Interior Mapping – Part 2

In part 1, we discussed the requirements and rationale behind Interior Mapping. In this second part, we’ll discuss the technical implementation of what I’m calling (for lack of a better title) “Tangent-Space Interior Mapping”.

Coordinates and Spaces

In the original implementation, room volumes were defined in object-space or world-space. This is by far the easiest coordinate system to work in, but it quickly presents a problem! What about buildings with angled or curved walls? At the moment, the rooms are bounded by building geometry, which can lead to extremely small rooms in odd corners and uneven or truncated walls!

In reality, outer rooms are almost always aligned with the exterior of the building. Hallways rarely run diagonally and are seldom narrower at one end than the other! We would rather have all our rooms aligned with the mesh surface, and then extruded inward towards the “core” of the building.

Cylindrical Building

Curved rooms, just by changing the coordinate basis.

In order to do this, we can just look for an alternative coordinate system for our calculations which lines up with our surface (linear algebra is cool like that). Welcome to Tangent Space! Tangent space is already used elsewhere in shaders. Even wonder why normal-maps are that weird blue color? They actually represent a series of directions in tangent-space, relative to the orientation of the surface itself. Rather than “Forward”, a Z+ component normal map points “Outward”. We can simply perform the raycast in a different coordinate basis, and suddenly the entire problem becomes surface-relative in world-space, while still being axis-aligned in tangent space! A neat side-effect of this is that our room volumes now follow the curvature of the building, meaning that curved facades will render curved hallways running their length, and always have a full wall parallel to the building exterior.

While we’re at it, what if we used a non-normalized ray? Most of the time, a ray should have a normalized direction. “Forward” should have the same magnitude as “Right”. If we pre-scale our ray direction to match room dimensions, then we can simplify it out of the problem. So now, we’re performing a single raycast against a unit-sized axis-aligned cube!

Room Textures

The original publication called for separate textures for walls, floors, and ceilings. This works wonderfully, but I find it difficult to work with. Keeping these three textures in sync can get difficult, and atlasing multiple room textures together quickly becomes a pain. Alternative methods such as the one proposed by Zoe J Wood in “Interior Mapping Meets Escher” utilizes cubemaps, however this makes atlasing downright impossible, and introduces new constraints on the artists building interior assets.
Andrew Willmott briefly touched on an alternative in “From AAA to Indie: Graphics R&D”, which used a pre-projected interior texture for the interior maps in SimCity. This was the format I decided to use for my implementation, as it is highly author-able, easy to work with, and provides results only slightly worse than full cubemaps. A massive atlas of room interiors can be constructed on a per-building basis, and then randomly selected. Buildings can therefore easily maintain a cohesive interior style with random variation using only a single texture resource.

Finally, The Code

I’ve excluded some of the standard Unity engine scaffolding, so as to not distract from the relevant code. You won’t be able to copy-paste this, but it should be easier to see what’s happening as a result.

v2f vert (appdata v) {
   v2f o;
   // First, let's determine a tangent basis matrix.
   // We will want to perform the interior raycast in tangent-space,
   // so it correctly follows building curvature, and we won't have to
   // worry about aligning rooms with edges.
   half tanSign = v.tangent.w * unity_WorldTransformParams.w;
   half3x3 objectToTangent = half3x3(,
      cross(v.normal, v.tangent) * tanSign,

   // Next, determine the tangent-space eye vector. This will be
   // cast into an implied room volume to calculate a hit position.
   float3 oEyeVec = v.vertex - WorldToObject(_WorldSpaceCameraPos);
   o.tEyeVec = mul(objectToTangent, oEyeVec);

   // The vertex position in tangent-space is just the unscaled
   // texture coordinate.
   o.tPos = v.uv;

   // Lastly, output the normal vertex data.
   o.vertex = UnityObjectToClipPos(v.vertex);
   o.uv = TRANSFORM_TEX(v.uv, _ExteriorTex);

   return o;

fixed4 frag (v2f i) : SV_Target {
   // First, construct a ray from the camera, onto our UV plane.
   // Notice the ray is being pre-scaled by the room dimensions.
   // By distorting the ray in this way, the volume can be treated
   // as a unit cube in the intersection code.
   float3 rOri = frac(float3(i.tPos,0) / _RoomSize);
   float3 rDir = normalize(i.tEyeVec) / _RoomSize;

   // Now, define the volume of our room. With the pre-scale, this
   // is just a unit-sized box.
   float3 bMin = floor(float3(i.tPos,-1));
   float3 bMax = bMin + 1;
   float3 bMid = bMin + 0.5;

   // Since the bounding box is axis-aligned, we can just find
   // the ray-plane intersections for each plane. we only 
   // actually need to solve for the 3 "back" planes, since the 
   // near walls of the virtual cube are "open".
   // just find the corner opposite the camera using the sign of
   // the ray's direction.
   float3 planes = lerp(bMin, bMax, step(0, rDir));
   float3 tPlane = (planes - rOri) / rDir;

   // Now, we know the distance to the intersection is simply
   // equal to the closest ray-plane intersection point.
   float tDist = min(min(tPlane.x, tPlane.y), tPlane.z);

   // Lastly, given the point of intersection, we can calculate
   // a sample vector just like a cubemap.
   float3 roomVec = (rOri + rDir * tDist) - bMid;
   float2 interiorUV = roomVec.xy * lerp(INTERIOR_BACK_PLANE_SCALE, 1, roomVec.z + 0.5) + 0.5;

   // If the room texture is an atlas of multiple variants, transform
   // the texture coordinates using a random index based on the room index.
   float2 roomIdx = floor(i.tPos / _RoomSize);
   float2 texPos = floor(rand(roomIdx) * _InteriorTexCount) / _InteriorTexCount;

   interiorUV /= _InteriorTexCount;
   interiorUV += texPos;

   // lastly, sample the interior texture, and blend it with an exterior!
   fixed4 interior = tex2D(_InteriorTex, interiorUV);
   fixed4 exterior = tex2D(_ExteriorTex, i.uv);

   return lerp(interior, exterior, exterior.a);

And that’s pretty much all there is to it! The code itself is actually quite simple and, while there are small visual artifacts, it provides a fairly convincing representation of interior rooms!


Interior + Exterior Blend


There’s definitely more room for improvement in the future. The original paper supported animated “cards” to represent people and furniture, and a more realistic illumination model may be desirable. Still, for an initial implementation, I think things came out quite well!

Interior Mapping – Part 1

Rendering convincing environments in realtime has always been difficult, especially for games which take place at a “human” scale. Games consist of a series of layered illusions and approximations, all working (hopefully) together to achieve a unified goal; to represent the world in which the game takes place. In the context of a simplified or fantastical world, this isn’t too bad. It’s a matter of creating a unified style and theme that feels grounded in the reality of the particular game. The fantastic narrative platformer “Thomas was Alone”, for example, arguably conveys a believable world using just shape and color. As soon as a game takes place in an approximation of our real world however, the cracks start to appear. There are a tremendous number of “details” in the real world. Subtle differences on seemingly identical surfaces that the eye can perceive, even if not consciously.

uncanny valley

This CG incarnation of Dwayne Johnson as the titular “Scorpion King” is a prime example of “The Uncanny Valley”

We as humans are exceptionally good at identifying visual phenomena, and more importantly, its absence. You may have heard this referred to as “The Uncanny Valley”; when something is too realistic to be considered cute or cartoony, but too unrealistic to look… right… It’s extremely important to include some representation of those “missing” pieces, even if they’re not 100% accurate in order to preserve the illusion.

While not nearly as noticeable at first glance, missing details in an environment are equally important to preserving the illusion of a living, breathing, virtual world.

Take, for example, this furniture store from GTA IV.

GTA Furniture Store.png

A nice looking furniture store, though something’s missing…

This is a very nice piece of environment art. It’s visually interesting, it fits the theme and location, and it seems cohesive within the world… though something is amiss. The view through the windows is clearly just a picture of a store, slapped directly onto the window pane, like a sticker on the glass! There’s no perspective difference between the individual windows on different parts of the facade. The view of the interior is always head-on, even if the camera is at an angle to the interior walls. This missing effect greatly weakens the illusion.

From this, the question arises…

How do we convey volume through a window, without creating tons of work for artists, or dramatically altering the production pipeline?

Shader Tricks!

The answer (as you may have guessed from the header) lies in shader trickery! To put it simply, Shaders are tiny programs which take geometric information as input, mush it around a bunch, and output a color. Our only concern is that the final output color looks correct in the scene. What happens in the middle frankly doesn’t matter much. If we offset the output colors, we can make it look like the input geometry is offset too! If outputs are offset non-uniformly, it can be made to appear as though the rendered image is skewed, twisted, or distorted in some way.

uncanny valley

If you’ve ever seen at 3D sidewalk art, you’ve seen a real-world implementation of parallax mapping.

The school of techniques collectively known as “Parallax Mapping” do just this. Input texture coordinates are offset based on the observer angle, and a per-texel “depth” value. By determining the point where our camera ray intersects the surface height-field, we can create what amounts to a 3D projection of an otherwise 2D image. “Learn OpenGL” provides an excellent technical explanation of parallax mapping if you’re curious.

While the theory is perfect for our needs, the methodology is lacking. Parallax mapping is not without its issues! Designed to be a general-purpose solution, it suffers from a number of visible artifacts when used in our specific case. It works best on smoother height-fields, for instance. Large differences in height between texels can create weird visual distortions! There are a number of alternatives to get around this issue (such as “Steep Parallax Mapping”), but many are iterative, and result in odd “step” artifacts as the ratio of depth to iteration count increases. In order to achieve a convincing volume for our buildings using an unmodified parallax shader, we’d need to use so many iterations that it would quickly become a performance nightmare.

Interior Mapping

Parallax mapping met nearly all of our criteria, but still wasn’t suitable for our application. Whenever a general solution fails, it’s usually a good idea to sit down and consider the simplest possible specific solution that will suffice.


For each point on the true geometry (blue), select a color at the point of intersection between the camera ray, and an imaginary room volume (red).

In our case, we want rectangular rooms inset into our surface. The keyword here is “rectangular”. The generality of parallax mapping means that an iterative numeric approach must be used, since there is no analytical way to determine where our camera ray intersects a general height-field. If we limit the problem to only boxes, then an exact solution is not only possible, but trivial! Furthermore, if these boxes are guaranteed to be axis-aligned, the computation becomes extremely simple! Then, it’s just a matter of mapping the point of intersection within our room volume to a texture, and outputting the correct color!

interior mapping example

Example of “Interior Mapping” from the original publication by Joost van Dongen.

Originally published in 2008, the now well known “Interior Mapping”, by Joost van Dongen seems like a prime candidate! In this approach, the facade of a building mesh is divided into “rooms”, and a raycast is performed for each texel. Then, the coordinate at the point of intersection between our camera ray and the room volume can be used to sample a set of “Room Textures”, and voila! This, similar to parallax mapping, offsets input texture coordinates to provide a projection of a wall, ceiling, and floor texture within each implicit “room volume”, resulting in a geometrically perfect representation of an interior without the added complexity of additional geometry and material work!

In part 2, we’ll discuss modifications to the original implementation for performance and quality-of-life improvements!

Abusing Blend Modes for Fun and Profit!

Today I decided to do a quick experiment.

Hardware “blend modes” have existed since the dawn of hardware-accelerated graphics. Primarily used for effects like transparency, they allow a developer to specify the way new colors are drawn into the buffer through a simple expression.

color = source * SrcFactor + destination * DstFactor

The final output color is the sum of a “source factor” term multiplied by the value output by the fragment shader, and a “destination factor” term multiplied by the color already in the buffer.

For example, if I wanted to simply add the new color into the scene, I could use blend modes of One One; Our coefficients would be negligible and we would end up with

color = source + destination

If I wanted a linear alpha blend between the source color and destination color, I could select the terms SrcAlpha, OneMinusSrcAlpha, which would perform a linear interpolation between the two colors.

But what happens when we have non-standard colors? Looking back at the blend expression, logic would dictate that we can express any two-term polynomial as long as the terms are independent, and the coefficients are one of the supported “blend factors”! By pre-loading our destination buffer with a value, the second term can be anything we need, and the alpha channel of our source can be packed with a coefficient to use as the destination factor if need be.

This realization got me thinking. “Subtract” blend modes aren’t explicitly supported in OpenGL, however a subtraction is simply the addition of a negative term. If our source term were negative, surely blend factors of One One would simply subtract the source from the destination color! That isn’t to say that this is guaranteed to work without issues! If the render target is a traditional 24 or 32-bit color buffer, then negative values may have undefined behavior! A subtraction by addition of a negative would only work assuming the sum is calculated independently somewhere in hardware, before it’s packed into the unsigned output buffer.

Under these assumptions, I set out to try my hand at a neat little trick. Rendering global object “thickness” in a single pass.

Why though?

Thickness is useful for a number of visual effects. Translucent objects, for example, could use the calculated thickness to approximate the degree to which light is absorbed along the path of the ray. Refraction could be more accurately approximated utilizing both incident, and emergent light calculations. Or, you could define the shape of a “fog volume” as an arbitrary mesh. It’s actually quite a useful thing to have!

Single pass global thickness maps

So here’s the theory. Every pixel in your output image is analogous to a ray cast into your scene. It can be thought of as a sweep backwards along the path of light heading towards the camera. What we really want to determine is the point where that ray enters and exits our object. Knowing these two points, we also essentially know the distance travelled through the volume along that ray.

It just so happens that we know both of these things! The projective-space position of a fragment must be calculated before a color can be written into a buffer, so we actually know the location of every fragment, or continuing the above analogy, ray intersection on the surface. This is also true of the emergent points, which all lie on the back-faces of our geometry! If we can find the distance the ray has traveled before entering our volume, and the distance the ray has traveled before exiting it, the thickness of the volume is just the difference of the two!

So how is this possible in a single pass? Well, normally when we render objects, we explicitly disable the “backfaces”; triangles pointing away from our camera. This typically speeds things up quite a bit, because backfaces almost certainly lie behind the visible portion of our model, and shading them is simply a waste of time. If we render them however, our fragment program will be executed both on the front and back faces! By writing the distance from the camera, or “depth” value as the color of our fragment, and negating it for front-faces, we can essentially output the “back minus front” thickness value we need!

DirectX provides a convenient semantic for fragment programs. float:VFACE. This value will be set to 1 when the fragment is part of a front-face, and -1 when the fragment is part of a back-face. Just render the depth, multiplied by the inverted value of the VFACE semantic, and we’ve got ourselves a subtraction!

Cull Off // disable front/back-face culling
Blend One One // perform additive (subtractive) blending
ZTest Off // disable z-testing, so backfaces aren’t occluded.

fixed4 frag (v2f i, fixed facing : VFACE) : SV_Target {
return -facing * i.depth;

Unity Implementation

From here, I just whipped up a quick “Camera Replacement Shader” to render all opaque objects in the scene using our thickness shader, and drew the scene to an off-screen “thickness buffer”. Then, in a post-effect, just sample the buffer, map it to a neat color ramp, and dump it to the screen! In just a few minutes, you can make a cool “thermal vision” effect!


The subtraction blend isn’t necessarily supported on all hardware. It relies on a lot of assumptions, and as such is probably not appropriate for real applications. Furthermore, this technique really only works on watertight meshes. Meshes with holes, or no back-faces will have a thickness of negative infinity, which is definitely going to cause some problems. There are also a number of “negative poisoning” artifacts, where the front-face doesn’t necessarily overlap a corresponding backface, causing brief pixel flickering. I think this occasional noise looks cool in the context of a thermal vision effect, but there’s a difference between a configurable “glitch” effect, and actual non-deterministic code!

Either way, I encourage everyone to play around with blend-modes! A lot of neat effects can be created with just the documented terms, but once you get into “probably unsafe” territory, things start to get really interesting!

GPU Isosurface Polygonalization

Isosurfaces are extremely useful when it comes to data visualization. From medical imaging to fluid flow analysis, they are an excellent tool for understanding complex volumetric data. Games have also adopted some of these techniques for their on purposes From the more rigid implementation in the ubiquitous game Minecraft to the Gels in Portal 2, these techniques serve the same basic purpose.

I wanted to try my hand at a general-purpose implementation, but before we dive into things, we must first answer a few basic questions.

What is an isosurface?

An isosurface can be thought of as the solution of a continuous function which produces a constant output in 3D. If you’re visualizing an electromagnetic field for example, you might generate an isosurface for a given potential, so you can easily determine its overall shape. This technique can be applied to other arbitrary values as well. Given CT scan data, a radiologist could construct an isosurface at the density of a specific type of tissue, extracting a 3D representation of bones or organs to view them separately, rather than having to manipulate a less intuitive stack of images.

What will we use this system for?

I don’t work in the medical field, nor would I trust the accuracy of my implementation when it comes to making a diagnosis. I work in entertainment and computer graphics, and as you would imagine, the requirements are quite different. Digital artists can already craft far better visuals than any procedure yet known; the real challenge is dynamic data. Physically simulated fluids, player-modifiable terrain, mechanics such as these present a significant challenge for traditional artists! What we really need is a generalized system for extracting and rendering isosurfaces in real time to fill in the gaps.

What are the requirements for such a system?

Given our previous use case, we can derive a few basic requirements. In no particular order…

  1. The system must be intuitive. Designers have other things to do besides tweaking simulation volumes and fiddling with configurations.
  2. The system must be flexible. If someone suggests a new mechanic which relies heavily on procedural geometry, it should be easy to get up and running.
  3. The system must be compatible. The latest experimental extensions are fun, but if you want to release something that anyone can enjoy, it needs to run on 5 year old hardware.
  4. The system must be fast. At 60 fps, you only have 16ms to render everything in your game. We can’t spend 10 of that drawing a special effect.

Getting Started!

Let’s look at requirement no. 4 first. Like many problems in computing, surface polygonalization can be broken down into repeated instances of much smaller problems. At the end of the day, the desired output is a series of interconnected polygons which appear to make up a complex surface. If each of these component polygons is accounted for separately, we can dramatically reduce the scope of the problem. Instead of generating a polygonal surface, we are now generating a single polygon, which is a much less daunting task. As with any discretization process, it is necessary to define a regular sample interval at which our continuous function will be evaluated. In the simplest 3D case, this will take the form of a regular grid of cells, and each of these cells will form a single polygon. Suddenly, this polygonalization process becomes massively parallel. With this new outlook, the problem becomes a perfect fit for standard graphics hardware!

For compatibility, I chose to implement this functionality in the Geometry Shader stage of the rendering pipeline, as it allows for the creation of arbitrary geometry given some basic input data. A Compute Shader would almost definitely be a better option in terms of performance and maintainability, but my primary development system is OSX, which presents a number of challenges when it comes to the use of Compute Shaders. I intend to update this project in the future, once Compute Shaders become more common.

If the field is evaluated at a number of regular points and a grid is drawn between them, we can construct a set of hypothetical cubes with a single sample at each of its 8 vertices. By comparing the values at each vertex, it is trivial to determine the plane of intersection between the theoretical isosurface and the cubic sample volume. If properly evaluated, the local solutions for each sample volume will form integral parts of the global surface implicitly, without requiring any global information.

This is the basic theory behind the ubiquitous Marching Cubes algorithm, first published in 1987 and still commonly used today. While it is certainly battle-tested, there are a number of artifacts in the output geometry that can make surfaces appear rough. The generated triangles are also often non-uniform and narrow, leading to additional artifacts later in the rendering process. Perhaps a more pressing issue is the sheer number of cases to be evaluated. For every sample cell, there are 256 possible planar intersections. The fantastic implementation by Paul Bourke wisely recommends the use of a look-up table, pre-computing these cases. While this may work well in traditional implementations, it crumbles under the parallel architecture of modern GPUs. Graphics hardware tends to excel at executing large batches of identical instructions, but begins to falter as soon as complex conditional branching is involved and operations have to be evaluated individually. In my tests, I found that the look-up tables performed no better, if not worse than explicit evaluation, as the complier could not easily expand and unroll the program flow, and evaluation could not be easily batched. Bearing this in mind, we ideally need to implement a method with as few logical branches as possible.

Marching Tetrahedra is a variant of the Marching Cubes algorithm which divides each cube into 5 (or 6 for a slightly different topology) tetrahedra. By limiting our integral sample volume to four vertices instead of 8, the number of possible cases drops to 16. In my tests, I got a 16x performance improvement using this technique (though realized savings are heavily dependent on the hardware used), confirming our suspicions. Unfortunately, marching tetrahedra can have some strange surface features, and has a number of artifacts of its own, especially with dynamic sampling grids.

Because of this, I ended up settling on naive surface nets, a simple dual method which generates geometry spanning multiple voxel sample volumes. An excellent discussion of the differences between these three meshing algorithms can be found here. Perhaps my favorite advantage of this method is the relative uniformity of the output geometry. Surface nets tend to be smooth surfaces comprised of quads of relatively equal size. In addition, I find it easier to comprehend and follow than other meshing algorithms, as its use of look-up-tables, and possible cases is fairly limited.

Implementation Details


The sample grid is actually defined as a mesh, with a single disjoint vertex placed at each integral sample coordinate. These vertices aren’t actually drawn, but instead are used as input data to a series of shaders. Therefore, the shader can be considered to be executed “per-voxel”, with its only input being the coordinate of the minimum bounding corner. One disadvantage commonly seen in similar systems is a fundamental restriction on surface resolution due to a uniform sample grid. In order to skirt around this limitation, meshing is actually performed. In projected space, rather than world space, so each voxel is a truncated frustum similar to that of the camera, rather than a cube. This not only eliminates a few extra transformations in the shader code, but provides LoD implicitly by ensuring each output triangle is of a fixed pixel size, regardless of its distance to the camera.

Once the sample mesh was created, I used a simple density function for the potential field used by this system. It provides a good amount of flexibility, while still being simple to comprehend and implement. Each new source of “charge” added to the field would contribute additively to the overall potential.  However, this quickly raises a concern! Each contributing charge must be evaluated at all sample locations, meaning our shader must, in some way, iterate through all visible charges! As stated earlier, branching and loops which cannot be unrolled can cause serious performance hiccups on most GPUs!

While I was at it, I also implemented a Vertex Pre-pass. Due to the nature of GPU parallelism, each voxel is evaluated in complete isolation. This has the unfortunate side-effect of solving for each voxel vertex position up to 6 times (once for each neighboring voxel). The surface net algorithm utilizes an interpolated surface vertex position, determined from the intersections of the surface and the sample volume. This interpolation can get expensive if repeated 6 times more than necessary! To remedy this, I instead do a pre-pass calculating the interpolated vertex position, and storing it as a normalized coordinate within the voxel in the pixel color of another texture. When the geometry stage builds triangles, it can simply look up the normalized vertex positions from this table, and spit them out as an offset from the voxel min coordinate!

The geometry shader stage is then fairly straightforward, reading in vertex positions, checking the case of the input voxel, looking up the vertex positions of its neighbors, and bridging the gap with a triangle.

Was it worth it?

Short answer, no.

I am extremely proud of the work I’ve done, and the end result is quite cool, but it’s not a solution I would recommend in a production setting. The additional complexity far outweighs any potential performance benefit, and maintainability, while not terrible, takes a hit as well. In addition, the geometry shader approach doesn’t work nearly as well as I had hoped. Geometry shaders are notoriously cache-unfriendly, and my implementation is no exception. Combine this with the rather unintuitive nature of working with on-GPU procedural geometry in a full-scale project, and you’ve got yourself a recipe for very unhappy engineers.

I think the concept of on-GPU surface meshing is fascinating, and I’m eager to look into Compute Shader implementations, but as it stands, the geometry stage is not the way to go.

I’ve made the source available on my GitHub if you’d like to check it out!

Keeping track of AssetBundles

Sometimes it’s necessary to load and unload content at runtime. Games which take place in an expansive, explorable world, for instance, probably shouldn’t load the entire play-space up front. Every area, from the dankest catacomb to the loftiest castle, would need to be read from long-term storage and buffered into memory before the game could be played. Doing so would contribute to long load-times, and cause often insurmountable memory problems! Unfortunately, we don’t yet live in an era where a game-maker can reasonably expect players to have computers, consoles, and mobile devices capable of loading 30 gigabytes of dirt textures into memory.

The Unity game engine has long struggled with this problem. Multiple solutions for asset streaming exist, but all are far from perfect.

Perhaps the most convenient streaming solution in Unity is the Resources API. Added in the early days of the engine, it allows assets to be loaded and unloaded using two simple functions.

public static T Resources.Load<T>(string path);

public static void Resources.UnloadAsset(Object assetToUnload);

Look at that! It’s a simple interface which can easily be rolled into any custom asset management code you might want! It’s easy to read, intuitive, requires no special handling… and it’s terrible…

What’s wrong with Resources?

Unity itself strongly recommends against using the Resources API. To reiterate the official arguments, the Resources API makes it much more difficult to manage memory carefully. This might seem like a moot point, but on memory constrained platforms like Mobile devices, this can cause issues. Fragmentation of the heap is a very real concern, especially when loading and unloading many large objects. Additional, the official injunction omits the fact that the Resources API compiles all referenced resources into a single bundle at build-time, meaning the maximum total size of your resources are restricted to the maximum size of a single file on the user’s machine! On most machines, this will be either 2 or 4 gigabytes! Not nearly enough to represent that massive world you and your team have been planning!

So what’s the alternative?

Unity added an alternative method for streaming assets sometime in the late 2000’s known as “Asset Bundles“. AssetBundles are compressed archives of content which can be loaded and unloaded on the fly in your application. Games can even download AssetBundles from a server to perform partial updates, and pull down large chunks of data at a time (useful for games like MMOs, which can’t feasibly store the entire world at once).

Yes, AssetBundles are wonderful, but all that flexibility isn’t without its downsides. Long gone are the days of “Resources.Load()”. AssetBundles require a slew of management. Loading bundles, loading dependencies, shifting around manifests, and compulsively version-checking your data are unfortunate requirements of the system. To poorly paraphrase Uncle Ben from the 2002 film adaptation of Spider-Man, “With great interface flexibility, comes a need for specificity.”

What can be done?

The job of a programmer is one of abstraction. I set out to write a facade over the AssetBundle system as an “exploratory exercise.” Surely, there’s a way to keep track of assets, their dependencies, and their lifespans without having to manually modify each piece of code we write!

I decided to adapt a concept from the Objective-C runtime. Apple’s OSX and iOS SDKs contain a memory-managment system known as ARC (Automatic Reference Counting). Essentially, what ARC does is keep a counter running for every instance of an object in your application. When an object is referenced somewhere, that counter is incremented. When that reference goes out of scope, the counter is decremented. When the counter reaches zero, the object is destroyed.

This has the net effect of “automatically” keeping track of when an object is referenced, and when it is no longer needed, similar to the “garbage collection” systems present in many environments (including Unity’s .NET execution environment). Unlike Garbage Collection however, it tends to incur a less noticeable runtime performance cost, as instances are freed from memory continuously rather than in an occasional “clean up” pass. This is not without its downsides, but that’s another story for another post.

The important thing is that the theory can be applied to AssetBundles. Game assets are a good use case for a system like this. Assets must be loaded on demand, but in environments with limited memory, must be freed as soon as possible. Dependencies are also a very real concern. AssetBundles may depend on others for content, and it’s important to know which bundles can and can’t be unloaded safely during the execution of an application. I figured it was worth a shot, and spent a few hours looking into things…

Automatic Reference-Counted StreamedAsset API

The StreamedAsset API is fairly simple and not without its own set of issues, however as a first experiment, it does a solid job of mimicking the simplicity of the old “Resources” API.

First, I define a generic class called “StreamedAsset”. This class will act as a wrapper to handle reference-counting of an internally managed Object instance!

public sealed partial class StreamedAsset<AssetType> where AssetType : Object {

public readonly string bundleName;
public readonly string assetName;

public StreamedAsset(string bundleName, string assetName) {
this.bundleName = bundleName;
this.assetName = assetName;


~StreamedAsset() {

public static implicit operator AssetType(StreamedAsset<AssetType> streamedAsset) {
LoadAsset(streamedAsset.bundleName, streamedAsset.assetName);
return m_loadedAssets[streamedAsset.assetName] as AssetType;


First, a StreamedAsset contains a “bundleName” and “assetName” field. These fields contain the name of the AssetBundle containing the desired asset, and the name of that asset itself, respectively.

You’ll notice two lines in both the initializer, and finalizer of this class…


These methods increment and decrement the reference counter for the bundle and asset within that bundle. This makes it externally impossible to construct an instance of “StreamedAsset” without updating the reference counters corresponding to the asset data. The finalizer is called automatically whenever this object is destroyed by the garbage collector, so our reference counters will be correctly decremented sometime after this object goes out of scope, whenever the runtime decides it’s safe to free unnecessary memory.

You’ll also notice the “Implicit Operator” towards the end of the definition. This is how we unwrap our StreamedAsset references. The implementation of this conversion operator means that they are implicitly convertible to the contained generic type. This allows StreamedAsset instances to be used identically to a traditional reference to our asset data.

// Works

renderer.material = new Material();

// Also works!

renderer.material = new StreamedAsset<Material>(“myBundle”, “matName”);

Included in this implicit operator is a call to the “LoadAsset” function. This function will do nothing if the object is loaded, but has the effect of lazily loading the asset in question. Therefore, StreamedAsset references can be instantiated, duplicated, and passed around the application without ever actually loading the asset from the AssetBundle! You can define a thousand references to a thousand assets, but until they’re actually used for something, they remain unloaded.  Placing the lazy load function in the implicit unwrap operator also allows assets to be unloaded when they’re referenced, but not used (though this behavior is not implemented in this demo project).

Now, our asset exists, but what about the actual loading and unloading?

StreamedAsset Internals

I made the StreamedAsset API a partial class, allowing it to be separated into multiple files. While the management of asset data and the asset reference type should be contained in different places, the management should still be completely internal to the StreamedAsset type. They are fundamentally inseparable, and it should not be externally accessible!

public sealed partial class StreamedAsset<AssetType> {

private static SynchronizationContext m_unitySyncContext;

private static Dictionary<string, AssetBundle> m_loadedBundles = new Dictionary<string, AssetBundle>();
private static Dictionary<string, Object> m_loadedAssets = new Dictionary<string, Object>();

private static Dictionary<string, uint> m_bundleRefCount = new Dictionary<string, uint>();
private static Dictionary<string, uint> m_assetRefCount = new Dictionary<string, uint>();

static StreamedAsset() {
m_unitySyncContext = SynchronizationContext.Current;

private static void ReferenceAsset(string name) {
if (!m_assetRefCount.ContainsKey(name))
m_assetRefCount.Add(name, 0);
m_assetRefCount[name] ++;


private static void DereferenceAsset(string name) {
m_assetRefCount[name] –;

if (m_assetRefCount[name] <= 0) {
// Dereferencing is handled through finalizers, which are run on
// background threads. Execute the unloading on the Unity sync context.
m_unitySyncContext.Post(_ => {
}, null);

private static void ReferenceBundle(string name) {
if (!m_bundleRefCount.ContainsKey(name))
m_bundleRefCount.Add(name, 0);
m_bundleRefCount[name] ++;


private static void DereferenceBundle(string name) {
m_bundleRefCount[name] –;

if (m_bundleRefCount[name] <= 0) {
// Dereferencing is handled through finalizers, which are run on
// background threads. Execute the unloading on the Unity sync context.
m_unitySyncContext.Post((context) => {
UnloadBundle(context as string);
}, name);

private static void LoadBundle(string bundleName) {
if (m_loadedBundles.ContainsKey(bundleName))

var path = System.IO.Path.Combine(Application.streamingAssetsPath, bundleName);
var bundle = AssetBundle.LoadFromFile(path);
m_loadedBundles.Add(bundleName, bundle);

private static void UnloadBundle(string bundleName) {
var bundle = m_loadedBundles[bundleName];

if (bundle != null) {

private static void LoadAsset(string bundleName, string assetName) {
if (m_loadedAssets.ContainsKey(assetName))

var asset = m_loadedBundles[bundleName].LoadAsset(assetName);
m_loadedAssets.Add(assetName, asset);

private static void UnloadAsset(string assetName) {
var asset = m_loadedAssets[assetName];

// if (asset != null) {
//  Resources.UnloadAsset(asset);
// }

This portion of the StreamedAsset class does exactly what it looks like, implementing functions to load, and unload asset bundles, as well as method to increment and decrement reference counts.

An important thing to note is the use of a “SynchronizationContext” to unload assets. The finalizer for object instances in Unity is executed on a background thread dedicated to garbage collection. As a result, all functions called from a finalizer will be executed on this background thread. Unfortunately, Unity’s Scripting API is not thread-safe! The static initializer is therefore used to capture a reference to the SynchronizationContext of Unity’s main thread, and all requests to unload assets are handled through this context.

Another note is the use of the “Resources.UnloadUnusedAssets()” method. While the Resources API is deprecated, a number of asset management functions are still grouped under the “Resource” umbrella. “Resources.UnloadAsset()”, and “Resources.UnloadUnusedAssets()” can actually be used to unload assets loaded from AssetBundles. This is never explicitly documented, however it is supported, and is clearly intended functionality. In the example, “UnloadUnusedAssets()” is used because it also unloads the dependencies of the dereferenced asset. This function has large performance implications, and should probably not be called so liberally, but as stated earlier, is useful for prototyping.

That’s really all there is to it! A custom build script can be included to construct bundles of streamed assets for the target platform, and copy them into the StreamingAssets “magic directory”. From there, the StreamedAsset API can be used to load them on the fly without ever having to worry about the nitty gritty of managing references!

Room for Improvement!

This is clearly not a perfect solution! The first and foremost issue is that the lifecycle of a streamed asset is tied to the lifecycle of its “StreamedAsset” references, rather than the asset itself.

For example, assigning a implicitly converted StreamedAsset directly to a built-in Unity component will cause that asset to be unloaded as soon as garbage is collected. Assets must instead be maintained at a higher level than function-scope.

private StreamedAsset m_goodMat = new StreamedAsset<Material>(“myBundle”, “myMat”);

void Start() {

var badMat = new StreamedAsset<Material>(“myBundle”, “myMat”);

obj1.GetComponent<Renderer>().material = badMat

// At this point, “badMat” will go out of scope, and the material will become null at some point in the future.

obj2.GetComponent<Renderer>().material = m_goodMat;

// “m_goodMat” however will share the life cycle of this behaviour instance, and will persist for the lifespan of the object.


This is an unfortunate downside of the implicit conversion. As far as I can tell, it isn’t possible to retroactively embed automatic reference counting in the UnityEngine.Material instance itself, though a higher-level “management” solution warrants further investigation.

Final Words

I’m fairly happy with this little experiment, though I would advise against using it in a production environment without further testing. Regardless, I think reference-counted assets have potential in larger games. Not having to worry explicitly about asset loading and unloading trades performance for ease of use, but for many games, I’m willing to bet that’s a worthwhile trade. I’m curious to see what else can be done with a more “automatic” system such as this!

Messing With Shaders – Realtime Procedural Foliage


The programmable rendering pipeline is perhaps one of the largest advances in the history of realtime computer graphics. Before its introduction, graphics libraries like OpenGL and DirectX were limited to the “fixed function pipeline”, a programmer would shove in geometric data, and the application would draw it however it saw fit. Developers had little to no control over the output of their application beyond a few “render mode” settings. This was fine for rendering relatively simple scenes, solid objects, and simplistic lighting, but as visual fidelity increased and hardware become more powerful it quickly became necessary to allow for a more customizable rendering.

The process of rendering a 3D object in the modern programmable pipeline is typically broken down into a number of steps. Data is copied into fast-access graphics memory, then transformed through a series of stages before the graphics hardware eventually rasterizes that data to the display. In its most basic form, there are two of these stages the developer can customize. The “Vertex Program” manipulates data on a per-vertex level, such as positions and texture coordinates, before handing the results on to the “Fragment Program”, which is responsible for determining the properties of a given fragment (like a pixel containing more than just color information). The addition of just these two stages opened the floodgates for interesting visual effects. Approximating reflections for metallic objects, cel-shading effects for cartoon characters, and more! Since then, even more optional stages have been inserted into the pipeline for an even greater variety of effects.

I’ve spent a considerable amount of time experimenting with vertex and fragment programs in the past, but this week I decided to spend a few hours working with the other, less common stages, mainly “Geometry Programs”. Geometry programs are a more recent innovation, and have only began to see extensive use in the last decade or so. They essentially allow developers to not only modify vertex data as it’s received, but to construct entirely new vertices based on the input primitives (triangles, quads, etc.) As you can easily imagine, this presents incredible potential for new effects, and is something I personally would like to become more experienced with.

In four or five hours, I managed to write a relatively complex effect, and the rest of this post will detail, at a high level, what I did to achieve it.


Procedurally generated geometry for ivy growing on a simple building.

This is my procedural Ivy shader. It is a relatively simple two-pass effect which will apply artist-configurable ivy to any surface. What sets this effect apart from those I’ve written in the past is that it actually constructs new geometry to add 3D leaves to the surface extremely efficiently.

One of the major technical issues when it comes to rendering things like foliage is that the level of geometric detail required to accurately represent leaves is quite high. While a digital environment artist could use a 3D modeling program to add in hundreds of individual leaves, this is not necessarily a good use of their time. Furthermore, it quickly becomes unmaintainable if anyone decides that the position, density, or style of foliage should change in the future. I don’t know about you, but I don’t want to be the one to have to tell a team of environment artists that all of the ivy in an entire game needs to be slightly different. In this situation, the key is to work smarter, not harder. While procedural art is often controversial in the game industry, I think most developers would agree that artist-directed procedural techniques are an invaluable tool.

First and foremost, my foliage effect is composed of two separate rendering passes. First, a triplanar-mapped base texture is blended onto the object based on the desired density of the ivy. This helps to make the foliage feel much more dense, and helps to hide the seams where the leaves meet the base geometry.

Next in a second rendering pass, the geometry program transforms every input triangle into a set of quads lying on that triangle with a uniform, psuedo-random distribution. First, it is necessary to determine the number of leaf quads to generate. In order to maintain a consistent density of leaf geometry, the surface area of the triangle is calculated quickly using the “half cross-product formula”, and is then multiplied by the desired number of leaves per square meter of surface area. Then, for each of these leaves, a random sample point on the triangle is picked, and a triangle strip is emitted. It does this by sampling a noise function seeded with the world-space centroid of the triangle and the index of the leaf quad being generated. These noise values are then used to generate barycentric coordinates, which in turn are used to interpolate the position and normal of the triangle at that point, essentially returning a random world-space position and its corresponding normal vector.

Now, all that’s needed is to determine the orientation of the leaf, and output the correct triangle-strip primitive. Even this is relatively simple. By using the world-space surface normal and world “up” vector, a simple “change of vector basis” matrix is constructed. Combining this with a slightly randomized scale factor, and a small offset to orientation (to add greater variety to patches of leaves), we can transform normalized quad vertices into the exact world-space positions we want for our leaves!


// Defines a unit-size square quad with its base at the origin. doing
// this allows for very easy scaling and positioning in the next steps.
static const float3 quadVertices[4] = {
   float3(-0.5, 0.0, 0.0),
   float3( 0.5, 0.0, 0.0),
   float3(-0.5, 0.0, 1.0),
   float3( 0.5, 0.0, 1.0)


// Change of basis matrix converts from XYZ space to leaf-space
float3x3 leafBasis = float3x3(
   leafX.x, leafY.x, leafZ.x,
   leafX.y, leafY.y, leafZ.y,
   leafX.z, leafY.z, leafZ.z

// constructs a random rotation matrix from Euler angles in the range 
// (-10,10) using wPos as a seed value.
float3x3 leafJitter = randomRotationMatrix(wPos, 10);

// Combine the basis matrix by the random rotation matrix to get the
// complete leaf transformation. Note, we could use a 4x4 matrix here
// and incorporate the translation as well, but it's easier to just add
// the world position as an offset in the final step.
float3x3 leafMatrix = mul(leafBasis, leafJitter);

// lastly, we can just output four vertices in a triangle strip
// to form a simple quad, and we'll be on our merry way.
for ( int i = 0; i < 4; i ++ ) {
   FS_INPUT v;
   v.vertex = UnityWorldToClipPos( 
      float4( mul(leafMatrix, quadVertices[i] * scale), 1) + wPos 

At this point, the meat of the work is done! We’ve got a geometry shader outputting quads on our surface. The last thing needed is to texture them, and it works!


I briefly touched on artist-configurable effects in the introduction, and I’d like to quickly address that too. I opted to go with the simplest solution I could think of, and it ended up being incredibly effective.


Configuring procedural geometry using painted vertex weights.

The density and location of ivy is controlled through painted vertex-colors. This allows artists to simply paint sections of their model they would like to be covered in foliage, and the shader will use this to weight the density and distribution of the procedural geometry. This way, an environment artist could use the tools they’re familiar with to quickly sketch out what parts of a model they would like to be effected by the shader. It will take an experienced artist less than a minute to get a rough draft working in-engine, and changes to the foliage can be made just as quickly!

At the moment, only the density of the foliage is mapped this way (All other parameters are uniform material properties), but I intend to expand the variety of properties which can be expressed this way, allowing for greater control over the final look of the model.


This ended up being an extremely informative project, but there are many things still left to do! For one, the procedural foliage does not take lighting into account. I built this effect in the Unity game engine, and opted out of using the standard “Surface Shader” code-generation system, which while very useful in 99% of cases, is extremely limiting in situations such as this. I would also like to improve the resolution of leaf geometry, applying adaptive runtime tessellation to the generated primitives in order to give them a slight curve, rather than displaying flat billboards. Other things, such as color variation on leaves could go a long way to improving the effect, but for now I’m quite satisfied with how it went!

Whelp, on to the next one!


Packed Geometry Maps

Geometry Map.png

Here, I present a novel implementation of well established techniques I am calling “Packed Geometry Maps”. By swizzling color channels and exploiting current compression techniques, packed geometry maps represent normal, height, and occlusion information in a single texture asset. This texture is fully backwards-compatible with standard normal maps in the Unity game engine, and can be automatically generated from traditional input texture maps through an extension of the Unity editor, reducing texture memory requirements by 2/3, and requiring no changes to developer workflow.

Normal Maps are great, but they could be better.

Normal maps are not a new invention. The initial concept was first published sometime around 1998 as a technique for storing high-frequency surface data to be re-mapped onto low resolution geometry. Since then, normal maps in some for or another have become ubiquitous in realtime computer graphics, where the geometric resolution of models may be limited.

The basic concept is to store surface normals (the direction a surface is “facing”), in an image. When you need to know this information, such as when calculating shading on an object, you simply look up the surface normal from the texture (and depending on your implementation, re-project it into a different basis). The end result is low resolution models being shaded as though they were higher detail, capturing smaller bumps and cracks in their surface which aren’t actually represented by geometry.

Normal Map Example.png

A comparison of a simple sphere without and with normal mapping. The underlying geometry is the same for both.

Normal maps are fantastic, but as computer hardware advances other techniques have become more frequent in addition to normal mapping. Runtime tessellation for example, produces additional geometric detail to better represent rough surfaces, and produce more accurate silhouettes, at the expense of additional texture data. Ambient occlusion maps are used to represent light absorption and scattering in complex geometry, and all of these new effects require new texture data as input. The normal map should not be discounted or ignored, but there’s definitely room for improvement here!

What’s In A Normal Map?

Surface normals are typically represented by colors, using the normal RGB format. The red channel of color is used to represent the “x” component of the normal vector (usually the coordinate along the surface tangent), green for “y” (bitangent), and blue for “z” (normal). These components all sum up to create a unit-length vector, used in lighting calculations, etc. which is stored in the final normal map texture.

You might have already noticed an optimization here. If the normal vector is known to be unit-length, then why are we storing three components? In most situations, the Z component of the surface normal in tangent-space is positive, and close to one. Using these assumptions, we can remove any ambiguity in solving for the Z component, and it can be reconstructed using only the X and Y components, which contain more critical signed data.

This allows us to completely remove one color channel from our normal map texture, while still preserving all the information required to properly shade an object!

In fact, the Unity game engine already does this! Unity uses a texture compression format called DXT5nm, a specific use-case of standard DXT5 compression. This confers many advantages in terms of memory usage by sacrificing precision and image quality. The DXT5 compression format is unique in that it preserves a fixed compression ratio, and the overall quality of the output image is dependent on the content of the image itself. I won’t get into details here, but images containing shades of a single color have a higher compressed quality than images with multiple colors. Unity disposes of the “Z” component of normal maps, and swizzles the X component into the alpha channel to take advantage of the inherent strengths of the format, and reduce the number of visible artifacts after compression.

What About Those Two Extra Channels?

So, the Unity game engine simply writes zeroes into two of the color channels used by normal maps. Granted, this increases the quality of the texture due to the subtleties of DXT5 compression, but we could easily store two more low-frequency channels without a significant loss in quality. For this, I’ve opted to store ambient occlusion in red, and displacement in blue. By placing ambient occlusion (which tends to have the lowest contrast of all the input maps) in red, we can generally reduce the number of artifacts visible in the green channel of the surface normals, where they are most noticeable. Imprecision and compression artifacts in displacement and occlusion maps are also much less visible, due to the low-frequency nature of the types of data typically stored here.

Geometry Vs Traditional.png

The traditional normal, ambient occlusion, and displacement maps (left), compared to packed geometry maps (right). While geometry maps exhibit some artifacts, they’re not particularly noticeable on matte surfaces.


Worst-case example of Packed Geometry Maps applied to a mirror sphere.

This technique tends to work quite well. The majority of the texture artifacts produced are relatively subtle on mostly matte surfaces, however it still looks considerably worse on highly metallic and reflective surfaces, and therefore is mostly recommended for environment textures. On mostly non-metal surfaces such as dirt, grass, or wood, the differences between a packed geometry map and a traditional multi-texture setup are largely imperceptible, and given the constant compression ratio of DXT5, will require 1/3 the texture memory of the equivalent traditional input maps.

Won’t This Change Our Workflow?

As a matter of fact, no. Since packed geometry maps are essentially a combination of the color channels of various input textures, they can be produced automatically as part of the import pipeline of your game engine.

I’ve written a seamless extension to the Unity engine which will automatically build geometry maps for input textures as they are imported, updating them as source assets change, and saving them as unique standalone assets that can be incorporated in the final build.

By adding this extension to your project, geometry maps will be generated automatically as you import and update source texture maps. The utility is configurable and allows manual generation, as well as adding asset labels to generated geometry maps for easy searches. Certain directories can also be excluded if they contain textures your team does not want generating maps, and the file suffixes used to identify different texture types can be configured to best suit your naming scheme. It supports live reloading when source assets are changed, and is designed to be unobtrusive as possible.

There are a few minor issues left with the editor extension, so I’m not ready to release it just yet, but I’ll post it to Github in the near future!

Using these new Geometry Maps are just as simple! I’ve written a CG-Include file which defines a simple function for unpacking geometry maps, which should serve as a mostly drop-in replacement for Unity’s “UnpackNormal” function. Unlike “UnpackNormal” however, the “UnpackGeometryMap” function only requires a single texture sample for all three input maps, and returns a struct type for convenient access.

#include “GeometryMap.cginc”

// defines three values
//     geo.normal
//     geo.displacement
//     geo.occlusion
GeometryMapSample geo = UnpackGeometryMap(tex2D(_GeoMap, IN.uv));

Including Geometry Map support in your new shaders is a snap and, should you choose, geometry maps are fully backwards-compatible with Unity’s built-in normal maps, and can simply be dropped in the “Normal Map” field of any standard shader.

In Summary

By utilizing unused texture color channels, it is possible to sacrifice some image quality for considerable savings on texture memory. A single texture map can be generated to produce backwards compatible geometry maps, representing normal, displacement, and occlusion information without considerable changes to artists’ workflow. This technique is particularly applicable to realtime applications designed to run on low-end hardware where texture memory is of significant concern, and is suitable for representing most non-metal surfaces.

In the future, I’ll look into other compression techniques that don’t degrade with additional color channels. Even if the compressed size of a single map is larger, there could still be potential for savings when compared to several DXT5 textures.