Skip to content

Latest commit

 

History

History
61 lines (37 loc) · 1.93 KB

README.md

File metadata and controls

61 lines (37 loc) · 1.93 KB
Raw

This repository contains benchmarks for taco and scripts to reproduce the results from the paper The Tensor Algebra Compiler.

Build taco-bench

Tell taco-bench where taco is located:

export TACO_INCLUDE_DIR=<taco-directory>/include
export TACO_LIBRARY_DIR=<taco-build-directory>/lib

Build taco-bench with CMake 2.8.3 or greater:

cd <taco-bench-directory>
mkdir build
cd build
cmake ..
make

Installing and building with other products

Do the following steps before you build taco-bench with cmake to benchmark against several libraries.

EIGEN

  1. Download and extract Eigen's source code from https://eigen.tuxfamily.org/dox/GettingStarted.html.
  2. Specify the variable EIGEN_DIR.

GMM++

  1. Use the installation guide http://getfem.org/gmm/install.html.
  2. Specify the variable GMM_INCLUDE_DIR.

UBLAS

  1. Follow the documentation to install ublas: http://www.boost.org/doc/libs/1_45_0/libs/numeric/ublas/doc/index.htm.
  2. Specify the variable UBLAS_INCLUDE_DIR.

OSKI

  1. Follow the user guide to install and tune OSKI: http://bebop.cs.berkeley.edu/oski/downloads.html.
  2. Specify OSKI_INCLUDE_DIR and OSKI_LIBRARY_DIR.

POSKI

  1. Follow the user guide to install and tune pOSKI: http://bebop.cs.berkeley.edu/poski/.
  2. Specify POSKI_INCLUDE_DIR and POSKI_LIBRARY_DIR.

Note: Install first OSKI and then use this installation to install POSKI.

MKL

  1. Commercial licensed product from Intel: https://software.intel.com/en-us/mkl.
  2. Specify MKL_ROOT.

Note: use the mklvars.sh script of Intel to set properly your environment.

Benchmarking taco against your product

  1. Modify CMakeList.txt: add an environment variable YOUR, compilation options, YOUR_INCLUDE, and YOUR_LIBRARY directories to your project.

  2. Implement the expression using your project. Modify your4taco.h file.

  3. Use the TACO_BENCH macro to benchmark and validate method to compare against expected results.