Has Nvidia Found a Way to get a Free Pass in AA comparrisons
- Thread starter Hellbinder
- Start date
Why yes, it is.kyleb said:
Which is why I'm not agreeing with him.Joe DeFuria
Legend
Lol....
Sorry Russ, looks like you (and I) were in fact wrong. I should've known better than to try and stick my neck out for ya.
Sorry Russ, looks like you (and I) were in fact wrong. I should've known better than to try and stick my neck out for ya.
Hrm. Technically yes, as there's some combinational logic in between the framebuffer and the RAMDAC, but I think the concept is the same.
Regardless, I expect that whatever method 3dfx used is nearly identical to what the current NVIDIA method is.
Regardless, I expect that whatever method 3dfx used is nearly identical to what the current NVIDIA method is.
Russ and Joe. You guys were essentially right because I don't believe the output of that special logic goes back to the frame buffer. Also, take into consideration that Nvidia didn't implement this until they hired 3dfx employees. I would guess the implementation is nearly identical.
3dcgi said:
Right, which addresses the post filtering, not the T-Buffer FSAA. That's the difference. 3dfx *never* used the post filter for FSAA--the suggestion is ludicrous. The V3's 16-22-bit mode had nothing whatever to do with FSAA of any sort but was the the V3's "answer" to a display mode near 24-bits but without the performance hit. At the time the V3 shipped the TNT was nVidia's prime 3D chip, which did 24-bit 3D, albeit at a slide-show pace.
Joe DeFuria said:Lol....
Sorry Russ, looks like you (and I) were in fact wrong. I should've known better than to try and stick my neck out for ya.
You know, the funny thing about all of this is that from the moment I heard a long while ago that the nv30's 2x and QC FSAA modes weren't showing up on screen shots, I *knew* they were doing something with the post filter--and in fact I said so on more than one post here at the time. What I found exceedingly strange about it was that instead of talking about it as 3dfx copiously did when it pioneered the technique for the V3 (which had super image quality at the time as a result), nVidia officially clamped a "trade secret" lid over it and refused to discuss the generalities. And then I remembered 3dfx had never used the post filter blending for FSAA--that indeed the V3 never did any type of FSAA at all--and it suddenly made sense why nVidia wouldn't want to talk about it. *chuckle*
Walt, I think you're off in your own tangent here, confused by semantics.
3dfx used some magical post filter thing on the V3 on a single frame buffer to undither the framebuffer and get "22bit" quality. I'm not talking about this, and neither is anybody but you.
Both the NVIDIA and the 3dfx method of AA center around combining samples from separate frame buffers, mixing them and directly sending them them to the RAMDAC. 3dfx called it "t-buffer", NVIDIA is just calling it "super secret"
3dfx used some magical post filter thing on the V3 on a single frame buffer to undither the framebuffer and get "22bit" quality. I'm not talking about this, and neither is anybody but you.
Both the NVIDIA and the 3dfx method of AA center around combining samples from separate frame buffers, mixing them and directly sending them them to the RAMDAC. 3dfx called it "t-buffer", NVIDIA is just calling it "super secret"
RussSchultz said:Why yes, it is.kyleb said:
well i happen to have a voodoo5 with the shipping drivers cd laying right here on my shelf which i could test. granted i am afraid that might require installing an antiquated operating system and is really more effort than i would care to put forth on this issue at this point. however, if no one else is going to provide any evidence to support the claim one way or another i am sure i could find the time to do it. however, without such evidence i can not see how you find rolling your eyes at me for my opinion to be justified.
And what exactly would you test with your V5 and driver CD to prove either one of us right or wrong? :?
nVidia isn't exactly calling it, "super secret." They've announced that this is what is done (subsample combination done at scanout), they just haven't given it any name. It's been done since the GeForce4.
However, I think the technique needs to be laid to rest. It makes lots of sense for 2x AA, but 4x and higher modes the memory usage penalty really kills the bandwidth advantage (which gets smaller and smaller as the level of FSAA increases).
However, I think the technique needs to be laid to rest. It makes lots of sense for 2x AA, but 4x and higher modes the memory usage penalty really kills the bandwidth advantage (which gets smaller and smaller as the level of FSAA increases).
Joe DeFuria
Legend
Walt,
That page and quote doesn't address the 22 bit post filter, It addresses FSAA: (The image depicts a 2 chip VSA-100, creating a 4X AA image).
You see the box that says "Combine the pixels of the buffers to form the final Anti-Aliased image?"
There is "speical circuitry", OUTSIDE of the physical "T-Buffers", that takes the contents of each buffer, and combines them to form the AA'd image.
There are basically 2 levels of "blending" operations.
1) Each CHIP creates 2 samples each. On-chip (in the t-buffer), the 2 samples are combined to create essentially a 2X AA image.
2) Each chip sends it's 2X AA image to the "specical circuitry", which does the final blend to form the 4X AA image. The "final image" does not appear to actually reside in memory. (Or by the picture, it's possible that the master VSA chip reads the 2 individual samples from it's own chip, and the 2X AA image from the second chip, and 'combines' the three of them, though that doesn't seem logical.)
Hence, just "grabbing" frame-buffer" memory will not give you the correct result. You'll likely get whatever is sitting in the "master chip" frame buffer, which in this case, is either a pseudo 2X AA image, or just one sample.
That page and quote doesn't address the 22 bit post filter, It addresses FSAA: (The image depicts a 2 chip VSA-100, creating a 4X AA image).
You see the box that says "Combine the pixels of the buffers to form the final Anti-Aliased image?"
There is "speical circuitry", OUTSIDE of the physical "T-Buffers", that takes the contents of each buffer, and combines them to form the AA'd image.
There are basically 2 levels of "blending" operations.
1) Each CHIP creates 2 samples each. On-chip (in the t-buffer), the 2 samples are combined to create essentially a 2X AA image.
2) Each chip sends it's 2X AA image to the "specical circuitry", which does the final blend to form the 4X AA image. The "final image" does not appear to actually reside in memory. (Or by the picture, it's possible that the master VSA chip reads the 2 individual samples from it's own chip, and the 2X AA image from the second chip, and 'combines' the three of them, though that doesn't seem logical.)
Hence, just "grabbing" frame-buffer" memory will not give you the correct result. You'll likely get whatever is sitting in the "master chip" frame buffer, which in this case, is either a pseudo 2X AA image, or just one sample.
Chalnoth,
why would it make a difference? 4x or 2x, you're still avoiding the downsampling. It should scale linearly.
Kyleb,
Read a little bit for comprehension. The reason we're talking about whether or not it is properly recorded is because of the method used. 3dfx had the same problems when the V5 came out and they addressed it with a hypersnap plugin, NVIDIA can do (has done?) the exact same thing because their method is essentially identical.
why would it make a difference? 4x or 2x, you're still avoiding the downsampling. It should scale linearly.
Kyleb,
Read a little bit for comprehension. The reason we're talking about whether or not it is properly recorded is because of the method used. 3dfx had the same problems when the V5 came out and they addressed it with a hypersnap plugin, NVIDIA can do (has done?) the exact same thing because their method is essentially identical.
RussSchultz said:
Russ, I think you've totally misunderstood the issue, at least as I see it. When nVidia first began emulating 3dfx's T-Buffer FSAA in its GFx products, it used supersampling--rendered to a high resolution and scaled it down. It was very slow, and looked not 1/10th as good as what 3dfx did with the T-buffer. Indeed, 3dfx did *hardware jittering* in the V5, that is, they took more than one frame *at the same resolution* and hardware-jittered the result in the VSA-100 hardware. They blended in the T-Buffer, which was also in the the V5 hardware. nVidia never, ever had anything similar (and the results showed that plainly, IMO.)
The post filter blending is something entirely different and was used by 3dfx to essentially upsample the display output in the V3 to a psuedo 22-bits of accuracy *in the post filter*--it wasn't real 22-bit accuracy, but close, and allowed 3dfx's 16-bit display (non-FSAA) to look much better while at the same time minimizing the performance hit. It was far faster than TNT2 24-bit output, and looked 90% as good. The only place this was carried through in the V5 was in the V5's 16-bit output mode, where it could be selected for as a display output option. In the V5, you could have 16-bits, 16/22 bits, or 32-bits, your choice. It's possible the post filter may have operated on the 16-bit FSAA result of the T-buffer, when the 16/22-bit mode was selected as the display output, but in 32-bits the post filter was not employed at all for any reason (since it wasn't needed, obviously.) Hence the post filter in 3dfx products was never employed primarily for FSAA--and in 32-bit V5 output the post filter wasn't used at all, for either FSAA or normal display.
The difference here, and this is just a theory based on the fact that normal screen shot software wouldn't catch nv30's 2x and QC FSAA modes, just like screen shot software at the time absolutely butchered the V3's normal non-FSAA output until it was written to grab the post filter blending as well, is that while nv30 seems to employ the technique as the primary technology for its 2x and QC modes, 3dfx never used it for FSAA. Since nVidia has decided to call it a "trade secret", *chuckle*, I guess we'll never know.
The strange part about it is that if nVidia had said "We found we could use post-filter blending and get as good FSAA results at 2x an QC FSAA that we'd get with standard FSAA that we decided to use it instead," I not only wouldn't have cared--I'd have thought it was pretty clever (as long as the results panned out.) But the recurring theme from them seems to be to hide and obscure what they are doing--which just doesn't "trip my trigger" so to speak...
4x is slower than 2x. If you're doing subsample blending at the scanout stage, you need to filter for every screen refresh. If you do the blending in the frame buffer, you only need to filter per frame.RussSchultz said:
If the 'frames per refresh' ratio is lower than (N-1) / (N+1) where N is the number of samples, scanout downfiltering will need more bandwidth than frame buffer filtering.
Walt, I'm sorry. I'm just going to have to disagree with you in the end.
I know what the post filter for the V3 is, and it is exactly as you describe.
I know what the t-buffers are, but they're not what you describe. They're the individual buffers that got jittered to, not the combining mechanism.
The combining mechanism for the V5 FSAA was went beyond the post filter of the V3, but it was still a "post processing step"(i.e. the results never got written back to the frame buffers). Instead of undithering from one framebuffer like the V3 did, it takes samples from the multiple t-buffers (frame buffers) and combines them to form a single sample that it then sends to the RAMDAC.
Both the current NVIDIA super sampling mechanism and 3dfx super sampling mechanism are nearly identical in concept, especially in the post processing step. (excepting the NVIDIA ones doesn't seem to do anything but ordered grid)
I've restated the same thing 3 times now, and so have you. So we'll just have to disagree.
I know what the post filter for the V3 is, and it is exactly as you describe.
I know what the t-buffers are, but they're not what you describe. They're the individual buffers that got jittered to, not the combining mechanism.
The combining mechanism for the V5 FSAA was went beyond the post filter of the V3, but it was still a "post processing step"(i.e. the results never got written back to the frame buffers). Instead of undithering from one framebuffer like the V3 did, it takes samples from the multiple t-buffers (frame buffers) and combines them to form a single sample that it then sends to the RAMDAC.
Both the current NVIDIA super sampling mechanism and 3dfx super sampling mechanism are nearly identical in concept, especially in the post processing step. (excepting the NVIDIA ones doesn't seem to do anything but ordered grid)
I've restated the same thing 3 times now, and so have you. So we'll just have to disagree.
WaltC, you're only talking about the 3dfx post filter here, ie. the one used do eliminate dithering artifacts occurring with 16bit rendering (why it's '22 bit' is simple: it takes 4 pixels in RGB-565 format and adds them, which results in a RGB-787 color value. This combination is conditional however and depends on some contrast threshold)
But there is another 'post filter' used to combine the AA samples for the final output to the screen. And that's why it was also difficult to grab an AAed screenshot on a V4/V5 at first.
But there is another 'post filter' used to combine the AA samples for the final output to the screen. And that's why it was also difficult to grab an AAed screenshot on a V4/V5 at first.