EGL doesn't work with NVIDIA Tesla

[Ingun Jon]

I run a test for my WebGL application using Chromium.

I give the argument --use-gl=egl to Chromium so it gets the GPU acceleration in headless mode.

I ran it in various environments that has GPU and it all worked fine with high fps. (e.g 60 on MacOS, 50~55 on Ubuntu on ec2 g4dn ...)

But it doesn't work in AWS Codebuild LINUX_GPU_CONTAINER environment. It seems to fall back to default software renderer SwiftRenderer judging by it's poor fps rate 1~2. It's the same number I get from other environment when I use SwiftRenderer.

Codebuild LINUX_GPU_CONTAINER's got Tesla gpu with the driver

[Container] 2020/11/27 11:01:20 Running command nvidia-smi 392 Fri Nov 27 11:01:20 2020 393 +-----------------------------------------------------------------------------+ 394 | NVIDIA-SMI 418.87.00 Driver Version: 418.87.00 CUDA Version: N/A | 395 |-------------------------------+----------------------+----------------------+ 396 | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | 397 | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | 398 |===============================+======================+======================| 399 | 0 Tesla V100-SXM2... On | 00000000:00:1B.0 Off | 0 | 400 | N/A 37C P0 37W / 300W | 0MiB / 16130MiB | 0% Default | 401 +-------------------------------+----------------------+----------------------+ 402 | 1 Tesla V100-SXM2... On | 00000000:00:1C.0 Off | 0 | 403 | N/A 39C P0 36W / 300W | 0MiB / 16130MiB | 0% Default | 404 +-------------------------------+----------------------+----------------------+ 405 | 2 Tesla V100-SXM2... On | 00000000:00:1D.0 Off | 0 | 406 | N/A 39C P0 38W / 300W | 0MiB / 16130MiB | 0% Default | 407 +-------------------------------+----------------------+----------------------+ 408 | 3 Tesla V100-SXM2... On | 00000000:00:1E.0 Off | 0 | 409 | N/A 40C P0 42W / 300W | 0MiB / 16130MiB | 0% Default | 410 +-------------------------------+----------------------+----------------------+ 411 412 +-----------------------------------------------------------------------------+ 413 | Processes: GPU Memory | 414 | GPU PID Type Process name Usage | 415 |=============================================================================| 416 | No running processes found | 417 +-----------------------------------------------------------------------------+

I suspect that chromium should link to NVIDIA's egl implementation that is shipped with NVIDIA driver instead of mesa or angle's... but I don't know how to test it.

How can I enable the gpu acceleration?

[Ingun Jon]

I found that chromium is trying to connect to XCB and fails.

[Container] 2020/11/30 16:14:19 Phase complete: PRE_BUILD State: SUCCEEDED

[Container] 2020/11/30 16:14:19 Phase context status code:  Message:

[Container] 2020/11/30 16:14:19 Entering phase BUILD

[Container] 2020/11/30 16:14:19 Running command dbus-run-session -- bash -c 'DBUS_SYSTEM_BUS_ADDRESS="$DBUS_SESSION_BUS_ADDRESS" node index.js'

[1130/161419.945622:INFO:cpu_info.cc(53)] Available number of cores: 32

[1130/161419.945660:INFO:cpu_info.cc(53)] Available number of cores: 32

[1130/161419.945692:VERBOSE1:zygote_main_linux.cc(217)] ZygoteMain: initializing 0 fork delegates

[1130/161419.945718:VERBOSE1:zygote_main_linux.cc(217)] ZygoteMain: initializing 0 fork delegates

[1130/161419.994646:VERBOSE1:webrtc_internals.cc(118)] Could not get the download directory.

[1130/161419.999335:ERROR:gl_surface_egl.cc(772)] EGL Driver message (Critical) eglInitialize: xcb_connect failed

[1130/161419.999454:ERROR:gl_surface_egl.cc(772)] EGL Driver message (Critical) eglInitialize: xcb_connect failed

[1130/161419.999517:ERROR:gl_surface_egl.cc(772)] EGL Driver message (Critical) eglInitialize: xcb_connect failed

[1130/161419.999580:ERROR:gl_surface_egl.cc(772)] EGL Driver message (Critical) eglInitialize: xcb_connect failed

[1130/161419.999636:ERROR:gl_surface_egl.cc(772)] EGL Driver message (Error) eglInitialize: eglInitialize

[1130/161419.999692:ERROR:gl_surface_egl.cc(1313)] eglInitialize Default failed with error EGL_NOT_INITIALIZED

[1130/161419.999752:ERROR:gl_initializer_linux_x11.cc(160)] GLSurfaceEGL::InitializeOneOff failed.

[1130/161419.999818:VERBOSE1:gpu_init.cc(361)] gl::init::InitializeGLNoExtensionsOneOff failed

[1130/161420.000941:VERBOSE1:device_data_manager_x11.cc(216)] X Input extension not available

[1130/161420.001793:VERBOSE2:vaapi_video_encode_accelerator.cc(240)] VaapiVideoEncodeAccelerator():

[1130/161420.001886:ERROR:viz_main_impl.cc(150)] Exiting GPU process due to errors during initialization

[1130/161420.001904:VERBOSE2:vaapi_video_encode_accelerator.cc(966)] DestroyTask():

[1130/161420.002048:VERBOSE2:vaapi_video_encode_accelerator.cc(256)] ~VaapiVideoEncodeAccelerator():

[1130/161420.004241:VERBOSE2:thread_state.cc(415)] [state:0x55d4ad9a2100] ScheduleGCIfNeeded

[1130/161420.010979:VERBOSE2:thread_state.cc(415)] [state:0x55d4ad9a2100] ScheduleGCIfNeeded

[1130/161420.011165:VERBOSE2:thread_state.cc(415)] [state:0x55d4ad9a2100] ScheduleGCIfNeeded

[1130/161420.011326:VERBOSE1:va_stubs.cc(779)] dlopen(libva.so.2) failed.

[1130/161420.011416:VERBOSE1:va_stubs.cc(780)] dlerror() says:

libva.so.2: cannot open shared object file: No such file or directory

To my understanding some X-server independent EGL implementation should be used in headless mode... I don't understand why XCB's in the way.

The environment is

Cloud Service: AWS Codebuild LINUX_GPU_CONTAINER.

Docker Image: standard 4.0 based on ubuntu:18.04 AS core

[Ken Russell]

Unfortunately Chromium hasn't yet added support for GPU acceleration of headless: http://crbug.com/765284 . It hasn't risen to the top of the priority list. If you can run regular (non-headless) Chromium on cloud instances that have a console, that should work.

[Ingun Jon]

Then can you explain me why I am getting headless gpu acceleration in other environments?

For example, I can get gpu acceleration in p2.xlarge AWS instance. It's got NVIDIA Tesla k80 and the driver. I double-checked by getting report from chrome://gpu and measuring fps of my own product. (report in the attachment)

I run chromium with following CLI arguments

execFileSync(chromiumBinPath, ["--no-sandbox", "--headless", "--use-gl=egl", "--print-to-pdf", "--enable-logging=stderr", "--v=999", "chrome://gpu"])

[Ingun Jon]

...and the environment is

Ubuntu 18.04 LTS (Not a docker image)

The hardware acceleration only activates when I have the NVIDIA driver installed.

[Ken Russell]

Without adding a bunch of logging to Chromium and setting up and deploying that configuration, I can't guess which code path this is taking nor why it's taking a different code path than the other AWS configuration. Please feel free to add more information to http://crbug.com/765284 with your configurations that are working and not working, and what you expect from each.

-Ken

[Jonathan Cardoso]

Hey Ingun, sorry to resurrect this pretty old thread, but did you have any luck making it work? I am facing a similar issue atm