双线性插值算法ARM NEON优化
【摘要】
C语言版本双线性插值算法
inline double bilinear_interp(double x, double y, double v11, double v12, double v21, double v22) { return (v11 * (1 - y) + v12 * y) * (1 - x) + (v21 * (1 - y) + v22 * y)...
inline double bilinear_interp(double x, double y, double v11, double v12,
double v21, double v22) {
return (v11 * (1 - y) + v12 * y) * (1 - x) + (v21 * (1 - y) + v22 * y) * x;
}
使用ARM NOEN优化后的双线性插值版本
inline uint8x8_t bilinear_interp_NEON(double x, double y, uint8x8_t v11,uint8x8_t v12,
uint8x8_t v21,uint8x8_t v22)
{
uint16x8_t v11_16 = vmovl_u8(v11);
uint16x8_t v12_16 = vmovl_u8(v12);
uint16x8_t v21_16 = vmovl_u8(v21);
uint16x8_t v22_16 = vmovl_u8(v22);
///convert v11 to two float32x4
uint16x4_t v_16_low = vget_low_u16(v11_16);
uint16x4_t v_16_high = vget_high_u16(v11_16);
uint32x4_t v_32_low = vmovl_u16(v_16_low);
uint32x4_t v_32_high = vmovl_u16(v_16_high);
float32x4_t v11_32f_low = vcvtq_f32_u32(v_32_low);
float32x4_t v11_32f_high = vcvtq_f32_u32(v_32_high);
//v12
v_16_low = vget_low_u16(v12_16);
v_16_high = vget_high_u16(v12_16);
v_32_low = vmovl_u16(v_16_low);
v_32_high = vmovl_u16(v_16_high);
float32x4_t v12_32f_low = vcvtq_f32_u32(v_32_low);
float32x4_t v12_32f_high = vcvtq_f32_u32(v_32_high);
//v21
v_16_low = vget_low_u16(v21_16);
v_16_high = vget_high_u16(v21_16);
v_32_low = vmovl_u16(v_16_low);
v_32_high = vmovl_u16(v_16_high);
float32x4_t v21_32f_low = vcvtq_f32_u32(v_32_low);
float32x4_t v21_32f_high = vcvtq_f32_u32(v_32_high);
//v22
v_16_low = vget_low_u16(v22_16);
v_16_high = vget_high_u16(v22_16);
v_32_low = vmovl_u16(v_16_low);
v_32_high = vmovl_u16(v_16_high);
float32x4_t v22_32f_low = vcvtq_f32_u32(v_32_low);
float32x4_t v22_32f_high = vcvtq_f32_u32(v_32_high);
float32_t fx = (float32_t)x;
float32_t fy = (float32_t)y;
float32_t one_fx = 1-fx;
float32_t one_fy = 1-fy;
float32x4_t tmp1,tmp2,tmp3,tmp4,tmp5,tmp;
uint32x4_t result_32_low,result_32_high;
uint16x4_t result_16_low,result_16_high;
//for low 32x4
tmp1 = vmulq_n_f32(v11_32f_low, one_fy);
tmp2 = vmulq_n_f32(v12_32f_low, fy);
tmp3 = vaddq_f32(tmp1, tmp2);
tmp4 = vmulq_n_f32(tmp3, one_fx);
tmp1 = vmulq_n_f32(v21_32f_low, one_fy);
tmp2 = vmulq_n_f32(v22_32f_low, fy);
tmp3 = vaddq_f32(tmp1, tmp2);
tmp5 = vmulq_n_f32(tmp3, fx);
tmp = vaddq_f32(tmp4, tmp5);
result_32_low = vcvtq_u32_f32(tmp);
result_16_low = vqmovn_u32(result_32_low);
//for high 32x4
tmp1 = vmulq_n_f32(v11_32f_high, one_fy);
tmp2 = vmulq_n_f32(v12_32f_high, fy);
tmp3 = vaddq_f32(tmp1, tmp2);
tmp4 = vmulq_n_f32(tmp3, one_fx);
tmp1 = vmulq_n_f32(v21_32f_high, one_fy);
tmp2 = vmulq_n_f32(v22_32f_high, fy);
tmp3 = vaddq_f32(tmp1, tmp2);
tmp5 = vmulq_n_f32(tmp3, fx);
tmp = vaddq_f32(tmp4, tmp5);
result_32_high = vcvtq_u32_f32(tmp);
result_16_high = vqmovn_u32(result_32_high);
uint16x8_t result_16 = vcombine_u16(result_16_low,result_16_high);
uint8x8_t result_8 = vqmovn_u16(result_16);
return result_8;
}
使用ARM NEON后，一次可以处理8个像素，成倍提高了运行的速度。实践中需要特别注意对边界的处理（行的开始和结尾处）。
到这里我们还不能满足。要追求更快！！！
注意上面的代码中虽然用了ARM NEON，但是在ARM 指令集的操作中用到了浮点操作。所以，还可以继续使用浮点数定点化的优化方式，优化后的代码如下：
inline uint8x8_t bilinear_interp_NEON_FixedPoint(double x, double y, uint8x8_t v11,uint8x8_t v12,
uint8x8_t v21,uint8x8_t v22)
{
uint16x8_t v11_16 = vmovl_u8(v11);
uint16x8_t v12_16 = vmovl_u8(v12);
uint16x8_t v21_16 = vmovl_u8(v21);
uint16x8_t v22_16 = vmovl_u8(v22);
uint16x4_t v_16_low = vget_low_u16(v11_16);
uint16x4_t v_16_high = vget_high_u16(v11_16);
uint32x4_t v11_32_low = vmovl_u16(v_16_low);
uint32x4_t v11_32_high = vmovl_u16(v_16_high);
v_16_low = vget_low_u16(v12_16);
v_16_high = vget_high_u16(v12_16);
uint32x4_t v12_32_low = vmovl_u16(v_16_low);
uint32x4_t v12_32_high = vmovl_u16(v_16_high);
v_16_low = vget_low_u16(v21_16);
v_16_high = vget_high_u16(v21_16);
uint32x4_t v21_32_low = vmovl_u16(v_16_low);
uint32x4_t v21_32_high = vmovl_u16(v_16_high);
v_16_low = vget_low_u16(v22_16);
v_16_high = vget_high_u16(v22_16);
uint32x4_t v22_32_low = vmovl_u16(v_16_low);
uint32x4_t v22_32_high = vmovl_u16(v_16_high);
unsigned int intX = x*4096;
unsigned int intY = y*4096;
unsigned int one_x = 4096-intX;
unsigned int one_y = 4096-intY;
uint32_t intX_32 = (uint32_t) intX;
uint32_t intY_32 = (uint32_t) intY;
uint32_t oneX_32 = (uint32_t) one_x;
uint32_t oneY_32 = (uint32_t) one_y;
uint32x4_t tmp1,tmp2,tmp3,tmp4,tmp5,tmp;
uint16x4_t result_16_low, result_16_high;
//for low 4 numbers
tmp1 = vmulq_n_u32(v11_32_low,oneY_32);
tmp2 = vmulq_n_u32(v12_32_low, intY_32);
tmp3 = vaddq_u32(tmp1, tmp2);
tmp4 = vmulq_n_u32(tmp3, oneX_32);
tmp1 = vmulq_n_u32(v21_32_low, oneY_32);
tmp2 = vmulq_n_u32(v22_32_low, intY_32);
tmp3 = vaddq_u32(tmp1, tmp2);
tmp5 = vmulq_n_u32(tmp3, intX_32);
tmp = vaddq_u32(tmp4, tmp5);
result_16_low = vshrn_n_u32(tmp,16); //shift right 16 bytes
result_16_low = vrshr_n_u16(result_16_low,8); //shift right 8 bytes, totally 24 bytes
//for high 4 numbers
tmp1 = vmulq_n_u32(v11_32_high,oneY_32);
tmp2 = vmulq_n_u32(v12_32_high, intY_32);
tmp3 = vaddq_u32(tmp1, tmp2);
tmp4 = vmulq_n_u32(tmp3, oneX_32);
tmp1 = vmulq_n_u32(v21_32_high, oneY_32);
tmp2 = vmulq_n_u32(v22_32_high, intY_32);
tmp3 = vaddq_u32(tmp1, tmp2);
tmp5 = vmulq_n_u32(tmp3, intX_32);
tmp = vaddq_u32(tmp4, tmp5);
result_16_high = vshrn_n_u32(tmp,16); //shift right 16 bytes
result_16_high = vrshr_n_u16(result_16_high,8); //shift right 8 bytes, totally 24 bytes
uint16x8_t result_16 = vcombine_u16(result_16_low,result_16_high);
uint8x8_t result_8 = vqmovn_u16(result_16);
return result_8;
}
加入浮点定点化之后的优化，时间能进一步提升一倍左右。
文章来源: blog.csdn.net，作者：网奇，版权归原作者所有，如需转载，请联系作者。
原文链接：blog.csdn.net/jacke121/article/details/55253542
【版权声明】本文为华为云社区用户转载文章，如果您发现本社区中有涉嫌抄袭的内容，欢迎发送邮件进行举报，并提供相关证据，一经查实，本社区将立刻删除涉嫌侵权内容，举报邮箱：
cloudbbs@huaweicloud.com
- 点赞
- 收藏
- 关注作者
作者其他文章
评论（0）