[d3d11] Recycle command lists from deferred contexts

Can save a few memory allocations and deallocations at runtime.

src/d3d11/d3d11_cmdlist.cpp

@@ -4,10 +4,11 @@
 namespace dxvk {
     
   D3D11CommandList::D3D11CommandList(
-          D3D11Device*  pDevice,
+          D3D11Device*               pDevice,
-          UINT          ContextFlags)
+          D3D11CommandListAllocator* pAllocator)
-  : m_device      (pDevice),
+  : m_device(pDevice), m_allocator(pAllocator) {
-    m_contextFlags(ContextFlags) { }
+
+  }
   
   
   D3D11CommandList::~D3D11CommandList() {
@@ -25,8 +26,14 @@ namespace dxvk {
 
   ULONG STDMETHODCALLTYPE D3D11CommandList::Release() {
     ULONG refCount = --m_refCount;
-    if (!refCount)
+
+    if (!refCount) {
+      Reset();
+
+      m_allocator->RecycleCommandList(this);
       m_device->Release();
+    }
+
     return refCount;
   }
   
@@ -81,6 +88,15 @@ namespace dxvk {
     
     MarkSubmitted();
   }
+
+
+  void D3D11CommandList::Reset() {
+    m_chunks.clear();
+
+    m_contextFlags  = 0;
+    m_submitted     = false;
+    m_warned        = false;
+  }
   
   
   void D3D11CommandList::MarkSubmitted() {
@@ -92,4 +108,42 @@ namespace dxvk {
     }
   }
   
+
+  D3D11CommandListAllocator::D3D11CommandListAllocator(D3D11Device* pDevice)
+  : m_device(pDevice) {
+
+  }
+
+
+  D3D11CommandListAllocator::~D3D11CommandListAllocator() {
+    for (uint32_t i = 0; i < m_listCount; i++)
+      delete m_lists[i];
+  }
+
+
+  D3D11CommandList* D3D11CommandListAllocator::AllocCommandList(
+          UINT                  ContextFlags) {
+    std::lock_guard<std::mutex> lock(m_mutex);
+    D3D11CommandList* result = nullptr;
+    
+    if (m_listCount)
+      result = m_lists[--m_listCount];
+    else
+      result = new D3D11CommandList(m_device, this);
+
+    result->SetContextFlags(ContextFlags);
+    return result;
+  }
+
+
+  void D3D11CommandListAllocator::RecycleCommandList(
+          D3D11CommandList*     pCommandList) {
+    std::lock_guard<std::mutex> lock(m_mutex);
+
+    if (m_listCount < m_lists.size())
+      m_lists[m_listCount++] = pCommandList;
+    else
+      delete pCommandList;
+  }
+
 }

src/d3d11/d3d11_cmdlist.h

@@ -1,16 +1,21 @@
 #pragma once
 
+#include <mutex>
+#include <vector>
+
 #include "d3d11_context.h"
 
 namespace dxvk {
+
+  class D3D11CommandListAllocator;
   
   class D3D11CommandList : public D3D11DeviceChild<ID3D11CommandList, NoWrapper> {
     
   public:
     
     D3D11CommandList(
-            D3D11Device*  pDevice,
+            D3D11Device*               pDevice,
-            UINT          ContextFlags);
+            D3D11CommandListAllocator* pAllocator);
     
     ~D3D11CommandList();
     
@@ -26,6 +31,10 @@ namespace dxvk {
             ID3D11Device **ppDevice) final;
     
     UINT STDMETHODCALLTYPE GetContextFlags() final;
+
+    void SetContextFlags(UINT ContextFlags) {
+      m_contextFlags = ContextFlags;
+    }
     
     void AddChunk(
             DxvkCsChunkRef&&    Chunk);
@@ -38,9 +47,11 @@ namespace dxvk {
     
   private:
     
-    D3D11Device* const m_device;
+    D3D11Device*                m_device;
-    UINT         const m_contextFlags;
+    D3D11CommandListAllocator*  m_allocator;
-    
+
+    uint32_t m_contextFlags = 0;
+
     std::vector<DxvkCsChunkRef> m_chunks;
 
     std::atomic<bool> m_submitted = { false };
@@ -48,8 +59,56 @@ namespace dxvk {
 
     std::atomic<uint32_t> m_refCount = { 0u };
 
+    void Reset();
+
     void MarkSubmitted();
     
   };
   
+
+  /**
+   * \brief Command list allocator
+   *
+   * Creates and recycles command list instances
+   * in order to reduce deferred context overhead.
+   */
+  class D3D11CommandListAllocator {
+
+  public:
+
+    D3D11CommandListAllocator(
+            D3D11Device*  pDevice);
+
+    ~D3D11CommandListAllocator();
+
+    /**
+     * \brief Allocates a command list
+     *
+     * \param [in] ContextFlags Flags of the parent context
+     * \returns The command list
+     */
+    D3D11CommandList* AllocCommandList(
+            UINT                  ContextFlags);
+
+    /**
+     * \brief Recycles a command list
+     *
+     * Automatically called when the command list
+     * in question reaches a ref count of zero.
+     * \param [in] pCommandList The command list
+     */
+    void RecycleCommandList(
+            D3D11CommandList*     pCommandList);
+
+  private:
+
+    D3D11Device* m_device;
+
+    std::mutex                        m_mutex;
+    std::array<D3D11CommandList*, 64> m_lists;
+
+    uint32_t m_listCount = 0;
+
+  };
+  
 }

src/d3d11/d3d11_context_def.cpp

@@ -309,7 +309,7 @@ namespace dxvk {
   
   
   Com<D3D11CommandList> D3D11DeferredContext::CreateCommandList() {
-    return new D3D11CommandList(m_parent, m_contextFlags);
+    return m_parent->AllocCommandList(m_contextFlags);
   }
   
   

src/d3d11/d3d11_device.cpp

@@ -39,7 +39,8 @@ namespace dxvk {
     m_dxvkAdapter   (m_dxvkDevice->adapter()),
     m_d3d11Formats  (m_dxvkAdapter),
     m_d3d11Options  (m_dxvkAdapter->instance()->config()),
-    m_dxbcOptions   (m_dxvkDevice, m_d3d11Options) {
+    m_dxbcOptions   (m_dxvkDevice, m_d3d11Options),
+    m_commandListAllocator(this) {
     m_initializer = new D3D11Initializer(this);
     m_context     = new D3D11ImmediateContext(this, m_dxvkDevice);
     m_d3d10Device = new D3D10Device(this, m_context);

src/d3d11/d3d11_device.h

@@ -14,6 +14,7 @@
 
 #include "../util/com/com_private_data.h"
 
+#include "d3d11_cmdlist.h"
 #include "d3d11_counter_buffer.h"
 #include "d3d11_initializer.h"
 #include "d3d11_interfaces.h"
@@ -411,6 +412,10 @@ namespace dxvk {
       DxvkCsChunk* chunk = m_csChunkPool.allocChunk(flags);
       return DxvkCsChunkRef(chunk, &m_csChunkPool);
     }
+
+    D3D11CommandList* AllocCommandList(UINT ContextFlags) {
+      return m_commandListAllocator.AllocCommandList(ContextFlags);
+    }
     
     const D3D11Options* GetOptions() const {
       return &m_d3d11Options;
@@ -465,6 +470,7 @@ namespace dxvk {
     D3D11StateObjectSet<D3D11RasterizerState>   m_rsStateObjects;
     D3D11StateObjectSet<D3D11SamplerState>      m_samplerObjects;
     D3D11ShaderModuleSet                        m_shaderModules;
+    D3D11CommandListAllocator                   m_commandListAllocator;
     
     Rc<D3D11CounterBuffer> CreateUAVCounterBuffer();
     Rc<D3D11CounterBuffer> CreateXFBCounterBuffer();