[cairo] Very slow performance

[Soeren Sandmann]

Keith Packard <keithp at keithp.com> writes:

> Around 18 o'clock on Dec 28, Soeren Sandmann wrote:
> 
> > In my tests when I did the MMX stuff in xorg, 128 bit reads/writes
> > were not an improvement over 64 bit read/writes on the systems I
> > tested on. On some - but not all - systems, 64 bit was an improvement
> > over 32 bit.
> 
> Were these tests to AGP or PCI video memory?  
> Or main machine memory?

They were done on AGP memory I think (on a laptop with an i830
chipset), so perhaps not that relevant for Cairo.

> I have been lead to believe that MMX/SSE copy and fill code can run
> faster than regular instructions because of the way they connect to
> the main memory interface.

The improvement from the MMX code was much greater on drawables in
system memory than in AGP space, so this does sound plausible to me.

I am attaching the benchmark I used. It should be pretty easy to
modify to test main memory.

Compile it with 

        gcc `pkg-config --cflags --libs glib-2.0` fbread.c 

Søren

-------------- next part --------------
A non-text attachment was scrubbed...
Name: fbread.c
Type: application/octet-stream
Size: 4395 bytes
Desc: benchmark
Url : http://lists.freedesktop.org/archives/cairo/attachments/20041228/d0c63c7b/fbread.obj