[Mlir-commits] [mlir] [mlir][vector] Clean up VectorTransferOpInterface (PR #72353)

[Nicolas Vasilache]

================
@@ -46,246 +48,249 @@ def VectorUnrollOpInterface : OpInterface<"VectorUnrollOpInterface"> {
 
 def VectorTransferOpInterface : OpInterface<"VectorTransferOpInterface"> {
   let description = [{
-    Encodes properties of a transfer read or write operation.
+    Encodes properties of a `vector.transfer_read` or `vector.transfer_write`
+    operation. Vector transfer ops have:
+
+    - A shaped value that the op reads from/writes to: a memref or a tensor.
+    - A vector, either as a result or as an operand.
+    - Indicies that describe where the transfer from/to the shaped value starts.
+    - An optional mask.
+    - An optional in_bounds array to indicate transfer dimensions that are
+      guaranteed to be in-bounds.
+    - A permutation map to indicate transposes and broadcasts.
+
+    The "vector rank" is the rank of the vector type. E.g.:
+    ```
+    // Transfer with shaped value rank 2 and vector (transfer) rank 1.
+    %0 = vector.transfer_read %arg0[%c3, %c3], %f0
+        {permutation_map = affine_map<(d0, d1) -> (d0)>}
+        : memref<?x?xf32>, vector<128xf32>
+    ```
+
+    The "vector transfer rank" is the number of dimensions that participate in
+    the transfer and matches the number of results in the permutation map. In
+    most cases, the vector rank matches the vector transfer rank; the only
+    exception is when a vector is flattened as part of the transfer (see
----------------
nicolasvasilache wrote:

transfer_read can also broadcast when a `0` is specified: https://github.com/llvm/llvm-project/blob/main/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td#L1238
This is actually supported by some GPU hardware.
I wouldn't mind getting rid of the embedded broadcast, the feature has not paid for its complexity.

https://github.com/llvm/llvm-project/pull/72353