Phylanx Report: March 2019

The nineteenth month of work focused on work in the following areas:

• Traveler Team Visit (March 7th and 8th)
• Support for the Keras Back-end

Organizational Activities and Results

• Weekly group-meeting, minutes:
  • March 7th: https://drive.google.com/open?id=1SOL-eIk3WOep3rVfFhVVIDHzgKC5d0ER
  • March 14th: https://drive.google.com/open?id=1ppwKttDjjGjoIAPlXSwp_KhpEybFYzXh
  • March 21st: https://drive.google.com/open?id=1HVuDeo0wBM8Rh5O8_mpatxwMKHdUbRiN
  • March 28th: https://drive.google.com/open?id=1abRtIbzIYdqBkPBtaG4EVkEEgE6XA_0V
• Phylanx Seminars:
  • http://phylanx.stellar-group.org/index.php/2019/03/20/phylanx-seminar-ordering/

Development Activities and Results

• General
  • Much of this month’s work focused on creating primitives which supports the Keras backend
    • Includes adding 3D support for some primitives
• APEX
  • Work on policy to determine whether to launch primitives synchronously or asynchronously
• Traveler
  • Traveler Tree now called “Atria”
  • Added list view to Atria. It shows an ordered list of high execution time primitives
  • Formal evaluation sessions of Atria
  • Atria UI fixes for code linking, tree collapsing
• Tiling
  • Theory:
    • Looked at the Latent Dirichlet allocation problem and found a suitable matrix-based learning algorithm
      • Set of operations performed by this algorithm consists of only simple matrix operations like multiplication, inverse etc. and they are suitable for distributed execution
      • As far as we are aware there are no distributed implementations of this particular algorithm, hence it is possible that a distributed implementation using Phylanx may lead to a small (student) paper
      • There are few variants of this algorithm, which claim better overall performance. It seems to us as a starting point this algorithm is a better choice due to its simplicity.
    • Partitioning interval hypergraphs
      • Motivated by sparse matrix tiling (partitioning) problem in the context of matrix vector multiplications
      • Problem is NP complete, if the intervals are contiguous and disjoint
      • Extending the intervals to two or more dimensions does not make this version of the problem any more interesting
  • Implementation:
    • Implemented two separate prototype distributed objects in HPX.
• Primitives & Algorithms
  • Added the following primitives:
    • ReLU #870
    • Sigmoid #872
    • Hard_Sigmoid #872
    • One_hot_operation #871
    • Softplus #869
    • Size #867
    • L2_normalize operation #881
    • elu_operation #882
    • batch_dot #877
  • 3D support added to the following primitives:
    • Reshape #823
    • Softmax #850
  • Slicing was adapted to support 32 bit indices #814
  • Added Truncated Normal Random Number Distribution #865
• Python
  • Added support for lazy evaluations of Phylanx functions
  • Support code generation for:
    • default values and named arguments
    • if expressions
    • list comprehension
    • for loops with variable iteration spaces
  • Bug fixes:
    • Plain returns
    • enforcing Python version in the build system

Repository Activity March 1st – March 31th:

Code statistics

-------------------------------------------------------------------------------
Language                     files          blank        comment           code
-------------------------------------------------------------------------------
C++                            417          16445           5149          80452
C/C++ Header                   272           5925           3462          29609
Python                         118           2272           1631           5838
CMake                          102            895            758           3725
YAML                             2             24             77            641
Dockerfile                       2              6             27             53
Markdown                         2             17              0             50
CSS                              1             14             10             38
HTML                             2              5              6             22
INI                              1              0              0              8
-------------------------------------------------------------------------------
SUM:                           919          25603          11120         120436
-------------------------------------------------------------------------------

Impact on Other Projects

• Blaze Tickets:
  • Add support for `mean` calculations (Issue #243)
• HPX Pull Requests:
  • Disable more code if networking is not available/not enabled (PR #3754)
  • Introduced cache_aligned_data and cache_line_data helper structure (PR #3745)
  • Add latch member function count_up and reset, requested by hpxMP (PR #3729)
  • Adding support for passing command line options through the HPX_COMMANDLINE_OPTIONS environment variable (PR #3726)
• Blaze Tensor
  • Fixing reduce operations to not operate on first rowslice twice (PR #18)
  • Changing the implementation of of the for loop in HPX backend (PR #12)
• hpxMP
  • Initial Support for OpenMP 5.0 (PR #50)
  • boost intrusive pointer for omp_task_data , all barriers replaced with latch (PR #49)
  • Using Latch replace Barrier (PR #48)