Advances in LPV, Volumetric Lighting and a new Engine Architecture

Welcome once again, and thanks for clickin on my journal! 🙂

When we think of light propagation volumes we think of awesomeness with a ton of light bleeding, right? Well I do, so I’ve been trying to limit the amount of light bleeding and I find the current result acceptable without being too hackyish.

The first change is the injection. Usually I would inject the lighting in the same position as the voxels into a sh lighting map, but, once I do this I don’t have much information on which “direction” the lighting came from if I was to propagate the lighting. In that case I would have three choices, light bleeding, expensive occlusion calculation or not propagating (But the last one wouldn’t be any fun…). So, what if instead of injecting the lighting at the voxel position I inject it with a small offset, this offset would then be proportional to the approximated normal. In this way the actual lighting information would be injected into the empty space.

The 2^nd change I made was simply discarding propagation in any occluded voxels/cells as the new method doesn’t require it. Now the issue with this is if the cell size ( How much a voxel/cell occupies in whatever unit your system uses ) is way too big compared to the world the propagation will visually fail and look horrible, so to look good a bit of performance is sucked.

The last change is when sampling the final indirect gi I apply a small offset as at all the lighting information is in the “empty” cells, now one might say that this is a crude approximation but I don’t find it that horrible.

So, there you have it, that’s my current recipe to a LPV system without bleeding, there are still lots of things to fix but it’s a start.

In my last entry I talked about a cascaded LPV system, however this has slightly changed. You can still configure multiple cascades; however the way it works is slightly different. In each cascade the system will create two grids, a high frequency grid and a low frequency grid (The dimensions of the grid is still intact). The low frequency grid represents the low frequency lighting information, and the high frequency grid will represent the slightly higher frequency lighting information. The two grids are treated as separate grids with different cell sizes but when rendered the energy proportion is taken into account.

So I’m fairly happy how my LPV system has progressed and I find the results acceptable, now obviously there’s the issue with the “blocky” look ( If you want an acceptable performance 🙂 ), which I’ll try and mess around with and share my results later on.

Now, let’s steer slightly away from that and think about volumetric fog! Yes! That’s right!

Volumetric Lighting!

So to make the volumetric lighting feel more “part” of the scene I integrated the indirect gi system. Currently I have a very basic volumetric lighting setup, raymarch from the camera to the world space position at the pixel and slowly accumulate the lighting (The method I used to calculate the lighting is based on “Lords of the Fallen’s” [Game] Volumetric Lighting). So each raymarch I also sample the indirect gi from the propagated lighting map and multiply that in. And I’m really liking the results!

(I know the roughness / specular looks wrong, I still need to integrate the rougness / specular maps from the sponza scene) (And I seriously improved the quality of the gifs…)

Now! The only issue with this is… performance! All of that added together is asking your hardware to commit suicide, at least, mine did. Since I’m an addict to the game Dota 2, I was having a casual game with some friends and decided to program in the background, now for some reason I was writing and reading from an unbound UAV in my compute shader ( I didn’t realize this ). The result was the gpu completely freezing ( I could still talk and hear my friends, whilst freaking out ), I waited for the tdr duration however the tdr did not occur. So in the end I had to force shut down and restart quickly in order to participate in the game ( We won though! ). I was actually scared to start it again even though I bound the uav…

Looking aside from that I’ve also implemented some basic debugging tools for the lpv system, such as getting the lighting information from each cell position ( It’s extremely simple to implement, but really helps a lot ):

Previously my engine has a pretty horrible architecture, because I’m horrible at architecture, I’m a horrible person. So I decided to attempt at improving the architecture of the engine. I decided to split the engine up in:

• Helpers : Just general things, common math stuff / etc…
• Native Modules : Shaders, Containers, etc
• User Modules : An example would be a custom voxel filler or whatever, depends on the type
• Chains : Responsible for higher level actions, such as Shadow Mapping, Voxel GI, etc…
• Device : Basically combining chains and working with them

Now I’m not saying that this is ideal or even good, but I find it nice and functional. Now the user modules are a bit special, the user modules are custom modules that the programmer can create. However each module has to derive from a module type. An example is the gi system, the gi system has a special module type that allows the modification of the lighting maps before the propagation. The programmer would then inherit from this type and override the pure virtual functions, and then push this module to a queue. I made a small module that would “approximate” the indirect radiance from the “sky” (Assuming that there is one) just to test around. The native c++ code is farily straight forward. Although this specific module type has a bunch of predefinitions and preprocessors in a shader file to ease the process, the shader code for this testing module:

#include "module_gridshfill.hlsl"

// Our basic module definition
MODULE((8, 1, 8), (uint3 CellPos : MODULE_CELLID) {
    // Testing data, This is just magic and stuff, not correct at all
    float fFactor = 0.01f;
    float3 f3Color = float3(0.658, 0.892, 1);
    float3x4 f3x4AmbientSH =
    {
        fFactor.xxxx * f3Color.x,
        fFactor.xxxx * f3Color.y,
        fFactor.xxxx * f3Color.z
    };

        // Raymarch Down
    [loop] for (CellPos.y = g_fVoxelGridSize-1; CellPos.y >= 0; CellPos.y--)
    {
        // Get the voxel
        VoxelData voxel = FETCH(CellPos - uint3(0, 1, 0));

        // If this voxel is occupied break the march
        // TODO: Semi occluded voxels (1<w>0)
        if (voxel.fOcclusion > 0)
        {
            break;
        }


        // Write the new value on top of the current value
        WRITE_ADDITION(CellPos, f3x4AmbientSH);
    }
});

Now some of the stuff will change for sure although it works fine for now. The result of the above is an indirect radiation from the “sky”. And it looks alright! So I’m pretty happy with the module system.

In the complete other hand I suddenly have this weird crave to work on my scripting language again… (I know I know, just use an existing one… But where would the fun be in that!? 🙂 ) And I soon need to reimplement some sort of physics engine into this version of my engine. So, there’s still lots of fun!

Looking away from some more or less small changes and additions, that’s more or less it folks! It’s been a heavy week though, lots of things happening. Fx my dog found out that a full day barbecue party is extremely tiring, he didn’t want to walk or anything, slept like a stone… (He loves walks).

See you next time!

Cascaded Light Propagation Volumes, VS RC 2015, Retarded Calculators + More stuff

Well, let’s begin shall we! 🙂 ( This article isn’t very focused, it’s just small notes and such )

======

For a while I’ve been thinking about working on cascaded light propagation volumes, so I finally did. For now I just have a 64 (detailed) + a 32 ( less detailed ) grid that are filled using the voxel caches. Although I have not worked on the energy ratio yet ( My solution is hacky ), I like the result.

(Images scaled to fit, originally rendered at resolution 1920×1080. The whitish color is because I’ve got some simple volumetric lighting going on, although it doesn’t respond to the LPV yet) (PS. Still lots of work, so there are issues + light bleeding ) ( And there’s no textures on the trees, for… reasons and stuff )

I’ve also worked on my BRDF shading model which is based on Disneys solution, and integrated my BRDF shading model into the LPV system ( Although it’s a simplified version, as we don’t need all the detail and some computations are meaningless in this context ). And I really think it made the indirect colors feel more part of the scene.

A poor quality gif showing how the light propagates through the scene:

======

On the complete other side, as I’m rewriting the engine I felt like upgrading to the RC Version of VS 2015 ( And dear god I recommend it to anyone ). And so I needed to recompile lots of libraries, such as SFML ( + most dependecies ), AntTweakBar, +small stuff. Now the AntTweakBar case was special, as it really only supports SFML 1.6. It contains a minified version of the SFML 1.6 events that it then uses, although when the memory layout changes in SFML 2.3 it all fucks up (Sorry). So I had to change some of the minified internal version of SFML to make it work, for anyone here is the modified part of the minified sfml (It’s hackyish, mostly c&p from the sfml sources, so there’s most likely errors and such, but for now it does the job ):

EDIT: See the code here -> http://www.gamedev.net/blog/1882/entry-2261240-cascaded-light-propagation-volumes-vs-rc-2015-retarded-calculators-more-stuff/

On top of that the performance of my engine in VS 2015 strangely increased by a few milliseconds which really surprised me. I’m not completely sure what it is. And in VS 2013 I had a strangely huge overhead when starting my application inside VS which made the file io incredibly slow, in VS 2015 this issue is gone and this huge waiting time is gone ( 20 seconds to a minute… ) :).

I finally got to redesign my gbuffer, and while there’s lots of work to be done, it all fits nicely, general structure:

2Channel: x = Depth, y = Packed(metallicness, anisotropicness),
4Channel: xy = Normal, z = Packed(subsurface, thickness), w = Packed(specular, roughness)
4Channel: xyz = Diffuse, z = Packed(clear_coat, emmision)

The tangent is then reconstructed later, and it’s pretty cheap and works fine for my needs. Now all the user has to do is call GBuffer_Retrieve(…) from their shaders and then all the data is decompressed which they then can use, the final data container looks somewhat like the following:

struct GBufferData
{
	float3 Diffuse;
	float3 PositionVS;
	float3 TangentVS;
	float3 NormalVS;
	float3 Position;
	float3 Normal;
	float3 Tangent;
	float SpecPower;
	float Roughness;
	float Metallic;
	float Emmision;
	float ClearCoat;
	float Anisotropic;
	float SubSurface;
	float Thickness;
};

Now, you might say “But what if I don’t want to use it all, huge overhead”, which is true, but, compilers! The cute little compiler will optimize out any computations that aren’t needed, so if you don’t need a certain element decompressed, it wont be (Yay)! So all of that fits together nicely.

But at the same time I think I’ve got an issue with the performance concerning filling the gbuffer stage, as it’s huge compared to everything else. Perhaps it’s the compression of the gbuffer, not sure yet.

But, it’s acceptable for now, although I think I can squeeze some cute little milliseconds out of it :).

On a side note I’ve also been trying to work on some basic voxel cone tracing but it’s far from done. And I seriously underestimated the performance issues, but it’s pretty fun.

======

Now due to family related issues I had to take my brother to our beach house ( Nothing fancy ), and there I allocated some time to work on my retarded calculator! It’s a small application based on a very basic neural network, I didn’t have time to work on my bias nodes or even my activation function, for now the output of the neuron is simply weight[i] * data, although it actually produces acceptable results. The network is composed of 4 layers:

• 10 Neurons
• 7 Neurons
• 5 Neurons
• 1 Neuron

Again, this was just for fun, I didn’t even adapt the learning rate during the back propagation, it was just to fill out a bit of time. The output from the application:

Starting trianing of neural network
  Train iteration complete, error 0.327538
  Train iteration complete, error 0.294999
  Train iteration complete, error 0.266
  Train iteration complete, error 0.240112
  Train iteration complete, error 0.216965
  Train iteration complete, error 0.196237
  Train iteration complete, error 0.177651
  Train iteration complete, error 0.160962
  Train iteration complete, error 0.145959
  Train iteration complete, error 0.132454
  Train iteration complete, error 0.120285
  ......... a few milliseconds later
  Training completed, error falls within treshold of 1e-06!

===============================

  Final testing stage
  Feeding forward the neural network
  Final averaged testing error: 0.0178298

===============================

Please enter a command...
>> f var(a0)
  Input:
    #0 -> 2
    #1 -> 4
    #2 -> 3
    #3 -> 1
    #4 -> 4
    #5 -> 5
    #6 -> 2
    #7 -> 3
    #8 -> 4
    #9 -> 1
  Feeding forward the neural network
  Layer Dump:
    #0 = 29.346

>> e var(a0) algo({sum(I)})
  Evaluating error: (a0)
    Error: 0.345961

======

So, overall, I’m pretty happy with it all. But I haven’t been able to allocate enough time ( You know, life and stuff, school or whatever everybody suddenly expects of you ). But if anybody is reading this, can you comment on the colors of the images, meaning do you find it natural or cartoony, I find them a bit cartoony. Well, thanks for even reaching the bottom!