|
| Author |
Message |
Jamie
Member |
I liked the discussion on 68060 texture mapping in a previous topic, i have some idea that i never tested.
************************************************** **********
here is a 7 cycle 2 pixel loop which has 8.8 precision for U and 8.16 for V:
************************************************** **********
.pixelLoop:
rept 4
add.l a0,d0 ;cycle 1 poep
lsr.l #8,d2 ;cycle 1 soep
add.l a1,d1 ;cycle 2 poep
move.b (a2,d2.w),d4 ;cycle 2 soep
move.l d1,d3 ;cycle 3 poep
move.b d4,(a3)+ ;cycle 3 soep
move.w d0,d3 ;cycle 4 poep
add.l a0,d0 ;cycle 4 soep
lsr.l #8,d3 ;cycle 5 poep
add.l a1,d1 ;cycle 5 soep
move.b (a2,d3.w),d4 ;cycle 6 poep
move.l d1,d2 ;cycle 6 soep
move.b d4,(a3)+ ;cycle 7 poep
move.w d0,d2 ;cycle 7 soep
endr
subq.l #8,d7
bgt.b .pixelLoop
rts
* I like the fact that i don't need to have some ror and eor operation in the outerloop for preparing addx pixel instruction loop, when you have lot of polygon the outerloop is really important too.
* I can have multisize texture mapping with this loop
* I can't have tiled texture
************************************************** **********
here is a ? cycle pixel loop which has 8.24 precision for UV
************************************************** **********
.pixelLoop:add.l d0,(a0)
add.l d1,(a1)
move.w (a1),d2
move.b (a0),d2
move.b (a2,d2.l),(a3)+
* It's really accurate
* few register is needed
* texture is UV tiled
* V can be multisize
* I'm sure it's slow:)
|
Kalms
Member |
If you accept dropping V precision to 8.8 in your first example, you can use:
add.l d1,d0 ; 1 p
move.b (a0,d2.w),d4 ; 1 s
move.l d0,d2 ; 2 p
lsr.w #8,d2 ; 2 s
add.l d1,d0 ; 3 p
rol.l #8,d2 ; 3 s
move.b (a0,d3.w),d5 ; 4 p
move.l d0,d3 ; 4 s
lsr.w #8,d3 ; 5 p
move.b d4,(a1)+ ; 5 s
rol.l #8,d3 ; 6 p
move.b d5,(a1)+ ; 6 s
-> 6 cycles for 2 pixels.
here, VVvv and UUuu are packed into the same register (counter in d0, gradient in d1).
Same restrictions apply as for your suggested loop, except that you might need to do the [reasonably cheap, but still] UV-packing per line (due to too large U->V overflow errors otherwise).
|
Jamie
Member |
add.l a0,a2
add.l a1,a3
move.l a2,d0
move.l a3,d1
add.l a0,a2
add.l a1,a3
move.l a2,d2
move.l a3,d3
ror.l #8,d0
rol.l #8,d1
ror.l #8,d2
rol.l #8,d3
move.b d0,d1
move.b d2,d3
move.b (a4,d1.w),(a5)+
move.b (a4,d3.w),(a5)+
-> 4,5 but you can have 24 bit precision, multisize and tile
|
Jamie
Member |
Finally 4 cycle with 8:16 for u and 8:24 for v
XX U1 P1 P2
V1 P1 P2 P3
move.l d1,d2
lsr.l d5,d3
and.l d4,d2
add.l a1,d1
move.b (a2,d3.l ),d6
or.l d0,d2
move.b d6,(a3)+
add.l a0,d0
move.l d1,d3
lsr.l d5,d2
and.l d4,d3
add.l a1,d1
move.b (a2,d2.l ),d6
or.l d0,d3
move.b d6,(a3)+
add.l a0,d0
|
Jamie
Member |
maybe an interesting solution for uv 16:16
d0 = U1 P1 P2 V1
d1 = P1 P2 xx xx
rol.l #8,d0
move.b (a0,d0.w),(a1)+
ror.l #8,d0
add.l d3,d1
addx.l d2,d0
|
|
|
