[Mlir-commits] [mlir] [mlir][AMDGPU] Add PermlaneOp (PR #154345)

[Tim Gymnich]

================
@@ -656,6 +656,60 @@ def AMDGPU_SwizzleBitModeOp : AMDGPU_Op<"swizzle_bitmode",
   }];
 }
 
+def AMDGPU_PermlanePerm : I32EnumAttr<"PermlanePerm",
+    "The possible permutations for a permlane operation",
+    [
+      I32EnumAttrCase<"swap_16",  0>,
+      I32EnumAttrCase<"swap_32",  1>,
+    ]> {
+  let genSpecializedAttr = 0;
+  let cppNamespace = "::mlir::amdgpu";
+}
+
+def AMDGPU_PermlanePermAttr : EnumAttr<AMDGPU_Dialect, AMDGPU_PermlanePerm,
+  "permlane_perm">;
+
+def AMDGPU_PermlaneOp : AMDGPU_Op<"permlane", [Pure, AllTypesMatch<["result", "src"]>]>,
+Arguments<(ins AnyIntegerOrFloatOr1DVector:$src, 
+               AMDGPU_PermlanePermAttr:$kind,
+               DefaultValuedAttr<BoolAttr, "false">:$fetch_inactive,
+               DefaultValuedAttr<BoolAttr, "false">:$bound_ctrl)> {
+  let summary = "AMDGPU permlane op";
+  let description = [{
+    High-level wrapper on `rocdl.permlane.*` variants.
+
+    Supports arbitrary int/float/vector types, which will be repacked to i32 and
+    one or more `rocdl.permlane.*` ops during lowering.
+    The following lane permutations are supported:
+    - Swap the data between odd and even rows of 16 lanes (`swap_16`)
+    - Swap the data between the first 32 lanes and the last 32 lanes (`swap_32`)
+
+    Format example:
+    ```
+    %0 = amdgpu.permlane %src swap_16 : f16
+    %1 = amdgpu.permlane %src swap_32 { fetch_inactive = true, bound_ctrl = true } : f16
+    ```
+
+    Operands:
+    * `$src`: Vector register to permute across lanes
+    * `$kind`: The kind of permutation operation.
+    * `$fetch_inactive`: Optional. Used to dertermine behavior of invalid lanes (disabled thread or out-of-range).
+      `fetch_inactive = false`: If source lane is invalid, use `bound_ctrl` to determine the source value.
+      `fetch_inactive = true`: If the source lane is disabled, fetch the source value anyway
+      (ignoring `bound_ctrl`). If the source lane is out-of-range, behavior is decided by `bound_ctrl`.
+    * `$bound_ctrl`: Optional. Used to determine what a thread should do if its source operand is from 
+      a disabled thread or invalid input: use the value zero, or disable the write.
+      `bound_ctrl = false`: Do not write when source is invalid or out-of-range.
+      `bound_ctrl = true`: Use zero as input if source is invalid or out-of-range
+
+    Note: Lowering is only supported on gfx950 and up.
+  }];
+  let results = (outs AnyIntegerOrFloatOr1DVector:$result);
----------------
tgymnich wrote:

done

https://github.com/llvm/llvm-project/pull/154345