Phylanx Report: April 2019

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

• Performance improvements
• Keras Backend
• hpxMP improvements

Organizational Activities and Results

• Weekly group-meeting, minutes:
  • April 4th: https://drive.google.com/open?id=101MCAlV8m4lDlXDRRlf7kl7OGmpqx_rh
  • April 11th: https://drive.google.com/open?id=1kiN8L5OrJQlihZFOLmIX9HN-JNI0_A1c
  • April 18th: https://drive.google.com/open?id=1G3SQKRAAuNxlzlTyCCcIcHFrYUlsixk8
  • April 25th: https://drive.google.com/open?id=1itonnjZaaj1uDN_Y2TmasrTStNnBw1HI
• Phylanx Seminars:
  • http://phylanx.stellar-group.org/index.php/2019/05/01/phylanx-seminar-distributed-object-in-hpx/

Development Activities and Results

• General
  • Work on Keras backend
    • Implemented 83 Keras functions
  • Performance
    • Implemented flags to control thread stealing and idle back-off
    • Optimizations for local-only primitive calls
  • hpxMP
    • Replaced hpx::spin_lock with hpx::latch
    • Began implementing OpenMP 5.0 features
• Traveler
  • Refactoring the Atria tree visualization
    • Modular architecture to support our deployment goals
    • Completed a proof of concept of the architecture
    • Working to add prior features
  • Gantt chart
    • Switching to Golden Layout
      • Matches the aggregator that Alex has been working on
• Tiling
  • Implementation:
    • Implemented a generic distributed object in HPX
      • Two versions:
        • Initialization using all-to-all communication
        • Initialization using a meta-object
      • Supports construction on a subset of nodes/ localities.
  • Future Work:
    • Implement a distributed matrix using the  generic distributed object
    • Enable distributed object’s fetch to return a sub-block
    • Determine how to link a Python / PhySL object to a distributed object in HPX
• Primitives & Algorithms
  • We have merged the following primitives into master:
    • softsign (PR #923)
    • resize_images (PR #908)
    • switch_operation (PR #935)
    • ctc_decode (PR #938)
    • binary crossentropy (PR #945)
    • sort (PR #948)
    • conv1d (PR #955)
    • 3D support was added to Boolean primitives (PR #954)
  • Support for arrays to fold left / fold right was added (PR #932)
• Python
  • Enable implementation of plugins in Python:
    • implement ‘in_top_k’ plugin
  • Extend OpenSCoP back-end to support multi-dimensional arrays.
  • Study the feasibility of JIT compilation using LLVM compiler- primarily to enable polyhedral optimizations.
  • Implement more functions to support Keras’ Phylanx back-end.
  • Support code generation for the ‘zip’ operation

Repository Activity March 30th – April 30th:

Code statistics

-------------------------------------------------------------------------------
Language                     files          blank        comment           code
-------------------------------------------------------------------------------
C++                            435          17324           5500          87673
C/C++ Header                   284           6096           3527          30264
Python                         135           2517           1754           6447
CMake                          105            914            770           3811
YAML                             2             24             77            641
Dockerfile                       2              6             27             53
Markdown                         2             17              0             46
CSS                              1             14             10             38
HTML                             2              5              6             22
INI                              1              0              0              8
-------------------------------------------------------------------------------
SUM:                           969          26917          11671         129003
-------------------------------------------------------------------------------

Impact on Other Projects

• HPX Pull Requests:
  • Adding shortcut for hpx::get_ptr<>(sync, id) for a local, non-migratable objects (PR #3827)
  • Allow setting default scheduler mode via the configuration database (PR #3814)
  • Adding a distributed object abstraction (PR #3807)
  • Update hpxcxx to look at lib64 and use python3 (PR #3803)
  • Adding flags to scheduler allowing to control thread stealing and idle back-off (PR #3768)
• Blaze Tensor Pull Requests:
  • Adding norms for tensors (PR #21)
  • Adding version and GIT hash to build system (PR #20)
  • Fixing vectorized implementation of blaze::sum(Tensor) (PR #19)