diff options
author | Vijay Vasudevan <vrv@google.com> | 2016-09-06 17:30:55 -0800 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2016-09-06 18:47:10 -0700 |
commit | 91ce95d497ec2957535b2ce6a965cd8269d723e5 (patch) | |
tree | a97799f5b2b92cc676e7890aa0f206c25f705490 /WORKSPACE | |
parent | 970c408a0c34726d5361b8b7e90fe02376f7e022 (diff) |
Change shape ops that produce int32s to also output int64s in the type
system, with the default remaining int32 for backwards compatibility,
and change some ops that consume only int32s to also take in int64.
Adds changes to the python op_def_library to properly handle the addition
of new types with unchanged defaults.
This changes the specification, not the implementations. The implementations
will come later with additional registrations. Note that in some cases
we are not yet ready for registering the int64 types, because there is
already an int64 type registered that assumes data is in device memory.
We'll need to fix the problem of choosing kernels based on memorytype
in a follow up CL before being able to register multiple kernels for
the same type but with different memorytypes.
Change: 132386744
Diffstat (limited to 'WORKSPACE')
0 files changed, 0 insertions, 0 deletions