According to Nvidia’s documentation, FFmpeg has several filters that can improve the utilization of the GPU, including hwupload_cuda and scale_npp/scale_cuda, but neither of these can be used with tvai_fi.
The former will throw the error
Impossible to convert between the formats supported by the filter 'Parsed_hwupload_cuda_0' and the filter 'auto_scale_0'
Error reinitializing filters!
Failed to inject frame into filter network: Function not implemented
Could you try piping the output to your other FFmpeg? Here’s a previous reply showing an example of piping to work around the lack of libx264 / 265 support, but the same principle applies:
This method should not be effective. Nvidia’s filter is designed to run operations on the GPU to reduce the overhead of data transfer. Pipe should unsupported it.
We cannot ship a version of ffmpeg with the scale_npp filter because it would require the CUDA SDK to be installed and distributed with the app.
Could I see the full command you are running to see if there are any configuration changes that need to be made for hwupload_cuda to work? Our build of ffmpeg should not have any issues with hardware accelerated decoding, it’s just the scaling filters that we cannot bundle.
There’s no reason to use hwupload_cuda alone, so it shouldn’t be a problem. Why can’t scale_npp/scale_cuda be supported? Dependencies such as cuda do not need to be built-in, only need to be installed before use.
scale_npp is only available in ffmpeg builds that were compiled with the NVIDIA Performance Primitives installed at the time of compilation, and scale_cuda is available with standard builds of ffmpeg as long as ffnvcodec is linked during compilation.
