CryEngine - Neon Noir: Hardware agnostic ray tracing benchmark

Elfotografoalocado · Nov 14, 2019

VFX_Veteran said:
It's using the SVOGI technique. It's only used on surfaces with mirror reflections. This article is really leading people astray. It's not the same thing as RTX -- at all.

1) The reflections are only mirror
2) They fade away as the camera moves away from the reflective surface
3) No blurred reflections using importance sampling
4) No ray-traced diffuse bounced light or environment sky lighting
5) No soft shadows
6) No emissive lighting (i.e. area lights)
7) No ambient occlusion

etc.. etc.. etc..

Don't get excited over this..

There is no single game out there that has all of those techniques simultaneously. Some do only reflections (BFV, Control), others GI (Metro Exodus) and others shadows. (Shadow of the Tomb Raider, Modern Warfare)

So I don't see how it's not comparable, quite honestly. It's doing exactly what BFV and Control do.

Nooblet · Nov 14, 2019

Actually Control does reflections (on both transparent and opaque surfaces), GI (and as a result AO), as well shadows.

chris 1515 · Nov 14, 2019

VFX_Veteran said:
chris 1515 - why would I do a partial solution coupled with a full accurate solution using hardware? What's the end game here, reflections or a fully capable ray-tracer from the moment a ray gets cast through a pixel and returns the final pixel color after traversing the entire scene over ALL objects in the scene? That technique sounds moot to me especially for PC gamers who can get the real cards.

Because it is always some compromise. It can help to make compromise with RTX or future raytracing AMD or Intel GPU. Sometimes in Battlefield 5 RTX, some part of the image use no reflection at all for keeping correct perfomance in other game they talk about using SSR as a fallback. Here it will be able to give mesh tracing where it counts and use SVOGI where it is ok far from the player or the screen.

This is how realtime works. This is a compromise and this is how raytracing at 4k 30 fps will work on next-generation console for example. Raytrace the part where you don't lose tons of performance and do a compromise elsewhere. This is not offline rendering. And this how with raytracing hardware you can have shadows and reflections at the same times at 4k 30 fps.

VFX_Veteran · Nov 14, 2019

Elfotografoalocado said:
There is no single game out there that has all of those techniques simultaneously. Some do only reflections (BFV, Control), others GI (Metro Exodus) and others shadows. (Shadow of the Tomb Raider, Modern Warfare)

So I don't see how it's not comparable, quite honestly. It's doing exactly what BFV and Control do.

Well, let's go through the list I made:

1) The reflections are only mirror -- none of the RTX games that implement reflections have this problem.
2) They fade away as the camera moves away from the reflective surface -- none of the RTX games that implement reflections have this problem.
3) No blurred reflections using importance sampling -- Control and BV5
4) No ray-traced diffuse bounced light or environment sky lighting -- Control and Metro
5) No soft shadows -- Control and Shadow of the Tomb Raider
6) No emissive lighting (i.e. area lights) - Control and Metro
7) No ambient occlusion -- Control and Metro

Control/Metro with DLC does way more than others without slowing the game's performance dramatically. If we target 4k/30FPS (which could be done in graphics settings), we could probably throw in reflections.

trugc · Nov 14, 2019

VFX_Veteran said:
Have you read the papers on techniques used for the CryEngine or even tried to implement it? Have you ever dealt with path tracers before and know how they work?

Yes I do.
What's the point you are trying to make here?

Nooblet said:
Actually Control does reflections (on both transparent and opaque surfaces), GI (and as a result AO), as well shadows.

It seems like Control is not using stochastic sampling for reflection. It doesn't have anisotropic stretching effects.

VFX_Veteran · Nov 14, 2019

chris 1515 said:
Because it is always some compromise. It can help to make compromise with RTX or future raytracing AMD or Intel GPU. Sometimes in Battlefield 5 RTX, some part of the image use no reflection at all for keeping correct perfomance in other game they talk about using SSR as a fallback. Here it will be able to give mesh tracing where it counts and use SVOGI where it is ok far from the player or the screen.

This is how realtime works. This is a compromise and this is how raytracing at 4k 30 fps will work on next-generation console for example. Raytrace the part where you don't lose tons of performance and do a compromise elsewhere. This is not offline rendering. And this how with raytracing hardware you can have shadows and reflections at the same times at 4k 30 fps.

So now you are defending a half-solution to the full solution? Ok.

chris 1515 · Nov 14, 2019

VFX_Veteran said:
Well, let's go through the list I made:

1) The reflections are only mirror -- none of the RTX games that implement reflections have this problem.
2) They fade away as the camera moves away from the reflective surface -- none of the RTX games that implement reflections have this problem.
3) No blurred reflections using importance sampling -- Control and BV5
4) No ray-traced diffuse bounced light or environment sky lighting -- Control and Metro
5) No soft shadows -- Control and Shadow of the Tomb Raider
6) No emissive lighting (i.e. area lights) - Control and Metro
7) No ambient occlusion -- Control and Metro

Control/Metro with DLC does way more than others without slowing the game's performance dramatically. If we target 4k/30FPS (which could be done in graphics settings), we could probably throw in reflections.

Wait next year when CryEngine will support RTX and other raytracing hardware to compare. Currently they have a solution working without raytracing hardware. Raytracing hardware is a must have, the compromise are too big to not use it. But they will have to make compromise too on consoles and PC too because there is limited power.

VFX_Veteran said:
So now you are defending a half-solution to the full solution? Ok.

I just think and I know realitme rendering is a compromise and here they are doing the right compromise to be able to run on hardware without raytracing hardware. Next year they will probably choose other compromise when the engine will be raytracing hardware ready.

This is the reality of realtime rendering.

VFX_Veteran · Nov 14, 2019

trugc said:
Yes I do.
What's the point you are trying to make here?

Can you explain then how this solution is a feasible replacement to the real solution when it has so many limitations?

Current Limitations Large scale AO and indirect shadows may be cast properly only by static geometry. GI is not working on some forward rendering components like particles or water. Some artifacts like ghosting, aliasing, light leaking and noise may be noticeable in some cases. Procedural vegetation and merged vegetation do not cast occlusion or secondary shadows. If the camera is teleported to a completely new location it may take up to a few seconds until occlusion is working properly. Only objects and materials with enabled shadow map casting may produce proper bounced light. For dynamic objects indirect light bounce is working only in the areas near voxelized static geometry. Bounce light may have a noticeable delay of 1-2 frames. r_Supersampling = 2 makes GI look strange but setting lower LowSpecMode (2X lower) pretty much restores the look and speed. Temporal AA (r_AntialiasingMode 2/3) works just fine.

Lastly, the paper came out years ago. Why has this implementation not been adopted globally (i.e. like PBR) if it's such a feasible technique?

KKRT · Nov 14, 2019

VFX_Veteran said:
Well, let's go through the list I made:

1) The reflections are only mirror -- none of the RTX games that implement reflections have this problem.
2) They fade away as the camera moves away from the reflective surface -- none of the RTX games that implement reflections have this problem.
3) No blurred reflections using importance sampling -- Control and BV5
4) No ray-traced diffuse bounced light or environment sky lighting -- Control and Metro
5) No soft shadows -- Control and Shadow of the Tomb Raider
6) No emissive lighting (i.e. area lights) - Control and Metro
7) No ambient occlusion -- Control and Metro

Control/Metro with DLC does way more than others without slowing the game's performance dramatically. If we target 4k/30FPS (which could be done in graphics settings), we could probably throw in reflections.

This demo is not a Path Tracer, its ray tracing reflection tech that combines RT with SVOGI.
Your first point is just a setting, rest just do not apply here.
Its great semi step to using cheaper ray-tracing tech in next few years, the less you spend on reflection, then you have some resources for shadows, ambient occlusion or some form of GI.
We do not have and wont have for some time enough compute power yet for using RT as intended, but it shouldnt stop with using RT tech or semi RT Tech with rasterization.
Smooth tech transition its easier and more beneficial than doing stuff the old way.

VFX_Veteran · Nov 14, 2019

KKRT said:
This demo is not a Path Tracer, its ray tracing reflection tech that combines RT with SVOGI.
Your first point is just a setting, rest just do not apply here.
Its great semi step to using cheaper ray-tracing tech in next few years, the less you spend on reflection, then you have some resources for shadows, ambient occlusion or some form of GI.
We do not have and wont have for some time enough compute power yet for using RT as intended, but it shouldnt stop with using RT tech or semi RT Tech with rasterization.
Smooth tech transition its easier and more beneficial than doing stuff the old way.

I know it's not a path-tracer. I actually read the real-time papers since I do the work now. I'm all for what the devs are doing now with Control, Metro and future games. I highly doubt this technique will be used in the PC space since it came out years ago.

trugc · Nov 14, 2019

VFX_Veteran said:
Can you explain then how this solution is a feasible replacement to the real solution when it has so many limitations?

Lastly, the paper came out years ago. Why has this implementation not been adopted globally (i.e. like PBR) if it's such a feasible technique?

I'm talking about their so-called mesh tracing approach, not vanilla SVOGI. You are referencing the doc on cone tracing in SVO. How's that related to this demo?

chris 1515 · Nov 14, 2019

VFX_Veteran said:
Can you explain then how this solution is a feasible replacement to the real solution when it has so many limitations?

Currently animation on RTX/DXR are a hack because they can't regenerate the BVH every frame. Transparency is a problem too. LOD is a problem too but it is in the futures features of DXR. Realtime is make of compromise until we have enough power to do all the lighinng part with raytracing it will be the case.

By someone doing the rendering part of the Unity BMW demo at GDC.

https://auzaiffe.wordpress.com/2019...altime-rendering-when-is-raytracing-worth-it/

On the other side, the current raytracing APIs is a high level one: very little is exposed to the developers (BVH structures are opaque, no control over the recursive TraceRay coherence and dispatch, cheating is required to support skinned meshes and particles, transparents are a nightmare, material LOD along recursion is hard, etc.).

Transparency is sometimes not included inside RT, this is coming from a GDC 2019 presentation for Unreal Engine

I can't find it again but I will search I have read a blog post of a dev analysing some RTX game utilising Renderdocs and it shows all the compromise to do with animation, LOD and transparency.

A SIGGRAPH 2019 Nvidia presentation talk about the feature to add to solve some problem like being able to update part of the BVH each frame.

EDIT: There is only one raytracing architecture with the ability to refresh the BVH every frame, it is ImgTec raytracing (ex Caustic Graphics).

EDIT:

https://drive.google.com/file/d/1B5fNRFwv2LsGlCBJ8oKYRiiDUtLMR4TY/view

Presentation with the current limitation of RTX/DXR and some solution

REFITTING

Given an animation where vertices move but the mesh topology stays the same, we can refit a
BVH instead of rebuilding: just read the new triangle data and recompute bounding boxes.

For example, in RTX refitting is ~10x faster than a full build.
Quality may degrade over a long animation → sometimes trees are periodically refreshed
with a rebuild.

They think about adding

BVH CONSTRUCTION
HARDWARE

The only place with BVH construction hardware is imgtec technology the other ideas are research paper and they will create their own solution from the most interesting paper.

TREE UPDATE HARDWARE

Small field, ~10 papers
k-D tree builders (Nah, 2014; Liu, 2015)
Refitter units (Nah, 2015; Woop, 2006)
Imagination Technologies SHG (McCombe 2014)
Binned SAH sweep unit (Doyle, 2013)
MergeTree (Viitanen, 2015)
PLOCTree (Viitanen, 2018)

Intel seems to have the solution to one of the RTX/DXR problem wih LOD with more flexibility for the future discreet GPU they will launch in the future. They have a paper I need to find the link.

T002 Tyrant · Nov 14, 2019

So what they're really saying is that the Switch can support this 😲

at 1fps 😅

VFX_Veteran · Nov 14, 2019

trugc said:
I'm talking about their so-called mesh tracing approach, not vanilla SVOGI. You are referencing the doc on cone tracing in SVO. How's that related to this demo?

Ok. The demo is phrased with the buzzword "ray-tracing without dedicated hardware". In the demo, they only use this technique for reflections (which shows some limitations). I'm just trying to get clarification on why people think this is a complete replacement to what RTX is doing. If you are saying that this mesh tracing approach is completely different than the SVOGI system and doesn't have the same disadvantages, then I'll read up on mesh tracing and keep my mouth shut. Are you saying that this mesh tracing approach is a replacement of the RTX solution using hardware and can be extended to implement all the above points I mentioned earlier that we didn't see in the demo?

trugc · Nov 14, 2019

VFX_Veteran said:
Ok. The demo is phrased with the buzzword "ray-tracing without dedicated hardware". In the demo, they only use this technique for reflections (which shows some limitations). I'm just trying to get clarification on why people think this is a complete replacement to what RTX is doing. If you are saying that this mesh tracing approach is completely different than the SVOGI system and doesn't have the same disadvantages, then I'll read up on mesh tracing and keep my mouth shut. Are you saying that this mesh tracing approach is a replacement of the RTX solution using hardware and can be extended to implement all the above points I mentioned earlier that we didn't see in the demo?

I usually don't pay much attention to what they boast about in PR articles. I doubt this is a better approach than current RTX implementation as well, but I also think this is an interesting direction that's worth some more investigations. Mesh tracing is not some kind of replacement of original SVOGI. It's simply using SVO as an acceleration structure (just like BVH). The implementation of their prototype is pretty simple. Each leaf node references a list of triangles presented in the voxel. If you hit a leaf node while traversing through the octree, ray-triangle intersections are preformed for each triangle. So everything you mentioned above is perfectly doable, since it's just standard ray tracing with different data structure. The real question is at what cost? But again this requires further experiments to see.

Some advantages of this approach that immediately come to my mind are:
1. In some cases if you don't need accurate intersection info(like reflections on super rough surfaces) or RT/BVH fails in efficiency(like vegetations, stuff with thin structures), you can leverage the prefiltered radiance and visibility info stored in SVO.

2. If vendors can figure out some ways to allow for hardware accelerations on custom data structures, this will open up a lot of grounds for further optimizations, as you can choose whether or to what extent you want to perform traversal, rather than handle everything inside a black box.

VFX_Veteran · Nov 14, 2019

trugc said:
I usually don't pay much attention to what they boast about in PR articles. I doubt this is a better approach than current RTX implementation as well, but I also think this is an interesting direction that's worth some more investigations. Mesh tracing is not some kind of replacement of original SVOGI. It's simply using SVO as an acceleration structure (just like BVH). The implementation of their prototype is pretty simple. Each leaf node references a list of triangles presented in the voxel. If you hit a leaf node while traversing through the octree, ray-triangle intersections are preformed for each triangle. So everything you mentioned above is perfectly doable, since it's just standard ray tracing with different data structure. The real question is at what cost? But again this requires further experiments to see.

Some advantages of this approach that immediately come to my mind are:
1. In some cases if you don't need accurate intersection info(like reflections on super rough surfaces) or RT/BVH fails in efficiency(like vegetations, stuff with thin structures), you can leverage the prefiltered radiance and visibility info stored in SVO.

2. If vendors can figure out some ways to allow for hardware accelerations on custom data structures, this will open up a lot of grounds for further optimizations, as you can choose whether or to what extent you want to perform traversal, rather than handle everything inside a black box.

To add, what if I fill my arrays with EVERY triangle in the scene? How much is the acceleration structure updated in the demo? Has it ever been tested in a fully playing game to see what the real performance would look like? Am I only restricted to just 1 primary ray cast or can I recurse as many times as I want? How about sampling area lights instead of just looking for ray-object intersections?

I feel it's not worth the time since RTX is finally out and it addresses all of those questions. Yea, some people may have to wait it out before 2080Ti performance cost $300 or so, but that won't be too long.

trugc · Nov 14, 2019

VFX_Veteran said:
To add, what if I fill my arrays with EVERY triangle in the scene? How much is the acceleration structure updated in the demo? Has it ever been tested in a fully playing game to see what the real performance would look like? Am I only restricted to just 1 primary ray cast or can I recurse as many times as I want? How about sampling area lights instead of just looking for ray-object intersections?

I am not sure about any of this. All of these are heavily related to their specific implementation. You probably could ask someone working on the rendering part at Crytek for clarification. I remember Vlad is on Twitter but he's not quite active there.

VFX_Veteran said:
I feel it's not worth the time since RTX is finally out and it addresses all of those questions. Yea, some people may have to wait it out before 2080Ti performance cost $300 or so, but that won't be too long.

IMHO it's worth the efforts right now. RTX is several orders of magnitude faster than previous software approaches, but still not fast enough. It's also not some kind of simple plug-in then everything just works. Take BFV for example, a lot of efforts go into optimizing the ray tracing performance, like screen space tracing to avoid unnecessary hardware query, ray batching/grouping for more coherent ray traversal(even when there is a hardware scheduler designed specifically for such case), tiled based ray allocation to keep RT budget at low cost, culling out small triangles/meshes when building BVH, and so on. Whether the current implementation of RTX is the most efficient one is still debatable, so it's reasonable to try other approaches. It's also possible that next-gen console might not be powerful enough to perform full ray tracing directly.

Oh just found Dictator posted a video on this. That's quick!

VFX_Veteran · Nov 14, 2019

trugc said:
It's also possible that next-gen console might not be powerful enough to perform full ray tracing directly.

Oh that's definitely the case. I was speaking with regards to the PC. If you start mixing different approaches within a complex gaming engine, I can see things going very bad.

Barnabas Reynolds · Nov 14, 2019

Elfotografoalocado said:
There is no single game out there that has all of those techniques simultaneously. Some do only reflections (BFV, Control), others GI (Metro Exodus) and others shadows. (Shadow of the Tomb Raider, Modern Warfare)

So I don't see how it's not comparable, quite honestly. It's doing exactly what BFV and Control do.

You're wrong about Control, and I'm pretty sure Metro introduced emissives along with a DLC release.

modiz · Nov 14, 2019

you know what, nvm.

Dictator · Nov 14, 2019

trugc said:
I am not sure about any of this. All of these are heavily related to their specific implementation. You probably could ask someone working on the rendering part at Crytek for clarification. I remember Vlad is on Twitter but he's not quite active there.

IMHO it's worth the efforts right now. RTX is several orders of magnitude faster than previous software approaches, but still not fast enough. It's also not some kind of simple plug-in then everything just works. Take BFV for example, a lot of efforts go into optimizing the ray tracing performance, like screen space tracing to avoid unnecessary hardware query, ray batching/grouping for more coherent ray traversal(even when there is a hardware scheduler designed specifically for such case), tiled based ray allocation to keep RT budget at low cost, culling out small triangles/meshes when building BVH, and so on. Whether the current implementation of RTX is the most efficient one is still debatable, so it's reasonable to try other approaches. It's also possible that next-gen console might not be powerful enough to perform full ray tracing directly.

Oh just found Dictator posted a video on this. That's quick!

Thanks for posting!
Yeah the novelty of the crytek solution's accelleration structure and the fallbacks are interesting. Fallback from triangle trace to cone trace for those more rough surfaces? I like that idea quite a lot actually... But it does make me wonder how worth it could be for material correctness / denoisers themselves too will be advancing at the same time. Maybe we might not need that prefiltering!

Nostremitus · Nov 14, 2019

Ran it with my RX 480, it... Uh, hovered around the 40s, drops to 20s, spiked to 50s at 1080p...

Ran low 70s to high 80s at 720p... So I guess my mClassic with get used a lot once everything starts using raytracing... ...until I eventually upgrade...

trugc · Nov 15, 2019

Dictator said:
Thanks for posting!
Yeah the novelty of the crytek solution's accelleration structure and the fallbacks are interesting. Fallback from triangle trace to cone trace for those more rough surfaces? I like that idea quite a lot actually... But it does make me wonder how worth it could be for material correctness / denoisers themselves too will be advancing at the same time. Maybe we might not need that prefiltering!

Scenes with lots of vegetation might be a good case, which don't require accurate intersections and tracing on SVO can be much faster.

Gitaroo · Nov 15, 2019

all this features are very impression for a strictly software base solution, hope they make it into the hunt eventually. How is that game running these day? Can you lock 4k 60fps with a 2080ti?

CryEngine - Neon Noir: Hardware agnostic ray tracing benchmark

Digital Foundry