Ripping off the veil: The mysterious PS3 hardware scaler exposed

The whole thing does have a sort of Hardy Boys-ish "The Case of the Curious Hardware Scaler" feel to it, doesn't it? I suppose this puts Vysez in the Joe Hardy/Shaun Cassidy role?

The next time we're wrong won't be the first time, but we're feeling pretty good there's a real hw scaler in there. I suspect at some point we'll be able to revisit this one for a post-mortem, and I'm sure we all look forward to that.

 

The GC demonstrates quite awful banding, all the time... but that's cause even the back buffer can be 24 bits..
And also how many modern GPUs that render to a 24 bit render target do you know?

 

Seeing that the HDCP 1.3 revision was finalized on Dec. 21 2006 and the timing of this update someone might want to look here to see if they can find a possible answer.

 

The article says:

As for the rest, there's certainly a lot of mystery about it all, but re. Sony PR or whatever..I don't think there's any point second-guessing their thoughts on this, since their approach seems to be to acknowledge no issues and just quietly fix them. Unless you're Watch Impress and have access to Ken Kutaragi..he seems more than happy to discuss anything and everything with them in finer technical detail, but that asides..

My own theory, or one of them, is that Sony is cautious about what it enables or how it enables it because they have an eye on cost reduction. There may be hardware in there that allows for more, but they may have an eye on a point in time where it won't be there, or can be substituted for something cheaper (and perhaps less capable). So they're being careful about what they commit too. The other commonly presented explanation is software issues. I haven't a clue, it is very mysterious.

 

As you know, the banding is due to the 6:6:6:6 mode, not the YUV conversion. And if you look closely, I was talking of the the front-buffer (in main-memory), not the back-buffer or any render targets in the embedded RAM...

 

Yeah, in fact I wrote it, if you look closely too now, going back to the 24 bit buffer, what would they buy you? maybe an extra meg of free mem, that's it (if we just talk about the front buffer)

 

First excuse me for my English…
Is it this up-scaling to the 1920x1080 resolution a real up-scaling processing ?
For a real up-scaling with a good interpolation, every pixel (almost all of them) of final image at the resolution of 1920x1080 have to be recalculated with an interpolation process, because the two sampling matrices don’t coincide. Is this consideration correct ?

In this case with only some columns of the final image are “interpolated”, but how ?
What kind of estimation is done ?
Is every single pixel of column calculated as the average of the two nearest pixel of the some row ?

 

Tipically in a scaling process all the final pixels need to be re-computed from the original image.

It's a resampling problem which is usually addressed via some kind of (pixels) interpolation.
How this interpolation operates and how many pixels this process needs to interpolate to generate a 'final' pixel is an implementation detail and high quality resampling filters require to interpolate many pixels to just generate one final output pixel.

 

So it is correct say that in this case only a discret columns of pixel are recalculate instead of all image, isn't it ?
Can the developers choose this implementation detail or the scaler of PS3 ( if there is one...) do it all alone whitout flexibility?
Can you tell me what this scaling do in detail ?

There is a real hardware scaler in PS3 or this implementation in the new devkit is a short cut to repair the problem of PS3 with something ?
Are the results comparable with a classic hardware scaling solution ?

 

What is 'this case'?

Umh..

Anything else you wanna ask?

 

"this case" => PS3

I know it. I'm a snoop ("curiosone" in italian, i don't know if the traslation is correct...)

 

I see, but I can't really answer to this question.

With a nickname like that I had no doubts about your nationality

 

Would it not have made more sense to scale to 1920x540? This is basically 1080i resolution anyway!

 

Is it significant in the context of this article that the Firmware 1.5 backwards compatibility fix was described as resolving a "horizontal scaling issue"?

 

That's not true. Sometimes it's dithering.

Now is there a possibility that you guys will have enough time to add "hardware" scaler support to HS? I realize supporting the unorthodox aspect ratios and giving up some fillrate, bandwidth, and memory will take away from valuable coding and testing time. If support won't be there at the game's launch, are you thinking about possibly working on a patch?

 

Just read up on this exposed "scaling" ability right now, and as lame as it is, it should (barely) do the job for Sony.

The biggest problem was that 1080i-only TVs (99% are RP and direct-view CRTs) would have to settle for 480p in games that output 720p only. Asking developers to software scale would require quite a bit of space. Using 960x1080 would let devs use almost the same space, and the 1080i-only TV's don't have much more than 1000 lines of horizontal resolution anyway (sometimes less). This solution allows devs to accomodate these users with minimal effort, assuming pixel alignment doesn't impact their game.

However, saying 960x1080 is as good as or better than 1280x720 in general is wrong. If you draw a circle without AA, for example, 960x1080 will show horrible aliasing on the left and right, and little aliasing on the top and bottom. 1280x720, on the other hand, will show medium aliasing everywhere, which is much more preferable.

And I agree that calling this a hardware scalar is being very generous. Horizontal scaling can be done with the RAMDAC, and that's probably what they're doing because most 1080i-only TVs will only accept component. It's the vertical resizing and ability to create new lines that really defines a scalar.

 

I am not sure about that. First of all, any decent scaler will make use of interpolation (or anti-aliasing) which will hide aliasing.

The problem is that scaled image won't have the sharpness of a native 1920x1080 image.

Fortunately, for those of us with two eyes, there is already a superimposition of horizontally shifted images and resulting _horizontal_ low pass filtering may very well hide horizontal sharpness, hence the lack of information won't matter much.

That's not the case for vertical scaling.

Obviously it depends on screen size and viewing distance (as well as focus).

 

The best horizontal scaler in the world couldn't fix that because it has no information about other lines. Do you even know what anti-aliasing is?

Huh? Are you cross eyed?

Sorry if I'm sounding harsh, but this post of yours is one of the worst cases I've ever seen here of using BS to try looking smart.

 

Aliasing means introducing artifacts (unwanted frequencies) while sampling. It is generally used with undersampling but can also happen when oversampling (0-order hold or whatever).

If you scale horizontally, you won't introduce vertical artifacts because you don't touch vertical frequencies. Hence you don't need other lines.

Just close one of your eyes when you are looking on the screen. Then open that one and close the other. Let me know if you don't notice any horizontal shifting.

Don't be sorry on my account, but I do recommend a course on signal processing.

 

Aliasing does not occur while sampling the framebuffer. It occurs when the rasterizer samples points in a polygon. Without anti-aliasing enabled, you get a single boolean answer to whether the pixel is in the polygon or not based on the pixel centre.

Do you know what the spatial frequency of a polygon edge is? Infinite. Of course a display does not have infinite resolution, so that's why one sample per pixel is fine for a first approximation. However, AA (multiple samples per pixel) is needed to get closer to the ideal picture.

You should realize, however, that AA is about sampling during the rasterization stage. A scaler simply transmits the framebuffer image to the TV. There is no magic reconstruction filter to obtain information lost during rasterization.

The artifacts are due to the rendering, not the framebuffer output. I'm talking about 960 pixels having less information than 1280 pixels. Scalers don't introduce aliasing unless they're broken (like the PS2 emulation in PS3 before the patch).

That's completely irrelevant, because when you actually use both of your eyes they're aligned fine. The question boils down to this: Do you have reduced ability to resolve vertical edges and features with both eyes? Do circular dots appear blurred horizontally so as to appear elliptical?

The answer is no for anyone with properly functioning eyes, so your point is meaningless. There is no horizonal LPF imposed by stereoscopic vision.

Buddy, I'm doing my PhD at Caltech in EE. I'm not deficient in knowledge about signal processing. Unfortunately, I've seen plenty of people misuse their knowledge of signal processing in 3D graphics. I've seen them think they can apply Nyquist's theorem to AA, but their logic is flawed.

You are another example.