Matthew Johnson 42a1ad4307 change dtype promotion behavior for jit-invariance
Here are two desiderata for jax.numpy dtype promotion behavior:
1. follow what NumPy does
2. be invariant to `@jit`

The latter is much more important, so whenever the two are in tension we
prefer the latter. (Also we already can't do a perfect job following
what NumPy does, e.g. around its value-dependent dtype promotion logic.)

Issue #732 showed our code had a special behavior that essentially
handled a case of the former desideratum but also broke the latter. #732
also showed us (again) that our tests really should cover Python
scalars.

In summary, in this commit:
* revise jax.numpy dtype promotion behavior to be invariant to `@jit`
* add Python scalar types to lax_numpy tests
* simplify and update kron implementation to fix dtype issues
2019-05-19 18:49:16 -07:00
..
2019-05-13 20:36:45 -07:00
2019-05-08 13:41:25 -07:00
2019-04-12 10:48:11 -04:00
2019-05-13 08:48:13 -07:00
2019-05-13 19:53:50 -04:00
2018-11-21 20:22:51 -08:00
2018-11-21 20:22:54 -08:00
2019-05-16 10:36:30 -07:00
2019-05-17 13:37:43 -07:00