Realtime Graphics / Game Engines

[Darren S]

I feel like I've just been schooled.... 😁

Good post mate! I even followed the vast majority of it.

 

[SharkyUK]

I feel like I've just been schooled.... 😁

Good post mate! I even followed the vast majority of it.

It's easy when you know how... 🫣 🤥
(I haven't got a bl00dy clue!)

 

[SharkyUK]

 

[SharkyUK]

I haven't touched my path tracing project for a while and decided to upgrade it to use the latest version of CUDA this evening. It didn't work out so well, so it's now going to be a job for another time! 🤣 However, I did make a few tweaks and improvements in readiness for the forthcoming 50xx series GPU from nVidia and I'm looking forward to getting my hands on one or two of those. There are also one or two features that I want to introduce into the renderer to see if I can get them running in real-time and without too much of a hit on performance.

I've not got much to show and the changes I have made won't show up that well in the generated imagery, but here are a few more renders for shizzles and giggles.

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

 

[SharkyUK]

And some real-time Lego renders, because fun. 🤣

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

SIPT2 Render by Andy Eder, on Flickr

 

[SharkyUK]

I decided to spend a couple of days fixing some long-standing bugs with my path tracer. The main culprit has been causing me headaches for a couple of years now and has been a b!tch to debug. I found it after a 14 hour debugging session. I was using variable "nl" when I should have been using "n". But had only made this error in one place in the code. Every other occurrence was fine.

Feeling a bit happier now it has been fixed.

This first image was rendered using Blender (Cycles render engine) - which I use as a benchmark for my own stuff.

(Blender)

And this next image was rendered using SIPT2 - my real time path tracer...

(SIPT2)

I'm quite happy with that.

A few more random renders for good measure.

 

[Darren S]

As always - very impressive stuff that my brain would probably comprehend 1% of!

 

[SharkyUK]

As always - very impressive stuff that my brain would probably comprehend 1% of!

You'd have spotted the bug straight away, mate. It's quite obvious when you think about it. When calculating a ray's path at the interface between different materials (involving dielectrics - which are non-conducting and transparent and can both reflect and refract a light ray), and assuming the ray does not end up in total internal reflection, when calculating the probability of of either generating a reflection or transmission/refraction ray, you need to use the underlying geometry's raw normal. I was using a corrected normal based on microfacet theory by mistake. The corrected normal for microfacets should be used for shading and calculations in general, but not when it comes to deciding if the ray should be reflected or refracted. It messes up the probability distribution, adversely affects light radiance weighting, and breaks the law of energy conservation. And can muck up the rendering. The issue was that the messed up render is not too dissimilar to a physically correct render, hence why I kept missing the issue. Piece of cake, y'see.

 

[Matt Cup]

You'd have spotted the bug straight away, mate. It's quite obvious when you think about it. When calculating a ray's path at the interface between different materials (involving dielectrics - which are non-conducting and transparent and can both reflect and refract a light ray), and assuming the ray does not end up in total internal reflection, when calculating the probability of of either generating a reflection or transmission/refraction ray, you need to use the underlying geometry's raw normal. I was using a corrected normal based on microfacet theory by mistake. The corrected normal for microfacets should be used for shading and calculations in general, but not when it comes to deciding if the ray should be reflected or refracted. It messes up the probability distribution, adversely affects light radiance weighting, and breaks the law of energy conservation. And can muck up the rendering. The issue was that the messed up render is not too dissimilar to a physically correct render, hence why I kept missing the issue. Piece of cake, y'see.

Indubitably!

 

[SharkyUK]

With plenty of time on my hands, I've been updating and reworking a few bits on my path tracer project, as well ass adding a few new features.

The under-the-hood stuff for the path tracer operates in linear RGB HDR (high dynamic range) space (it's not spectral yet, but that's a HUGE rewrite). At some point, for display on regular display devices or saving to non-HDR file formats, the HDR image needs to be transformed (tone mapped) to LDR (low dynamic range). This can be quite a tricky process as you have to try and preserve image detail in dark areas as well as excessively bright areas, and then somehow map those extreme colour intensities in to a [0,1] range (or [0,255]) RGB.

TLDR; take the rendered image from the software and make it displayable on HDR and regular display devices by dicking around with the colour range.

This is done through a process called tone mapping. In the past, I used a naïve Reinhard Luminance tone mapper (or simply cut any values off that fell out of the [0,1] range. This didn't produce the best results. So, I decided to pick a few different tone mappers and implement them.

Implemented ACES (which gives a nice punchy look), AgX (which is more realistic and probably the best one for use in a render + vfx + post-production workflow), and in improved Reinhard.

Probably not the best 3D scene to show the differences, although they can be observed in the image below.

I also decided to rework my post-process bloom effect as I wasn't happy with my old version. It's a lot quicker now, and produces much better results. It is not physically realistic (far too expensive for a real time renderer) but it does the job and is user-controllable.

I also added a reworked "sheen". It's based on "Practical Multiple-Scattering Sheen Using Linearly Transformed Cosines" (Zeltner, Burley, and Chiang, SIGGRAPH 2022) that introduces a method for real-time, physically-based rendering of fuzzy or dusty materials. It models sheen as a volumetric layer containing fiber-like particles (using the SGGX distribution) and approximates the resulting multiple scattering using an LTC fit, replacing expensive simulation. Good for clothing and things like velvet. Quite happy how this turned out.

In addition, I reworked my transmission code that models light passing through a surface for things like curtains, or strong sunlight shining through leaves. Good for backlighting! Here is an example of a bright light under a piece of fabric.

And a bright light shining through stained glass:

Another area to update was the coat (clearcoat) processing; this is good for adding a gloss/lacquer layer on top of an existing surface/material. For example, the lacquer layer over car paint, or even moisture/wetness on skin. Or varnish on wood. That sort of thing. I gave the McLaren in the image below a dull Papaya Orange paint layer with high roughness - which meant it appeared diffuse with no highlights - matte orange if you like. I then applied the coat on top which simulates the lacquer clear coat, and suddenly the paintwork has reflections and gloss as you would expect to find in the real world.

A woman:

A sweaty woman:

Bored yet?

 

[SharkyUK]

A few more random images from the latest version of the project...

Ta muchly.

 

[charltjr]

Yes yes, but when are we going to get photorealistic real time AI p**n women? Chop chop man.

 

[SharkyUK]

Yes yes, but when are we going to get photorealistic real time AI p**n women? Chop chop man.

It's the best I can do for now. Will try harder. Promise.

 

[SharkyUK]

I finally found a couple of hours to migrate my path tracer code base from CUDA 11.8 to 13.1 - fascinating, right?!

 

[SharkyUK]

Well, the upgrade to CUDA 13.1 didn't last long... now on 13.2...

Whilst I am still looking for my next job, I decided to update my path tracer and implemented an improved light scattering system. Rays travelling through a 3D scene now interact with participating media (e.g. particles in the atmosphere) if present. It's quite an expensive process, but gives some cool results - such as realistic fog. Mixed with a 3D texture, it can also be used for rendering volumetric clouds, flames, etc.

 

[Mr Squashie]

No doubt you've seen this already, but some of these demos are ridiculous. The Starfield one is especially impressive given that the game doesn't even use basic ray tracing, let alone path tracing.

 

[dann2707]

Well, the upgrade to CUDA 13.1 didn't last long... now on 13.2...

Whilst I am still looking for my next job, I decided to update my path tracer and implemented an improved light scattering system. Rays travelling through a 3D scene now interact with participating media (e.g. particles in the atmosphere) if present. It's quite an expensive process, but gives some cool results - such as realistic fog. Mixed with a 3D texture, it can also be used for rendering volumetric clouds, flames, etc.

View attachment 1780175
View attachment 1780176
View attachment 1780177
View attachment 1780178
View attachment 1780179
View attachment 1780180
View attachment 1780181

I think you collectively destroyed Primarks WiFi on Friday when I was in there loitering around waiting for the Mrs and I tried loading this page up 😂

Was genuinely like being back on dial up when the images just bit by bit load up progressively 😂😂😂

 

[SharkyUK]

No doubt you've seen this already, but some of these demos are ridiculous. The Starfield one is especially impressive given that the game doesn't even use basic ray tracing, let alone path tracing.

Yeah - there is some crazy stuff coming. DLSS5 has ridiculous potential (despite the early hate it is getting from folks losing their s**t over it - which is hilarious to see! It is such a misunderstood technology). I think it is going to be a tough ask to have it up and running before the end of the year in mainstream DLL and driver releases... but I also have to say that the speed at which the AI and ML stuff is evolving is staggering. DLSS5 ability to "understand" lighting and material response is on another level. Path tracing without having to path trace.

It's stupidly hungry though in terms of the processing oomph that is needed. It will be interesting to see where the major performance increases come from in order to see it running on a single GPU with a respectable modicum of performance; even on high-end 50x0 series GPUs. It's not just DLSS5; there are some other changes and new tech coming along to support it and that looks pretty special, too.

I think you collectively destroyed Primarks WiFi on Friday when I was in there loitering around waiting for the Mrs and I tried loading this page up 😂

Was genuinely like being back on dial up when the images just bit by bit load up progressively 😂😂😂

Sorry mate! I have a few more coming soon. I've been working on extending my path traced volume support to align with with DreamWorks Animation developed - as seen in movies such as the newer Star Wars, Marvel movies, etc. Admittedly, my version is cutback a bit as I am trying to run it in real time rather than have a data centre of 10,000 GPUs render it over the course of a week!

 

[dann2707]

I have a few more coming soon. I've been working on extending my path traced volume support to align with with DreamWorks Animation developed - as seen in movies such as the newer Star Wars, Marvel movies, etc. Admittedly, my version is cutback a bit as I am trying to run it in real time rather than have a data centre of 10,000 GPUs render it over the course of a week!

 

[Crispy.]

View attachment 1780794

If you ever have a name change on here, it should be GIF Nonce What do you search to find the likes of these?! haha