-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
PS4 Pro provides checkerboard rendering in hardware - no cost to devs
- Thread starter Blanquito
- Start date
Futurematic
Member
They added dedicated checkerboard rendering hardware to the GPU. Your use of made-up term "uprendering" is still silly, but even if you had the details wrong you were right that Sony had a high quality / low resource way to hit 4K. They did it primarily via hardware and not software like you speculated, but hey at least you nailed the ease of it part.
Which, incidentally, makes the comment in the leaked dev docs about developers not being able to hit 1800p checkerboard (900p base image) to contact Sony make a lot more sense. It's calling out incompetent developers lol, because any developer should be able to hit 900p on the OG PS4.
Self_Destructive
Member
Why is it silly when the patent actually uses it?
https://patents.google.com/patent/US20160005344A1/en
And this is not 4K.
The checkerboard technique by itself is, at least if we're on the semantic train where upscaling and upsampling are equivalent.
If by "upscaling" we specifically mean "render an image filling in all samples on a rectangular grid, then upsample it to a larger image with all samples on a rectangular grid filled in", then I guess "upscaling" would be a bit imprecise. I would guess that the tweet is addressing it in this sense and/or is accounting for further temporal reprojection.
It's uprendering dammit I don't care that no one else is accepting the name lol it's exactly what is happening
Futurematic
Member
Because the patent also makes up a word to try and explain what it is doing. Which is complicated, so it makes sense why they invented a word, but also not what the poster was referring to in many many previous posts.
I could file a patent with all kinds of made up words that would be granted as long as the rest of it clearly explained what I was doing.
(Edit: I am utterly bemused in a fun and charming fashion given, um, well y'all know and no offence, that onQ123 managed to get the most important grafPS4 Pro can do 4K smartlydead right.)
To put a but more color on that, from the leaked dev docs they seem to want at least 1800p - at least if you can't reach that, they were asking you to contact them. 1800p is 2.5x 1080p.
The kit also supports a range of intermediate resolutions beyond that, up to native 4K.
So is it sort of like "upscaling" that Xbox does to output everything at 1080p? If so how in particular is this any different /better /worse?In a rough, very small nutshell.
Render the red pixels.
Calculate everything else.
fabricated backlash
Member
Yep,its blast processing all over again.
The checkerboard pattern and upsample filter produces pretty good results and should play pretty nicely with temporal AA. But, yes: it's rendering fewer unique new pixels each frame, and (all other things being equal) will not produce results as good as full spatial 4K rendering.
I think there's a conceptual mix-up here; if they were setting 1800p as a minimum and checkerboard rendering was used, that would be a post-filter value, and would imply a spatial res of 2x900p. Much like how, starting from 2x1080p, you can upscample into 2160p.
They were talking about that res in the context of simple scaling - scaling up from a base 1800p every frame. I assume similar recommendations apply about base res for checkerboard rendering - e.g something like a unique 1800p worth of new pixels per frame in a checkerboard pattern.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Every frame is rendered at half res. What would you call that?
No, an upscale like that is just an interpolation of pixel values. This reprojects samples from one frame to another in the 'holes' of the alternate checkerboard pattern.
That would result in a post-filter resolution of around 40% higher than 4K, which would be a somewhat strange arbitrary line to draw. We're also currently being told that that's not what games are actually doing.
It doesn't have to. Part of the excitement is a spatial scaling technique that gets pretty good results by itself. The checkerboard approach should play nicely with temporal reprojection, however.
The misleading thing about calling it upscaling is that this technique results in true 4K rendering when things are still, since all of the pixels will actually come from rendering in that case. Spatial upscaling can never create additional detail, it just spreads the existing detail over more pixels. In this technique, the additional detail is actually rendered, but it takes two frames to get there. When there is motion, the previous set of pixels will be inaccurate and have to be approximated, and that is more like upscaling (although the checkerboard and temporal aspects will make the imperfections look quite different than with normal, spatial-only, upscaling).
"Uprendering" may be a little better, since people will not automatically believe they understand what it means, while "upscaling" will make them think they understand when they actually don't. But it's hardly a self-explanatory term, either. You'd need to say something like "checkerboard rendering with temporal interpolation/reprojection" to really convey what's happening.
fabricated backlash
Member
But it's not the same method, it's not the same technique. Get Reggie in here to explain it to you.
It invents detail. It doesn't render it. If it would render it, it'd be native. That's what good upscaling does. It looks at the samples and creates an aproximation based on the surrounding samples.
Shit people this is stuff that was used in Photography for decades, with meatier solutions being quite impressive.
It's only now that you can get the same processes in real time which in itself is impressive, but it's not rendering true 4K. Even when still, it's just approximating half of the resolution.
But having seen these techniques for years, a good implementation will always be more cost effective than to brute force it.
It's just harder for PR folks to sell than just straight up power. Stop buying into cheap PR:
This is upsampling. That's literally what it does. Different algorythm to other reconstructions, but still a reconstruction nonetheless.
justsomeguy
Member
Different tech probably but that eshift is a thing of beauty, I've seen it side by side with a real 4K and it's wonderful. Too laggy for me for gaming though so I went with the Sony hw45
Dictator93
Member
Wait a minute, how is that possible exactly?
Is it taking alternately checkered frames and combining them?
justsomeguy
Member
There's a separate thread discussing it. Rainbow 6, qb potentially use it or similar it seems. So yes it IS an existing thing and it has been slated before on xb1.
However the difference is in your starting res. Up scaling 720 to 1080 vs up scaling 1080 to 4K makes a difference according to discussion in that thread)
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Yes.
It's still not native 4k either. There will be artefacts.
What I was referring to as upscaling is just the sample pattern and spatial upsampling technique. Temporal reprojection can be used to help fill in the gaps, but that's not specific to checkerboard patterns, and I'd be surprised if it's part of the hardware support.
Dictator93
Member
Where did you get that information from?
This is not the same thing though, that uses an alternating and offset 2XMSAA to try and reconstruct a combined image.
As I understand it, this is what we know:
In a rough, very small nutshell.
Render the red pixels.
Calculate everything else.
A purely spatial upscale, not a reprojection from previous frames that were offset differently.
fabricated backlash
Member
You missed my Reggie joke... I'll see myself out.
I know it's not native, I was making fun of people trying to spin this up sampling method as something ground breaking new tech that has just been decreed upon the world, and was unheard of before.
It's the feel of 4K.
justsomeguy
Member
Faux K
4Kinda
We definitely need a new term 😂
Plus one for when Scorpio comes out with its own techniques
Nobody with an ounce of sanity ever expected it to be. You're not dropping any inconvenient truths here. You're just being pedantic.
If you're not going to be content until all games are rendered in native 4k then feel free sit on the bench for the next several years. Just try not to pretend like this checkerboard 4k business is akin to a console not being able to hit 1080 in 2014 and beyond. That's ridiculous.
.
It's not about what a specific console can do at a given year, it's about people looking at an image and judging it from there instead of reading stuff on the internet and finding out the image wasn't native and using it as shit flinging ammo.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
That's how one does checkered rendering. If you don't invert the checker pattern each frame you get no temporal benefits, reprojection or not.
fabricated backlash
Member
Exactly. It just pisses me off that games that used similar tech before were getting lambasted, and now suddenly it's gods gift to mankind.
It always has been a great idea to up scale images for enhanced fidelity. Unlike cheap and mostly free solutions however, Good algorythms can do wonders, and in some cases be almost indistinguishable to native resolution (for still images)
Sure if you combine them side by side with Native images, you will see artifacts, as it's always just an approximation, but like Cerny correctly stated. Brute forcing 4k on a console budget with 2016 tech is just not possible. And it'd be downright stupid to try.
These solutions have allowed digital photography to produce images for big banner ads ever since professional cameras had only 1 to 2 Megapixel and still get respectable results.
Back then it took minutes to upsample a single RAW File, now we're at a point where similar and technically even better stuff can be done about 60 times a second.
It's a clever solution and frankly the only viable one to get any results on consoles today.
It's at least honest.
Struggling to hit 4K is a bit more impressive than struggling to hit 1080p 😂
LuvOfThaGame
Banned
Am I mistaken or did Respawn say this is exactly what TF2 is using?
"fauxK", (stolen from the projector community with their 4k shift tech)
Dictator93
Member
But where are you getting that informatoin beyond trying to deduce from what other techniques that are not this one?
What you are saying has a whole lot of implications and assumptions.
Where if anywhere does it say that the hardware unit compares information from previous frames to generate the new one? All I am seeing is that it is a spatial upscale with just a different method (from where the sample patterns come from).
This is my basic MS Paint understanding. Correct me if this is off base:
Resolution is about information in the image.
In traditional upscaling, you have a unique 'x' amount of information in the source frame. The output still has the same amount of information.
In these newer techniques, you're taking your 'x' amount of information from the current frame, and some portion of 'x' from prior frames (in the above, the orange pixels are meant to represent usable samples in the current frame, black ones which are not valid anymore or cannot be mashed into being valid).
So basically with the newer techniques, you should have more unique information in the final image compared to a traditional upscale, and so should get closer to your target (native) amount of unique info in the final frame.
Indeed in the perfect case of a still scene you could reach the same amount of information in the final image as a native render. tldr - the diff between a reference native frame and a frame using these kinds of techniques should be size-ably smaller than in a simple traditional upscale, the size of the diff scaling with the motion speed in the scene.
Resolution is about information in the image.
In traditional upscaling, you have a unique 'x' amount of information in the source frame. The output still has the same amount of information.
In these newer techniques, you're taking your 'x' amount of information from the current frame, and some portion of 'x' from prior frames (in the above, the orange pixels are meant to represent usable samples in the current frame, black ones which are not valid anymore or cannot be mashed into being valid).
So basically with the newer techniques, you should have more unique information in the final image compared to a traditional upscale, and so should get closer to your target (native) amount of unique info in the final frame.
Indeed in the perfect case of a still scene you could reach the same amount of information in the final image as a native render. tldr - the diff between a reference native frame and a frame using these kinds of techniques should be size-ably smaller than in a simple traditional upscale, the size of the diff scaling with the motion speed in the scene.
QB ended up having garbage IQ so let's hope not I guess?
It reminds me of how people wanted to tell us how 4k fits supernatural to 1080p in terms of scaling: just double every pixel.
Is that your Last Word on the subject?
Truth be told, on a 4K TV, these techniques are pretty great. I can only run older PC games in native 4K on my PC, for instance - this type of rendering would be an amazing addition for PC gamers as well if more titles used similar techniques.
Dictator93
Member
You have a good understanding of how a temporal technique could leverage offset / jittered / alternated samples to reconstruct a higher spatial resolution image over a step of time (over multiple / 2 frames), but the thing is - and as I mention above - I have yet to see any description from Sony that describes that this is what the hardware scaler is doing.
Rather, I have just seen that hardware scaler does math within one frame to fill in the missing blanks of 1 frame, not with information from multiple / the previous.
And if one were to presume the hardware scaler somehow does fill in the blank checkered areas by leveraging information from the previous frame, doesn't that require a lot of assumptions about engine design and the necessary information that needs to be fed to the hardware scaler?
You would need to feed it motion vector information so it knows which pixels to toss out, and how to weight, right?
QBs, Killzone's, or Rainbow Six's way of generating a higher resolution from a lower resolution image is different from this from everything I have seen. I will post this again and again until I see otherwise, but no where has there been concrete informatoin that the hardware scaler is using information from previous frames to upscale.
Let's remember that everything is relative. In a worst case when we're talking about applying these techniques toward 4K, from base per-frame resolutions in excess of 2x1080p, this will be a substantial bump over current 1080p IQ levels.
Perceptually these techniques may even fair better vs the native target res than the same applied to a target of 1080p from sub-1080p - these techniques should do better with more source resolution to work with.
I don't think we've any idea of what hardware supports are in there, except that there are seemingly some.
Cerny did cite temporal sampling in his spiel about more efficient ways of getting to '4K'.
I think the actual applied techniques will vary from title to title, but temporal sampling with checkerboard patterns certainly seems to have been namechecked.
Now whether that entire process has been baked into custom hardware, taking the inputs you describe and handling the whole thing outside of the shader pipeline...I've no idea. It's possible, it's possible they've baked certain approaches wholesale into fixed logic such that the whole thing can be treated as a post-process from standard buffers. But we'll have to wait for details I guess. I think it's more possible that some discrete common parts of typical approaches to this kind of processing have been brought out into hardware, and then between the shader pipeline and this hardware, the cost of doing it is reduced vs doing everything in shaders. But that it'll still be fairly custom from the devs' end about exactly what is done.
This is my basic MS Paint understanding. Correct me if this is off base:
Resolution is about information in the image.
In traditional upscaling, you have a unique 'x' amount of information in the source frame. The output still has the same amount of information.
In these newer techniques, you're taking your 'x' amount of information from the current frame, and some portion of 'x' from prior frames (in the above, the orange pixels are meant to represent usable samples in the current frame, black ones which are not valid anymore or cannot be mashed into being valid).
So basically with the newer techniques, you should have more unique information in the final image compared to a traditional upscale, and so should get closer to your target (native) amount of unique info in the final frame.
Indeed in the perfect case of a still scene you could reach the same amount of information in the final image as a native render. tldr - the diff between a reference native frame and a frame using these kinds of techniques should be size-ably smaller than in a simple traditional upscale, the size of the diff scaling with the motion speed in the scene.
You know you want to say "Uprendered"
Devs are likely going to be using temporal reprojection in this manner, but it's not what the DF article is talking about it. They're describing a specific way of taking a checkerboard of 2x2 squares - that is, the individual squares on the checkerboard are 2 pixels wide, rather than the checkerboard pattern skipping every other pixel - and extrapolating the data from the rendered 2x2 squares into the unrendered 2x2 squares, independent of any temporal sampling.
I'll have to go read the article fully... intuitively I'm not fully getting how that would work out. I shall read the article though!