[Open-graphics] HELP: Weird discrepancy between fbZPutImage and memcpy

[Stephen Warren]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Timothy Normand Miller wrote:
> On 7/11/07, Stephen Warren <swarren at wwwdotorg.org> wrote:
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> Timothy Normand Miller wrote:
>> > Note:  We looked at it with a PCI bus analyzer, and what we see is
>> > that the x.org code results in long bursts, while ours results in
>> > individual transactions, which would explain why we're so slow.  What
>> > could x.org code be doing that would result in bursts, while ours
>> > doesn't?
>>
>> Probably, the dcache generates bursts during write-back/flush and CPU
>> generates single-beat accesses when writing to uncached areas (or
>> write-through cache?)
> 
> Is there something you can do when mapping memory that will set
> whether it's cached or not?  We're just doing a simple mmap in our
> test software, while we're using some xf86 function in the X server.

I imagine so, but I'm sorry that I don't know exactly what, beyond
anything the "man mmap" might say.

One thing to check; PCI BARs have a flag that says whether they're
"memory like" (or something like that) which could automatically affect
the region's cache-ability (either in HW, or due to the kernel's
interpretation of that data). I assume that the device's registers and
frame buffer memory are in separate PCI bars; that would be a sensible
setup.

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2 (MingW32)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFGlWlkhk3bo0lNTrURAmNWAKCKnfhoysyneC8BWHX8JlTfDniJ8ACfQzqr
BdnABB6JFmr7Cxz20L2hico=
=UGco
-----END PGP SIGNATURE-----