Repository navigation
Missing byteswap when converting array with record dtype to opposite-endian dtype. #6314
Description
Consider the following self-contained snippet of code:
import numpy as np
dtype = np.dtype([("V", "<f8"), ("A", "<f8")])
be_dtype = dtype.newbyteorder('>')
value = np.array((1.1, 1.2), dtype=dtype)
print np.asarray(value, dtype=be_dtype)
print np.asarray([value], dtype=be_dtype)
When I run this under Python 2.7 and NumPy 1.9.2 on my OS X 10.9 machine, I get the following output:
(1.1, 1.2)
[(-1.5423487136706484e-180, 4.667261458438315e-62)]
I'd expect to get this output:
(1.1, 1.2)
[(1.1, 1.2)]
Doing the same with a scalar dtype gives me the expected behaviour, and replacing value with the tuple (1.1, 1.2) also gives me the expected behaviour. It looks as though there's a byteswap operation being missed at some level during the conversion.
Activity
ahaldane commentedon Sep 14, 2015
This happens because
VOID_setitemcopies structured arrays as if they were void buffers, so no casting or byteswapping occurs. For example try this:This would be fixed by #6053
MAINT: improve struct assignment, allow multi fields views
MAINT: improve struct assignment, allow multi fields views