[host] improve capture performance

Contrary to the MS documentation, benchmarking shows a substantial
increase in performance when releasing the captured frame as soon
as possible. This change makes it possible to achieve 60FPS at 4K
resolutions.

host/Capture/DXGI.cpp

@@ -435,11 +435,7 @@ void DXGI::DeInitialize()
     m_mfTransform->DeleteInputStream(0);
   }
 
-  if (m_releaseFrame)
+  ReleaseFrame();
-  {
-    m_releaseFrame = false;
-    m_dup->ReleaseFrame();
-  }
 
   if (m_pointer)
   {
@@ -611,26 +607,6 @@ GrabStatus Capture::DXGI::GrabFrameTexture(struct FrameInfo & frame, struct Curs
   {
     while (true)
     {
-      if (m_releaseFrame)
-      {
-        m_releaseFrame = false;
-        status = m_dup->ReleaseFrame();
-
-        switch (status)
-        {
-        case S_OK:
-          break;
-
-        case DXGI_ERROR_INVALID_CALL:
-          DEBUG_ERROR("Frame was already released");
-          return GRAB_STATUS_ERROR;
-
-        case DXGI_ERROR_ACCESS_LOST:
-          WaitForDesktop();
-          return GRAB_STATUS_REINIT;
-        }
-      }
-
       status = m_dup->AcquireNextFrame(1000, &frameInfo, &res);
       if (status == DXGI_ERROR_WAIT_TIMEOUT)
       {
@@ -745,6 +721,7 @@ GrabStatus Capture::DXGI::GrabFrameTexture(struct FrameInfo & frame, struct Curs
 
   if (!texture)
   {
+    ReleaseFrame();
     DEBUG_ERROR("Failed to get src ID3D11Texture2D");
     return GRAB_STATUS_ERROR;
   }
@@ -752,6 +729,30 @@ GrabStatus Capture::DXGI::GrabFrameTexture(struct FrameInfo & frame, struct Curs
   return GRAB_STATUS_OK;
 }
 
+GrabStatus Capture::DXGI::ReleaseFrame()
+{
+  if (!m_releaseFrame)
+    return GRAB_STATUS_OK;
+
+  m_releaseFrame = false;
+  switch (m_dup->ReleaseFrame())
+  {
+  case S_OK:
+    break;
+
+  case DXGI_ERROR_INVALID_CALL:
+    DEBUG_ERROR("Frame was already released");
+    return GRAB_STATUS_ERROR;
+
+  case WAIT_ABANDONED:
+  case DXGI_ERROR_ACCESS_LOST:
+    WaitForDesktop();
+    return GRAB_STATUS_REINIT;
+  }
+
+  return GRAB_STATUS_OK;
+}
+
 GrabStatus Capture::DXGI::GrabFrameRaw(FrameInfo & frame, struct CursorInfo & cursor)
 {
   GrabStatus result;
@@ -762,7 +763,10 @@ GrabStatus Capture::DXGI::GrabFrameRaw(FrameInfo & frame, struct CursorInfo & cu
   {
     result = GrabFrameTexture(frame, cursor, src, timeout);
     if (result != GRAB_STATUS_OK)
+    {
+      ReleaseFrame();
       return result;
+    }
 
     if (timeout)
     {
@@ -785,6 +789,10 @@ GrabStatus Capture::DXGI::GrabFrameRaw(FrameInfo & frame, struct CursorInfo & cu
   m_deviceContext->CopyResource(m_texture, src);
   SafeRelease(&src);
 
+  result = ReleaseFrame();
+  if (result != GRAB_STATUS_OK)
+    return result;
+
   if (m_surfaceMapped)
   {
     m_deviceContext->Unmap(m_texture, 0);
@@ -855,6 +863,7 @@ GrabStatus Capture::DXGI::GrabFrameH264(struct FrameInfo & frame, struct CursorI
         result = GrabFrameTexture(frame, cursor, src, timeout);
         if (result != GRAB_STATUS_OK)
         {
+          ReleaseFrame();
           return result;
         }
 
@@ -898,6 +907,7 @@ GrabStatus Capture::DXGI::GrabFrameH264(struct FrameInfo & frame, struct CursorI
       SafeRelease(&sample);
       SafeRelease(&buffer);
 
+      ReleaseFrame();
       EnterCriticalSection(&m_encodeCS);
     }
 

host/Capture/DXGI.h

@@ -117,6 +117,7 @@ namespace Capture
     bool InitH264Capture();
 
     GrabStatus GrabFrameTexture(struct FrameInfo & frame, struct CursorInfo & cursor, ID3D11Texture2DPtr & texture, bool & timeout);
+    GrabStatus ReleaseFrame();
     GrabStatus GrabFrameRaw    (struct FrameInfo & frame, struct CursorInfo & cursor);
     GrabStatus GrabFrameH264   (struct FrameInfo & frame, struct CursorInfo & cursor);
 

host/Service.cpp

@@ -157,6 +157,10 @@ void Service::DeInitialize()
   m_cursorDataSize = 0;
   m_frameSize      = 0;
 
+  WaitForSingleObject(m_cursorThread, INFINITE);
+  CloseHandle(m_cursorThread);
+  CloseHandle(m_cursorEvent);
+
   m_ivshmem->DeInitialize();
 
   if (m_capture)
@@ -165,10 +169,6 @@ void Service::DeInitialize()
     m_capture = NULL;
   }
 
-  WaitForSingleObject(m_cursorThread, INFINITE);
-  CloseHandle(m_cursorThread);
-  CloseHandle(m_cursorEvent );
-
   m_memory = NULL;
   m_initialized = false;
 }
@@ -213,10 +213,8 @@ bool Service::Process()
     Sleep(2);
   }
 
-  struct FrameInfo  frame;
+  struct FrameInfo  frame  = {0};
-  struct CursorInfo cursor;
+  struct CursorInfo cursor = {0};
-  ZeroMemory(&frame , sizeof(struct FrameInfo ));
-  ZeroMemory(&cursor, sizeof(struct CursorInfo));
   frame.buffer     = m_frame[m_frameIndex];
   frame.bufferSize = m_frameSize;
 
@@ -323,8 +321,6 @@ bool Service::Process()
 
   // update the flags
   INTERLOCKED_AND8((volatile char *)flags, KVMFR_HEADER_FLAG_RESTART);
-//  _mm_sfence();
-
   return true;
 }