scala-blas

Benchmarks of BLAS libraries with Scala interface

Description

This project contains benchmarks for matrix-matrix multiplication of various libraries that have Scala interface. Currently, they are

• Breeze, which can use any BLAS library with CBLAS interface by means of underlying Netlib-java
  • OpenBLAS, open source CPU optimized library
  • NVBLAS, proprietary Fortran-BLAS wrapper for few functions from GPU optimized library CUBLAS
  • Intel MKL, proprietary CPU optimized library
  • f2jblas, reference Java BLAS implementation
• BIDMat, based on Intel MKL and GPU CUDA
  • Own wrapper and optimizations for GPU CUBLAS
  • Own wrapper and optimizations for Intel MKL

Sources

Sources contain Scala code snippets that are used to run benchmarks

Results

Spreadsheets with results, each tab is a different hardware configuration: https://docs.google.com/spreadsheets/d/1lWdVSuSragOobb0A_oeouQgHUMx378T9J5r7kwKSPkY/edit#gid=0

Glossary

• BLAS - Basic Linear Algebra Subprograms, general term for linear algebra in software, might be misleading becuase it does not mean any specific interface
• Fortran BLAS - Reference Fortran interface for BLAS
• CBLAS - reference C interface for BLAS
• CUBLAS - proprietary (non-standard) interface for BLAS
• NVBLAS - proprietary reference Fortran-BLAS wrapper for few functions from GPU optimized library CUBLAS
• Netlib - repository of software for scientific computing

Relevant links

• Netlib, Reference BLAS and CBLAS http://www.netlib.org/blas/
• Netlib-java https://github.com/fommil/netlib-java
• Breeze https://github.com/scalanlp/breeze
• BIDMat https://github.com/BIDData/BIDMat/
• OpenBLAS https://github.com/xianyi/OpenBLAS
• CUDA http://www.nvidia.com/object/cuda_home_new.html
• NVBLAS http://docs.nvidia.com/cuda/nvblas