flush denormal in the tracer op, test=develop (#32350)

* flush denormal in the tracer op, test=develop

* add cmake dependencies, test=develop

* add a macro, test=develop

* fix the windows case, test=develop

paddle/fluid/imperative/CMakeLists.txt

@@ -4,7 +4,7 @@ cc_library(prepared_operator SRCS prepared_operator.cc DEPS proto_desc operator
 cc_library(layer SRCS layer.cc DEPS prepared_operator math_function imperative_flag variable_helper op_registry)
 add_subdirectory(jit)
 cc_library(amp SRCS amp_auto_cast.cc DEPS layer )
-cc_library(tracer SRCS tracer.cc DEPS layer engine program_desc_tracer amp)
+cc_library(tracer SRCS tracer.cc DEPS layer engine program_desc_tracer amp denormal)
 cc_library(basic_engine SRCS basic_engine.cc DEPS layer gradient_accumulator)
 cc_library(engine SRCS basic_engine.cc partial_grad_engine.cc DEPS layer gradient_accumulator)
 cc_library(imperative_profiler SRCS profiler.cc)

paddle/fluid/imperative/tracer.cc

@@ -19,6 +19,7 @@
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/imperative/amp_auto_cast.h"
 #include "paddle/fluid/imperative/op_base.h"
+#include "paddle/fluid/platform/denormal.h"
 #include "paddle/fluid/platform/profiler.h"
 #include "paddle/fluid/string/string_helper.h"
 
@@ -134,6 +135,7 @@ void Tracer::TraceOp(const std::string& type, const NameVarBaseMap& ins,
                      const platform::Place& place, bool trace_backward,
                      const std::map<std::string, std::string>& inplace_map) {
   platform::RecordEvent op_type_record_event(type);
+  platform::ScopedFlushDenormal flush;
   VLOG(1) << "Trace Op: " << type;
   if (FLAGS_use_mkldnn) {
     // if both lists are empty all ops are enabled (default for

paddle/fluid/platform/denormal.cc

@@ -28,7 +28,7 @@
 #endif
 
 #if !defined(GCC_WITHOUT_INTRINSICS) && !defined(PADDLE_WITH_ARM) && \
-    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS)
+    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS) && !defined(_WIN32)
 #define DENORM_USE_INTRINSICS
 #endif