diff options
author | Peter Hawkins <phawkins@google.com> | 2018-09-28 09:17:51 -0700 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2018-09-28 09:21:44 -0700 |
commit | 97498f64ef097096b756c6b262f3ae38965e8685 (patch) | |
tree | 4688366a590da0876eef852feff9c484667f3bed /tensorflow/compiler | |
parent | 35459cbaa0f654393b242c5357f6939b05267ab8 (diff) |
[TF:XLA] Add comment explaining why there is no PrimitiveTypeToDataType function.
PiperOrigin-RevId: 214945748
Diffstat (limited to 'tensorflow/compiler')
-rw-r--r-- | tensorflow/compiler/tf2xla/type_util.h | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/tensorflow/compiler/tf2xla/type_util.h b/tensorflow/compiler/tf2xla/type_util.h index bda667eb1f..6354216eee 100644 --- a/tensorflow/compiler/tf2xla/type_util.h +++ b/tensorflow/compiler/tf2xla/type_util.h @@ -25,6 +25,14 @@ namespace tensorflow { // Converts a Tensorflow DataType to an XLA PrimitiveType. Status DataTypeToPrimitiveType(DataType data_type, xla::PrimitiveType* type); +// N.B.: there is intentionally no function to convert an XLA PrimitiveType to +// a TensorFlow DataType. The mapping from TF types to XLA types is not +// one-to-one: for example, both DT_INT8 and DT_QINT8 map to xla::S8. So the +// inverse would not be a well-defined function. If you find that you want the +// inverse mapping, then most likely you should be preserving the original +// TensorFlow type, rather than trying to convert an XLA type into a TensorFlow +// type. + } // namespace tensorflow #endif // TENSORFLOW_COMPILER_TF2XLA_TYPE_UTIL_H_ |