MMX instructions

Intel product with MMX released in 1997
 

packed operations:

add, sub, mul, compare, shift, logic, conversion
multiply-accumulate

saturate arithmetic

Applications

MMX instruction set

padd         // wrap around and sat
psub
pcmpeq
pcmgt
pmul
pmadd       // mul and add adj pair
psra
psrl
punpck      // merge pack
packss      // pack sat
pand
pandn
por
pxor
mov

Explanation of some important instructions

pmadd
A3 A2 A1 A0
x
B3 B2 B1 B0

A3*B3 A2*B2 A1*B1 A0*B0

A3*B3+A2*B2 A1*B1+A0*B0


pcmpgt
 

51 03 05 23
>
73 02 05 06

0  1  0  1

punpck
 
. . . . B3 B2 B1 B0  . . . . A3 A2 A1 A0

B3 A3 B2 A2 B1 A1 B0 A0

Example of the use of MMX instructions

Vector multiply matrix
vec[.]
mat[.][.]
res[.]

for i=0; i<COL; i++
  a = 0
  for j=0; j<ROW; j++
    a = a + vec[j] * mat[j][i]
  res[i] = a


mul_vec_mat 4 x 2  in MMX
 
 
MMX instruction
Result
punpck mm6, mm0 EFGH ABCD
CGDH
punpck mm1, mm7 EFGH ABCD
AEBF
pmadd  mm6, mm7 AEBF V0V1V0V1  A*V0+E*V1 B*V0+F*V1
pmadd  mm1, mm7 CGDH V0V1V0V1  C*V0+G*V1 D*V0+H*V1
padd   mm2, mm1 mm2 (a)  = A*V0+E*V1 B*V0+F*V1
padd   mm3 ,mm6 mm3 (a) = C*V0+G*V1 D*V0+H*V1

Performance comparison

 
program
Pentium
MMX
vector dot product
input vector: 16 elements, signed 16-bit
accum 32 bits		 32 loads
16 muls
15 adds
4 MAC per loop: 12 inst. loop control
(3 inst. per iter., inc, cmp, branch)
1 inst. store result
--------------
total 76 inst.  (opt. 200 cycles)		 8 loads
4 pmadds
3 padds
four loop unroll (no branch)
---------
total 19 inst.  (opt. 12 cycles)
matrix mul 16 x 16 4236 cycles 367 cycles

Another example of using MMX instructions

Chroma keying: merge images x,y to z
 
for i=0; i<size; i++
  if x[i] == Blue then z = y[i]
                  else z = x[i]


x is the image with a blue background
y is the background image
 
 
MMX instructions
Result
mov    mm3, x
mov    mm4, y
pcmpeg mm1, mm3 mm1  B B B B B B B B
mm3  G R B B G R B B
mm1  0 0 1 1 0 0 1 1
pand   mm4, mm1 mm4  Y Y Y Y Y Y Y Y
mm1  0 0 1 1 0 0 1 1
mm4  0 0 Y Y 0 0 Y Y
pandn  mm1, mm3 mm1  0 0 1 1 0 0 1 1
mm3  X X X X X X X X
mm1  X X 0 0 X X 0 0
por    mm4, mm1 mm4  X X Y Y X X Y Y

 

Speed up on applications

general 1.5-2 times with MMX   (MPEG-1 video)
 

References

A. Peleg and U. Weiser, MMX technology extension to the Intel architecture, IEEE Micro, August 1996, pp. 42-50.