GPU cubemap filtering

Prefiltered cubemaps are widely used in games. The idea is to match cubemap blurriness at each mip to appearance of the BRDF you use.

It’s all started with Modified Cubemapgen
Then there’s also cmft which is quite faster
Finally there’s Lys

All these tools are separate utilities which also have a command-line support (except Lys).

Sometimes you want to filter the cubemap in engine, without exporting stuff/running separate utils and you want it to be fast. Cubemapgen may take minutes to filter, cmft is much better here, but sometimes you still need more control and performance. Sometimes you also have to filter it in something like WebGL, where running separate utils isn’t very acceptable (yes, there is Emscripten, but the resulting code is often too fat).

At first I thought that this is a very slow and complicated process (why otherwise cubemapgen does it sooo slow?) and was quite afraid of implementing it myself, but turns out, it’s not THAT hard.

Here’s the result:

You may notice some blurred seam appearance on the rightmost sphere in all versions – this is the inevitable consequence of using the seamless cubemap filtering trick on a DX9-level GAPI. DX10+ doesn’t have this problem and looks smoother. For WebGL implementation I had to support DX9, because latest Firefox still uses it to emulate WebGL (Chrome and IE11 render through DX11).

As you can see, Blinn-Phong importance sampled version looks extremely close to cubemapgen offline output. This version takes around 1.5 sec on GTX560 to compute from 128×128 source cubemap. “Simple” version takes even less than 1 sec.

So, how it is done?
We don’t just blur the cubemap. We imagine a sphere, reflecting it and associate each mip with different roughness value. Rougher surfaces have more noisy microfacets reflecting light in many directions at each visible point. The rougher the surface, the wider will be the cone of incoming light:
So what we should do is to average this cone of lighting.
The simple averaging is not enough though – there is also a complicated weight factor depending on deviation from the original normal based on many factors. This factor actually makes your BRDF to have the falloff it has, instead of just being an averaged circle.

So what, do you have to cast a million rays and then multiply each by a weight? No.
Importance sampling is a technique, the idea of which is to just generate the rays with density depending on your BRDF’s shape. Usually you’ll end up having more rays with direction similar to the surface normal and less deviating rays. Simply averaging the lighting from each ray will naturally give you more intensity near the center, because there were more rays there.

Here’s a difference between “Simple” version (top), which uses simple cosine-weighted distribution of rays and Blinn-Phong version (bottom):

As you can see, getting the correct ray distribution can be important for getting nice highlight falloff instead of just circular spots.

The Blinn-Phong is, of course, not ideal and quite old. GGX is considered more realistic, but I haven’t used it yet.

Different BRDF have different requirements for the ray count. That’s the point of having a “Simple” version – despite of having less correct highlights, it requires MUCH less rays for an artifact-free result (because it’s more uniform).

So the basic algorithm is:
– generate uniformly random directions on the hemisphere;
– focus the directions depending on your BRDF intensity-direction relation at given roughness;
– render lower mip cubemap faces, while reading higher mip data;
– for each direction
—transform the direction to be relative to the original lookup vector;
—light += texCUBE(higher mip, transformed direction)
–light /= numDirs;

You don’t have to pre-generate the directions though – all this can be done in shader sufficiently fast.
Here is a good read on generating points on the hemisphere. It even has an interactive demo.
The Hammersley method relies on bitwise operations though, which are not always accessible (not in WebGL/DX9). In such old-school GAPIs you have to either precompute the directions, or use some other way to generate random numbers. There are many other ways, one you can see in my source code, others on ShaderToy or somewhere else. Such randoms will likely be less uniform.

Ideally, when writing each mip, you should probably read the highest available mip with different cones of rays, but it’s not very fast, and you’ll get some significant aliasing trying to sample a high resolution cubemap with a very wide cone of limited rays.
Instead, it is actually quite sufficient to read just a 2x larger mip. This larger mip must resemble the highest mip as much as possible, so something simple like automatic mipmap generation (bilinear downsample) will do. Note, that you must NOT cone-sample the mip that was already cone-sampled, because you’ll get noticeable over-blurring.

My version of filtering is now in the PlayCanvas engine repository, and it’s open source.
Sadly, there’s a bug in ANGLE, which prevents us from using it on Win Chrome/FF, ironically only IE11 works correctly.

The source in here:

The latest version with additional bells and whistles will be always here:

Notes on shadow bias

These are notes for myself about shadow mapping bias.
Good summary about all aspects of shadow mapping:

My results:

I’m not sure what’s wrong about Receiver Plane depth bias. What is interesting, it does work OK when there is no interpolation between samples.
In this presentation, there’s a comparison, but it also uses samples without interpolation: (page 39).
Here they also get strange artifact with it similar to one I have on sphere:
MJP also says that “When it works, it’s fantastic. However it will still run into degenerate cases where it can produce unpredictable results”.
So, maybe I implemented it wrong, or maybe I was unlucky enough to quickly get degenerate cases, but I’m not really willing to try this technique anymore.

Normal offset:
Also this may better explain why it works:

There are 2 ways how to implement Normal Offset bias. One way is to inset geometry by normal when rendering the shadow map. The insetting amount is also scaled by slope aka dot(N,L) and also can be scaled by distance factor with FOV included for using with perspective projection.
Second way is to render shadow map normally, but add (instead of subtract) same scaled vertex normal to fragment position just before multiplying it by shadow map matrix and comparing.
The second method has less impact on shadow silhouette distortion and gives better results. It is, however, not easy to do with deferred rendering, because you need vertex normal, not normal mapped one!
Unity 5 seems to use 1st version exactly because it can’t hold vertex normal in G-Buffer.

Funnily enough, Infamous Second Son is OK with storing it there:

And they use it exactly for normal offset (and other stuff too):

You can also try to calculate face normal from depth, BUT you’ll get unpredictable results on edges.
Even Intel guy couldn’t solve that:

Hole appears after insetting geometry:


2nd variant doesn’t suffer from this (you can still see a tiny hole there though… but it also exists with just constant bias, so it’s not a normal bias problem).
Real-time demo with Normal Offset 2nd version:
No acne, no peter-panning, yay!
Use RMB + WASD to fly around. Feel free to look into source.

You can tweak both normal/constant bias in browser console using

Designing an Ubershader system

OK, so you probably know what ubershaders are? Unfortunately there is no wiki on this term, but mostly by it we mean very fat shaders containing all possible features with compile-time branching that allows them to be then specialized into any kind of small shader with a limited amount of tasks. But it can be implemented very differently, so here I’ll share my experience on this.


So, you can use #ifdef, #define and #include in your shaders? Or you’re going to implement it yourself? Anyway, it’s the first idea anyone has.

Why it sucks:
  • Too many #ifdefs make your code hard to read. You have to scroll the whole ubershader to see some scattered compile-time logic.
  • How do you say “compile this shader with 1 spot light and that shader with 2 directional lights”? Or 2 decals instead of 6? One PCF shadow and one hard? You can’t specify it with #ifdefs elegantly, only by copy-pasting code making it even less readable.

Terrible real-life examples: 1, 2

Code generation from strings

Yet another approach I came across and have seen in some projects. Basically you use your language of choice and use branching and loops to generate new shader string.

Why it sucks:
  • Mixing shader language with other languages looks like total mess
  • Quotes, string additions, spaces inside strings and \n’s are EVERYWHERE flooding your vision
  • Still have to scroll a lot to understand the logic

Terrible real-life examples: 1, 2

Code generation from modules

So you take your string-based code generation and try to decouple all shader code from engine code as much as possible. And you definitely don’t want to have hundreds of files with 1-2 lines each, so you start to think how to accomplish it.
So you make some small code chunks like this one, some of them are interchangeable, some contain keywords to replace before adding.

Why naive approach sucks:
  • All chunks share the same scope, can lead to conflicts
  • You aren’t sure what data is available for each chunk
  • Takes time to understand what generated shader actually does

Code generation from modules 2.0

So you need some structure. The approach I found works best is:

struct internalData {
some data

void shaderChunk1(inout internalData data) {
float localVar;
read/write data

float4 main() {
internalData data;
return colorCombinerShaderChunk(data);

So you just declare an r/w struct for all intermediate and non local data, like diffuse/specular light accumulation, global UV offset or surface normal used for most effects.
Each shader chunk is then a processing function working with that struct and a call to it, put between other calls. Most compilers will optimize out unused struct members, so basically you should end up with some pretty fast code, and it’s easy to change the parts of your shader. Shader body also looks quite descriptive and doesn’t require you to scroll a lot.
The working example of such system is my contribution to PlayCanvas engine: 1, 2

Examples of generated code: vertex, pixel, pixel 2, pixel 3

So, I’m not saying this is the best approach. But for me, it’s the easiest one to use/debug/maintain so far.