图像预处理SSE加速，90度旋转和垂直镜像

</pre><p>图像预处理操作的90度整倍数旋转、镜像，都可由以下两个基本操作组合得出，测试比单点操作速度提升约3-5倍。</p><p></p><p></p><pre code_snippet_id="1889082" snippet_file_name="blog_20160920_3_6201408" name="code" class="cpp">//镜像-1*16单字节矩阵 
void MirrorMatrix(__m128i *pSrc_tmp ,__m128i *pDst_tmp , __m128i sort)
{
	_mm_storeu_si128(pDst_tmp, _mm_shuffle_epi8(_mm_loadu_si128(pSrc_tmp),sort));
}
           

//转置-16*16单字节矩阵 
void TransposeMatrix(__m128i** pSrc_tmp , __m128i** pDst_tmp)
{
	int i,j,k,bitcount = 16;
	__m128i src[16],dest[16];	//对应原始地址的数据
	__m128i tlh1[16];			//前8位l，后8位h == tlh3再用此地址
	__m128i tlh2[16];			//前4位l_l，后4位l_h,后4位h_l,后4位h_h
	for (i = 0;i<16;i++)
	{
		src[i] = _mm_loadu_si128(pSrc_tmp[i]); 
	}
	for (i = 0;i<8;i++)
	{
		tlh1[i] = _mm_unpacklo_epi8(src[i*2],src[i*2+1]);		
		tlh1[8+i] = _mm_unpackhi_epi8(src[i*2],src[i*2+1]);		
	}
	k = 4;
	for (i = 0;i<k;i++)
	{
		for (j = 0;j<2;j++)
		{
			tlh2[2*j*k+i] = _mm_unpacklo_epi16(tlh1[2*j*k+i*2],tlh1[2*j*k+i*2+1]);
			tlh2[(2*j+1)*k+i] = _mm_unpackhi_epi16(tlh1[2*j*k+i*2],tlh1[2*j*k+i*2+1]);		
		}
	}
	k = 2;
	for (i = 0;i<k;i++)
	{
		for (j = 0;j<4;j++)
		{
			tlh1[2*j*k+i] = _mm_unpacklo_epi32(tlh2[2*j*k+i*2],tlh2[2*j*k+i*2+1]);
			tlh1[(2*j+1)*k+i] = _mm_unpackhi_epi32(tlh2[2*j*k+i*2],tlh2[2*j*k+i*2+1]);		
		}	
	}
	for (j = 0;j<8;j++)
	{
		dest[2*j] = _mm_unpacklo_epi64(tlh1[2*j],tlh1[2*j+1]);
		dest[2*j+1] = _mm_unpackhi_epi64(tlh1[2*j],tlh1[2*j+1]);		
	}
	for (i = 0;i<16;i++)
	{
		_mm_storeu_si128(pDst_tmp[i], dest[i]);
	}
}