Nvidia or AMD AF Filtering broken? (was ATI 5870 AF filtering broken?)

Discussion in '3D Hardware, Software & Output Devices' started by Tamlin, Jun 29, 2010.

  1. gamervivek

    Regular Newcomer

    Joined:
    Sep 13, 2008
    Messages:
    732
    Likes Received:
    223
    Location:
    india

    http://www.tomshardware.co.uk/geforce-gtx-570-gf110-performance,review-32062-5.html

    meh, much ado about nothing.:mad:
     
  2. Alexko

    Veteran Subscriber

    Joined:
    Aug 31, 2009
    Messages:
    4,514
    Likes Received:
    934
  3. Xenus

    Veteran

    Joined:
    Nov 2, 2004
    Messages:
    1,316
    Likes Received:
    6
    Location:
    Ohio
    Indeed much ado about nothing is usually someone makes a big stink it's fixed in a few days people have been complaining about this for quite a while now.
     
  4. Silent_Buddha

    Legend

    Joined:
    Mar 13, 2007
    Messages:
    16,987
    Likes Received:
    6,237
    Not a bug fix. The quote seems to imply that AMD will sacrifice texture sharpness by modifying LOD in order to match LOD and texture quality on Nvidia cards.

    Regards,
    SB
     
  5. gamervivek

    Regular Newcomer

    Joined:
    Sep 13, 2008
    Messages:
    732
    Likes Received:
    223
    Location:
    india
    I think it's more of a driver thing. nvidia's numerous release beta drivers are no angels either.

    for instance 9800gtx didn't show the same image as gtx 280 in the 2nd comparison here as it does in the 1st one:

    http://www.techenclave.com/graphic-cards/whose-iq-do-you-prefer-ati-125145-4.html#post998496

    I don't know about others, but shimmering is like a bonus game effect for me. Razor sharp textures over blurry colours anyday.

    and searching for "shimmering ati" on yt gave me this
    http://www.youtube.com/watch?v=JVjM_vrsBO8

    http://www.youtube.com/watch?v=BObbmHqQiVM

    :roll:

    I wonder how many people who have gone on to 68xx from nvidia noticed the shimmering only because they read about it, considering that I don't even remember how bad games used to look without AA and AF on my nvidia card.
     
  6. chavvdarrr

    Veteran

    Joined:
    Feb 25, 2003
    Messages:
    1,165
    Likes Received:
    34
    Location:
    Sofia, BG
    Hopefully they'll make it as switch accessible to user, this way it will be win-win.
     
  7. DavidGraham

    Veteran

    Joined:
    Dec 22, 2009
    Messages:
    3,257
    Likes Received:
    3,468
  8. no-X

    Veteran

    Joined:
    May 28, 2005
    Messages:
    2,334
    Likes Received:
    291
    mipmap's slider was always at high-quality by default, wasn't it?
     
  9. Silent_Buddha

    Legend

    Joined:
    Mar 13, 2007
    Messages:
    16,987
    Likes Received:
    6,237
    I'm just glad to see the left panel tree navigation back. :)

    Regards,
    SB
     
  10. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,873
    Likes Received:
    2,808
    Location:
    Finland
    But the filtering tester has quite high frequency texture, while the example (Trackmania) is far from it, and the textures on the back are clearly sharper again.
    Incidently, the areas really prone to shimmer on AMD (in 3DCenter tester anyway) are in the middle of the scene at least on the ground2 texture, with a calm area in front, between them and in the back - having softer LOD on those areas would definately help on the shimmering, and in this case, it seems like nVidia is doing just that - having softer LOD settings at specific "ranges", as the textures a bit further back seem to be sharper again
     
  11. Psycho

    Regular

    Joined:
    Jun 7, 2008
    Messages:
    745
    Likes Received:
    39
    Location:
    Copenhagen
    Ok... Just did a directed testing program to see if anyone was cheating with the number of AF samples or LOD level..
    Lets start with a picture, which should give the idea: [​IMG]
    Each block on the screen is textured in scale 1:1 in the x direction and 1:n in the y direction, n being the first number to the right. So with high enough AF level we should get as many samples per pixel as the number indicates.
    Texture is binary white noise (only containing 0 and 255) and cycling channels through the mip levels, and always using point sampling. If we try with AF=1 we get the expected result (except the usual skew):
    [​IMG]
    So the idea is that with enough pixels we should be able to record as many different shades of grey as there are AF samples per pixel (all possible sums of 0 and 255). 2nd number is the measured number of samples. As we can se we don't have quite enough pixel to get all-white and all-black every time, hence the 14 samples where it should be 16.

    Now, the above image is for HD6870 (AI texture quality doesn't matter) and the result is just by the book. In detail we record the following pixel values for each level:
    Code:
     1   1   0 255
     2   2   0 128 255
     3   4   0 64 128 191 255
     4   4   0 64 128 191 255
     5   8   0 32 64 96 128 159 191 223 255
     6   8   0 32 64 96 128 159 191 223 255
     7   8   0 32 64 96 128 159 191 223 255
     8   8   0 32 64 96 128 159 191 223 255
     9  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    10  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    11  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    12  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    13  15   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    14  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    15  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    16  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    17  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    18  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    19  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    20  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    21  15   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    22  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    23  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    24  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239

    Only strange thing is that it's not going down in LOD at 17, but at 23. This could obviously lead to some under sampling at those angles.
    Lowering the contrast of the texture didn't do any difference to the number of samples, so no obvious analyzing of the incoming textures going on..

    So no cheating going on here, but I could imagine those jumps in number of samples could lead to the banding (which I don't think is seen on 6870?).

    But now lets see what the GTX460 produces:
    [​IMG]
    Code:
     1   1   0 255
     2   2   0 127 255
     3   6   0 42 85 127 170 213 255
     4   4   0 64 127 191 255
     5  14   0 26 51 52 77 102 103 127 152 153 178 203 204 229 255
     6   9   0 42 43 85 127 128 170 212 213 255
     7  18   0 18 36 37 55 73 91 109 110 127 145 146 164 182 200 218 219 237 255
     8   8   0 32 64 96 127 159 191 223 255
     9  26   0 15 28 30 43 56 58 71 84 86 99 113 114 127 141 142 156 169 171 184 197 199 212 225 227 240 255
    10  21   0 1 25 26 51 52 76 77 102 103 127 128 152 153 178 179 203 204 229 230 254 255
    11  25   0 12 23 35 46 47 58 69 70 81 93 104 116 127 139 151 162 174 185 197 208 209 220 232 243 255
    12  23   0 21 22 42 43 64 85 86 106 107 127 128 148 149 169 170 190 191 212 213 233 234 254 255
    13  34   19 30 39 41 50 58 61 69 78 80 89 97 100 108 117 119 127 136 138 147 155 158 166 175 177 186 194 197 205 214 216 225 233 236 244
    14  18   18 36 37 55 73 91 109 110 127 128 145 146 164 182 200 218 219 236 237
    15  39   17 26 34 35 43 51 52 60 68 69 77 85 86 94 102 103 111 119 120 127 135 136 144 152 153 161 169 170 178 186 187 195 203 204 212 220 221 229 238 246
    16  16   0 16 32 48 64 80 96 112 127 143 159 175 191 207 223 239 255
    17  14   16 32 48 64 80 96 112 127 143 159 175 191 207 223 239
    18  15   0 16 32 48 64 80 96 112 127 143 159 175 191 207 223 239
    19  14   16 32 48 64 80 96 112 127 143 159 175 191 207 223 239
    20  15   0 16 32 48 64 80 96 112 127 143 159 175 191 207 223 239
    21  15   16 32 48 64 80 96 112 127 143 159 175 191 207 223 239 255
    22  14   16 32 48 64 80 96 112 127 143 159 175 191 207 223 239
    23  15   0 16 32 48 64 80 96 112 127 143 159 175 191 207 223 239
    24  15   0 16 32 48 64 80 96 112 127 143 159 175 191 207 223 239

    It has way to many different values at the odd levels. While some values are just a result of inconsistent rounding (a bit like what I was reporting some time ago), we also have quite some actually different values. So either it's taking too many samples (more than at a higher level!) or it is not sticking to the requested point sampling.
    But from level 16 and on it behaves just like the AMD card, including the same selection of mip levels.

    If I change the texture scale the choice of mip levels will change accordingly - no difference between the cards. So no sign of different LOD bias. (Well, only tried doubling the scale - some slight adjustment may show something).

    Test is in OpenGL, so of course it could behave differently in DX9/10/11. The executable for the standard case is here for now: http://loonies.dk/aftester.exe
     
    #211 Psycho, Dec 18, 2010
    Last edited by a moderator: Dec 18, 2010
  12. Mize

    Mize 3dfx Fan
    Moderator Legend Veteran

    Joined:
    Feb 6, 2002
    Messages:
    5,073
    Likes Received:
    1,138
    Location:
    Cincinnati, Ohio USA
    Can we get the short summary?
     
  13. Psycho

    Regular

    Joined:
    Jun 7, 2008
    Messages:
    745
    Likes Received:
    39
    Location:
    Copenhagen
    :)
    Expected/Required number of samples on Radoen. Too many / weird behavior on Geforce.
    No obvious difference in choice of texture detail level.
     
  14. Neb

    Neb Iron "BEAST" Man
    Legend

    Joined:
    Mar 16, 2007
    Messages:
    8,391
    Likes Received:
    3
    Location:
    NGC2264
    I remember it like that.
     
  15. Mize

    Mize 3dfx Fan
    Moderator Legend Veteran

    Joined:
    Feb 6, 2002
    Messages:
    5,073
    Likes Received:
    1,138
    Location:
    Cincinnati, Ohio USA
    So something is dorked - intentionally or otherwise - that makes shimmer go away but is somehow a kludge?

    Is this like that added random noise that causes textures to be rendered with slight differences each time that was discussed in another thread?
     
  16. Thowllly

    Regular

    Joined:
    Feb 6, 2002
    Messages:
    551
    Likes Received:
    4
    Location:
    Norway
    Tested my 4850 for reference:

    16xAF: Same as the 6870, except at 1:9 scaling it uses 8 samples instead of 16, and it doesn't go down in LOD at all.

    8xAF: Tops out at 8 samples at 1:5. Goes down in LOD at 1:13.

    4xAF: Tops out at 4 samples at 1:3. Goes down in LOD at 1:7 and 1:13.

    2xAF: Tops out at 2 samples at 1:2. Goes down in LOD at 1:4, 1:7 and 1:13.



    So it is actually undersampling. But only because it goes down in LOD too late, not because it uses fewer than specified samples as I used to think. ( of course, it is still possible that it uses fewer samples than specified on textures with lower contrast)
     
  17. Alexko

    Veteran Subscriber

    Joined:
    Aug 31, 2009
    Messages:
    4,514
    Likes Received:
    934
    Or the non-deterministic anti-aliasing algorithm in Fermi… :lol:
     
  18. fellix

    fellix Hey, You!
    Veteran

    Joined:
    Dec 4, 2004
    Messages:
    3,505
    Likes Received:
    424
    Location:
    Varna, Bulgaria
    Just for reference - output from 7800GTX:

    [​IMG]

    Code:
     1   1   0 255
     2   2   0 128 255
     3   8   0 43 85 86 128 169 170 212 255
     4   4   0 64 128 191 255
     5  14   0 26 51 52 77 102 103 128 152 153 178 203 204 229 255
     6  13   0 42 43 84 85 86 127 128 169 170 171 212 213 255
     7  20   0 17 34 37 54 71 74 91 108 111 128 144 147 164 181 184 201 218 221 238 255
     8   8   0 32 64 96 128 159 191 223 255
     9  26   0 16 28 32 44 56 60 72 84 88 100 112 116 128 139 143 155 167 171 183 195 199 211 223 227 239 255
    10  26   0 24 26 48 50 52 74 76 78 100 102 104 126 128 129 151 153 155 177 179 181 203 205 207 229 231 255
    11  32   0 13 23 26 36 46 49 59 69 72 82 92 95 105 115 118 128 137 140 150 160 163 173 183 186 196 206 209 219 229 232 242 255
    12  36   2 21 23 25 42 44 46 63 65 67 84 86 88 105 107 109 126 128 129 146 148 150 167 169 171 188 190 192 209 211 213 230 232 234 251 253 255
    13  34   20 28 36 40 48 56 60 68 76 80 88 96 100 108 116 120 128 135 139 147 155 159 167 175 179 187 195 199 207 215 219 227 235 239 247
    14  37   18 20 36 38 40 54 56 58 72 74 76 90 92 94 108 110 112 126 128 129 143 145 147 161 163 165 179 181 183 197 199 201 215 217 219 233 235 237
    15  39   17 26 34 35 43 51 52 60 68 69 77 85 86 94 102 103 111 119 120 128 135 136 144 152 153 161 169 170 178 186 187 195 203 204 212 220 221 229 238 246
    16  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    17  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    18  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    19  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    20  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    21  15   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    22  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    23  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    24  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    Display driver version: 258.96
    OS: Windows 7 x64 SP1
     
  19. Mize

    Mize 3dfx Fan
    Moderator Legend Veteran

    Joined:
    Feb 6, 2002
    Messages:
    5,073
    Likes Received:
    1,138
    Location:
    Cincinnati, Ohio USA
    Let's hope it's limited to AA! Would suck to have non-deterministic GPGPU results :)
     
  20. hoom

    Veteran

    Joined:
    Sep 23, 2003
    Messages:
    3,010
    Likes Received:
    540
    If I'm understanding this right, the lines are effectively equivalent to parts of the road on a racing game but displayed as separate rectangles?
    Bottom to top looking perpendicular to the texture (down to ground) through to somewhere out towards the horizon?

    With available sample rates being 1:2:4:8:16 we should see the right column being bottom to top 1,2,4,4,8,8,8,8,16,...,16.

    [strike]How do you tell that it has changed LoD after 4th & why should it be at 17 instead of 23?[/strike]

    Edit: Ahh, the LoD change is when the colour output changes from white to red etc.

    5770:
    Code:
     1   1   0 255
     2   2   0 128 255
     3   4   0 64 128 191 255
     4   4   0 64 128 191 255
     5   8   0 32 64 96 128 159 191 223 255
     6   8   0 32 64 96 128 159 191 223 255
     7   8   0 32 64 96 128 159 191 223 255
     8   8   0 32 64 96 128 159 191 223 255
     9  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    10  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    11  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    12  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    13  15   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    14  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    15  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    16  16   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    17  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    18  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    19  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    20  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    21  15   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239 255
    22  14   16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    23  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
    24  15   0 16 32 48 64 80 96 112 128 143 159 175 191 207 223 239
     
    #220 hoom, Dec 19, 2010
    Last edited by a moderator: Dec 19, 2010
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...