GPU Matrix Transpose

This repository implements and benchmarks different matrix transpose algorithms for GPU's using CUDA. Definitely check out the corresponding blog post

Repository structure

The data folder contains the original benchmark data from the tested architectures that was used in the experimental analyses.

The lib folder contains C-functions used by all tested algorithms with the corresponding header file.

The src folder contains C files with the different algorithms for matrix transposition.

Setup project

After cloning the repository you can run

make

and the files in src are compiled, and the benchmark test is started and stored in the stats folder (that is created by the Makefile). Waring: It can take a lot of time for the benchmarks to finish!

If you only want to compile the files in src create a folder bin and run

make compile_c

This should compile all C files in src and store them into the bin folder without starting the benchmarks. Compiled binaries follow the naming convention of <ALGORITHM>-<OPTIMIZATION LEVEL>

In order to run the experiments on the Marzola cluster run

make marzola

The script compiles the source code and launches SLURM jobs on the cluster. Results are stored in the stats/ folder.

Validate implementations

The correctness of the provided implementations can be verified by running the compiled binaries in 'debug mode'. After compilation you can run

./bin/<BINARY> <MATRIX SIZE> --debug

For example

./bin/naive-0 2 --debug

Should output a randomly initialized matrix with dimension 2^2 and the corresponding transposed matrix. Additionaly the execution time and the effective bandwidth are displayed.