Using rocm-xio for Copy Engine

.gitmodules

@@ -1,3 +1,3 @@
-[submodule "ext/shader_sdma"]
-	path = ext/shader_sdma
-	url = https://github.com/AARInternal/shader_sdma.git
+[submodule "ext/rocm-xio"]
+	path = ext/rocm-xio
+	url = https://github.com/ROCm/rocm-xio.git

examples/06_message_passing/message_passing_host_initiated.py

@@ -8,7 +8,7 @@
 the consumer (GPU 1) remains a device kernel.
 
 Key difference from message_passing_put.py:
-- Producer: Host uses anvil to initiate SDMA transfers from Python
+- Producer: Host uses sdma_ep (rocm-xio) to initiate SDMA transfers from Python
 - Consumer: Same device kernel waiting for data
 
 This shows how to orchestrate GPU-to-GPU transfers from Python without

iris/device/__init__.py

@@ -0,0 +1,13 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2025-2026 Advanced Micro Devices, Inc. All rights reserved.
+
+"""
+Device-side utilities for Iris.
+
+This module provides low-level device-side functions for use in Triton kernels,
+including SDMA queue management and packet construction utilities.
+"""
+
+from . import sdma_utils
+
+__all__ = ["sdma_utils"]