[Mlir-commits] [mlir] [MLIR][Linalg] Remove matmul_transpose variants (PR #147961)

[Renato Golin]

================
@@ -144,4 +144,126 @@ std::pair<int64_t, int64_t> getFmrFromWinogradConv2DFmr(WinogradConv2DFmr fmr);
 #define GET_OP_CLASSES
 #include "mlir/Dialect/Linalg/IR/LinalgRelayoutOps.h.inc"
 
+namespace mlir::linalg {
+
+/// Specialization of `linalg.matmul` op that has a transpose map on A
+class MatmulTransposeAOp : public MatmulOp {
----------------
rengolin wrote:

My answer to @banach-space above is that we may actually want to have those specializations for a bit longer, because they can actually be helpful, and maybe even never deprecate them. This could be a first implementation of the aliasing mechanism that I [described before](https://discourse.llvm.org/t/rfc-mlir-linalg-operation-tree/83586).

I'm ok either way, just want us to chose something with a bit more time to discuss on a more generic level than this specific PR.

https://github.com/llvm/llvm-project/pull/147961