Mandelbrot Benchmark FPU/SSE2 released

Index > Windows > Mandelbrot Benchmark FPU/SSE2 released

Goto page Previous 1, 2, 3 ... 7, 8, 9 ... 18, 19, 20 Next

Author

Thread

Kuemmel

Joined: 30 Jan 2006
Posts: 200
Location: Stuttgart, Germany

Kuemmel 16 Dec 2007, 16:32

Hi, for all the people who get lost in all the numbers here...I made a graph showing the impressive improvements with the 64bit versions in efficiency per MHz per Core...of course the current 32bit version is sub-optimal and only comparable to the X1 64 version. It shows how good Core 2 Duo architecture really works here:

16 Dec 2007, 16:32

Madis731

Joined: 25 Sep 2003
Posts: 2139
Location: Estonia

Madis731 16 Dec 2007, 18:18

I can see the same patterns here like in Tom's and Anand's articles:
http://www.tomshardware.com/2007/11/19/the_spider_weaves_its_web/page22.html
http://www.anandtech.com/showdoc.aspx?i=3153&p=6

Like 'come again!?!?' AMD Phenom 9900 (@2.6GHz) is beaten by Intel Core 2 Quad Q6600 (@2.4GHz). Wait a minute. Even in power consumption?
...
Yes - even in that - http://www.anandtech.com/showdoc.aspx?i=3153&p=10

16 Dec 2007, 18:18

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20299
Location: In your JS exploiting you and your system

revolution 16 Dec 2007, 18:23

I think the AMD execs will be crying over Phenom. It was supposed to be something special, instead it is something below average.

16 Dec 2007, 18:23

Madis731

Joined: 25 Sep 2003
Posts: 2139
Location: Estonia

Madis731 16 Dec 2007, 18:30

I wouldn't say average right away - its actually the next step from them and its rather good - just Intel did something even more remarkable Very Happy

16 Dec 2007, 18:30

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20299
Location: In your JS exploiting you and your system

revolution 16 Dec 2007, 18:36

There's only two players, AMD and Intel, one is below ave and the other above ave Wink

AMD need to lower their price to match the performance else no one will buy!

16 Dec 2007, 18:36

Kuemmel

Joined: 30 Jan 2006
Posts: 200
Location: Stuttgart, Germany

Kuemmel 19 Dec 2007, 18:42

@Xorpd!
I recently looked again at the quickman-code, it seems to me that he counts iterations in a different way. We only add iterations when a point is not diverged, but I think the quickman-code still adds iteration counts if one of the two points in an SSE register is already diverged...it can be quite a difference, so in the end he will get more iteration counts than us and finally a higher speed value...what wouldn't make it comparable...got some time to check his code also ?...I'm not 100 % sure, though.

Okay, in the end it's also a matter of definition. In a non SSE code you would stop iteration in an SSE code you wouldn't due to speed, though it can be still seen as a computed but useless iteration...

19 Dec 2007, 18:42

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20299
Location: In your JS exploiting you and your system

revolution 19 Dec 2007, 19:53

Here are some nice high-precision Madelbrot images. If you haven't seen them before you will be amazed.

Deepzooming with Fractint
Fractint Fractal Gallery

19 Dec 2007, 19:53

Xorpd!

Joined: 21 Dec 2006
Posts: 161

Xorpd! 19 Dec 2007, 22:00

Oh well, I composed a big long reply and the forum software seems to have timed me out. What a waste of time.

19 Dec 2007, 22:00

f0dder

Joined: 19 Feb 2004
Posts: 3175
Location: Denmark

f0dder 20 Dec 2007, 14:47

Xorpd!: use FireFox and the It's All Text! extension (possibly with a bbcode-knowing editor) to avoid being bitten by timeouts and crashes. It's worth it Smile

20 Dec 2007, 14:47

bitRAKE

Joined: 21 Jul 2003
Posts: 4020
Location: vpcmpistri

bitRAKE 21 Dec 2007, 05:34

The DDRAW.INC file in the current version of FASM is different: (Un)LockSurface has changed to (Un)Lock. A major change is in the removal of the RET macro - which was just plain crazy in the first place, iyam.

(Sorry, I have not read the whole thread - maybe, this has been stated already. Having multiple include paths for the same files is not my idea of fun - recipe for error actually.)

I don't believe it fair to compare a single core running 16 threads to multiple cores running 16 threads. Unless the goal is to see how bad single cores perform on software designed for multiple cores. :-/

21 Dec 2007, 05:34

bitRAKE

Joined: 21 Jul 2003
Posts: 4020
Location: vpcmpistri

bitRAKE 21 Dec 2007, 07:44

I changed the frame_counter_loop:

Code:

NUMBER_OF_THREADS = 2

frame_counter_loop:
    stdcall Frame_Loop, NUMBER_OF_THREADS

...

Frame_Loop:
        label .threads dword at esp+4

        xor ebx,ebx
        mov esi,[.threads]
        lea edi,[threadhandles]
.0:
        if ALGO=0
          invoke  CreateThread,NULL, 0, thread_draw_fpu, ebx, \
                  REALTIME_PRIORITY_CLASS or CREATE_SUSPENDED, tId
        else if ALGO=1
          invoke  CreateThread,NULL, 0, thread_draw_sse2, ebx, \
                  REALTIME_PRIORITY_CLASS or CREATE_SUSPENDED, tId
        else if ALGO=2
          invoke  CreateThread,NULL, 0, thread_draw_sse3, ebx, \
                  REALTIME_PRIORITY_CLASS or CREATE_SUSPENDED, tId
        else if ALGO=3
          invoke  CreateThread,NULL, 0, thread_draw_sse3_vodnaya, ebx, \
                  REALTIME_PRIORITY_CLASS or CREATE_SUSPENDED, tId
        end if

        mov [edi],eax
        add edi,4
        add ebx,LINE_INTERLEAVE
        dec esi
        jne .0
        mov [current_line],ebx ; start drawing here next

        mov ebx,[.threads]
        inc esi ; affinity mask
        lea edi,[threadhandles]
.1:     invoke  SetThreadAffinityMask,dword [edi],esi
        shl esi,1
        add edi,4
        dec ebx
        jne .1

        mov ebx,[.threads]
        lea edi,[threadhandles]
.2:     invoke  ResumeThread,dword [edi]
        add edi,4
        dec ebx
        jne .2

        invoke WaitForMultipleObjectsEx,[.threads],threadhandles,1,-1,0

        mov ebx,[.threads]
        lea edi,[threadhandles]
.3:     invoke  CloseHandle, dword [edi]
        add edi,4
        dec ebx
        jne .3

        retn 4

Virtually no change in performance, but the program is smaller. Laughing

Seems there is a slight improvement (+0.25% Razz

) when the number of threads is one greater than the number of cores.

Is this correct in the orginal code?

Code:

mov [current_line], LINE_INTERLEAVE*15 ; start drawing here next

Shouldn't it be *16?

21 Dec 2007, 07:44

Kuemmel

Joined: 30 Jan 2006
Posts: 200
Location: Stuttgart, Germany

Kuemmel 21 Dec 2007, 10:45

bitRAKE wrote:

Virtually no change in performance, but the program is smaller. Seems there is a slight improvement (+0.25% ) when the number of threads is one greater than the number of cores.

Is this correct in the orginal code?
Code:
mov [current_line], LINE_INTERLEAVE*15 ; start drawing here next    
Shouldn't it be *16?

Ups, may be, may be not, didn't look at the code for more than a year...actually, the 0.53 version from me can't be seen as a masterpeice of coding Wink

It was my first x86 assembly programming and involved lots of help here from others and lots of copy and paste from examples...so quite messy...

Thanks a lot for your help, so it seems you modified it and compiled it may be with the lates FASM version ? Could you attach the file here ? Would be nice, it could be a new start version for me to enhance it !

21 Dec 2007, 10:45

bitRAKE

Joined: 21 Jul 2003
Posts: 4020
Location: vpcmpistri

bitRAKE 21 Dec 2007, 17:03

There was another bug in the stack frame I suggested. This file has ALGO=1 changed to new frame (that was good for more than +1% Laughing

). Few other changes, too.

Lots of room for improvement, imho.

Description:		Download
Filename:	KMB_V0.53_MT(bitRAKE).asm
Filesize:	39.16 KB
Downloaded:	292 Time(s)

21 Dec 2007, 17:03

Xorpd!

Joined: 21 Dec 2006
Posts: 161

Xorpd! 24 Dec 2007, 06:23

A couple of random notes:

The phenom reviews had a stupid version of memory bandwidth test. See http://www.bit-tech.net/hardware/2007/12/13/amd_phenom_9500_9600_9700_and_9900/5
Firefox is pitiful. I looked at my web page on a friend's computer and it didn't render my table (updated, BTW) for beans.
[current_line] refers to the highest-addressed line currently being rendered. Thus the code is good, comment a little fuzzy.
Room for improvement for sure. Demonstrably 2X at the least, as demonstrated by quickman and some of my 64-bit versions. Some clear bugs in the Windows stuff, maybe someone with more experience can offer some detailed criticisms.

Saving this in notepad before posting...

24 Dec 2007, 06:23

asmfan

Joined: 11 Aug 2006
Posts: 392
Location: Russian

asmfan 25 Dec 2007, 09:38

2 bitRAKE - algo 4 fails with error at this point:

Code:

        ; get color word
        shl             ecx, 2
        mov             eax, [edx+ecx]
        mov             [ebx], eax ; (!!!) exactly this

WinXP 32bit sp2+

_________________
Any offers?

25 Dec 2007, 09:38

madmatt

Joined: 07 Oct 2003
Posts: 1045
Location: Michigan, USA

madmatt 25 Dec 2007, 10:15

asmfan wrote:

2 bitRAKE - algo 4 fails with error at this point:

Code:

        ; get color word
        shl             ecx, 2
        mov             eax, [edx+ecx]
        mov             [ebx], eax ; (!!!) exactly this

WinXP 32bit sp2+

same thing happens to me too, here is a crash dump from DrWatson:
(Look for the word 'FAULT' to know where the crash occurred, around the center of the listing)

Code:

Application exception occurred:
        App: C:\Documents and Settings\Matt Childress\My Documents\Matts Stuff\Assembly Code Examples\KMB_V0.53_MT(bitRAKE).exe (pid=4016)
        When: 12/25/2007 @ 05:10:48.046
        Exception number: c0000005 (access violation)

*----> State Dump for Thread Id 0x920 <----*

eax=00d4d484 ebx=89f2a890 ecx=00000000 edx=004036ab esi=000003a0 edi=00000020
eip=0040157f esp=007eff00 ebp=007effa4 iopl=0         nv up ei pl zr na po nc
cs=001b  ss=0023  ds=0023  es=0023  fs=003b  gs=0000             efl=00000246

function: KMB_V0.53_MT(bitRAKE)
        00401464 c9               leave
        00401465 c21000           ret     0x10
        00401468 90               nop
        00401469 90               nop
        0040146a 90               nop
        0040146b 90               nop
        0040146c 90               nop
        0040146d 90               nop
        0040146e 90               nop
        0040146f 90               nop
        00401470 90               nop
        00401471 90               nop
        00401472 90               nop
        00401473 90               nop
        00401474 90               nop
        00401475 90               nop
        00401476 90               nop
        00401477 90               nop
        00401478 90               nop
        00401479 90               nop
        0040147a 90               nop
        0040147b 90               nop
        0040147c 90               nop
        0040147d 90               nop
        0040147e 90               nop
        0040147f 90               nop
        00401480 8d6c24ec         lea     ebp,[esp-0x14]
        00401484 81ec93000000     sub     esp,0x93
        0040148a 896d0c           mov     [ebp+0xc],ebp
        0040148d 897d08           mov     [ebp+0x8],edi
        00401490 897504           mov     [ebp+0x4],esi
        00401493 895d00           mov     [ebp],ebx
        00401496 83e4c0           and     esp,0xffffffc0
        00401499 a19e364000       mov   eax,[KMB_V0.53_MT(bitRAKE)+0x369e (0040369e)]
        0040149e 6bc018           imul    eax,eax,0x18
        004014a1 dd80c0314000     fld     qword ptr [eax+0x4031c0]
        004014a7 dd80c8314000     fld     qword ptr [eax+0x4031c8]
        004014ad dd80d0314000     fld     qword ptr [eax+0x4031d0]
        004014b3 dd5c2410         fstp    qword ptr [esp+0x10]
        004014b7 dd5c2408         fstp    qword ptr [esp+0x8]
        004014bb dd1c24           fstp    qword ptr [esp]
        004014be dd442408         fld     qword ptr [esp+0x8]
        004014c2 dc2424           fsub    qword ptr [esp]
        004014c5 d9e1             fabs
        004014c7 da3530314000 fidiv dword ptr [KMB_V0.53_MT(bitRAKE)+0x3130 (00403130)]
        004014cd dd5c2418         fstp    qword ptr [esp+0x18]
        004014d1 c744243800000000 mov     dword ptr [esp+0x38],0x0
        004014d9 baab364000       mov     edx,0x4036ab
        004014de 8b1db8304000     mov   ebx,[KMB_V0.53_MT(bitRAKE)+0x30b8 (004030b8)]
        004014e4 81c390010000     add     ebx,0x190
        004014ea 8b35a4304000     mov   esi,[KMB_V0.53_MT(bitRAKE)+0x30a4 (004030a4)]
        004014f0 8b4514           mov     eax,[ebp+0x14]
        004014f3 0fafc6           imul    eax,esi
        004014f6 01c3             add     ebx,eax
        004014f8 81ee60090000     sub     esi,0x960
        004014fe 8b4514           mov     eax,[ebp+0x14]
        00401501 83c001           add     eax,0x1
        00401504 89442434         mov     [esp+0x34],eax
        00401508 dd442418         fld     qword ptr [esp+0x18]
        0040150c da4d14           fimul   dword ptr [ebp+0x14]
        0040150f dc442410         fadd    qword ptr [esp+0x10]
        00401513 dd5c2420         fstp    qword ptr [esp+0x20]
        00401517 c744243000000000 mov     dword ptr [esp+0x30],0x0
        0040151f ded9             fcompp
        00401521 ded9             fcompp
        00401523 ded9             fcompp
        00401525 dd0520314000 fld qword ptr [KMB_V0.53_MT(bitRAKE)+0x3120 (00403120)]
        0040152b dd442418         fld     qword ptr [esp+0x18]
        0040152f b860090000       mov     eax,0x960
        00401534 da4c2430         fimul   dword ptr [esp+0x30]
        00401538 dc0424           fadd    qword ptr [esp]
        0040153b dd442420         fld     qword ptr [esp+0x20]
        0040153f d9c1             fld     st(1)
        00401541 d9c1             fld     st(1)
        00401543 d9c1             fld     st(1)
        00401545 d8ca             fmul    st,st(2)
        00401547 d9c1             fld     st(1)
        00401549 d8ca             fmul    st,st(2)
        0040154b d9ca             fxch    st(2)
        0040154d decb             fmulp   st(3),st
        0040154f d9c0             fld     st(0)
        00401551 d8c2             fadd    st,st(2)
        00401553 d9cb             fxch    st(3)
        00401555 d8c0             fadd    st,st(0)
        00401557 d9ca             fxch    st(2)
        00401559 dee9             fsubrp  st(1),st
        0040155b d9ca             fxch    st(2)
        0040155d dff5             fcomip  st,st(5)
        0040155f d8c2             fadd    st,st(2)
        00401561 d9c9             fxch    st(1)
        00401563 d8c3             fadd    st,st(3)
        00401565 d9c9             fxch    st(1)
        00401567 7705             ja      KMB_V0.53_MT(bitRAKE)+0x156e (0040156e)
        00401569 83e801           sub     eax,0x1
        0040156c 75d5             jnz     KMB_V0.53_MT(bitRAKE)+0x1543 (00401543)
        0040156e b960090000       mov     ecx,0x960
        00401573 29c1             sub     ecx,eax
        00401575 014c2438         add     [esp+0x38],ecx
        00401579 c1e102           shl     ecx,0x2
        0040157c 8b040a           mov     eax,[edx+ecx]
FAULT ->0040157f 8903             mov     [ebx],eax         ds:0023:89f2a890=????????
        00401581 8344243001       add     dword ptr [esp+0x30],0x1
        00401586 83c304           add     ebx,0x4
        00401589 817c243058020000 cmp     dword ptr [esp+0x30],0x258
        00401591 758c             jnz     KMB_V0.53_MT(bitRAKE)+0x151f (0040151f)
        00401593 83451401         add     dword ptr [ebp+0x14],0x1
        00401597 01f3             add     ebx,esi
        00401599 8b442434         mov     eax,[esp+0x34]
        0040159d 394514           cmp     [ebp+0x14],eax
        004015a0 0f8562ffffff     jne     KMB_V0.53_MT(bitRAKE)+0x1508 (00401508)
        004015a6 b801000000       mov     eax,0x1
        004015ab f00fc105a6364000 lock xadd [KMB_V0.53_MT(bitRAKE)+0x36a6 (004036a6)],eax
        004015b3 83c001           add     eax,0x1
        004015b6 894514           mov     [ebp+0x14],eax
        004015b9 3d58020000       cmp     eax,0x258
        004015be 0f8c1affffff     jl      KMB_V0.53_MT(bitRAKE)+0x14de (004014de)
        004015c4 8b442438         mov     eax,[esp+0x38]
        004015c8 f0010528314000 lock add [KMB_V0.53_MT(bitRAKE)+0x3128 (00403128)],eax
        004015cf 8b6d0c           mov     ebp,[ebp+0xc]
        004015d2 8b7d08           mov     edi,[ebp+0x8]
        004015d5 8b7504           mov     esi,[ebp+0x4]
        004015d8 8b5d00           mov     ebx,[ebp]
        004015db 8d6514           lea     esp,[ebp+0x14]
        004015de c20400           ret     0x4
        004015e1 0000             add     [eax],al
        004015e3 0000             add     [eax],al
        004015e5 0000             add     [eax],al
        004015e7 0000             add     [eax],al
        004015e9 0000             add     [eax],al
        004015eb 0000             add     [eax],al
        004015ed 0000             add     [eax],al
        004015ef 0000             add     [eax],al
        004015f1 0000             add     [eax],al
        004015f3 0000             add     [eax],al
        004015f5 0000             add     [eax],al
        004015f7 0000             add     [eax],al
        004015f9 0000             add     [eax],al
        004015fb 0000             add     [eax],al
        004015fd 0000             add     [eax],al
        004015ff 0000             add     [eax],al
        00401601 0000             add     [eax],al
        00401603 0000             add     [eax],al
        00401605 0000             add     [eax],al
        00401607 0000             add     [eax],al
        00401609 0000             add     [eax],al
        0040160b 0000             add     [eax],al
        0040160d 0000             add     [eax],al
        0040160f 0000             add     [eax],al
        00401611 0000             add     [eax],al
        00401613 0000             add     [eax],al
        00401615 0000             add     [eax],al
        00401617 0000             add     [eax],al
        00401619 0000             add     [eax],al
        0040161b 0000             add     [eax],al
        0040161d 0000             add     [eax],al
        0040161f 0000             add     [eax],al
        00401621 0000             add     [eax],al
        00401623 0000             add     [eax],al
        00401625 0000             add     [eax],al
        00401627 0000             add     [eax],al
        00401629 0000             add     [eax],al
        0040162b 0000             add     [eax],al
        0040162d 0000             add     [eax],al
        0040162f 0000             add     [eax],al
        00401631 0000             add     [eax],al
        00401633 0000             add     [eax],al
        00401635 0000             add     [eax],al
        00401637 0000             add     [eax],al
        00401639 0000             add     [eax],al
        0040163b 0000             add     [eax],al
        0040163d 0000             add     [eax],al
        0040163f 0000             add     [eax],al
        00401641 0000             add     [eax],al
        00401643 0000             add     [eax],al
        00401645 0000             add     [eax],al
        00401647 0000             add     [eax],al
        00401649 0000             add     [eax],al
        0040164b 0000             add     [eax],al
        0040164d 0000             add     [eax],al
        0040164f 0000             add     [eax],al
        00401651 0000             add     [eax],al
        00401653 0000             add     [eax],al
        00401655 0000             add     [eax],al
        00401657 0000             add     [eax],al
        00401659 0000             add     [eax],al
        0040165b 0000             add     [eax],al
        0040165d 0000             add     [eax],al
        0040165f 0000             add     [eax],al
        00401661 0000             add     [eax],al
        00401663 0000             add     [eax],al
        00401665 0000             add     [eax],al
        00401667 0000             add     [eax],al
        00401669 0000             add     [eax],al
        0040166b 0000             add     [eax],al
        0040166d 0000             add     [eax],al
        0040166f 0000             add     [eax],al
        00401671 0000             add     [eax],al
        00401673 0000             add     [eax],al
        00401675 0000             add     [eax],al
        00401677 0000             add     [eax],al
        00401679 0000             add     [eax],al

25 Dec 2007, 10:15

bitRAKE

Joined: 21 Jul 2003
Posts: 4020
Location: vpcmpistri

bitRAKE 25 Dec 2007, 16:07

need to add a line to thread setup: dec ebx.

Code:

        mov [edi],eax
        add edi,4
        add ebx,LINE_INTERLEAVE
        dec esi
        jne .0
        dec ebx ;<--------
        mov [current_line],ebx

As stated above I misunderstood the code comment and erroneously advanced the line index. But I don't see how this is the source of the crash as it causes a line to be skipped and not advancement beyond the screen buffer. Confused

ALGO=1 is the only one I made changes to directly.

25 Dec 2007, 16:07

Kuemmel

Joined: 30 Jan 2006
Posts: 200
Location: Stuttgart, Germany

Kuemmel 26 Dec 2007, 09:15

Here we go,

I made an optimized version trying to implement some ideas from Quickman and Xorpd for SSE2. My basic thought was to compute 4 points in one iteration loop with independet instruction lines, like Quickman, with minimizing local variables for intermediates (thanks to Bitrake for the help). I found that at least at the Core 2 Duo interleaving instruction by instruction pays less of than interleaving bigger blocks...the CPU seems to do the job by itself, I guess, lots of try and error at least for me...

Still way to go to catch up with Xorpds 64bit X4 version, and I still got to see how the lack of registers for the 32bit version can be compensated or not...at the moment it looks quite awfull regarding memory access but pays off, of course I think for sure that's not the end of optimization...here's the result:

KMB_0.53_SSE2: 520,206 Million Iterations/s
KMB_0.53B_SSE2: 817,714 Million Iterations/s
(Core 2 Duo, 1,87 GHz)

...so at least a gain of about 57 %. The code is attached (also the include files, in case of problems compiling)...any comments welcome !!!

Description:		Download
Filename:	KMB_V0.53B_MT.rar
Filesize:	144.52 KB
Downloaded:	223 Time(s)

26 Dec 2007, 09:15

Madis731

Joined: 25 Sep 2003
Posts: 2139
Location: Estonia

Madis731 26 Dec 2007, 21:59

846.045MiPS T7200 SSE2
lower-case "i" is iterations - not instructions Very Happy

PS. The "back" button seems to work most of the times - browser retains the value of the input-box. When I suspect some errors, I go back and copy the text => then check the boards that it *really* went there and:
1) forget the text in the clipboard
2) or try to post again, when it failed

26 Dec 2007, 21:59

bitRAKE

Joined: 21 Jul 2003
Posts: 4020
Location: vpcmpistri

bitRAKE 26 Dec 2007, 23:10

here are some interesting location in .log file format for Quickman.

Description:	Quickman log file	Download
Filename:	bitrake.log.txt
Filesize:	818 Bytes
Downloaded:	209 Time(s)

26 Dec 2007, 23:10

Goto page Previous 1, 2, 3 ... 7, 8, 9 ... 18, 19, 20 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