NVIDIA GameWorks Integration

[MASSIVE Can]

In this case, a large part of the fps drop comes from shadow map rendering, the ShadowDepths pass. Like I mentioned, when VXGI is enabled, the directional light shadows include all objects that are covered by the shadow map, not just those affecting the current camera view. You can see that using "vis ShadowDepthAtlas" or "vis WholeSceneShadowmap" and toggling VXGI on and off. It's probably possible to do some culling still or limit the range of cascades that are used for VXGI. Since you observe such a significant slowdown, it would be great if you could send me the scene, or something else that reproduces the same kind of slowdown, and then I can try to optimize the code (or settings).

VXGI 2.0 is generally faster than 1.0, but no one said you will see 2x speedup overall. In the GDC presentation, I mentioned that voxelization can be almost twice as fast in some cases, mostly because it can be done in a single pass now. Other things like voxel processing and tracing have been somewhat improved. But not shadow map rendering, as that's part of the engine, not VXGI. (Actually, it's probably possible to implement single-pass directional shadow map rendering using Maxwell-2 FastGS, which should give a speedup for everyone using UE).

thank you for this important information. If there is a way to have a more optimized VXGI I will do whatever I can. If you send me an email or something like that I can send a scene to reproduce. And also dynamic shadow cost is a big deal in Unreal for open world. If its performance will be improved too, with VXGI, Unreal would be a better position for open world to compare with CryEngine.

[0lento]

Hey, I'm kinda curious since you specify 'HBAO+ / HBAO+ 4' for some of the builds.
It's a bit confusing since there's already a feature in the VXGI builds that's called hbao.

It's HBAO+ from VXGI2 branch. I specify it separately as it's a separate GameWorks tech from Nvidia, it's not coupled with VXGI technically in any way, they just chose to distribute them at same merged branch.

Meaning, you can technically strip out all VXGI code from the nvidia VXGI branch and have only HBAO+ left and it still works (we've done that for UE4.15 when I didn't need VXGI for anything).

[Chesire]

Hi Alexey,

Not sure if this is easily fixed but making a typo when entering a voxelsize in the console (example ~32 as opposed to 32) will crash the editor.

[Derjyn]

VXAO and HBAO+ are in fact two different bits. HBAO+ also is in VXGI builds by default, just not sure if it's HBAO+4. Maxime, 0lento or Alexey would know better on the differences there. And ShadowWorks is something totally different from the ambient occlusion stuff.

(Actually, it's probably possible to implement single-pass directional shadow map rendering using Maxwell-2 FastGS, which should give a speedup for everyone using UE).

What would that process look like? Is that something simple to do?

[ax448]

I updated all three 4.19 GameWorks merges I've done to include latest fixes from Nvidia (for TXAA3 and VXGI2) and official UE 4.19.2:

https://github.com/0lento/UnrealEngi...4.19-Flow-VXGI (Flow + VXGI 2)
https://github.com/0lento/UnrealEngi...4.19-GameWorks (Blast + Flow + HairWorks 1.4 + HBAO+ 4 + TXAA 3 + VXGI 2)
https://github.com/0lento/UnrealEngi...9-GameWorks-VL (Blast + Flow + HairWorks 1.4 + HBAO+ 4 + TXAA 3 + Volumetric Lighting + VXGI 2)

Hey, I'm kinda curious since you specify 'HBAO+ / HBAO+ 4' for some of the builds.
It's a bit confusing since there's already a feature in the VXGI builds that's called hbao.

I'm assuming what's in the VXGI builds is actually the Voxel AO / VXAO.
While when you specify HBAO+, it's the screen-space HBAO from a separate branch?
( aka the one from here https://developer.nvidia.com/shadowworks , but without shadowworks itself, which doesn't seem to have any kind of UE4 implementation yet. )

[Alexey.Panteleev]

Alexey.Panteleev I have just tested the updated branch. I noticed VXGI voxelization and tracing frame times are good but when I enabled VXGI (Only Directional Light) with optimized parameters, my fps drops 97 to 48 on GTX 980Ti because of shadow depths. Is It usual or there is another problem? I uploaded GPU profiler results. First: Shadows and VXGI enabled (GI only). Second: Shadows enabled but not XVGI. Third: Shadows and VXGI disabled. Fourth: Shadows disabled VXGI enabled (GI only). Also, I noticed that in fourth scenario GPU visualizer shows 3 ms less than viewport values. Is this "half of the fps" situation nature of VXGI? I thought that VXGI 2.0 would be much faster than old one, up to 2x according to updates. But fps which I had drop doesn't verify this is true.

In this case, a large part of the fps drop comes from shadow map rendering, the ShadowDepths pass. Like I mentioned, when VXGI is enabled, the directional light shadows include all objects that are covered by the shadow map, not just those affecting the current camera view. You can see that using "vis ShadowDepthAtlas" or "vis WholeSceneShadowmap" and toggling VXGI on and off. It's probably possible to do some culling still, or limit the range of cascades that are used for VXGI. Since you observe such a significant slowdown, it would be great if you could send me the scene, or something else that reproduces the same kind of slowdown, and then I can try to optimize the code (or settings).

VXGI 2.0 is generally faster than 1.0, but no one said you will see 2x speedup overall. In the GDC presentation, I mentioned that voxelization can be almost twice as fast in some cases, mostly because it can be done in a single pass now. Other things like voxel processing and tracing have been somewhat improved. But not shadow map rendering, as that's part of the engine, not VXGI. (Actually, it's probably possible to implement single-pass directional shadow map rendering using Maxwell-2 FastGS, which should give a speedup for everyone using UE).

[Gomo]

0lento - I've managed to package project using your branch.

[MASSIVE Can]

Alexey.Panteleev I have just tested the updated branch. I noticed VXGI voxelization and tracing frame times are good but when I enabled VXGI (Only Directional Light) with optimized parameters, my fps drops 97 to 48 on GTX 980Ti because of shadow depths. Is It usual or there is another problem? I uploaded GPU profiler results. First: Shadows and VXGI enabled (GI only). Second: Shadows enabled but not XVGI. Third: Shadows and VXGI disabled. Fourth: Shadows disabled VXGI enabled (GI only). Also, I noticed that in fourth scenario GPU visualizer shows 3 ms less than viewport values. Is this "half of the fps" situation nature of VXGI? I thought that VXGI 2.0 would be much faster than old one, up to 2x according to updates. But fps which I had drop doesn't verify this is true.

[CGICollector]

r.VXGI.AmbientOcclusionMode= 1 means you can only use VXAmbientOcclusion.
r.VXGI.AmbientOcclusionMode= 0 means you can use Full VXGI+AmbientOcclusion.

[pedrosunye]

It won't work on a 780 because it's Kepler. You need at least 750 Ti (Maxwell-1) or 9xx to run VXGI in non-AO mode. Also, are you running the latest version that I pushed yesterday, or something older?

thanks for the quick reply. I just tried your latest build and the issues are gone, no more red light and no more crashes. Is there anyway I can use VXGI in this build with my 780, since r.VXGI.AmbientOcclusionMode= 0 means no VXGI and my GPU doesn't support r.VXGI.AmbientOcclusionMode= 1?

[CGICollector]

Then I keep trying. VS2015 may be having problems. Thanks for the info.

[XenthorXi]

Please, can somebody test trueSKY with VXGI 2?

I'm using Truesky with VXGI 2.0 since VXGI 2.0 came out and have no issue ^^

[CGICollector]

Please, can somebody test trueSKY with VXGI 2?

[0lento]

0lento - your branch seams to work fine. I had some issues but they turn out to be my bugs.
So for tried on three projects - no crashes so far.

That's nice to hear, but do test the packaging too. Like mentioned I couldn't use shipping conf on Volumetric Lighting for some reason on my initial merge. There could be some issues with other techs on shipping builds too as I haven't tested that conf extensively.

Development conf on each GameWorks tech seemed to package just fine (I've test packaged and ran the packaged version for at least one sample from each GameWorks tech to make sure everything seems to be in order on packaged builds as well).

[0lento]

I updated all three 4.19 GameWorks merges I've done to include latest fixes from Nvidia (for TXAA3 and VXGI2) and official UE 4.19.2:

https://github.com/0lento/UnrealEngi...4.19-Flow-VXGI (Flow + VXGI 2)
https://github.com/0lento/UnrealEngi...4.19-GameWorks (Blast + Flow + HairWorks 1.4 + HBAO+ 4 + TXAA 3 + VXGI 2)
https://github.com/0lento/UnrealEngi...9-GameWorks-VL (Blast + Flow + HairWorks 1.4 + HBAO+ 4 + TXAA 3 + Volumetric Lighting + VXGI 2)