Skip to content

STEllAR-GROUP/blaze_tensor

Repository files navigation

3D Tensors for Blaze

This project implements 3D datastructures (tensors) that integrate well with the Blaze library.

All the highlights listed for Blaze apply to BlazeTensor as well:

  • high performance through the integration of BLAS libraries and manually tuned HPC math kernels
  • vectorization by SSE, SSE2, SSE3, SSSE3, SSE4, AVX, AVX2, AVX-512, FMA, and SVML
  • parallel execution by OpenMP, HPX, C++11 threads and Boost threads
  • the intuitive and easy to use API of a domain specific language
  • unified arithmetic with dense 3D tensors
  • thoroughly tested 3D tensor arithmetic
  • completely portable, high quality C++ source code

The implemented facilities are verified using a thorough testing environment. The CircleCI contiguous integration service tracks the current build status for the master branch: CircleCI.

What's implemented:

Datastructures

  • blaze::DynamicArray<N, T>: a resizable, row-major ND dense array data structure of arbitrary types
  • blaze::CustomArray<N, T, ...>: a non-owning ND dense array data structure usable to refer to some other ND dense array
  • blaze::DynamicTensor<T>: a resizable, row-major 3D dense array data structure of arbitrary types
  • blaze::CustomTensor<T, ...>: a non-owning 3D dense array data structure usable to refer to some other 3D dense array
  • blaze::StaticTensor<T, O, M, N>: a statically sized 3D dense array data structure of arbitrary types
  • blaze::UniformTensor<T>: a dynamically sized uniform (all elements have the same value) 3D dense array data structure of arbitrary types

Views

  • blaze::SubTensor<...>: a view representing a contigous subregion of a dense 3D array that is statically or dynamically sized
  • blaze::PageSlice<...>: a view representing a slice of 'thickness' one along the row/column plane of a 3D dense array
  • blaze::ColumnSlice<...>: a view representing a slice of 'thickness' one along the page/column plane of a 3D dense array
  • blaze::RowSlice<...>: a view representing a slice of 'thickness' one along the column/page plane of a 3D dense array
  • blaze::DilatedSubvector<...>: a view representing a sub-vector of any Blaze vector-like data structure with an additional 'dilation' (step-size between the elements of the underlying vector)
  • blaze::DilatedSubmatrix<...>: a view representing a sub-matrix of any Blaze matrix-like data structure with two additional argumnts: 'rowdilation' and 'columndilation' (step-size between the rows and the columns of the underlying matrix, respectively).
  • blaze::DilatedSubtensor<...>: a view representing a sub-tensor of any Blaze tensor-like data structure with three additional argumnts: 'pagedilation',
    'rowdilation' and 'columndilation' (step-size between the pages, rows and the columns of the underlying tensor, respectively).
  • blaze::QuatSlice<...>: a view representing a slice of 'thickness' one along the page/row/column tensor of a 4D dense array

Operations

  • All element-wise arithmetic operations that are supported by the Blaze library: element-wise addition, subtraction, division, Schur-multiplication, scalar multiplication, boolean comparison operations, and many mathematical operations like sqrt, cqrt, abs, sign, floor, etc.
  • Reduction operations like min and max, also unary and binary overloads for blaze::map.
  • Matrix expand operations that produce tensors from expanding (broadcasting) matrices in page direction (blaze::expand()).
  • Matrix and Tensor flattening (blaze::ravel()).

We have created a list of things that need to be implemented: TODO: Things to implement. This is a good starting point if you would like to help developing BlazeTensor.

Building and installing BlazeTensor

In order to use BlazeTensor you will need a proper installation of the Blaze library. Please see here for instructions on how to install it. Please note that you will need a very recent version (preferrably top of master) of Blaze. If you want to run the BlazeTensor tests you will also need the Blaze source directory.

  1. Clone this repository
  2. Create a build directory
    $ mkdir build
    $ cd build
    
  3. Call cmake with the relevant options:
    $ cmake -DCMAKE_INSTALL_PREFIX=/opt/BlazeTensor -Dblaze_DIR=<blaze_dir> <srcdir>
    
    here: <blaze_dir> should refer to the directory that contains the file blaze-config.cmake that was created during installation of Blaze.
  4. Build and install:
    $ make
    $ make install
    
  5. If you want to build the tests, additionally specify -DBLAZETENSOR_WITH_TESTS=ON and -Dblazetest_DIR=<blazesrc/blazetest> on the cmake command line. Run the tests with make tests.

BlazeTensor is a header only C++ library. Projects depending on it should make sure the headers are being found by the compiler. If your depending project uses cmake, just add find_package(BlazeTensor) to your scripts and refer to the target named BlazeTensor::BlazeTensor.

Communication

A channel on the freenode IRC network is used for discussions on BlazeTensor: #ste||ar on freenode (via SSL). Feel free to use the Github issue tracker for questions, bug reports, and feature requests.

License

BlazeTensor is released under the terms of the New (Revised) BSD license.

Acknowledgements

We would like to acknowledge the NSF, DoD, and the Center for Computation and Technology (CCT) at Louisiana State University (LSU).

BlazeTensor is currently funded by:

  • The National Science Foundation through awards 1737785 (Phylanx).

    Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

  • The Defense Technical Information Center (DTIC) under contract FA8075-14-D-0002/0007

    Neither the United States Government nor any agency thereof, nor any of their employees makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights.