bitmap-fonts, osd, neon, draw, yuv

🔵 Internationalization

English | 中文文档

@think3r 2019-08-18 20:45:21
Reference:

1. RealView® 编译工具-汇编程序指南(pdf)
2. ARM Neon: conditional store suggestion
3. how to use arm neon vbit intrinsics?
4. arm neon 相关文档和指令意义 - github
5. ARM Neon Intrinsics各函数介绍
6. ARM NEON Intrinsics -- gcc.gnu.org
7. 总结各种RGB转YUV的转换公式
8. 在线 Arm neon 汇编调试器
9. NEON加速之memcpy在ARM平台的优化

0x00 Introduction

Accelerate the following two multimedia OSD character overlay functions through neon technology under the arm platform :

1. Expand 1bit dot-matrix-fonts(bitmap-fonts) to 8bit / 16bit ;

2. Render dot matrix fonts on several YUV formats;

   

0x01 Support

1. NEON-Accelerate :
   1. Dot-matrix(1bit) expanded to 1Byte and 2Byte;
      • The output of 2Byte can be any value, such as ARGB1555, ARGB4444, RGB565 etc.
   2. Draw OSD directly on the following YUV format :
      • YUV420 support : I420, YV12, NV12, NV21;
      • YUV422 support : YUYV, UYVY;
2. Support character types :
   • ASCII, base size 8x16;
   • Chinese (HZK16), base size 16*16;
3. Support character specifications:
   1. Dot-matrix expanded :
      • 1Byte : x2, x4;
        • TODO : x1, x3
      • 2Byte : x1, x2, x3, x4, contains two implementations:
        • Normal cpu expanded : @ ./osd_Draw/cpu_osd.c
        • Arm-neon cpu expanded with neon (Performance optimized) : @./osd_Draw/neon_osd.c
   2. YUV-OSD-Overlay : x2, x4 ;
      • Does not support x1 & x3, because YUV420 sampling 1/2 on horizontal and vertical ;

0x02 Build

./
├── CMakeLists.txt      # 分离的 cmake : 上层文件;
├── CMakeLists_one.txt  # 单独的 Cmake 文件(完整);
├── build               # cmake 构建文件夹
├── demo
│   ├── ASCII8          #英文字库
│   ├── CMakeLists.txt  # 分离的 cmake : demo 编译文件;
│   ├── HZK16           #汉字字库 GB2312
│   ├── inc
│   │   ├── build_time.h
│   │   ├── common.h
│   │   ├── neon_intrinsics_test.h
│   │   └── osd_test.h
│   ├── Makefile
│   ├── Makefile.android.clang  #高版本 ndk
│   ├── Makefile.android.gcc    #低版本 ndk
│   └── src
│       ├── build_time.c
│       ├── neon_intrinsics_test.c  #noen 内嵌函数测试, 与本项目无关
│       ├── osd_test.c
│       └── test.c
└── osd_Draw
    ├── CMakeLists.txt  # 分离的 cmake : 库文件编译;
    ├── cpu_osd.c      #cpu 拓展 osd 字符点阵, 仅支持 u16
    ├── cpu_osd.h
    ├── neon_osd.c     #neon 加速实现 : 拓展 osd 字符点阵和 yuv 绘制字符
    ├── neon_osd.h
    ├── osd_base.c     #初始化相关
    └── osd_base.h

• The following three compilation methods are provided:

  1. Makefile base on ndk-r20+ clang --> Makefile.android.clang;
     • To compile @64bit version;
  2. Makefile base on ndk-r10e gcc --> Makefile.android.gcc;
     • To compile @32bit version;
  3. CMake --> CMakeLists.txt 和 CMakeLists_one.txt
     • To compile @32bit version;

• Build steps (take the clang compilation of ndk-r20 as an example) :

  1. Open the content you want to test in test.c:
     • test_CreatOsdDot_u16_func() ---> u16 dot-expand;
     • test_DrawOsd_YUV_func() ---> yuv superimposed osd character test;
  2. Pick Makefile : cp Makefile.android.clang Makefile
  3. vi Makefile and chang ndk-r20 path;
  4. make clean && make
  5. Send the executable-binary and necessary yuv files to the phone, then run ./osd_demo

• u8 和 u16 点阵简易查看方式: beyond Compare 4:

  1. 16 进制比较生成的两个 bin 文件 osdOut.bin 和 osdOut_copy.bin (文件内容相同);
  2. 视图 -> 布局 中的 每组字节数, 配置为程序运行输出的 pitch 值; 如下图所示:

  

• yuv-overlay example:

  

0x03 Performance

u16 dot expand

• Test data size : 256,0000

• Compiler optimization level : -O2

• Test phone :

  • xiaomi-6, Snapdragon-835
  • Redmi K30 Pro, Snapdragon-865

• Phone environment: termux

• Compiler: ndk-r20-clang @ununtu

• The test procedure is @64bit;

• result 835 :

  front scale	cpu all	cpu average	neon all	neon average
  8 * 16	433 ms	0.169 us	132 ms	0.052 us
  16 * 32	1206 ms	0.471 us	341 ms	0.133 us
  24 * 48	2164 ms	0.845 us	565 ms	0.221 us
  32 * 64	2686 ms	1.049 us	861 ms	0.334 us

• result 865 :

  front scale	cpu all	cpu average	neon all	neon average
  16 * 32	490 ms	0.191 us	113 ms	0.044 us