tree b2b7207734841186df5913474fb1e8c144731afa
parent e2bb4b2acd601b0db3946e20e3ce72e8dc5dd42c
author Derek Murray <mrry@google.com> 1587494119 -0700
committer TensorFlower Gardener <gardener@tensorflow.org> 1587494281 -0700

[tf.unique()] Optimize the hash table implementation in `UniqueOp::Compute()`.

This change makes two improvements to the `UniqueOp` implementation:

1. Use `absl::flat_hash_map` instead of `std::unordered_map`.
2. For the `tstring` implementation, use `StringPiece` as the key instead of `tstring`, which avoids copying the strings into the map.

In addition, this change switches the microbenchmarks in unique_op_test.cc to use the SINGLE_THREADED_EXECUTOR, which removes thread scheduling overhead from the microbenchmark, and reduces noise in the results.

Microbenchmark results show a saving of between 0% and 65% on BM_Unique_INT32, between 8% and 26% on BM_Unique_INT32_Repeat, and between 17% and 40% on BM_Unique_STRING.

PiperOrigin-RevId: 307647292
Change-Id: If4367df37b856bf1c4cf91fcb34eea479014077f