Visillect/MetaImageMark

Microbenchmarking framework for comparison of image processing algorithm implementation

MetaImageMark

A framework to benchmark Image Processing Libraries.

Example of benchmark report.

Application areas

The framework can be used for various tasks, including:

1. Image processing code optimization.
2. Performance comparison of various image processing libraries.
3. Investigation of the dependence of the speed of operations depending on the compilation parameters and/or hardware.

MetaImageMark has been successfully used to identify weaknesses of MinImg image processing library and in its further optimization.

Running benchmarks

Building cmake project

To run benchmarks, you need to build a cmake project. The project depends on google/benchmark. There are two ways to get a dependency: allow cmake to download it itself (option -DDOWNLOAD_DEPENDENCIES=ON), or specify the path to the benchmark (-Dbencharm_DIR=/path/to/benchmarkConfig.cmake if benchmark is not installed in the system).

To run the benchmark of the selected library, you have to specify the path to this library.

OpenCV

If OpenCV is not installed in your system, specify

-DOpenCV_DIR=/path/to/OpenCVConfig.cmake

MinImg

To run MinImg benchmarks you have to specify path to MinImg install directory:

-Dminimg_DIR=/path/to/minimg/install

Running benchmarks

Before starting, read the section about noise suppression.

The executable files of the corresponding libraries will be located in the directory bin.[BuildType]. Run the desired benchmark using the command:

./bench_[library]_[operation] --benchmark_out=out.json.

Building a report

A python program is used to build the report. Usage example:

python3 -m report_builder build/minimg.json build/opencv.json --time -o transpose.html

Available benchmarks

Operation	Benchmark Name	MinImg	OpenCv
im = im1 + im2	BinarySum	+	+
im = im1 - im2	BinaryDiff	+	+
im = abs(im1 - im2)	BinaryADF	+	+
im = im1 * im2	BinaryMult	+	+
im = im1 ^ im2	BinaryPow	+	-

Operation	MinImg	OpenCv
BoxFilter	+	+
GaussFilter	+	+
Transpose	+	+
Downscale	+	+
LinTransform	+	+
Convert	+	+

Report builder

To analyze the benchmark data, a report builder is used. It allows you to visualize benchmarking data and conduct a comparative analysis of operations.

See usage examples here.

Noise suppression

Detailed information about the factors affecting noise is described here.

Processor shielding

$ sudo apt-get install cpuset
$ sudo cset shield --cpu=0,1 --kthread=on
$ sudo cset shield --exec -- mybench --params
$ sudo cset shield --reset

CPU frequency scaling

$ sudo cpupower frequency-set --governor performance
$ ./mybench
$ sudo cpupower frequency-set --governor powersave