[Mlir-commits] [mlir] [mlir][ArmSME] Switch to an attribute-based tile allocation scheme (PR #73253)

Benjamin Maxwell llvmlistbot at llvm.org
Fri Nov 24 02:24:57 PST 2023


================
@@ -21,6 +21,99 @@ include "mlir/Interfaces/SideEffectInterfaces.td"
 include "mlir/Dialect/LLVMIR/LLVMOpBase.td"
 include "mlir/Interfaces/InferTypeOpInterface.td"
 
+//===----------------------------------------------------------------------===//
+// ArmSME op interfaces
+//===----------------------------------------------------------------------===//
+
+def ArmSMETileType : I32EnumAttr<"ArmSMETileType", "Arm SME tile type",
+    [
+      I32EnumAttrCase<"ZAB", 0, "za.b">,
+      I32EnumAttrCase<"ZAH", 1, "za.h">,
+      I32EnumAttrCase<"ZAS", 2, "za.s">,
+      I32EnumAttrCase<"ZAD", 3, "za.d">,
+      I32EnumAttrCase<"ZAQ", 4, "za.q">,
+    ]>{
+  let cppNamespace = "mlir::arm_sme";
+  let genSpecializedAttr = 0;
+}
+
+def ArmSMETileOpInterface : OpInterface<"ArmSMETileOpInterface"> {
+  let description = [{
+    An interface for operations that use or allocate Arm SME tiles. These
+    operations need to be assigned a tile ID an i32 attribute, which specifies
+    which virtual tile within the ZA storage to use. The number of tiles
+    available depends on the type of the tile. This is summarized below:
+
+    | Tile Vector Types                                                       | Possible Tile IDs   |
+    |-------------------------------------------------------------------------|---------------------|
+    | `vector<[16]x[16]xi8>`                                                  | 0                   |
+    | `vector<[8]x[8]xi16>`, `vector<[8]x[8]xf16>`, or `vector<[8]x[8]xbf16>` | 0 and 1             |
+    | `vector<[4]x[4]xi32>` or `vector<[4]x[4]xf32>`                          | 0 to 3 (inclusive)  |
+    | `vector<[2]x[2]xi64>` or `vector<[2]x[2]xf64>`                          | 0 to 7 (inclusive)  |
+    | `vector<[1]x[1]xi128>`                                                  | 0 to 15 (inclusive) |
+
+    Operations that allocate a new tiles (such as arm_sme.get_tile), are used as
+    the roots for tile allocation, with all operations that (transitively)
+    depend on a root being assigned the same tile ID.
+  }];
+  let methods = [
+    InterfaceMethod<
+      "Sets the tile ID for this operation.",
+      /*returnType=*/"void",
+      /*methodName=*/"setTileId",
+      /*arguments=*/(ins "mlir::IntegerAttr":$tileId),
+      /*methodBody=*/[{}],
+      /*defaultImpl=*/ [{
+        if (!tileId)
+          return;
+        ::mlir::Operation* op = this->getOperation();
+        op->setAttr("tile_id", tileId);
+      }]
+    >,
+    InterfaceMethod<
+      "Returns the (possibly null) tile ID assigned to this operation.",
----------------
MacDue wrote:

Yep, just that no tile ID has been allocated yet. 

https://github.com/llvm/llvm-project/pull/73253


More information about the Mlir-commits mailing list