[Cegcc-devel] slowness of memcpy

[Vincent T. <vt...@un...>]

Hey,

I'm porting to Windows CE a canvas library which uses the GAPI library as 
backend. It uses a lot the memcpy function (I use a backbuffer and copy 
that buffer to the frame buffer of the device). I've looked on internet 
about fast memcpy functions (asm optimized), as i've also read that memcpy 
is slow on arm.

I've found an asm implementation of memcpy which uses arm v5te (at least) 
instructions and compared it with memcpy which is in cegcc 0.51. I've 
attached a test program. Just run 'make' to create the program. Note that 
I compile with -march=armv5te -O3 -fomit-frame-pointer. I don't know if 
it's the best flags for best performance wrt that test.

On big data (like 320 * 240 * 2, like when one copies the entire screen of 
a qvga device, which is my case), the bandwidth is around 23 MB/s with the 
cegcc memcpy, and around 70 MB/s with the asm implementation. My device is 
the Ipaq H3970. So on that device, it's almost 3 times faster than memcpy 
in newlib. Even on non aligned data.

I use a lot memcpy in that canvas library, but I think that that function 
is used a lor around and everyone would benefit such optimized 
implementation.

My question is then: would it be possible to optimized memcpy with or 
without using that asm implementation of memcpy, as it is very slow ? If 
you want to use that asm implementation, look at the copyright notice in 
memcpy_libc.S

I've quickly compared with the memcpy which is in uclibc and it seems that 
this one is slower than memcpy_libc.S, but I have maybe did wrong 
modifications to make it work. The uclibc one is nevertheless faster than 
std implmentation, especially on non aligned data:

~20 MB/s with newlib
~27 MB/s with uclibc
~70 MB/s with memcpy_libc.S (strange though, the difference is really 
huge)

regards

Vincent Torri