[Mlir-commits] [mlir] [mlir][linalg] unfold projected permutation. (PR #114704)

[Andrzej Warzyński]

================
@@ -147,14 +150,19 @@ LogicalResult DecomposeProjectedPermutation::matchAndRewrite(
       op.isSingleYieldOp() || !op.isAllParallelLoops())
     return failure();
 
-  // All maps need to be projected permutations.
+  // If the map of an operand is not a `projected permutation` then
+  // it cannot be decomposed to mere transpose and broadcast.
+  // The requirement that all maps be `projected permutation` may be
+  // over-restrictive but since we need to determine shape of the
+  // iteration space as well, reject if any map violates assumption.
   for (auto &opOperand : op->getOpOperands()) {
     auto map = op.getMatchingIndexingMap(&opOperand);
     if (!map.isProjectedPermutation(false))
       return failure();
   }
 
-  // Currently we handle only static shapes.
+  // Decomposing linalg.generic involves creating `tensor.empty`
+  // which cannot have dnyamic shapes.
----------------
banach-space wrote:

`tensor.empty` with dynamic shapes _are allowed_, see e.g.: https://github.com/llvm/llvm-project/blob/58a17e1bbc54357385d0b89cfc5635e402c31ef6/mlir/test/Dialect/Tensor/ops.mlir#L49

What's a bit "tricky" is computing the required sizes. You could leave that as a TODO.

https://github.com/llvm/llvm-project/pull/114704