Just some comments as I looked at it anyway.
This comment is off. There is no enclosing linalg generic op. Maybe explain that it updates the part of the enclosing tiled_loop ops result specified by the second operand with the values from the first operand.
If the assumption is that it always is canonicalized away to an in-place update, then the terminator does not have any of these effects. However, I don't know this part well enough to judge whether it models the naive state or the expected bufferized behavior.
|956 ↗||(On Diff #358963)|
Comment is now off