Try to reduce size of GEBP kernel for non-ARM targets.

Reference issue

What does this implement/fix?

This is an experiment to try and prevent MSVC from running out of heap memory when building TensorFlow.

Additional information

Merge request reports

Loading