fastinlinefuncs.h

3D reconstruction, medical image processing from colons, using intel image processing for based clas

#ifndef _FASTINLINEFUNCS_H_
#define _FASTINLINEFUNCS_H_

#include <xmmintrin.h>
/*
	fast inline functions
*/

int inline FloatToInt(float *pSrc) {
	return _mm_cvttss_si32(_mm_load_ss(pSrc));
};

int inline FloatToInt(float fSrc) {
	return _mm_cvttss_si32(_mm_load_ss(&fSrc));
};

float inline IntToFloat(int iSrc) {
	__m128 vTemp=_mm_setzero_ps();
	float fRetval;
	_mm_store_ss(&fRetval, _mm_cvtsi32_ss(vTemp, iSrc));
	return fRetval;
};

float inline Sqrt(float fSrc) {
	float fRetval;
	_mm_store_ss(&fRetval, _mm_sqrt_ss(_mm_load_ss(&fSrc)));
	return fRetval;
}

#define __abs(fSrc) (((fSrc)>0) ? (fSrc) : -(fSrc))

/*
row0 = a3,a2,a1,a0
row1 = b3,b2,b1,b0
row2 = c3,c2,c1,c0
row3 = d3,d2,d1,d0

tmp0 = b3,a3,b2,a2
tmp1 = d3,c3,d2,c2
tmp2 = b1,a1,b0,a0
tmp3 = d1,c1,d0,c0

row0 = d0,c0,b0,a0
row1 = d1,c1,b1,a1
row2 = d2,c2,b2,a2
row3 = d3,c3,b3,a3
*/

#define _mm_transpose(row0,row1,row2,row3)  {				\
		__m64 tmp3, tmp2, tmp1, tmp0;						\
															\
															\
		tmp0	= _mm_unpackhi_pi16(row0,row1);				\
		tmp2	= _mm_unpacklo_pi16(row0,row1);				\
		tmp1	= _mm_unpackhi_pi16(row2,row3);				\
		tmp3	= _mm_unpacklo_pi16(row2,row3);				\
															\
		(row0)	= _mm_unpacklo_pi32(tmp2,tmp3);				\
		(row1)	= _mm_unpackhi_pi32(tmp2,tmp3);				\
		(row2)	= _mm_unpacklo_pi32(tmp0,tmp1);				\
		(row3)	= _mm_unpackhi_pi32(tmp0,tmp1);				\
}

#endif