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

[Renato Golin]

================
@@ -57,18 +58,31 @@ FailureOr<Operation *> mlir::linalg::transposeMatmul(RewriterBase &rewriter,
       dynamicDims);
   auto transposeOp = rewriter.create<linalg::TransposeOp>(
       loc, input, empty, ArrayRef<int64_t>{1, 0});
-  Operation *newMatmulOp;
+  Value newLHS, newRHS;
+  AffineMap mapLHS, mapRHS, mapOut;
+  AffineExpr d0, d1, d2;
+  auto context = rewriter.getContext();
+  bindDims(context, d0, d1, d2);
   if (transposeLHS) {
-    newMatmulOp = rewriter.create<linalg::MatmulTransposeAOp>(
-        loc, matmulOp.getResultTypes(),
-        ValueRange{transposeOp->getResult(0), matmulOp.getInputs()[1]},
-        matmulOp.getOutputs());
+    newLHS = transposeOp->getResult(0);
+    newRHS = matmulOp.getInputs()[1];
+    mapLHS = AffineMap::get(3, 0, {d2, d0}, context);
+    mapRHS = AffineMap::get(3, 0, {d2, d1}, context);
+    mapOut = AffineMap::get(3, 0, {d0, d1}, context);
   } else {
-    newMatmulOp = rewriter.create<linalg::MatmulTransposeBOp>(
-        loc, matmulOp.getResultTypes(),
-        ValueRange{matmulOp.getInputs()[0], transposeOp->getResult(0)},
-        matmulOp.getOutputs());
+    newLHS = matmulOp.getInputs()[0];
+    newRHS = transposeOp->getResult(0);
+    mapLHS = AffineMap::get(3, 0, {d0, d2}, context);
+    mapRHS = AffineMap::get(3, 0, {d1, d2}, context);
+    mapOut = AffineMap::get(3, 0, {d0, d1}, context);
----------------
rengolin wrote:

Ok, turns out the specializations make the changes in `TransposeMatmul.cpp` moot, so I reverted and it's not back to the original file. This means anyone using the builders out there will continue to have the same behaviour. 

So we're down to two problems: previous IR that needs regex-changing, and the pattern matchers. The IR changes are specific to each downstream, so better done in there as a one off. The pattern matchers will also be specific to usage and additional attributes, so can be on the downstreams until we figure a generic way to hold them all.

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