SWDEV-399098 - Change hip_init lock to use std::call_once

Updating hip_init lock to use std::call_once fixed Performance
drops in TF benchmarks for FP16

Change-Id: Ib1125ac66806b232057ba183e296ea4d0642d08d


[ROCm/clr commit: 2f83719d12]

projects/clr/hipamd/src/hip_context.cpp

@@ -26,13 +26,13 @@
 #include "utils/versions.hpp"
 
 std::vector<hip::Device*> g_devices;
-amd::Monitor g_hipInitlock{"hipInit lock"};
+std::once_flag g_ihipInitialized;
 namespace hip {
 thread_local TlsAggregator tls;
 amd::Context* host_context = nullptr;
 
 //init() is only to be called from the HIP_INIT macro only once
-bool init() {
+void init(bool* status) {
   amd::IS_HIP = true;
   GPU_NUM_MEM_DEPENDENCY = 0;
 #if DISABLE_DIRECT_DISPATCH
@@ -42,7 +42,8 @@ bool init() {
 #endif
   AMD_DIRECT_DISPATCH = flagIsDefault(AMD_DIRECT_DISPATCH) ? kDirectDispatch : AMD_DIRECT_DISPATCH;
   if (!amd::Runtime::init()) {
-    return false;
+    *status = false;
+    return;
   }
   ClPrint(amd::LOG_INFO, amd::LOG_INIT, "Direct Dispatch: %d", AMD_DIRECT_DISPATCH);
 
@@ -52,7 +53,10 @@ bool init() {
   for (unsigned int i=0; i<devices.size(); i++) {
     const std::vector<amd::Device*> device(1, devices[i]);
     amd::Context* context = new amd::Context(device, amd::Context::Info());
-    if (!context) return false;
+    if (!context) {
+      *status = false;
+      return;
+    }
 
     // Enable active wait on the device by default
     devices[i]->SetActiveWait(true);
@@ -62,14 +66,18 @@ bool init() {
     } else {
       auto device = new Device(context, i);
       if ((device == nullptr) || !device->Create()) {
-        return false;
+        *status = false;
+        return;
       }
       g_devices.push_back(device);
     }
   }
 
   amd::Context* hContext = new amd::Context(devices, amd::Context::Info());
-  if (!hContext) return false;
+  if (!hContext) {
+    *status = false;
+    return;
+  }
 
   if (CL_SUCCESS != hContext->create(nullptr)) {
     hContext->release();
@@ -77,7 +85,8 @@ bool init() {
   host_context = hContext;
 
   PlatformState::instance().init();
-  return true;
+  *status = true;
+  return;
 }
 
 Device* getCurrentDevice() {

projects/clr/hipamd/src/hip_internal.hpp

@@ -71,29 +71,28 @@ typedef struct ihipIpcEventHandle_st {
 
 const char* ihipGetErrorName(hipError_t hip_error);
 
-extern amd::Monitor g_hipInitlock;
-#define HIP_INIT(noReturn) {\
-    amd::ScopedLock lock(g_hipInitlock);                     \
-    if (!amd::Runtime::initialized()) {                      \
-      if (!hip::init() && !noReturn) {                       \
-        HIP_RETURN(hipErrorInvalidDevice);                   \
-      }                                                      \
-    }                                                        \
-    if (hip::tls.device_ == nullptr && g_devices.size() > 0) {  \
-      hip::tls.device_ = g_devices[0];                          \
-      amd::Os::setPreferredNumaNode(g_devices[0]->devices()[0]->getPreferredNumaNode());  \
-    }                                                        \
+extern std::once_flag g_ihipInitialized;
+#define HIP_INIT(noReturn)                                                                         \
+  {                                                                                                \
+    bool status = true;                                                                            \
+    std::call_once(g_ihipInitialized, hip::init, &status);                                         \
+    if (!status  && !noReturn) {                                                                   \
+      HIP_RETURN(hipErrorInvalidDevice);                                                           \
+    }                                                                                              \
+    if (hip::tls.device_ == nullptr && g_devices.size() > 0) {                                     \
+      hip::tls.device_ = g_devices[0];                                                             \
+      amd::Os::setPreferredNumaNode(g_devices[0]->devices()[0]->getPreferredNumaNode());           \
+    }                                                                                              \
   }
 
-#define HIP_INIT_VOID() {\
-    amd::ScopedLock lock(g_hipInitlock);                     \
-    if (!amd::Runtime::initialized()) {                      \
-      if (hip::init()) {}                                    \
-    }                                                        \
-    if (hip::tls.device_ == nullptr && g_devices.size() > 0) {  \
-      hip::tls.device_ = g_devices[0];                          \
-      amd::Os::setPreferredNumaNode(g_devices[0]->devices()[0]->getPreferredNumaNode());  \
-    }                                                        \
+#define HIP_INIT_VOID()                                                                            \
+  {                                                                                                \
+    bool status = true;                                                                            \
+    std::call_once(g_ihipInitialized, hip::init, &status);                                         \
+    if (hip::tls.device_ == nullptr && g_devices.size() > 0) {                                     \
+      hip::tls.device_ = g_devices[0];                                                             \
+      amd::Os::setPreferredNumaNode(g_devices[0]->devices()[0]->getPreferredNumaNode());           \
+    }                                                                                              \
   }
 
 
@@ -532,7 +531,7 @@ namespace hip {
   /// Device representing the host - for pinned memory
   extern amd::Context* host_context;
 
-  extern bool init();
+  extern void init(bool* status);
 
   extern Device* getCurrentDevice();