danilhendrasr/video-decoding-benchmark

Compare NVIDIA Video Codec SDK's, PyAV's, and OpenCV's performance on video decoding.

Video Decoding Benchmark

A program to benchmark the performance of NVIDIA NVDEC (through VideoProcessingFramework), PyAV, and OpenCV (without hardware acceleration) on video decoding operation. The following metrics are observed:

1. Frame Processing Time (the time it takes to decode 1 frame)
2. CPU Utilization Across All Cores
3. Memory Utilization
4. GPU Utilization
5. GPU Memory Utilization

Prerequisites

• Machine with supported NVIDIA GPU
• NVIDIA Driver
• Docker
• Git LFS

Running the Benchmark

1. Make sure you got all the prerequisites on your system (see section above)
2. Clone the repo
3. Download NVIDIA Video Codec SDK and place the zip file in the root directory of this project
4. Run git lfs pull to download the sample videos from GitHub's Git Large File Storage
5. Build the docker image

   docker build -t videc-benchmark .

   2 build args are available:
   • PROJECT_PATH, to control where in the image you want this project to be copied to, this arg defaults to /videc-benchmark.
   • VIDEO_CODEC_SDK_VERSION, to specify which version of NVIDIA Video Codec SDK that you've downloaded from step 3, this arg defaults to 11.1.5.
6. Run the docker image:

   docker run --gpus all -it --pid host videc-benchmark bash

7. Enter the following command inside the container's terminal:

   scripts/run-benchmark.sh -i ./videos/45-seconds.mp4 -w $x

   Notes:
   • You should replace $x with an integer. The program will use the first $x frames as warmup during the benchmark, meaning it won't collect any data during the processing of the first $x frames.
   • The ./videos/45-seconds.mp4 part is the path to input file. You can look into the videos directory to see files that are available to be used as input.
8. Copy the benchmark result from the container to the host:

   docker cp <container ID>:/videc-benchmark/benchmark-results ./<directory name>

Existing Report

I've done the benchmark previously and I've compiled a report document. Click here if you need to read it.