Add CMake mlas_private_compile_definitions variable for internal MLAS definitions. (microsoft#26225)

### Description
<!-- Describe your changes. -->

Add CMake `mlas_private_compile_definitions` variable for internal MLAS
definitions.

### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

Refactor so that it is easier to add new private compile definitions for
MLAS and related targets.

Author: edgchen1

cmake/onnxruntime_mlas.cmake

@@ -5,6 +5,9 @@ set(MLAS_ROOT ${ONNXRUNTIME_ROOT}/core/mlas)
 set(MLAS_SRC_DIR ${MLAS_ROOT}/lib)
 set(MLAS_INC_DIR ${MLAS_ROOT}/inc)
 
+# mlas_private_compile_definitions contains compile definitions that are private to onnxruntime_mlas and targets which
+# use internal MLAS headers like mlasi.h.
+set(mlas_private_compile_definitions)
 #
 # All hardware agnostic source files here
 # hardware specific files would cause trouble in
@@ -133,9 +136,9 @@ function(setup_mlas_source_for_windows)
       )
 
       if (onnxruntime_USE_ARM_NEON_NCHWC)
-		setup_arm_neon_nchwc()	
+		setup_arm_neon_nchwc()
 	  endif()
-      
+
 	  if (onnxruntime_USE_KLEIDIAI)
         setup_kleidiai()
       endif()
@@ -293,11 +296,12 @@ endfunction()
 
 function (setup_arm_neon_nchwc)
   target_sources(onnxruntime_mlas PRIVATE
-   ${MLAS_SRC_DIR}/sconv.h  
+   ${MLAS_SRC_DIR}/sconv.h
    ${MLAS_SRC_DIR}/sconv_kernel_neon.cpp
    ${MLAS_SRC_DIR}/spool_kernel_neon.cpp
   )
-  target_compile_definitions(onnxruntime_mlas PRIVATE MLAS_USE_ARM_NEON_NCHWC)
+  list(APPEND mlas_private_compile_definitions MLAS_USE_ARM_NEON_NCHWC)
+  set(mlas_private_compile_definitions ${mlas_private_compile_definitions} PARENT_SCOPE)
 endfunction ()
 
 if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
@@ -445,25 +449,25 @@ else()
           ${MLAS_SRC_DIR}/eltwise_kernel_neon.cpp
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_neon_int8_i8mm.cpp
         )
-        
+
         # Conditionally add the SVE implementation if compiler supports it
         if (onnxruntime_USE_SVE)
           list(APPEND mlas_platform_srcs ${MLAS_SRC_DIR}/sve/mlasi_sve.h)
           list(APPEND mlas_platform_srcs ${MLAS_SRC_DIR}/sve/elementwise_sve.cpp)
           set_source_files_properties(${MLAS_SRC_DIR}/sve/elementwise_sve.cpp PROPERTIES COMPILE_FLAGS " -march=armv8.2-a+sve+fp16 ")
-          target_compile_definitions(onnxruntime_mlas PRIVATE MLAS_USE_SVE)
+          list(APPEND mlas_private_compile_definitions MLAS_USE_SVE)
         endif()
 
         if (onnxruntime_USE_ARM_NEON_NCHWC)
-		  setup_arm_neon_nchwc()	
+		  setup_arm_neon_nchwc()
 		endif()
-        
+
 		if (onnxruntime_USE_KLEIDIAI)
           setup_kleidiai()
         endif()
         set_source_files_properties(${MLAS_SRC_DIR}/sqnbitgemm_kernel_neon_int8.cpp
                                     PROPERTIES COMPILE_FLAGS " -march=armv8.2-a+dotprod")
-        set_source_files_properties(${MLAS_SRC_DIR}/sqnbitgemm_kernel_neon_int8_i8mm.cpp 
+        set_source_files_properties(${MLAS_SRC_DIR}/sqnbitgemm_kernel_neon_int8_i8mm.cpp
 				    PROPERTIES COMPILE_FLAGS " -march=armv8.2-a+i8mm ")
 
         if (NOT APPLE)
@@ -806,6 +810,8 @@ foreach(mlas_target ${ONNXRUNTIME_MLAS_LIBS})
     target_include_directories(${mlas_target} PRIVATE ${MLAS_INC_DIR} ${MLAS_SRC_DIR})
     onnxruntime_add_include_to_target(${mlas_target} ${GSL_TARGET})
 
+    target_compile_definitions(${mlas_target} PRIVATE ${mlas_private_compile_definitions})
+
     set_target_properties(${mlas_target} PROPERTIES FOLDER "ONNXRuntime")
 endforeach()
 

cmake/onnxruntime_unittests.cmake

@@ -1325,9 +1325,7 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
       ${BENCHMARK_DIR}/layer_normalization.cc)
     target_include_directories(onnxruntime_benchmark PRIVATE ${ONNXRUNTIME_ROOT} ${onnxruntime_graph_header} ${ONNXRUNTIME_ROOT}/core/mlas/inc)
     target_compile_definitions(onnxruntime_benchmark PRIVATE BENCHMARK_STATIC_DEFINE)
-    if (onnxruntime_USE_SVE)
-      target_compile_definitions(onnxruntime_benchmark PRIVATE MLAS_USE_SVE)
-    endif()
+    target_compile_definitions(onnxruntime_benchmark PRIVATE ${mlas_private_compile_definitions})
     if(WIN32)
       target_compile_options(onnxruntime_benchmark PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:--compiler-options /wd4141>"
                         "$<$<NOT:$<COMPILE_LANGUAGE:CUDA>>:/wd4141>")
@@ -1355,9 +1353,7 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
     target_include_directories(onnxruntime_mlas_benchmark PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc)
     target_link_libraries(onnxruntime_mlas_benchmark PRIVATE benchmark::benchmark onnxruntime_util ${ONNXRUNTIME_MLAS_LIBS} onnxruntime_common ${CMAKE_DL_LIBS})
     target_compile_definitions(onnxruntime_mlas_benchmark PRIVATE BENCHMARK_STATIC_DEFINE)
-    if (onnxruntime_USE_SVE)
-      target_compile_definitions(onnxruntime_mlas_benchmark PRIVATE MLAS_USE_SVE)
-    endif()
+    target_compile_definitions(onnxruntime_mlas_benchmark PRIVATE ${mlas_private_compile_definitions})
     if(WIN32)
       target_link_libraries(onnxruntime_mlas_benchmark PRIVATE debug Dbghelp)
       # Avoid using new and delete. But this is a benchmark program, it's ok if it has a chance to leak.
@@ -1655,9 +1651,7 @@ endif()
         XCODE_ATTRIBUTE_CODE_SIGNING_ALLOWED "NO"
       )
     endif()
-    if (onnxruntime_USE_SVE)
-      target_compile_definitions(onnxruntime_mlas_test PRIVATE MLAS_USE_SVE)
-    endif()
+    target_compile_definitions(onnxruntime_mlas_test PRIVATE ${mlas_private_compile_definitions})
     target_include_directories(onnxruntime_mlas_test PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc ${ONNXRUNTIME_ROOT}
             ${CMAKE_CURRENT_BINARY_DIR})
     target_link_libraries(onnxruntime_mlas_test PRIVATE GTest::gtest GTest::gmock ${ONNXRUNTIME_MLAS_LIBS} onnxruntime_common)