Question

I got some vectors containing unsigned chars that represent pixels from a frame.
I got this function working without the MMX improvement, but I frustrated whit MMX that doesnt work … So:

I need to add two unsigned chars (the sum need to be done as a 16bit instead of a 8bit cause unsigned char goes from 0-255 as known) and divide them by two (shift right 1). The code I have done so far is below, but the values are wrong, the adds_pu16 doesnt add the 16bit just 8:

  MM0 = _mm_setzero_si64();        //all zeros
  MM1 = TO_M64(lv1+k);             //first 8 unsigned chars
  MM2 = TO_M64(lv2+k);             //second 8 unsigned chars

  MM3 =_mm_unpacklo_pi8(MM0,MM1);  //get first 4chars from MM1 and add Zeros
  MM4 =_mm_unpackhi_pi8(MM0,MM1);  //get last 4chars from MM1 and add Zeros

  MM5 =_mm_unpacklo_pi8(MM0,MM2);  //same as above for line 2
  MM6 =_mm_unpackhi_pi8(MM0,MM2);

  MM1 = _mm_adds_pu16(MM3,MM5);    //add both chars as a 16bit sum (255+255 max range)
  MM2 = _mm_adds_pu16(MM4,MM6);

  MM3 = _mm_srai_pi16(MM1,1);      //right shift (division by 2)
  MM4 = _mm_srai_pi16(MM2,1);

  MM1 = _mm_packs_pi16(MM3,MM4);   //pack the 2 MMX registers into one

  v2 = TO_UCHAR(MM1);              //put results in the destination array

---

New developments:
Thanks for that king_nak!!
I wrote a simple version of what I am trying to do:

---

int main()
{
char A[8]={255,155,2,3,4,5,6,7};
char B[8]={255,155,2,3,4,5,6,7};
char C[8];
char D[8];
char R[8];

__m64* pA=(__m64*) A;

__m64* pB=(__m64*) B;

__m64* pC=(__m64*) C;

__m64* pD=(__m64*) D;

__m64* pR=(__m64*) R;

_mm_empty();

__m64 MM0 = _mm_setzero_si64();

__m64 MM1 = _mm_unpacklo_pi8(*pA,MM0);

__m64 MM2 = _mm_unpackhi_pi8(*pA,MM0);

__m64 MM3 = _mm_unpacklo_pi8(*pB,MM0);

__m64 MM4 = _mm_unpackhi_pi8(*pB,MM0);

__m64 MM5 = _mm_add_pi16(MM1,MM3);

__m64 MM6 = _mm_add_pi16(MM2,MM4);

printf("SUM:\n");

*pC= _mm_add_pi16(MM1,MM3);

*pD= _mm_add_pi16(MM2,MM4);

for(int i=0; i<8; i++) printf("\t%d ", (C[i])); printf("\n");

for(int i=0; i<8; i++) printf("\t%d ", D[i]); printf("\n");

printf("DIV:\n");

*pC= _mm_srai_pi16(MM5,1);

*pD= _mm_srai_pi16(MM6,1);

for(int i=0; i<8; i++) printf("\t%d ", (C[i])); printf("\n");

for(int i=0; i<8; i++) printf("\t%d ", D[i]); printf("\n");

MM1= _mm_srai_pi16(MM5,1);    
MM2= _mm_srai_pi16(MM6,1);

printf("Final Result:\n");
*pR= _mm_packs_pi16(MM1,MM2);
for(int i=0; i<8; i++) printf("\t%d ", (R[i])); printf("\n");

return(0);
}

---

And the results are:

SUM:

-2  1   54  1   4   0   6   0 

8   0   10  0   12  0   14  0 

DIV:

-1  0   -101    0   2   0   3   0 

4   0   5   0   6   0   7   0 

Final Result:

127     127     2   3   4   5   6   7 

Well the small numbers are ok while the big numbers which give 127 are wrong. This is a problem, what am I doing wrong :s

Answer 1

I think i found the problem:
The arguments of the unpack instructions are in the wrong order. If you look at the registers as a whole, it looks like the individual chars are zero-extended to shorts, but in fact, they are zero-padded. Just swap around mm0 and the other register in each case and it should work.

Also, you don’t need saturated add, a normal PADDW is sufficient. The maximum value you will get is 0xff+0xff=0x01fe, which doesn’t have to be saturated.

Edit: What’s more, PACKSSWB doesn’t quite do what you want. PACKUSWB is the correct instruction, saturation will get you wrong results.

Here’s a solution (Also replaced the shifts with logical ones and used different pseudo-registers in some places):

mm0=pxor(mm0,mm0) =[00,00,00,00,00,00,00,00]
mm1 =[a0,10,ff,18,7f,f0,ff,cc]
mm2 =[c0,20,ff,00,70,26,ff,01]
mm3=punpcklbw(mm1,mm0) =[00a0,0010,00ff,0018]
mm4=punpckhbw(mm1,mm0) =[007f,00f0,00ff,00cc]
mm5=punpcklbw(mm2,mm0) =[00c0,0020,00ff,0000]
mm6=punpckhbw(mm2,mm0) =[0070,0026,00ff,0001]
mm5=paddw(mm3,mm5) =[0160,0030,01fe,0018]
mm6=paddw(mm4,mm6) =[00ef,0116,01fe,00cd]
mm3=psrlwi(mm5,1) =[00b0,0018,00ff,000c]
mm4=psrlwi(mm6,1) =[0077,008b,00ff,0066]
mm1=packuswb(mm3,mm4) =[b0,18,ff,0c,77,8b,ff,66]