Proper 10bit videos support on GPU

[NicolasHug]

This issue is about enabling correct support for 10bit videos on GPU. Currently, we error out on such videos. And before #777 was merged, we would sometimes return incorrect values as previously reported in e.g. #585, #598, and #736.

To properly support 10-bits videos on GPU we need two main changes:

• Run the color-conversion step on uint16 temporary buffers, call the proper NPP functions with a custom color conversion matrix, and cast the uint16 into uint8. This is what I'm working on in Support 10 bits videos on CUDA #790.
• Sometimes, NVDEC cannot decode the video, so it actually relies on the CPU to do the decoding. For those videos, all that's left to do is the color-conversion step. We can either rely on our existing CPU implementation for that, or move the decoded frame to the GPU and run the color-conversion there. Support CPU fallback for videos that don't get decoded by nvdec #792 implements the former as it's easier, but we should try to do the second option eventually.

[shiftc]

for 10/12bit videos，maybe you can just let ffmpeg do the convertion：use avfilter's scale_cuda filter, with "format" set to AV_PIX_FMT_NV12, then the output cuda frame's sw_format will be NV12.

[shiftc]

for 10/12bit videos，maybe you can just let ffmpeg do the convertion：use avfilter's scale_cuda filter, with "format" set to AV_PIX_FMT_NV12, then the output cuda frame's sw_format will be NV12.

Since video formats varies, hiding the pixel_format conversion work within ffmpeg also has additional benefits.

[NicolasHug]

Thanks @shiftc , I wasn't aware of the existence of scale_cuda. Looks like there's also scale_npp!
I'll look into it

[shiftc]

Looks like there's also scale_npp!

scale_np doesn't support 10/12bit input, see sources here

[dvrogozh]

Thanks @shiftc , I wasn't aware of the existence of scale_cuda. Looks like there's also scale_npp!

@NicolasHug : I am working on enabling ffmpeg decoding+filters to better handle XPU cases (step out from low level VAAPI color conversion). Actually this pipeline proves to be very interesting because it's highly device agnostic and will allow to easily switch and support any ffmpeg backend, cuda, vaapi, vulkan, etc. Of course there is some vendor/hw api specific code but it's getting isolated into few things:

1. Function to create ffmpeg HW device context from the torch device (should return AVBufferRef)
2. Settings to select and tune HW backend specific filters (scale vs. scale_vaapi vs. scale_cuda)
3. Function to covert output AVFrame (from filters) to the tensor

I do have a working draft already and working on a cleanup. I will submit PRs on that by end of this week. We can also discuss offline if you want to know details.

[dvrogozh]

I will submit PRs on that by end of this week.

@NicolasHug, I've submitted to PRs:

• Move filter graph to stand alone class #831
  • I hope that's something we can consider to merge as it will give us flexibility to use ffmpeg filters for HW backends in multiple scenarios: to support more media filters beyond scale and color conversion, experiment with scale_cuda and scale_npp
• Full (dec+filters) ffmpeg-vaapi pipeline for Intel GPU (xpu pytorch backend) #832
  • That's a draft PR which demonstrates usage of Move filter graph to stand alone class #831 with ffmpeg-vaapi filters. I hope this will give a jump start to try scale_cuda and scale_npp as implementation will be very close as I noted in a previous comment.

[dvrogozh]

Both scale_cuda and scale_npp has some limitations. The one which is relevant for torchcodec is that they both don't currently support color conversion from yuv formats to rgb ones. scale_cuda does support 10-bit to 8-bit color conversion though which further can be paired with the current torchcodec step to convert nv12 to rgb24. Couple links for the reference:

• https://github.com/FFmpeg/FFmpeg/blob/n8.0/libavfilter/vf_scale_npp.c#L46
• https://github.com/FFmpeg/FFmpeg/blob/n8.0/libavfilter/vf_scale_cuda.cu (look for Convert_a_b kernels in this file)