[dxvk] Recycle CS chunks

src/d3d11/d3d11_context_imm.cpp

@@ -266,10 +266,8 @@ namespace dxvk {
   
   
   void D3D11ImmediateContext::EmitCsChunk() {
-    if (m_csChunk->commandCount() > 0) {
-      m_csThread.dispatchChunk(std::move(m_csChunk));
-      m_csChunk = new DxvkCsChunk();
-    }
+    if (m_csChunk->commandCount() > 0)
+      m_csChunk = m_csThread.dispatchChunk(std::move(m_csChunk));
   }
   
 }

src/dxvk/dxvk_cs.cpp

@@ -40,17 +40,37 @@ namespace dxvk {
   }
   
   
-  void DxvkCsThread::dispatchChunk(Rc<DxvkCsChunk>&& chunk) {
+  Rc<DxvkCsChunk> DxvkCsThread::dispatchChunk(Rc<DxvkCsChunk>&& chunk) {
+    Rc<DxvkCsChunk> nextChunk = nullptr;
+    
     { std::unique_lock<std::mutex> lock(m_mutex);
-      m_chunks.push(std::move(chunk));
+      m_chunksQueued.push(std::move(chunk));
       m_chunksPending += 1;
       
-      m_condOnSync.wait(lock, [this] {
-        return m_stopped.load() || (m_chunksPending < MaxChunksInFlight);
-      });
+      // If a large number of chunks are queued up, wait for
+      // some of them to be processed in order to avoid memory
+      // leaks, stuttering, input lag and similar issues.
+      if (m_chunksPending >= MaxChunksInFlight) {
+        m_condOnSync.wait(lock, [this] {
+          return (m_chunksPending < MaxChunksInFlight / 2)
+              || (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;
   }
   
   
@@ -64,30 +84,31 @@ namespace dxvk {
   
   
   void DxvkCsThread::threadFunc() {
+    Rc<DxvkCsChunk> chunk;
+    
     while (!m_stopped.load()) {
-      Rc<DxvkCsChunk> chunk;
-      
       { std::unique_lock<std::mutex> lock(m_mutex);
+        if (chunk != nullptr) {
+          m_chunksPending -= 1;
+          m_chunksUnused.push(std::move(chunk));
+          
+          m_condOnSync.notify_one();
+        }
         
         m_condOnAdd.wait(lock, [this] {
-          return m_stopped.load() || (m_chunks.size() != 0);
+          return m_stopped.load() || (m_chunksQueued.size() != 0);
         });
         
-        if (m_chunks.size() != 0) {
-          chunk = std::move(m_chunks.front());
-          m_chunks.pop();
+        if (m_chunksQueued.size() != 0) {
+          chunk = std::move(m_chunksQueued.front());
+          m_chunksQueued.pop();
+        } else {
+          chunk = nullptr;
         }
       }
       
-      if (chunk != nullptr) {
+      if (chunk != nullptr)
         chunk->executeAll(m_context.ptr());
-        
-        { std::unique_lock<std::mutex> lock(m_mutex);
-          m_chunksPending -= 1;
-        }
-        
-        m_condOnSync.notify_one();
-      }
     }
   }
   

src/dxvk/dxvk_cs.h

@@ -65,7 +65,7 @@ namespace dxvk {
    * Stores a list of commands.
    */
   class DxvkCsChunk : public RcObject {
-    constexpr static size_t MaxCommands  = 64;
+    constexpr static size_t MaxCommands  = 1024;
     constexpr static size_t MaxBlockSize = 64 * MaxCommands;
   public:
     
@@ -140,7 +140,7 @@ namespace dxvk {
   class DxvkCsThread {
     // Limit the number of chunks in the queue
     // to prevent memory leaks, stuttering etc.
-    constexpr static uint32_t MaxChunksInFlight = 128;
+    constexpr static uint32_t MaxChunksInFlight = 16;
   public:
     
     DxvkCsThread(const Rc<DxvkContext>& context);
@@ -152,8 +152,9 @@ 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
      */
-    void dispatchChunk(Rc<DxvkCsChunk>&& chunk);
+    Rc<DxvkCsChunk> dispatchChunk(Rc<DxvkCsChunk>&& chunk);
     
     /**
      * \brief Synchronizes with the thread
@@ -173,7 +174,8 @@ namespace dxvk {
     std::mutex                  m_mutex;
     std::condition_variable     m_condOnAdd;
     std::condition_variable     m_condOnSync;
-    std::queue<Rc<DxvkCsChunk>> m_chunks;
+    std::queue<Rc<DxvkCsChunk>> m_chunksQueued;
+    std::queue<Rc<DxvkCsChunk>> m_chunksUnused;
     std::thread                 m_thread;
     
     uint32_t                    m_chunksPending = 0;