[Mlir-commits] [mlir] [MLIR][NVVM] Fix undef in cp.async.bulk.tensor.reduce Op (PR #157423)

[Srinivasa Ravi]

================
@@ -1641,46 +1641,77 @@ CpAsyncBulkTensorSharedCTAToGlobalOp::getIntrinsicIDAndArgs(
   is_im2col ? CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, im2col)                \
             : CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, tile)
 
-#define GET_CP_ASYNC_BULK_TENSOR_ID(op, dims, is_im2col)                       \
-  [&]() -> auto {                                                              \
-    switch (dims) {                                                            \
-    case 1:                                                                    \
-      return CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 1, tile);                    \
-    case 2:                                                                    \
-      return CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 2, tile);                    \
-    case 3:                                                                    \
-      return CP_ASYNC_BULK_TENSOR_REDUCE(op, 3, is_im2col);                    \
-    case 4:                                                                    \
-      return CP_ASYNC_BULK_TENSOR_REDUCE(op, 4, is_im2col);                    \
-    case 5:                                                                    \
-      return CP_ASYNC_BULK_TENSOR_REDUCE(op, 5, is_im2col);                    \
-    default:                                                                   \
-      llvm_unreachable("Invalid TensorDim in CpAsyncBulkTensorReduceOp.");     \
-    }                                                                          \
-  }()
+#define GET_CP_ASYNC_BULK_TENSOR_ID(iid, op, dims, is_im2col)                  \
----------------
Wolfram70 wrote:

Discussed with @durga4github offline and removed the macros to use an intrinsic table instead. Please take a look, thanks!

https://github.com/llvm/llvm-project/pull/157423