Optimized command submission

Command submission now does not synchronize with the device every single
time. Instead, the command list and the fence that was created for it are
added to a queue. A separate thread will then wait for the execution to
complete and return the command list to the device.

src/dxvk/dxvk_device.cpp

@@ -12,7 +12,8 @@ namespace dxvk {
     m_features        (features),
     m_memory          (new DxvkMemoryAllocator(adapter, vkd)),
     m_renderPassPool  (new DxvkRenderPassPool (vkd)),
-    m_pipelineManager (new DxvkPipelineManager(vkd)) {
+    m_pipelineManager (new DxvkPipelineManager(vkd)),
+    m_submissionQueue (this) {
     m_vkd->vkGetDeviceQueue(m_vkd->device(),
       m_adapter->graphicsQueueFamily(), 0,
       &m_graphicsQueue);
@@ -214,12 +215,8 @@ namespace dxvk {
         waitSemaphore, wakeSemaphore, fence->handle());
     }
     
-    // TODO Delay synchronization by putting these into a ring buffer
-    fence->wait(std::numeric_limits<uint64_t>::max());
-    commandList->reset();
-    
-    // FIXME this must go away once the ring buffer is implemented
-    m_recycledCommandLists.returnObject(commandList);
+    // Add this to the set of running submissions
+    m_submissionQueue.submit(fence, commandList);
     m_statCounters.increment(DxvkStat::DevQueueSubmissions, 1);
     return fence;
   }
@@ -232,4 +229,9 @@ namespace dxvk {
       throw DxvkError("DxvkDevice::waitForIdle: Operation failed");
   }
   
+  
+  void DxvkDevice::recycleCommandList(const Rc<DxvkCommandList>& cmdList) {
+    m_recycledCommandLists.returnObject(cmdList);
+  }
+  
 }

src/dxvk/dxvk_device.h

@@ -9,6 +9,7 @@
 #include "dxvk_image.h"
 #include "dxvk_memory.h"
 #include "dxvk_pipemanager.h"
+#include "dxvk_queue.h"
 #include "dxvk_recycler.h"
 #include "dxvk_renderpass.h"
 #include "dxvk_sampler.h"
@@ -30,6 +31,8 @@ namespace dxvk {
    * contexts. Multiple contexts can be created for a device.
    */
   class DxvkDevice : public RcObject {
+    friend class DxvkSubmissionQueue;
+    
     constexpr static VkDeviceSize DefaultStagingBufferSize = 64 * 1024 * 1024;
   public:
     
@@ -308,6 +311,11 @@ namespace dxvk {
     
     DxvkStatCounters m_statCounters;
     
+    DxvkSubmissionQueue m_submissionQueue;
+    
+    void recycleCommandList(
+      const Rc<DxvkCommandList>& cmdList);
+    
   };
   
 }

src/dxvk/dxvk_memory.cpp

@@ -58,23 +58,20 @@ namespace dxvk {
   : m_heap  (heap),
     m_memory(memory),
     m_mapPtr(mapPtr),
-    m_size  (size),
-    m_free  (size) {
-    TRACE(this);
+    m_size  (size) {
     // Mark the entire chunk as free
     m_freeList.push_back(FreeSlice { 0, size });
   }
   
   
   DxvkMemoryChunk::~DxvkMemoryChunk() {
-    TRACE(this);
     m_heap->freeDeviceMemory(m_memory);
   }
   
   
   DxvkMemory DxvkMemoryChunk::alloc(VkDeviceSize size, VkDeviceSize align) {
-    // Fast exit if the chunk is full already
-    if (size > m_free)
+    // If the chunk is full, return
+    if (m_freeList.size() == 0)
       return DxvkMemory();
     
     // Select the slice to allocate from in a worst-fit
@@ -82,8 +79,12 @@ namespace dxvk {
     auto bestSlice = m_freeList.begin();
     
     for (auto slice = m_freeList.begin(); slice != m_freeList.end(); slice++) {
-      if (slice->length > bestSlice->length)
+      if (slice->length == size) {
         bestSlice = slice;
+        break;
+      } else if (slice->length > bestSlice->length) {
+        bestSlice = slice;
+      }
     }
     
     // We need to align the allocation to the requested alignment
@@ -99,7 +100,6 @@ namespace dxvk {
     // We can use this slice, but we'll have to add
     // the unused parts of it back to the free list.
     m_freeList.erase(bestSlice);
-    m_free -= size;
     
     if (allocStart != sliceStart)
       m_freeList.push_back({ sliceStart, allocStart - sliceStart });
@@ -108,6 +108,7 @@ namespace dxvk {
       m_freeList.push_back({ allocEnd, sliceEnd - allocEnd });
     
     // Create the memory object with the aligned slice
+    m_delta++;
     return DxvkMemory(this, m_heap,
       m_memory, allocStart, allocEnd - allocStart,
       reinterpret_cast<char*>(m_mapPtr) + allocStart);
@@ -117,8 +118,6 @@ namespace dxvk {
   void DxvkMemoryChunk::free(
           VkDeviceSize  offset,
           VkDeviceSize  length) {
-    m_free += length;
-    
     // Remove adjacent entries from the free list and then add
     // a new slice that covers all those entries. Without doing
     // so, the slice could not be reused for larger allocations.
@@ -137,6 +136,7 @@ namespace dxvk {
       }
     }
     
+    m_delta--;
     m_freeList.push_back({ offset, length });
   }
   

src/dxvk/dxvk_memory.h

@@ -130,8 +130,7 @@ namespace dxvk {
     VkDeviceMemory  const m_memory;
     void*           const m_mapPtr;
     VkDeviceSize    const m_size;
-    VkDeviceSize          m_free = 0;
-    
+    size_t m_delta = 0;
     std::vector<FreeSlice> m_freeList;
     
   };

src/dxvk/dxvk_queue.cpp

@@ -0,0 +1,62 @@
+#include "dxvk_device.h"
+#include "dxvk_queue.h"
+
+namespace dxvk {
+  
+  DxvkSubmissionQueue::DxvkSubmissionQueue(DxvkDevice* device)
+  : m_device(device),
+    m_thread([this] () { this->threadFunc(); }) {
+    
+  }
+  
+  
+  DxvkSubmissionQueue::~DxvkSubmissionQueue() {
+    m_stopped.store(true);
+    m_condOnAdd.notify_one();
+    m_thread.join();
+  }
+  
+  
+  void DxvkSubmissionQueue::submit(
+    const Rc<DxvkFence>&        fence,
+    const Rc<DxvkCommandList>&  cmdList) {
+    { std::unique_lock<std::mutex> lock(m_mutex);
+      
+      m_condOnTake.wait(lock, [this] {
+        return m_entries.size() < 4;
+      });
+      
+      m_entries.push({ fence, cmdList });
+    }
+    
+    m_condOnAdd.notify_one();
+  }
+  
+  
+  void DxvkSubmissionQueue::threadFunc() {
+    while (!m_stopped.load()) {
+      Entry entry;
+      
+      { 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) {
+          entry = std::move(m_entries.front());
+          m_entries.pop();
+        }
+      }
+      
+      m_condOnTake.notify_one();
+      
+      if (entry.fence != nullptr) {
+        entry.fence->wait(std::numeric_limits<uint64_t>::max());
+        entry.cmdList->reset();
+        m_device->recycleCommandList(entry.cmdList);
+      }
+    }
+  }
+  
+}

src/dxvk/dxvk_queue.h

@@ -0,0 +1,52 @@
+#pragma once
+
+#include <condition_variable>
+#include <mutex>
+#include <queue>
+#include <thread>
+
+#include "dxvk_cmdlist.h"
+#include "dxvk_sync.h"
+
+namespace dxvk {
+  
+  class DxvkDevice;
+  
+  /**
+   * \brief Submission queue
+   * 
+   * 
+   */
+  class DxvkSubmissionQueue {
+    
+  public:
+    
+    DxvkSubmissionQueue(DxvkDevice* device);
+    ~DxvkSubmissionQueue();
+    
+    void submit(
+      const Rc<DxvkFence>&        fence,
+      const Rc<DxvkCommandList>&  cmdList);
+    
+  private:
+    
+    struct Entry {
+      Rc<DxvkFence>       fence;
+      Rc<DxvkCommandList> cmdList;
+    };
+    
+    DxvkDevice*             m_device;
+    
+    std::atomic<bool>       m_stopped = { false };
+    
+    std::mutex              m_mutex;
+    std::condition_variable m_condOnAdd;
+    std::condition_variable m_condOnTake;
+    std::queue<Entry>       m_entries;
+    std::thread             m_thread;
+    
+    void threadFunc();
+    
+  };
+  
+}

src/dxvk/meson.build

@@ -18,6 +18,7 @@ dxvk_src = files([
   'dxvk_memory.cpp',
   'dxvk_pipelayout.cpp',
   'dxvk_pipemanager.cpp',
+  'dxvk_queue.cpp',
   'dxvk_renderpass.cpp',
   'dxvk_resource.cpp',
   'dxvk_sampler.cpp',