Announcement

Collapse
No announcement yet.

Upcoming Push for dynamic GI after Neo & Scorpio Announcements?

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

    Originally posted by Maximum-Dev View Post
    I'm blown away by SVOTI, again.

    Having these light bounces without waiting even 1 second, and still running so smooth to develop any games with, is truly amazing...

    http://www.dsogaming.com/screenshot-...dible-visuals/
    even more amazing is this:

    https://www.youtube.com/watch?v=VqXL7QMWN-8

    Comment


      Hmm, I wonder why they put resources into a technique that is so old and clearly inferior to what the competition is doing.
      ArtStation

      Comment


        They need to drop LPV and go with one of the better solutions

        Comment


          Maybe Epic can use combined ideas from VXGI, SVOTI, AHR, Enlighten to create superior LPV or just help RyanTorant to code AHR into usable stage and integrate/replace LPV. Just my solution.
          tox.chat - Skype alternative
          blender.org - 3D suite

          Comment


            they also have a "multiple bounces for skylight" card. hmm... interesting to see where this leads.

            Comment


              Originally posted by AE_3DFX View Post
              Maybe Epic can use combined ideas from VXGI, SVOTI, AHR, Enlighten to create superior LPV or just help RyanTorant to code AHR into usable stage and integrate/replace LPV. Just my solution.
              If they could, there would be some major wizardry involved. They're pretty incompatible techniques as far as I'm aware. VXGI/SVOGI are completely different solutions than what LPVs do, while Enlighten actually prebakes the GI and uses material magic to make it react to lights properly.

              The only way I could see that happening is if it automatically uses LPVs for lower spec (which they -should- be able to do on any hardware that even meets the min spec of the UE4 seeing as how my 6 year old laptop could run Crysis 2 with LPVs), while using the more complex SVOGI/VXGI tracing system for higher specs on dynamic objects, while doing an Enlighten style bake for static objects. Even typing that out, I'm not even sure it would make sense doing it like that, there's a lot of repeated and redundant lighting systems in there.


              The push for LPVs though is an interesting one, I am curious as to what they have in mind. It is worth noting however that the trello board says "investigating", so it may just be some research on their part to see if it's viable to use LPVs and not any specific action in mind yet. If nothing comes up, I wouldn't be surprised if a SVOGI-like system was reconsidered now that a considerable amount of time has passed since the initial reveal of the tech 4 years ago.

              Comment


                Originally posted by Chariots
                I recently did a test, on a gtx970 VXGI has 20ms
                You must have it configured wrong, it only has a 10 ms hit on my GTX 780 in a VERY heavy scene I'm working in with a lot of moving foliage. In the included sci-fi hallway demo they modified, VXGI is a 9ms hit. VXGI is meant to run faster on 900 series GPUs, it should be even faster for you. On the 980 Ti I tested it on once before, it was only a 4-6ms hit as well, it was getting pretty negligible.


                Originally posted by Chariots
                Crytek uses light probes in conjunction with SVOGI, probably(?) why they have it down to 2-4ms on gtx780, its not the vanilla SVOGI that nvidia published back in 2011 I think.
                Crytek's SVOGI runs independently of light probes, it doesn't use them except for the areas that SVOGI doesn't cover like in the distance and for reflections if specular tracing is disabled. It has nothing to do with the performance. Also, vanilla SVOGI from the original elemental demo was running at 30 fps in 1080p on a GTX 580, it's not too far fetched to say that after 4 years of work a GTX 780 would take a 2-4ms hit.
                Last edited by Zero-Night; 07-19-2016, 02:53 PM.

                Comment


                  Originally posted by Chariots
                  I recently did a test, on a gtx970 VXGI (on lowest possible settings) has 20ms, DFGI has 35~ms, LPV has 4ms cost for the gpu on a scene that gives 120fps capped otherwise. Crytek uses light probes in conjunction with SVOGI, probably(?) why they have it down to 2-4ms on gtx780, its not the vanilla SVOGI that nvidia published back in 2011 I think. LPV is probably the most performant way of doing dynamic gi in unreal without doing major architectural changes, or doing risky R&D with DFGI. Which is kind of a shame, as I think DFGI looks better than VXGI on open areas.
                  The issue with crytek is that if they see something is expensive, they work their way to make it cheap. But they don't settle down to move it to backlog. Same with their SVOTI. And that's still a WIP, I don't know what a beast it will be later on but point being. But point being, it's the old saying that says Either do one thing, and do a great job. Or do a lot of things and none of them will be done as good.
                  Artstation
                  Join the support channel
                  Gumroad Store

                  Comment


                    Originally posted by Chariots
                    I've used Pirates Island scene from the marketplace, Its actually a very small scene. I'm getting 120fps capped on 1080p in that without VXGI (its still dynamic lighting). I did try it on simpler scenes, and it gives much better result, and getting 20ms hit on a small but full-ish scene puzzled me so I asked about it on gameworks section, but the answer I got was that it was an accurate case.
                    Yeah I saw the post I just didn't have time to reply since roughly 1/3rd of the posts on that thread are discussing performance. Looking at the screenshots of the scene on the marketplace (not buying it just for the sake of running a performance test), there's no reason why it should be running so slow unless it was a misconfiguration on your end. A 20ms hit is -not- standard for VXGI, my 3gb GTX 780 doesn't even technically hit the minimum recommended requirements for using it and I'm still getting half that performance hit that you are, while my tests with a 980 Ti on other machines show it having the same performance impact as LPVs do on that hardware.


                    Originally posted by Chariots
                    I think they are using them to increase the quality of the final image? I'm not really sure to be honest. I assumed that it allowed them to run SVOGI on an even lower setting to get better performance. This site kinda mentions it also: http://byzantos.com/cryengine5-svoti/
                    Only on lower specs when specular tracing is disabled. It's not a requirement and doesn't have any impact on performance from my testing.

                    Comment


                      Originally posted by Chariots
                      Here is what I did for configuration:

                      I mean, is there any other configuration that I'm not aware of? Is it even possible to set it up wrong?
                      VXGI with the settings tuned high enough gets the same quality as Lightmass itself does. By default to show off the tech, Nvidia set the default settings for everything higher than probably necessary.

                      Do the following:

                      r.VXGI.Mapsize 32
                      r.VXGI.Multibounceenable 0
                      r.VXGI.HighQualityEmittanceDownsamplingEnable 0
                      r.VXGI.StoreEmittanceInHdrFormat 0

                      In the post process:

                      VXGI Diffuse:
                      Number of cones - 4
                      Tracing sparsity - 4
                      Cone rotation off
                      Random cone offsets off
                      Tracing step - 2

                      VXGI Specular:
                      Probably not a good idea to have this on for a GTX 970, but if for whatever reason you really want it:
                      Filter - off
                      Seriously though, disable specular tracing if you want to squeeze a bit of performance out. It doesn't have much of one though, I find it's usually the difference of 1ms when the filter is off.

                      The results:

                      Click image for larger version

Name:	results.png
Views:	1
Size:	103.2 KB
ID:	1112286


                      Seeing as how my hardware doesn't even make the minimum spec for VXGI to begin with, I would say that's not a bad result.

                      If you're still trying to get more performance out of it, disable VXGI on things that update every frame that won't impact the GI results that much, like the ocean material. It's a per-material setting.

                      Originally posted by Chariots
                      Quoting the site I linked:

                      Especially the integration of CryEngine’ lightprobes seems to be responsible for a good chunk of the magic here. Without this, you end up at the highest SVOTI config and oddly enough, this level fails to deliver the dramatic impact of the one below.

                      I haven't done and tests on removing the lightprobes themselves though.
                      Right. The lightprobes are good for lower spec or disabling SVOGI, but there's a reason that SVOGI has an option to disable the lightprobes entirely, they change the look visually and nothing more. They don't impact performance. At all. It doesn't run faster with them, it doesn't run faster without them. The lightprobes are basically glorified ambient lights that use an HDR image, and don't look correct with a major shift in lighting to begin with, that's why SVOGI has the option to disable it when enabled.
                      Last edited by Zero-Night; 07-19-2016, 03:26 PM. Reason: Added more console commands

                      Comment


                        Yeah that's why I attached the image, it's easier to see the results that way, and it's the only kind of proof I can provide to show it does indeed work. The step/sparsity settings are some of the more important ones to have set, they have a pretty significant performance benefit using them. The tracing sparsity is the more important one, if you had to only pick one then that's the better option. It's more important to have the sparsity set to 4 for exterior scenes with a lot going on, if it's interior you can probably push it up a bit along with the cones to get more accurate results. Thankfully it's all in the post process settings so you can set it differently in different parts of the map.

                        Your 970 should be superior to my 780 in every way for this. If you're getting a 20ms hit, then I would definitely say it's something wrong with the scene in some way messing with VXGI. Even with a mapsize of 128, I still don't get a 20ms performance hit on my 780, in 1080p with a mapsize of 128 it's a 17ms hit. So for there to be such an enormous gap between what I'm seeing and what you're seeing, it must be either the scene having something causing problems or a configuration issue.

                        Comment


                          Originally posted by Chariots
                          Only thing that comes to my mind is maybe the scene has lots of intersecting meshes, could that be messing up with voxelization perhaps? Otherwise, there is really nothing notable on the map. some low poly trees, 5-6 huts at most, some rocks, and a path that takes 10 seconds to cross. I'll post some stats on gameworks thread when I get back, we kinda derailed the thread a bit

                          I see this is as just another proof of Crytek dabbling in black magic. 2-3ms on 780 is unnatural.
                          It shouldn't matter if there's intersecting meshes really, it only voxelizes meshes that the light actually touches directly unless you have multibounce on. I'm not sure, but I would recommend you try building a scene from scratch with the assets and see if you can pinpoint a problematic mesh, I'm sure Nvidia would be interested to find out what's causing such a huge problem.


                          Back on topic though, 2-4ms on a 780 for Crytek's SVOGI is an amazing feat and they should be commended for it. I'm sure if the UE4 had realtime GI with that low of a performance hit, many users would switch over to using it overnight.

                          Comment


                            I know nothing about coding beyond some basic HTML and PHP tags but since the cryengine source code is available can't anybody take the advantage of looking into it? can't any team of programmers just have a quick look at it and bring back SVOGI and apply the improvements they learn from Crytek's integration instead of spending time on LPVs ? Can't we make a patreon page and fund them to do so?
                            Artstation
                            Join the support channel
                            Gumroad Store

                            Comment


                              I'm fairly sure the only thing stopping that is the license to the CryENGINE preventing it. There's nothing actually really stopping them from doing it on a technical level, it's just illegal. The license I believe stops you from taking components from one engine and putting it in another.

                              Comment


                                There are bunch of whitepapers on various types of real-time GI. A major chunk of work is optimization as implementing whitepaper is relatively trivial (if engine's architecture allows it).

                                Comment

                                Working...
                                X