Re: [Libvisual-devel] Next release cycle plans!

[Dennis S. <sy...@yo...>]

On Fri, 2005-02-11 at 20:44 -0300, Duilio Javier Protti wrote:
> > 		A better autogen.sh, and review our build trees.
> 
> I can work on autogen.sh this weekend.

Excelent!

> > 	Libvisual:
> > 		Setup our own memcpy/memset wrappers and accelerate
> > 			these with mmx/sse etc etc.
> 
> I see on the Linux kernel (include/asm-i386/string.h):
> 
> static __inline__ void *__memcpy3d(void *to, const void *from, size_t
> len)
> {
> 	if (len < 512)
> 		return __constant_memcpy(to, from, len);
> 	return _mmx_memcpy(to, from, len);
> }
> 
> so it looks like MMX pays well on buffers > 512 bytes long

I did some serious playing with this a few weeks ago. MMX does matter, 
because you can load store 8 * 8 bytes, (all the mmx regs). However
what does matter more is cache control, by using 3dnow, or sse.

That way I can pump 2 gigabyte a second on my machine, without cache 1 
gigabyte would take 1.4 sec.


> I must finish the gtk1-widget!!! I will work on this

Sounds good, I changed a lot in the gtk2 version, (not one
big function, but a function for every widget). You want to look
at this.