ENH: cache dtype.__hash__

Computing the type of a dhash can be slow for complex (e.g. structured)
dtypes.  Hashing dtypes can be useful in some applications, such as
when doing type-based dispatching, and speed can be critical in those
cases.  This enhancement caches the once-computed hash value in the
dtype structure, so as to save time on further lookups.  The cached
value is invalidated in the rare cases where the dtype is mutated.

Benchmarks numbers:

python3.4 -m timeit -s "import numpy as np; t=np.dtype('uint64')" "hash(t)"
* before patch: 1000000 loops, best of 3: 0.498 usec per loop
* after patch: 10000000 loops, best of 3: 0.0616 usec per loop

python3.4 -m timeit -s "import numpy as np; t=np.dtype([(s, 'f') for s in 'abcdefghij'])" "hash(t)"
* before patch: 100000 loops, best of 3: 4.43 usec per loop
* after patch: 10000000 loops, best of 3: 0.0603 usec per loop

Closes #5339.

1 files changed, 3 insertions, 0 deletions

diff --git a/doc/release/1.10.0-notes.rst b/doc/release/1.10.0-notes.rst
index d5cd99203..a7c0e2852 100644
--- a/doc/release/1.10.0-notes.rst
+++ b/doc/release/1.10.0-notes.rst
@@ -61,6 +61,9 @@ C API
 The changes to *swapaxes* also apply to the *PyArray_SwapAxes* C function,
 which now returns a view in all cases.
 
+The dtype structure (PyArray_Descr) has a new member at the end to cache
+its hash value.  This shouldn't affect any well-written applications.
+
 recarray field return types
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Previously the returned types for recarray fields accessed by attribute and by