Accelerating VMAF with CUDA

This post is about how to compile VMAF with CUDA support and run the filter from FFMPEG. The entire compilation process has been tested on Ubuntu 25.15, without using Docker and with a minimum of pre-built libraries.

Install compilation tools and required libraries:

sudo apt install autoconf automake build-essential yasm cmake git-core unzip libtool wget meson ninja-build pkg-config libunistring-dev libnuma-dev libc6-dev libsdl2-dev zlib1g-dev

Install CUDA 13.1:

wget https://developer.download.nvidia.com/compute/cuda/13.1.1/local_installers/cuda_13.1.1_590.48.01_linux.run

chmod +x cuda_13.1.1_590.48.01_linux.run

sudo ./cuda_13.1.1_590.48.01_linux.run --toolkit

# add path to nvcc compiler
export PATH+=':/usr/local/cuda/bin/'

Install Nvidia Video SDK headers:

git clone https://git.videolan.org/git/ffmpeg/nv-codec-headers.git

cd nv-codec-headers

sudo make install

Download sources and configure VMAF:

wget https://github.com/Netflix/vmaf/archive/v2.3.1.tar.gz

tar xvf v2.3.1.tar.gz

mkdir -p vmaf-2.3.1/libvmaf/build

cd vmaf-2.3.1/libvmaf/build

meson setup ../libvmaf/ -Denable_tests=false -Denable_docs=false --buildtype=release --default-library=static

ninja -vC .

sudo ninja install

Download FFMPEG sources:

git clone https://git.ffmpeg.org/ffmpeg.git

cd ffmpeg

#894da5ca7d742e4429ffb2af534fcda0103ef593 - v8.0.1

git checkout 894da5ca7d742e4429ffb2af534fcda0103ef593

Configure and compile FFMPEG:

cd ffmpeg

./configure \
--extra-libs='-lpthread -lm' \
--ld=g++ \
--enable-gpl \
--enable-gnutls \
--enable-nvenc \
--enable-cuvid \
--enable-nvdec \
--enable-cuda-nvcc \
--enable-cuda \
--enable-zlib \
--enable-libvmaf \
--enable-nonfree \
--enable-ffplay \
--enable-filter=libvmaf_cuda \
--enable-filter=scale_cuda \
--disable-filter=scale_npp \
--pkg-config-flags=--static \
--enable-static \
--disable-shared \
--extra-cflags=-fPIC \
--extra-cflags=-I/usr/local/cuda/include/ \
--extra-ldflags=-L/usr/local/cuda/lib64/ \
--extra-ldflags=-L/usr/local/lib

make -j

sudo make install

---

Run VMAF calculation with scaling the original to 1920×1080:

ffmpeg -y -hwaccel cuda -hwaccel_output_format cuda -r 30 -i reference.mp4 -hwaccel cuda -hwaccel_output_format cuda -r 30 -i transcoded.mp4 -filter_complex "[0:v]scale_cuda=w=1920:h=1080:format=yuv420p[ref],[1:v]scale_npp=format=yuv420p[enc];[ref][enc]libvmaf_cuda=log_path=vmaf.log" -f null /dev/null

Or run calculation with the VMAF NEG model and using uncompressed original video:

ffmpeg -y -hwaccel cuda -hwaccel_output_format cuda -r 30 -i reference.y4m -hwaccel cuda -hwaccel_output_format cuda -r 30 -i transcoded.mp4 -filter_complex "[0:v]split=2[ref1][ref2];[1:v]split=2[enc1][enc2];[ref1]hwupload_cuda[ref_hw1];[ref2]hwupload_cuda[ref_hw2];[ref_hw1]scale_cuda=w=1920:h=1080:format=yuv420p[ref_sc1];[ref_hw2]scale_cuda=w=1920:h=1080:format=yuv420p[ref_sc2];[ref_sc1][enc1]libvmaf_cuda=log_path=vmaf.log[out1];[ref_sc2][enc2]libvmaf_cuda=log_path=vmafneg.log:model=path=/model/vmaf_v0.6.1neg.json[out2]" -map "[out1]" -map "[out2]" -f null /dev/null

Please let me know if you have any issue using these steps.