[Mlir-commits] [mlir] [mlir][tosa] Remove FullyConnectedOp from TOSA Dialect (PR #126152)

[Suraj Sudhir]

================
@@ -1,164 +0,0 @@
-//===- TosaDecomposeConv2D.cpp --------------------------------------------===//
----------------
sjarus wrote:

These were added by Arm TOSA backend compiler teams from what I recall; contributor name is in GitLens : @aardeb  

Whether conv2d 1x1 should remain as it is, or be decomposed to a matmul is a backend compilation and TOSA numerics stability driven question. The tosa.matmul has neither a bias nor an acc_type . 

The basis for this pass should be a foundation of validated functional testing using the TOSA reference model confirming that the numerics suggest the use of conv2d or matmul here. E.g. in TFLite, this conversation yielded the decision to use conv2d: https://github.com/tensorflow/tensorflow/pull/62310 . 

https://github.com/llvm/llvm-project/pull/126152