prog8/benchmark-program/b_3d.p8
2024-09-25 23:32:45 +02:00

110 lines
3.2 KiB
Lua

%import textio
%import math
rotate3d {
const ubyte WIDTH = 40
const ubyte HEIGHT = 30
sub benchmark(uword max_time) -> uword {
uword anglex
uword angley
uword anglez
uword frames
txt.nl()
cbm.SETTIM(0,0,0)
while cbm.RDTIM16()<max_time {
matrix_math.rotate_vertices(msb(anglex), msb(angley), msb(anglez))
draw_edges() ; doesn't really draw anything in the benchmark, but does do the screen calculations
anglex+=500
angley+=215
anglez+=453
frames++
}
return frames
}
sub draw_edges() {
; plot the points of the 3d cube
; first the points on the back, then the points on the front (painter algorithm)
ubyte @zp i
word @zp rz
word @zp persp
byte @shared sx
byte @shared sy
for i in 0 to len(matrix_math.xcoor)-1 {
rz = matrix_math.rotatedz[i]
if rz >= 10 {
persp = 600 + rz/64
sx = matrix_math.rotatedx[i] / persp as byte + WIDTH/2
sy = matrix_math.rotatedy[i] / persp as byte + HEIGHT/2
;; txt.setcc(sx as ubyte, sy as ubyte, 46, 7)
}
}
for i in 0 to len(matrix_math.xcoor)-1 {
rz = matrix_math.rotatedz[i]
if rz < 10 {
persp = 600 + rz/64
sx = matrix_math.rotatedx[i] / persp as byte + WIDTH/2
sy = matrix_math.rotatedy[i] / persp as byte + HEIGHT/2
;; txt.setcc(sx as ubyte, sy as ubyte, 81, 7)
}
}
txt.chrout('.')
}
}
matrix_math {
; vertices
word[] @split xcoor = [ -40, -40, -40, -40, 40, 40, 40, 40 ]
word[] @split ycoor = [ -40, -40, 40, 40, -40, -40, 40, 40 ]
word[] @split zcoor = [ -40, 40, -40, 40, -40, 40, -40, 40 ]
; storage for rotated coordinates
word[len(xcoor)] @split rotatedx
word[len(ycoor)] @split rotatedy
word[len(zcoor)] @split rotatedz
sub rotate_vertices(ubyte ax, ubyte ay, ubyte az) {
; rotate around origin (0,0,0)
; set up the 3d rotation matrix values
word wcosa = math.cos8(ax)
word wsina = math.sin8(ax)
word wcosb = math.cos8(ay)
word wsinb = math.sin8(ay)
word wcosc = math.cos8(az)
word wsinc = math.sin8(az)
word wcosa_sinb = wcosa*wsinb / 128
word wsina_sinb = wsina*wsinb / 128
word Axx = wcosa*wcosb / 128
word Axy = (wcosa_sinb*wsinc - wsina*wcosc) / 128
word Axz = (wcosa_sinb*wcosc + wsina*wsinc) / 128
word Ayx = wsina*wcosb / 128
word Ayy = (wsina_sinb*wsinc + wcosa*wcosc) / 128
word Ayz = (wsina_sinb*wcosc - wcosa*wsinc) / 128
word Azx = -wsinb
word Azy = wcosb*wsinc / 128
word Azz = wcosb*wcosc / 128
ubyte @zp i
for i in 0 to len(xcoor)-1 {
; don't normalize by dividing by 128, instead keep some precision for perspective calc later
rotatedx[i] = Axx*xcoor[i] + Axy*ycoor[i] + Axz*zcoor[i]
rotatedy[i] = Ayx*xcoor[i] + Ayy*ycoor[i] + Ayz*zcoor[i]
rotatedz[i] = Azx*xcoor[i] + Azy*ycoor[i] + Azz*zcoor[i]
}
}
}