Skip to content

dev call 20220707

Jump to bottom
Florian Angerer edited this page Jul 7, 2022 · 2 revisions

Dev Call: 7 July 2022

Present

Matti, Simon, Ronan, Du Toit, Gianluca, Antonio, Florian, Tim, Mohaned

Agenda

Conferences

  • Antonio and Simon are going to SciPy 2022
  • Ronan is going to attend EuroPython

NumPy Migration Progress

  • GraalVM team's branches are now online
  • It should be easy to try it out. Just checkout the appropriate branch in HPy and install it. Then checkout NumPy branch and build/install it.
  • Tim already pushed the game-of-life example and some description on how to run it in numpy-hpy's readme.
  • Current NumPy/HPy changes are barely in a shape that can be merged.
  • How to approach upstreaming of changes: try to push heap-type conversion first
  • Ronan did not upstream his changes because PyPy couldn't support that types are changed after creation. NumPy patches buffer procs or something after the type was initialized.
  • Did the GraalVM team see some fundamental performance blockers? * That's hard to answer by now. The GraalVM team has seen performance regressions due to generic APIs. For example, in CPyhon you can usee PyTuple_GetItem which assumes a tuple as receiver and directly accesses the underlying storage. In HPy, you need to use HPy_GetItem_i which is generic. * Florian also expects some penalty because HPy does not _allow_ borrowed and stolen references. I.e. there will be more inc/dec-refs because of that. However, that can probably be fixed by restructuring code and using trackers.

Misc

  • SciPy 2022 HPy BoF session topics: * NumPy devs will also be there we could advertise our migration efforts
  • Pillow progress (Du Toit): started to port Pillow; seems to be 20% slower. Need to investigate why. HPy args parser cannot handle tuple unpacking (intentionally); the idea is to parse the tuple as 'O' and unpack manually
  • HPyContext layout * Florian will create benchmarks to see if additional indirection is expensive * We could use NumPy array printing because that's stack intensive
  • HPyContext backwards compatibility (PR-331) * Matti confirms that NumPy uses a similar approach * Antonio suggests that we could also do this on a file-granularity (e.g. have public_api.0.h and public_api.1.h and the added context members are added at the end of the context). * In general: adding a preprocessor to autogen for public_api.h would also be nice for the debug mode. We could use pcpp (a pure Python impl) for that.
  • Antonio came up with the idea to produce real universal binaries by compiling to web assembly.

Home

Dev Call Minutes

Next Level Manifesto

HPy Sprints

Misc Notes

Clone this wiki locally