This benchmarking suite allows a direct comparison of popular linear algebra frameworks in C++. The libraries are easy-to-install using the conda package manager:
conda env create -f environment.yml
The environment.yml installs the following libraries:
- xtensor with xsimd
- Eigen3
- Armadillo
- Blitz++
After setting up the environment, it is advised to create a build
directory, and execute cmake
:
mkdir build
cd build
cmake .. -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX -DBENCHMARK_ALL=ON
You should see a message for each found library, similar to the following:
COMPILING WITH
======================================
Found eigen : /home/myuser/miniconda3/envs/bench/include/eigen3
Found Blitz : /home/myuser/miniconda3/envs/bench/include | /home/myuser/miniconda3/envs/bench/lib/libblitz.so
Found Armadillo : /home/myuser/miniconda3/envs/bench/include | /home/myuser/miniconda3/envs/bench/lib/libarmadillo.so
Found xtensor : /home/myuser/miniconda3/envs/bench/include
Found xsimd : /home/myuser/miniconda3/envs/bench/include
This allows you to make sure you're compiling with the correct, up-to-date versions of the libraries.
To build and run the benchmarks, just use the following command:
make xbenchmark
If you are only interested in specific benchmarks, build with make xtensor_benchmark
and then run manually ./xtensor_benchmark --benchmark_filter=my_benchmark
. The backend to the benchmarks is the popular google-benchmark suite, so look there for more documentation.