I wanted to collect some thoughts on dropping agp_gart.
intel_express is still lugging around agp_gart for modern chipsets.
Classic Intel AGP chipsets exist as PCI devices with an extra
layer to speed up transfers. Without agp_gart the bus simply operates
at the slower pci speeds.[1]
We've been carrying around agp_gart a bit longer than we should of
as intel_extreme performs all of its gtt work through agp_gart.
Since us leveraging 3d hardware rendering on classic Intel AGP
chipsets doesn't seem likely, i'd personally like to see agp_gart
disappear as it layers on lots of added complexity.
If I were to rip out agp_gart (and still have a functional
intel_extreme driver) would anyone care?
-- Alex
[1] http://utah-glx.sourceforge.net/gart/README
(also small chat I had with a linux graphics stack developer)
20:54 < kallisti5> I read @ utah-glx that agp is accessed just like PCI and the
gart is an optional layer
20:55 < kallisti5> right now our intel driver depends on the intel_gart stuff
through haswell+
20:57 < kallisti5> .. which is getting awkward in the pcie + soc age:-)
20:58 <+airlied> kallisti5: you'd have to know the haiku driver architecture to
answer that
20:58 <+airlied> and since you are the only person who knows that :)
20:59 < kallisti5> lol. from a "should it work" standpoint
20:59 <+airlied> we've dropped using a separate gart driver ages ago
20:59 < kallisti5> is it sane to program drivers for old intel chipsets without
gart?
20:59 < kallisti5> aka. we have an intel driver not gart aware. just treats old
intel as pci
21:00 < kallisti5> linux can't really drop gart as some might use it for mesa
accel. but we have minimal accelerated overlay support in our intel driver
21:01 <+airlied> the agp driver has two modes of operation on Linux
21:01 <+airlied> one for integrated GPU, one for AGP chipsets with an discrete
GPU
21:01 <+airlied> in theory since i915 intel the gart driver is redunant
21:02 <+airlied> and you can just have the actual intel driver program the hw
itself
21:02 < kallisti5> ok. that's what I thought
21:02 < kallisti5> our gtt code is in our gart
21:02 < kallisti5> that should be easy to shuffle over to the driver
21:03 <+airlied> you probably want to leave the old code alone
21:03 <+airlied> unless you really don't care about AGP, which you probably
don't since it's hw that is pretty dead