|
Elite Member
|
引用:
Originally posted by giligula
[恕刪]
毛兄
可以請你看看同時在 .INI 中開啟 AF 及 control panel 也同時開啟
結果會是哪種嗎?
9000Pro 的話
是會遵照 .ini 檔
我的看法和你一樣
效能的差別就是在這邊不同
也覺得是驅動程式的問題
可能是 bug
但「偷吃步」的機會也很高
這個消息有些勁爆
好像還沒在國外的網站上出現
|
呵呵，我有試過改了ini之後(LevelOfAnisotropy-->8)
如果在cp裡面設定8xafq的話，會是bilinear
而如果勾依應用程式喜好，就會出現正常的圖(如上面那張)
由於裝了A.D.情況一樣，我猜這個機制應該不是"偵測性"的(也就是都有此種機制)
呵呵，我再給你看一段他寫給我的，有點長(他真的是對我太好了  )
他認為nv在ut2003詭異的表現的原因，
差不多全文的重點我都已經貼出來了，我把我覺得重要的地方加粗體 
(直接接上篇：)
引用:
You also need some background information. UT2003 is rather advanced game, it supports something called EARLY Z rejection. What does it do? It’s a function that when used with supported hardware that reduces over draw. What’s over draw? A simple version, game has 2 items on screen, but 1 is covered by the first item. We can’t see the second item, but the graphic card doesn’t know that. So it renders both. The work done by rendering the covered item goes waste. That’s overdraw or, useless work. Now back to early z rejection. Since everything is 3D, the graphic card has a buffer (they call it Z buffer) that stores all the 3D coordinates in a game. That buffer needs to be accessed all the time, and 40% or more of the graphic card bandwidth is used to access that buffer. There is a 2D and 3D buffer. R300 divides Z buffer to 256 different blocks (for easy access), before a frame is drawn, the 2d coordinates and 3d has to agree.
R300 has hardware support for early Z, which means based on the coordinates of different items on screen in the Z buffer, it guesses what part of an item is not visible or completely invisible. That part of data is then deleted before it’s accessed. So when the buffer is fetched, only coordinates of visible items are sent, and this reduces bandwidth consumption and the graphics card will use the extra bandwidth for other functions and hence speed bump. Just remember fill rate=bandwidth. And fill rate also = performance.
NVIDIA claims their NV30 and NV35 something similar to early Z, but I really think it’s a lie because NV3X could never keep pace with R300 in this game despite its massive fill rate and bandwidth advantage, which would suggest NV3X waste a lot of fill rate drawing things we can’t see. And personally, because of this, NV3X should never be faster than R300 under this title, but it does, especially with AF/AA. Isn’t that suspicious?
And because it wastes fill rate drawing things we can’t see, a trilinear filtering puts it on a bit of strain than R300, so they want to reduce Trilinear filtering as much as possible, just like ATI. But unlike ATI, NVIDIA can’t filter 1 texture layer in Trilinear and bilinear the rest. It’s all or nothing. If you force Radeon to filter everything in Trilinear, performance is 30 to 40% less comparing to having 1 layer filtered in Trilinear! I’d say probably the same or worse for NVIDIA. So they have decided not to Trilinear filter but with bilinear instead (you know, benchmark wars). It is not noticeable in 80 to 90% of games. But NVIDIA always criticize Radeon 8500 method of AF because it can’t do Trilinear, and because it’s so fast. While telling everyone Trilinear looks better than bilinear and ATI is cheating.
|
|
2003-07-28, 02:44 AM
#51
