Add amdxdna native HAL backend

Adds the amdxdna HAL driver with Linux KMQ and Windows MCDM native shims, PDIX and XADX executable schemas, async queue execution, transfer-queue support, native command recording/submission paths, HRX runtime integration, and amdxdna host CI coverage.

Includes unit and CTS coverage for allocator/buffer/device/event/semaphore/executable paths, command-buffer planning and caches, transfer queue behavior, XADX/PDIX artifact handling, and platform shim utilities.

Author: jtuyls

.github/workflows/ci_core_linux.yml

@@ -53,6 +53,16 @@ jobs:
             run_tests: true
             ctest_exclude_regex: ""
             package: true
+          - variant: AMDXDNA Host
+            sanitizer: none
+            assertions: false
+            run_tests: true
+            ctest_exclude_regex: ""
+            cmake_options: |
+              -DIREE_HAL_DRIVER_AMDGPU=OFF
+              -DIREE_HAL_DRIVER_AMDXDNA=ON
+              -DLIBHRX_BUILD_PASSTHROUGH=OFF
+            package: false
           - variant: ASAN
             sanitizer: asan
             assertions: true
@@ -99,6 +109,7 @@ jobs:
       ctest_label_exclude_regex: "runtime-resource=|manual"
       cmake_options: |
         -DIREE_HAL_AMDGPU_TARGETS=gfx942;gfx1151;gfx1201
+        ${{ matrix.cmake_options }}
       package: ${{ matrix.package }}
 
   gpu_linux:

CMakeLists.txt

@@ -70,14 +70,21 @@ option(IREE_ENABLE_ASSERTIONS "Force unset of NDEBUG compile option." OFF)
 option(IREE_DEV_MODE
   "Configure settings optimized for IREE development instead of CI/release."
   OFF)
+option(IREE_HAL_DRIVER_AMDXDNA
+  "Enables the amdxdna runtime HAL driver."
+  OFF)
 option(LIBHRX_BUILD
   "Build libhrx and HRX compatibility targets." ON)
 if(LIBHRX_BUILD)
-  if(DEFINED IREE_HAL_DRIVER_AMDGPU AND NOT IREE_HAL_DRIVER_AMDGPU)
+  if(NOT DEFINED IREE_HAL_DRIVER_AMDGPU)
+    if(NOT IREE_HAL_DRIVER_AMDXDNA)
+      set(IREE_HAL_DRIVER_AMDGPU_DEFAULT ON)
+    endif()
+  elseif(NOT IREE_HAL_DRIVER_AMDGPU AND NOT IREE_HAL_DRIVER_AMDXDNA)
     message(FATAL_ERROR
-      "LIBHRX_BUILD=ON requires IREE_HAL_DRIVER_AMDGPU=ON.")
+      "LIBHRX_BUILD=ON requires at least one accelerator driver "
+      "(IREE_HAL_DRIVER_AMDGPU=ON or IREE_HAL_DRIVER_AMDXDNA=ON).")
   endif()
-  set(IREE_HAL_DRIVER_AMDGPU_DEFAULT ON)
 endif()
 
 set(IREE_USE_LINKER "" CACHE STRING "")
@@ -185,6 +192,9 @@ endif()
 if(IREE_HAL_DRIVER_HIP)
   message(STATUS "  - hip")
 endif()
+if(IREE_HAL_DRIVER_AMDXDNA)
+  message(STATUS "  - amdxdna")
+endif()
 if(IREE_HAL_DRIVER_LOCAL_SYNC)
   message(STATUS "  - local-sync")
 endif()

README.md

@@ -185,6 +185,13 @@ ctest --test-dir build/hrx-tests/share/hrx-system/tests \
   --output-on-failure -L runtime-resource=amd-gpu
 ```
 
+NPU runners use the AMD NPU resource label:
+
+```bash
+ctest --test-dir build/hrx-tests/share/hrx-system/tests \
+  --output-on-failure -L runtime-resource=amd-npu
+```
+
 ## Release Artifacts
 
 Tagged releases publish one tarball per artifact, plus manifests and environment

libhrx/.bazel_to_cmake.cfg.py

@@ -100,6 +100,7 @@ def hrx_cts_test(self, name, deps=[], **kwargs):
         defines = kwargs.get("defines")
         includes = kwargs.get("includes")
         tags = kwargs.get("tags")
+        target_compatible_with = kwargs.get("target_compatible_with")
         full_deps = (
             [
                 ":core",
@@ -131,6 +132,7 @@ def hrx_cts_test(self, name, deps=[], **kwargs):
         if platform_deps_block:
             self._converter.body += platform_deps_block
         labels_block = self._convert_string_list_block("LABELS", tags)
+        self._emit_platform_guard_begin(target_compatible_with)
         self._converter.body += (
             f"hrx_cc_test(\n"
             f"{name_block}"
@@ -143,6 +145,7 @@ def hrx_cts_test(self, name, deps=[], **kwargs):
             f"{labels_block}"
             f")\n\n"
         )
+        self._emit_platform_guard_end(target_compatible_with)
 
 
 def convert_unmatched_target(converter, target):

libhrx/cts/BUILD.bazel

@@ -16,6 +16,11 @@ package(
 
 AMD_GPU_CTS_TAGS = ["runtime-resource=amd-gpu"]
 
+AMDGPU_TARGET_COMPATIBLE_WITH = select({
+    "//runtime/config/hal:driver_amdgpu": [],
+    "//conditions:default": ["@platforms//:incompatible"],
+})
+
 iree_amdgpu_binary_variants_embed_data(
     name = "amdgpu_executable_test_kernels",
     testonly = True,
@@ -26,6 +31,7 @@ iree_amdgpu_binary_variants_embed_data(
     identifier = "hrx_cts_amdgpu_executable_test_kernels",
     tags = ["manual"],
     target = "amdgcn-amd-amdhsa",
+    target_compatible_with = AMDGPU_TARGET_COMPATIBLE_WITH,
     target_selectors_flag = "//runtime/src/iree/hal/drivers/amdgpu:targets",
 )
 
@@ -112,6 +118,7 @@ hrx_cts_test(
     name = "executable",
     srcs = ["tests/executable/executable_test.cpp"],
     tags = AMD_GPU_CTS_TAGS,
+    target_compatible_with = AMDGPU_TARGET_COMPATIBLE_WITH,
     deps = [
         ":amdgpu_executable_test_kernels",
         "//build_tools/amdgpu:target_map_h",

libhrx/cts/CMakeLists.txt

@@ -9,6 +9,7 @@
 
 iree_add_all_subdirs()
 
+if(IREE_HAL_DRIVER_AMDGPU)
 iree_amdgpu_binary_variants_embed_data(
   NAME
     amdgpu_executable_test_kernels
@@ -29,6 +30,7 @@ iree_amdgpu_binary_variants_embed_data(
   TESTONLY
   PUBLIC
 )
+endif()
 
 iree_cc_library(
   NAME
@@ -195,6 +197,7 @@ hrx_cc_test(
     "runtime-resource=amd-gpu"
 )
 
+if(IREE_HAL_DRIVER_AMDGPU)
 hrx_cc_test(
   NAME
     hrx_cts_executable
@@ -211,6 +214,7 @@ hrx_cc_test(
   LABELS
     "runtime-resource=amd-gpu"
 )
+endif()
 
 hrx_cc_test(
   NAME

libhrx/src/libhrx/CMakeLists.txt

@@ -83,6 +83,16 @@ if(IREE_HAL_DRIVER_AMDGPU)
     HRX_HAS_IREE_AMDGPU_DRIVER
   )
 endif()
+if(IREE_HAL_DRIVER_AMDXDNA)
+  target_link_libraries(libhrx_src_libhrx_hrx PRIVATE
+    iree::hal::drivers::amdxdna::registration
+    iree::hal::drivers::amdxdna
+    iree::base::tooling::flags
+  )
+  target_compile_definitions(libhrx_src_libhrx_hrx.objects PRIVATE
+    HRX_HAS_IREE_AMDXDNA_DRIVER
+  )
+endif()
 # IREE's allocator selector is exposed as a compile definition on the base
 # library target. libhrx is compiled through IREE's object-library wrapper, so
 # forward that definition to the object target that actually compiles the