[Mlir-commits] [mlir] [mlir][vector] Refine vectorisation of tensor.extract (PR #109580)

[Andrzej Warzyński]

================
@@ -1012,7 +1019,10 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
   bool foundIndexOp = false;
   bool isContiguousLoad = isContiguousLoadIdx(linalgOp, extractOpTrailingIdx,
                                               foundIndexOp, resType);
-  isContiguousLoad &= foundIndexOp;
+  // TODO: Support generating contiguous loads for column vectors - that will
+  // require adding a permutation map to tranfer_read Ops.
+  bool isRowVector = resType.getShape().back() != 1;
+  isContiguousLoad &= (foundIndexOp && isRowVector);
----------------
banach-space wrote:

See for _which_ indices we call `isLoopInvariantIdx`: https://github.com/llvm/llvm-project/blob/1c47fa9b620d0abb280647b4f361ada43784d00e/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp#L974-L991 

In particular, look at this line:
```cpp
  // The trailing index is not analysed here
  auto leadIndices = indices.drop_back(1);
```

The trailing index is analysed separately further down: https://github.com/llvm/llvm-project/blob/1c47fa9b620d0abb280647b4f361ada43784d00e/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp#L1013

https://github.com/llvm/llvm-project/pull/109580