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

[Georgios Pinitas]

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

The preferred path in TfLite was essentially defined with the same argument but I can read many comments expecting matmul on the preferred path. The fact that conv2d has a bias addition in TOSA itself is really a design issue in my point of view but don't think that this is a discussion for this thread.

But I suppose expecting this to be a backend optimization is fine (as with many of these decompositions).

@Jerry-Ge could you please update the commit message with the fact that this pass is removed alongside with a justification? Thank you.

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