MMX Question

[Farabi]

I had 4 data of bytes, I want to convert it to a 4 data of words. How to do that using MMX? All data are integers.

[qWord]

hi,
MMX is outdated and should no longer used - use SSEx instead. For you question, see the unpack instructions:

Code Select Expand

mov eax,04030201h
movd xmm0,eax
pxor xmm1,xmm1
punpcklbw xmm0,xmm1
movq QWORD ptr ...,xmm0 ; store 4 WORDs
note that you could process 8 bytes in one XMM register. Also, it works only for unsigned integers.
(For an array of DWORDs, there are much better implementation possible.)

[Farabi]

 :t Thanks qWord

[frktons]

hi,
MMX is outdated and should no longer used - use SSEx instead. For you question, see the unpack instructions:

Code Select Expand

mov eax,04030201h
movd xmm0,eax
pxor xmm1,xmm1
punpcklbw xmm0,xmm1
movq QWORD ptr ...,xmm0 ; store 4 WORDs
note that you could process 8 bytes in one XMM register. Also, it works only for unsigned integers.
(For an array of DWORDs, there are much better implementation possible.)

With these instructions you are creating 8 words in a 16 byte register.
If you need only 4 words, you can also use MMX registers with the same
instructions, if I correctly remember.

This is the transformation of qWord code:

Code Select Expand


mov eax,04030201h
movd mm0,eax
pxor mm1,mm1
punpcklbw mm0,mm1
movq QWORD ptr ...,mm0 ; store 4 WORDs


[frktons]

This is how Intel shows the functioning of PUNPCKLBW.

[Farabi]

This is how Intel shows the functioning of PUNPCKLBW.

Im still figuring out what this instruction used for in real life.

[frktons]

I had 4 data of bytes, I want to convert it to a 4 data of words. How to do that using MMX? All data are integers.

Im still figuring out what this instruction used for in real life.

Why do you need to convert 4 bytes into 4 words? That's why they created PUNPCKLBW.

[Farabi]

I had 4 data of bytes, I want to convert it to a 4 data of words. How to do that using MMX? All data are integers.

Im still figuring out what this instruction used for in real life.

Why do you need to convert 4 bytes into 4 words? That's why they created PUNPCKLBW.

I need that to make the substraction easy, sometime I need to multiply the byte data by 100 and then divide it by 200. Using byte data type it wont fit.

[frktons]

I need that to make the substraction easy, sometime I need to multiply the byte data by 100 and then divide it by 200. Using byte data type it wont fit.

And then you can use PUNPCKLBW to have 4/8 single bytes moved
into 4/8 words in one instruction. Isn't it useful?

[dedndave]

Code Select Expand

shr al,1
or, if you are nit-picking

Code Select Expand

shr al,1
adc al,0

:P

[Farabi]

I need that to make the substraction easy, sometime I need to multiply the byte data by 100 and then divide it by 200. Using byte data type it wont fit.

And then you can use PUNPCKLBW to have 4/8 single bytes moved
into 4/8 words in one instruction. Isn't it useful?

Yeah it should be usefull if I understand it right.

[dedndave]

if you use PCMGTB first, you could use it to sign-extend bytes to words   :P

[frktons]

It is not too difficult to understand:

Code Select Expand


mov eax,04030201h
movd mm0,eax
pxor mm1,mm1
punpcklbw mm0,mm1
movq QWORD ptr ...,mm0 ; store 4 WORDs


1) you move your 4 bytes into a register (EAX)
2) you move EAX to an MMX register
3) you put zero to another MMX register
4) you intermix bytes from dest MMX and src MMX
5) you have 4 words into the MMX dest register

[frktons]

Have a look at Intel Manuals:

http://www.intel.it/content/www/it/it/search.html?keyword=developer%CA%B9s+manual

[Farabi]

This is how Intel shows the functioning of PUNPCKLBW.

O I get it. If I want to fill the high order of the word fill it on the src, if the other wise, fill it on the dest.