Hi Guys! Is there a plan of record for low-precision hardware support in Tensorflow? The obvious answer I guess is "Eigen" but eigen doesn't do anything below fp16.
Low precision fucntion is better supported in TF Lite (via gemmlowp), but there is also a new library called ruy (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/ruy) that might be the future of low pecision function in TF/TFLite ?
We're looking at hardware that may include some native low precision modes so I'm exploring for the appropriate places to add it. Do you guys know? Or is lowp something only the TFLite team cares about?
tensorflow/tools/pip_package/setup.py. my understanding is that all of the native code and SWIG stuff is handled by Bazel, and the only way to install the tensorflow package is via binary wheels, so why all this header file shuffling?