Attempting to use scale_npp filter results in "impossible to convert between formats"

[rednaxela]

As best as I can tell, trying to use the scale_npp filter is always resulting in an error such as the following for me:

Impossible to convert between the formats supported by the filter 'graph 0 input from stream 0:0' and the filter 'auto-inserted scaler 0'
Error opening filters!

An example command line is

ffmpeg -i input_file.mkv -filter:v "scale_npp=w=1316:h=1080:format=same:interp_algo=lanczos" -codec:v libx264 output_file.mkv

This seems to happen regardless of what sort of input file I try. Initially was using a rgb24 color space input file, and thinking that might have been the problem tried a yuv420p input file, with the same result.

This is based upon a local build of the current git version, built under MSYS2 and which includes libnpp support.

A full log is attached.

Maybe this is user error rather than a bug? I can't see what I've done wrong though.

[rednaxela]

Full log up to when the error occurs.

[Andrey Turkin]

libavfilter is not too smart when it comes to HW acceleration. Try something like "hwupload_cuda,scale_npp=w=1316:h=1080:format=yuv420p:interp_algo=lanczos,hwdownload,format=yuv420p". Basically, you need to explicitly send frames to GPU and back, and you need to manually tell libavfilter the pixel format of the frames on their way back to main memory (which is a shame but it's how things are for now).

[rednaxela]

Ahhh, thanks for the clarification about that. With the addition of hwupload_cuda and hwdownload, things are working nicely.