[dxvk] Introduce sequence numbers for CS submissions

src/d3d11/d3d11_context_imm.cpp

@@ -548,8 +548,7 @@ namespace dxvk {
     // recorded prior to this function will be run
     FlushCsChunk();
     
-    if (m_csThread.isBusy())
-      m_csThread.synchronize();
+    m_csThread.synchronize(DxvkCsThread::SynchronizeAll);
   }
   
   

src/d3d9/d3d9_device.cpp

@@ -4759,8 +4759,7 @@ namespace dxvk {
     // recorded prior to this function will be run
     FlushCsChunk();
 
-    if (m_csThread.isBusy())
-      m_csThread.synchronize();
+    m_csThread.synchronize(DxvkCsThread::SynchronizeAll);
   }
 
 

src/dxvk/dxvk_cs.cpp

@@ -111,23 +111,33 @@ namespace dxvk {
   }
   
   
-  void DxvkCsThread::dispatchChunk(DxvkCsChunkRef&& chunk) {
+  uint64_t DxvkCsThread::dispatchChunk(DxvkCsChunkRef&& chunk) {
+    uint64_t seq;
+
     { std::unique_lock<dxvk::mutex> lock(m_mutex);
+      seq = ++m_chunksDispatched;
       m_chunksQueued.push(std::move(chunk));
-      m_chunksPending += 1;
     }
     
     m_condOnAdd.notify_one();
+    return seq;
   }
   
   
-  void DxvkCsThread::synchronize() {
+  void DxvkCsThread::synchronize(uint64_t seq) {
+    // Avoid locking if we know the sync is a no-op, may
+    // reduce overhead if this is being called frequently
+    if (seq > m_chunksExecuted.load(std::memory_order_acquire)) {
       std::unique_lock<dxvk::mutex> lock(m_mutex);
 
-    m_condOnSync.wait(lock, [this] {
-      return !m_chunksPending.load();
+      if (seq == SynchronizeAll)
+        seq = m_chunksDispatched.load();
+
+      m_condOnSync.wait(lock, [this, seq] {
+        return m_chunksExecuted.load() >= seq;
       });
     }
+  }
   
   
   void DxvkCsThread::threadFunc() {
@@ -139,7 +149,7 @@ namespace dxvk {
       while (!m_stopped.load()) {
         { std::unique_lock<dxvk::mutex> lock(m_mutex);
           if (chunk) {
-            if (--m_chunksPending == 0)
+            m_chunksExecuted++;
             m_condOnSync.notify_one();
             
             chunk = DxvkCsChunkRef();

src/dxvk/dxvk_cs.h

@@ -380,6 +380,8 @@ namespace dxvk {
     
   public:
 
+    constexpr static uint64_t SynchronizeAll = ~0ull;
+
     DxvkCsThread(const Rc<DxvkContext>& context);
     ~DxvkCsThread();
     
@@ -389,41 +391,43 @@ namespace dxvk {
      * Can be used to efficiently play back large
      * command lists recorded on another thread.
      * \param [in] chunk The chunk to dispatch
+     * \returns Sequence number of the submission
      */
-    void dispatchChunk(DxvkCsChunkRef&& chunk);
+    uint64_t dispatchChunk(DxvkCsChunkRef&& chunk);
     
     /**
      * \brief Synchronizes with the thread
      * 
-     * This waits for all chunks in the dispatch
-     * queue to be processed by the thread. Note
-     * that this does \e not implicitly call
-     * \ref flush.
+     * This waits for all chunks in the dispatch queue to
+     * be processed by the thread, up to the given sequence
+     * number. If the sequence number is 0, this will wait
+     * for all pending chunks to complete execution.
+     * \param [in] seq Sequence number to wait for.
      */
-    void synchronize();
+    void synchronize(uint64_t seq);
     
     /**
-     * \brief Checks whether the worker thread is busy
+     * \brief Retrieves last executed sequence number
      *
-     * Note that this information is only reliable if
-     * only the calling thread dispatches jobs to the
-     * worker queue and if the result is \c false.
-     * \returns \c true if there is still work to do
+     * Can be used to avoid synchronization in some cases.
+     * \returns Sequence number of last executed chunk
      */
-    bool isBusy() const {
-      return m_chunksPending.load() != 0;
+    uint64_t lastSequenceNumber() const {
+      return m_chunksExecuted.load();
     }
 
   private:
     
     const Rc<DxvkContext>       m_context;
 
+    std::atomic<uint64_t>       m_chunksDispatched = { 0ull };
+    std::atomic<uint64_t>       m_chunksExecuted   = { 0ull };
+    
     std::atomic<bool>           m_stopped = { false };
     dxvk::mutex                 m_mutex;
     dxvk::condition_variable    m_condOnAdd;
     dxvk::condition_variable    m_condOnSync;
     std::queue<DxvkCsChunkRef>  m_chunksQueued;
-    std::atomic<uint32_t>       m_chunksPending = { 0u };
     dxvk::thread                m_thread;
     
     void threadFunc();