Author: fijal Branch: extradoc Changeset: r5617:afc5d24e8d8a Date: 2016-03-24 16:59 +0200 http://bitbucket.org/pypy/extradoc/changeset/afc5d24e8d8a/
Log: work on the last para diff --git a/talk/scipyUS2016/proposal.rst b/talk/scipyUS2016/proposal.rst --- a/talk/scipyUS2016/proposal.rst +++ b/talk/scipyUS2016/proposal.rst @@ -29,8 +29,15 @@ We have been working for a number of years on an alternative implementation of the ndarray, tightly integrated to the PyPy machinary. Reimplementing numpy completely comes with costs, like a constant need to -replicate updates when a new version of numpy is released. -We've been recently experimenting with reusing -more of C numpy through the C API and we have ideas how to merge both -approaches to have the best of both worlds. +replicate updates when a new version of numpy is released and difficulty to +integrate with the rest of the numeric stack. We've been recently experimenting +with a quite different approach -- to have a good enough implementation of CPython +C API to be able to just use numpy, which is in the experimental stage. +That comes with a problem of speed of array access which now has to traverse +the slow C API. We propose a solution that would merge the two, where we +would be able to both use the current numpy as well as reimplement parts +that will be performance critical. +We hope that this approach will fill a gap of using pypy in the numeric +world, where jit-level performance can provide middle ground between +user-input heavy solutions like cython and the slowness of CPython. _______________________________________________ pypy-commit mailing list [email protected] https://mail.python.org/mailman/listinfo/pypy-commit
