週末、いつものように、TensorFlow のソースコードを眺めていたら、見つけました。
TensorFlow XLA : XRL
First checkin of XRT primitives to call XLA computations directly via引用します。 First checkin of XRT primitives to call XLA computations directly via the TensorFlow infrastructure. These primitives will allow direct access to XLA via Cloud TPU, and also provide an easy path to integrate custom XLA computations with other (distributed) TensorFlow computations on CPU and GPU. The API is *very experimental* and subject to change, perhaps substantially in the short term.
なんか、Cloud TPU 用なんですね。
XLA computation を直接たたくような感じ？
