.tHE .sIRIUS .cYBERNETICS .cORPORATION

 (Really) Basic Things About 3D Graphics  In this article I would like to tell you some things I learned during last 23 months. It's dedicated to people who understand some basics, but that have a mishmash in it :) So if you're experienced 3D coder, skip this article... After reading I recommend to get any serious docs about 3D coding, for example from my homepage (hey!!! that's selfpromotion!!! kill him! kill him!!! :)) to get deeper into this problematic. OK, let's start: Matrices  No, no, I'm not going to explain here what they are and how to calculate with them. I only want to say, they are a MUST for doing 3D things! So take your school book about Linear Algebra and start to study :) Rows vs Columns  Or, more concretly, I mean this: /a b c\ /a d g\ /x\ [x y z] * d e f and b e h * y \g h i/ \c f i/ \z/ ^ ^ ^ ^ point matrix matrix point You know the result in both cases is the same: (x*a+y*d+z*g x*b+y*e+z*h z*c+y*f+z*i) resp. /a*x+d*y+g*z\ b*x+e*y+h*z \c*x+f*y+i*z/ Where the 1st column/row means final xcomposant, the 2nd ycomposant and the 3rd zcomposant. That's clear. You see that (x,y,z) point didn't change (except the fact that is in column instead of row) and the matrix is transposed (rows= columns and vice versa; btw in the case of inverse rotation is the transposition equivalent to inverse operation). These two examples are two various possibilities how to write matrix & point/ vector multiplication. In this case it doesn't matter what kind of multiplication we use. But, let's look at this example: imagine that we have three matrices: A (rotation around zaxis), B (around yaxis) and C (around xaxis) and one point with (x,y,z) coordinates. We want to rotate around z, then y and finally around x the axis. And here comes the problem: you know matrix multiplication isn't comutative (A.B != B.A in general case) and we swapped the point and matrix in the example above, so what is the right order??? Ok, ok, here's solution :) new [x y z] = [x y z] * A * B * C or /x\ /x\ new y = C * B * A * y \z/ \z/ Got the idea? If we use point/vector as row, we start with point/vector on the left side and we write matrices from left to right and vice versa. I think that I mustn't write that we always multiple from left to right :) Some of the texts I red had mistakes in order of this multiplication. Homogeneous vs Nonhomogeneous Coordinate System  Or, in another jargon: affine vs linear transformation. Probably you know how looks rotation matrix for 2D around the origin (0,0): / cosFI sinFI\ rotated [x y] = [x y] * \sinFI cosFI/ that means we got result: rotated x = x*cosFI  y*sinFI rotated y = x*sinFI + x*cosFI Basic stuff, isn't? :) Similarly for 3D: matrix for rotating around xaxis: /1 0 0 \ 0 cosFI sinFI \0 sinFI cosFI/ etc... You know, if we multiple A, B, C matrices from the example above, we'll get again the 3x3 matrix. This looks nearly perfect, but ... again that wellknown BUT :) All things like rotation, scaling, stretching around the origin are LINEAR transformations. Imagine we want to TRANSLATE point, that means we want to move from original (x,y,z)point to, let's say, (x+10,y10,z+5). And we want to do with matrices. So the only way how to do it is as follows: /1 0 0\ [x y z 1] * 0 1 0 0 0 1 \10 10 5/ But hey !! 4x3 matrix?? We can't use it! (becoz of inverse matrices, see below) So what?? Now you probably think: "I can fuck up such translation matrices, I'll add it manually before rotatematrix multiplication" Yes, you can, but look at this: we want to rotate point around zaxis (A), then move to the origin (0,0,0) (M), then rotate around yaxis (B), then rotate around xaxis (C) and move back to the original (x,y,z) position. (resp to the position rotated around zaxis). In ideal case it could look like this: new [x y z] = [x y z] * A * M * B * C * M^(1) (M^(1) means inverse) But as we could see, it's impossible becoz of multiplying 3x3 matrices with 4x3. So... and here it comes... AFFINE transformation :) All the problem is with non square 4x3 matrix. So what about "extending" to the 4x4? /1 0 0 0\ [x y z 1] * 0 1 0 0 0 0 1 0 \10 10 5 1/ Hmmm... that looks nice... let's applicate it to the rotation matrix: /1 0 0 0\ [x y z 1] * 0 cosFI sinFI 0 0 sinFI cosFI 0 \0 0 0 1/ And we got what? Square matrix !!! Now we can multiply/inverse/anything_you_want these matrices and all is ok! Final note: if we use 4D space (4x4 matrices), it's called homogeneous coordinates, if we use only 3x3 matrices, we're working with nonhomogeneous coordinates. Ofcoz there are applications in (x,y) space, but there is better to get final formulas instead of working with matrices. Righthanded vs Lefthanded Coordinate system  Probably you know that stupid theorem with fingers from physics :) Only thing you need to know is how looks these systems: +y +y +z/   /   /   /  +x / +x /  / / /+z righthanded lefthanded There are only two differences: 1)direction of zaxis > sign of z value 2)direction of rotation around zaxis (clock vs anticlockwise) I prefer lefthanded one since if we put eye/camera in origin (0,0,0), objects with positive zcoordinates are visible and objects wit negative zcoordinates are behind us (nonvisible) Additive vs Cumulative Rotating  There are two different ways of rotating the point in 2D/3D space. You have rotation matrix, cos/sin tables and you want to rotate around whole circle (360 degrees/2pi radians/256 units). Additive rotating: you start with alpha/beta/gama == 0. Calculate rotation matrix with these angles, save them somewhere and use it with every point. Then increase alpha/beta/gama and next frame you'll calculate new rotation matrix. So every frame you'll do the matrix precalcing and for every point you'll do 9 muls (or less if you skip any angle or optimize your rotation matrix) Cumulative rotating: you start again with alpha/beta/gama == 0. But you'll define how much they will increase, e.g. if one degree, we precalc ONLY sin (1) and cos (1) values and these values we'll use in our matrix (remember we'll do this only ONCE  in the init part of the program). And, here comes the main idea, after calculating of every point, we put them to the screen, but to their original coordinates, too. So what will happen in the next frame? Our "pointbuffer" contains rotated coordinates => we need again only 1 degree to rotate and that's the reason why we don't need to precalc the rotation matrix again. Again, here are some pros and cons... Against additive rotating is the speed  remember in every frame we need to calculate 14 muls, that means it takes "some" CPU time... But a big pro is (instead of cumulative rotating) an ACCURACY > in cumulative rotating you need to "refresh" original values (e.g. in 90/180 degrees) bcoz you'll see nice morphingdeformations :)) and in generall, with more muls you need more refreshing... Final Words  Yes, we're at the end of this article. I hope I helped someone, if not, it was at least a good excercise for me :) In case of positive feedback I can write some more things for newbies in 3D gfx, bcoz it's really cool stuff :) MiKRO  mk20432@decef.elf.stuba.sk XE/XL/MegaSTE/Falcon mikro.atari.org  
