neon----深度转化

摘要

一、移位相乘

二、RGB到GRAY的转化

三、RGB565与RGB888互相转

总结

---

摘要

图像深度的转化, 主要是neon移位和乘法的操作。

---

一、移位和相乘操作

VSHR_N: 右移动 right shifts each element in a vector by an immediate value, and places the results in the destination vector。

VMLA (Vector Multiply Accumulate): 相乘 multiplies corresponding elements in two vectors and adds the results to the corresponding element of the destination vector.

二、RGB到GRAY的转化

对于彩色转灰度，心理学公式：

gray = R*0.299 + G*0.587 + B*0.114

在此0.299+0.587+0.144=1，刚好是满偏，这是通过不同的敏感度以及经验总结出来的公式，在实际应用时，希望避免低速的浮点运算，所以需要整数算法。注意到系数都是小数点后3位，我们可以将它们缩放1000倍来实现整数运算算法，如下

gray = (R*299 + G*587 + B*114) / 1000

但是涉及到除法，为了能在后续实现移位，将1000扩展到1024，得到式子如下：

gray = (R* 306 + G*601 + B*117) / 1024 = (R* 306 + G*601 + B*117) >>10

适当的还可以在精简，压缩到8位以内，现在变成这样子：

gray = (R*75 + G*147 + B*36) >>8

程序使用C语言和neon内联函数进行rgb转灰度算法优化。

//未使用neon优化的算法
void rgb2gray (const char* src, int n, char* dest)
{
  int i;
  for (i=0; i<n; i++)
  {
    int r = *src++;   // load red
    int g = *src++;   // load green
    int b = *src++;   // load blue
    // 不同的权证
    int tmp = (r*77) +(g*151) +(b*28);
    // 移位操作
    *dest = tmp >>8;
    dest++;
  }
}

//使用内联neon函数优化后的算法
void neon_rgb2gray (const char* src, int n, char* dest)
{
  int i;
  uint8x8_t rdup = vdup_n_u8 (77);            //将77（8位）存入rfac中
  uint8x8_t gdup = vdup_n_u8 (151);
  uint8x8_t bdup = vdup_n_u8 (28);
  n /= 8;
  for (i=0; i<n; i++)
  {
    uint16x8_t  temp;                         //uint16位*8
    uint8x8x3_t rgb  = vld3_u8 (src);         //结构体rgb
    uint8x8_t result;
    temp = vmull_u8(rgb.val[0],      rdup);   //长向量乘法
    temp = vmlal_u8(temp, rgb.val[1], gdup);
    temp = vmlal_u8(temp, rgb.val[2], bdup);
    result = vshrn_n_u16(temp, 8);            //右移
    vst1_u8(dest, result);                   //存储
    src  += 8*3;
    dest += 8;
  }
}

三、RGB565与RGB888互相转

图像是以RGB888格式显示图像的，24位图每个像素保存了32bit的数据，即RGB888+ Alpha， Alpha就是半透明填充字节……但是对于真彩的图像而言，肉眼在16bit的时候已经难以分辨了。因此，有些时候，可以讲RGB888转换为RGB565来存储，减少了存储器的容量的同时，降低了数据量；在后端显示的时候，再次把RGB565转换为RGB888，实现数据宽度的匹配。

RGB888转RGB565，只要提取相应单色高位即可(R5 G6 B5)，但会导致低位的缺失，影响精度，而且无法恢，RGB565转RGB888，方法只要补充相应单色低位即可(R3 G2 B3)。

RGB888用unsigned int 32位字节存储
0 0 0 0 0 0 0 0 R7 R6 R5 R4 R3 R2 R1 R0 G7 G6 G5 G4 G3 G2 G1 G0 B7 B6 B5 B4 B3 B2 B1 B0

RGB565用unsigned short 16位字节存储
R7 R6 R5 R4 R3 G7 G6 G5 G4 G3 G2 B7 B6 B5 B4 B3

unsigned short RGB888ToRGB565(unsigned int n888Color)
{
	unsigned short n565Color = 0;
	// 获取RGB
	unsigned char cRed   = (n888Color & RGB888_RED) >> 19;
	//右移八位再次移动两位,保留6位
    unsigned char cGreen = (n888Color & RGB888_GREEN) >> 10;
    //保留高5位
	unsigned char cBlue  = (n888Color & RGB888_BLUE) >> 3;
 
	// 连接
	n565Color = (cRed << 11) + (cGreen << 5) + (cBlue << 0);

	return n565Color;
}
 
unsigned int RGB565ToRGB888(unsigned short n565Color)
{
	unsigned int n888Color = 0;
 
	//左移动3位,再右移动11位
	unsigned char cRed   = (n565Color & RGB565_RED) >> 8;
	//左移动2位,再右移动5位
    unsigned char cGreen = (n565Color & RGB565_GREEN) >> 3;
    //五位凑齐8位
	unsigned char cBlue  = (n565Color & RGB565_BLUE) << 3;
 
	// 连接
	n888Color = (cRed << 16) + (cGreen << 8) + (cBlue << 0);
	return n888Color;
}

VMOVL (Vector Move Long) takes each element in a doubleword vector and sign or zero-extendsthem to twice their original length. The results are stored in a quadword vector

int i;
const int pixelsPerLoop = 8;
for(i = 0; i < pixelCount; i += pixelsPerLoop, inPixel32 += pixelsPerLoop, outPixel16 += pixelsPerLoop) {
    //Read all r,g,b pixels into 3 registers
    uint8x8x4_t rgba  = vld4_u8(inPixel32);
    //Right-shift
    uint8x8_t r = vshr_n_u8(rgba.val[0], 3);
    uint8x8_t g = vshr_n_u8(rgba.val[1], 2);
    uint8x8_t b = vshr_n_u8(rgba.val[2], 3);

    //b channel
    uint16x8_t r5_g6_b5 = vmovl_u8(b);
    //r channel
    uint16x8_t r16 = vmovl_u8(r);
    //排列565即移动11位, 取r
    r16 = vshlq_n_u16(r16, 11);
    r5_g6_b5 |= r16;

    //gchannel
    uint16x8_t g16 = vmovl_u8(g);
    //排列565,即移动5
    g16 = vshlq_n_u16(g16, 5);

    r5_g6_b5 |= g16;

    //Now write back to memory
    vst1q_u16(outPixel16, r5_g6_b5);        
}

VSRI_N right shifts each element in the second input vector by an immediate value, and inserts the results in the destination vector. It does not affect the highest n significant bits of the elements in the destination register. Bits shifted out of the right of each element are lost.

void setBgr888ToRgb565()
{
  uint8_t *src = image_src;
  uint16_t *dst = image_dst;
  int count = PIXEL_NUMBER;
  while (count >= 8) {
    uint8x8x3_t vsrc;
    uint16x8_t vdst;
    vsrc = vld3_u8(src);
    //the b channel, u8--u16
    vdst = vshll_n_u8(vsrc.val[0], 8);
    //g channel shift the 5 bit then insert to vdst
    vdst = vsriq_n_u16(vdst, vshll_n_u8(vsrc.val[1], 8), 5);
    //r channel shift the 11 bit then insert to vdst
    vdst = vsriq_n_u16(vdst, vshll_n_u8(vsrc.val[2], 8), 11);
    vst1q_u16(dst, vdst);
    dst += 8;
    src += 8*3;
    count -= 8;
  }
}

---

总结

移位和相乘的操作