Good news! The change for including the Py_TPFLAGS_HAVE_VERSION_TAG for cython types by default is trivial and will likely make it into cython 0.19. See:
https://groups.google.com/group/cython-users/browse_thread/thread/5110f344ea2e6e86?hl=en To see what kind of effect we can expect in sage I checked 5.19b4, both vanilla and with the flag added (the easiest way is probably to include the flag in local/lib/python/ site-packages/Cython/Compiler/TypeSlots.py, line 380 and then doing 'sage -ba' to rebuild the library) and it's quite noticable: make ptest: Vanilla 5.9beta4 Total time for all tests: 1423.2 seconds cpu time: 1870.8 seconds cumulative wall time: 5026.0 seconds make ptest: 5.9beta4 with cython attribute caching Total time for all tests: 1308.8 seconds cpu time: 1794.7 seconds cumulative wall time: 4825.2 seconds Timing of make ptest can be tricky, so picking for instance devel/sage/ sage/combinat/sf/k_dual.py, which seems to quite heavily use python classes: vanilla: sage -t devel/sage/sage/combinat/sf/k_dual.py [247 tests, 7.56 s] ---------------------------------------------------------------------- All tests passed! ---------------------------------------------------------------------- Total time for all tests: 7.6 seconds cpu time: 7.2 seconds cumulative wall time: 7.6 seconds attribute caching: sage -t devel/sage/sage/combinat/sf/k_dual.py [247 tests, 6.71 s] ---------------------------------------------------------------------- All tests passed! ---------------------------------------------------------------------- Total time for all tests: 6.8 seconds cpu time: 6.3 seconds cumulative wall time: 6.7 seconds To illustrate, consider class P(sage.structure.parent.Parent): @cached_method def P(self): return 1 On vanilla, we get sage: p=P() sage: p.t=1 sage: p.P() 1 sage: p.t 1 sage: timeit('p.P()',number=200000,repeat=100) 200000 loops, best of 100: 91.8 ns per loop sage: timeit('p.t',number=200000,repeat=100) 200000 loops, best of 100: 98.9 ns per loop i.e., cached_method is a little faster, because the method P is found earlier in the MRO (instance attributes can be shadowed by data escriptors anywhere in the MRO, so the whole MRO needs to be walked for instance attributes; the worst case) With attribute caching this problem goes away, so everything is faster: sage: timeit('p.P()',number=200000,repeat=100) 200000 loops, best of 100: 80 ns per loop sage: timeit('p.t',number=200000,repeat=100) 200000 loops, best of 100: 50.3 ns per loop but since the absence of a class attribute t in the MRO gets cached as well, the lookup of the instance attribute is now by far the fastest. It's just a dict lookup, no computation or calling necessary (which still causes a little overhead for the cached method). We should upgrade to Cython 0.19 as soon as it's out. Or if someone feels like it, we can already patch the cython in sage. -- You received this message because you are subscribed to the Google Groups "sage-devel" group. To unsubscribe from this group and stop receiving emails from it, send an email to sage-devel+unsubscr...@googlegroups.com. To post to this group, send email to sage-devel@googlegroups.com. Visit this group at http://groups.google.com/group/sage-devel?hl=en. For more options, visit https://groups.google.com/groups/opt_out.
