On Sunday, 12 October 2014 at 16:07:19 UTC, Laeeth Isharc wrote:
Any thoughts on speed in 2014 of pyd vs using cython to talk to D directly via C/C++ interface? I saw this old coment here:
pyd is basically just a convenience layer on top of the C interface. The part that would most likely screw with performance would be conversion from D objects to python objects and vice versa since that normally does a value copy. At the time of that comment, pyd's strategy for arrays was just iterate and convert each element. Now pyd can take advantage of the buffer protocol if it is exposed (2.7+), so converting e.g. numpy arrays should be just a memcopy.
If you want by ref semantics instead of by value, use PydObject instead of a D type in your function signature. You can also turn off all conveniences and just use the C api with raw_only (https://bitbucket.org/ariovistus/pyd/wiki/CeleriD)
