x86 Code Optimization

[blueshogun96]

Hey everyone, I was wondering if anyone could point me in the right direction to some tutorials and/or documents on optimizing C/C++ code with MMX, SSE, SIMD instructions? I searched google, but modt of the information i found pertained to pure assembly.

 

[Zilla]

http://www.linuxjournal.com/article/7269 : For GCC.

 

[NecroRomancist]

I'm using a cool lib on a emulator project i'm doing.
http://sourceforge.net/projects/simdx86

It supports anything from mmx to sse3 .Pretty nice

 

[Slougi]

Try here

 

[blueshogun96]

thanks guys, this will do

 

[Cyberman]

one thing I should mention Microsofts optimizations for these particular instructions aren't very good. In fact they are horrible. If you plan on the MMX instructions being in a loop consider using ASM only otherwise you will get nasty gotcha's from Microsoft's compilor.

Cyb

 

[smcd]

one thing I should mention Microsofts optimizations for these particular instructions aren't very good. In fact they are horrible. If you plan on the MMX instructions being in a loop consider using ASM only otherwise you will get nasty gotcha's from Microsoft's compilor.

Cyb

Agreed, MMX/SSE/SSE2 from MS's compiler (tried VC++2003) is icky.

 

[blueshogun96]

one thing I should mention Microsofts optimizations for these particular instructions aren't very good. In fact they are horrible. If you plan on the MMX instructions being in a loop consider using ASM only otherwise you will get nasty gotcha's from Microsoft's compilor.

Cyb

Hmm. Is that so? Ok. But how can I use ASM files in a VS 2003 project?

 

[Falcon4ever]

do a search on inline asm (it will show up some links to msdn)

 

[Cyberman]

I believe you can use it inline, there are some things you need to be careful with in dealing with MMX etc code. FIRST you must at the end SWITCH back to FPU register usage or the CPU's behavior will invoke an deadly embrace problem if it hits an FPU instruction mix.
I recomend grabing AMD's instruction reference it's what I used to mess with MMX instructions (and it certain made more sense that Intels poo of a manual). Be sure to remember how parameters are passed and if they are by reference or value. I don't recomend passing any C++ objects unless you are a glutton for severe traumatic punishment.

asm
{
}

I believe it's been a few years for moi

You could look at the VirtualDub source to find examples of it.


Cyb

 

[smcd]

The assembly doesnt have to be inline even. You can use ml to assemble MASM-style assembly, or you can define a custom build step to use nasm to assemble some source too.

 

[Falcon4ever]

we need samples people!!!
I wanted to learn programming ASM a few months ago, so I looked up what the easiest way was and in my case it was inline asm. But I've been very curious about how to do it the other ways

edit/note @ sethmcdoogle
I was doing it in MS VS2K5

 

[smcd]

we need samples people!!!
I wanted to learn programming ASM a few months ago, so I looked up what the easiest way was and in my case it was inline asm. But I've been very curious about how to do it the other ways

There are a couple emulators that do this (use assembly source files not just inline source) already, I forget the names - gens i think? Anyhow, let me get home from work and I'll make a very crude example project (VC++ 2003 is the only MS compiler I have currently).

EDIT: OK, see crappy attached project using custom build steps to assemble (using ml.exe) and link (using link.exe) an assembly file in a project and presto, it's a crappy demo for VC++ 2003!

 

[blueshogun96]

Hey, sorry it's been so long since I replied to your post. Getting online has been nothing but a b@#%& lately. Thanks for your code example. I really learned alot from it. :arabia:

I've been wanting to learn how to do this for ages now.

 

[smcd]

hopefully i didn't botch it up too badly if you want another example, check the hq2x example project http://www.hiend3d.com/hq2x.html the optimized asm one

 

[Runik]

IIRC ZSnes is done in assembly, and only uses C for some GUI parts ...
Browse their CVS files on SourceForge to get an idea of their work

On my own project I'm using inline assembly for some critical parts, and some naked assembly functions too (mainly for a future dynarec implementation )

 

[Doomulation]

Snes9x also has an asm core afaik

 

[smcd]

I was trying for simple examples to not overwhelm people but yeah zsnes was another project that came to mind