Vector Addition using SIMD

#include 

#define VECTOR_SIZE         4
typedef float v4sf __attribute__ ((vector_size(sizeof(float)*VECTOR_SIZE))); // vector of four singl
e floats

typedef union f4vector
{
    v4sf    v;
    float   f[VECTOR_SIZE];
} f4vector;


void print_vector (f4vector *v)
{
    printf("%f,%f,%f,%f\n", v->f[0], v->f[1], v->f[2], v->f[3]);
}

int main()
{
    union f4vector a, b, c;

    a.v = (v4sf){1., 2., 3., 4.};
    b.v = (v4sf){5., 6., 7., 8.};
    c.v = a.v + b.v;

    print_vector(&a);
    print_vector(&b);
    print_vector(&c);
}

Compile with the following command:

gcc -ggdb -mtune=pentium3 -march=pentium3 -c -O3 -ffast-math -mfpmath=sse -msse5 sse.c

To test, just link the object code to binary:

gcc -lm sse.o -o sse

$ ./sse
1.000000,2.000000,3.000000,4.000000
5.000000,6.000000,7.000000,8.000000
6.000000,8.000000,10.000000,12.000000

The assembled code:

$ objdump -dS ./sse.o | grep -2 c.v | tail -8
  7c:   0f 58 c1                addps  %xmm1,%xmm0
  7f:   0f 29 45 c8             movaps %xmm0,-0x38(%ebp)
--
 120:   f2 0f 11 44 24 04       movsd  %xmm0,0x4(%esp)
 126:   e8 00 00 00 00          call   12b <_main+0xdb>
        c.v = a.v + b.v;

        print_vector(&a);

As we can see, it's very optimized where adding 4 components of vector a and b is done in one SSE instruction (addps) instead of multiple instructions if we don't use -msse and -mfpmath=sse



How fast is the program?

$ time ./sse
1.000000,2.000000,3.000000,4.000000
5.000000,6.000000,7.000000,8.000000
6.000000,8.000000,10.000000,12.000000

real    0m0.109s
user    0m0.046s
sys     0m0.030s