2020 07 06 Open NEST Developer Video Conference

previous | list | next

Agenda

1. Welcome
2. Review of NEST User Mailing List
3. Project team round
4. In-depth discussion
   • Hackathon wrap-up

Mailing list

• https://nest-simulator.org/mailinglist
• all threads answered

Project team round

Here we discuss topics that need broader attention, for example questions that came up but are outside a single project's scope, larger planned changes/PRs that affect all teams or pending work that is blocked by external factors.

1. Models / NESTML
   • no particular points to discuss, some progress during the hackathon
   • sometimes hard to find second reviewers, e.g. #1672 (Tsodyks)
2. PyNEST
   • nothing major here, handling remaining 2.20.1 issue
3. Kernel
   • #1544 (weight normalization) relates to models
4. Installation
   • #1072 decided to close
5. Infrastructure
   • nothing particular to report
   • travis currently without issues
   • factoring out https://github.com/nest/nest-extension-module also would need travis checking (#1662)
6. Documentation
   • many updates during the hackathon (Debian, Homebrew)
   • rendering issue fixed in kernel docs
   • many model doc updates
7. EBRAINS
   • no updates
8. (Feature) Random number generation
   • no updates
9. (Feature) Automated Testing
   • mymodule factoring out many in progress or in review
   • work in progress

In-depth discussion: Hackathon Wrap-Up

• many discussions and decisions
• significant progress in the work on issues and reviewed and merged pull-requests
• everyone is encouraged to actively comment on the open discussion issues especially discussions related to NEST 3.0

In-depth discussion: Models with exp() function

#795 discusses various faster/less acurate implementations for the exp() function

• neuron models that do many calls may become slow with the default implementation
• how to let the user choose?
• must be model specific
  • for some models, the calculation of exp() can be moved to pre-compile time.
  • check explicitly if this is optimized by compiler, may not be recognized
  • some models have 1-exp(~0) which requires very accurate evaluation (std. library has a dedicated expm1() function, which usually requires human thought and often can not automatically be optimized)
  • usually the argument of exp is not constant, but only evaluated on certain grid points, this can also benefit from pre-computation.
  • neuron model should come with its best integrator, during these thoughts it becomes
  • also implementations of exp() in FPGA available
    • usually no hardware support for exp(), but if available and sufficient accuracy, then it should be used.
  • use fastexp()? hint to solver?
  • implementation of exp() must be choose-able by the model implementor.
    • considering general configuraion for these choices when calling NESTML
• usually choice of implementation also entangled with choice of solver
• place of implementation within NEST should be in libnestutil
  • table-based exp() may need a different table for different situations (may still be pre-computed during model generation, or when user changes time constant)

Announcement: Virtual Hackathon

In order to find a date for the next hackathon, please note your interest and possible dates in the Doodle