gistfile1.cpp

// clx /O2 1.cpp 
#include <stdio.h>
#include <windows.h>
#include <tmmintrin.h> // sse3
#include <smmintrin.h> // sse4
#include <math.h>

const int BLOCK = 16*1024*1024;
const int RUNS = 100*1000*1000;

int * a = new int[BLOCK];
int * b = new int[BLOCK];

int sign(int x)
{
    return (x<0) ? -1 : 1;
}

void dump(const char * n, __m128i v)
{
	printf("%s: %d, %d, %d, %d\n", n, v.m128i_i32[0], v.m128i_i32[1], v.m128i_i32[2], v.m128i_i32[3]);
}

int sum(__m128i v)
{
	return v.m128i_i32[0] + v.m128i_i32[1] + v.m128i_i32[2] + v.m128i_i32[3];
}

static inline __m128i LLR4(const int * pa, const int * pb)
{
	__m128i a = *(__m128i*)pa;
	__m128i b = *(__m128i*)pb;

	// sign(a,b)*min(abs(a),abs(b))
	__m128i absa = _mm_abs_epi32(a);
	__m128i absb = _mm_abs_epi32(b);
	__m128i absm = _mm_min_epi32(absa, absb);
	__m128i ab = _mm_mullo_epi32(a, b);
	__m128i rr = _mm_sign_epi32(absm, ab);
	return rr;
}

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));
	b = (b^mask[c])+constant[c];

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

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

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

int Test1()
{
	int x = 0;
	for (int i=0;i<RUNS;i++)
	{
		register int j = i & (BLOCK-1);
		x += LLR(a[j],b[j]);
	}
	return x;
}

int Test2()
{
	int x = 0;
	for (int i=0;i<RUNS;i++)
	{
		register int j = i & (BLOCK-1);
		x += LLR_2(a[j],b[j]);
	}
	return x;
}

int Test3()
{
	int x = 0;
	for (int i=0;i<RUNS;i++)
	{
		register int j = i & (BLOCK-1);
		x += LLR_3(a[j],b[j]);
	}
	return x;
}

int Test4()
{
	__m128i t = { 0 };
	for (int i=0;i<RUNS;i+=4)
	{
		register int j = i & (BLOCK-1);
		t = _mm_add_epi32(t, LLR4(a+j, b+j));
	}
	return sum(t);
}

void Bench(int (*fnTest)())
{
	const int TRIALS = 10;
	float mintime = 1000000.0f;
	int x;
	for (int i=0; i<TRIALS; i++)
	{
		__int64 t1, t2;
		QueryPerformanceCounter((LARGE_INTEGER*)&t1);
		x = fnTest();
		QueryPerformanceCounter((LARGE_INTEGER*)&t2);
		t2 -= t1;
		QueryPerformanceFrequency((LARGE_INTEGER*)&t1);
		mintime = min(mintime, t2/(t1*1.));
	}
	printf("%f sec, x=%d\n", mintime, x);
}

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

	SetPriorityClass(GetCurrentProcess(),REALTIME_PRIORITY_CLASS);
	SetThreadPriority(GetCurrentThread(),THREAD_PRIORITY_TIME_CRITICAL);

	Bench(Test1);
	Bench(Test2);
	Bench(Test3);
	Bench(Test4);
}