SL-18154 WIP -- CPU sampling (AMD uProf) profile guided optimizations to reduce CPU usage of background threads.

indra/llcommon/llqueuedthread.cpp

@@ -477,9 +477,14 @@ void LLQueuedThread::processRequest(LLQueuedThread::QueuedRequest* req)
                 mRequestQueue.post([=]
                     {
                         LL_PROFILE_ZONE_NAMED("processRequest - retry");
-                        while (LL::WorkQueue::TimePoint::clock::now() < retry_time)
+                        if (LL::WorkQueue::TimePoint::clock::now() < retry_time)
                         {
-                            std::this_thread::yield(); //note: don't use LLThread::yield here to avoid 
+                            auto sleep_time = std::chrono::duration_cast<std::chrono::milliseconds>(retry_time - LL::WorkQueue::TimePoint::clock::now());
+                            
+                            if (sleep_time.count() > 0) 
+                            {
+                                ms_sleep(sleep_time.count());
+                            }
                         }
                         processRequest(req);
                     });

indra/llcommon/llthread.cpp

@@ -42,6 +42,7 @@
 
 
 #ifdef LL_WINDOWS
+
 const DWORD MS_VC_EXCEPTION=0x406D1388;
 
 #pragma pack(push,8)
@@ -133,6 +134,15 @@ void LLThread::threadRun()
 {
 #ifdef LL_WINDOWS
     set_thread_name(-1, mName.c_str());
+
+#if 0 // probably a bad idea, see usage of SetThreadIdealProcessor in LLWindowWin32)
+    HANDLE hThread = GetCurrentThread();
+    if (hThread)
+    {
+        SetThreadAffinityMask(hThread, (DWORD_PTR) 0xFFFFFFFFFFFFFFFE);
+    }
+#endif
+
 #endif
 
     LL_PROFILER_SET_THREAD_NAME( mName.c_str() );

indra/llcommon/lltimer.cpp

@@ -92,6 +92,7 @@ U32 micro_sleep(U64 us, U32 max_yields)
 U32 micro_sleep(U64 us, U32 max_yields)
 {
     LL_PROFILE_ZONE_SCOPED
+#if 0
     LARGE_INTEGER ft;
     ft.QuadPart = -static_cast<S64>(us * 10);  // '-' using relative time
 
@@ -99,6 +100,9 @@ U32 micro_sleep(U64 us, U32 max_yields)
     SetWaitableTimer(timer, &ft, 0, NULL, NULL, 0);
     WaitForSingleObject(timer, INFINITE);
     CloseHandle(timer);
+#else
+    Sleep(us / 1000);
+#endif
 
     return 0;
 }

indra/llwindow/llwindowwin32.cpp

@@ -467,7 +467,6 @@ LLWindowWin32::LLWindowWin32(LLWindowCallbacks* callbacks,
         // process deprioritization during profiles
     // force high thread priority
     HANDLE hProcess = GetCurrentProcess();
-    HANDLE hThread = GetCurrentThread();
 
     if (hProcess)
     {
@@ -484,6 +483,20 @@ LLWindowWin32::LLWindowWin32(LLWindowCallbacks* callbacks,
             }
         }
     }
+#endif
+
+#if 0  // this is also probably a bad idea, but keep it in your back pocket for getting main thread off of background thread cores (see also LLThread::threadRun)
+    HANDLE hThread = GetCurrentThread();
+
+    SYSTEM_INFO sysInfo;
+
+    GetSystemInfo(&sysInfo);
+    U32 core_count = sysInfo.dwNumberOfProcessors;
+
+    if (max_cores != 0)
+    {
+        core_count = llmin(core_count, max_cores);
+    }
 
     if (hThread)
     {
@@ -499,6 +512,9 @@ LLWindowWin32::LLWindowWin32(LLWindowCallbacks* callbacks,
             {
                 LL_INFOS() << "Failed to set thread priority: " << std::hex << GetLastError() << LL_ENDL;
             }
+
+            // tell main thread to prefer core 0
+            SetThreadIdealProcessor(hThread, 0);
         }
     }
 #endif

indra/newview/llappviewer.cpp

@@ -2181,6 +2181,12 @@ bool LLAppViewer::initThreads()
     // get the number of concurrent threads that can run
     S32 cores = std::thread::hardware_concurrency();
 
+    U32 max_cores = gSavedSettings.getU32("EmulateCoreCount");
+    if (max_cores != 0)
+    {
+        cores = llmin(cores, (S32) max_cores);
+    }
+
     // The only configurable thread count right now is ImageDecode
     // The viewer typically starts around 8 threads not including image decode, 
     // so try to leave at least one core free