换了快几倍的ARM，为什么性能没提上去

magicboatrocket

我有个用opencv处理图像的程序，之前在200M的ARM核（TI DM6446）上跑，monavista gcc工具链，性能不怎么样。
现在换了块600M的ARM11（freescale 的imx31 多媒体处理器）的处理器，在一般的arm  linux gcc工具链上编译，用的同一编译徒然，性能没有提高，反而降低了。
:em12: :em12: 这是为什么

有人对工具链gcc有深入一点的了解吗，是不是gcc编译器没有产生充分利用处理器硬件特性的机器码。比如我的imx31有VPF,浮点运算单元，gcc却没有产生利用这一特性的代码.


[ 本帖最后由 magicboatrocket 于 2009-10-20 23:08 编辑 ]

magicboatrocket

说到memory， 大小是差不多，肯定是够程序用的。不说我还不注意，查了下，先前的系统的是SDRAM，64M，后来的系统是用mobile DDR内存，64M。
SDRAM， mobile DDR的内存性能比较如何？有人有insight吗


imx31板子有提供编译器，但我无法确定他提供的编译器好不好。平时对arm-linux－gcc工具也就关注能不能把程序编译过去，对它产生的指令的质量就没什么了解了。
想问T-bagwell兄台，怎么测试软浮点和硬浮点运算的问题呢？反汇编，查硬浮点指令？

[ 本帖最后由 magicboatrocket 于 2009-10-21 13:45 编辑 ]

magicboatrocket

memory方面应该不会影响太大。觉得还是编译器的原因

Mobile DDR for automotive
While there are a lot of memory options for vehicular electronics, here we have sought to demonstrate that Mobile DDR is a good, high-quality choice for many -applications because it functions in a broad temperature range ("40 to 105C), consumes relatively little power, and is available in a variety of densities and configurations for the automotive market long haul.

http://www.embedded.com/design/208403917?pgno=2

magicboatrocket

谢谢T-bagwell兄，opencv里的确很多浮点运算，这些浮点运算对Intel的架构有做优化，但对ARM就没有了。有些浮点运算我在做移植的时候直接用标准库函数了。

写几个测试函数，但我要怎么知道编译器是用硬浮点还是软浮点来算的呢

magicboatrocket

呵呵，确实是比较头疼的问题


imx31工具链，没有看到--with-float=soft的选项，不过感觉这个工具链相对montavista工具链比较简单

[root@localhost ~]# arm-linux-gcc  -v
Reading specs from /opt/mtwk/usr/local/gcc-3.4.3-glibc-2.3.4/arm-linux/lib/gcc/arm-linux/3.4.3/specs
Configured with: /usr/src/RPM/BUILD/crosstool/source/gcc-3.4.3/configure --target=arm-linux --host=i686-host_pc-linux-gnu --prefix=/opt/mtwk/usr/local/gcc-3.4.3-glibc-2.3.4/arm-linux --with-headers=/opt/mtwk/usr/local/gcc-3.4.3-glibc-2.3.4/arm-linux/arm-linux/include --with-local-prefix=/opt/mtwk/usr/local/gcc-3.4.3-glibc-2.3.4/arm-linux/arm-linux --disable-nls --enable-threads=posix --enable-symvers=gnu --enable-__cxa_atexit --enable-languages=c,c++ --enable-shared --enable-c99 --enable-long-long
Thread model: posix
gcc version 3.4.3
[root@localhost ~]#

DM6446带的montavista的工具链，有--with-float=soft，但事实是用它编译出来的程序运行起来比imx31带的工具链编译出来的程序速度快。

[root@localhost bin]# ./arm-linux-gcc -v
Reading specs from /opt/mv_pro_4.0/montavista/pro/devkit/arm/v5t_le/bin/../lib/gcc/armv5tl-montavista-linuxeabi/3.4.3/specs
Configured with: ../configure --host=i686-pc-linux-gnu --target=armv5tl-montavista-linuxeabi --prefix=/opt/montavista/foundation/devkit/arm/v5t_le --exec-prefix=/opt/montavista/foundation/devkit/arm/v5t_le --bindir=/opt/montavista/foundation/devkit/arm/v5t_le/bin --sbindir=/opt/montavista/foundation/devkit/arm/v5t_le/sbin --sysconfdir=/opt/montavista/foundation/devkit/arm/v5t_le/etc --datadir=/opt/montavista/foundation/devkit/arm/v5t_le/share --includedir=/opt/montavista/foundation/devkit/arm/v5t_le/include --libdir=/opt/montavista/foundation/devkit/arm/v5t_le/lib --libexecdir=/opt/montavista/foundation/devkit/arm/v5t_le/libexec --localstatedir=/opt/montavista/foundation/devkit/arm/v5t_le/var --sharedstatedir=/opt/montavista/foundation/devkit/arm/v5t_le/share --mandir=/opt/montavista/foundation/devkit/arm/v5t_le/man --infodir=/opt/montavista/foundation/devkit/arm/v5t_le/info --build=i686-pc-linux-gnu --program-transform-name=s,^,arm_v5t_le-, --enable-cross --with-sysroot=/opt/montavista/foundation/devkit/arm/v5t_le/target --enable-shared --enable-languages=c,c++ --enable-__cxa_atexit --enable-threads=posix --disable-multilib --with-gxx-include-dir=${prefix}/lib/gcc/armv5tl-montavista-linuxeabi/3.4.3/../../../../target/usr/include/c++/3.4.3 --with-numa-policy=yes --with-float=soft --with-cpu=arm10tdmi --with-interwork --with-arch=armv5t --with-tune=arm10tdmi --libexecdir=/opt/montavista/foundation/devkit/arm/v5t_le/lib
Thread model: posix
gcc version 3.4.3 (MontaVista 3.4.3-25.0.30.0501131 2005-07-23)
[root@localhost bin]#

magicboatrocket

用arm-linux-objdump -p 查看，硬浮点

[root@localhost lib]# arm-linux-objdump -p  libjpeg.a
In archive libjpeg.a:

jcapimin.o:     file format elf32-littlearm
private flags = 0: [APCS-32] [FPA float format]

magicboatrocket

从老外那里得知
gcc 4.2 toolchain has support to do ARM vfp
the compiler options you would want to use would be: -mfpu=vfp -mfloat-abi=softvfp

我的toolchain才3.4.3 版本的，应该还不直持硬件的浮点单元。
上面用arm-linux-objdump -p 看格式是错误的方法。

关键部分我再慢慢详细的测一下

1. 
   
2. static void timestamp( char* buffer) {
   
3. 
   
4.     static int startSecond = 0;
   
5.     static int startMs = 0;
   
6.     struct timeval tv;
   
7.     int deltaSecond, deltaMs;
   
8. 
   
9.     gettimeofday(&tv, NULL);
   
10. 
    
11.     /* Running for the first time? */
    
12.     if(startSecond == 0) {
    
13. 
    
14.         /* Copy to prev so that we get 0 delta. */
    
15.         startSecond = tv.tv_sec;
    
16.         startMs = tv.tv_usec;
    
17. 
    
18.     }
    
19. 
    
20.     /* Calculate the delta (in microseconds). */
    
21.     deltaSecond = tv.tv_sec - startSecond;
    
22.     deltaMs = tv.tv_usec - startMs;
    
23. 
    
24. 
    
25. 
    
26.     /* Create the string giving offset from start in seconds. */
    
27.        snprintf(buffer, BUFFER_SIZE, "%u.%u",deltaSecond,deltaMs);
    
28. }
    

复制代码

[ 本帖最后由 magicboatrocket 于 2009-10-21 23:18 编辑 ]

magicboatrocket

问题解决

感谢 “ARM VFP的一点体会”文档的作者，写得相当棒！！！
http://linux.chinaunix.net/bbs/thread-1125926-1-1.html

利用这个文档，我找到一个支持vfp的4.1版本的gcc。把所有库，程序
加 CFLAGS="-mfpu=vfp -mfloat-abi=softfp" ，用新的toolchian重新遍过。
现在性能满足了。

居然有差不多20 ~ 25倍的差距:wink:
VFP硬浮点奶真强大也！