Setup

If you are on a compute cluster supporting Environment Modules, load the correct modules:

$ module load DEVELOP
$ module unload intel
$ module unload gcc
$ module load gcc/7
$ module unload python
$ module load python/3.6.0
$ module unload cuda
$ module load cuda/102

To be able to compile the sources google benchmark is needed. Visit Google Benchmark and follow the installation instructions. Make sure to compile Google Benchmark with the same compiler that is used to compile these benchmark-framework sources (recommended: gcc 7).

Besides of google benchmark, you need to have the following installed for building the sources:

• CUDA toolkit (recommended version: 10.2)
• cmake (should already be present for building Google Benchmark)

In order to build the sources follow these steps:

$ git clone git@git.rwth-aachen.de:elias.geiger/benchmarking-framework.git
$ cd benchmarking-framework

Open CMakeLists.txt and point it to the google benchmark installation:

#TODO point BENCHMARK to location of google benchmark
set(BENCHMARK "/home/elias/Programs/benchmark")

Build the project:

$ mkdir build && cd build
$ cmake ../
$ make

Usage of the tuning capability

Make sure to install kernel_tuner first. In order to start the tuning utility, follow these steps:

$ cd kernels/kernel_tuner
$ ./matmul_batch_local.sh

Vice versa on the cluster:

$ cd kernels/kernel_tuner
$ sbatch matmul_batch.sh

The resulting statistics are placed into /kernels/optimized

Usage of the benchmarking capability

In order to start the benchmarking utility, follow these steps:

$ cd build
$ ./benchmark

Vice versa on the cluster:

$ cd benchmarks
$ sbatch gbench_batch.sh

The resulting statistics are placed into /benchmarks