and hurt the frame rate when the camera is close to them. With this amount of smoke, it may be more efficient to use a screen space technique such as ray marching in a volume field rather than rendering individual smoke particle billboards. In any case, at least the smoke doesn't obscure the fires too much in these images.

Some trees have been consumed by fire, and now the grass has started to burn outward as the fire spreads.

I've thought about making the tree fire spreading system multi-threaded, but so far I don't have a working solution. The problem is that the tree fires affect other scene elements such as grass, nearby trees, smoke, players, objects, etc. Some of these, such as blackening and removing of leaves, requires modifying GPU state, which can only be done from the master thread. In fact, most of the CPU time is spent updating all of these other scene objects. The fire logic applied to each branch itself has good cache performance and is very fast.

It's now a full forest fire, with nearly half the trees and over half the grass burned or burning. Note the volumetric, depth-aware gray smoke puff at the center of the screen.

Here are some videos showing how fires spread through individual trees, and through groups of trees. I watched a number of videos of real trees burning on YouTube and tried to adjust the constants to get a believable effect. At first I though the 3DWorld fire spread too quickly, but apparently dry trees can be consumed by flames in a matter of seconds in the real world. I also remember watching a tree fire years ago when I lived in Berkeley. The trunk of a tree in the parking lot of an apartment building was burning. I watched as someone ran to his parked car and quickly drove it out from under the tree just before the crown burst into flames. There sure was a lot of smoke!

The first video shows me burning three trees in the office building scene. Rocket explosions start fires almost all the time, but plasma cannon explosions only start fires half the time. The trees are over concrete so there's no grass under them to burn, making this a good test of isolated trees.

The next four videos have a lot of trees and bushes. I recorded these videos before I was finished optimizing the code, and 3DWorld can't quite keep up with the target 60 FPS video recording rate. I got around 40 FPS in the house scene and as little as 25-30 FPS in the forest scene when streaming video compression was running at the same time as rendering. (With the latest optimizations I now get over 40 FPS in the forest scene.) This means that the videos play back at around 2x the realtime rate after the fires have spread a bit. It takes 3DWorld 2 seconds to generate 1s worth of recorded video (60 frames / 30 frames per second). That's okay though, it takes a while for the trees to burn so speeding up the process shortens the videos and makes them more interesting.

Keep in mind that video compression seems to ruin the nice alpha blending and antialiased leaf and fire edges. The raw frames sent to the screen at 1080p resolution look much better. Unfortunately, the high resolution videos are over 100MB and are too large to include here.

This scene contains a yard with three trees, two hedges, and lots of grass. The hedges are each composed of dozens of small square shaped trees. The dense packing of these trees makes fire spread quickly through the hedges. I start by setting the trees on fire, which quickly ignites the grass as well. The small tree burns very quickly, and the large tree on the hill burns for a long time. There is strong wind enabled, which affects the fire spreading direction and also makes the leaves fall in a diagonal direction.

These three videos show fires spreading in the forest scene. This is a square region of light forest with a total of 140 trees + bushes and 1.86M blades of grass. Each tree has around 200 branches composed of a total of about 800 cylinders and 4,000 to 16,000 leaves. Fire burns the leaves to black over time. When leaves are fully burned they fall off the trees and land in the grass. The max number of active leaves is set to 3000, so when many leaves are falling, the leaf objects are recycled before they reach the ground.  As the percentage of burned leaves increases, the tree bark transitions from a light brown to black color. A single fire will spread to every grass blade and every tree in this scene within 5 to 10 minutes.

The first forest video shows how fires and started and how they spread through trees and grass. Near the end I disable the terrain mesh (ground) so that the burning tree roots can be seen.

This video shows fire spreading between the small branches of nearby trees. I set one tree trunk on fire and wait for it to spread. Note that I've turned off grass blades to prevent the grass from burning and igniting other trees. Fire will only spread to dense clusters of trees and will stop at the gaps.

Here I set the grass and trees on fire, then put the fires out with rain. Rain and snow have two effects on fire. First, they reduces fire intensity by a fixed amount per unit time, decreasing the temperature and radius of each fire until it's extinguished. Second, they temporarily increase the wetness factor of the scene. This effectively increases the hitpoints (resistance to heat damage) of the tree branches and grass, making it more difficult for fire to spread. Once the rain has stopped and the scene dries out, the fires can be restarted and will once again spread quickly across the trees and grass.

I've also just recently implemented burning of plants. It's pretty simple though, the plant leaves just turn black and disappear after they've taken enough heat damage. It's not much of an effect, but it does remove those pesky bits of green that remain after everything else has burned to black.

That's it for now. I might work on burning other objects next, such as building structures and other non-vegetation items. I'm not quite sure how to model those objects though. There's currently no way of tagging scene objects such as chairs and walls with parameters that control their flammability or fuel content. There's also no obvious way to burn or destroy these objects. Do they turn black? Do they explode? Break apart? Melt and change shape? Disappear? This requires a more detailed material model. I already have objects that can and do explode when burned, but no system for implementing the other effects.

Monday, October 16, 2017

Grass Fire Simulation and Rendering

About two weeks ago I was looking through some graphics blogs and was inspired by this article on fire spreading in the game Far Cry. I decided to implement something similar in 3DWorld, since it seemed like a good addition to the existing physics and terrain systems. I already have support for isolated fires, but so far I didn't have a good/efficient system for fire spreading.

Note: This work was started the week before the California wildfires and is unrelated to that disaster.

I decided to start with only grass fires that spread across the terrain mesh but don't spread to other scene objects. The existing uniform mesh x-y grid structure seemed like a fine place to store fire related data. Each grid cell contains the following fire fields:
  • Hitpoints: The amount of fire/heat damage this cell can take before it catches on fire; this is a function of humidity (grass wetness), current precipitation (rain/snow), and some randomness
  • Available Fuel: The amount of grass fuel available for the fire to burn; consumed over time as the fire burns; this is a function of grass density (blade count)
  • Burn Intensity: The intensity of the fire from 0.0 to 1.0; this controls fuel consumption rate, heat/damage generated, spreading rate, and flame density + height
Weapons such as the plasma cannon and rocket launcher can be used to start fires in grass covered mesh cells. Explosions have a probability of creating one or more nearby fires. Once an isolated fire is burning in close proximity to the mesh, it will start doing damage to reduce the hitpoints of the containing cell. When the hitpoints reach zero, the fire will start to burn with a small intensity. The intensity increases over time until all fuel is consumed, at which point the intensity will decrease over time to zero. High intensity cells do damage to their four neighbors, which will eventually also catch on fire. Rain and snow quickly extinguish fires by reducing their intensity to zero over the course of a few seconds. They also wet the grass, making it more difficult to set it on fire later.

Here are some screenshots of fires burning in a scene consisting of dense grass with light tree cover. The fires burn the grass and flowers but not the trees. Fire leaves behind burnt black grass blades and gray ash covered ground where the green grass used to be.

Grass fire burning a circular area between the trees. Oops, looks like flowers aren't burning.
Ring of fire, after flowers were fixed so that they burn as well.

Fire burning the grass and flowers nearby, with heat haze producing rippling effects on the screen.

Fire won't burn areas that are over dirt or rock, or cells that are underwater. Also, once the grass has been burned, it remains black and can't be set on fire again. This image shows multiple fires. If you look closely, there is a shallow pool of water under the grassy area in the bottom right, which is why it's not burning. In addition, the steep slopes are covered in dirt and don't burn. The source of the nearest fire was an explosion in bottom center of the image where the grass is pushed away in a circle.

Multiple fires burning in this scene. The puddle near the bottom right produces wet grass that doesn't burn.

The rate and direction of fire spreading is affected by wind speed and direction. The fire will spread more quickly in the direction of strong winds. In the absence of strong wind, fire spreads approximately uniformly in all directions that contain grass, forming a circular burn area and a doughnut shaped ring of active fire. The fire at the center of the circle has consumed all of the grass fuel and has burned out.

This fire implementation is integrated into the physics and gameplay system of 3DWorld through the following effects:
  • Green grass blades are slowly turned black at randomly varying rates
  • Green terrain texture is temporarily turned black and will fade to gray ash over time
  • Light black/gray smoke is produced that will float toward the sky
  • The player takes damage while walking over fire in gameplay mode
  • Objects tagged as explode-able explode when fire contacts them (see video below)
  • A burning sound loop is played when the player is near a fire
  • Heat from nearby fire produces a wavy screen space effect (see image above)
  • Heat sensors can detect the heat from fires
Here's another image showing a fire burning in the back yard of my house scene. It was like that when I got there, I swear! This plasma cannon I'm holding has nothing to do with it!

Fire from the plasma cannon spreading in the back yard.

Given enough time, the fire will eventually spread to all the grass around the house, but it currently won't set anything else on fire.

I recorded two videos of the grass fire effect. The first video shows how a fire spreads, destroying an exploding tank in its path. Near the end I stand in the fire and take damage until I burn to death. Sorry, there's no sound this time.

The second video shows fire spreading in a dense grassy meadow with sparse tree cover. It will climb the hills and reach any area connected by a path through the grass until the entire scene is burned black. You can see how the fire produces emissive lighting when I cycle to night time. At the end of the video, I enable rain, which wets the grass and quickly extinguishes all fires.

This system is very efficient. The entire scene can be burning and I still get about 40 frames per second. More reasonable fires will give me at least 100 FPS. Most of the frame time is taken by the grass and terrain texture updates, as these require new data to be sent to the GPU each frame in many small batches. To reduce the time spent in these operations, grid cells are split into batches and updated sparsely. Each cell is assigned a random number fro 0 to 31 every frame, with the following update cycle:
0: Update grass color (green => black)
1: Update terrain texture (grass => dirt or rock, depending on elevation)
2: Update terrain color modulation (white => black, temporary)
3: Generate a smoke puff that rises
4: Damage nearby objects such as exploding tanks
Any other number does no update. Therefore, expensive updates are only done for 1/32 of the active fire cells each frame. This is enough to keep cell update time around 1ms in most cases.

Fire is drawn as a collection of up to 6 variable sized, camera facing quads (billboards) per active cell, for the cells within the camera's view frustum. The number of quads is a function of burning intensity with some random variation. Each quad is textured with a sequence of 16 images from a texture atlas that gives a burning flame animation. They're drawn with additive blending and emissive lighting to produce bright, intense colors. Enabling additive blending and disabling depth testing allows thousands of fire billboards to be drawn without having to sort them back to front for alpha blending. I also use the depth buffer from the previous frame to blend the bottom edges of the flame quads to transparent where they meet the ground, which avoids ugly artifacts at locations where the quads intersect the terrain mesh under them.

For future work, I would like to extend this system to non-grass types of fires. For example, fire could be made to spread to trees and other scene objects. This is more difficult because it's a 3D effect rather than a 2D affect applied to a uniform grid on the ground. It's not clear if my billboard rendering will look good when the fire is up in the air on a tree branch. It's also unclear how to make other scene objects appear burned. I'll be sure to post more screenshots and videos if I ever figure this out.

Sunday, September 24, 2017

Slow Motion Physics Video

I haven't done much work on the 3DWorld source code since my last post. I've been mostly experimenting with the physics system and trying to reduce instability. I took some videos of slow motion physics to help me analyze the quality and problems in my system, and I'll show two of these here.

3DWorld has an interactive timestep adjustment feature that allows the user to slow down the physics timestep by an arbitrary factor, which can be used to make objects move in slow motion. The timestep can also be reduced to 0 to disable physics and lock all objects in place. This doesn't affect the player movement or weapon firing, only dynamic object motion. I can use this feature for analysis and debugging of physics problems. For example, I can fire a rocket at an object, then freeze physics and walk up to the object for a closer view of the collision, then let the rocket continue to explode in slow motion.

The following video shows me breaking windows in 3DWorld with the physics timestep set slow enough that the individual triangular glass fragments can be seen spinning though the air and colliding with the scene. They don't collide with each other. The sharp glass fragments damage the player, which is why there is so much blood in the air and on the ground. Blood moves in slow motion as well. Bloody glass fragments will turn a reddish color.

Note that I'm recording video in realtime on 7 of 8 threads (3.5 of 4 cores) and only letting 3DWorld use one thread for physics + rendering + screen capture. Even with this setup, the video compression can't always keep up with the 60 FPS target framerate. This means that 3DWorld sometimes can't produce the results at the recording rate, which is why the video plays back somewhat faster than realtime. However, the physics is slowed down by ~10x, so you can see the fragments moving just fine. Sorry, I still haven't figured out how to capture and record 3DWorld's audio stream, which is why there's no sound.

That big gray sphere at the end is my skull, and the black sphere is one of my eyes. The camera follows the other eye as it bounces on the floor. I have no idea what happened to my red nose. Yes, I bled to death from cuts due to broken glass. It happens, especially when using the baseball bat on windows. I don't recommend trying this at home. Oh, and if you're curious, you can get hurt by wood fragments as well.

If you look closely, there's some instability in the glass fragments resting on the windowsill and on the ground. Some of them rotate in place even though they're no longer falling. Why is that? Fragments rotate about a random axis at a rate that's a function of object velocity. This means that angular velocity scales with linear velocity. Objects that have come to rest on a flat horizontal surface such as the ground are flagged as "stopped". This allows the physics system to skip some of the work when processing stopped objects, and it also disables object rotation. Objects that aren't moving but aren't in a stopped state have a small negative Z velocity impulse related to the effect of gravity that pulls them down. This velocity produces a small rotation that looks odd.

The problem is that the stopped state logic is approximate, and doesn't work well when an object collides with multiple scene faces at the same time. This is what happens near the outside corner of the windowsills and the inside corner between the wall and the floor. They're not put into a stopped state due to the possible collision with a vertical face that may push the object away at the next timestep. If they land far enough away from vertical surfaces, everything is fine. These fragments look like triangles, but the physics system uses bounding spheres for them. This means that a triangle that appears to be resting on the ground may actually have its bounding sphere touching the wall.

In addition, the wood fragments from the bench are 1D polygons. I hadn't implemented texturing of thick 3D triangle fragments yet. I've partially fixed the glass instability now - the fragments sometimes jitter a small amount, but they don't rotate. I've also added textured thick wood fragments. These look better than flat polygons, but still don't look much like real wood splinters. It's progress, but far from complete. The updated video is below.

I was careful not to die this time.

There are still some issues to fix, but it looks a bit better. Well, at least if you know what to look for. For example, real glass fragments don't land at odd angles, they land with the flat side down. Unfortunately, that's also a bit too simple and doesn't look right. Drawing triangle fragments that intersect each other on the ground looks even worse. No, they land with the flat side down, but only if they're not landing on top of each other. If fragments land on each other, they form a chaotic pile. This requires collision detection between thousands of fragments. It's something to possibly add later.

I'll post more screenshots or videos if I make progress later.

Tuesday, August 29, 2017

Bouncy Ball Physics

Recently, I've watched some videos of physics simulations in other game engines. This got me interested in experimenting with physics and collisions of large number of objects in 3DWorld. 3DWorld supports full physics for spherical objects, including gravity, elastic and inelastic collisions, momentum, friction forces, air resistance, etc. All "large" (non-particle) spheres collide with all objects in the scene, including each other. Spheres are simpler to work with than other object shapes due to their rotational invariance and other factors. Each frame runs multiple physics simulation + collision detection steps on all active (moving) objects. The active object test significantly reduces computation time for common cases where most of the objects have stopped moving. In theory, the collision detection should prevent objects from intersecting each other and the scene.

I added a config file option that reduces weapon fire delay to 0, which allows me to throw a ball every frame. The maximum number of active balls is set to 20,000. At ~100 FPS, it only takes a few minutes to generate 20K balls. Here is a video of me throwing some 10K balls into the office building scene in realtime. See the balls fall and bounce in parabolic trajectories. Note that I've disabled player collision detection to prevent the ball recoil from pushing me around wildly.

Collision detection works for dynamic as well as static scene objects. I can push balls around on the ground using the player model or by pushing a movable object such as a crate into them. They'll go up and down inside an elevator, and fall when pushed over a ledge. I can place objects on a glass surface or in a glass box, shatter the glass, and watch them fall. Even though the balls are stopped, the breaking glass will re-activate their physics and put them back into motion.

Balls will also stack when added to a container so that they're non-overlapping, and will spread out to fill in the gaps just like real world spheres. I have a colored glass box on the third floor of the office building scene that I use for lighting experiments. It's fairly large, so it takes a while to fill with balls. I marked the glass as shatterable in the config file so that I can break it with weapon fire. Here is the video showing the glass plate and glass box experiments. Note that the lighting is static/baked into the scene because the glass box isn't supposed to be destroyable, which is why the lighting looks odd near the end of the video.

This is a high resolution screenshot showing the balls stacked in the glass box. See how closely together they're packed. The collision force of the glass on the balls keeps them inside the volume. It takes several seconds for the simulation to converge to this solution and the balls to stop moving. When the glass is broken, the balls will spill out to form a single layer on the floor.

1000 bouncy balls in a colored glass enclosure. Collision forces keep the balls from intersecting each other and the glass.

There are a number of different ways to apply forces to objects in 3DWorld. Weapon explosions produce force radiating outward from a central point, which affects dynamic objects. Here I use a rocket launcher (named "Seek and Destroy") to create explosions that push the balls around the scene. The following video shows the effect of these explosions on the physics state. This also shows off my neat explosion shockwave postprocessing shader effect. The shader distorts the image by simulating a spherical region of compressed gas with higher index of refraction forming a lens.

The rocket launcher is a crazy weapon to use with zero fire delay! I wouldn't recommend this in real gameplay, it's usually instant death. The small random rocket firing error causes two rockets to occasionally collide in mid-air - they're also dynamic objects. This results in an explosion and chain reaction that sometimes propagates back to the player along the chain of close proximity in-flight rockets, resulting in a high amount of area damage. Oops. Better stick to bullets, or some other non-explosive weapon.

Wednesday, August 16, 2017

Screenshot Gallery

I've been busy with other things and haven't worked on 3DWorld much lately. It's been a while since my last blog post. I've mostly been making minor bug fixes, feature additions, and improvements. I don't have enough to say about any particular topic to create a separate post. Instead of writing another wall of text, I'll just show some interesting screenshots. Most of these are new, but a few were captured some time ago and never posted anywhere.

First up, I've added ambient occlusion to tiled terrain scenes. Well, I've had ambient occlusion for a few years now, but this version works with procedurally generated terrain as well as terrain read from heightmaps. And it uses height values computed on the GPU, which is faster. The runtime penalty of ambient occlusion is much lower now, so why not use it? It certainly adds depth to the scene.

Terrain generated using domain warping and precomputed ambient occlusion to darken the deep ravines.

3DWorld gives the user control over temperature, vegetation, atmosphere, and many other physics parameters. These variables can also be set automatically by traveling to other planets through universe mode. Not all terrain is covered with green grass, trees, and water. Here are some other biomes.

Barren moon terrain with no clouds, atmosphere, or water, only small ice caps on the mountain peaks.

I even implemented a lava mode for the water shaders so that volcanic planets can be shown.

Hot lava/rocky planet with strong wind and dense, low clouds.

I've been experimenting with volumetric spotlights that cast glowing particle cones in a dark room. This is a simple and efficient technique that draws a cone using a custom shader rather than expensive GPU ray marching. I found this blog post that explains the technique. It looks okay, but I'm not sure where to use this effect in 3DWorld. There aren't a lot of spotlights in dark, smoky rooms. I'll keep it around for future use.

Experimental volumetric fog effect for spotlights in dark basements.

This is one of my favorite 3DWorld screenshots. It was taken a few months ago. I tried to get all of the different universe objects in the same screenshot: sun, stars, nebula, planet, rings, and asteroids. There's some nice contrast between the yellow tinted foreground and purple tinted background.

One of my favorite universe screenshot images, though it's kind of dark. Stars, nebula, asteroids, and planet with rings.

Finally, here is an older screenshot showing the moment after a cluster grenade has exploded. There are a huge number of particles here, somewhere around 10,000. Each one has physics including gravity and collision detection, and all of them receive dynamic light + shadows. The triangle particles also emit their own light. Smoke and fire are drawn in many depth sorted layers with low alpha to produce a volumetric effect, which is dynamically lit with sun shadows. The light of the fires is what tints the otherwise gray-black smoke a yellowish color.

Cluster grenade explosion screenshot with ~100 light emitting particles, ~10,000 colliding physics particles, smoke, and fire.

That's it for this post. I'll put up more content when I have something new to show.

Monday, June 26, 2017

Procedural City With Buildings

I would like to have large cities and other man-made objects in 3DWorld. Terrain, vegetation, clouds, and water look fine, but the scene needs more. My large array of museum building models from a previous post is interesting, but looks too repetitive, nothing like a real city.

I started looking into procedural city creation a few weeks ago. There are tons of articles and demos of generated cities online. Most of them seem to start with city planning, then add roads (often in a grid), and finally add buildings. This works fine with flat open land areas, but probably not so well in 3DWorld procedural scenes. It's not clear how this top-down approach works with water, mountains, cliffs, and other non-flat terrain features. So I gave up on planning and roads, and went straight to building generation.

I got some of my procedural building ideas from this post by Shamus Young on building generation for the Procedural City project. 3DWorld buildings need to look good both during the day and at night, which means they need textures. At some point I'll probably also need to add detail features such as railings, antennas, AC units, etc.

To simplify things, I have only two textures per building, one for the sides and another for the roof. The easiest textures to use are brick/stone walls and arrays of office building windows. I was able to find some good ones online. A few of them also have normal maps. I'm using per-material texture scales and custom texture mapping to make the windows and bricks/blocks appear to be a consistent size across textures and building shapes. The textures are very repetitive so that they tile properly across the building, which means they don't have any features that really stand out. No doors, etc.
Buildings using plain brick or stone wall textures don't even have windows. I may need to improve on these issues in the future. It would be nice to have a texture library of the same exterior building materials and style but a variety of tiles that can be interchanged. For example, blank walls, rows of windows, door(s), etc. These would need to tile properly so that a window next to a door looks seamless. Unfortunately, it's very difficult to find acceptable texture sets online that are both free and high quality.

I've added the following building types/shapes to 3DWorld, starting with the simplest and most common:
  1. Low brick buildings (rectangle, L, T, and U shapes) with 1-3 levels
  2. Office buildings of 1-8 stacked cubes of decreasing size, possibly with cut corners
  3. Round (cylindrical or ellipsoid) office towers, possibly with a flat side
  4. Polygon shaped office towers with 3-8 sides of 1 or 2 different side lengths
  5. Large buildings composed of multiple cubes of various heights attached together
  6. Large buildings composed of angled geometric pieces attached together
Here are some example screenshots showing how building generation has advanced over time. The first one was early in the development process, and the last few are the most recent.

Early cube buildings of types 1 and 2.

Mixed cube, cylinder, and polygon buildings of types 1-4.

I have a config file format that defines different building classes, which I call materials. Each class has a large number of parameters including frequency, textures, placement rules, size variables, shape probabilities, distribution parameters, etc. The width, height, aspect ratio, number of levels, number of splits, number of sides, allowed wall angles, rotation angles, altitude ranges, and others can all be user configured. The user assigns material parameters (with ranges) to the various building types and probabilities to control their ratios for each city. I asked the system to generate 100K buildings over a radius of several square miles, and it came back with around 60K buildings. Buildings over water and on high mountains are skipped, which means that buildings cover approximately 60% of the surface within the city radius. Generation time is about half a second using 4 CPU cores.

Buildings are packed together and checked so that they don't intersect each other. They're placed at the correct height value by querying the terrain height generation function, which is where most of the generation time is spent. This works correctly for both procedural terrain and height values read from a texture. Their bases are adjusted to remove gaps between the bottom of the buildings and the ground. Trees and plants are placed in the gaps between buildings. The scales aren't completely right though. Buildings are too small compared to the player and trees, which make trees look odd when placed near tall office skyscrapers. Buildings are probably too large compared to mountains and other terrain features though. All of these sizes are configurable in 3DWorld, so I'll have to work on this later.

Buildings are pretty well integrated into 3DWorld now. They're generated on-the-fly when needed each time the terrain changes. The algorithm can be re-seeded by the user to generate a new city. I've implemented shadows from and onto buildings, and building self-shadowing for both ground and tiled terrain mode. The shader used supports dynamic and indirect lighting, and fog. I've added player collision detection with buildings; you can even walk on a building's roof. I've mostly finished the line intersection tests for building ray tracing so that they appear correctly in overhead map view:

Overhead map view of buildings, showing that ray casting is working.

Rendering is efficient because it supports block-based distance and view frustum culling, back face culling, and dynamic level of detail (LOD) for complex buildings. Draw time is 2.7ms for around 5000 visible buildings such as in the screenshots below. I could probably use multiple threads for rendering, but it's much more complex and doesn't help much. I'm currently generating the vertex data (mostly quads) on the CPU and sending it to the GPU every frame. I could precompute this once and store it on the GPU, but that could take a lot of GPU memory for 60K buildings at full detail. Keep in mind that most buildings have flat faces, so the normals are different for the corner vertices, which means the vertex data can't be shared. The data I actually have to send to the GPU is very small compared to the full set, only the front faces of the nearby buildings within the view frustum. I'm guessing that's around 1-2% of the total number of vertices/triangles. Also, it's difficult to do culling and batching of buildings by texture if the data is stored on the GPU. If I eventually add transparency from glass windows, I'll need to sort back to front by depth, and that will also complicate things.

Here are several more images of my generated city. What you see is only a tiny fraction of the total city. It extends off into the distance out to a radius several times what the view/fog distance is. The player is free to walk or fly out to the edge of the city to explore it all. It's also possible to specify more than one city at different locations.

Mixed building types 1-5 on rock and grass terrain.

Cube buildings of types 1-2 among various types of trees. Trees are probably too large compared to the buildings.

Here are some screenshots showing close-ups of interesting building types. I believe the most complex building is composed of 192 triangles, though at most 128 are visible and drawn at once. Note that every one of the 60K buildings is unique. I could use instancing, but that doesn't seem to be required to get good performance.

Various buildings with a variety of shapes, of all types. The building in the bottom left is a cube with corners cut off.

Triangular building of type 4 with cut corners on a hillside. The trees are probably too large compared to the buildings.

Complex angled office building of type 6 on a snowy peak.

Maybe I should also place larger/higher buildings in the city center, and smaller buildings near the perimeter. That can actually be done using the config file, where I can define a different placement rectangle or circle radius for each building type. I can even create multiple cities using different center and radius values, or multiple districts within one city that each have a different mix of building types. It's worth experimenting to get a more consistent and realistic look to the city. ... Okay, here is what it looks like when I place the office towers in the city center:

City skyline: Tall office buildings in the city center, smaller brick and stone buildings in the outskirts.

Huge city in a flat area with a few sharp peaks in the distance.

The next step is to generate roads or some other system to tie the buildings together. I don't know quite how to do that yet, since it either requires flattening the terrain, or curving the roads to match it. This is particularly difficult because the terrain hasn't been generated at the time of building placement. It's generated in chunks when needed, whereas the entire city is created at once. It's also unclear whether the roads should form a uniform grid, or curve to follow the terrain contour. A grid road network is likely easier, though it requires the buildings to be oriented parallel to the grid axes rather than randomly rotated like they are now. There are a lot of choices with assorted trade-offs. It's will be interesting to see how all of this turns out.

Saturday, June 17, 2017

Procedural Clouds in Tiled Terrain

I've shown tiled terrain clouds in many screenshots over various blog posts, but I've never had a post that shows off the different types/layers of clouds. The 3DWorld sky is rendered back to front as several layers:
  • Blue sky with gradient for atmospheric scattering effect (day) / starfield (night)
  • Sun flare when the sun is in view (day) / moon (night)
  • Fog/haze layer that increases in intensity with distance to blend the terrain with the sky
  • Procedural 2D wispy cloud plane layer that casts shadows on the ground
  • Procedural volumetric 3D clouds that move with the wind
Both cloud layers are drawn using screen aligned quads and noise evaluation in shaders on the GPU. The 2D cloud layer casts shadows on the terrain and grass by tracing a ray from the sun to the point on the ground and intersecting it with the cloud plane to compute light intensity. I haven't figured out how to efficiently produce shadows for the 3D cloud layer. The two types of clouds use different lighting calculations because the shaders have access to different types of data. For example, the cloud plane doesn't have local cloud density information, so it can't compute volumetric lighting like the 3D cloud system can. They don't always blend together perfectly, but in most cases it looks good enough.

3DWorld has a set of config file parameters, key bindings, and UI sliders to control the amount of clouds on the 2D and 3D layers. Clouds are also affected by weather conditions such as wind, rain, and snow. I've added an "auto" mode with day/night cycle that dynamically varies the weather and cloud cover over time to simulate a real environment.

Here is a set of screenshots showing the various weather and cloud conditions that can be selected. I've turned off the birds in the sky for most of these images so that the clouds can be seen clearly.

Sunny day with sun lens flare and minimal cloud cover.

Light, wispy, high clouds, with distant fog.

Medium wispy clouds and haze. These clouds cast dynamic shadows on the ground.

High, dense clouds, with small areas where the sun peeks out.

Scattered small 3D cloud puffs.

Mixed 3D low lying volumetric clouds and high 2D cloud layers.

Heavy cloud cover with mixed cloud types. The lighting doesn't really match between cloud types.

Storm clouds with rain. Are those city buildings in the background?

Clouds colored red at sunset.

Night time clouds in a starry sky.