neon intrinsic之图像融合

最近在做360环视的过程中碰到了图像融合的问题，由于板子上cpu算力有限，最终尝试用neon加速一下，最后总体算下来大概有5ms的加速，融合源图像越大，加速效果越明显，大致思路如下：
1.由于图像是切片引用过来的，然而传进寄存器需要一块连续的内存，因此这里深拷贝clone一下，图像过小，不建议使用这种方式，因为clone也会增加耗时，融合带来的提升会比clone小；
2.为了一次性处理八个数据，由于精度要求不高，所以采用定点优化将将权重取整到0~255，后右移八位还原

Mat StichManage::m_merge_neon(cv::Mat& imA, cv::Mat& imB, Mat& weight){
    Mat imA_clone = imA.clone();
    Mat imB_clone = imB.clone();
    uint8x8_t rfac = vdup_n_u8(255);

    int sum = imA.rows*imA.cols/8; 
    int yushu = imA.rows*imA.cols % 8;//不足八位的单独处理

    uchar* p_data_a = imA_clone.data;
    uchar* p_data_b = imB_clone.data;
    uchar* p_weight = weight.data;
    uint16x8_t temp1, temp2, temp3;
    uint8x8x3_t rgb1, rgb2;
    uint8x8_t weight1, weight2;
    for(int i = 0; i < sum; i++)
    {
        rgb1 = vld3_u8(p_data_a);
        rgb2 = vld3_u8(p_data_b);

        weight1 = vld1_u8(p_weight);
        weight2 = vsub_u8(rfac, weight1);

        temp1 = vmull_u8(rgb1.val[0], weight1);
        temp1 = vmlal_u8(temp1, rgb2.val[0], weight2);
        rgb1.val[0] = vshrn_n_u16 (temp1, 8);

        temp2 = vmull_u8(rgb1.val[1], weight1);
        temp2 = vmlal_u8(temp2, rgb2.val[1], weight2);
        rgb1.val[1] = vshrn_n_u16 (temp2, 8);

        temp3 = vmull_u8(rgb1.val[2], weight1);
        temp3 = vmlal_u8(temp3, rgb2.val[2], weight2);
        rgb1.val[2] = vshrn_n_u16 (temp3, 8);

        vst3_u8(p_data_a, rgb1);
        p_data_a += 24;
        p_data_b += 24;
        p_weight += 8;
    }
    for(int i = 0; i < yushu; i++)
    {
    	float weight1 = *p_weight / 255.0f;
    	float weight2 = 1 - weight1;
    	for(int j = 0; j< 3; j++)
    	{
    	    *p_data_a =  （*p_data_a）* weight + （*p_data_b）* weight2;
    	    p_data_a ++;
    	    p_data_b ++;
    	}
    	p_weight ++;
    }
    return imA_clone;
}