Intel QSV fills up VRAM and then stops

[l00tzOMG]

Hi,

I did some further testing on the HW decoding capabilities, and I experienced the following. Using an Intel UHD770 GPU and prefer_hw=3 (so it is using the QSV decoder), the GPU starts to fill up the VRAM without flushing it and eventually stops decoding when it runs out of all VRAM (dedicated+shared). This does not happen when DXVA2 is enforced (e.g., prefer_hw=4). I guess it's an Intel driver bug (latest Intel driver and W11 x64) and probably not LMASH related, but would it be possible to change the order from QSV>DXVA2 to the other way around for Intel GPUs? DXVA2 is also significantly faster.

Thanks again!

[Asd-g]

This order (CUDA->QSV->DXVA2...) is for every GPU vendor (NVIDIA, Intel, AMD..).
The order CUDA->QSV->software is probably since hw decoding is added. I just preserve the order and added the new options. I will keep the order the same because people are more or less used to it.
Actually I will check if it's relatively easy to add option so one can define own order.

Can you test 1194.0.0.0 if you still have VRAM leaking?

I tested QSV decoder for h264 video and I have no issues. Also it's ~3-4x faster than DXVA2 for me.

[l00tzOMG]

Thanks for the reply. With the older 1194 version, the VRAM stays at a solid 800 MB and does not increase at all.

[Asd-g]

I forgot to mention that r1194 has bug with prefer_hw=2 (I don't remember if prefer_hw=3 is ok). Use Subtitle(LWLDECORER) to see the actual decoder.

[l00tzOMG]

Thanks, I have used =3 and for this Subtitle(LWLDECODER) returns h264. After changing to =2 it remained the same.

[l00tzOMG]

I also tried the version linked in the attached feed (#77). That returns h264_qsv for both cases. With this, the memory leak occurs as well as in the latest release.

On the side, I forgot to mention that I'm using Prefetch(12) during all my tests, if that matters.

[Asd-g]

With r1194 using LWLibavVideoSource(..., decoder="h264_qsv") - do you have vram leak?

[l00tzOMG]

Yes, with that switch r1194 also starts to fill up the memory until it runs out and stops.

[l00tzOMG]

You might be aware of this, but I found this interesting and might be a related issue with QSV memory leak in FFmpeg:

https://trac.ffmpeg.org/ticket/11426

Solution is also included.

[Asd-g]

Thanks for the link.
I have no time lately but I didn't forget this.

[l00tzOMG]

No worries.