On 02/11/2019 02:51 PM, Andy Ritger wrote:
On Fri, Feb 08, 2019 at 03:43:25PM -0700, Kyle Brenneman wrote:
On 2/8/19 2:33 PM, Andy Ritger wrote:
On Fri, Feb 08, 2019 at 03:01:33PM -0500, Adam Jackson wrote:
On Fri, 2019-02-08 at 10:19 -0800, Andy Ritger wrote:
(1) If configured for PRIME GPU offloading (environment variable or
application profile), client-side libglvnd could load the possible
libGLX_${vendor}.so libraries it finds, and call into each to
find which vendor (and possibly which GPU) matches the specified
string. Once a vendor is selected, the vendor library could optionally
tell the X server which GLX vendor to use server-side for this
client connection.
I'm not a huge fan of the "dlopen everything" approach, if it can be
avoided.
Yes, I agree.
I'm pretty sure libglvnd could avoid unnecessarily loading vendor libraries
without adding nearly so much complexity.
If libglvnd just has a list of additional vendor library names to try, then
you could just have a flag to tell libglvnd to check some server string for
that name before it loads the vendor. If a client-side vendor would need a
server-side counterpart to work, then libglvnd can check for that. The
server only needs to keep a list of names to send back, which would be a
trivial (and backward-compatible) addition to the GLXVND interface.
Also, even without that, I don't think the extra dlopen calls would be a
problem in practice. It would only ever happen in applications that are
configured for offloading, which are (more-or-less by definition)
heavy-weight programs, so an extra millisecond or so of startup time is
probably fine.
But why incur that loading if we don't need to?
As I noted, we can still avoid loading extra loads even with an (almost)
strictly client-based design. You don't need to do any sort of
server-based device enumeration, all you need is something in the server
to add a string to a list that the client can query.
But, there's no reason that query can't be optional, and there's no
reason it has to be coupled with anything else.
I think I'd rather have a new enum for GLXQueryServerString
that elaborates on GLX_VENDOR_NAMES_EXT (perhaps GLX_VENDOR_MAP_EXT),
with the returned string a space-delimited list of <profile>:<vendor>.
libGL could accept either a profile or a vendor name in the environment
variable, and the profile can be either semantic like
performance/battery, or a hardware selector, or whatever else.
This would probably be a layered extension, call it GLX_EXT_libglvnd2,
which you'd check for in the (already per-screen) server extension
string before trying to actually use.
That all sounds reasonable to me.
At the other extreme, the server could do nearly all the work of
generating the possible __GLX_VENDOR_LIBRARY_NAME strings (with the
practical downside of each server-side GLX vendor needing to enumerate
the GPUs it can drive, in order to generate the hardware-specific
identifiers).
I don't think this downside is much of a burden? If you're registering
a provider other than Xorg's you're already doing it from the DDX
driver (I think? Are y'all doing that from your libglx instead?), and
when that initializes it already knows which device it's driving.
Right. It will be easy enough for the NVIDIA X driver + NVIDIA server-side GLX.
Kyle and I were chatting about this, and we weren't sure whether people
would object to doing that for the Xorg GLX provider: to create the
hardware names, Xorg's GLX would need to enumerate all the DRM devices
and list them all as possible <profile>:<vendor> pairs for the Xorg
GLX-driven screens. But, now that I look at it more closely, it looks
like drmGetDevices2() would work well for that.
So, if you're not concerned with that burden, I'm not. I'll try coding
up the Xorg GLX part of things and see how it falls into place.
That actually is one of my big concerns: I'd like to come up with something
that can give something equivalent to Mesa's existing DRI_PRIME setting, and
requiring that logic to be in the server seems like a very poor match. You'd
need to take all of the device selection and enumeration stuff from Mesa and
transplant it into the Xorg GLX module, and then you'd need to define some
sort of protocol to get that data back into Mesa where you actually need it.
Or else you need to duplicate it between the client and server, which seems
like the worst of both worlds.
Is this actually a lot of code? I'll try to put together a prototype so
we can see how much it is, but if it is just calling drmGetDevices2() and
then building PCI BusID-based names, that doesn't seem unreasonable to me.
The fact that it's required *at all* tells you that a server-based
design doesn't match the reality of existing drivers. I've also seen
ideas for GLX implementations based on EGL or Vulkan, which probably
wouldn't be able to work with server-side device enumeration.
And like I pointed out, adding that requirement doesn't give you
anything that you can't do with a client-based interface.
By comparison, if libglvnd just hands the problem off to the vendor
libraries, then you could do either. A vendor library could do its device
enumeration in the client like Mesa does, or it could send a request to
query something from the server, using whatever protocol you want --
whatever makes the most sense for that particular driver.
More generally, I worry that defining a (vendor+device+descriptor) list as
an interface between libglvnd and the server means baking in a lot of
unnecessary assumptions and requirements for drivers that we could otherwise
avoid without losing any functionality.
Is the GLX_VENDOR_MAP_EXT string ajax described that constraining?
In and of itself, maybe, maybe not. Requiring that it exist in the first
place (and all of the requirements that it implies) is the problem.
But, something like that might work for a way to let libglvnd filter out
client-side vendors if we wanted more granularity than just checking for
the existence of a server-side vendor. In that case, though, the
requirements are much looser. A list of strings is all you'd need, and a
vendor only has to use it if it makes sense to do so.
Also, is Mesa the only client-side vendor library that works with the Xorg
GLX module? I vaguely remember that there was at least one other driver that
did, but I don't remember the details anymore.
Two follow-up questions:
(1) Even when direct-rendering, NVIDIA's OpenGL/GLX implementation sends
GLX protocol (MakeCurrent, etc). So, we'd like something client-side
to be able to request that server-side GLXVND route GLX protocol for the
calling client connection to a specific vendor (on a per-screen basis).
Do you think it would be reasonable for GLX_EXT_libglvnd2 to define a
new protocol request, that client-side libglvnd uses, and sends either
the profile or vendor name from the selected '<profile>:<vendor>'?
(2) Who should decide which vendor/gpu gets the semantic name
"performance" or "battery"? They are relative, so I don't know that
vendors can decide for themselves in isolation. It kind of feels
like it should be GLXVND's job, but I don't know that it has enough
context to infer. I'm curious if anyone else has ideas.
Ultimately, I think this should be up to the user. If you've got a system
with two high-power GPU's, then both of them would have a reasonable claim
for "performance," but a user ought to be able to designate one or the
other.
It should definitely be user configurable, but it would be nice if
we could make a reasonable inference in the case that the user hasn't
provided explicit configuration.
Thanks,
- Andy
- ajax
_______________________________________________
mesa-dev mailing list
mesa-...@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
xorg-devel@lists.x.org: X.Org development
Archives: http://lists.x.org/archives/xorg-devel
Info: https://lists.x.org/mailman/listinfo/xorg-devel
