Upcoming Global Illumination improvements in Ogre-Next

Note: This work is being sponsored by Open Source Robotics for the Ignition Project

Ogre-Next offers a wide amount of Global Illumination solutions.

Some better than others, but VCT (Voxel Cone Tracing) stands out for its high quality at an acceptable performance (on high end GPUs).

However the main problem right now with our VCT implementation is that it’s hard to use and needs a lot of manual tweaking:

  1. Voxelization process is relatively slow. 10k triangles can take 10ms to voxelize on a Radeon RX 6800 XT, which makes it unsuitable for realtime voxelization (only load time or offline)
  2. Large scenes / outdoors need very large resolution (i.e. 1024x32x1024) or just give up to large quality degradations
  3. It works best on setting up static geometry on a relatively small scene like a room or a house.

If your game is divided in small sections that are paged in/out (i.e. PS1 era games like Resident Evil, Final Fantasy 7/8/9, Grim Fandango) VCT would be ideal.

But in current generation of games with continous movement over large areas, VCT falls short, not unless you do some insane amount of tricks.

So we’re looking to improve this and that’s where our new technique Cascaded Image VCT (CIVCT… it wouldn’t be a graphics technique if we didn’t come up with a long acronym) comes in:

  1. Voxelizes much faster (10x to 100x), enabling real time voxelization. Right now we’re focusing on static meshes but it should be possible to support dynamic stuff as well
  2. User friendly
  3. Works out of the box
  4. Quality settings easy to understand
  5. Adapts to many conditions (indoor, outdoor, small, large scenes)

That would be pretty much the holy grail of real time GI.

Step 1: Image Voxelizer

Our current VctVoxelizer is triangle based: It feeds on triangles, and outputs a 3D voxel (Albedo + Normal + Emissive). This voxel is then fed to VctLighting to produce the final GI result:

Right now we’re using VctVoxelizer voxelizes the entire scene. This is slow.

Image Voxelizer is image-based and consists in two steps:

  1. Reuse VctVoxelizer to separately voxelize each mesh offline and save results to disk (or during load time). At 64x64x64 a mesh would need between 2MB and 3MB of VRAM per mesh (and disk space) depending on whether the object contains emissive materials. Some meshes require much lower resolution though. This is user tweakable. You’d want to dedicate more resolution to important/big meshes, and lower resolution for the unimportant ones.
    • This may sound too much, but bear in mind it is a fixed cost independent of triangle count. A mesh with a million triangles and a mesh with a 10.000 triangles will both occupy the same amount of VRAM.
    • Objects are rarely square. For example desk table is often wider than it is tall or deep. Hence it could just need 64x32x32, which is between 0.5MB and 0.75MB
  2. Each frame, we stitch together these 3D voxels of meshes via trilinear interpolation into a scene voxel. This is very fast.

This feature has been fast thanks to Vulkan, which allows us to dynamically index an arbitrary number of bound textures in a single compute dispatch.

OpenGL, Direct3D 11 & Metal* will also support this feature but may experience degraded performance as we must perform the voxelization in multiple passes. How much of a degradation depends on the API, e.g. OpenGL actually will let us dynamically index the texture but has a hard limit on how many textures we can bind per pass.

(*) I’m not sure if Metal supports dynamic texture indexing or not. Needs checking.

Therefore this is how it changed:

This step is done offline or at loading time:

This step can be done every frame or when the camera moves too much, or if an object is moved

Downside

There is a downside of this (aside from VRAM usage): We need to voxelize each mesh + material combo. Meaning that if you have a mesh and want to apply different materials, we need to consume 2-3MB per material

This is rarely a problem though because most meshes only use one set of materials. And for those that do, you may be able to get away with baking a material set that is either white or similar; the end results after calculating GI may not vary that much to worth the extra VRAM cost.

Non-researched solutions:

  • For simple colour swaps (e.g. RTS games, FPS with multiplayer teams), this should be workaroundeable by adding a single multiplier value, rather than voxelizing the mesh per material
  • It should be possible to apply some sort of BC1-like compression, given that the mesh opaqueness and shape is the same. The only thing that changes is colour; thus a delta colour compression scheme could work well

Trivia

At first I panicked a little while developing the Image Voxelizer because the initial quality was far inferior than that of the original voxelizer.

The problem was that the original VCT is a ‘perfect’ voxelization. i.e. if a triangle touches a single voxel, then that voxel adopts the colour of the triangle. Its neighbour voxels will remain empty. Simple.
That results in a ‘thin’ voxel result.

However in IVCT, voxels are interpolated into a scene voxel that will not match in resolution and may be arbitrarily offsetted by subpixels. It’s not aligned either.

The result is that certain voxels have 0.1, 0.2 … 0.9 of the influence of the mesh. This generates ‘fatter’ voxels.

In 2D we would say that the image has a halo around the contours

Once I understood what was going on, I tweaked the math to thin out these voxels by looking at the alpha value of the interpolated mesh and applying an exponential curve to get rid of this halo.

And now it looks very close to the reference VCT implementation!

Step 2: Row Translation

We want to use cascades (a similar concept from shadow mapping, i.e. Cascaded Shadow Maps. In Ogre we call it Parallel Split Shadow Maps but it’s the same thing) concentric around the camera.

That means when the camera moves, once the camera has moved too much, we must move the cascades and re-voxelize.

But we don’t need to voxelize the entire thing from scratch. We can translate everything by 1 voxel, and then revoxelize the new row:

As the camera wants to move, once it moved far enough, we must translate the voxel cascade

When we do that, there will be a region that is no longer covered (it will be covered by a higher, lower quality cascade) marked in grey, and a row of missing information we must revoxelize, marked in red.

Given that we only need to partially update the voxels after camera movement, it makes supporting cascades very fast

Right now this step is handled by VctImageVoxelizer::partialBuild

Step 3: Cascades

This step is currently a work in progress. The implementation is planned to have N cascades (N user defined). During cone tracing, after we reach the end of a cascade we move on to the next cascade, which covers more ground but has coarser resolution, hence lower quality.

Wait isn’t this what UE5’s Lumen does?

AFAIK Lumen is also a Voxel Cone Tracer. Therefore it’s normal there will be similarities. I don’t know if they use cascades though.

As far as I’ve read, Lumen uses an entirely different approach to voxelizing which involves rasterizing from all 6 sides, which makes it very user hostile as meshes must be broken down to individual components (e.g. instead of submitting a house, each wall, column, floor and ceiling must be its own mesh).

With Ogre-Next you just provide the mesh and it will just work (although with manual tunning you could achieve greater memory savings if e.g. the columns are split and voxelized separately).

Wait isn’t this what Godot does?

Well, I was involved in SDFGI advising Juan on the subject, thus of course there are a lot of similarities.

The main difference is that Godot generates a cascade of SDFs (signed distance fields), while Ogre-Next is generating a cascade of voxels.

This allows Godot to render on slower GPUs (and is specially better at specular reflections), but at the expense of accuracy (there’s a significant visual difference when toggling between Godot’s own raw VCT implementation and its SDFGI; but they both look pretty) but I believe these quality issues could be improved in the future.

Having an SDF of the scene also offers interesting potential features such as ‘contact shadows’ in the distance.

Ogre-Next in the future may generate an SDF as well as it offers many potential features (e.g. contact shadows) or speed improvements. Please understand that VCT is an actively researched topic and we are all trying and exploring different methods to see what works best and under what conditions.

The underlying techniques aren’t new, but what made it possible are the new APIs and the raw power provided by current generation of GPUs that can keep up with them (although the current GPU shortage might delay the widespread adoption of these techniques).

Since this technique will be used in Ignition Gazebo for simulations, I had to err on the side of accuracy.

When is it coming?

CIVCT isn’t done yet but hopefully it should be ready 1-2 more weekends (I can only work on this during the weekends). Maybe 3? (I hope not!). I want to release Ogre-Next 2.3 RC0 in the meantime, and when CIVCT is ready a proper Ogre-Next 2.3 release.

The reason it’s taking so little time is because we’re improving on our existing technology and reusing lots of code. We’re just changing a few details to make it faster and more use friendly now that Vulkan gives us that freedom (but again, we plan on supporting this feature on all our API backends).

These improvements are currently living in vct-image branch but has no sample yet showcasing it as it is WIP.

Btw! Remember there is an active poll to decide on Ogre-Next 2.3 name. Don’t forget to vote!

Ogre 2.2.5 Cerberus Released and EGL Headless support!

This is a special release! Most Ogre 2.1.x and 2.2.x releases, it only contains maintenance fixes and no new features.

Thus efforts to port from 2.2.4 to 2.2.5 should be minimum. And this still holds true.

But there is a new feature!

This feature was sponsored by Open Source Robotics Corporation and was written to be used by the Ignition Project

EGL Headless

OpenGL traditionally requires a window. Without a window, OpenGL cannot be used. This implies either X11 or Wayland is installed and running; which can be a problem when running on cloud servers, VMs, embedded devices, and similar environments.

Direct3D11 doesn’t have this flaw, but it does not run on Linux.

Vulkan also doesn’t have this flaw, but its support is new (coming in Ogre 2.3) and is not yet robust and tested enough. Additionally SW implementations have yet to catch up.

Ogre can use the NULL RenderSystem to run as a server without a window, however this doesn’t actually render anything. It’s only useful to pretend there is a screen so that apps (mostly games) can reuse and turn client code into server code. It’s also useful for mesh manipulation and conversion tools which need to read Ogre meshes but don’t actually render anything.

Fortunately, Khronos introduced a workaround with EGL + PBuffers (not to be confused with 2000-era PBuffers which competed against FBOs) where an offscreen dummy ‘window’ could be created to satisfy OpenGL’s shenanigans.

Because PBuffer support in some EGL drivers are not well tested (e.g. sRGB support was only added in EGL 1.5, which Mesa does not support) Ogre creates a 1×1 PBuffer alongside the Context and uses an FBO internally for the ‘Window’ class. By using a dummy 1×1 PBuffer tied with the GL context, OpenGL context creation becomes conceptually free of window interfaces, like in D3D11 and Vulkan.

Switchable interfaces: GLX and EGL

When Ogre is built with both OGRE_GLSUPPORT_USE_GLX and OGRE_GLSUPPORT_USE_EGL_HEADLESS, toggling between GLX and EGL can be done at runtime.

This is how it looks:

Originally the GLX interface will be selected:

But after switching it to EGL Headless, only a couple options appear (since stuff like Resolution, VSync, Full Screen no longer make sense)

And like in D3D11/Vulkan, it is possible to select the GPU. /dev/dri/card0 is a dedicated AMD Radeon HD 7770 GPU, /dev/dri/card1 is a dedicated NVIDIA GeForce 1060. Yes, they can coexist:

NVIDIA seems to expose 2 “devices” belonging to the same card. ‘EGL_NV_device_cuda … #0’ is a headless device. Trying to use ‘EGL_EXT_device_drm #1’ will complain that it can’t run in headless mode. It seems it is meant for use with GLX.

‘EGL_EXT_device_drm #2’ is the AMD card.

EGL_MESA_device_software is SW emulation

We chose not to include the marketing names in device selection because Linux drivers (propietary and open source) have the tendency of changing the exposed OpenGL marketing labels quite often in subtle ways. This could break config settings quite often (i.e. the saved chosen device can no longer be found after a driver upgrade), increasing maintenance burden when this feature is meant for automated testing and similar.

Complete X11 independence

Users who need to be completely free of X11 dependencies can build with OGRE_GLSUPPORT_USE_EGL_HEADLESS + OGRE_CONFIG_UNIX_NO_X11.

This will force-disable OGRE_GLSUPPORT_USE_GLX as it is incompatible. GLX requires X11.

Headless SW Rasterization

It is possible to select the Mesa SW rasterization device. So even if there is no HW support, you can still use SW.

Please note Mesa SW at the time of writing supports up to OpenGL 3.3, which is the bare minimum to run Ogre. Some functionality may not be available.

Update: It has been called to my attention that llvmpipe (aka SW emulation) supports OpenGL 4.5 since Mesa 20.3.0

More info

This new feature seems to be very stable and has been tested on NVIDIA, AMD (Mesa drivers) and Intel.
Nonetheless it is disabled by default (i.e. OGRE_GLSUPPORT_USE_EGL_HEADLESS is turned off) which means it should not affect users who are not caring about headless support.

For more details, please see the README of the EglHeadless tutorial.

Running EglHeadless sample should result in a CLI interface:

OpenGL ES 3.x may be around the corner?

With EGL integration, it should be possible to create an EGL window and ask for an ES 3.x context instead of an OpenGL one. There is a lot of similarities between ES 3 and OpenGL 3.3, and we already have workarounds for it as they’re the same ones we use for macOS.

While I don’t have very high hopes for Android, WebGL2 may be another story.

If such feature is added into the roadmap, it would probably be for 2.3 though.

RenderDoc integration

Functions RenderSystem::startGpuDebuggerFrameCapture and RenderSystem::endGpuDebuggerFrameCapture were added to programmatically capture a RenderDoc frame. This was necessary for RenderDoc to work with headless rendering, but it works with all APIs in most platforms.

Users can call RenderSystem::getRenderDocApi if they wish to perform more advanced manipulation:

if( rs->loadRenderDocApi() )
    RENDERDOC_API_1_4_1 *apiHandle = rs->getRenderDocApi();

About the 2.2.5 release

For a full list of changes see the Github release

Source and SDK is in the download page.

Discussion in forum thread.

Thanks again to Open Source Robotics Corporation for sponsoring this feature for their Ignition Project