[2/N] Remove unnecessary once flag usage (#145057)

Fixes #ISSUE_NUMBER

Pull Request resolved: https://github.com/pytorch/pytorch/pull/145057
Approved by: https://github.com/albanD

aten/src/ATen/native/NNPACK.cpp

@@ -48,24 +48,18 @@ bool _nnpack_available() {
 namespace at::native {
 
 static bool init_nnpack() {
-  static c10::once_flag once_;
-  static bool nnpack_successfully_initialized_ = false;
+  const static nnp_status nnpack_status = nnp_initialize();
+  auto nnpack_successfully_initialized_ = (nnp_status_success == nnpack_status);
 
-  c10::call_once(once_, []() {
-    const nnp_status nnpack_status = nnp_initialize();
-    nnpack_successfully_initialized_ = (nnp_status_success == nnpack_status);
-
-    if (nnpack_status != nnp_status_success) {
-      if (nnpack_status == nnp_status_out_of_memory) {
-        LOG(WARNING) << "Could not initialize NNPACK! Reason: Out of memory.";
-      } else if (nnpack_status == nnp_status_unsupported_hardware) {
-        LOG(WARNING) << "Could not initialize NNPACK! Reason: Unsupported hardware.";
-      } else {
-        LOG(WARNING) << "Could not initialize NNPACK! Reason: Unknown error!";
-      }
+  if (nnpack_status != nnp_status_success) {
+    if (nnpack_status == nnp_status_out_of_memory) {
+      LOG(WARNING) << "Could not initialize NNPACK! Reason: Out of memory.";
+    } else if (nnpack_status == nnp_status_unsupported_hardware) {
+      LOG(WARNING) << "Could not initialize NNPACK! Reason: Unsupported hardware.";
+    } else {
+      LOG(WARNING) << "Could not initialize NNPACK! Reason: Unknown error!";
     }
-  });
-
+  }
   return nnpack_successfully_initialized_;
 }
 

aten/src/ATen/native/quantized/cpu/init_qnnpack.cpp

@@ -4,15 +4,10 @@
 #include <c10/util/Exception.h>
 #include <pytorch_qnnpack.h>
 
-#include <c10/util/CallOnce.h>
-
 namespace at::native {
 
 void initQNNPACK() {
-  static c10::once_flag once;
-  static enum pytorch_qnnp_status qnnpackStatus =
-      pytorch_qnnp_status_uninitialized;
-  c10::call_once(once, []() { qnnpackStatus = pytorch_qnnp_initialize(); });
+  static enum pytorch_qnnp_status qnnpackStatus = pytorch_qnnp_initialize();
   TORCH_CHECK(
       qnnpackStatus == pytorch_qnnp_status_success,
       "failed to initialize QNNPACK");

torch/csrc/autograd/engine.cpp

@@ -59,9 +59,8 @@ static void forked_autograd_child() {
 // Should be called before unsafe for forks (thread pool) calls
 static void track_bad_autograd_forks() {
 #if !defined(WIN32)
-  static c10::once_flag flag;
-  c10::call_once(
-      flag, [&] { pthread_atfork(nullptr, nullptr, forked_autograd_child); });
+  static auto result [[maybe_unused]] =
+      pthread_atfork(nullptr, nullptr, forked_autograd_child);
 #endif
 }
 

torch/csrc/cuda/Module.cpp

@@ -73,8 +73,8 @@ static void forked_child() {
 // has some working functions (e.g. device_count) but cannot fully initialize.
 static void poison_fork() {
 #ifndef WIN32
-  static c10::once_flag flag;
-  c10::call_once(flag, [] { pthread_atfork(nullptr, nullptr, forked_child); });
+  static auto result [[maybe_unused]] =
+      pthread_atfork(nullptr, nullptr, forked_child);
 #endif
 }
 

torch/csrc/cuda/memory_snapshot.cpp

@@ -98,8 +98,7 @@ CapturedTraceback* getFromContext(
 }
 
 void _initRecordAnnotations() {
-  static c10::once_flag ra_init;
-  c10::call_once(ra_init, [&] {
+  static auto init_placeholder [[maybe_unused]] = [&] {
     // Save user annotations to CCA memory snapshot tool
     at::addThreadLocalCallback(
         at::RecordFunctionCallback(
@@ -114,7 +113,8 @@ void _initRecordAnnotations() {
                   {{"name", fn.name()}, {"stage", "END"}});
             })
             .scopes({at::RecordScope::USER_SCOPE}));
-  });
+    return true;
+  }();
 }
 
 } // namespace

torch/csrc/distributed/c10d/NCCLUtils.cpp

@@ -390,11 +390,9 @@ std::unordered_map<std::string, std::string> NCCLComm::ncclCommDump() {
 #endif
 
 std::string getNcclVersion() {
-  static c10::once_flag ncclGetVersionFlag;
-  static std::string versionString;
-
-  c10::call_once(ncclGetVersionFlag, []() {
+  static std::string versionString = []() {
     int version = 0;
+    std::string versionString;
     ncclResult_t status = ncclGetVersion(&version);
     // can't compute the version if call did not return successfully or version
     // code < 100 (corresponding to 0.1.0)
@@ -417,7 +415,8 @@ std::string getNcclVersion() {
       }
 #endif
     }
-  });
+    return versionString;
+  }();
 
   return versionString;
 }

torch/csrc/distributed/c10d/ProcessGroupMPI.cpp

@@ -226,8 +226,6 @@ void ProcessGroupMPI::AsyncWork::populateException() {
 // Static global states
 int ProcessGroupMPI::mpiThreadSupport_ = 0;
 std::mutex ProcessGroupMPI::pgGlobalMutex_;
-// We only want to initialize once
-c10::once_flag ProcessGroupMPI::onceFlagInitMPI;
 
 void ProcessGroupMPI::mpiExit() {
   std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
@@ -235,8 +233,8 @@ void ProcessGroupMPI::mpiExit() {
 }
 
 void ProcessGroupMPI::initMPIOnce() {
-  // Initialize MPI environment
-  c10::call_once(onceFlagInitMPI, []() {
+  // Initialize MPI environment. We only want to initialize once.
+  static bool init_mpi_flag [[maybe_unused]] = []() {
     int mpi_was_initialized = 0;
     MPI_CHECK(MPI_Initialized(&mpi_was_initialized));
     if (mpi_was_initialized == 0) {
@@ -256,7 +254,8 @@ void ProcessGroupMPI::initMPIOnce() {
     } else {
       TORCH_WARN_ONCE("MPI was previously initialized.");
     }
-  });
+    return true;
+  }();
 }
 
 c10::intrusive_ptr<ProcessGroupMPI> ProcessGroupMPI::createProcessGroupMPI(

torch/csrc/distributed/c10d/ProcessGroupMPI.hpp

@@ -258,7 +258,6 @@ class TORCH_API ProcessGroupMPI : public Backend {
   // Global states
   static void initMPIOnce();
   static void mpiExit();
-  static c10::once_flag onceFlagInitMPI;
 
   static std::mutex pgGlobalMutex_;
   static int mpiThreadSupport_;

torch/csrc/mps/Module.cpp

@@ -1,7 +1,6 @@
 #define PYBIND11_DETAILED_ERROR_MESSAGES
 
 #include <ATen/ATen.h>
-#include <c10/util/CallOnce.h>
 #include <pybind11/pytypes.h>
 #include <torch/csrc/Generator.h>
 #include <torch/csrc/THP.h>
@@ -35,9 +34,8 @@ static void forked_mps_child() {
 // Should be called before the first mps call.
 static void track_bad_mps_fork() {
 #ifndef WIN32
-  static c10::once_flag flag;
-  c10::call_once(
-      flag, [] { pthread_atfork(nullptr, nullptr, forked_mps_child); });
+  static auto result [[maybe_unused]] =
+      pthread_atfork(nullptr, nullptr, forked_mps_child);
 #endif
 }
 } // namespace

torch/csrc/mtia/Module.cpp

@@ -1,7 +1,6 @@
 #include <ATen/ATen.h>
 #include <c10/core/DeviceType.h>
 #include <c10/core/Stream.h>
-#include <c10/util/CallOnce.h>
 #include <torch/csrc/Generator.h>
 #include <torch/csrc/Stream.h>
 #include <torch/csrc/python_headers.h>
@@ -28,8 +27,8 @@ static void forked_child() {
 // has some working functions (e.g. device_count) but cannot fully initialize.
 static void poison_fork() {
 #ifndef WIN32
-  static c10::once_flag flag;
-  c10::call_once(flag, [] { pthread_atfork(nullptr, nullptr, forked_child); });
+  static auto result [[maybe_unused]] =
+      pthread_atfork(nullptr, nullptr, forked_child);
 #endif
 }
 

torch/csrc/xpu/Module.cpp

@@ -32,8 +32,8 @@ static void forked_child() {
 // has some working functions (e.g. device_count) but cannot fully initialize.
 static void poison_fork() {
 #ifndef WIN32
-  static c10::once_flag flag;
-  c10::call_once(flag, [] { pthread_atfork(nullptr, nullptr, forked_child); });
+  static auto result [[maybe_unused]] =
+      pthread_atfork(nullptr, nullptr, forked_child);
 #endif
 }