[Mlir-commits] [mlir] [MLIR][XeGPU] Extend op definitions to support 3D+: dpas, dpas_mx (PR #199809)

[Sang Ik Lee]

================
@@ -915,9 +915,21 @@ def XeGPU_DpasOp : XeGPU_Op<"dpas", [Pure, AllElementTypesMatch<["lhs", "rhs"]>,
 
   let description = [{DPAS performs matrix multiplication on matrix A of `mxk`
     size, B of `kxn` size, and accumulate on matrix C of `mxn` to the same size
-    matrix , `m=8`, `n=16` and `k=8 * 32/bit_width_of_elem_type`. So for fp16
-    data type, the matrices are `A: vector<8x16xf16>`, `B: vector<16x16xf16>`,
-    and `C/D: vector<8x16xf32>`.
+    matrix.
+
+    The operands can be 2D, 3D, or 4D vectors. When the vectors have more than 2
----------------
silee2 wrote:

Do the batch dimensions (if they are used) need to be the same across A, B and C?
What if B is shared across batches, then does it need to be broadcasted first before passing as an operand? 

https://github.com/llvm/llvm-project/pull/199809