wtf.cpp

// cl /O2 wtf.cpp

#include <stdio.h>
#include <windows.h>

static inline int LLR(int a, int b)
{
    if (a>0) 
        return (b>0) ? __min(a,b) : -__min(a,-b);
    else
        return (b>0) ? -__min(-a,b) : __min(-a,-b);
}

static inline int LLR_2(int a, int b)
{
    int sign, numbers[2];
    sign = 1-2*((unsigned)a^b)>>31;
    a *=  1-2*((unsigned)a)>>31;
    b *= 1-2*((unsigned)b)>>31;
    numbers[0] = b;
    numbers[1] = a;
    a -= b;
    return sign*numbers[((unsigned)a)>>31];
}

static unsigned int mask[] = {0,(unsigned)-1};
static unsigned int constant[] = {0, 1};

static inline int LLR_3(int a, int b)
{
    int sign,c, numbers[2];
    sign = (_rotl(a^b,1) & 1);
    c = ((_rotl(a,1) & 1));
    a = a^mask[c]+constant[c];

    c = ((_rotl(b,1) & 1));
    a = a^mask[c]+constant[c];

    numbers[0] = b;
    numbers[1] = a;

    c = ((_rotl(a-b,1) & 1));

    return numbers[c]^mask[sign]+constant[sign];
}

void main()
{
    SetPriorityClass(GetCurrentProcess(),REALTIME_PRIORITY_CLASS);
    SetThreadPriority(GetCurrentThread(),THREAD_PRIORITY_TIME_CRITICAL);

    int a[1024], b[1024];
    srand(0);
    for ( int i=0; i<1024; i++ )
    {
        a[i] = rand() - RAND_MAX / 2;
        b[i] = rand() - RAND_MAX / 2;
  }

    int x;
    __int64 t1,t2;

    // bench reference version
    x = 0;
    QueryPerformanceCounter((LARGE_INTEGER*)&t1);
    for (int i=0;i<100*1000*1000;i++)
    {
    	register int j = i & 1023;
		x += LLR(a[j],b[j]);
    }
    QueryPerformanceCounter((LARGE_INTEGER*)&t2);
    t2 -= t1;
    QueryPerformanceFrequency((LARGE_INTEGER*)&t1);
    printf("%f sec, x=%d\n",t2/(t1*1.), x);

    // bench v2
    x = 0;
    QueryPerformanceCounter((LARGE_INTEGER*)&t1);
    for (int i=0;i<100*1000*1000;i++)
    {
    	register int j = i & 1023;
		x += LLR_2(a[j],b[j]);
    }
    QueryPerformanceCounter((LARGE_INTEGER*)&t2);
    t2 -= t1;
    QueryPerformanceFrequency((LARGE_INTEGER*)&t1);
    printf("%f sec, x=%d\n",t2/(t1*1.), x);

    // bench v3
    x = 0;
    QueryPerformanceCounter((LARGE_INTEGER*)&t1);
    for (int i=0;i<100*1000*1000;i++)
    {
    	register int j = i & 1023;
		x += LLR_3(a[j],b[j]);
    }
    QueryPerformanceCounter((LARGE_INTEGER*)&t2);
    t2 -= t1;
    QueryPerformanceFrequency((LARGE_INTEGER*)&t1);
    printf("%f sec, x=%d\n",t2/(t1*1.), x);
}