In a Linux or Android build check if the compiler support bfloat16 an…

…d float16 (#18813)

### Description
Restrict clang version because we have an upcoming change that requires
clang version >=16 , which will mainly affect Android build.

cmake/CMakeLists.txt

@@ -354,13 +354,7 @@ if (onnxruntime_USE_ROCM)
   endif()
 endif()
 
-if (APPLE)
-    if (NOT CMAKE_OSX_ARCHITECTURES)
-        message("Building ONNX Runtime for ${CMAKE_HOST_SYSTEM_PROCESSOR}")
-    endif()
-elseif (NOT WIN32 AND NOT APPLE)
-    message("Building ONNX Runtime for ${CMAKE_SYSTEM_PROCESSOR}")
-endif()
+
 
 # Single output director for all binaries
 set(RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin CACHE PATH "Single output directory for all binaries.")
@@ -493,6 +487,14 @@ endif()
 
 include(adjust_global_compile_flags.cmake)
 
+if (APPLE)
+  if (NOT CMAKE_OSX_ARCHITECTURES)
+    message("Building ONNX Runtime for ${CMAKE_HOST_SYSTEM_PROCESSOR} CPU ARCH")
+  endif()
+elseif (NOT WIN32 AND NOT APPLE)
+  message("Building ONNX Runtime for ${onnxruntime_target_platform} CPU ARCH")
+endif()
+
 # We need to link with libatomic on systems that do not have built-in atomics, or
 # don't have built-in support for 8 byte atomics
 # Derived from https://github.com/protocolbuffers/protobuf/blob/master/cmake/CMakeLists.txt
@@ -639,7 +641,16 @@ else()
   check_cxx_compiler_flag(-Wunused-variable HAS_UNUSED_VARIABLE)
   check_cxx_compiler_flag(-Wuseless-cast HAS_USELESS_CAST)
   check_function_exists(reallocarray HAS_REALLOCARRAY)
-
+  if (NOT APPLE AND NOT CMAKE_SYSTEM_NAME STREQUAL "Emscripten" AND onnxruntime_target_platform STREQUAL "aarch64")
+   check_cxx_compiler_flag(-march=armv8.2-a+bf16 HAS_ARM64_BFLOAT16)
+   if(NOT HAS_ARM64_BFLOAT16)
+     message(FATAL_ERROR  "The compiler doesn't support BFLOAT16!!!")
+   endif()
+   check_cxx_compiler_flag(-march=armv8.2-a+fp16 HAS_ARM64_FLOAT16)
+   if(NOT HAS_ARM64_FLOAT16)
+     message(FATAL_ERROR  "The compiler doesn't support FLOAT16!!!")
+   endif()
+  endif()
   if (HAS_TAUTOLOGICAL_POINTER_COMPARE)
     #we may have extra null pointer checkings in debug build, it's not an issue
     list(APPEND ORT_WARNING_FLAGS -Wno-tautological-pointer-compare)

cmake/adjust_global_compile_flags.cmake

@@ -300,6 +300,31 @@ if (MSVC)
   endif()
 else()
   if (NOT APPLE)
+    #XXX: Sometimes the value of CMAKE_SYSTEM_PROCESSOR is set but it's wrong. For example, if you run an armv7 docker
+    #image on an aarch64 machine with an aarch64 Ubuntu host OS, in the docker instance cmake may still report
+    # CMAKE_SYSTEM_PROCESSOR as aarch64 by default. Given compiling this code may need more than 2GB memory, we do not
+    # support compiling for ARM32 natively(only support cross-compiling), we will ignore this issue for now.
+    if(NOT CMAKE_SYSTEM_PROCESSOR)
+      message(WARNING "CMAKE_SYSTEM_PROCESSOR is not set. Please set it in your toolchain cmake file.")
+      # Try to detect it
+      if("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU" OR "${CMAKE_C_COMPILER_ID}" STREQUAL "Clang")
+        execute_process(
+		COMMAND "${CMAKE_C_COMPILER}" -dumpmachine
+		OUTPUT_VARIABLE GCC_DUMP_MACHINE_OUT OUTPUT_STRIP_TRAILING_WHITESPACE
+		ERROR_VARIABLE _err
+		RESULT_VARIABLE _res
+		)
+		if(NOT _res EQUAL 0)
+			message(SEND_ERROR "Failed to run 'gcc -dumpmachine':\n ${_res}")
+		endif()
+		string(REPLACE "-" ";" GCC_DUMP_MACHINE_OUT_LIST "${GCC_DUMP_MACHINE_OUT}")
+		list(LENGTH GCC_DUMP_MACHINE_OUT_LIST GCC_TRIPLET_LEN)
+		if(GCC_TRIPLET_LEN EQUAL 4)
+		  list(GET GCC_DUMP_MACHINE_OUT_LIST 0 CMAKE_SYSTEM_PROCESSOR)
+          message("Setting CMAKE_SYSTEM_PROCESSOR to ${CMAKE_SYSTEM_PROCESSOR}")
+        endif()
+      endif()
+    endif()
     set(onnxruntime_target_platform ${CMAKE_SYSTEM_PROCESSOR})
   endif()
   if (onnxruntime_BUILD_FOR_NATIVE_MACHINE)