EasyRaider
Regular
Gollum said:(...) Just my opinion of course...
And my opinion too
AA or AF?
- Thread starter Shad66
- Start date
And my opinion too
Freak'n Big Panda
Regular
Personally I perfer AF to AA. I play all my games @ 1600x1200 with 8X AF. But as mentioned before it depend on personal prefrence.
Supersampling > *
Especially 3dfx's solution
And 3dfx had a pretty fun method of combining AA and AF such that if AA is on, AF has negligible (<1FPS) performance hit... (works on V4/5 too, with Aquo's ICD) Up to 4x Aniso per AA sample, so 4x AA results in instant free 16x Aniso on everything
Especially 3dfx's solution
And 3dfx had a pretty fun method of combining AA and AF such that if AA is on, AF has negligible (<1FPS) performance hit... (works on V4/5 too, with Aquo's ICD)
Up to 4x Aniso per AA sample, so 4x AA results in instant free 16x Aniso on everything Tagrineth said:Supersampling > *
Especially 3dfx's solution
And 3dfx had a pretty fun method of combining AA and AF such that if AA is on, AF has negligible (<1FPS) performance hit... (works on V4/5 too, with Aquo's ICD)
pfft for a card that couldnt do proper trilinear without a massive perfromance hit, you really beleive it has a hardware supported AF routine? Dont beleive those LOD bias adjustments = true AF.
Tagrineth said:Supersampling > *
Especially 3dfx's solution
And 3dfx had a pretty fun method of combining AA and AF such that if AA is on, AF has negligible (<1FPS) performance hit... (works on V4/5 too, with Aquo's ICD)
That's not what I recall any of my past experiences with Supersampling, not 3dfx's RGSS or any other OGSS alternative. SSAA so far was never better than 2x Level anisotropic (or 16-tap if you prefer).
I had Acquoes exposing ColorMiplevels in SS:SE and I for one couldn't detect anything higher than 2x Level equivalent texture filtering; you don't even need to hack the drivers for that. It's already present on default with normal Supersampling.
I'd say that with any modern video card (GeForce4 Ti, Radeon 9700), there's essentially no reason not to run with both enabled. This will become even more the case in the future. I almost always run with 8-degree aniso and 2x MSAA on my Ti 4200.
However, when turned off, AA is easier to spot for me than aniso.
However, when turned off, AA is easier to spot for me than aniso.
Randell said:
It CAN do Trilinear sampling, thank you. Even 3dfx exposed true trilinear in Direct3D.
The problem is in the poor, neglected GLide driver... in letting GLide die, they basically made the VSA-100 drivers wrap to a Voodoo3 driver.
Oh, and 3dfx themselves would've used 3dhq's current method of AF... take the AA subsamples and re-jitter them for new texture 'samples'. Granted it isn't true AF, and Quake3 is cheap because it uses true MIP levels (not just resampled... they actually manually made the lower MIP levels, which is why Kyro cards look like Bilinear when they use Tri + S3TC in Quake-engine games), but it works pretty well from what I've seen of it. But because of the way it works (not taking 'real' new samples), coloured MIP levels probably won't pick it up as well as human eyes. ;P
I beg your pardon? It's technically true trilinear wether TC is enabled or not and the difference is immediately detectable, between plain bilinear. It just doesn't show the smooth colour transitions trilinear should when you expose colormiplevels in q3a.
Despite the very old algorithm used and dog slow performance, it was capable of true 2xLevel trilinear anisotropic filtering too.
In what exactly? Spectre and upwards were to be capable of adaptive 128-tap anisotropic filtering. Care to elaborate what purpose it would served if you get already 2xLevel aniso equivalent texture filtering via Supersampling? (assuming you meant for V5 class cards).
I never really bought that supposed "Feline" implementation, but that's just me.
EasyRaider
Regular
Chalnoth said:
But 8-degree aniso is slooooow (with GF4). I only use it when I've got fillrate to burn. Any degree at all is too much in UT2003.
Voodoo's (3+ ?) could do trilinear without multitexturing.
Ailuros said:
It's "true" trilinear in the sense that it does blend MIP levels, however the K2 creates those MIP levels on the fly rather than using the ones provided.
Ailuros said:
Yes, they were to be capable of true Aniso, but they would've also provided effectively free Aniso with multisampled AA via re-jittered subsamples.
Not the way how I recall it. Make that under selected occassions and I might agree. With two textures, 2xRGMS + 32-tap anisotropic, the performance cost would have been minimal and 2xRGMS for free.
All of that wonderful stuff on a dual chip Spectre with 12.8GB/s bandwidth and a not lower than 500$ pricetag for it's time.
Can we get back to reality and stop dealing with relics and vaporware?
Chalnoth,
You can optimize just the base textures (ie texturing stage 0), and leave the remaining three on Levelx1. I don't care if someone might call it illegitimate or a "cheat", it works for me as a good balance between IQ and performance.
I run it in 1024x768x32 with 4xOGMS in Convolution Mode and optimized 4xLevel aniso.
Ailuros said:
Yes, and I'd do that, if I played any Direct3D games commonly.
The only problem is that turning off the first stage is what gives the largest performance improvement.
The reason for this is simple, but unfortunate: The second texture stage will not go through if the first stage has any sort of anisotropic filtering enabled. I got this from 3DMark2k1 fillrate tests, where disabling texture stages 1,3 does nothing to affect the tests, whereas disabling stages 0,2 results in fillrates identical to those without anisotropic filtering.
I confirmed that anisotropic was still used when stages 0,2 were enabled through some quick benchmarks: it's still slower with just stages 1,3 having aniso enabled.
Hopefully the NV30 will fix this problem, as it should be as simple as having the hardware to compute a different level of anisotropic for each texture pipeline. But, even if it's not fixed, and it is an 8x2 pipeline, I don't see how it will underperform the R9700 with aniso enabled (provided similar computational power per pixel pipeline).
Update:
It kind of makes me wonder, however, whether or not ATI does something similar. That is, whether or not they disable anisotropic filtering entirely for certain texture stages. I believe I heard something alluding to that a couple of months ago, but nothing concrete.
OpenGL guy
Veteran
Some us some evidence, please, instead of making us rely on what you believe you heard.Chalnoth said:
Would you've been confident if he addressed angles then OpenGlguy?
Chalnoth,
I replied with UT2003 in mind. Why do have to to have two texturing stages simultaneously optimized? What I meant was:
Level 8x aniso
where:
texturing stage 0 (unchecked, thus using 8x aniso)
texturing stage 1, 2, 3 (checked/limited to Level 1x)
(note: no disabled trilinear)
As far as next generation products concerns my tip goes rather in the direction of alternative algorithms, not necessarily TexMax and possibly truly adaptive, as in varying samples of aniso depending on the surfaces.
Slightly OT:
Can someone give me a fair explanation (if Mr.Vogel is reading any input would be welcome) why openGL with aniso vs D3D is like night and day in favour of openGL in the IQ department?
Chalnoth,
I replied with UT2003 in mind. Why do have to to have two texturing stages simultaneously optimized? What I meant was:
Level 8x aniso
where:
texturing stage 0 (unchecked, thus using 8x aniso)
texturing stage 1, 2, 3 (checked/limited to Level 1x)
(note: no disabled trilinear)
As far as next generation products concerns my tip goes rather in the direction of alternative algorithms, not necessarily TexMax and possibly truly adaptive, as in varying samples of aniso depending on the surfaces.
Slightly OT:
Can someone give me a fair explanation (if Mr.Vogel is reading any input would be welcome) why openGL with aniso vs D3D is like night and day in favour of openGL in the IQ department?
Similar threads
