Date:2018.10.11
1、libmpeg2介绍及其源码下载
project homepage: http://libmpeg2.sourceforge.net/
sourcecode download: https://sourceforge.net/projects/libmpeg2/
github sourcecode download: https://github.com/aholtzma/mpeg2dec/tree/master
开源协议： GNU General Public License version 2.0（GPLv2）
摘抄自：http://libmpeg2.sourceforge.net/
About libmpeg2
libmpeg2 is a free library for decoding mpeg-2 and mpeg-1 video streams. It is released under the terms of the GPL license.
The main goals in libmpeg2 development are:
- Conformance - libmpeg2 is able to decode all mpeg streams that conform to certain restrictions: “constrained parameters” for mpeg-1, and “main profile” for mpeg-2. In practice, this is what most people are using. For streams that follow these restrictions, we believe libmpeg2 is 100% conformant to the mpeg standards - and we have a pretty extensive test suite to check this.
- Speed - there has been huge efforts there, and we believe libmpeg2 is the fastest library around for what it does. Please tell us if you find a faster one! With typical video streams as found on DVD’s, and doing only decoding with no display, you should be able to get about 110 fps on a PIII/666, or 150 fps on an Athlon/950. This is less than 20 cycles per output pixel. In a real player program, the display routines will probably take as much time as the actual decoding!Speed - there has been huge efforts there, and we believe libmpeg2 is the fastest library around for what it does. Please tell us if you find a faster one! With typical video streams as found on DVD’s, and doing only decoding with no display, you should be able to get about 110 fps on a PIII/666, or 150 fps on an Athlon/950. This is less than 20 cycles per output pixel. In a real player program, the display routines will probably take as much time as the actual decoding!
- Portability - most of the code is written in C, and when we use platform-specific optimizations (typically assembly routines, currently used for the motion compensation and the inverse cosine transform stages) we always have a generic C routine to fall back on. This should be portable to all architectures - at least we have heard reports from people running this code on x86, ppc, sparc, arm and sh4. Assembly-optimized implementations are available on x86 (MMX) and ppc (AltiVec) architectures. Ultrasparc (VIS) is probably the next on the list - we’ll see.
- Reuseability - we do not want libmpeg2 to include any project-specific code, but it should still include enough features to be used by very diverse projects. We are only starting to get there - the best way to help here is to give us some feedback!
The project homepage is at http://libmpeg2.sourceforge.net/
优化说明：
开源mpeg2解码器libmpeg2进行了x86、arm、ppc、alpha和sparc架构的优化，没有进行aarch64的优化。
2、多平台编译测试
当前最新版本是libmpeg2-0.5.1。
libmpeg2解码库主要在Windows、Linux和ARM平台进行编译和性能测试。
2.1 Linux平台
编译很简单：
./configure
make
编译生成的mpeg2解码器路径为： ./src/.libs/mpeg2dec
./mpeg2dec -h
libmpeg2-0.5.1 - by Michel Lespinasse <walken@zoy.org> and Aaron Holtzman
usage: /home/soaringlee/Desktop/codecs/libmpeg2-0.5.1/src/.libs/lt-mpeg2dec [-h] [-o <mode>] [-s [<track>]] [-t <pid>] [-p] [-c]
[-v] [-b <bufsize>] <file>
-h display help and available video output modes
-s use program stream demultiplexer, track 0-15 or 0xe0-0xef
-t use transport stream demultiplexer, pid 0x10-0x1ffe
-p use pva demultiplexer
-c use c implementation, disables all accelerations
-v verbose information about the MPEG stream
-b set input buffer size, default 4096 bytes
-o video output mode
null
nullslice
nullskip
nullrgb16
nullrgb32
pgm
pgmpipe
md5
2.2 Windows平台
源码中没有windows平台的vs工程，首先需要建立库和demo的vs工程。
代码提取可以参考我的博客：
libmpeg2开源解码器的Windows工程搭建和编译
2.3 ARM平台
ARM32位：
./configure --host=arm-hisiv500-linux
make
ARM64位：
./configure --host=aarch64-unknown-linux
make
3、性能测试
3.1 Windows平台
测试配置：Intel® Core™ i5-6500 CPU @3.20GHZ
win32
| 分辨率 | 目标码率 | 平均解码帧率（汇编优化）/fps |
|---|---|---|
| 720P | 512k | 532 |
| 720P | 1024k | 490 |
| 720P | 1536k | 473 |
| 720P | 2048k | 440 |
3.2 Linux平台
测试配置：Intel® Xeon® CPU E5-2667 v3 @3.2GHZ 32核
Linux32位平均性能：
| 分辨率 | 目标码率 | 平均解码帧率（纯C）fps | 平均解码帧率（汇编优化）fps |
|---|---|---|---|
| 720P | 512k | 874 | 2483 |
| 720P | 1024k | 730 | 2086 |
| 720P | 1536k | 669 | 1456 |
| 720P | 2048k | 640 | 1518 |
Linux64位平均性能：
| 分辨率 | 目标码率 | 平均解码帧率（纯C）fps | 平均解码帧率（汇编优化）fps |
|---|---|---|---|
| 720P | 512k | 1049 | 3267 |
| 720P | 1024k | 934 | 2400 |
| 720P | 1536k | 839 | 1975 |
| 720P | 2048k | 821 | 2086 |
3.3 ARM平台
测试配置：Hisi3519 ARMv7 32位架构
双核： A7@800MHZ, 32KB I-Cache
A17@1.25GHZ,32KB I-Cache
内存： 1GB
测试配置：Hisi3559 ARMv8 64位架构
ARM32位平均性能：
| 分辨率 | 目标码率 | 平均解码帧率（纯C）fps | 平均解码帧率（汇编优化）fps |
|---|---|---|---|
| 720P | 512k | 105 | 189 |
| 720P | 1024k | 94 | 150 |
| 720P | 1536k | 90 | 141 |
| 720P | 2048k | 90 | 140 |
ARM64位平均性能：
| 分辨率 | 目标码率 | 平均解码帧率（纯C）fps |
|---|---|---|
| 720P | 512k | 248 |
| 720P | 1024k | 213 |
| 720P | 1536k | 201 |
| 720P | 2048k | 198 |
4、主要耗时模块分析
通过gprof工具对Linux平台和Arm平台进行分析可知，libmpeg2解码器的主要耗时模块包括：插值模块、IDCT模块和Iquant模块。
THE END!