fix(brain): fix winodws build

GitOrigin-RevId: 74eaea594e

CMakeLists.txt

@@ -106,8 +106,8 @@ if(MSVC OR WIN32)
     if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang-cl")
         message(FATAL_ERROR "only support clang-cl for windows build, pls check detail: scripts/cmake-build/BUILD_README.md")
     endif()
-    # add flags for sse/avx for X86
-    set(WIN_FLAGS "-msse4.2 -mavx -mavx2")
+    # add flags for enable sse instruction optimize for X86, enable avx header to compile avx code
+    set(WIN_FLAGS "-msse4.2 -O2 -D_AVX_ -D_AVX2_ -D__AVX__ -D__AVX2__ -D__FMA__")
     # if u CPU is cascadelake series, u can enable for performance
     # set(WIN_FLAGS "{WIN_FLAGS} -march=cascadelake -mtune=cascadelake")
     # set(WIN_FLAGS "{WIN_FLAGS} -mavx512cd -mavx512vl -mavx512dq -mavx512bw -mavx512vbmi -mavx512vnni")

src/opr/impl/utility.cpp

@@ -476,7 +476,9 @@ void AssertEqual::scn_do_execute_finish(const DeviceTensorND &) {
     auto err = DTypeScalar::make_from_raw(
             m_hv.dtype(), m_hv.raw_ptr()).get_cast<float>();
     if (m_param.verbose) {
-        fprintf(stderr,
+        //! FIXME: stderr will be slow when build windows with VS clang-cl (test in VM),
+        //! but I can`t find the root case. fix it when you figure out
+        fprintf(stdout,
                 "AssertEqual: err=%g (name=%s id=%zu)\n", err, cname(), id());
     }
     if (!(err >= 0 && err <= m_param.maxerr)) {