README.md

simple-gemm

Collection of simple General Matrix Multiplication - GEMM implementations

C = a . A x B + C
if a = 1 and C = zeros
C = A x B

A and B are initialized with random numbers C is initialized with zeros

Arguments are always 3 matrix dimensions: args = [A_rows, A_cols (= B_rows), B_cols]

e.g. 5 5 5 or 10 10 10

CPU multithreading:

• GemmDenseThreads: native Julia Threads implementation

  $ cd GemmDenseThreads
  $ julia -t 4 gemm-dense-threads.jl 5 5 5    
  

• GemmDenseThreads.py: native Python Numba Threads implementation

  $ cd python/GemmDenseThreads
  $ NUMBA_NUM_THREADS=4 python3 GemmDenseThreads.py 5 5 5    
  

• GemmDenseBlas: uses LinearAlgebra.jl (super-fast), if compiled with OpenBLAS set OPENBLAS_NUM_THREADS

  $ cd GemmDenseThreads
  $ OPENBLAS_NUM_THREADS=4 julia gemm-dense-blas.jl 5 5 5    
  

GPU :

• GemmDenseCUDA : uses CUDA.jl which uses the optimized cuBLAS (very fast) on NVIDIA GPUs

  $ cd GemmDenseCUDA
  $ julia gemm-dense-cuda.jl 5 5 5
  

Citation

If you find the repository useful, please cite the reference 2023 IPDPSW paper:

@INPROCEEDINGS{10196600,
  author={Godoy, William F. and Valero-Lara, Pedro and Dettling, T. Elise and Trefftz, Christian and Jorquera, Ian and Sheehy, Thomas and Miller, Ross G. and Gonzalez-Tallada, Marc and Vetter, Jeffrey S. and Churavy, Valentin},
  booktitle={2023 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW)}, 
  title={Evaluating performance and portability of high-level programming models: Julia, Python/Numba, and Kokkos on exascale nodes}, 
  year={2023},
  volume={},
  number={},
  pages={373-382},
  doi={10.1109/IPDPSW59300.2023.00068}}