dword to ascii, ascii to dword

Index > Main > dword to ascii, ascii to dword

Goto page Previous 1, 2, 3, 4 Next

Author

Thread

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 11 Feb 2013, 07:12

Fixed in the above post.

11 Feb 2013, 07:12

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 11 Feb 2013, 09:37

Here is slightly faster version of my code (at least on my work computer Intel code duo):

Code:

johnfound_ascii2d:
        xor     eax, eax
        mov     ecx, [esp+4]
        mov     edx, eax
        mov     [esp+4], eax

        cmp     byte [ecx], '-'
        jne     .loop
        add     ecx, 1
        sub     dword [esp+4], 1
.loop:
        mov     dl, [ecx]
        inc     ecx
        test    dl, dl
        jz      .ends

        lea     eax, [5*eax]
        sub     dl, '0'
        shl     eax, 1

        add     eax, edx
        jmp     .loop

.ends:
        xor     eax, [esp+4]
        sub     eax, [esp+4]
        retn 4

11 Feb 2013, 09:37

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20451
Location: In your JS exploiting you and your system

revolution 11 Feb 2013, 11:12

JohnFound wrote:

Code:

        shl     eax, 1

        add     eax, edx

What about:

Code:

lea eax,[eax*2+edx]

11 Feb 2013, 11:12

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 11 Feb 2013, 11:17

revolution wrote:

What about:
Code:
lea eax,[eax*2+edx]    

Yes, it is better! But not faster on my tests. Sad

Strange.

_________________
Tox ID: 48C0321ADDB2FE5F644BB5E3D58B0D58C35E5BCBC81D7CD333633FEDF1047914A534256478D9

11 Feb 2013, 11:17

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20451
Location: In your JS exploiting you and your system

revolution 11 Feb 2013, 11:20

IMO testing for "faster" is rather futile for minor changes. There are far too many factors that can change the results that make it a very unsatisfying experience.

11 Feb 2013, 11:20

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 11 Feb 2013, 11:27

But this is really faster from 2620ms for the previous variant to 1545ms for this (for 100000000 conversions):

Code:

johnfound_ascii2d:
        push    esi

        xor     eax, eax
        mov     esi, [esp+8]
        mov     edx, eax
        mov     ecx, eax

        cmp     byte [esi], '-'
        jne     .loop

        add     esi, 1
        sub     ecx, 1

.loop:
        lodsb
        sub     al, '0'
        jc      .ends

        lea     edx, [5*edx]
        lea     edx, [2*edx+eax]
        jmp     .loop

.ends:
        xor     edx, ecx
        sub     edx, ecx
        mov     eax, edx
        pop     esi
        retn 4

This evening I will test on Intel Atom. I expect, the result will be different.

11 Feb 2013, 11:27

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20451
Location: In your JS exploiting you and your system

revolution 11 Feb 2013, 11:30

JohnFound wrote:

But this is really faster from 2620ms for the previous variant to 1545ms for this (for 100000000 conversions): ...

How much of that data is coming from cache and how much is fresh from disk / other source?

So many things to check before we know if it will be faster in the "real world".

11 Feb 2013, 11:30

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 11 Feb 2013, 11:33

revolution, you are right of course. But the faster algorithm is always faster. Of course, it is not so important if the bottleneck is somewhere else - for example on the disk operations. That is why I always argue that it is always better to optimize for size, not for speed.
But here it is a game - isn't it? Smile

11 Feb 2013, 11:33

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20451
Location: In your JS exploiting you and your system

revolution 11 Feb 2013, 11:38

JohnFound wrote:

But the faster algorithm is always faster.

Well that is also a myth it seems. Even theoretically asymptotically faster algorithms can be slower than naive algorithms if the input size is small. One must know the domain of the expected source data before one can conclude the "best"* algorithm to apply.

* where the definition of best is left open to interpretation.

11 Feb 2013, 11:38

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 11 Feb 2013, 11:51

Revolution, when I use the word "faster" I mean "In the same conditions" of course. Another type of comparison is impossible at all. And of course on some conditions even theoretically faster algorithm can be slower - for example Bubble sort is faster than Quick sort on very small arrays.

11 Feb 2013, 11:51

AsmGuru62

Joined: 28 Jan 2004
Posts: 1671
Location: Toronto, Canada

AsmGuru62 11 Feb 2013, 12:05

@uart777:
Why load the address into EDX?
It can be shorter:

Code:

mov cx, [table + eax*2]

11 Feb 2013, 12:05

HaHaAnonymous

Joined: 02 Dec 2012
Posts: 1178
Location: Unknown

HaHaAnonymous 11 Feb 2013, 19:59

[ Post removed by author. ]

Last edited by HaHaAnonymous on 28 Feb 2015, 21:34; edited 2 times in total

11 Feb 2013, 19:59

sleepsleep

Joined: 05 Oct 2006
Posts: 13069
Location: ˛　　　　　　　　　　　　　　　　　　　　　　　　　　　　　⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣Posts: 0010456

sleepsleep 12 Feb 2013, 00:18

ok guys,
will come back to work after 2 days holiday,

what i would do is, copy paste ur code and run it,

i would use my proc and every of ur guys submitted proc and cmp the EAX returned,

to verify that each proc return valid value,

then i will test the speed =)

will announce the winner on 16th Feb 2013

u can use MMX SSE1 SSE2 or etc if u think that could help.

*plez note that, i am not suppose to do anything with your code in case they are not in the form that i proposed, i only will copi paste and call it then time it.

12 Feb 2013, 00:18

edfed

Joined: 20 Feb 2006
Posts: 4353
Location: Now

edfed 12 Feb 2013, 01:53

what does give my atoi function?

Code:

mov esi,test
call atoi
mov [result],eax
...
test db '-1235697',0

;on return:
;value from signed decimal string at [esi] in eax
;esi will point directlly after the command.
;carry set if valid decimal number

atoi:
.seekend:
        lodsb
        case al,' ',@f
        case al,'.',@f
        case al,',',@f
        case al,'!',@f
        casne al,0,.seekend
@@:
        dec esi
        push esi
        dec esi
        xor ebx,ebx
        mov edx,1
.next:
        movzx eax,byte[esi]
        dec esi
        case al,'-',.neg
        case al,' ',.end
        case al,',',.end
        case al,'.',.end
        casl al,'0',@f
        casg al,'9',@f
        sub al,'0'
        imul eax,edx
        add ebx,eax
        imul edx,10
        jmp .next
@@:
        pop esi
        clc
        mov eax,ebx
        ret
.neg:
        mov al,[esi]
        casne al,' ',@b
        neg ebx
.end:
        pop esi
        stc
        mov eax,ebx
        ret
.null:
        xor ebx,ebx
        clc
        mov eax,ebx
        ret

need that:

Code:

macro case var,value,jmp {
        cmp var,value
        je jmp
}

macro casne var,value,jmp {
        cmp var,value
        jne jmp
}

macro casa var,value,jmp {
        cmp var,value
        ja jmp
}

macro casle var,value,jmp {
        cmp var,value
        jle jmp
}

macro casg var,value,jmp {
        cmp var,value
        jg jmp
}

macro casge var,value,jmp {
        cmp var,value
        jge jmp
}

macro casl var,value,jmp {
        cmp var,value
        jl jmp
}

12 Feb 2013, 01:53

AsmGuru62

Joined: 28 Jan 2004
Posts: 1671
Location: Toronto, Canada

AsmGuru62 12 Feb 2013, 13:32

I will also test these routines -- keep posting!

12 Feb 2013, 13:32

sleepsleep 12 Feb 2013, 14:40

if got any better method or what you see more suitable, please let us know, so everybody will get more fair assessment,

and for our future game as well

12 Feb 2013, 14:40

AsmGuru62

Joined: 28 Jan 2004
Posts: 1671
Location: Toronto, Canada

AsmGuru62 12 Feb 2013, 17:46

What if my code is the slowest!?...
First I'll measure it and then post it -- you know, not to be embarassed...

Basically, my idea is an unrolled loop.
The maximum of digits is 10, so I call my macro for converting a digit ten times.
Should be simple enough. No jumps -- well, only one jump if no more digits found.
But the measurement will actually decide if it is really fast code or not.
Maybe not... we''ll see. It is all very intriguing!

12 Feb 2013, 17:46

HaHaAnonymous

Joined: 02 Dec 2012
Posts: 1178
Location: Unknown

HaHaAnonymous 12 Feb 2013, 20:48

[ Post removed by author. ]

Last edited by HaHaAnonymous on 28 Feb 2015, 21:33; edited 2 times in total

12 Feb 2013, 20:48

AsmGuru62

Joined: 28 Jan 2004
Posts: 1671
Location: Toronto, Canada

AsmGuru62 12 Feb 2013, 23:16

I will still post my code, even if it will be slow.

12 Feb 2013, 23:16

HaHaAnonymous

Joined: 02 Dec 2012
Posts: 1178
Location: Unknown

HaHaAnonymous 12 Feb 2013, 23:25

[ Post removed by author. ]

Last edited by HaHaAnonymous on 28 Feb 2015, 21:33; edited 1 time in total

12 Feb 2013, 23:25

Goto page Previous 1, 2, 3, 4 Next

< Last Thread | Next Thread >

Forum Rules:

You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
You cannot attach files in this forum
You can download files in this forum