Add a small int64 workaround for elementwise ops that don't support it

microsoft / tensorflow-directml-plugin

DirectML PluggableDevice plugin for TensorFlow 2

Apache License 2.0

185 stars 25 forks source link

Add a small int64 workaround for elementwise ops that don't support it #265

Closed PatriceVignola closed 2 years ago

PatriceVignola commented 2 years ago

It turns out that DIVIDE, MODULUS_FLOOR, DIFFERENCE_SQUARE, MODULUS_TRUNCATE and POW don't yet support emulated int64 in DirectML, so we can't entirely get rid of int64 workarounds.