Miscellaneous optimization... and removed isnormal (...) checks from frame pacing statistics

Author: Kaldaien

include/SpecialK/framerate.h

@@ -842,9 +842,9 @@ struct game_pacer_s {
   DWORD   last_paced_time = 0UL;
   HANDLE  event           = 0;
 
-  bool   isSupported (void) noexcept;
-  bool   wantPacing  (void) noexcept;
-  void   signalEvent (void) noexcept;
+  bool   isSupported (ULONG64 frames_drawn = SK_GetFramesDrawn ()) noexcept;
+  bool   wantPacing  (ULONG64 frames_drawn = SK_GetFramesDrawn ()) noexcept;
+  void   signalEvent (void)                                        noexcept;
 } game_pace;
 
 

src/core.cpp

@@ -2904,8 +2904,18 @@ SK_Inject_PostHeartbeatToSKIF (void)
     //   but alt-tab can still flicker.
     if (hWndSKIF != game_window.hWnd && IsWindow (hWndSKIF))
     {
-      DWORD                                   dwPid = 0x0;
-      SK_GetWindowThreadProcessId (hWndSKIF, &dwPid);
+      static HWND   hWndLast = 0;
+      static DWORD dwPidLast = 0;
+
+      DWORD dwPid = 0x0;
+
+      if (hWndLast == hWndSKIF)
+              dwPid = dwPidLast;
+      else
+      {      
+        SK_GetWindowThreadProcessId (hWndSKIF, &dwPid);
+                          hWndLast = hWndSKIF;  dwPidLast = dwPid;
+      }
 
       if ( dwPid != 0x0 &&
            dwPid != GetCurrentProcessId () )
@@ -4730,13 +4740,11 @@ SK_EndBufferSwap (HRESULT hr, IUnknown* device, SK_TLS* pTLS)
     if ( config.render.framerate.present_interval == 0 &&
          config.render.framerate.target_fps        > 0.0f )
     {
-      SK_AutoHandle hTimer (
-        INVALID_HANDLE_VALUE
-      );
+      static thread_local HANDLE hTimer = (HANDLE)-1;
 
       SK_Framerate_WaitUntilQPC (
         qpcTimeOfSwap.QuadPart + __SK_LatentSyncPostDelay,
-                    hTimer.m_h  );
+                        hTimer  );
     }
   }
 

src/framerate.cpp

@@ -2162,6 +2162,13 @@ SK_Framerate_SanitizeTimerResolution (void) noexcept
   if (! config.render.framerate.max_timer_resolution)
     return;
 
+  // Only do this occasionally, it defeats the purpose to
+  //   call it frequently.
+  static int
+        tick_tock       = 0;
+  if (++tick_tock % 64 != 0)
+    return;
+
   auto _SetTimerResolution =
     ( ZwSetTimerResolution_Original != nullptr ) ?
       ZwSetTimerResolution_Original              :
@@ -2211,7 +2218,7 @@ extern bool                                  reset_frame_history;
 
 void
 SK::Framerate::Limiter::wait (void) noexcept
-{
+{  
   // This is actually counter-productive in testing... when the thread priority
   //   is lowered after this goes out of scope it may reschedule the thread.
   //SK_Thread_ScopedPriority prio_scope (THREAD_PRIORITY_TIME_CRITICAL);
@@ -2506,6 +2513,7 @@ SK::Framerate::Limiter::wait (void) noexcept
       }
     }
 
+    // Trigger a timer phase shift.
     SK_Framerate_SanitizeTimerResolution ();
 
     // Create an unnamed waitable timer.
@@ -4502,7 +4510,7 @@ SK::Framerate::TickEx ( bool     /*wait*/,
 
 
   // Prevent inserting infinity into the dataset
-  if ( std::isnormal (dt) )
+  //if ( std::isnormal (dt) )
   {
     if (snapshots->frame_history.addSample (1000.0 * dt, now))
     {
@@ -4597,7 +4605,7 @@ SK::Framerate::TickEx ( bool     /*wait*/,
       } break;
     }
 
-    if (std::isnormal (sample) && now.QuadPart > 0)
+    if (/*std::isnormal(sample) &&*/ now.QuadPart > 0)
     {
       container->addSample (
         sample, now
@@ -4610,25 +4618,31 @@ SK::Framerate::TickEx ( bool     /*wait*/,
 
 int sk_config_t::fps_osd_s::getTimingMethod (void)
 {
-  if (__SK_IsDLSSGActive)
-  {      
-    if (config.render.framerate.streamline.enable_native_limit)
+  switch (timing_method)
+  {
+    case SK_FrametimeMeasures_LimiterPacing:
     {
-      if (__target_fps_now <= 0.0f && timing_method == SK_FrametimeMeasures_LimiterPacing)
+      if (__SK_IsDLSSGActive)
+      {      
+        if (config.render.framerate.streamline.enable_native_limit)
+        {
+          if (__target_fps_now <= 0.0f)
+          {
+            return SK_FrametimeMeasures_NewFrameBegin;
+          }
+        }
+
+        else if (__target_fps_now > 0.0f)
+        {
+          return SK_FrametimeMeasures_PresentSubmit;
+        }
+      }
+
+      else if (__target_fps_now <= 0.0f)
       {
         return SK_FrametimeMeasures_NewFrameBegin;
       }
-    }
-
-    else if (__target_fps_now > 0.0f && timing_method == SK_FrametimeMeasures_LimiterPacing)
-    {
-      return SK_FrametimeMeasures_PresentSubmit;
-    }
-  }
-
-  else if (__target_fps_now <= 0.0f && timing_method == SK_FrametimeMeasures_LimiterPacing)
-  {
-    return SK_FrametimeMeasures_NewFrameBegin;
+    } break;
   }
 
   return timing_method;
@@ -4860,7 +4874,7 @@ SK::Framerate::Stats::sortAndCacheFrametimeHistory (void) //noexcept
     {
       if (datum.when.QuadPart >= 0)
       {
-        if (isnormal (datum.val))
+        //if (isnormal (datum.val))
         {
           kWriteBuffer.second.emplace_back (datum.val);
         }
@@ -4891,6 +4905,7 @@ SK_Framerate_WaitUntilQPC (LONGLONG llQPC, HANDLE& hTimer) noexcept
   if (llQPC < SK_QueryPerf ().QuadPart)
     return;
 
+  // Trigger a timer phase shift.
   SK_Framerate_SanitizeTimerResolution ();
 
   if ((LONG_PTR)hTimer < 0)
@@ -5066,10 +5081,10 @@ SK_Framerate_EnergyControlPanel (void)
 }
 
 bool
-game_pacer_s::wantPacing (void) noexcept
+game_pacer_s::wantPacing (ULONG64 frames_drawn) noexcept
 {
   return 
-    config.render.framerate.pace_game_thread && isSupported () && event != 0 && last_frame_id > SK_GetFramesDrawn () - 2;
+    config.render.framerate.pace_game_thread && isSupported (frames_drawn) && event != 0 && last_frame_id > frames_drawn - 2;
 }
 
 void
@@ -5080,10 +5095,10 @@ game_pacer_s::signalEvent (void) noexcept
 }
 
 bool
-game_pacer_s::isSupported (void) noexcept
+game_pacer_s::isSupported (ULONG64 frames_drawn) noexcept
 {
   static constexpr auto SK_Reflex_MinimumFramesBeforeNative = 150;
 
   return
-    last_frame_id != 0 && !config.nvidia.reflex.native && SK_GetFramesDrawn () > SK_Reflex_MinimumFramesBeforeNative;
+    last_frame_id != 0 && !config.nvidia.reflex.native && frames_drawn > SK_Reflex_MinimumFramesBeforeNative;
 }

src/plugins/unity.cpp

@@ -2981,9 +2981,9 @@ SK_Unity_PaceGameThreadDxgi (IDXGISwapChain *pSwapChain, DXGI_FRAME_STATISTICS *
   game_pace.last_frame_id =
     SK_GetFramesDrawn ();
 
-  if (game_pace.wantPacing ())
+  if (game_pace.wantPacing (game_pace.last_frame_id))
   {
-    static HANDLE hTimer = (HANDLE)-1;
+    static thread_local HANDLE hTimer = (HANDLE)-1;
 
     auto& rb =
       SK_GetCurrentRenderBackend ();

src/render/d3d11/d3d11_state_tracker.cpp

@@ -201,7 +201,7 @@ SK_D3D11_ShouldTrackRenderOp ( ID3D11DeviceContext* pDevCtx,
     const auto frame_id =
       SK_GetFramesDrawn ();
 
-    if (!game_pace.wantPacing () || game_pace.last_paced_time < SK_timeGetTime ())
+    if (!game_pace.wantPacing (frame_id) || game_pace.last_paced_time < SK::ControlPanel::current_time)
     {
       static ULONG64
           last_frame = 0;

src/render/dxgi/dxgi.cpp

@@ -2144,12 +2144,14 @@ SK_ImGui_DrawD3D12 (IDXGISwapChain* This)
   }
 }
 
+#ifdef USE_D3D11_GPU_WAIT
 static HANDLE SK_ImGui_D3D11GPUEvent = 0;
+#endif
 
 void
 SK_ImGui_WaitD3D11 (void)
 {
-#if 0
+#ifdef USE_D3D11_GPU_WAIT
   if (SK_ImGui_D3D11GPUEvent != 0)
     WaitForSingleObject (SK_ImGui_D3D11GPUEvent, 100UL);
 #endif
@@ -2178,15 +2180,21 @@ SK_ImGui_DrawD3D11 (IDXGISwapChain* This)
   if (! pTLS)
     return;
 
+#ifdef USE_D3D11_GPU_WAIT
   SK_RunOnce (
     SK_ImGui_D3D11GPUEvent =
       SK_CreateEvent (nullptr, FALSE, TRUE, nullptr);
   );
 
-  SK_ComQIPtr <IDXGIDevice2>
-      pDXGIDev2 (pDev);
-  if (pDXGIDev2 != nullptr)
-      pDXGIDev2->EnqueueSetEvent (SK_ImGui_D3D11GPUEvent);
+  // This is currently unused... it's for a more advanced framerate limiter.
+  if (SK_ImGui_D3D11GPUEvent != nullptr)
+  {
+    SK_ComQIPtr <IDXGIDevice2>
+        pDXGIDev2 (pDev);
+    if (pDXGIDev2 != nullptr)
+        pDXGIDev2->EnqueueSetEvent (SK_ImGui_D3D11GPUEvent);
+  }
+#endif
 
 #define _SetupThreadContext()                                                   \
                              pTLS->imgui->drawing                      = FALSE; \
@@ -2336,8 +2344,16 @@ SK_ImGui_DrawD3D11 (IDXGISwapChain* This)
     }
   }
 
-  if (pDXGIDev2 != nullptr)
-      pDXGIDev2->EnqueueSetEvent (SK_ImGui_D3D11GPUEvent);
+#ifdef USE_D3D11_GPU_WAIT
+  // This is currently unused... it's for a more advanced framerate limiter.
+  if (SK_ImGui_D3D11GPUEvent != nullptr)
+  {
+    SK_ComQIPtr <IDXGIDevice2>
+        pDXGIDev2 (pDev);
+    if (pDXGIDev2 != nullptr)
+        pDXGIDev2->EnqueueSetEvent (SK_ImGui_D3D11GPUEvent);
+  }
+#endif
 }
 
 HRESULT

src/widgets/latency.cpp

@@ -880,8 +880,11 @@ SK_ImGui_DrawGraph_Latency (bool predraw)
 
   ImGui::EndGroup   ();
 
+  const auto frames_drawn =
+    SK_GetFramesDrawn ();
+
   if ( ReadULong64Acquire (&SK_Reflex_LastFrameMarked) <
-       ReadULong64Acquire (&SK_RenderBackend::frames_drawn) - 2 && !game_pace.wantPacing () )
+                               frames_drawn - 2 && !game_pace.wantPacing (frames_drawn) )
   {
     static bool bRTSS64 =
       SK_GetModuleHandle (

src/window.cpp

@@ -4688,9 +4688,12 @@ GetWindowInfo_Detour (HWND hwnd, PWINDOWINFO pwi)
       pwi->cyWindowBorders = 0;
     }
 
+    static HMODULE hModSystemDXGI =
+      LoadLibraryExW (L"dxgi.dll", nullptr, LOAD_LIBRARY_SEARCH_SYSTEM32);
+
     // DXGI checks on this during SwapChain creation...
     //   lie to DXGI if we have to, so that SwapChain creation succeeds.
-    if (StrStrIW (SK_GetCallerName ().c_str (), L"dxgi.dll"))
+    if (SK_IsCallingDLL (hModSystemDXGI))
     {
       pwi->dwExStyle &= ~WS_EX_TOPMOST;