[dxvk] Off-load command buffer submission to separate thread

Reduces load on the CS thread a bit, which may yield a small
performance improvement.

src/dxvk/dxvk_device.cpp

@@ -262,8 +262,10 @@ namespace dxvk {
   VkResult DxvkDevice::presentImage(
     const Rc<vk::Presenter>&        presenter,
           VkSemaphore               semaphore) {
-    std::lock_guard<std::mutex> queueLock(m_submissionLock);
-    VkResult status = presenter->presentImage(semaphore);
+    DxvkPresentInfo presentInfo;
+    presentInfo.presenter = presenter;
+    presentInfo.waitSync  = semaphore;
+    VkResult status = m_submissionQueue.present(presentInfo);
 
     if (status != VK_SUCCESS)
       return status;
@@ -278,32 +280,22 @@ namespace dxvk {
     const Rc<DxvkCommandList>&      commandList,
           VkSemaphore               waitSync,
           VkSemaphore               wakeSync) {
-    VkResult status;
-    
-    { // Queue submissions are not thread safe
-      std::lock_guard<std::mutex> queueLock(m_submissionLock);
-      std::lock_guard<sync::Spinlock> statLock(m_statLock);
-      
-      m_statCounters.merge(commandList->statCounters());
-      m_statCounters.addCtr(DxvkStatCounter::QueueSubmitCount, 1);
-      
-      status = commandList->submit(
-        m_graphicsQueue.queueHandle,
-        waitSync, wakeSync);
-    }
-    
-    if (status == VK_SUCCESS) {
-      // Add this to the set of running submissions
-      m_submissionQueue.submit(commandList);
-    } else {
-      Logger::err(str::format(
-        "DxvkDevice: Command buffer submission failed: ",
-        status));
-    }
+    DxvkSubmitInfo submitInfo;
+    submitInfo.cmdList  = commandList;
+    submitInfo.queue    = m_graphicsQueue.queueHandle;
+    submitInfo.waitSync = waitSync;
+    submitInfo.wakeSync = wakeSync;
+    m_submissionQueue.submit(submitInfo);
+
+    std::lock_guard<sync::Spinlock> statLock(m_statLock);
+    m_statCounters.merge(commandList->statCounters());
+    m_statCounters.addCtr(DxvkStatCounter::QueueSubmitCount, 1);
   }
   
   
   void DxvkDevice::waitForIdle() {
+    m_submissionQueue.synchronize();
+
     if (m_vkd->vkDeviceWaitIdle(m_vkd->device()) != VK_SUCCESS)
       Logger::err("DxvkDevice: waitForIdle: Operation failed");
   }

src/dxvk/dxvk_device.h

@@ -339,6 +339,7 @@ namespace dxvk {
      * presenter's \c presentImage method.
      * \param [in] presenter The presenter
      * \param [in] semaphore Sync semaphore
+     * \returns Status of the operation
      */
     VkResult presentImage(
       const Rc<vk::Presenter>&        presenter,
@@ -347,11 +348,11 @@ namespace dxvk {
     /**
      * \brief Submits a command list
      * 
-     * Synchronization arguments are optional. 
+     * Submits the given command list to the device using
+     * the given set of optional synchronization primitives.
      * \param [in] commandList The command list to submit
      * \param [in] waitSync (Optional) Semaphore to wait on
      * \param [in] wakeSync (Optional) Semaphore to notify
-     * \returns Synchronization fence
      */
     void submitCommandList(
       const Rc<DxvkCommandList>&      commandList,
@@ -366,7 +367,8 @@ namespace dxvk {
      * to lock the queue before submitting command buffers.
      */
     void lockSubmission() {
-      m_submissionLock.lock();
+      m_submissionQueue.synchronize();
+      m_submissionQueue.lockDeviceQueue();
     }
     
     /**
@@ -376,7 +378,7 @@ namespace dxvk {
      * itself can use them for submissions again.
      */
     void unlockSubmission() {
-      m_submissionLock.unlock();
+      m_submissionQueue.unlockDeviceQueue();
     }
 
     /**
@@ -430,7 +432,6 @@ namespace dxvk {
     sync::Spinlock              m_statLock;
     DxvkStatCounters            m_statCounters;
     
-    std::mutex                  m_submissionLock;
     DxvkDeviceQueue             m_graphicsQueue;
     DxvkDeviceQueue             m_presentQueue;
     

src/dxvk/dxvk_queue.cpp

@@ -5,7 +5,8 @@ namespace dxvk {
   
   DxvkSubmissionQueue::DxvkSubmissionQueue(DxvkDevice* device)
   : m_device(device),
-    m_thread([this] () { threadFunc(); }) {
+    m_submitThread([this] () { submitCmdLists(); }),
+    m_finishThread([this] () { finishCmdLists(); }) {
     
   }
   
@@ -15,61 +16,132 @@ namespace dxvk {
       m_stopped.store(true);
     }
     
-    m_condOnAdd.notify_one();
-    m_thread.join();
+    m_appendCond.notify_all();
+    m_submitCond.notify_all();
+
+    m_submitThread.join();
+    m_finishThread.join();
   }
   
   
-  void DxvkSubmissionQueue::submit(const Rc<DxvkCommandList>& cmdList) {
-    { std::unique_lock<std::mutex> lock(m_mutex);
-      
-      m_condOnTake.wait(lock, [this] {
-        return m_entries.size() < MaxNumQueuedCommandBuffers;
+  void DxvkSubmissionQueue::submit(DxvkSubmitInfo submitInfo) {
+    std::unique_lock<std::mutex> lock(m_mutex);
+
+    m_finishCond.wait(lock, [this] {
+      return m_submitQueue.size() + m_finishQueue.size() <= MaxNumQueuedCommandBuffers;
+    });
+
+    m_pending += 1;
+    m_submitQueue.push(std::move(submitInfo));
+    m_appendCond.notify_all();
+  }
+
+
+  VkResult DxvkSubmissionQueue::present(DxvkPresentInfo presentInfo) {
+    this->synchronize();
+    
+    std::unique_lock<std::mutex> lock(m_mutexQueue);
+    return presentInfo.presenter->presentImage(presentInfo.waitSync);
+  }
+
+
+  void DxvkSubmissionQueue::synchronize() {
+    std::unique_lock<std::mutex> lock(m_mutex);
+
+    m_submitCond.wait(lock, [this] {
+      return m_submitQueue.empty();
+    });
+  }
+
+
+  void DxvkSubmissionQueue::lockDeviceQueue() {
+    m_mutexQueue.lock();
+  }
+
+
+  void DxvkSubmissionQueue::unlockDeviceQueue() {
+    m_mutexQueue.unlock();
+  }
+
+
+  void DxvkSubmissionQueue::submitCmdLists() {
+    env::setThreadName("dxvk-submit");
+
+    std::unique_lock<std::mutex> lock(m_mutex);
+
+    while (!m_stopped.load()) {
+      m_appendCond.wait(lock, [this] {
+        return m_stopped.load() || !m_submitQueue.empty();
       });
       
-      m_submits += 1;
-      m_entries.push(cmdList);
-      m_condOnAdd.notify_one();
+      if (m_stopped.load())
+        return;
+      
+      DxvkSubmitInfo submitInfo = std::move(m_submitQueue.front());
+      lock.unlock();
+
+      // Submit command buffer to device
+      VkResult status;
+
+      { std::lock_guard<std::mutex> lock(m_mutexQueue);
+
+        status = submitInfo.cmdList->submit(
+          submitInfo.queue,
+          submitInfo.waitSync,
+          submitInfo.wakeSync);
+      }
+      
+      // On success, pass it on to the queue thread
+      lock = std::unique_lock<std::mutex>(m_mutex);
+
+      if (status == VK_SUCCESS) {
+        m_finishQueue.push(std::move(submitInfo));
+        m_submitQueue.pop();
+        m_submitCond.notify_all();
+      } else {
+        Logger::err(str::format(
+          "DxvkSubmissionQueue: Command submission failed with ",
+          status));
+        m_pending -= 1;
+      }
     }
   }
   
   
-  void DxvkSubmissionQueue::threadFunc() {
+  void DxvkSubmissionQueue::finishCmdLists() {
     env::setThreadName("dxvk-queue");
 
+    std::unique_lock<std::mutex> lock(m_mutex);
+
     while (!m_stopped.load()) {
-      Rc<DxvkCommandList> cmdList;
+      m_submitCond.wait(lock, [this] {
+        return m_stopped.load() || !m_finishQueue.empty();
+      });
+
+      if (m_stopped.load())
+        return;
       
-      { std::unique_lock<std::mutex> lock(m_mutex);
-        
-        m_condOnAdd.wait(lock, [this] {
-          return m_stopped.load() || (m_entries.size() != 0);
-        });
-        
-        if (m_entries.size() != 0) {
-          cmdList = std::move(m_entries.front());
-          m_entries.pop();
-        }
-        
-        m_condOnTake.notify_one();
-      }
+      DxvkSubmitInfo submitInfo = std::move(m_finishQueue.front());
+      lock.unlock();
       
-      if (cmdList != nullptr) {
-        VkResult status = cmdList->synchronize();
+      VkResult status = submitInfo.cmdList->synchronize();
+      
+      if (status == VK_SUCCESS) {
+        submitInfo.cmdList->signalEvents();
+        submitInfo.cmdList->reset();
         
-        if (status == VK_SUCCESS) {
-          cmdList->signalEvents();
-          cmdList->reset();
-          
-          m_device->recycleCommandList(cmdList);
-        } else {
-          Logger::err(str::format(
-            "DxvkSubmissionQueue: Failed to sync fence: ",
-            status));
-        }
-        
-        m_submits -= 1;
+        m_device->recycleCommandList(submitInfo.cmdList);
+      } else {
+        Logger::err(str::format(
+          "DxvkSubmissionQueue: Failed to sync fence: ",
+          status));
       }
+
+      lock = std::unique_lock<std::mutex>(m_mutex);
+      m_pending -= 1;
+
+      m_finishQueue.pop();
+      m_finishCond.notify_all();
     }
   }
   

src/dxvk/dxvk_queue.h

@@ -6,17 +6,45 @@
 
 #include "../util/thread.h"
 
+#include "../vulkan/vulkan_presenter.h"
+
 #include "dxvk_cmdlist.h"
 
 namespace dxvk {
   
   class DxvkDevice;
+
+  /**
+   * \brief Queue submission info
+   * 
+   * Stores parameters used to submit
+   * a command buffer to the device.
+   */
+  struct DxvkSubmitInfo {
+    Rc<DxvkCommandList> cmdList;
+    VkQueue             queue;
+    VkSemaphore         waitSync;
+    VkSemaphore         wakeSync;
+  };
   
+  
+  /**
+   * \brief Present info
+   *
+   * Stores parameters used to present
+   * a swap chain image on the device.
+   */
+  struct DxvkPresentInfo {
+    Rc<vk::Presenter>   presenter;
+    VkSemaphore         waitSync;
+  };
+
+
   /**
    * \brief Submission queue
    */
   class DxvkSubmissionQueue {
-    
+
   public:
     
     DxvkSubmissionQueue(DxvkDevice* device);
@@ -30,35 +58,84 @@ namespace dxvk {
      * \returns Pending submission count
      */
     uint32_t pendingSubmissions() const {
-      return m_submits.load();
+      return m_pending.load();
     }
     
     /**
-     * \brief Submits a command list
+     * \brief Submits a command list asynchronously
      * 
-     * Submits a command list to the queue thread.
-     * This thread will wait for the command list
-     * to finish executing on the GPU and signal
-     * any queries and events that are used by
-     * the command list in question.
-     * \param [in] cmdList The command list
+     * Queues a command list for submission on the
+     * dedicated submission thread. Use this to take
+     * the submission overhead off the calling thread.
+     * \param [in] submitInfo Submission parameters 
      */
-    void submit(const Rc<DxvkCommandList>& cmdList);
+    void submit(
+            DxvkSubmitInfo  submitInfo);
+    
+    /**
+     * \brief Presents an image synchronously
+     *
+     * Waits for queued command lists to be submitted
+     * and then presents the current swap chain image
+     * of the presenter. May stall the calling thread.
+     * \param [in] present Present parameters
+     * \returns Status of the operation
+     */
+    VkResult present(
+            DxvkPresentInfo present);
+    
+    /**
+     * \brief Synchronizes with queue submissions
+     * 
+     * Waits for all pending command lists to be
+     * submitted to the GPU before returning.
+     */
+    void synchronize();
+
+    /**
+     * \brief Locks device queue
+     *
+     * Locks the mutex that protects the Vulkan queue
+     * that DXVK uses for command buffer submission.
+     * This is needed when the app submits its own
+     * command buffers to the queue.
+     */
+    void lockDeviceQueue();
+
+    /**
+     * \brief Unlocks device queue
+     *
+     * Unlocks the mutex that protects the Vulkan
+     * queue used for command buffer submission.
+     */
+    void unlockDeviceQueue();
     
   private:
-    
+
     DxvkDevice*             m_device;
     
     std::atomic<bool>       m_stopped = { false };
-    std::atomic<uint32_t>   m_submits = { 0u };
-    
+    std::atomic<uint32_t>   m_pending = { 0u };
+
     std::mutex              m_mutex;
-    std::condition_variable m_condOnAdd;
-    std::condition_variable m_condOnTake;
-    std::queue<Rc<DxvkCommandList>> m_entries;
-    dxvk::thread             m_thread;
+    std::mutex              m_mutexQueue;
     
-    void threadFunc();
+    std::condition_variable m_appendCond;
+    std::condition_variable m_submitCond;
+    std::condition_variable m_finishCond;
+
+    std::queue<DxvkSubmitInfo> m_submitQueue;
+    std::queue<DxvkSubmitInfo> m_finishQueue;
+
+    dxvk::thread            m_submitThread;
+    dxvk::thread            m_finishThread;
+
+    VkResult submitToQueue(
+      const DxvkSubmitInfo& submission);
+
+    void submitCmdLists();
+
+    void finishCmdLists();
     
   };