[FFmpeg-user] The option of -hwaccel device doesn't works well(FFmpeg n3.4)

[James Girotti]

Hello,

Let me start by saying that I only have one NVENC capable GPU, so I cannot
test this. I hope to point you in the right direction though.

On Thu, Nov 2, 2017 at 7:38 PM, Baek Seung Hoon <seung51hoon at gmail.com>
wrote:

> Could you tell me how can I select gpu device with FFmpeg CLI in this
> problem?
>

For all of the nvenc encoders there is a '-gpu' option: "Selects which
NVENC capable GPU to use. First GPU is 0, second is 1, and so on. (from -2
to INT_MAX) (default any)"

The cuvid decoders have similar '-gpu' options.


> *====================================================*
> *Below is test command*
> 1. The first gpu device(index is 0) is always working although input option
> is hwaccel_device 1.
>     $) ffmpeg -hwaccel_device 1 -y -i ./input.mp4 -filter_complex
> "[0:v]hwupload_cuda,scale_npp=w=960:h=540:format=yuv420p:
> interp_algo=lanczos,setsar=1[map0]"
> -map [map0] -c:v h264_nvenc -vprofile high -map 0:a -c:a copy  ./output.mp4
>
>
The above command doesn't use a hw-decoder (not a real problem if that's
what you want.) But I'm not sure if the '-hwaccel_device' option does
anything without something like '-hwaccel cuvid' and specifying '-c:v
h264_cuvid' before '-i ./input.mp4'.


> 2. If I remove "hwupload_cuda" options in filter_complex, hwaccel_device
> option works well.
>     $) ffmpeg -hwaccel_device 1 -hwaccel cuvid -y -c:v h264_cuvid -i
> ./input.mp4 -filter_complex
> "[0:v]scale_npp=w=960:h=540:format=yuv420p:interp_algo=
> lanczos,setsar=1[map0]"
> -map [map0] -c:v h264_nvenc -vprofile high -map 0:a -c:a copy ./output.mp4
>

This command is significantly different from the previous. It's more than
just removing "hwupload_cuda". This command uses hw-decoding, which is why
you don't have to "hwupload_cuda" because the frames are already in GPU
memory. Also, likely why '-hwaccel_device 1' works.

I see that you are using scale_npp to resize your video and I would be
amiss if I did not mention that the cuvid decoders have resizing and
cropping built in now. You don't need to use scale_npp to scale/resize, you
could do something as simple as:

ffmpeg -hwaccel_device 1 -hwaccel cuvid -c:v h264_cuvid -resize 960x540 -i
./input.mp4 -vf "setsar=sar=1" -c:v h264_nvenc -vprofile high -map 0:a -c:a
copy ./output.mp4

I hope that help!
-J