Support building ARM64EC onnxruntime.dll (#6999)

cmake/onnxruntime_common.cmake

@@ -63,6 +63,8 @@ else()
 endif()
 if(onnxruntime_target_platform STREQUAL "ARM64")
     set(onnxruntime_target_platform "ARM64")
+elseif(onnxruntime_target_platform STREQUAL "ARM64EC")
+    set(onnxruntime_target_platform "ARM64EC")
 elseif(onnxruntime_target_platform STREQUAL "ARM" OR CMAKE_GENERATOR MATCHES "ARM")
     set(onnxruntime_target_platform "ARM")
 elseif(onnxruntime_target_platform STREQUAL "x64" OR onnxruntime_target_platform STREQUAL "x86_64" OR onnxruntime_target_platform STREQUAL "AMD64" OR CMAKE_GENERATOR MATCHES "Win64")
@@ -71,6 +73,15 @@ elseif(onnxruntime_target_platform STREQUAL "Win32" OR onnxruntime_target_platfo
     set(onnxruntime_target_platform "x86")
 endif()
 
+if(onnxruntime_target_platform STREQUAL "ARM64EC")
+    if (MSVC)
+        link_directories("$ENV{VCINSTALLDIR}/Tools/MSVC/$ENV{VCToolsVersion}/lib/ARM64EC")
+        link_directories("$ENV{VCINSTALLDIR}/Tools/MSVC/$ENV{VCToolsVersion}/ATLMFC/lib/ARM64EC")
+        link_libraries(softintrin.lib)
+        add_compile_options("/bigobj")
+    endif()
+endif()
+
 file(GLOB onnxruntime_common_src CONFIGURE_DEPENDS
     ${onnxruntime_common_src_patterns}
     )

cmake/onnxruntime_mlas.cmake

@@ -54,7 +54,7 @@ if(MSVC)
       )
       list(APPEND mlas_platform_srcs ${obj_filename})
     endforeach()
-  elseif(onnxruntime_target_platform STREQUAL "ARM")
+  elseif((onnxruntime_target_platform STREQUAL "ARM") OR (onnxruntime_target_platform STREQUAL "ARM64EC"))
     set(mlas_platform_srcs
       ${ONNXRUNTIME_ROOT}/core/mlas/lib/arm/sgemmc.cpp
     )

onnxruntime/core/mlas/inc/mlas.h

@@ -35,7 +35,7 @@ Abstract:
 // Define the target architecture.
 //
 
-#if defined(_M_AMD64) || defined(__x86_64__)
+#if (defined(_M_AMD64) && !defined(_M_ARM64EC)) || defined(__x86_64__)
 #define MLAS_TARGET_AMD64
 #endif
 #if defined(_M_IX86) || defined(__i386__)
@@ -47,7 +47,7 @@ Abstract:
 #if defined(_M_ARM64) || defined(__aarch64__)
 #define MLAS_TARGET_ARM64
 #endif
-#if defined(_M_ARM) || defined(__arm__)
+#if defined(_M_ARM) || defined(_M_ARM64EC) || defined(__arm__)
 #define MLAS_TARGET_ARM
 #endif
 #if defined(__VSX__)

onnxruntime/core/providers/cpu/tensor/cast_op.cc

@@ -22,7 +22,7 @@
 #include "Eigen/src/Core/arch/Default/BFloat16.h"
 #include "Eigen/src/Core/arch/Default/Half.h"
 
-#if defined(_M_AMD64)
+#if defined(_M_AMD64) && !defined(_M_ARM64EC)
 #include "core/mlas/inc/mlas.h"
 #endif
 
@@ -207,7 +207,7 @@ struct TensorCaster<std::string, DstType> {
   }
 };
 
-#if defined(_M_AMD64)
+#if defined(_M_AMD64) && !defined(_M_ARM64EC)
 // specializations to use optimized and Windows x64-specific
 // MlasConvertHalfToFloatBuffer() routine for MLFloat16 -> float conversion
 

tools/ci_build/build.py

@@ -268,6 +268,10 @@ def parse_arguments():
         "--arm64", action='store_true',
         help="Create ARM64 makefiles. Requires --update and no existing cache "
         "CMake setup. Delete CMakeCache.txt if needed")
+    parser.add_argument(
+        "--arm64ec", action='store_true',
+        help="Create ARM64EC makefiles. Requires --update and no existing cache "
+        "CMake setup. Delete CMakeCache.txt if needed")
     parser.add_argument(
         "--msvc_toolset", help="MSVC toolset to use. e.g. 14.11")
     parser.add_argument("--android", action='store_true', help='Build for Android')
@@ -679,7 +683,7 @@ def generate_build_tree(cmake_path, source_dir, build_dir, cuda_home, cudnn_home
         "-Donnxruntime_MIGRAPHX_HOME=" + (migraphx_home if args.use_migraphx else ""),
         # By default - we currently support only cross compiling for ARM/ARM64
         # (no native compilation supported through this script).
-        "-Donnxruntime_CROSS_COMPILING=" + ("ON" if args.arm64 or args.arm else "OFF"),
+        "-Donnxruntime_CROSS_COMPILING=" + ("ON" if args.arm64 or args.arm64ec or args.arm else "OFF"),
         "-Donnxruntime_DISABLE_CONTRIB_OPS=" + ("ON" if args.disable_contrib_ops else "OFF"),
         "-Donnxruntime_DISABLE_ML_OPS=" + ("ON" if args.disable_ml_ops else "OFF"),
         "-Donnxruntime_DISABLE_RTTI=" + ("ON" if args.disable_rtti else "OFF"),
@@ -1626,7 +1630,7 @@ def is_cross_compiling_on_apple(args):
 
 
 def build_protoc_for_host(cmake_path, source_dir, build_dir, args):
-    if (args.arm or args.arm64 or args.enable_windows_store) and \
+    if (args.arm or args.arm64 or args.arm64ec or args.enable_windows_store) and \
             not (is_windows() or is_cross_compiling_on_apple(args)):
         raise BuildError(
             'Currently only support building protoc for Windows host while '
@@ -1738,7 +1742,7 @@ def main():
     args = parse_arguments()
     cmake_extra_defines = (args.cmake_extra_defines
                            if args.cmake_extra_defines else [])
-    cross_compiling = args.arm or args.arm64 or args.android
+    cross_compiling = args.arm or args.arm64 or args.arm64ec or args.android
 
     # If there was no explicit argument saying what to do, default
     # to update, build and test (for native builds).
@@ -1837,7 +1841,7 @@ def main():
             update_submodules(source_dir)
         if is_windows():
             if args.cmake_generator == 'Ninja':
-                if args.x86 or args.arm or args.arm64:
+                if args.x86 or args.arm or args.arm64 or args.arm64ec:
                     raise BuildError(
                         "To cross-compile with Ninja, load the toolset "
                         "environment for the target processor (e.g. Cross "
@@ -1847,7 +1851,7 @@ def main():
                 cmake_extra_args = [
                     '-A', 'Win32', '-T', 'host=x64', '-G', args.cmake_generator
                 ]
-            elif args.arm or args.arm64:
+            elif args.arm or args.arm64 or args.arm64ec:
                 # Cross-compiling for ARM(64) architecture
                 # First build protoc for host to use during cross-compilation
                 if path_to_protoc_exe is None:
@@ -1855,8 +1859,10 @@ def main():
                         cmake_path, source_dir, build_dir, args)
                 if args.arm:
                     cmake_extra_args = ['-A', 'ARM']
-                else:
+                elif args.arm64:
                     cmake_extra_args = ['-A', 'ARM64']
+                elif args.arm64ec:
+                    cmake_extra_args = ['-A', 'ARM64EC']
                 cmake_extra_args += ['-G', args.cmake_generator]
                 # Cannot test on host build machine for cross-compiled
                 # builds (Override any user-defined behaviour for test if any)