Le 16.01.2004 19:41:39, Mark Vojkovich a écrit :
On Thu, 15 Jan 2004, manu wrote:
> Le 16.01.2004 00:55:48, Mark Vojkovich a ?crit?:
> > I don't see anything out of the ordinary. This looks like
> > perfectly normal data for a graphics engine programmed by
> > memory mapped I/O.
> >
> > Mark.
>
> OK then my mistake. I was just surprised by the bad results of hw
accel
> compared to software ops. I guess this is due to the overhead of
> programming the hw, or am I completely wrong?
The only places that I notice where software is faster than
hardware is for things like 1x1 rectangles. The overhead of
programming the hardware is higher than just writing a single
dword to the framebuffer in that case. Those types of hardware
operations can be faster than software if you are using DMA
and processing large batches at a time, but rarely are for memory
mapped I/O.
> Anyway does this is taken into account when drawing something? I
mean
> it is sometimes stupid to use hw accel for some ops, are the
drivers
> able to do that correctly?
Mixing software and hardware rendering is bad because software
and hardware rendering can only be synchronized by waiting for the
hardware to go idle. Attempts to improve x11perf performance by
falling back to software for small primitives will generally hurt
real-life performance.
Thanks for the clarifications, now I understand.
Bye
Manu
_______________________________________________
XFree86 mailing list
XFree86@xxxxxxxxxxx
http://XFree86.Org/mailman/listinfo/xfree86
