feat: Add BFloat16 support for Gemm and MatMul CPU operators

cmake/onnxruntime_mlas.cmake

@@ -809,6 +809,9 @@ endif()
 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} ${GSL_TARGET})
+    if (CPUINFO_SUPPORTED AND NOT CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
+      onnxruntime_add_include_to_target(${mlas_target} cpuinfo::cpuinfo)
+    endif()
 
     target_compile_definitions(${mlas_target} PRIVATE ${mlas_private_compile_definitions})
 

onnxruntime/core/common/endian.h

@@ -0,0 +1,27 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+namespace onnxruntime {
+
+// the semantics of this enum should match std::endian from C++20
+enum class endian {
+#if defined(_WIN32)
+  little = 0,
+  big = 1,
+  native = little,
+#elif defined(__GNUC__) || defined(__clang__)
+  little = __ORDER_LITTLE_ENDIAN__,
+  big = __ORDER_BIG_ENDIAN__,
+  native = __BYTE_ORDER__,
+#else
+#error onnxruntime::endian is not implemented in this environment.
+#endif
+};
+
+static_assert(
+    endian::native == endian::little || endian::native == endian::big,
+    "Only little-endian or big-endian native byte orders are supported.");
+
+}  // namespace onnxruntime