[d3d11] Implemented CS chunk submission for deferred contexts

Changes were necessary due to the fact that chunks in deferred
contexts are not directly submitted to the CS thread.

src/d3d11/d3d11_cmdlist.cpp

@@ -34,4 +34,23 @@ namespace dxvk {
     return m_contextFlags;
   }
   
+  
+  void D3D11CommandList::AddChunk(Rc<DxvkCsChunk>&& Chunk) {
+    m_chunks.push_back(std::move(Chunk));
+  }
+  
+  
+  void D3D11CommandList::EmitToCommandList(ID3D11CommandList* pCommandList) {
+    auto cmdList = static_cast<D3D11CommandList*>(pCommandList);
+    
+    for (auto chunk : m_chunks)
+      cmdList->m_chunks.push_back(chunk);
+  }
+  
+  
+  void D3D11CommandList::EmitToCsThread(const Rc<DxvkCsThread>& CsThread) {
+    for (auto chunk : m_chunks)
+      CsThread->dispatchChunk(Rc<DxvkCsChunk>(chunk));
+  }
+  
 }

src/d3d11/d3d11_cmdlist.h

@@ -1,11 +1,9 @@
 #pragma once
 
-#include "d3d11_device_child.h"
+#include "d3d11_context.h"
 
 namespace dxvk {
   
-  class D3D11Device;
-  
   class D3D11CommandList : public D3D11DeviceChild<ID3D11CommandList> {
     
   public:
@@ -25,11 +23,21 @@ namespace dxvk {
     
     UINT GetContextFlags();
     
+    void AddChunk(Rc<DxvkCsChunk>&& Chunk);
+    
+    void EmitToCommandList(
+            ID3D11CommandList*  pCommandList);
+    
+    void EmitToCsThread(
+      const Rc<DxvkCsThread>&   CsThread);
+    
   private:
     
     D3D11Device* const m_device;
     UINT         const m_contextFlags;
     
+    std::vector<Rc<DxvkCsChunk>> m_chunks;
+    
   };
   
 }

src/d3d11/d3d11_context.h

@@ -568,12 +568,21 @@ namespace dxvk {
     template<typename Cmd>
     void EmitCs(Cmd&& command) {
       if (!m_csChunk->push(command)) {
-        EmitCsChunk();
+        EmitCsChunk(std::move(m_csChunk));
+        
+        m_csChunk = new DxvkCsChunk();
         m_csChunk->push(command);
       }
     }
     
-    virtual void EmitCsChunk() = 0;
+    void FlushCsChunk() {
+      if (m_csChunk->commandCount() != 0) {
+        EmitCsChunk(std::move(m_csChunk));
+        m_csChunk = new DxvkCsChunk();
+      }
+    }
+    
+    virtual void EmitCsChunk(Rc<DxvkCsChunk>&& chunk) = 0;
     
   };
   

src/d3d11/d3d11_context_def.cpp

@@ -7,7 +7,8 @@ namespace dxvk {
     Rc<DxvkDevice>  Device,
     UINT            ContextFlags)
   : D3D11DeviceContext(pParent, Device),
-    m_contextFlags(ContextFlags) {
+    m_contextFlags(ContextFlags),
+    m_commandList (CreateCommandList()) {
     
   }
   
@@ -60,8 +61,13 @@ namespace dxvk {
   }
   
   
-  void D3D11DeferredContext::EmitCsChunk() {
+  Com<D3D11CommandList> D3D11DeferredContext::CreateCommandList() {
-    
+    return new D3D11CommandList(m_parent, m_contextFlags);
+  }
+  
+  
+  void D3D11DeferredContext::EmitCsChunk(Rc<DxvkCsChunk>&& chunk) {
+    m_commandList->AddChunk(std::move(chunk));
   }
 
 }

src/d3d11/d3d11_context_def.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include "d3d11_cmdlist.h"
 #include "d3d11_context.h"
 
 namespace dxvk {
@@ -42,7 +43,11 @@ namespace dxvk {
     
     const UINT m_contextFlags;
     
-    void EmitCsChunk() final;
+    Com<D3D11CommandList> m_commandList;
+    
+    Com<D3D11CommandList> CreateCommandList();
+    
+    void EmitCsChunk(Rc<DxvkCsChunk>&& chunk) final;
     
   };
   

src/d3d11/d3d11_context_imm.cpp

@@ -55,7 +55,7 @@ namespace dxvk {
         dev->createCommandList());
     });
     
-    EmitCsChunk();
+    FlushCsChunk();
   }
   
   
@@ -251,7 +251,7 @@ namespace dxvk {
   void D3D11ImmediateContext::SynchronizeCsThread() {
     // Dispatch current chunk so that all commands
     // recorded prior to this function will be run
-    EmitCsChunk();
+    FlushCsChunk();
     
     m_csThread.synchronize();
   }
@@ -266,9 +266,8 @@ namespace dxvk {
   }
   
   
-  void D3D11ImmediateContext::EmitCsChunk() {
+  void D3D11ImmediateContext::EmitCsChunk(Rc<DxvkCsChunk>&& chunk) {
-    if (m_csChunk->commandCount() > 0)
+    m_csThread.dispatchChunk(std::move(chunk));
-      m_csChunk = m_csThread.dispatchChunk(std::move(m_csChunk));
   }
   
 }

src/d3d11/d3d11_context_imm.h

@@ -50,7 +50,7 @@ namespace dxvk {
     
     void SynchronizeDevice();
     
-    void EmitCsChunk() final;
+    void EmitCsChunk(Rc<DxvkCsChunk>&& chunk) final;
     
   };
   

src/dxvk/dxvk_cs.cpp

@@ -40,9 +40,7 @@ namespace dxvk {
   }
   
   
-  Rc<DxvkCsChunk> DxvkCsThread::dispatchChunk(Rc<DxvkCsChunk>&& chunk) {
+  void DxvkCsThread::dispatchChunk(Rc<DxvkCsChunk>&& chunk) {
-    Rc<DxvkCsChunk> nextChunk = nullptr;
-    
     { std::unique_lock<std::mutex> lock(m_mutex);
       m_chunksQueued.push(std::move(chunk));
       m_chunksPending += 1;
@@ -56,21 +54,10 @@ namespace dxvk {
               || (m_stopped.load());
         });
       }
-      
-      if (m_chunksUnused.size() != 0) {
-        nextChunk = std::move(m_chunksUnused.front());
-        m_chunksUnused.pop();
-      }
     }
     
     // Wake CS thread
     m_condOnAdd.notify_one();
-    
-    // Allocate new chunk if needed
-    if (nextChunk == nullptr)
-      nextChunk = new DxvkCsChunk();
-    
-    return nextChunk;
   }
   
   
@@ -90,8 +77,6 @@ namespace dxvk {
       { std::unique_lock<std::mutex> lock(m_mutex);
         if (chunk != nullptr) {
           m_chunksPending -= 1;
-          m_chunksUnused.push(std::move(chunk));
-          
           m_condOnSync.notify_one();
         }
         

src/dxvk/dxvk_cs.h

@@ -152,9 +152,8 @@ namespace dxvk {
      * Can be used to efficiently play back large
      * command lists recorded on another thread.
      * \param [in] chunk The chunk to dispatch
-     * \returns New chunk for the next submissions
      */
-    Rc<DxvkCsChunk> dispatchChunk(Rc<DxvkCsChunk>&& chunk);
+    void dispatchChunk(Rc<DxvkCsChunk>&& chunk);
     
     /**
      * \brief Synchronizes with the thread
@@ -175,7 +174,6 @@ namespace dxvk {
     std::condition_variable     m_condOnAdd;
     std::condition_variable     m_condOnSync;
     std::queue<Rc<DxvkCsChunk>> m_chunksQueued;
-    std::queue<Rc<DxvkCsChunk>> m_chunksUnused;
     std::thread                 m_thread;
     
     uint32_t                    m_chunksPending = 0;