[d3d11] Enabled command stream thread

src/d3d11/d3d11_context.cpp

@@ -14,7 +14,6 @@ namespace dxvk {
       Rc<DxvkDevice>  device)
   : m_parent  (parent),
     m_device  (device),
-    m_context (m_device->createContext()),
     m_csChunk (new DxvkCsChunk()) {
     // Create default state objects. We won't ever return them
     // to the application, but we'll use them to apply state.

src/d3d11/d3d11_context.h

@@ -518,7 +518,6 @@ namespace dxvk {
     D3D11Device* const m_parent;
     
     Rc<DxvkDevice>              m_device;
-    Rc<DxvkContext>             m_context;
     Rc<DxvkCsChunk>             m_csChunk;
     Rc<DxvkSampler>             m_defaultSampler;
     

src/d3d11/d3d11_context_imm.cpp

@@ -7,13 +7,16 @@ namespace dxvk {
   D3D11ImmediateContext::D3D11ImmediateContext(
     D3D11Device*    parent,
     Rc<DxvkDevice>  device)
-  : D3D11DeviceContext(parent, device) {
+  : D3D11DeviceContext(parent, device),
+    m_csThread(device->createContext()) {
     
   }
   
   
   D3D11ImmediateContext::~D3D11ImmediateContext() {
-    
+    Flush();
+    SynchronizeCs();
+    Synchronize();
   }
   
   
@@ -113,6 +116,7 @@ namespace dxvk {
             return DXGI_ERROR_WAS_STILL_DRAWING;
           
           Flush();
+          SynchronizeCs();
           Synchronize();
         }
       }
@@ -189,6 +193,7 @@ namespace dxvk {
         });
         
         Flush();
+        SynchronizeCs();
         Synchronize();
         
         physicalSlice = textureInfo->imageBuffer->slice();
@@ -247,16 +252,15 @@ namespace dxvk {
   
   
   void D3D11ImmediateContext::SynchronizeCs() {
-    // Dispatch recorded commands first,
     EmitCsChunk();
     
-    // TODO synchronize with CS thread
+    m_csThread.synchronize();
   }
   
   
   void D3D11ImmediateContext::EmitCsChunk() {
     if (m_csChunk->commandCount() > 0) {
-      m_csChunk->executeAll(m_context.ptr());
+      m_csThread.dispatchChunk(std::move(m_csChunk));
       m_csChunk = new DxvkCsChunk();
     }
   }

src/d3d11/d3d11_context_imm.h

@@ -45,6 +45,8 @@ namespace dxvk {
     
   private:
     
+    DxvkCsThread m_csThread;
+    
     void Synchronize();
     void SynchronizeCs();
     

src/dxvk/dxvk_cs.cpp

@@ -43,6 +43,7 @@ namespace dxvk {
   void DxvkCsThread::dispatchChunk(Rc<DxvkCsChunk>&& chunk) {
     { std::unique_lock<std::mutex> lock(m_mutex);
       m_chunks.push(std::move(chunk));
+      m_chunksPending += 1;
     }
     
     m_condOnAdd.notify_one();
@@ -53,7 +54,7 @@ namespace dxvk {
     std::unique_lock<std::mutex> lock(m_mutex);
     
     m_condOnSync.wait(lock, [this] {
-      return m_chunks.size() == 0;
+      return m_chunksPending == 0;
     });
   }
   
@@ -71,14 +72,20 @@ namespace dxvk {
         if (m_chunks.size() != 0) {
           chunk = std::move(m_chunks.front());
           m_chunks.pop();
-          
-          if (m_chunks.size() == 0)
-            m_condOnSync.notify_one();
         }
       }
       
-      if (chunk != nullptr)
+      if (chunk != nullptr) {
         chunk->executeAll(m_context.ptr());
+        
+        const bool doNotify = [this] {
+          std::unique_lock<std::mutex> lock(m_mutex);
+          return --m_chunksPending == 0;
+        }();
+        
+        if (doNotify)
+          m_condOnSync.notify_one();
+      }
     }
   }
   

src/dxvk/dxvk_cs.h

@@ -174,6 +174,8 @@ namespace dxvk {
     std::queue<Rc<DxvkCsChunk>> m_chunks;
     std::thread                 m_thread;
     
+    uint32_t                    m_chunksPending = 0;
+    
     void threadFunc();
     
   };