Make fully_connected op use reference implementation in certain cases.

   - The optimized implementation will generate incorrect results when
     batch_size > 1 and the number of input elements is the square of
     batch_size.
   - This change makes fully_connected fallback to reference
     impelementation in this cases.
   - The bug only affects float32 path.

Bug: 80425683
Bug: 109805089
Test: mma
Test: NeuralNetworksTest_static --gtest_filter=*fully*
Merged-In: Ib1aa9fb47e536c64756b4ff3904def5bb4fe0f48
Change-Id: Ib1aa9fb47e536c64756b4ff3904def5bb4fe0f48
(cherry picked from commit c78d7cc2580720eca5c91595b87615a213c0a99d)
5 files changed
tree: dab27d8c5172542187a1997f62dbd15922df20ae
  1. nn/
  2. CleanSpec.mk
  3. OWNERS