Libmpv hardware decoding performances on AMD GPU and Linux

[leon-costa]

Hello,

When I render a video to an OpenGL framebuffer with libmpv on an AMD GPU on Linux I cannot find any hwdec that has performances similar to using nvdec or cuda on an Nvidia GPU.

The video I tested with is encoded with HEVC at 5400x2700. On an Nvidia 1080 with hwdec=nvdec the call to mpv_render_context_render takes about 0.5 ms. With an AMD Radeon RX 7900 I cannot get anything less than 2.7 ms with any of the hwdec listed in the mpv documentation, which is also the timing I get when I select no as hwdec. (I can also get worse: with vaapi-copy or vdpau-copy I get about 3.3 ms)

Is it possible to get render times as good as with nvdec on an AMD GPU on Linux? Am I doing something wrong?

With the mpv CLI player I get variable "frame timings" when I play the same video and press "I" (I'm not sure what these timings represent though). The best average frame timing I could get with the AMD GPU is about 900 µs with hwdec=vdpau. With hwdec=vaapi I get about 1500 µs average. So it looks like vdpau could provide good performances. Strangely, with the Nvidia 1080 I get worse average timings with --hwdec=cuda -vo=gpu: about 2400 µs, so I'm not sure these values are relevant.

I'm using mpv 0.39.0.