FFmpeg GPU Transcoding & VMAF Benchmark

This project allows testing different FFmpeg configurations using GPU acceleration with an NVIDIA Quadro P400, and evaluating the visual quality of the generated videos with VMAF.

The goal is to compare multiple encoding settings on the same source video (containing diverse scenes: motion, static, dark, etc.), and produce a CSV report with both performance and quality metrics.

---

🎯 Goals

• Test different encoding configurations with FFmpeg and NVENC/NVDEC

• Measure the impact of each setting on:

  • output file size
  • encoding speed
  • perceived quality (VMAF and VMAF NEG)

• Generate a consolidated CSV report

---

📦 Project Contents

• Dockerfile

  • Builds FFmpeg 8.0 with NVENC/NVDEC, libx264/libx265, and libvmaf support
  • Installs VMAF v3.0.0 and NVENC headers
  • Based on CUDA, adapted for the NVIDIA Quadro P400

• Python script (ffmpeg_vmaf_benchmark.py)

  • Reads a CSV file describing FFmpeg encoding settings
  • Runs GPU-accelerated transcodes
  • Automatically computes VMAF and VMAF NEG
  • Produces a CSV report (file size, ratio, encoding time, quality metrics)

• Examples

  • Input CSV with encoding parameters
  • Output CSV with performance and quality results

---

🛠️ Requirements

• NVIDIA GPU with NVENC support (tested with Quadro P400)
• Docker and NVIDIA Container Toolkit installed
• A source video with varied scenes (motion, static, different lighting)

---

📦 Build the Docker image

docker build -t ffmpeg-vmaf-nvidia .

---

▶️ Usage

1. Prepare a configuration CSV

Example (encoding_settings.csv):

quality;flags
22;-c:v hevc_nvenc -cq:v 22
23;-c:v hevc_nvenc -cq:v 23
24;-c:v hevc_nvenc -cq:v 24
25;-c:v hevc_nvenc -cq:v 25

Each line defines a configuration to test:

• quality: identifier/quality parameter
• flags: FFmpeg options

---

2. Run the container

docker run --gpus all -it --rm \
    -v $(pwd):/workspace \
    ffmpeg-vmaf-nvidia bash

---

3. Launch the tests

Inside the container:

cd /workspace
python3 ffmpeg_vmaf_benchmark.py \
    --input_csv encoding_settings.csv \
    --source_video input.mp4 \
    --output_dir results_videos \
    --output_csv results.csv \
    --max_transcode_workers 1 \
    --max_vmaf_workers 2

---

📊 Example Results

Example of generated CSV (results.csv):

index,quality,flags,size,ratio_size,walltime,usertime,vmaf_min,vmaf_max,vmaf_mean,vmaf_harmonic_mean,vmaf_neg_min,vmaf_neg_max,vmaf_neg_mean,vmaf_neg_harmonic_mean
0,22,"-c:v hevc_nvenc -cq:v 22","13706.75 KB",0.926,00:00:02:943:430,00:00:00:000:424,95.48,100.0,97.88,97.87,91.57,100.0,97.11,97.10
1,23,"-c:v hevc_nvenc -cq:v 23","11826.97 KB",0.799,00:00:02:929:708,00:00:00:001:107,95.35,100.0,97.86,97.85,91.70,100.0,97.01,97.00

Key columns:

• size: generated file size
• ratio_size: output size / source size ratio
• walltime: actual transcoding time
• usertime: CPU time consumed
• vmaf / vmaf_neg: min, max, mean, harmonic mean scores

---

⚙️ Script Options

Option	Description	Default
--input_csv	CSV file with FFmpeg configs	required
--source_video	Source video path	required
--output_dir	Directory for encoded videos	out
--output_csv	CSV results file	output_results.csv
--max_transcode_workers	Max parallel transcodes	1
--max_vmaf_workers	Max parallel VMAF calculations	2
--sequential	Sequential execution	false
--vmaf_threads	Number of CPU threads for VMAF	1

---

📝 Notes

• The Dockerfile includes CUDA and dependencies for the Quadro P400.

• The project can be used with any NVIDIA GPU supporting NVENC.

• VMAF is computed twice:

  • Standard VMAF
  • VMAF NEG (stricter model).

---

🔗 Inspired by

This project was inspired by the article: Benchmarking FFMPEG's H.265 Options