Added processor group support for >64 processors

platform/windows/syspovtask.cpp

@@ -34,9 +34,9 @@
 //******************************************************************************
 
 #include "syspovtask.h"
-
+#if defined (_WIN32) || defined (_WIN64) || defined (WIN32) || defined (WIN64)
 #include <windows.h>
-
+#endif
 // this must be the last file included
 #include "base/povdebug.h"
 
@@ -59,22 +59,6 @@ void WinMemThreadCleanup();
 namespace vfePlatform
 {
 
-// TODO - Maybe move this somewhere else as well.
-static unsigned int GetNumberofCPUs (void)
-{
-    SYSTEM_INFO         sysinfo;
-    static unsigned int result = 0;
-
-    // we cache the result, since this function is called on each thread startup
-    // TODO - this isn't ideal on systems with hot-pluggable CPUs
-    if (result != 0)
-      return result;
-
-    GetSystemInfo (&sysinfo) ;
-    result = sysinfo.dwNumberOfProcessors;
-    return result;
-}
-
 }
 // end of namespace vfeplatform
 
@@ -91,12 +75,6 @@ using namespace vfePlatform;
 
 void Task::Initialize ()
 {
-    // NB This is not thread-safe, but we currently don't care.
-    static volatile unsigned int count = 0;
-    unsigned int numCPUs = GetNumberofCPUs();
-    // TODO - if numCPUs > 64, we need to do more than this
-    if (numCPUs > 1)
-        SetThreadIdealProcessor (GetCurrentThread(), (count++) % numCPUs);
 #ifndef _CONSOLE
     povwin::WinMemThreadStartup();
 #endif

source/backend/support/processorGroups.cpp

@@ -0,0 +1,105 @@
+#if defined (_WIN32) || defined (_WIN64) || defined (WIN32) || defined (WIN64)
+
+#include <mutex>
+#include <Windows.h>
+#include "processorGroups.h"
+
+namespace pov
+{
+	namespace processorGroups
+	{
+		// used to dynamically load processor group utility functions from Kernel32.dll at runtime.
+		typedef WORD(*PFN_GETACTIVEPROCESSORGROUPCOUNT)(void);
+		typedef WORD(*PFN_GETACTIVEPROCESSORCOUNT)(WORD);
+		typedef BOOL(*PFN_SETTHREADGROUPAFFINITY)(HANDLE, const GROUP_AFFINITY*, PGROUP_AFFINITY);
+		static bool g_Initialized = false;
+		static HMODULE g_Kernel32 = NULL;
+		static PFN_GETACTIVEPROCESSORGROUPCOUNT g_GetActiveProcessorGroupCount = NULL;
+		static PFN_GETACTIVEPROCESSORCOUNT g_GetActiveProcessorCount = NULL;
+		static PFN_SETTHREADGROUPAFFINITY g_SetThreadGroupAffinity = NULL;
+
+		void Initialize()
+		{
+			// Initialize internal state if it hasn't already been done.
+			if (!g_Initialized)
+			{
+				// Prevent multiple threads from entering.
+				static std::mutex mutex;
+				std::lock_guard<std::mutex> lock(mutex);
+				if (!g_Initialized)
+				{
+					// Default all pointers to null.
+					g_Kernel32 = NULL;
+					g_GetActiveProcessorGroupCount = NULL;
+					g_GetActiveProcessorCount = NULL;
+					g_SetThreadGroupAffinity = 0;
+
+					// Load Kernel32.dll and retrieve required entry points.
+					g_Kernel32 = LoadLibrary("Kernel32.dll");
+					if (g_Kernel32 != NULL)
+					{
+						g_GetActiveProcessorGroupCount = (PFN_GETACTIVEPROCESSORGROUPCOUNT)GetProcAddress(g_Kernel32, "GetActiveProcessorGroupCount");
+						g_GetActiveProcessorCount = (PFN_GETACTIVEPROCESSORCOUNT)GetProcAddress(g_Kernel32, "GetActiveProcessorCount");
+						g_SetThreadGroupAffinity = (PFN_SETTHREADGROUPAFFINITY)GetProcAddress(g_Kernel32, "SetThreadGroupAffinity");
+					}
+
+					g_Initialized = true;
+				}
+			}
+		}
+
+		int GetNumberOfProcessors()
+		{
+			// Safety check.
+			Initialize();
+
+			// Check to see if processor group support is present, else default to old Windows 7 function.
+			bool hasProcessorGroupsSupport = (g_Kernel32 && g_GetActiveProcessorGroupCount && g_GetActiveProcessorCount && g_SetThreadGroupAffinity);
+			if (hasProcessorGroupsSupport) return g_GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
+			else
+			{
+				SYSTEM_INFO sysInfo = {};
+				GetSystemInfo(&sysInfo);
+				return sysInfo.dwNumberOfProcessors;
+			}
+		}
+
+		void SetThreadAffinity(int index)
+		{
+			// Skip setting thread affinity if incoming value was never set or isn't positive (Task class defaults taskIndex to -1).
+			if (index == -1)
+				return;
+
+			// Safety check.
+			Initialize();
+
+			// Check to see if processor group support is present.
+			bool hasProcessorGroupsSupport = (g_Kernel32 && g_GetActiveProcessorGroupCount && g_GetActiveProcessorCount && g_SetThreadGroupAffinity);
+			if (hasProcessorGroupsSupport)
+			{
+				// Since thread index may continue to increase unbounded, wrap it back around within the range of available processors.
+				int processorIndex = index % g_GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
+
+				// Determine which processor group to bind the thread to.
+				auto totalProcessors = 0U;
+				for (auto i = 0U; i < g_GetActiveProcessorGroupCount(); ++i)
+				{
+					totalProcessors += g_GetActiveProcessorCount(i);
+					if (totalProcessors >= processorIndex)
+					{
+						GROUP_AFFINITY groupAffinity = { ~0ULL, static_cast<WORD>(i), { 0, 0, 0 } };
+						g_SetThreadGroupAffinity(GetCurrentThread(), &groupAffinity, nullptr);
+						break;
+					}
+				}
+			}
+		}
+
+		void Deinitialize()
+		{
+			if (g_Kernel32)
+				FreeLibrary(g_Kernel32);
+		}
+	}
+}
+#endif

source/backend/support/processorGroups.h

@@ -0,0 +1,35 @@
+#if defined (_WIN32) || defined (_WIN64) || defined (WIN32) || defined (WIN64)
+#ifndef POVRAY_BACKEND_PROCESSORGROUPS_H
+#define POVRAY_BACKEND_PROCESSORGROUPS_H
+
+// This module enables support for processor groups in Windows 8 and above.
+// The implementation only requires a thread index be passed into SetThreadAffinity
+// which will then be used to select an appropriate logical processor on a system with
+// one or more processor groups.
+//
+// The module is meant to support earlier versions of Windows by dynamically loading appropriate
+// dlls and entry points at runtime if they exist.  If support for processor groups is not present
+// (e.g. Windows 7) then the older SetThreadAffinityMask will be used.
+namespace pov
+{
+	namespace processorGroups
+	{
+		// Initializes internal state (e.g. loading Kernel32.dll and getting required entry points).
+		// This does not have to be called explicitly.  The utility functions below will also call Initialize
+		// if it has not already been called.
+		extern void Initialize();
+
+		// Returns the total number of available processors.
+		extern int GetNumberOfProcessors();
+
+		// Sets the thread affinity for the calling thread given the specified index.
+		// Here index can pertain to any linearly increasing value (e.g. thread index, task index).
+		extern void SetThreadAffinity(int index);
+
+		// Frees and deinitializes internal state.
+		extern void Deinitialize();
+	}
+}
+
+#endif
+#endif

source/backend/support/task.cpp

@@ -58,6 +58,10 @@
 #include "backend/control/messagefactory.h"
 #include "backend/scene/backendscenedata.h"
 
+#if defined (_WIN32) || defined (_WIN64) || defined (WIN32) || defined (_WIN32)
+// Support for processor groups.
+#include "processorGroups.h"
+#endif
 // this must be the last file included
 #include "base/povdebug.h"
 
@@ -77,7 +81,8 @@ Task::Task(ThreadData *td, const boost::function1<void, Exception&>& f) :
     realTime(-1),
     cpuTime(-1),
     taskThread(nullptr),
-    povmsContext(nullptr)
+    povmsContext(nullptr),
+    taskIndex(-1)
 {
     if (td == nullptr)
         throw POV_EXCEPTION_STRING("Internal error: TaskData is NULL in Task constructor");
@@ -171,7 +176,10 @@ POV_LONG Task::ElapsedThreadCPUTime() const
 }
 
 void Task::TaskThread(const boost::function0<void>& completion)
-{
+{	
+#if defined (_WIN32) || defined (_WIN64) || defined (WIN32) || defined (WIN64)
+    processorGroups::SetThreadAffinity(taskIndex);
+#endif
     int result;
 
     if((result = POVMS_OpenContext(&povmsContext)) != kNoErr)

source/backend/support/task.h

@@ -92,6 +92,9 @@ class Task
 
         inline POVMSContext GetPOVMSContext() { return povmsContext; }
 
+	/// assigned task index.
+	int taskIndex;
+
     protected:
 
         struct StopThreadException final {}; // TODO - consider subclassing from std::exception hierarchy.
@@ -126,7 +129,7 @@ class Task
         /// task thread
         std::thread *taskThread;
         /// POVMS message receiving context
-        POVMSContext povmsContext;
+        POVMSContext povmsContext;		
 
         inline void FatalErrorHandler(const Exception& e)
         {

source/backend/support/taskqueue.cpp

@@ -34,6 +34,7 @@
 //******************************************************************************
 
 // Unit header file must be the first file included within POV-Ray *.cpp files (pulls in config)
+#include "base/types.h"
 #include "backend/support/taskqueue.h"
 
 // C++ variants of C standard header files
@@ -59,7 +60,7 @@ namespace pov
 using std::list;
 using std::shared_ptr;
 
-TaskQueue::TaskQueue() : failed(kNoError)
+TaskQueue::TaskQueue() : failed(kNoError), nextTaskIndex(0)
 {
 }
 
@@ -169,6 +170,7 @@ ThreadData *TaskQueue::AppendTask(Task *task)
 
     Notify();
 
+    task->taskIndex = nextTaskIndex++;
     return task->GetDataPtr();
 }
 
@@ -195,7 +197,6 @@ void TaskQueue::AppendFunction(const boost::function1<void, TaskQueue&>& fn)
     std::lock_guard<std::recursive_mutex> lock(queueMutex);
 
     queuedTasks.push(TaskEntry(fn));
-
     Notify();
 }
 
@@ -229,7 +230,7 @@ bool TaskQueue::Process()
         switch(queuedTasks.front().GetEntryType())
         {
             case TaskEntry::kTask:
-            {
+            {				
                 activeTasks.push_back(queuedTasks.front());
                 queuedTasks.front().GetTask()->Start(boost::bind(&TaskQueue::Notify, this));
                 queuedTasks.pop();

source/backend/support/taskqueue.h

@@ -128,6 +128,8 @@ class TaskQueue final
         int failed;
         /// wait for data in queue or related operation to be processed
         std::condition_variable_any processCondition;
+	/// next task index.
+	int nextTaskIndex;
 
         TaskQueue(const TaskQueue&) = delete;
         TaskQueue& operator=(const TaskQueue&) = delete;

windows/pvengine.cpp

@@ -90,6 +90,7 @@
 #include "backend/control/benchmark.h"
 #include "pvdisplay.h"
 #include "backend/povray.h"
+#include "backend/support/processorGroups.h"
 
 #ifdef RTR_SUPPORT
 #include "rtrsupport.h"
@@ -5123,7 +5124,7 @@ void GenerateDumpMeta(bool brief)
     s += sprintf(s, "installdate=%s\n", InstalledOn) ;
   }
   s += sprintf(s, "cpuarchitecture=%u\n", (DWORD) sysinfo.wProcessorArchitecture) ;
-  s += sprintf(s, "numberofcpus=%u\n", sysinfo.dwNumberOfProcessors) ;
+  s += sprintf(s, "numberofcpus=%u\n", processorGroups::GetNumberOfProcessors()) ;
   s += sprintf(s, "processortype=%u\n", sysinfo.dwProcessorType) ;
   s += sprintf(s, "processorlevel=%u\n", (DWORD) sysinfo.wProcessorLevel) ;
   s += sprintf(s, "processorrevision=%u\n", (DWORD) sysinfo.wProcessorRevision) ;
@@ -5430,10 +5431,8 @@ int PASCAL WinMain (HINSTANCE hInst, HINSTANCE hPrev, LPSTR szCmdLine, int sw)
 
   hInstance = hInst ;
   hMainThread = GetCurrentThread () ;
-
-  GetSystemInfo (&sysinfo) ;
-  ThreadCount = sysinfo.dwNumberOfProcessors ;
-  NumberOfCPUs = sysinfo.dwNumberOfProcessors ;
+
+  ThreadCount = NumberOfCPUs = processorGroups::GetNumberOfProcessors();
 
   while (*s == ' ' || *s == '\t')
     s++ ;

windows/pvupdate.cpp

@@ -43,7 +43,9 @@
 #include <stdio.h>
 
 #include "pvengine.h"
-
+#if defined (_WIN32) || defined (_WIN64) || defined (WIN32) || defined (WIN64)
+#include "backend/support/processorGroups.h"
+#endif
 // this must be the last file included
 #include "syspovdebug.h"
 
@@ -155,7 +157,7 @@ int IsUpdateAvailable (bool SendSysinfo, char *CurrentVersion, std::string& NewV
     strcpy (s, "&NoInfo=false\n") ;
     GetSystemInfo (&sysinfo) ;
     s += sprintf (s, "CPUArchitecture=0x%04x\n", (DWORD) sysinfo.wProcessorArchitecture) ;
-    s += sprintf (s, "NumberOfCPUs=0x%04x\n", sysinfo.dwNumberOfProcessors) ;
+    s += sprintf (s, "NumberOfCPUs=0x%04x\n", pov::processorGroups::GetNumberOfProcessors()) ;
     s += sprintf (s, "ProcessorType=0x%04x\n", sysinfo.dwProcessorType) ;
     s += sprintf (s, "ProcessorLevel=0x%04x\n", (DWORD) sysinfo.wProcessorLevel) ;
     s += sprintf (s, "ProcessorRevision=0x%04x\n", (DWORD) sysinfo.wProcessorRevision) ;

windows/vs2015/povbackend.vcxproj

@@ -413,6 +413,7 @@
     <ClCompile Include="..\..\source\backend\render\rendertask.cpp" />
     <ClCompile Include="..\..\source\backend\render\tracetask.cpp" />
     <ClCompile Include="..\..\source\backend\scene\view.cpp" />
+    <ClCompile Include="..\..\source\backend\support\processorGroups.cpp" />
     <ClCompile Include="..\..\source\backend\support\task.cpp" />
     <ClCompile Include="..\..\source\backend\support\taskqueue.cpp" />
     <ClCompile Include="..\..\source\backend\lighting\photonestimationtask.cpp" />
@@ -452,6 +453,7 @@
     <ClInclude Include="..\..\source\backend\scene\view.h" />
     <ClInclude Include="..\..\source\backend\scene\viewthreaddata_fwd.h" />
     <ClInclude Include="..\..\source\backend\scene\view_fwd.h" />
+    <ClInclude Include="..\..\source\backend\support\processorGroups.h" />
     <ClInclude Include="..\..\source\backend\support\task.h" />
     <ClInclude Include="..\..\source\backend\support\taskqueue.h" />
     <ClInclude Include="..\..\source\backend\lighting\photonestimationtask.h" />

windows/vs2015/povbackend.vcxproj.filters

@@ -114,6 +114,9 @@
     <ClCompile Include="..\..\source\backend\control\parsertask.cpp">
       <Filter>Backend Source\Control</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\source\backend\support\processorGroups.cpp">
+      <Filter>Backend Source\Support</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\..\source\backend\configbackend.h">
@@ -209,5 +212,8 @@
     <ClInclude Include="..\..\source\backend\control\messagefactory_fwd.h">
       <Filter>Backend Headers\Control</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\source\backend\support\processorGroups.h">
+      <Filter>Backend Headers\Support</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>