Fast multiply / divide by 64?

[mcgeezer]

Hi all,

My code uses a lot of multiplying and dividing by 64.

for example:

lsl.w #6,d0
lsl.w #6,d1

---

lsr.w #6,d0
lsr.w #6,d1

Each one of those shifts is taking 18 cycles.

The maximum number i need to shift to and from is 200 so if I were to use a table I would need values going up to 12800, therefore I would need to hold a table on memory that would be over 25Kb (big waste).

Is there a better/faster method i can use for doing this multiplication without sapping 25Kb ram with a table?

Looking at doing it in a table I still would'nt improve on the cycle count anyway.

Cheers,
Geezer

[DanScott]

add.w d0,d0
move.w Shift6Table(PC,d0.w),d0


you're still looking at 18 cycles for that anyway

[alexh]

Are the shift instructions faster if the shift value is in a register rather than an immediate?

[Don_Adan]

Originally Posted by mcgeezer

Hi all,

My code uses a lot of multiplying and dividing by 64.

for example:

lsl.w #6,d0
lsl.w #6,d1

---

lsr.w #6,d0
lsr.w #6,d1

Each one of those shifts is taking 18 cycles.

The maximum number i need to shift to and from is 200 so if I were to use a table I would need values going up to 12800, therefore I would need to hold a table on memory that would be over 25Kb (big waste).

Is there a better/faster method i can use for doing this multiplication without sapping 25Kb ram with a table?

Looking at doing it in a table I still would'nt improve on the cycle count anyway.

Cheers,
Geezer

I see only one good enough solution for 64 multiply, but it needs many code changes. You can place D1.W at high word position of D0 and use lsl.l #6,D0 and later swap D0 or move.l D0,D1 and swap D1.

[mcgeezer]

Originally Posted by alexh

Are the shift instructions faster if the shift value is in a register rather than an immediate?

Not as far as I can see.
http://mrjester.hapisan.com/04_MC68/...eTimes/LSL.htm

[mcgeezer]

Originally Posted by Don_Adan

I see only one good enough solution for 64 multiply, but it needs many code changes. You can place D1.W at high word position of D0 and use lsl.l #6,D0 and later swap D0 or move.l D0,D1 and swap D1.

This is not a bad idea as I have to move x and y coordinates into the registers anyway. Looks like an lsl.l is 20 cycles as opposed to 18 for a .w

so at the moment i typically have.

move.w XPOS(a2),d0    ;12 
move.w YPOS(a2),d1    ;12
lsl.w #6,d0    ;18
lsl.w #6,d1    ;18
; do something to move object
lsr.w #6,d0    ;18
lsr.w #6,d1    ;18

; 96 cycles

move.w XPOS(a2),d0    ;12 
swap d0    ;4
move.w YPOS(a2),d0    ;12
lsl.l #6,d0    ;20

; do something to move object
lsr.l #6,d0    ;20

; 68 cycles.

Big saving!

[DanScott]

Depending on how your data is arranged, and if you have a reasonably large (and well arranged) array of words that need shifting, you could use the blitter with shift and A source masking to shift them all in one go.

[DanScott]

For the swap method, you should arrange XPos and YPos to be consecutive words, so you can just read them with a move.l XPos(a0),d0

[Don_Adan]

Originally Posted by DanScott

For the swap method, you should arrange XPos and YPos to be consecutive words, so you can just read them with a move.l XPos(a0),d0

Yes, it can save next 12 cycles too.

[mcgeezer]

Originally Posted by DanScott

For the swap method, you should arrange XPos and YPos to be consecutive words, so you can just read them with a move.l XPos(a0),d0

And infact they already are!!!

[Don_Adan]

And remember that divide by 64 is not safe. This is dependent on input data values and/or later data handling. You can/must use:
; do something to move object
lsr.l #6,d0 ;20
and.w #$00FF, D0