[util] Reimplement Signal

The new implementation is more akin to D3D12 fences or timeline
semaphores, and should make implementing similar concepts easier.

src/d3d11/d3d11_swapchain.cpp

@@ -26,7 +26,7 @@ namespace dxvk {
     m_device    (pDevice->GetDXVKDevice()),
     m_context   (m_device->createContext()),
     m_frameLatencyCap(pDevice->GetOptions()->maxFrameLatency) {
-    CreateFrameLatencySignals();
+    CreateFrameLatencySignal();
 
     if (!pDevice->GetOptions()->deferSurfaceCreation)
       CreatePresenter();
@@ -209,9 +209,8 @@ namespace dxvk {
     immediateContext->Flush();
 
     // Wait for the sync event so that we respect the maximum frame latency
-    uint32_t frameId = m_frameId++ % GetActualFrameLatency();
+    uint64_t frameId = ++m_frameId;
-    auto syncEvent = m_frameLatencySignals[frameId];
+    m_frameLatencySignal->wait(frameId - GetActualFrameLatency());
-    syncEvent->wait();
     
     if (m_hud != nullptr)
       m_hud->update();
@@ -312,7 +311,7 @@ namespace dxvk {
         m_hud->render(m_context, info.imageExtent);
       
       if (i + 1 >= SyncInterval)
-        m_context->queueSignal(syncEvent);
+        m_context->signal(m_frameLatencySignal, frameId);
 
       SubmitPresent(immediateContext, sync);
     }
@@ -370,9 +369,8 @@ namespace dxvk {
   }
 
 
-  void D3D11SwapChain::CreateFrameLatencySignals() {
+  void D3D11SwapChain::CreateFrameLatencySignal() {
-    for (uint32_t i = 0; i < m_frameLatencySignals.size(); i++)
+    m_frameLatencySignal = new sync::Fence(m_frameId);
-      m_frameLatencySignals[i] = new sync::Signal(true);
   }
 
 

src/d3d11/d3d11_swapchain.h

@@ -117,9 +117,9 @@ namespace dxvk {
 
     std::vector<Rc<DxvkImageView>> m_imageViews;
 
-    uint32_t                                                  m_frameId = 0;
+    uint64_t                m_frameId = DXGI_MAX_SWAP_CHAIN_BUFFERS;
-    std::array<Rc<sync::Signal>, DXGI_MAX_SWAP_CHAIN_BUFFERS> m_frameLatencySignals;
     uint32_t                m_frameLatencyCap = 0;
+    Rc<sync::Signal>        m_frameLatencySignal;
 
     bool                    m_dirty = true;
     bool                    m_vsync = true;
@@ -135,7 +135,7 @@ namespace dxvk {
     void RecreateSwapChain(
             BOOL                      Vsync);
 
-    void CreateFrameLatencySignals();
+    void CreateFrameLatencySignal();
 
     void CreatePresenter();
 

src/dxvk/dxvk_cmdlist.h

@@ -186,9 +186,11 @@ namespace dxvk {
      * 
      * The signal will be notified once the command
      * buffer has finished executing on the GPU.
+     * \param [in] signal The signal
+     * \param [in] value Signal value
      */
-    void queueSignal(const Rc<sync::Signal>& signal) {
+    void queueSignal(const Rc<sync::Signal>& signal, uint64_t value) {
-      m_signalTracker.add(signal);
+      m_signalTracker.add(signal, value);
     }
 
     /**

src/dxvk/dxvk_context.cpp

@@ -2375,8 +2375,8 @@ namespace dxvk {
   }
 
 
-  void DxvkContext::queueSignal(const Rc<sync::Signal>& signal) {
+  void DxvkContext::signal(const Rc<sync::Signal>& signal, uint64_t value) {
-    m_cmd->queueSignal(signal);
+    m_cmd->queueSignal(signal, value);
   }
 
 

src/dxvk/dxvk_context.h

@@ -992,9 +992,11 @@ namespace dxvk {
      * previously submitted commands have
      * finished execution on the GPU.
      * \param [in] signal The signal
+     * \param [in] value Signal value
      */
-    void queueSignal(
+    void signal(
-      const Rc<sync::Signal>&   signal);
+      const Rc<sync::Signal>&   signal,
+            uint64_t            value);
     
     /**
      * \brief Trims staging buffers

src/dxvk/dxvk_signal.cpp

@@ -12,14 +12,14 @@ namespace dxvk {
   }
     
   
-  void DxvkSignalTracker::add(const Rc<sync::Signal>& signal) {
+  void DxvkSignalTracker::add(const Rc<sync::Signal>& signal, uint64_t value) {
-    m_signals.push_back(signal);
+    m_signals.push_back({ signal, value });
   }
 
 
   void DxvkSignalTracker::notify() {
-    for (const auto& sig : m_signals)
+    for (const auto& pair : m_signals)
-      sig->notify();
+      pair.first->signal(pair.second);
   }
 
 

src/dxvk/dxvk_signal.h

@@ -16,9 +16,11 @@ namespace dxvk {
     
     /**
      * \brief Adds a signal to track
+     *
      * \param [in] signal The signal
+     * \param [in] value Target value
      */
-    void add(const Rc<sync::Signal>& signal);
+    void add(const Rc<sync::Signal>& signal, uint64_t value);
     
     /**
      * \brief Notifies tracked signals
@@ -32,7 +34,7 @@ namespace dxvk {
     
   private:
     
-    std::vector<Rc<sync::Signal>> m_signals;
+    std::vector<std::pair<Rc<sync::Signal>, uint64_t>> m_signals;
     
   };
   

src/util/sync/sync_signal.h

@@ -9,52 +9,80 @@ namespace dxvk::sync {
   /**
    * \brief Signal
    * 
-   * Acts as a simple CPU fence which can be signaled by one
+   * Interface for a CPU-side fence. Can be signaled
-   * thread and waited upon by one more thread. Waiting on
+   * to a given value, and any thread waiting for a
-   * more than one thread is not supported.
+   * lower value will be woken up.
    */
   class Signal : public RcObject {
     
   public:
     
-    Signal()
+    virtual ~Signal() { }
-    : m_signaled(false) { }
-    Signal(bool signaled)
-    : m_signaled(signaled) { }
-    ~Signal() { }
 
-    Signal             (const Signal&) = delete;
+    /**
-    Signal& operator = (const Signal&) = delete;
+     * \brief Last signaled value
+     * \returns Last signaled value
+     */
+    virtual uint64_t value() const = 0;
     
     /**
      * \brief Notifies signal
-     * Wakes any waiting thread.
+     *
+     * Wakes up all threads currently waiting for
+     * a value lower than \c value. Note that
+     * \c value must monotonically increase.
+     * \param [in] value Value to signal to
      */
-    void notify() {
+    virtual void signal(uint64_t value) = 0;
-      std::lock_guard<std::mutex> lock(m_mutex);
-      m_signaled.store(true);
-      m_cond.notify_one();
-    }
     
     /**
      * \brief Waits for signal
      * 
      * Blocks the calling thread until another
-     * thread wakes it up, then resets it to
+     * thread signals it with a value equal to
-     * the non-signaled state.
+     * or greater than \c value.
+     * \param [in] value The value to wait for
      */
-    void wait() {
+    virtual void wait(uint64_t value) = 0;
-      if (!m_signaled.exchange(false)) {
+    
-        std::unique_lock<std::mutex> lock(m_mutex);
+  };
-        m_cond.wait(lock, [this] {
+
-          return m_signaled.exchange(false);
+
-        });
+  /**
+   * \brief Fence
+   *
+   * Simple CPU-side fence.
+   */
+  class Fence : public Signal {
+
+  public:
+
+    Fence()
+    : m_value(0ull) { }
+
+    explicit Fence(uint64_t value)
+    : m_value(value) { }
+
+    uint64_t value() const {
+      return m_value.load(std::memory_order_acquire);  
     }
+
+    void signal(uint64_t value) {
+      std::unique_lock<std::mutex> lock(m_mutex);
+      m_value.store(value, std::memory_order_release);
+      m_cond.notify_all();
+    }
+    
+    void wait(uint64_t value) {
+      std::unique_lock<std::mutex> lock(m_mutex);
+      m_cond.wait(lock, [this, value] {
+        return value <= m_value.load(std::memory_order_acquire);
+      });
     }
 
   private:
 
-    std::atomic<bool>       m_signaled;
+    std::atomic<uint64_t>   m_value;
     std::mutex              m_mutex;
     std::condition_variable m_cond;