@RYStorm most of the Sony first parties are still going on their own tech. SMS, Naughty Dog, Team Asobi, Sucker Punch, San Diego, Bungie all have custom engines. And of course Guerilla as you already mentioned. 🙂

@mtothevizzah regretfully, we no longer do lighting on SPUs over here at SMS.

@Doomed_Daniel @ByteArcane I don’t know if any offhand, but you could use texassemble.exe to make some: https://github.com/microsoft/DirectXTex/wiki/Texassemble

@Doomed_Daniel @ByteArcane you can also check out the Sun Temple and Bistro scenes, I believe they both have .DDS textures

https://developer.nvidia.com/ue4-sun-temple

https://developer.nvidia.com/orca/amazon-lumberyard-bistro

@Doomed_Daniel @ByteArcane they’re here now: https://github.com/walbourn/directx-sdk-samples

In today's new blog post, I try to provide some intuition and clear up some misconceptions about early Z cull behavior on desktop GPUs. Have a read if that interests you!

https://therealmjp.github.io/posts/to-earlyz-or-not-to-earlyz/

@floooh @sjb3d of course, and many engines have over-complicated systems for doing such a thing. If you can share structs none of that is necessary.

@artemist @aeva I was referring to host memory specifically, not device memory. Sorry if that wasn’t clear.

@rygorous @aeva it looks like you’re correct though about Intel flip-flopping on LLC being used, so thank you for pointing that out: https://chipsandcheese.com/p/igpu-cache-setups-compared-including-m1

as far as coherency goes it should be fine either way, I think

@rygorous @aeva maybe it’s not snooping, or I’m using the wrong terminology? This is mostly coming from conversations with people smarter than me explaining that uncached CPU memory was not required for coherency. I’ve at least verified on my own discrete GPUs that both WC and WB memory allocated through D3D12 has equivalent GPU read perf and caching behavior. But happy to be corrected if I’m way off. 🙂

@aeva it should be fine, I think. Even discrete GPUs snoop CPU caches over PCIe, although (write combined) uncached memory is popular for CPU-write-only data to avoid polluting the CPU cache. I believe Intel and some other integrated GPUs have a shared last-level-cache between CPU and GPU which makes things very optimal for CPU->GPU traffic, and removes any reason to use uncached memory for the CPU.

@bitinn yeah not unexpected at all really, they face rather severe financial consequences ($5000 per user) if prosecuted. It looks like it’s gone from the Google Play store as well.

@mtothevizzah I used to work right by Centre Street in NYC where they would always film the “lawyers talking on the steps of the courthouse” scenes for Law & Order. Weirdly I was never murdered!

@bitinn

@archo VT page generation at runtime is very much a thing though, since you still get the massive compression benefit of shipping tiling textures that can be used over and over again throughout the game. But even then the VT system will still spit out normal + roughness maps, since you want that for dynamic lighting. Or you can go even further and do full texture-space shading using the bones of a runtime VT system.

@archo if you were to bake lighting/normals into the color map you would effectively be having a gigantic combinatorial explosion on your texture sizes, MegaTexture-style. That's a very tough pill to swallow for all kinds of reasons, both in terms of dev workflow and end user experience. People already complain about filling up their HDDs. Which is why that idea of shipping unique VTs never went anywhere post-Rage.

@archo I very much doubt that assumption would hold for all viewing angles or lighting conditions in which that wall material ends up getting used. Grass has tons of specular, especially up close! Very few materials have effectively no noticeable specular, especially at grazing angles.

@archo the whole “wet look” of the road in the left screenshot is coming from normal and roughness maps. I’m not sure how you would replicate that with just diffuse. Or even if you could, it would surely not look right from multiple view angles.

@archo those two scenes would look wildly different without roughness/normal/maps, and *especially* if you dropped specular entirely. Even if you had infinite resolution baked diffuse response through VT or similar they wouldn’t look the same, and you’d have no ability to do dynamic lighting.

@ataylor if it’s just a 2D grid then you’ll probably want to stress test scenes with lots of Z complexity, since that’s where the 2D techniques start to fall over vs. 3D grid approaches. Something like an alpha-tested fence in the foreground with stuff in the mid and far field within the same tile, with lights all throughout the Z range.