Move it Move it... (68000)

Gilloo · 28 November 2011, 09:43

I have a steam of data like that U0N0I0X0 and I want to transform it as faster as 68000 can do into UNIX (skip one byte data on two)

Here the code
the string U0N0I0X0 is pointed by A0 and the move_xxx is called in a infernal loop between two time() in C langage to get the time taken by each function...

move_optimized
movem D0-D3/A1,-(SP)
move.l A0,A1
move.l (A1)+,D0 ;
move.l (A1)+,D1 ; U0N0 I0X0

lsr.w #8,D0 ;
lsr.w #8,D1 ; U00N I00X
clr.l D2
move.l D2,D3
move.w D0,D2
move.w D1,D3
clr.w D0
clr.w D1 ; U000 I000 000N 000X
swap D1 ; U000 00I0 000N 000X
swap D2 ; U000 00I0 0N00 000X
or.l D1,D0 ; U0I0 00I0 0N00 000X
or.l D2,D0 ; UNI0 00I0 0N00 000X
or.l D3,D0 ; UNIX 00I0 0N00 000X

move.l D0,(A0)

movem (SP)+,D0-D3/A1
rts

move_not_optimized movem D0-D3/A1,-(SP)
move.b 0(A0),0(A0)
move.b 2(A0),1(A0)
move.b 4(A0),2(A0)
move.b 6(A0),3(A0)
movem (SP)+,D0-D3/A1
rts

And... it's lame !! the move_not_optimized is faster

musashi5150 · 28 November 2011, 09:51

Sounds like a job for MOVEP (unless you need it to run on ALL 680x0).

StingRay · 28 November 2011, 09:54

movep.l 0(a0),d0
move.l d0,(a0)

should do the trick I think. Not tested.

Edit: too slow again... Musashen was faster.

musashi5150 · 28 November 2011, 10:01

Quote:

Originally Posted by StingRay

Edit: too slow again... Musashen was faster.

Muhahaha!

But we'll call it even as you wrote some code too

Gilloo · 28 November 2011, 12:00

boiîing! movep is illegal on 68060... but I will try it for testing purpose on 68000 (thanks

)

StingRay · 28 November 2011, 12:13

Quote:

Originally Posted by Gilloo

boiîing! movep is illegal on 68060... but I will try it for testing purpose on 68000 (thanks

)

well, well:

Quote:

Originally Posted by Gilloo

I have a steam of data like that U0N0I0X0 and I want to transform it as faster as 68000 can do

Anyway, the common 060 patchers (Oxypatcher etc.) make the movep instruction available, even on 060. Anyway, time to think about another solution w/o movep.

28 November 2011, 15:09

Gilloo, as musashi5150 and StingRay mentioned this is task for MOVEP.
below you can find my 060 friendly routine (input text should be word/long aligned):

Code:

    lea    data_in(PC),a0
    lea    data_out(PC),a1

    move.l (A0)+,D0 ; U0N0
    move.l    D0,D2
    lsl.l    #8,D2
    or.l    D2,D0

    move.l (A0)+,D1 ; I0X0
    move.l    D1,D3
    lsl.l    #8,D3
    or.l    D3,D1

    swap.w    D1
    move.w    D1,D0

    move.l    D0,(A1)+


    even
data_in
    dc.b    'U',0,'N',0,'I',0,'X',0

data_out
    dc.l    0

Don_Adan · 29 November 2011, 17:02

Quote:

Originally Posted by Gilloo

I have a steam of data like that U0N0I0X0 and I want to transform it as faster as 68000 can do into UNIX (skip one byte data on two)

Here the code
the string U0N0I0X0 is pointed by A0 and the move_xxx is called in a infernal loop between two time() in C langage to get the time taken by each function...

move_optimized
movem D0-D3/A1,-(SP)
move.l A0,A1
move.l (A1)+,D0 ;
move.l (A1)+,D1 ; U0N0 I0X0

lsr.w #8,D0 ;
lsr.w #8,D1 ; U00N I00X
clr.l D2
move.l D2,D3
move.w D0,D2
move.w D1,D3
clr.w D0
clr.w D1 ; U000 I000 000N 000X
swap D1 ; U000 00I0 000N 000X
swap D2 ; U000 00I0 0N00 000X
or.l D1,D0 ; U0I0 00I0 0N00 000X
or.l D2,D0 ; UNI0 00I0 0N00 000X
or.l D3,D0 ; UNIX 00I0 0N00 000X

move.l D0,(A0)

movem (SP)+,D0-D3/A1
rts

move_not_optimized movem D0-D3/A1,-(SP)
move.b 0(A0),0(A0)
move.b 2(A0),1(A0)
move.b 4(A0),2(A0)
move.b 6(A0),3(A0)
movem (SP)+,D0-D3/A1
rts

And... it's lame !! the move_not_optimized is faster

For your example, simple:
move.b 2(A0),1(A0) ; 20c
move.b 4(A0),2(A0) ; 20c
move.b 6(A0),3(A0) ; 20c
rts

is one of the fastest (60 cycles for 68000) solution.

Movep version need 24cycles + 12cycles, and extra 24 cycles, if all (D0 too) registers are preserved. Other versions has no big sense to use for this example, especially for 68000 CPU.

Photon · 29 November 2011, 19:17

Quote:

Originally Posted by StingRay

well, well:

Anyway, the common 060 patchers (Oxypatcher etc.) make the movep instruction available, even on 060. Anyway, time to think about another solution w/o movep.

Heartily agree and yes optimize for 68000 unless you intend to break new performance records on AGA 060

Asman · 29 November 2011, 21:44

My approach ( a0 - start of data )

Code:

	move.l	a0,a1		;4 (added after Don Adan post)
	move.w	(a0)+,d0	;8
	move.b	(a0),d0		;8
	move.l	(a0)+,d1	;12
	move.b	(a0),d1		;8
	movem.w	d0-d1,-(a1)	;16
				;total 56 cycles

Don_Adan · 01 December 2011, 14:12

Quote:

Originally Posted by Asman

My approach ( a0 - start of data )

Code:

    move.w    (a0)+,d0    ;8
    move.b    (a0),d0        ;8
    move.l    (a0)+,d1    ;12
    move.b    (a0),d1        ;8
    movem.w    d0-d1,-(a0)    ;16
                ;total 52 cycles

Perhaps
movem.w D0-D2,-(A0) ; 20 c
or
subq.w #2,A0 ; 8c ?
or
move.l A0,A1 ; 4c
....
movem.w D0-D1, (A1) ; 16c

Asman · 01 December 2011, 21:41

Sorry Don Adan, I don't catch what you mean. For sure subq.l #2,a0 takes 8c. Please post full routine, thanks.

Codetapper · 01 December 2011, 22:05

lea -2(a0),a0 would save a couple wouldn't it?

StingRay · 02 December 2011, 07:51

No, both instructions have the same cycle count. And since lea x(ax),ax is 2 bytes larger than subq #x,ax, it's best to use that one here.

Don_Adan · 02 December 2011, 12:55

Quote:

Originally Posted by Asman

Sorry Don Adan, I don't catch what you mean. For sure subq.l #2,a0 takes 8c. Please post full routine, thanks.

Simple, you change A0, 6 bytes forward by "+" commands, but only 4 bytes back by "-" command :

Code:

 move.w (a0)+,d0 ;8
 move.b (a0),d0 ;8
 move.l (a0)+,d1 ;12
 move.b (a0),d1 ;8
 subq.w #2,A0 ; 8c
 movem.w d0-d1,-(a0) ;16 ;total 60 cycles

Code:

 move.l A0,A1 ; 4c
 move.w (a0)+,d0 ;8
 move.b (a0),d0 ;8
 move.l (a0)+,d1 ;12
 move.b (a0),d1 ;8
 movem.w d0-d1,(a1) ;16 ;total 56 cycles

Don_Adan · 02 December 2011, 12:59

Quote:

Originally Posted by StingRay

No, both instructions have the same cycle count. And since lea x(ax),ax is 2 bytes larger than subq #x,ax, it's best to use that one here.

Do you know how many cycles used addq,w #x,Ax? 4c or 8c?
Of course for 68000.

pmc · 02 December 2011, 13:30

According to the Motorola 68000 User Manual: addq.w #x,Ax == 4 cycles, addq.l #x,Ax == 8 cycles

For the #x,(Ax) variant you need to add the effective address calc time on top of that which is another 4 cycles for word sized operations or 8 cycles for longword sized operations.

Asman · 02 December 2011, 14:33

@Don Adan
Right. I edited my original post. Thank you.

TheDarkCoder · 04 December 2011, 15:38

Quote:

Originally Posted by pmc

According to the Motorola 68000 User Manual: addq.w #x,Ax == 4 cycles, addq.l #x,Ax == 8 cycles

For the #x,(Ax) variant you need to add the effective address calc time on top of that which is another 4 cycles for word sized operations or 8 cycles for longword sized operations.

pmc, you should stop believing in manuals!

You quoted a known mistake in the 68k user manual.
In reality addq.w,Ax == 8 cycles, addq.l #x,Ax == 8 cycles

We already discussed the issue:
http://eab.abime.net/showthread.php?...ghlight=addq.w

pmc · 04 December 2011, 17:36

LOL

I'm just too trusting in the documentation

Thanks for reminding me!

28 November 2011, 09:43	#1
Gilloo Registered User Join Date: Nov 2010 Location: Grenoble, Isère, Rhône-Alpes, France, Europe, Earth Posts: 287	Move it Move it... (68000) I have a steam of data like that U0N0I0X0 and I want to transform it as faster as 68000 can do into UNIX (skip one byte data on two) Here the code the string U0N0I0X0 is pointed by A0 and the move_xxx is called in a infernal loop between two time() in C langage to get the time taken by each function... move_optimized movem D0-D3/A1,-(SP) move.l A0,A1 move.l (A1)+,D0 ; move.l (A1)+,D1 ; U0N0 I0X0 lsr.w #8,D0 ; lsr.w #8,D1 ; U00N I00X clr.l D2 move.l D2,D3 move.w D0,D2 move.w D1,D3 clr.w D0 clr.w D1 ; U000 I000 000N 000X swap D1 ; U000 00I0 000N 000X swap D2 ; U000 00I0 0N00 000X or.l D1,D0 ; U0I0 00I0 0N00 000X or.l D2,D0 ; UNI0 00I0 0N00 000X or.l D3,D0 ; UNIX 00I0 0N00 000X move.l D0,(A0) movem (SP)+,D0-D3/A1 rts move_not_optimized movem D0-D3/A1,-(SP) move.b 0(A0),0(A0) move.b 2(A0),1(A0) move.b 4(A0),2(A0) move.b 6(A0),3(A0) movem (SP)+,D0-D3/A1 rts And... it's lame !! the move_not_optimized is faster

28 November 2011, 15:09	#7
CyprianK Posts: n/a	Gilloo, as musashi5150 and StingRay mentioned this is task for MOVEP. below you can find my 060 friendly routine (input text should be word/long aligned): Code: lea data_in(PC),a0 lea data_out(PC),a1 move.l (A0)+,D0 ; U0N0 move.l D0,D2 lsl.l #8,D2 or.l D2,D0 move.l (A0)+,D1 ; I0X0 move.l D1,D3 lsl.l #8,D3 or.l D3,D1 swap.w D1 move.w D1,D0 move.l D0,(A1)+ even data_in dc.b 'U',0,'N',0,'I',0,'X',0 data_out dc.l 0 Last edited by CyprianK; 29 November 2011 at 13:05.

29 November 2011, 21:44	#10
Asman 68k Join Date: Sep 2005 Location: Somewhere Posts: 828	My approach ( a0 - start of data ) Code: move.l a0,a1 ;4 (added after Don Adan post) move.w (a0)+,d0 ;8 move.b (a0),d0 ;8 move.l (a0)+,d1 ;12 move.b (a0),d1 ;8 movem.w d0-d1,-(a1) ;16 ;total 56 cycles Last edited by Asman; 02 December 2011 at 14:32. Reason: correction

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Move Em	amigapd	request.Old Rare Games	1	13 March 2011 02:22
[68000 ASM] Move memory to memory	AmigaBoy	Coders. General	7	08 December 2009 08:16
How to move files from PC with Catweasel ?	zimgr5	support.Hardware	2	20 February 2009 11:21
Please delete or move if necessary	JSemple3	support.Games	7	30 June 2003 13:59

28 November 2011, 09:51	#2
musashi5150 move.w #$4489,$dff07e Join Date: Sep 2005 Location: Norfolk, UK Age: 42 Posts: 2,351	Sounds like a job for MOVEP (unless you need it to run on ALL 680x0).

28 November 2011, 09:54	#3
StingRay move.l #$c0ff33,throat Join Date: Dec 2005 Location: Berlin/Joymoney Posts: 6,863	movep.l 0(a0),d0 move.l d0,(a0) should do the trick I think. Not tested. Edit: too slow again... Musashen was faster.

28 November 2011, 12:00	#5
Gilloo Registered User Join Date: Nov 2010 Location: Grenoble, Isère, Rhône-Alpes, France, Europe, Earth Posts: 287	boiîing! movep is illegal on 68060... but I will try it for testing purpose on 68000 (thanks )

01 December 2011, 21:41	#12
Asman 68k Join Date: Sep 2005 Location: Somewhere Posts: 828	Sorry Don Adan, I don't catch what you mean. For sure subq.l #2,a0 takes 8c. Please post full routine, thanks.

01 December 2011, 22:05	#13
Codetapper 2 contact me: email only! Join Date: May 2001 Location: Auckland / New Zealand Posts: 3,182	lea -2(a0),a0 would save a couple wouldn't it?

02 December 2011, 07:51	#14
StingRay move.l #$c0ff33,throat Join Date: Dec 2005 Location: Berlin/Joymoney Posts: 6,863	No, both instructions have the same cycle count. And since lea x(ax),ax is 2 bytes larger than subq #x,ax, it's best to use that one here.

02 December 2011, 13:30	#17
pmc gone Join Date: Apr 2007 Location: completely gone Posts: 1,596	According to the Motorola 68000 User Manual: addq.w #x,Ax == 4 cycles, addq.l #x,Ax == 8 cycles For the #x,(Ax) variant you need to add the effective address calc time on top of that which is another 4 cycles for word sized operations or 8 cycles for longword sized operations.

02 December 2011, 14:33	#18
Asman 68k Join Date: Sep 2005 Location: Somewhere Posts: 828	@Don Adan Right. I edited my original post. Thank you.

04 December 2011, 17:36	#20
pmc gone Join Date: Apr 2007 Location: completely gone Posts: 1,596	LOL I'm just too trusting in the documentation Thanks for reminding me!

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)