LOSTCIRCUITS

My first encounter with TRUFORM was at Computex in form of the Dolphin demo. The technology behind TRUFORM is actually quite interesting since it involves probabilistic calculations how a crude, low tesselation surface could be converted into something more realistic. The real issue here, once again, is not necessarily the next step towards photorealism, rather, it is the need to make more out of less. Essentially, any given number of vertices within a surface allows to predict what would lie between if the resolution were higher. The rest is almost trivial: take two vertices, divide the distance between them by three and get two additional assumed vertices. As a result, any triangle will have 6 additional vertices that can be used to simulate one additional vertex in the center. Needless to say that also the normals are calculated. Consequently, instead of having a single triangle, the same surface is split into 9 individual triangles without increasing bandwidth over the single triangle. The degree of tessellation can be changed depending on the application from a 4 x to a 64 x increase in vertices, that is each distance between vertices is divided by 2, 3, 4, ... 8 to generate 4, 9, 16, ... 64 additional vertices from a single triangle.

Three vertices and their normals form one triangle and define the angle to the adjacent triangle.	Additional vertex points are inserted along the connecting lines between vertices. the number of divisions depends on the level of tesselation.
Additional vertices are generated to form a regular geometric network.	The additional normals are added.

TRUFORM is, in most cases, quite accurate, there are some cases, though where the guesswork of TRUFORM will overstep its boundaries. A classic case is when two straight walls meet at an angle since TrueForm will smoothen the edge and possibly create a bump.

In some cases, TRUFORM can distort even surfaces.

Like most advanced effects, TRUFORM lives from dynamic changes of the picture, without motion it is rather impossible to capture its effect on the images. Suffice it to show the effect on the wireframe of the dolphin.

The dolphin without TRUFORM (above) and with 64x tesselation (below).

SMOOTHVISION

Smoothvision is ATI's version of antialiasing (AA) and it is a neat trick that ATI has pulled on how to generate an optical illusion of smooth edges. In short, there are two major techniques called multisampling and supersampling.

Supersampling renders the image at a higher resolution than the actual display and then scales it down to the output resolution. The advantage is that each pixelblock is divided into sub-blocks that are then averaged with the neighboring pixel colors. The main problem is that the virtual resolution of e.g. 4x supersampling is, in fact, 4 times as high as the output resolution and, consequently, requires a tremendous amount of rendering muscle (Voodoo5).

Multisampling is a different approach in that each scene is rendered multiple times at the original resolution but each copy is shifted in a random direction by ½ of a pixel width. Subsequently, the different samples are blended together to generate the output. The advantage is that not the entire image has to be averaged but only those pixels where the directional shift has generated a difference in color values. This technique borrows from edge detection algorithms (since only edges will generate a differential if shifted) and consequently, multisampling is better described as edge antialiasing. The advantage is that textures remain largely unaltered (they are slightly blurred in Supersampling techniques) and only pixels that belong to different triangles are averaged according to their subpixel content. This technique requires about the same bandwidth as multisampling and is used by the GeForce3.

SMOOTHVISION

According to ATI's white papers, SMOOTHVISION is a supersampling technique using 16 sample blocks to cover 8 or less pixels (depending on the level of AA chosen). Out of these eight pixels, four are chosen for a 4x multisampling (somewhat conflicting ... ). The sampling pattern is not-ordered, meaning that the pattern of points taken for sampling changes in a radom fashion from one image to the next. This is possible since only half of the pixels are sampled each time. For still images, the result can be less than convincing since it might look like no antialiasing happens at all. For any dynamic sequence, though, the effect is mind-boggling since the alternating random shifts of individual pixel groups are perceived by the eye as a smooth line without negative impact on clarity of textures. In other words, ATI is creating an optical illusion suggesting smooth lines by temporal superimposition of jagged lines with 50% jittering smooth parts embedded at random positions (This is about as trivial as I can explain it).

next page:    => FIC's version of the RADEON 8500 =>