R420 IQ Comparison

I've posted a request to do so in another thread, and dig seems up to it. I *think* he has a 9600 in one of his systems, and he has Mafia....

 

The problem is that M$'s comments had nothing to do with a look at R4x0 whatever, especially in a comparative sense with nV40. Therefore, whomever assumes something relative to an nV40/R4x0 comparison may not do that on the basis of what M$ said, since M$ did not say that "R4x0 produces an image inferior to nV40." That's not even close to what M$ said, which was that nV40 produced a better image than the rasterizer, because M$ had not yet updated the rasterizer to the capabilities of the "newer hardware." It must not be assumed that R4x0 is any less "new" than nV40, therefore, the most probable case is that the rasterizer hasn't been updated to properly serve either architecture at the present time.

As well, the THG article was simply trying to to put words in M$'s mouth relative to the context of the THG article, instead of in the context from which M$'s remarks were actually made (IE, the context for M$'s remarks had nothing to do with IQ comparisons between R4x0 and nV40, but were strictly made from the standpoint of the 60.72's rendering versus the DX9 rasterizer output.) R4x0 isn't in there anywhere, as per the context of M$ remarks as quoted by the unkown M$ employee who is said to have made them.

 

Walt, you might be interested in this.

 

That is interesting... What is the prominent green square in the nVidia image that is absent from the rast?

 

It looks like the driver is incorrectly rendering the light in the middle of the green square.

 

Maybe that’s NV’s version of adaptive-Trilinear.

 

WaltC, did you even read what Xmas said? He's talking about one specific aspect: "LOD selection algorithm". Not about the final image quality or even final texture filtering quality or anything else that is being brought up in reply.

The problem with your reply is that it has nothing do with what he said at all, except by way of ignoring it.

There is a method of LOD determination. The reference rasterizer seems to use it (accuracy of it is B). The R3xx and R4xx appear to use something of equal accuracy, if not the same exact method ( accuracy of it is C, and C=B). This has been said, and seems to be demonstrated by the ways R3xx and R4xx resemble the reference rasterizer, and doesn't seem to have evern been refuted by ATI personnel that I've seen. It also contradicts the entire premise of your reply...you haven't said anything that refutes, changes, recognizes, or addresses that at all.

This doesn't mean there is a problem with it, but it seems it happens not to be the most accurate possible representation of LOD, and that what nVidia uses is more accurate.
Is the extra accuracy important? What impact does it have on aliasing? Questions like these are important, and might or might not relate to there being issues with the presentation of LOD determination, but they are separate questions that don't happen to change the answer to the first one. Which is the one Xmas is talking about in his post.

Accurate discussion of the topic can start when your reply doesn't simply talk over things that are inconvenient.

 

Not going to get one. Hardware is always in a state of evolution, and software even more so. The API's themselves and the IHV's handling of those calls are always going to be in a state of flux as each tries to improve things in the ways they feel work best. Each will encorporate, improve, or re-interprety what advancements have come from the others, and each will keep trying to refine even the "tried and true" processes. And ALL of them are looking for unsought "breakthroughs" that might really give them an edge.

What we see now is exactly what we'll continue seeing--there's no path to follow. There's just the public kicking the players if they've gone wandering off too far. (And they have to kick REALLY hard to actually cause a reaction, too.)

 

Heh... Then I have to wonder why Xmas said he didn't understand me, since what I originally said had nothing to do with isolating an LOD selection algorithm, did it?... My original post, which he said he didn't understand, dealt with THG using an out-of-context quote attributed to an anonymous M$ employee (which THG clipped from Tech Report in a piece TR wrote about the 60.72 Forcenators.) The context of the THG article was one of comparing IQ between nV40 and R420, but the context of the M$ quote THG used in that context was one solely of comparing the 60.72 Forcenator driver output to the DX rasterizer output (and did not involve R420 in any fashion.) Thus, THG used it entirely out of context. That was all there was to understand, or misunderstand, in my original post.

What you seem to be suggesting is that Xmas didn't understand my original post because Xmas was talking about something my original post never talked about...? I mean, fine, if he wants to make a point independent of my posts--no problem. But why drag me into it?... But since he referenced my post by quote, and admitted to not understanding it, I felt compelled to reiterate what I'd said in the original post.

Which might be because he wasn't actually arguing with my original post at all, even though he thought he was? --This is all too convoluted for me, I'm afraid...

 

WaltC,

You've actually quoted a discussion of what the MS representative said, and it specified precisely that it was "level-of-detail computation for isotropic filtering" that they were talking about.

If you continue to repeat your mistake of thinking "what I originally said had nothing to do with isolating an LOD selection algorithm" when talking about what the MS representative said, that is your fault.

 

Come on, D, don't tell me I have to repeat what I originally said a third time. My point was that *whatever* the unnamed M$ source said in the quote, it had nothing to do with a comment on comparative IQ between R420 and nV40, which was the subject of the THG article, and which used that particular quote completely out of context.

 

Except that it does...

 

But it doesn't with AF since its only for isotropic filtering.

 

No, you don't have to repeat it, I already addressed it in my initial reply to you. I also blatantly illustrated an error you continue to make, and you continue to pretend the mistake doesn't exist instead of going through the inconvenient task of trying to make your point without that glaring mistake.

But it has everything to do with what Xmas said to you, which you specifically just said it did not.

What you're demonstrating is that, as far as you're concerned, you're having decided you were right is more important than the discussion, so ignoring mistakes or inconveniences to say that to others is OK, as long as you can continue to say you were right at the end. Or, I suppose, you're reading and replying very sloppily, over and over.

• You can make sense for your point by acknowledging the mistake, and then making your point while recognizing it, which is what you continue to fail to do because rhetoric seems much easier for you.
• You can't make sense by using rhetoric to repeat yourself without ever dealing with that mistake, or conversing based on snipping out commentary that makes your mistake blatantly obvious like You've actually quoted a discussion of what the MS representative said, and it specified precisely that it was "level-of-detail computation for isotropic filtering" that they were talking about when it is inconvenentily close to you saying Then I have to wonder why Xmas said he didn't understand me, since what I originally said had nothing to do with isolating an LOD selection algorithm, did it?

...

Which, strangely enough, brings us back to where I was in my initial reply (well, if you read it, at least), if you hadn't gone off on the useless tangent of trying to avoid admitting a mistake by talking over a simple point repeatedly.

 

Very simple. That green square is the glow effect of the light in the middle of the square. You can see that it's translucent, so alpha blending is used. The base LOD is not colored, so you don't see a colored square on the ATI shot, only the glow (refrast doesn't show the glow at all).

Right, MS did not say that.

Wrong. MS stated that the algorithm for level-of-detail computation for isotropic filtering used in NV40 produces a better result than the one used in the DX9 reference rasterizer.

And, as it happens, ATI uses a level-of-detail computation algorithm in R420 that produces virtually the same result as the one used in the refrast.

Wrong. They said the API is evolving, as is hardware, and they will continue to improve the API, and with it, the reference rasterizer. So the refrast is not perfect, and being close to it doesn't imply highest quality rendering. The reference rasterizer doesn't have to "serve an architecture". It's completely independent of the hardware.

The context of this page of the THG article is a comparison of isotropic filtering of R420, NV40 and the reference rasterizer. They have screenshots showing the result of the LOD selecion algorithm. R420 and the reference rasterizer are virtually identical in this point. So this statement from MS is not at all used out of context.