Fix performance

Signed-off-by: cyy <[email protected]>

Author: cyyever

libkineto/include/libkineto.h

@@ -94,14 +94,14 @@ class LibkinetoApi {
   }
 
   void initProfilerIfRegistered() {
-    static std::once_flag once;
     if (activityProfiler_) {
-      std::call_once(once, [this] {
+      static bool once_flag = [this] {
         if (!activityProfiler_->isInitialized()) {
           activityProfiler_->init();
           initChildActivityProfilers();
         }
-      });
+        return true;
+      }();
     }
   }
 

libkineto/src/DeviceUtil.cpp

@@ -8,39 +8,36 @@
 
 #include "DeviceUtil.h"
 
-#include <mutex>
-
 namespace KINETO_NAMESPACE {
 
-bool amdGpuAvailable = false;
-bool cudaGpuAvailable = false;
-
 bool isAMDGpuAvailable() {
 #ifdef HAS_ROCTRACER
-  static std::once_flag once;
-  std::call_once(once, [] {
+  static bool amdGpuAvailable = [] {
     // determine AMD GPU availability on the system
     hipError_t error;
     int deviceCount;
     error = hipGetDeviceCount(&deviceCount);
-    amdGpuAvailable = (error == hipSuccess && deviceCount > 0);
-  });
-#endif
+    return (error == hipSuccess && deviceCount > 0);
+  }();
   return amdGpuAvailable;
+#else
+  return false;
+#endif
 }
 
 bool isCUDAGpuAvailable() {
 #ifdef HAS_CUPTI
-  static std::once_flag once;
-  std::call_once(once, [] {
+  static bool cudaGpuAvailable = [] {
     // determine CUDA GPU availability on the system
     cudaError_t error;
     int deviceCount;
     error = cudaGetDeviceCount(&deviceCount);
-    cudaGpuAvailable = (error == cudaSuccess && deviceCount > 0);
-  });
-#endif
+    return (error == cudaSuccess && deviceCount > 0);
+  }();
   return cudaGpuAvailable;
+#else
+  return false;
+#endif
 }
 
 bool isGpuAvailable() {

libkineto/test/EventProfilerTest.cpp

@@ -316,8 +316,8 @@ class EventProfilerTest : public ::testing::Test {
         loggers_,
         onDemandLoggers_);
 
-    for (int i = 0; i < kEventGroupCount; i++) {
-      eventGroups_[i] = &eventGroups_[i];
+    for (auto& eventGroup : eventGroups_) {
+      eventGroup = &eventGroup;
     }
     for (int i = 0; i < kGroupSetCount; i++) {
       // Default size to 1 but can be changed by test