Nvidia GeForce products that are Kepler or Maxwell based are limited to two simultaneous video streams per system using NVENC hardware. Quadro products (Kepler or Maxwell) can encode more than 2 video streams.
How to around this limit and run more than 2 FFmpeg transcodings avoid “out of memory” error on Nvidia GeForce cards?
Requirements:
- Ubuntu 14.04
- Nvenc-compatible gpu
- Nvidia 375.39 driver
Install necessary software:
sudo apt-get install build-essential git yasm unzip wget sysstat nasm
Before proceeding installation NVIDIA drivers on Ubuntu you must disable Nouveau kernel driver. To disable the default Nouveau Nvidia driver follow guide: How to Disable Nouveau Nvidia Driver on Ubuntu
We need exact Nvidia driver version 375.39
Download driver from https://yadi.sk/d/yahf1Y-D3PJnzd, then install:
chmod +x ./NVIDIA-Linux-x86_64-375.39.run ./NVIDIA-Linux-x86_64-375.39.run
check driver driver installation :
nvidia-smi
Now we need patch GPU driver with patch from: https://github.com/keylase/nvidia-patch
mkdir ~/nvenc_backup cd ~/nvenc_backup cp /usr/lib/x86_64-linux-gnu/libnvidia-encode.so.375.39 ~/nvenc_backup/ wget https://raw.githubusercontent.com/keylase/nvidia-patch/master/patch.sh chmod +x patch.sh ./patch.sh ~/nvenc_backup/libnvidia-encode.so.375.39 /usr/lib/x86_64-linux-gnu/libnvidia-encode.so.375.39 reboot
Install NVENC SDK
Get the NVENC SDK from the NVIDIA developer’s site.
wget http://developer.download.nvidia.com/compute/nvenc/v5.0/nvenc_5.0.1_sdk.zip unzip nvenc_5.0.1_sdk.zip
Copy the NVENC headers to /usr/local/include to make it easier later
$ sudo cp nvenc_5.0.1_sdk/Samples/common/inc/*.h /usr/local/include
Download and install a light-weight library to communicate with the CUDA display driver
wget http://developer.download.nvidia.com/compute/redist/ffmpeg/1511-patch/cudautils.zip unzip cudautils.zip cd cudautils make
Download and install all the open source libraries
1. Get x264.
$ git clone git://git.videolan.org/x264.git $ cd x264
2. Configure x264
$ ./configure --disable-cli --enable-static --enable-shared --enable-strip
3. Build x264
$ make -j 10
4. Install x264
$ sudo make install $ sudo ldconfig
Get FFmpeg $ git clone git://source.ffmpeg.org/ffmpeg.git
Download the NVIDIA acceleration.
$ wget http://developer.download.nvidia.com/compute/redist/ffmpeg/1511-patch/ffmpeg_NVIDIA_gpu_acceleration.patch $ cd ffmpeg
Apply the NVIDIA acceleration patch. Note that this patch was created against the
git master commit:
commit b83c849e8797fbb972ebd7f2919e0f085061f37f
$ git reset --hard b83c849e8797fbb972ebd7f2919e0f085061f37f $ git apply ../ffmpeg_NVIDIA_gpu_acceleration.patch
Configure FFmpeg with NVENC, NVRESIZE and x264 support.
$ ./configure --enable-nonfree --enable-nvenc --enable-nvresize --extra-cflags=-I../cudautils --extra-ldflags=-L../cudautils --enable-gpl --enable-libx264
Build FFmpeg.
$ make -j 10
Install FFmpeg.
$ sudo make install $ sudo ldconfig
For testing FFmpeg transcode performance with nvenc and libx264 you can run:
time ffmpeg -y -i <mp4 input file> -vcodec nvenc -b:v 5M -acodec copy OUTPUT.mp4 time ffmpeg -y -i <INPUT.mp4> -vcodec libx264 -b:v 5M -acodec copy OUTPUT.mp4
