[Mplayer-users] fastmemcpy benchmark

[Dag B]

Arpi wrote:
> 
> Hi,
> 
> > Asus P5A, K6-2/500, G200/16MB. Linux 2.4.3.
> >
> > # ./fastmem.sh
> > k6 : v2-v1=915325227 = 1826477us  (54.750fps)   82.1MB/s
> > mmx: v2-v1=790536059 = 1577468us  (63.393fps)   95.1MB/s
> > mga-k6 : v2-v1=3012802007 = 6011851us  (16.634fps)   25.0MB/s
> > mga-mmx: v2-v1=3798322327 = 7579307us  (13.194fps)   19.8MB/s
> >
> > I had an idea about memory bandwidth of this system sucking big time. But
> > *this* bad...??
> Do you have MTRR registers set up? Or your cpu don't have MTRR?

using mgafb. mtrr is set up automatically with mgafb.
# cat /proc/mtrr 
reg00: base=0xe7000000 (3696MB), size=  16MB: write-combining, count=1
reg01: base=0xd0000000 (3328MB), size= 256MB: write-combining, count=1

Does this look right? AGP aperture set to 256MB in BIOS. My k62/500 supports
mtrr.

 
> > How come the mga-* test is so much worse than the others? Does this indicate
> > a bug? (Using the module in latest cvs.)
> It can't be a bug...
:-)

> It's problem with your hardware, especially memory handling or pci/agp bus.
> Try to play with pci retries latency ram speed and other funny things.

PCI retries enabled, overclockmem enabled. (Already was.)
Tuned memory options in BIOS to max. (7ns)

fastmem.sh/mmx gives the best results when running in a framebufferconsole,
with xdm running (not logged in to X) and no mga_vid module loaded. (This
enables pci retries and overclockmem)

In this case I can get 74fps and 111MB/s with the mmx (glibc?) version.
In X, I now get:

> ./fastmem.sh
k6 : v2-v1=981023203 = 1957572us  (51.084fps)   76.6MB/s
mmx: v2-v1=840943184 = 1678053us  (59.593fps)   89.4MB/s
mga-k6 : v2-v1=3010500372 = 6007259us  (16.647fps)   25.0MB/s
mga-mmx: v2-v1=3813158672 = 7608912us  (13.142fps)   19.7MB/s


> Results on my cel2-600 at 900 (100Mhz FSB) + 256mb pc133 + G400MAX 32MB SGRAM:
> 
[depressing results deleted]

> So, unfortunately, hardware DOES matter a lot.
> You can't workaround shit hardware with cool software.

Noted.  :-)
But something is really odd about the results above, because I definitely get
the best results with mplayer when using the mga output module. I'd expect
fastmem.sh to reflect that?!
Also see one of my other postings, indicating that AGP is working very well.

No reason to work ones ass off just to get the benchmarks correct. But how
come your benchmark results reflects reality, while testing with a G200
doesn't?

Dag B
p.s.
mga_vid now behaves nicely with the Xvideo extention. I.e. I can run xawtv in
fullscreen, even when having mga_vid loaded. Nice. Was this intentional?

_______________________________________________
Mplayer-users mailing list
Mplayer-users at lists.sourceforge.net
http://lists.sourceforge.net/lists/listinfo/mplayer-users