search

JeffyCN / libv4l-rkmpp

A rockchip-mpp V4L2 wrapper plugin for chromium V4L2 VDA/VEA
GNU Lesser General Public License v2.1
79 stars 26 forks source link

chromium v121 can't use egl to render video output #18

Closed amazingfate closed 4 months ago

amazingfate commented 4 months ago

I am running kernel 6.1.43 and gpu driver is panfork(mesa userspace driver and kernel arm mali driver) Run chromium v121 with patches from meta-rockchip:

DISPLAY=:0 chromium --enable-logging --vmodule=*media/gpu*=4 /home/jfliu/bbb_sunflower_1080p_60fps_normal.mp4

Here is log from ~/.config/chromium/chrome_debug.log: https://paste.armbian.com/kafubisuwi We can see:

ERROR:shared_image_factory.cc(960)] Could not find SharedImageBackingFactory with params: usage: DisplayRead|Scanout, format: NV12_LEGACY, share_between_threads: 0, gmb_type: platform, debug_label: MailboxVideoFrameConverter_Cid:10_Pid:1962718

EGL rendering related code should be at media/gpu/v4l2/legacy and I don't know if we can use that since google has dropped it. I think this line of patch makes chromium know it can render NV12 without color space conversion, but in fact it can't.

I also meet the same issue when testing mainline kernel v4l2 stateless decoding with the patch line above. Without that patch chromium can make use of libyuvimageprocessor to convert NV12 to AR24.

JeffyCN commented 4 months ago

it works well with the BSP kernel and packages. media/gpu/v4l2/legacy is no longer used, i think it should use media/gpu/v4l2/v4l2_video_decoder_backend_stateful.cc now.

but that doesn't looks like related to this issue, which could due to the difference between panfork and mali DDK(no vulkan, just EGL).

the mali should support importing NV12 texture directly, maybe you can try to add logs in:

gpu/command_buffer/service/shared_image/shared_image_factory.cc
  auto* egl_display = gl::GetDefaultDisplayEGL();
  if (use_gl && egl_display && egl_display->ext->b_EGL_KHR_image_base &&
      egl_display->ext->b_EGL_KHR_gl_texture_2D_image &&
      egl_display->ext->b_EGL_KHR_fence_sync &&
      gl::g_current_gl_driver->ext.b_GL_OES_EGL_image) {
    auto egl_backing_factory = std::make_unique<EGLImageBackingFactory>(
        gpu_preferences, workarounds, feature_info.get());
    factories_.push_back(std::move(egl_backing_factory));
  }

gpu/command_buffer/service/shared_image/egl_image_backing_factory.cc
gpu/command_buffer/service/shared_image/gl_common_image_backing_factory.cc
JeffyCN commented 4 months ago

wait, SCANOUT usage is invalid for egl image factory, so guessing it should use: gpu/command_buffer/service/shared_image/ozone_image_backing_factory.cc

so the OzoneImageBackingFactory::IsSupported returns false, and we need to check into it, please add logs there to find which line returns false.

amazingfate commented 4 months ago

Which ozone platform do you use? The above logs are running under ozone platform x11. When running under ozone platform wayland, I will get:

[3583:3583:0408/184429.803251:ERROR:gbm_wrapper.cc(402)] gbm format not supported: 842094158
[3583:3583:0408/184429.803378:ERROR:gbm_pixmap_wayland.cc(111)] Cannot create bo with format= YUV_420_BIPLANAR
[3583:3583:0408/184429.803503:ERROR:shared_image_factory.cc(971)] CreateSharedImage: could not create backing.
[3583:3583:0408/184429.803607:ERROR:shared_image_stub.cc(246)] SharedImageStub: Unable to create shared image
amazingfate commented 4 months ago

wait, SCANOUT usage is invalid for egl image factory, so guessing it should use: gpu/command_buffer/service/shared_image/ozone_image_backing_factory.cc

so the OzoneImageBackingFactory::IsSupported returns false, and we need to check into it, please add logs there to find which line returns false.

It's here that returning false: https://github.com/chromium/chromium/blob/121.0.6167.212/gpu/command_buffer/service/shared_image/ozone_image_backing_factory.cc#L296

amazingfate commented 4 months ago

CanCreateNativePixmapForFormat will check device_->CanCreateBufferForFormat, and finally check gbm_device_is_format_supported, which failed with format 842094158. Indeed mesa does not support NV12 in gbm. Here is a test program.

amazingfate commented 4 months ago

After adding the patch from https://gitlab.freedesktop.org/mesa/mesa/-/issues/7745, I can play video with rkmpp now. But frame rate is extramely low, especially for av1 videos, only 4 FPS. And cpu usage is high compared with v114.

JeffyCN commented 4 months ago

only av1? how about other video formats? what pixel format did it output?(nv16?)

guessing another renderer issue due to mesa and chromium(triggered software conversion)

amazingfate commented 4 months ago

The av1 video is outputing NV12. Another h264 1080p@60 video hits 60 fps for the a few seconds. Then frame dropped. CPU usage during playing is always high. From rk3588's top there are two chromium process both have above 120% cpu usage. fps log of h264 1080p@60: https://paste.armbian.com/rusabanigo.less image

amazingfate commented 4 months ago

AV1 video is here:

https://github.com/JeffyCN/libv4l-rkmpp/assets/5022901/2d0bfc47-2938-45fb-af8f-190e4f5b48a8

H264 video is from: https://mirrors.tuna.tsinghua.edu.cn/blender/demo/movies/BBB/bbb_sunflower_1080p_60fps_normal.mp4.zip

JeffyCN commented 4 months ago

both of them works well on my rk3588 evb with mali DDK, guessing something wrong with the rendering flow with mesa3d.

so the nv12 dmabuf import and rendering works with old chromium(imported in V4L2VDA)+mesa3d, but failed with new chromium(likely in ui/ozone/common/native_pixmap_egl_binding.cc)+mesa3d

can you add logs in that file's InitializeFromNativePixmap() to confirm?

amazingfate commented 4 months ago

Here is the patch I add: https://paste.armbian.com/iciwoyufux And I get:

[98724:98724:0410/184435.235520:INFO:native_pixmap_egl_binding.cc(172)] Going to check format_ == gfx::BufferFormat::YUV_420_BIPLANAR
[98724:98724:0410/184435.235601:INFO:native_pixmap_egl_binding.cc(191)] format_ == gfx::BufferFormat::YUV_420_BIPLANAR
[98724:98724:0410/184435.235784:INFO:native_pixmap_egl_binding.cc(212)] plane_ == gfx::BufferPlane::DEFAULT
JeffyCN commented 4 months ago

did it create valid eglimage(not null)?

if not, try to get error info from eglGetError

amazingfate commented 4 months ago

If eglimage is null, InitializeFromNativePixmap should return false, then I should get error Unable to initialize binding from pixmap, but there is no such error.

JeffyCN commented 4 months ago

the dmabuf to eglimage should be zero-copy, then it will be attached to ecternal texture, which is zero-copy too, then it will be rendered to surfaces by gpu. so i have no idea what goes wrong in mesa3d, maybe you can try to sample cpu usage with perf and find out which functions make cpu busy

amazingfate commented 4 months ago

Maybe it's the issue of panfork driver. Since panfork is not under good maintaince, I will toggle mainline panthor driver later. But mesa v24.1 does not support NV12 mapping like v23, I have to wait upstream solving this issue.

amazingfate commented 4 months ago

I managed to make mainline panthor gpu driver running with 6.1 rkr1 kernel: https://github.com/armbian/linux-rockchip/pull/171. When playing videos with chromium v121, gpu will crash and chromium falls back to ffmpeg decoder:

[3771:3771:0421/165958.593794:ERROR:command_buffer_proxy_impl.cc(131)] ContextResult::kTransientFailure: Failed to send GpuControl.CreateCommandBuffer.
[3725:3725:0421/165958.593767:ERROR:command_buffer_proxy_impl.cc(323)] GPU state invalid after WaitForGetOffsetInRange.
[3590:3590:0421/165958.620553:ERROR:gpu_process_host.cc(992)] GPU process exited unexpectedly: exit_code=11

When the same mesa userspace panfrost driver runs with mainline kernel, mainline chromium can use v4l2 stateless decoding without crash.

amazingfate commented 4 months ago

Finally I made it work with mainline panthor driver. I don't know why it didn't work before, and I can't reproduce the issue above. Here is my steps to make it work with this armbian image: