/BBLAS

Reference implementation of the Batched BLAS routines (Group API).

Primary LanguageC

Batched Basic Linear Algebra Subroutines

University of Manchester (UK)

University of Tennessee (US)

Download BBLAS Software

About

BBLAS is the reference implementation of the Batched BLAS standard specification. A current trend in high-performance computing is to decompose a large linear algebra problem into batches containing thousands of smaller problems, which can be solved independently, before collating the results. To standardize the interface to these routines, the community developed an extension to the BLAS standard (the batched BLAS), enabling users to perform thousands of small BLAS operations in parallel whilst making efficient use of their hardware. Please visit BBLAS workshops for more information on the standardization efforts.

The main folders & files

  • compute: contains the standard BBLAS group API functions

  • core: contains the auxiliary batched BLAS functions which perform on groups of same size problems

  • control: contains auxiliary functions for type conversions

  • test: contains testing routines associated with the BBLAS functions and provides an insight on how the BBLAS functions should be called/used.

  • include: contains header files

  • make.inc: a configuration file to specify a C/C++ compiler, compilation flags and a BLAS library. The default configuration should work when MKL is installed.

  • Makefile: the Makefile, normally, it should not be modified.

Requirements

BLAS & LAPACK

OR

Doxygen for documentation

  • Doxygen can be install on Unix systems by sudo apt-get install doxygen or downloaded on the doxygen page.

Compilation

After the configuration of make.inc, the compilation is very simple:

  • make [all] -- compiles lib test
  • make lib -- generates library files in lib/libbblas.{a,so} lib/libcore.{a,so}
  • make test -- generates tester files in test/test
  • make docs -- generates documentation docs/html
  • make generate -- generates routines of other precisions
  • make clean -- removes objects, libraries, and executables
  • make cleangen -- removes generated precision files
  • make distclean -- removes above, Makefile.*.gen, and anything else that can be generated

For MacOS the libraries have to be linked manually. Use the following in the .profile file.

export DYLD_FALLBACK_LIBRARY_PATH=/path/to/BBLAS/lib:$DYLD_FALLBACK_LIBRARY_PATH

Testing

At the end of the compilation, testing routines (in four precisions [c,d,s,z]) will be available in the folder BBLAS/test. The main testing driver is the binary ./test. It should be used with the kernel to test as the firt parameter followed by the kernel arguments. For example ./test dgemm_batch will run the double precision of version of dgemm_batch with default arguments. For help ./test -h will display the list of kernel availble for testing, while ./test dgemm_batch -h will display all the possible more help and details on how to test dgemm_batch, this holds for all the kernels. We use random matrices for the test. To simplify the way to provide matrices in a batch and the size of each group, we provide the following arguments:

  • --ng : the number groups [default: --ng=10].
  • --gs : the number matrices in the first group [default: --gs=100].
  • --incg : the increment of group sizes. The size of the i^th group is gs + i * incg [default: --incg=10].
  • --dim : M x N x K dimensions of th matrices in the first group [default: --dim=1000 x 1000 x 1000].
  • --incm : The increment of matrix size across the group. If the matrix size in the first group is M x N x K, the matrix size in the i^th group will be (M + incm * i) x (N + incm * i) x (K + incm * i) [default: --incm = 1].
  • --info[a|g|n|o] : The parameter to set an error handling option [default: --info=a].
    • a: which indicates that all errors will be specified on output.
    • g: which indicates that only a single error will be reported for each group, independently.
    • n: which indicates that no errors will be reported on output.
    • o: which indicates that the occurrence of errors will be specified on output as a single integer value.

In addition to these options, arguments like trans, transa, transb, side, diag, uplo, etc., can be set, and for the sake of simplicity, they have the same value in all the groups.

  • --transa=[n|t|c] : transposition [default: --transa=n].
  • --transa=[n|t|c] : transposition of A [default: --transa=n].
  • --transb=[n|t|c] : transposition of B [default: --transb=n].
  • --side=[l|r] : left or right side application [default: --side=l].
  • --uplo=[g|u|l] : general rectangular or upper or lower triangular matrix [default: --uplo=l].
  • --diag=[n|u] : non-unit diagonal or unit diagonal [default: --diag=n].

Citing

Feel free to use the following publications to reference BBLAS:

  • Jack Dongarra, Sven Hammarling, Nicholas J. Higham, Samuel D. Relton, Mawussi Zounon: Optimized Batched Linear Algebra for Modern Architectures. Euro-Par 2017: 511-522.

  • Jack Dongarra, Sven Hammarling, Nicholas J. Higham, Samuel D. Relton, Pedro Valero-Lara, Mawussi Zounon: The Design and Performance of Batched BLAS on Modern High-Performance Computing Systems, ICCS 2017: 495-504

  • Jack Dongarra, Iain Duff, Mark Gates, Azzam Haidar, Sven Hammarling, Nicholas J. Higham, Jonathan Hogg, Pedro Valero Lara, Mawussi Zounon, Samuel D. Relton, and Stanimire Tomov, A Proposed API for Batched Basic Linear Algebra Subprograms, Draft Report, May 2016.

Funding

Primary funding for BBLAS was provided the European Union grant:

People

The following people listed in alphabetical order contributed to the BBLAS reference implementation:

  • Jack Dongarra
  • Mark Gates
  • Srikara Pranesh
  • Samuel Relton
  • Pedro Valero Lara
  • Mawussi Zounon