Message Pump Optimization

Index > Windows > Message Pump Optimization

Author

Thread

bitshifter

Joined: 04 Dec 2007
Posts: 796
Location: Massachusetts, USA

bitshifter 27 Jun 2010, 10:33

I built this one for pure speed Razz

Also i was tired of seeing MSG struct in data section...

Code:

        sub     esp,28          ; allocate MSG struct
        mov     esi,esp           ; ESI -> MSG struct
        push    PM_REMOVE       ; PeekMessage::wRemoveMsg
        push    0                  ; PeekMessage::wMsgFilterMax
        push    0                  ; PeekMessage::wMsgFilterMin
        push    NULL             ; PeekMessage::hWnd
        push    esi            ; PeekMessage::lpMsg
    MainLoop:
        call    [PeekMessage]
        sub     esp,20  ; re-allocate PeekMessage params
        test    eax,eax ; PeekMessage returns zero if queue is empty
        jz      RenderFrame ; jumps back into MainLoop
        cmp     dword[esi+4],WM_QUIT ; MSG::message
        je      ExitLoop
       ;invoke  TranslateMessage,esi
        invoke  DispatchMessage,esi
        jmp     MainLoop
    ExitLoop:
        add     esp,28+20 ; release MSG struct and PeekMessage params

I think it kicks ass, what do you think?

_________________
Coding a 3D game engine with fasm is like trying to eat an elephant,
you just have to keep focused and take it one 'byte' at a time.

Last edited by bitshifter on 29 Jun 2010, 16:15; edited 1 time in total

27 Jun 2010, 10:33

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20520
Location: In your JS exploiting you and your system

revolution 27 Jun 2010, 10:37

How do you know if it is faster than other code?

Have you seen the code executed by the OS when you call PeekMessage?

You have no delay in your loop so it will use full CPU resources while achieving nothing. Perhaps you should consider giving time to other tasks to do the things that you are waiting for?

27 Jun 2010, 10:37

bitshifter

Joined: 04 Dec 2007
Posts: 796
Location: Massachusetts, USA

bitshifter 27 Jun 2010, 10:40

It has to be faster since we are using less instructions from the inner loop.
And yeah, it was meant to be a total resource hog Razz

27 Jun 2010, 10:40

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20520
Location: In your JS exploiting you and your system

revolution 27 Jun 2010, 10:49

bitshifter wrote:

It has to be faster since we are using less instructions from the inner loop.

That is a false assumption. You have to test it to see if it really is faster. Although I am unsure just how you will test that.

bitshifter wrote:

And yeah, it was meant to be a total resource hog

So the system slows down and you think that makes it faster? I am confused.

27 Jun 2010, 10:49

bitshifter

Joined: 04 Dec 2007
Posts: 796
Location: Massachusetts, USA

bitshifter 27 Jun 2010, 11:03

Basically i traded pushing 5 parameters each iteration for one subtraction.
In my book, thats always faster...
Its for a single threaded demo loop, not a normal application.

27 Jun 2010, 11:03

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20520
Location: In your JS exploiting you and your system

revolution 27 Jun 2010, 11:16

bitshifter wrote:

Basically i traded pushing 5 parameters each iteration for one subtraction.
In my book, thats always faster...

I suggest you test to see if it really is faster. Your assumption may be false and will never know until you test it. Modern CPUs are very complex with all sorts of weird things that happen internally. You might be surprised at the results.

bitshifter wrote:

Its for a single threaded demo loop, not a normal application.

Thanks for clarifying the usage case. But I want to ask, what is the loop waiting for? If it is waiting for user input then your 100% resource usage (to save a perhaps few nano seconds at most) is just power burned for no good reason. The user would never notice such a small increase in response time. If you are waiting for something else, then what?

27 Jun 2010, 11:16

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20520
Location: In your JS exploiting you and your system

revolution 27 Jun 2010, 11:23

One thing to note here is that Windows does not guarantee to leave the stack parameters untouched. The values may have been changed:

Code:

    MainLoop:
        call    [PeekMessage]
        sub     esp,20  ;<--- the values may have been changed!
        test    eax,eax
        jz      MainLoop

27 Jun 2010, 11:23

bitshifter

Joined: 04 Dec 2007
Posts: 796
Location: Massachusetts, USA

bitshifter 27 Jun 2010, 11:28

I never seen any winapi code play behind the RIP before?

27 Jun 2010, 11:28

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20520
Location: In your JS exploiting you and your system

revolution 29 Jun 2010, 11:27

bitshifter wrote:

I never seen any winapi code play behind the RIP before?

But are you willing to take the risk that all previous, current and future versions of Windows will never alter the incoming parameters?

29 Jun 2010, 11:27

r22

Joined: 27 Dec 2004
Posts: 805

r22 30 Jun 2010, 16:08

You forgot to align the MainLoop label to 16bytes.

30 Jun 2010, 16:08

ass0

Joined: 31 Dec 2008
Posts: 518
Location: ( . Y . )

ass0 30 Jun 2010, 17:58

why?

_________________

Nombre: Aquiles Castro.
Location2: about:robots

30 Jun 2010, 17:58

baldr

Joined: 19 Mar 2008
Posts: 1651

baldr 30 Jun 2010, 23:52

ass0,

Cache fill after miss after jmp MainLoop will probably perform better.

30 Jun 2010, 23:52

ass0

Joined: 31 Dec 2008
Posts: 518
Location: ( . Y . )

ass0 01 Jul 2010, 00:26

huh? but how it will filled? with several nop to reach the 16th byte?
In that case, processing each nop takes 1 cycle, then where is the needed to align?

Without to mention the file size grows.

_________________

Nombre: Aquiles Castro.
Location2: about:robots

01 Jul 2010, 00:26

LocoDelAssembly
Your code has a bug

Joined: 06 May 2005
Posts: 4624
Location: Argentina

LocoDelAssembly 01 Jul 2010, 01:13

ass0, but those NOPs will be executed only once in the entire program run, the remaining N-1 iterations of the loop will skip those NOPs.

Quote:

Without to mention the file size grows.

That is provided the extra bytes helped to exceed a 512-byte boundary, otherwise the file size will be unaffected.

01 Jul 2010, 01:13

ass0

Joined: 31 Dec 2008
Posts: 518
Location: ( . Y . )

ass0 01 Jul 2010, 01:16

wow!

_________________

Nombre: Aquiles Castro.
Location2: about:robots

01 Jul 2010, 01:16

asmfan

Joined: 11 Aug 2006
Posts: 392
Location: Russian

asmfan 01 Jul 2010, 07:18

This isn't bottleneck.
The bottleneck is message processing not picking:

Processing by:
1. jumptables (sequential)
2. message frequency analysis (sequential but more efficient)
3. binary trees on message processing (non-sequential long-time ago implemented by c/c++ compilers)

And I doubt about reusing used parameters unless some API states that it's func (const param).
Anyway doesn't worth optimization at all.

01 Jul 2010, 07:18

f0dder

Joined: 19 Feb 2004
Posts: 3175
Location: Denmark

f0dder 01 Jul 2010, 22:09

As mentioned before, you definitely shouldn't re-use stack parameters. I haven't bumped into cases where API calls modify stack parms, and there might be Microsoft coding guidelines against it - but the day it happens, you'll be pretty sorry.

Also, please stop doing CPU-greedy MessageLoops like the one above... there's really no reason to do it, and it sucks wrt. conserving laptop battery.

01 Jul 2010, 22:09

< 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