Information about Nvidia optimus support

[Mahler122]

Hello All,

I've been browsing the web for information regarding muxless graphics (aka Optimus) with new NVIDIA discrete graphics chips. The results have been alright, however I feel like there is a lack of aggregation of this information.

My setup is the following:

MSI GT70 laptop with,
Ivy Bridge i7-3610QM cpu with HD4000 iGPU (integrated GPU), and
nvidia GTX 675m dGPU (discrete GPU)

Currently with the latest nvidia binary blob and bumblebee, I get the following error when I try to use 'optirun'

[ERROR]Cannot access secondary GPU - error: [XORG] (EE) NVIDIA(0): Failed to assign any connected display devices to X screen 0

[ERROR]Aborting because fallback start is disabled.

This shows you that the issue is a little more deep than just nvidia not providing support. It seems like some changes are needed to Xorg in order to allow this kind of thing to work.

In fact it seems that the way a lot of laptops like mine with gtx 6xx discrete chips are wired, is for the dgpu to copy it's rendered image into the memory of the igpu on intel's chip. This is new, and there will need to be some software changes in linux before proper drivers can be written. There is already progress being made though on both fronts it seems!

In the kernel world, there is 'DMA-BUF', which would be the ideal case. This way NVIDIA could do something much closer to what is happening on windows. However there are the much heard of conflicts with the closed-source nature of NVIDIA's software.

Another way is being pursued by david airlie in the form of enhancements to the xorg server and freedesktop drivers to allow gpu offloading, which he is calling PRIME. here, here, and here.

I'm sitting here eager to test new software to get this stuff working however I'm not sure what is current or if anyone has tried various things before, so I'm creating this thread for this purpose. If anybody has any information about new software please post it here. I should mention that I would like to try the prime notes mentioned here like this guy, however I get the following error when cloning the mesa git tree he has.

matthew@sibelius:~/Sources$ git clone git://people.freedesktop.org/~airlied/mesa mesa_prime
Cloning into 'mesa_prime'...
remote: Counting objects: 496561, done.
remote: error: unable to find 536cb264f64e90e4123ad30b4cf0a155244c0eb0
remote: error: unable to find 77ed5b4ff21d5a99313380db94ff58b6d2a0be6a
remote: error: unable to find a91bdb3499de0eef0197803410a19f53f70cad51
remote: Compressing objects: 100% (91027/91027), done.
remote: fatal: unable to read 536cb264f64e90e4123ad30b4cf0a155244c0eb0
remote: aborting due to possible repository corruption on the remote side.
fatal: early EOF
fatal: index-pack failed

So, I'm waiting until something I can test is ready.

I will edit this post to make it more readable and contain more information such as links as they become available. Hopefully soon I will be able to post how to get it working.

[Lekensteyn]

Mesa is just one patch. Clone from git://anongit.freedesktop.org/mesa/mesa and apply this patch on it (I'm using the master branch, but 8.x.x should be fine too).

Something to keep in mind: when building the xf86-video-{intel,nouveau} drivers, you need to install the new X stack first. Otherwise, the compilation will pick up the old headers, causing X to segfault.

[Mahler122]

Actually, I found out from airlie himself that his mesa git repository is a remote of the real mesa tree so I should first clone the original mesa tree, then attach his repo as a remote, then pull from it. This worked.

However I'm now having typical build problems which result from not knowing exactly what you're doing..

Maybe I'm being a bit too optimistic, but I'm trying to build X alongside the X which ubuntu set up. I figured I could isolate the build environment, build, then setup the correct LD_LIBRARY_PATH and PATH variables so that when I call X, it calls the new version. I could then shut down my current X, change environment and start the new X. Do you think it's possible to actually have this work in the end?

Right now it's crashing and giving an error related to XKB.

[idoitprone]

Two things

first install bbswitch

sudo apt-get install bbswitch

this allow you to shutoff your nvidia card when it not in use

believe me, it will stop that noisy fan

second

I only got nvidia optimus working on open drivers. Nvidia support just sucks

[dino99]

Please send your complaint to nvidia for their documentation helping on linux.

http://www.phoronix.com/scan.php?pag...tem&px=MTEyMTc

[Mahler122]

Two things

first install bbswitch

I've had that installed for a while, although cooling on this laptop is pretty good so The fan wasn't very much louder before I had bbswitch.

I only got nvidia optimus working on open drivers. Nvidia support just sucks

Yeah, at the moment I'm just trying to reproduce the open source work. I want to see that this works in principle on my machine too.

[idoitprone]

cat /proc/acpi/bbswitch

?


nouveau driver does not support your card yet

[Mahler122]

matthew@sibelius:~$ cat /proc/acpi/bbswitch 
0000:01:00.0 OFF

nouveau driver does not support your card yet

This is unclear:
http://nouveau.freedesktop.org/wiki/CodeNames#NVC0
Mine is code NVCE

http://nouveau.freedesktop.org/wiki/InstallDRM#Firmware

My card code is not explicitly listed as needing proprietary firmware, but then again it's also not explicitly listed as having free firmware. It's in the NVC0 family though, so the claim is that the free firmware will work.

As far as I can tell there aren't any muxed systems with a gtx 675m with which people can appropriately test nouveau. It seems like everyone has the same problem as me with X. But then again perhaps I haven't searched deep enough.

Thus the first thing to do is see if I can get this xserver working before the release of xorg 1.13 which now looks like it will have the gpu offloading support: http://www.phoronix.com/scan.php?pag...tem&px=MTEzMzE

[Lekensteyn]

However I'm now having typical build problems which result from not knowing exactly what you're doing..

Maybe I'm being a bit too optimistic, but I'm trying to build X alongside the X which ubuntu set up. I figured I could isolate the build environment, build, then setup the correct LD_LIBRARY_PATH and PATH variables so that when I call X, it calls the new version. I could then shut down my current X, change environment and start the new X. Do you think it's possible to actually have this work in the end?

Right now it's crashing and giving an error related to XKB.

To fix the xkb issue, install "libxkbfile-dev" (may be necessary before building X).
You also need to rebuild the evdev module (and any other X module you use, e.g. the "xf86-input-wacom" module for Wacom tables) otherwise the devices will be unusuable, unabling you to use X because of the ABI mismatch.
My notes for PRIME are here: http://lekensteyn.nl/files/prime-instructions.txt

[idoitprone]

matthew@sibelius:~$ cat /proc/acpi/bbswitch 
0000:01:00.0 OFF

This is unclear:
http://nouveau.freedesktop.org/wiki/CodeNames#NVC0
Mine is code NVCE

http://nouveau.freedesktop.org/wiki/InstallDRM#Firmware

My card code is not explicitly listed as needing proprietary firmware, but then again it's also not explicitly listed as having free firmware. It's in the NVC0 family though, so the claim is that the free firmware will work.

As far as I can tell there aren't any muxed systems with a gtx 675m with which people can appropriately test nouveau. It seems like everyone has the same problem as me with X. But then again perhaps I haven't searched deep enough.

Thus the first thing to do is see if I can get this xserver working before the release of xorg 1.13 which now looks like it will have the gpu offloading support: http://www.phoronix.com/scan.php?pag...tem&px=MTEzMzE

Nouveau team barely added remedial support of the kepler series. It is not worth the effort to run nouveau drivers. Graphic cards are very complex these days and those gpu clocks are hard to reverse engineer.
You will not get acceptable performance because the starting clock is too slow.