如何在android上更快地将RGB565转换为YUV420SP? [英] How to convert RGB565 to YUV420SP faster on android?
问题描述
我需要显示一个 jpeg 图片，并将其转换为 YUV420SP.首先我使用SkBitmap解析jpeg并显示出来，然后我在android上使用下面的代码将RGB565转换为YUV420SP，但是转换640*480 RGB565图片需要75ms，所以任何人都知道将RGB565转换为YUV420SP的更快方法安卓?或者在android上将jpeg文件转换为YUV420SP的更快方法?
I need display a jpeg picture, and convert it to YUV420SP. First I use SkBitmap to parse jpeg and display it, then I use the code below to convert RGB565 to YUV420SP on android, but it spend 75ms to convert a 640*480 RGB565 picture, so anybody know the faster way to convert RGB565 to YUV420SP on android? or faster way to convert jpeg file to YUV420SP on android?
// Convert from RGB to YUV420
int RGB2YUV_YR[256], RGB2YUV_YG[256], RGB2YUV_YB[256];
int RGB2YUV_UR[256], RGB2YUV_UG[256], RGB2YUV_UBVR[256];
int RGB2YUV_VG[256], RGB2YUV_VB[256];
//
// Table used for RGB to YUV420 conversion
//
void InitLookupTable()
{
static bool hasInited = false;
if(hasInited)
return ;
hasInited = true;
int i;
for (i = 0; i < 256; i++)
RGB2YUV_YR[i] = (float) 65.481 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_YG[i] = (float) 128.553 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_YB[i] = (float) 24.966 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_UR[i] = (float) 37.797 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_UG[i] = (float) 74.203 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_VG[i] = (float) 93.786 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_VB[i] = (float) 18.214 * (i << 8);
for (i = 0; i < 256; i++)
RGB2YUV_UBVR[i] = (float) 112 * (i << 8);
}
int ConvertRGB5652YUV420SP(int w, int h, unsigned char *bmp, unsigned char *yuv)
{
unsigned char *u, *v, *y, *uu, *vv;
unsigned char *pu1, *pu2, *pu3, *pu4;
unsigned char *pv1, *pv2, *pv3, *pv4;
unsigned char rValue = 0, gValue = 0, bValue = 0;
uint16_t* bmpPtr;
int i, j;
printf("ConvertRGB5652YUV420SP begin,w=%d,h=%d,bmp=%p,yuv=%p\n", w, h, bmp, yuv);
struct timeval tpstart,tpend;
gettimeofday(&tpstart,NULL);
InitLookupTable();
gettimeofday(&tpend,NULL);
float timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+tpend.tv_usec-tpstart.tv_usec;
timeuse/=1000;
printf("InitLookupTable used time=%f\n", timeuse);
gettimeofday(&tpstart,NULL);
uu = new unsigned char[w * h];
vv = new unsigned char[w * h];
if (uu == NULL || vv == NULL || yuv == NULL)
return 0;
y = yuv;
u = uu;
v = vv;
// Get r,g,b pointers from bmp image data....
bmpPtr = (uint16_t*)bmp;
//Get YUV values for rgb values...
for (i = 0; i < h; i++) {
for (j = 0; j < w; j++) {
uint16_t color = *bmpPtr;
unsigned int r = (color>>11) & 0x1f;
unsigned int g = (color>> 5) & 0x3f;
unsigned int b = (color ) & 0x1f;
rValue = (r<<3) | (r>>2);
gValue = (g<<2) | (g>>4);
bValue = (b<<3) | (b>>2);
*y++ = (RGB2YUV_YR[rValue] + RGB2YUV_YG[gValue] + RGB2YUV_YB[bValue] +
1048576) >> 16;
*u++ = (-RGB2YUV_UR[rValue] - RGB2YUV_UG[gValue] + RGB2YUV_UBVR[bValue] +
8388608) >> 16;
*v++ = (RGB2YUV_UBVR[rValue] - RGB2YUV_VG[gValue] - RGB2YUV_VB[bValue] +
8388608) >> 16;
bmpPtr++;
}
}
gettimeofday(&tpend,NULL);
timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+tpend.tv_usec-tpstart.tv_usec;
timeuse/=1000;
printf("Get YUV values used time=%f\n", timeuse);
gettimeofday(&tpstart,NULL);
// Now sample the U & V to obtain YUV 4:2:0 format
// Get the right pointers...
u = yuv + w * h;
v = u + 1;
// For U
pu1 = uu;
pu2 = pu1 + 1;
pu3 = pu1 + w;
pu4 = pu3 + 1;
// For V
pv1 = vv;
pv2 = pv1 + 1;
pv3 = pv1 + w;
pv4 = pv3 + 1;
// Do sampling....
for (i = 0; i < h; i += 2) {
for (j = 0; j < w; j += 2) {
*u = (*pu1 + *pu2 + *pu3 + *pu4) >> 2;
u += 2;
*v = (*pv1 + *pv2 + *pv3 + *pv4) >> 2;
v += 2;
pu1 += 2;
pu2 += 2;
pu3 += 2;
pu4 += 2;
pv1 += 2;
pv2 += 2;
pv3 += 2;
pv4 += 2;
}
pu1 += w;
pu2 += w;
pu3 += w;
pu4 += w;
pv1 += w;
pv2 += w;
pv3 += w;
pv4 += w;
}
gettimeofday(&tpend,NULL);
timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+tpend.tv_usec-tpstart.tv_usec;
timeuse/=1000;
printf("Do sampling used time=%f\n", timeuse);
gettimeofday(&tpstart,NULL);
delete uu;
delete vv;
return 1;
}
int main(int argc, char **argv) {
unsigned char bmp[640*480*2] = {0};
unsigned char yuv[(640*480*3)/2] = {0};
struct timeval tpstart,tpend;
gettimeofday(&tpstart,NULL);
ConvertRGB5652YUV420SP(640, 480, bmp, yuv);
gettimeofday(&tpend,NULL);
float timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+tpend.tv_usec-tpstart.tv_usec;
timeuse/=1000;
printf("ConvertARGB2YUV420SP used time=%f\n", timeuse);
return 0;
}
android(armv6) 上的输出:
output on android(armv6):
ConvertRGB5652YUV420SP begin,w=640,h=480,bmp=0xbe7314fc,yuv=0xbe7c74fc
InitLookupTable used time=0.383000
Get YUV values used time=61.394001
Do sampling used time=11.918000
ConvertARGB2YUV420SP used time=74.596001
CPU 信息:
$ cat /proc/cpuinfo
cat /proc/cpuinfo
Processor : ARMv6-compatible processor rev 5 (v6l)
BogoMIPS : 791.34
Features : swp half thumb fastmult vfp edsp java
CPU implementer : 0x41
CPU architecture: 6TEJ
CPU variant : 0x1
CPU part : 0xb36
CPU revision : 5
Hardware : IMAPX200
Revision : 0000
Serial : 0000000000000000
推荐答案
在 ARMv7 上，使用 NEON.它将在不到 1 毫秒的时间内完成工作.(VGA)
On ARMv7, use NEON. It will do the job in less than 1ms. (VGA)
如果您坚持使用 ARMv6，请在 ARM 汇编中对其进行优化.(VGA约8ms)
If you are stuck with ARMv6, optimize it in ARM assembly. (about 8ms on VGA)
使用定点算法代替查找表.摆脱它们.
Use fixed-point arithmetic instead of the lookup tables. Get rid of them.
制作两个面具:
- 0x001f001f:掩码1
- 0x003f003f:mask2
然后一次将两个像素加载到 32 位寄存器中(这比读取 16 位要快得多)
then load two pixels at once into a 32bit register (which is a lot faster than 16bit read)
and red, mask1, pixel, lsr #11
and grn, mask2, pixel, lsr #5
and blu, mask1, pixel
现在你有三个寄存器，每个包含两个值 - 一个在低位，另一个在高位 16 位.
now you have three registers, each containing two values - one in the lower, and the other in the upper 16 bits.
从这里开始，smulxy 指令会创造一些奇迹.(16位乘法)
smulxy instructions will do some miracles from here on. (16bit multiply)
祝你好运.
PS:你的查找表也不是很好.为什么它们的长度都是 256?您可以将它们减少到 32(r 和 b 相关)和 64(g 相关)，这将增加缓存命中率.可能这仅适用于目标 40 毫秒，而无需求助于组装.是的，缓存未命中太痛苦了.
PS : your lookup table isn't that good either. Why are they all in length of 256? You could reduce them to 32 (r and b related) and 64 (g related) Which will increase the cache hit rate. Probably that will just do for the targeted 40ms without resorting to assembly. Yes, cache-misses are THAT painful.
这篇关于如何在android上更快地将RGB565转换为YUV420SP?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
