romansavrulin/ffmpeg

NVIDIA accelerated ffmpeg

Features

• NVENCODE acceleration
• NVDECODE acceleration
• video codec: x264
• video codec: x265
• audio codec: AAC

NVENCODE (nvenc) and NVDECODE (formerly CUVID) are packaged in the NVIDIA Video Codec SDK.

Hardware Accelerated Encoders:

List options of an encoder using ffmpeg -h encoder=XXXX

• h264_nvenc, nvenc, nvenc_h264
• nvenc_hevc, hevc_nvenc

Hardware Accelerated Decoders:

List options of a decoder using ffmpeg -h decoder=XXXX

• h264_cuvid
• hevc_cuvid
• mjpeg_cuvid
• mpeg1_cuvid
• mpeg2_cuvid
• mpeg4_cuvid
• vc1_cuvid
• vp8_cuvid
• vp9_cuvid

Hardware Accelerated Filters:

List options of a filter using ffmpeg -h filter=XXXX

• hwupload_cuda
• scale_cuda
• scale_npp
• thumnail_cuda

Usage

Run the container mounting the current directory to /workspace processing input.mp4 to output.mp4 without any hardware acceleration

$ docker run --rm -it --runtime=nvidia \
    --volume $PWD:/workspace \
    willprice/nvidia-ffmpeg -i input.mp4 output.avi

$ docker run --rm -it --runtime=nvidia \
    --volume $PWD:/workspace \
    willprice/nvidia-ffmpeg \
      -hwaccel_device 0 \
      -hwaccel cuvid \
      -c:v h264_cuvid \
      -i input.mp4 \
      -c:v hevc_nvenc
      out.mkv

Get a shell prompt inside the container, useful for debugging:

$ docker run --rm -it --runtime=nvidia \
    --volume $PWD:/workspace \
    --entrypoint bash
    willprice/nvidia-ffmpeg

Build

The docker image is a multistage build. The initial stage, the build stage, builds a statically linked ffmpeg binary that is then copied over into the runtime image. By statically linking we minimize the number of external dependencies and shrink the runtime image.

Resources

• FFmpeg hardware acceleration guide with examples
• Static FFmpeg build on Ubuntu 16.04 guide
• Using FFmpeg with GNU parallel
• Listing NVENC and NPP capabilities of FFmpeg
• Encoding HEVC using FFmpeg with NVENC
• FFmpeg cheatsheet
• FFmpeg-static build scripts