[PATCH] D79541: [mlir] Simplify and better document std.view semantics

[Tim Shen via Phabricator via llvm-commits]

timshen added inline comments.


================
Comment at: mlir/include/mlir/Dialect/StandardOps/IR/Ops.td:2910
   let description = [{
-    The "view" operation converts a 1-D memref with i8 element type,
-    to an N-D memref with arbitrary element type. In addition, the ViewOp
-    supports the following arguments:
-    *) A single dynamic offset operand can be specified which represents a
-       a dynamic offset within the base 1-D memref at which to create the
-       resulting memref view.
-    *) A dynamic size operand must be specified for each dynamic dimension
+    The "view" operation extracts an N-D contiguous memref (i.e. with empty or
+    identity layout map) with arbitrary element type from a 1-D contiguous
----------------
nicolasvasilache wrote:
> timshen wrote:
> > "contiguous memref" -> "slice"? To me "contiguous" means "no gaps between the accessed bytes", but it doesn't imply identity layout. Coming from Python/C++ background, "slice" implies identity layout.
> `slice` has potential conflicts in MLIR, there is a rank-reducing slice op in Linalg that is likely to be ported to std or to be folded within subview.
> 
> How about `"contiguous memref with empty or identity layout"` ?
Sure, but other occurrences of "contiguous" will have to be lengthened with "empty or identity layout" as well.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D79541/new/

https://reviews.llvm.org/D79541