Advanced general matrix multiplication (gemm)
First clone this repo by running:
git clone https://github.com/davidmallasen/gemm_adv.git
cd gemm_adv
Install googletest and googlemock for unit testing and build OpenBLAS and LIBXSMM to improve the naive implementation.
To build OpenBLAS in Ubuntu download the tar.gz from OpenBLAS, extract it and run make. Then run make PREFIX=/path/to/installation install.
To build LIBXSMM in Ubuntu download the tar.gz from GitHub, extract it and run make PREFIX=/path/to/installation STATIC=0 install.
Detailed steps can be seen in the Dockerfile.
Set the following environment variables:
export OPENBLAS_NUM_THREADS=1
export OPENBLAS_LIB=/path/to/openblas/lib
export OPENBLAS_INCLUDE=/path/to/openblas/include
export LIBXSMM_LIB=/path/to/libxsmm/lib
export LIBXSMM_INCLUDE=/path/to/libxsmm/include
export LD_LIBRARY_PATH=$OPENBLAS_LIB:$LIBXSMM_LIB:$LD_LIBRARY_PATH
Where OPENBLAS_LIB contains libopenblas.so, LIBXSMM_LIB contains libxsmm.so, OPENBLAS_INCLUDE contains cblas.h and LIBXSMM_INCLUDE contains libxsmm.h. We set the number of threads to 1 to have a better comparison with the naive implementation.
To build the system, in the root of the repo run:
./autogen.sh
./configure
make
Unit testing
To execute the tests for the naive implementation and the OpenBLAS replacement run make check.
To get the full output run ./tests/bin/gemm_test, ./tests/bin/openblas_gemm_test or ./tests/bin/libxsmm_gemm_test.
Performance comparisons
To measure the performance of the naive, OpenBLAS and LIBXSMM implementations you can run the measure_performance.sh script. For example:
./scripts/measure_performance.sh bin/naive data/naive_perf.dat
./scripts/measure_performance.sh bin/openblas data/openblas_perf.dat
./scripts/measure_performance.sh bin/libxsmm data/libxsmm_perf.dat
Performance data is given for an intel i7-9750H CPU running on one core. You can visualize the results in the images given or generate new plots using gnuplot with the .gp files.
Note: When using LIBXSMM, the matrices have to be transposed since the library expects column-major order. This could affect the comparison with the other implementations.
Deploying with docker
First create the distribution tarball by running make dist. Or download it from releases.
Then build and run the docker container:
sudo docker build -t gemm_adv-1:latest .
sudo docker run --rm -v `pwd`/data:/gemm_adv-1.0/data -it gemm_adv-1:latest
In the docker container you can run the same tests and performance comparisons as above.