[Mlir-commits] [mlir] [mlir][gpu][spirv] Add patterns for gpu.shuffle up/down (PR #139105)

[Jakub Kuderski]

================
@@ -1332,7 +1332,8 @@ def GPU_ShuffleOp : GPU_Op<
     %3, %4 = gpu.shuffle down %0, %cst1, %width : f32
     ```
 
-    For lane `k`, returns the value from lane `(k + 1) % width`.
+    For lane `k`, returns the value from lane `(k + cst1)`. The resulting value
----------------
kuhar wrote:

Previously this would shuffle down a circular buffer (because of the modulo), now the op is more similar to a logical shift -- this seems like a bigger change then what the PR title suggests

https://github.com/llvm/llvm-project/pull/139105