AMD engineer admits Linux graphics stack needs improvement

[or1k]

When discussing the bug about the relatively high power consumption on AMD APUs supporting hardware video decoding compared to Windows, an AMD engineer, Alex Deucher, the primary developer of the amdgpu driver, admitted that displaying video in Linux is fundamentally inefficient.

Linux video output now uses the following chain:

- Compressed video stream
- VCN (hardware video decoding module for AMD GPUs)
- Raw YUV data
- Palette conversion, scaling on the GFX module (essentially a 3D gas pedal in the GPU, which forces it to increase core and VRAM frequencies)
- RGB data
- Display output.

 

How it should work:

- Compressed video stream
- VCN
- Raw YUV data
- A display controller that will convert the palette, scale and display.

This can be solved more efficiently in Wayland compositors, but there is no implementation yet. This problem is solved in Microsoft Windows and Google Android, for there are full-fledged single compositors that provide the appropriate capabilities and APIs - something that is not yet available in Linux, because neither X.org nor Wayland can handle YUV streams directly.