On Tue, Dec 20, 2011 at 9:10 PM, Mark Wiebe <mwwi...@gmail.com> wrote: > On Tue, Dec 20, 2011 at 6:24 PM, Geoffrey Irving <irv...@naml.us> wrote: >> >> Hello, >> >> As a followup to the prior thread on bugs in user defined types in >> numpy, I converted my rational number class from C++ to C and switched >> to 32 bits to remove the need for unportable 128 bit numbers. It >> should be usable as a fairly thorough test case for user defined types >> now. It does rather more than a minimal test case would need to do, >> but that isn't a problem unless you're concerned about code size. Let >> me know if any further changes are needed before it's suitable for >> inclusion in numpy as a test case. The repository is here: >> >> https://github.com/girving/rational >> >> The tests run under either py.test or nose. >> >> For completeness, my branch fixing all but one of the bugs I found in >> numpy user defined types is here: >> >> https://github.com/girving/numpy/tree/fixuserloops >> >> The remaining bug is that numpy incorrectly releases the GIL during >> casts even though NPY_NEEDS_API is set. The resulting crash goes away >> if the line defining ACQUIRE_GIL is uncommented. With the necessary >> locks in place, all my tests pass with my branch of numpy. I haven't >> tracked this one down and fixed it yet, but it shouldn't be hard to do >> so. > > > Looks great. I've added some comments to the pull request for the > fixuserloops branch, which is here: > > https://github.com/numpy/numpy/pull/175 > > I would advise anyone with an interest in the low-level aspects of how > NumPy's handling of the GIL and multi-threading/concurrency should evolve to > take a look. Prior to anything I contributed, NumPy hardcoded whether to > release the GIL during ufuncs or not. I added a needs_api flag in a few
So releasing the GIL wasn't something the user could get at? (I'm curious if this is something that should be mentioned in the ufunc tutorial on the numpy docs.) > places to indicate whether the inner loop functions call the CPython API or > not. Note that for ABI compatibility reasons, this flag is not 100% > correctly integrated throughout NumPy. Could you expand a little bit on this ABI compatibility issue? > > What Geoffrey is proposing here conflicts with the way I imagined the flag > would be used, but supporting both of our ways of calling the inner loop > seems useful to me. Take a look at the pull request for more details. > > Cheers, > Mark > >> >> >> Geoffrey >> _______________________________________________ >> NumPy-Discussion mailing list >> NumPy-Discussion@scipy.org >> http://mail.scipy.org/mailman/listinfo/numpy-discussion > > > > _______________________________________________ > NumPy-Discussion mailing list > NumPy-Discussion@scipy.org > http://mail.scipy.org/mailman/listinfo/numpy-discussion > _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
