diff options
| author |  A. Unique TensorFlower <gardener@tensorflow.org> | 2018-06-27 09:15:47 -0700 |
|---|---|---|
| committer | TensorFlower Gardener <gardener@tensorflow.org> | 2018-06-27 09:18:50 -0700 |
| commit | 3c121430e66d199e9c56b148482cb598115f8413 (patch) | |
| tree | d60b0a16214ab38d5062b852d8850923b3b7dc5a /third_party/llvm | |
| parent | 9f89437deb1e6ef68ab3ad226bb9ca3a0263a9ad (diff) | |
[TF2XLA] Don't wrap single-output computations in a tuple and don't resolve constants
Two small-but-not-micro-optimizations:
1) If a computation only produces one output, elide the tuple table by not wrapping it in a single-element tuple.
2) Don't resolve constants at compile time. If we resolve constants, the constant values are not available on the device and this forces us to materialize them, which involves a H2D transfer so is slow. Instead, get the device to materialize them because this is cheap even if they are not used.
PiperOrigin-RevId: 202319850
Diffstat (limited to 'third_party/llvm')
0 files changed, 0 insertions, 0 deletions