Author Topic: Assembly Programmers - Help Axe Optimize! (Read 154120 times)

Quigibo · « **Reply #225 on:** July 12, 2011, 02:02:39 am »

Awesome wow! Yeah, forward djnz is about as rare as cpir. Although I think calc84maniac's original 4 level routine used them as well but for a different purpose.

Also on the same subject, although you'll be the only one who knows what I'm talking about, all 12 DispGraph forms work perfectly now.

Runer112 · « **Reply #226 on:** July 12, 2011, 08:48:57 am »

Just checking, have you actually tested the routines out? Because I didn't actually test those routines I gave you, I just modeled them after some routines I knew worked and hoped these would still work as well.

Quigibo · « **Reply #227 on:** July 12, 2011, 03:33:09 pm »

Yeah I tested everything. One of them had a problem that I fixed with the buffer ordering being switched though.

calc84maniac · « **Reply #228 on:** July 14, 2011, 11:31:27 pm »

Here's a peephole optimization suggestion: Keep track of whether the value in HL is a constant or not, and if so, what constant. For example, I have some code:

Code: [Select]

If condition
do stuff
Else
16->W
End

Obviously, after the Else, HL has to be 0. Thus the 16 can be reduced to a ld l,16 instead of ld hl,16. It might be possible to auto-optimize stuff like 1->A:2->B into 1->A+1->B, but you could always leave that to the user like usual.

Also, I found it a bit annoying that when I did something like If E<(96*256), the part in the parentheses wasn't reduced to a constant before doing the less-than operation. Could the look-ahead parsing be able to detect constants in parentheses?

Runer112 · « **Reply #229 on:** July 14, 2011, 11:34:17 pm »

Quote from: calc84maniac on July 14, 2011, 11:31:27 pm

Also, I found it a bit annoying that when I did something like If E<(96*256), the part in the parentheses wasn't reduced to a constant before doing the less-than operation. Could the look-ahead parsing be able to detect constants in parentheses?

This times a million.

ztrumpet · « **Reply #230 on:** July 14, 2011, 11:35:45 pm »

Quote from: Runer112 on July 14, 2011, 11:34:17 pm

Quote from: calc84maniac on July 14, 2011, 11:31:27 pm
Also, I found it a bit annoying that when I did something like If E<(96*256), the part in the parentheses wasn't reduced to a constant before doing the less-than operation. Could the look-ahead parsing be able to detect constants in parentheses?

This times a million.

This times a million and five.

Seriously, I thought Axe did this already. Apparently not... so, please?

Runer112 · « **Reply #231 on:** July 25, 2011, 11:16:41 am »

Quigibo, you read my mind. I was about to make a post with code for commands that deal with archived variables to work with variables in RAM too, but you added that in Axe 1.0.2 before I could finish! However, I'll make a post anyways because my p_GetArc routine is smaller.

I also have a few other things.

p_GetArc: 7 bytes smaller.

Code: (Old code: 76 bytes) [Select]

p_GetArc:
	.db __GetArcEnd-1-$
	push	de
	MOV9TOOP1()
	B_CALL(_ChkFindSym)
	jr	c,__GetArcFail
	dec	b
	inc	b
	jr	z,__GetArcRam
	B_CALL(_IsFixedName)
	ld	hl,9
	jr	z,__GetArcName
__GetArcStatic:
	ld	l,12
	and	%00011111
	jr	z,__GetArcDone
	cp	l
	jr	z,__GetArcDone
	ld	l,14
	jr	__GetArcDone
__GetArcName:
	add	hl,de
	bit	7,h
	jr	z,$+7
	res	7,h
	set	6,h
	inc	b
	B_CALL(_LoadDEIndPaged)
	ld	d,0
	inc	e
	inc	e
__GetArcDone:
	add	hl,de
	ex	de,hl
__GetArcStore:
	pop	hl
	ld	(hl),e
	inc	hl
	ld	(hl),d
	inc	hl
	ld	(hl),b
	ex	de,hl
	ret
__GetArcRam:
	and	%00011111
	jr	z,__GetArcStore
	cp	CplxObj
	jr	z,__GetArcStore
	inc	de
	inc	de
	jr	__GetArcStore
__GetArcFail:
	ld	hl,0
	pop	de
	ret
__GetArcEnd:

Code: (New code: 69 bytes) [Select]

p_GetArc:
	.db __GetArcEnd-1-$
	push	de
	MOV9TOOP1()
	B_CALL(_ChkFindSym)
	jr	c,__GetArcFail
	dec	b
	inc	b
	jr	z,__GetArcRam
	B_CALL(_IsFixedName)
	ld	hl,9
	jr	z,__GetArcName
	ld	l,12
__GetArcChkFloat:
	and	%00011111
	jr	z,__GetArcDone
	cp	CplxObj
	jr	z,__GetArcDone
	inc	l
	inc	l
	jr	__GetArcDone
__GetArcName:
	add	hl,de
	bit	7,h
	jr	z,$+7
	res	7,h
	set	6,h
	inc	b
	B_CALL(_LoadDEIndPaged)
	ld	d,0
	inc	e
	inc	e
__GetArcDone:
	add	hl,de
	ex	de,hl
	pop	hl
	ld	(hl),e
	inc	hl
	ld	(hl),d
	inc	hl
	ld	(hl),b
	ex	de,hl
	ret
__GetArcRam:
	ld	h,b
	ld	l,b
	jr	__GetArcChkFloat
__GetArcFail:
	ld	hl,0
	pop	de
	ret
__GetArcEnd:

p_ReadArc: Bumping an old request for larger but drastically faster archive reading routines. The routines would need to modified slightly to allow for reading from RAM as well, but that should be no problem. I would understand if you didn't want to add the app version, but the program version is immensely better in my opinion.

Quote from: Runer112 on June 10, 2011, 11:20:26 pm

And on the topic of stuff that involves port 6, I think it would be nice if the archive byte reading routine avoided using a B_CALL for a massive speed boost, especially for code compiled as programs:

p_ReadArc: 18 bytes (2x) larger, but ~1400 cycles (!!!10x!!!) faster

Code: (36 bytes, ~142 cycles) [Select]
p_ReadArc: .db __ReadArcEnd-1-$ ld c,a in a,(6) ld b,a ld a,h set 6,h res 7,h rlca rlca dec a and %00000011 add a,c out (6),a ld c,(hl) inc hl bit 7,h jr z,__ReadArcNoBoundary set 6,h res 7,h inc a out (6),a __ReadArcNoBoundary: ld l,(hl) ld h,c ld a,b out (6),a ret __ReadArcEnd:
p_ReadArcApp: 36 bytes (3x) larger, but ~1050 cycles (4x) faster

Code: (54 bytes, ~396 cycles) [Select]
p_ReadArcApp: .db __ReadArcAppEnd-1-$ push hl ld hl,$0000 ld de,ramCode ld bc,__ReadArcAppRamCodeEnd-__ReadArcAppRamCode ldir pop hl ld e,a ld c,6 in b,(c) ld a,h set 6,h res 7,h rlca rlca dec a and %00000011 add a,e call ramCode ld e,d inc hl bit 7,h jr z,__ReadArcAppNoBoundary set 6,h res 7,h inc a __ReadArcAppNoBoundary: call ramCode ex de,hl ret __ReadArcAppEnd: .db rp_Ans,__ReadArcAppEnd-p_ReadArcApp-3 __ReadArcAppRamCode: out (6),a ld d,(hl) out (c),b ret __ReadArcAppRamCodeEnd:

p_CopyArc: Modified to allow for sources in RAM.

Code: (Old code: 22 bytes) [Select]

p_CopyArc:
	.db __CopyArcEnd-1-$
	pop	ix
	pop	de
	ex	(sp),hl
	ld	b,a
	ld	a,h
	rlca
	rlca
	dec	a
	and	%00000011
	add	a,b
	set	6,h
	res	7,h
	pop	bc
	B_CALL(_FlashToRAM)
	jp	(ix)
__CopyArcEnd:

Code: (New code: 28 bytes) [Select]

p_CopyArc:
	.db __CopyArcEnd-1-$
	ex	(sp),hl
	pop	bc
	pop	de
	ex	(sp),hl
	or	a
	jr	z,__CopyArcRam
	push	bc
	ld	b,a
	ld	a,h
	rlca
	rlca
	dec	a
	and	%00000011
	add	a,b
	set	6,h
	res	7,h
	pop	bc
	B_CALL(_FlashToRAM)
	ret
__CopyArcRam:
	ldir
	ret
__CopyArcEnd:

Also, I'm not sure why I just realized this now, but why don't the 8-bit logic operations on variables just load the variable into a instead of de to save 2 bytes?

Quigibo · « **Reply #232 on:** July 26, 2011, 06:04:13 am »

Hmm, I'm still not sure if the extra speed is worth the size increase. I guess a new argument for the speed is to make file reads more consistent (a program using a file from archive might run slower than one reading from ram). But I will put this up in the poll since I'd like to know how may people this would benefit or hurt.

The 8-bit logical operators I don't do that optimization because then I'd need duplicate commands and have even more special casing. This is something that can easily be peephole optimized in the future however so it might become a non-issue.

I was trying to recursively parse constants in parenthesis in the last update, but it was extremely complicated so I gave up. I will have to modify the core number reading system to get it to work (which I was planning to do eventually anyway) so I will get to it then.

Quigibo · « **Reply #233 on:** July 27, 2011, 02:47:08 pm »

Wait.... Runer! What were you thinking! The App code for file reading can be the same as the program code, but just use port 7 instead of port 6 and set the high bits of hl for the $8000-$BFFF range. That's what the Axe app does.

EDIT: Also, another thing that the routines would need is to disable interrupts and then restore them afterwards... which I can use the "Safety" code for, but its going to be slower and even larger.

Runer112 · « **Reply #234 on:** July 27, 2011, 10:47:25 pm »

The app version might need to disable interrupts, but why would the program version need to? Both Axe's and the OS's interrupt handlers back up the page in the $4000-$7FFF bank and restore it upon returning.

Xeda112358 · « **Reply #235 on:** August 11, 2011, 01:10:51 pm »

I was toying around with some math routines while I was away and I was curious about the square root algorithms. Are the designed to return the square root rounded down, up, or just rounded? If it is rounded down and you want to round it to the nearest integer answer, here is a code I made a while ago (it isn't even close to what Axe needs, but it should only be taken as an example):

Code: [Select]

;===============================================================
sqrtE:
;===============================================================
;Input:
;     E is the value to find the square root of
;Outputs:
;     A is E-D^2
;     B is 0
;     D is the rounded result
;     E is not changed
;     HL is not changed
;Destroys:
;     C
;
        xor a               ;1      4         4
        ld d,a              ;1      4         4
        ld c,a              ;1      4         4
        ld b,4              ;2      7         7
sqrtELoop:
        rlc d               ;2      8        32
        ld c,d              ;1      4        16
        scf                 ;1      4        16
        rl c                ;2      8        32

        rlc e               ;2      8        32
        rla                 ;1      4        16
        rlc e               ;2      8        32
        rla                 ;1      4        16

        cp c                ;1      4        16
        jr c,$+4            ;4    12|15      48+3x
          inc d             ;--    --        --
          sub c             ;--    --        --
        djnz sqrtELoop      ;2    13|8       47
        cp d                ;1      4         4
        jr c,$+3            ;3    12|11     12|11
          inc d             ;--    --        --
        ret                 ;1     10        10
;===============================================================
;Size  : 29 bytes
;Speed : 347+3x cycles plus 1 if rounded down
;   x is the number of set bits in the result.
;===============================================================

The only reason that I mention this is that I know a lot of graphical algorithms would have better results if the square root was returned in rounded form as opposed to just rounded up or down.

Sorry if this was already covered and I missed it

Spoiler For Spoiler:

EDIT:

Quote from: Runer112 on July 12, 2011, 01:05:00 am

Wow, I just did something I didn't think was even possible. I found a good use for a forward djnz.

>.> Hehe, I use forward djnz in many-- if not most-- of my programs... It is one of the most useful tricks I use and is kind of my signature touch

I use it to save time and memory a lot, especially in instances like this:

Code: [Select]

     ld b,a
     or a \ jr nz,Next1
       ;code
Next1:
     djnz Next2
       ;code
Next2:
     djnz Next3
       ;code
;...et cetera

* Xeda112358 loves it

Runer112 · « **Reply #236 on:** August 11, 2011, 04:29:36 pm »

All of Axe math simply truncates, so I think the current square root algorithm is pretty good. Anyways you have to remember that Axe uses 16-bit math and that's an 8-bit square root function.

Xeda112358 · « **Reply #237 on:** August 16, 2011, 07:39:41 pm »

Yeah, I know, but I just wanted to give an example. It is really only the last few bytes that are important, though, and I wanted to give a simple, easy to follow example. Also, great job with the optimisations

I wish I could help more, but most of the codes are a bit beyond my optimisation abilities.

Runer112 · « **Reply #238 on:** August 30, 2011, 02:45:57 am »

I think I just performed the most ridiculous, impressive optimization I've ever performed on an Axe command. 27 bytes optimized down to 60% of its size: 17 bytes!

Code: (Old code: 27 bytes, ~220.5 cycles) [Select]

p_DKeyVar:
	.db __DKeyVarEnd-1-$
	dec	l
	ld	a,l
	rra
	rra
	rra
	and	%00000111
	inc	a
	ld	b,a
	ld	a,%01111111
	rlca
	djnz	$-1
	ld	h,a
	ld	a,l
	and	%00000111
	inc	a
	ld	b,a
	ld	a,%10000000
	rlca
	djnz	$-1
	ld	l,a
	ret
__DKeyVarEnd:

Code: (New code: 17 bytes, ~225 cycles) [Select]

p_DKeyVar:
	.db __DKeyVarEnd-1-$
	ld	a,l
	ld	hl,%0111111111110111
	rlc	h
	adc	a,l
	jr	c,$-3
	ld	l,%0000001
	rrc	l
	inc	a
	jr	nz,$-3
	ret
__DKeyVarEnd:

Quigibo · « **Reply #239 on:** August 30, 2011, 03:02:02 am »

O_O I don't even understand what's going on here. That's quite impressive!

EDIT: Also, a really obvious optimization I just noticed is that the return should be replaced by a jump to the direct key command so it doesn't have to return and re-call it.

Author Topic: Assembly Programmers - Help Axe Optimize! (Read 154120 times)

Quigibo

Re: Assembly Programmers - Help Axe Optimize!

Runer112

Re: Assembly Programmers - Help Axe Optimize!

Quigibo

Re: Assembly Programmers - Help Axe Optimize!

calc84maniac

Re: Assembly Programmers - Help Axe Optimize!

Runer112

Re: Assembly Programmers - Help Axe Optimize!

ztrumpet

Re: Assembly Programmers - Help Axe Optimize!

Runer112

Re: Assembly Programmers - Help Axe Optimize!

Quigibo

Re: Assembly Programmers - Help Axe Optimize!

Quigibo

Re: Assembly Programmers - Help Axe Optimize!

Runer112

Re: Assembly Programmers - Help Axe Optimize!

Xeda112358

Re: Assembly Programmers - Help Axe Optimize!

Runer112

Re: Assembly Programmers - Help Axe Optimize!

Xeda112358

Re: Assembly Programmers - Help Axe Optimize!

Runer112

Re: Assembly Programmers - Help Axe Optimize!

Quigibo

Re: Assembly Programmers - Help Axe Optimize!