Add some DX12 performance optimisations

* Switches some contended mutexes for R/W locks

renderdoc/core/resource_manager.h

@@ -151,7 +151,10 @@ struct ResourceRecord
     }
 
     if(!dataWritten)
-      recordlist.insert(m_Chunks.begin(), m_Chunks.end());
+    {
+      for(auto it = m_Chunks.begin(); it != m_Chunks.end(); ++it)
+        recordlist[it->first] = it->second;
+    }
   }
 
   void AddRef() { Atomic::Inc32(&RefCount); }
@@ -175,10 +178,10 @@ struct ResourceRecord
 
   void AddChunk(Chunk *chunk, int32_t ID = 0)
   {
-    LockChunks();
     if(ID == 0)
       ID = GetID();
-    m_Chunks[ID] = chunk;
+    LockChunks();
+    m_Chunks.push_back(std::make_pair(ID, chunk));
     UnlockChunks();
   }
 
@@ -232,16 +235,16 @@ struct ResourceRecord
   Chunk *GetLastChunk() const
   {
     RDCASSERT(HasChunks());
-    return m_Chunks.rbegin()->second;
+    return m_Chunks.back().second;
   }
 
   int32_t GetLastChunkID() const
   {
     RDCASSERT(HasChunks());
-    return m_Chunks.rbegin()->first;
+    return m_Chunks.back().first;
   }
 
-  void PopChunk() { m_Chunks.erase(m_Chunks.rbegin()->first); }
+  void PopChunk() { m_Chunks.pop_back(); }
   byte *GetDataPtr() { return DataPtr + DataOffset; }
   bool HasDataPtr() { return DataPtr != NULL; }
   void SetDataOffset(uint64_t offs) { DataOffset = offs; }
@@ -278,7 +281,7 @@ protected:
     return Atomic::Inc32(&globalIDCounter);
   }
 
-  std::map<int32_t, Chunk *> m_Chunks;
+  std::vector<std::pair<int32_t, Chunk *>> m_Chunks;
   Threading::CriticalSection *m_ChunkLock;
 
   map<ResourceId, FrameRefType> m_FrameRefs;

renderdoc/driver/d3d12/d3d12_command_queue.h

@@ -252,6 +252,10 @@ public:
   IMPLEMENT_FUNCTION_SERIALISED(virtual void STDMETHODCALLTYPE, ExecuteCommandLists,
                                 UINT NumCommandLists, ID3D12CommandList *const *ppCommandLists);
 
+  virtual void ExecuteCommandListsInternal(UINT NumCommandLists,
+                                           ID3D12CommandList *const *ppCommandLists,
+                                           bool InFrameCaptureBoundary);
+
   IMPLEMENT_FUNCTION_SERIALISED(virtual void STDMETHODCALLTYPE, SetMarker, UINT Metadata,
                                 const void *pData, UINT Size);
 

renderdoc/driver/d3d12/d3d12_command_queue_wrap.cpp

@@ -321,8 +321,15 @@ bool WrappedID3D12CommandQueue::Serialise_ExecuteCommandLists(SerialiserType &se
   return true;
 }
 
-void STDMETHODCALLTYPE WrappedID3D12CommandQueue::ExecuteCommandLists(
-    UINT NumCommandLists, ID3D12CommandList *const *ppCommandLists)
+void WrappedID3D12CommandQueue::ExecuteCommandLists(UINT NumCommandLists,
+                                                    ID3D12CommandList *const *ppCommandLists)
+{
+  ExecuteCommandListsInternal(NumCommandLists, ppCommandLists, false);
+}
+
+void WrappedID3D12CommandQueue::ExecuteCommandListsInternal(UINT NumCommandLists,
+                                                            ID3D12CommandList *const *ppCommandLists,
+                                                            bool InFrameCaptureBoundary)
 {
   ID3D12CommandList **unwrapped = m_pDevice->GetTempArray<ID3D12CommandList *>(NumCommandLists);
   for(UINT i = 0; i < NumCommandLists; i++)
@@ -342,7 +349,9 @@ void STDMETHODCALLTYPE WrappedID3D12CommandQueue::ExecuteCommandLists(
   if(IsCaptureMode(m_State))
   {
     SCOPED_LOCK(m_Lock);
-    SCOPED_LOCK(m_pDevice->GetCapTransitionLock());
+
+    if(!InFrameCaptureBoundary)
+      m_pDevice->GetCapTransitionLock().ReadLock();
 
     bool capframe = IsActiveCapturing(m_State);
     std::set<ResourceId> refdIDs;
@@ -529,6 +538,9 @@ void STDMETHODCALLTYPE WrappedID3D12CommandQueue::ExecuteCommandLists(
         m_QueueRecord->AddChunk(scope.Get());
       }
     }
+
+    if(!InFrameCaptureBoundary)
+      m_pDevice->GetCapTransitionLock().ReadUnlock();
   }
 }
 

renderdoc/driver/d3d12/d3d12_device.cpp

@@ -886,7 +886,7 @@ void WrappedID3D12Device::Unmap(ID3D12Resource *Resource, UINT Subresource, byte
 
   bool capframe = false;
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1119,7 +1119,7 @@ void WrappedID3D12Device::WriteToSubresource(ID3D12Resource *Resource, UINT Subr
 {
   bool capframe = false;
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1316,7 +1316,7 @@ void WrappedID3D12Device::StartFrameCapture(void *dev, void *wnd)
   // will check to see if they need to markdirty or markpendingdirty
   // and go into the frame record.
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_WRITELOCK(m_CapTransitionLock);
 
     initStateCurBatch = 0;
     initStateCurList = NULL;
@@ -1332,7 +1332,7 @@ void WrappedID3D12Device::StartFrameCapture(void *dev, void *wnd)
     initStateCurBatch = 0;
     initStateCurList = NULL;
 
-    ExecuteLists();
+    ExecuteLists(NULL, true);
     FlushLists();
 
     RDCDEBUG("Attempting capture");
@@ -1406,7 +1406,7 @@ bool WrappedID3D12Device::EndFrameCapture(void *dev, void *wnd)
 
   // transition back to IDLE and readback initial states atomically
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_WRITELOCK(m_CapTransitionLock);
     EndCaptureFrame(backbuffer);
 
     m_State = CaptureState::BackgroundCapturing;
@@ -1495,7 +1495,7 @@ bool WrappedID3D12Device::EndFrameCapture(void *dev, void *wnd)
 
         list->Close();
 
-        ExecuteLists();
+        ExecuteLists(NULL, true);
         FlushLists();
 
         byte *data = NULL;
@@ -2392,13 +2392,14 @@ void WrappedID3D12Device::CloseInitialStateList()
   initStateCurBatch = 0;
 }
 
-void WrappedID3D12Device::ExecuteList(ID3D12GraphicsCommandList2 *list, ID3D12CommandQueue *queue)
+void WrappedID3D12Device::ExecuteList(ID3D12GraphicsCommandList2 *list,
+                                      WrappedID3D12CommandQueue *queue, bool InFrameCaptureBoundary)
 {
   if(queue == NULL)
     queue = GetQueue();
 
   ID3D12CommandList *l = list;
-  queue->ExecuteCommandLists(1, &l);
+  queue->ExecuteCommandListsInternal(1, &l, InFrameCaptureBoundary);
 
   for(auto it = m_InternalCmds.pendingcmds.begin(); it != m_InternalCmds.pendingcmds.end(); ++it)
   {
@@ -2412,7 +2413,7 @@ void WrappedID3D12Device::ExecuteList(ID3D12GraphicsCommandList2 *list, ID3D12Co
   m_InternalCmds.submittedcmds.push_back(list);
 }
 
-void WrappedID3D12Device::ExecuteLists(ID3D12CommandQueue *queue)
+void WrappedID3D12Device::ExecuteLists(WrappedID3D12CommandQueue *queue, bool InFrameCaptureBoundary)
 {
   // nothing to do
   if(m_InternalCmds.pendingcmds.empty())
@@ -2426,7 +2427,7 @@ void WrappedID3D12Device::ExecuteLists(ID3D12CommandQueue *queue)
   if(queue == NULL)
     queue = GetQueue();
 
-  queue->ExecuteCommandLists((UINT)cmds.size(), &cmds[0]);
+  queue->ExecuteCommandListsInternal((UINT)cmds.size(), &cmds[0], InFrameCaptureBoundary);
 
   m_InternalCmds.submittedcmds.insert(m_InternalCmds.submittedcmds.end(),
                                       m_InternalCmds.pendingcmds.begin(),

renderdoc/driver/d3d12/d3d12_device.h

@@ -312,7 +312,7 @@ private:
 
   bool m_AppControlledCapture;
 
-  Threading::CriticalSection m_CapTransitionLock;
+  Threading::RWLock m_CapTransitionLock;
   CaptureState m_State;
 
   uint32_t m_SubmitCounter = 0;
@@ -351,7 +351,7 @@ private:
   {
     D3D12_CPU_DESCRIPTOR_HANDLE rtvs[8];
 
-    ID3D12CommandQueue *queue;
+    WrappedID3D12CommandQueue *queue;
 
     int32_t lastPresentedBuffer;
   };
@@ -397,7 +397,7 @@ public:
   D3D12ResourceManager *GetResourceManager() { return m_ResourceManager; }
   D3D12ShaderCache *GetShaderCache() { return m_ShaderCache; }
   ResourceId GetResourceID() { return m_ResourceID; }
-  Threading::CriticalSection &GetCapTransitionLock() { return m_CapTransitionLock; }
+  Threading::RWLock &GetCapTransitionLock() { return m_CapTransitionLock; }
   void ReleaseSwapchainResources(IDXGISwapChain *swap, IUnknown **backbuffers, int numBackbuffers);
   void FirstFrame(WrappedIDXGISwapChain4 *swap);
   FrameRecord &GetFrameRecord() { return m_FrameRecord; }
@@ -485,8 +485,9 @@ public:
 
   void AddCaptureSubmission();
 
-  void ExecuteList(ID3D12GraphicsCommandList2 *list, ID3D12CommandQueue *queue = NULL);
-  void ExecuteLists(ID3D12CommandQueue *queue = NULL);
+  void ExecuteList(ID3D12GraphicsCommandList2 *list, WrappedID3D12CommandQueue *queue = NULL,
+                   bool InFrameCaptureBoundary = false);
+  void ExecuteLists(WrappedID3D12CommandQueue *queue = NULL, bool InFrameCaptureBoundary = false);
   void FlushLists(bool forceSync = false, ID3D12CommandQueue *queue = NULL);
 
   void GPUSync(ID3D12CommandQueue *queue = NULL, ID3D12Fence *fence = NULL);

renderdoc/driver/d3d12/d3d12_device_wrap.cpp

@@ -748,7 +748,7 @@ HRESULT WrappedID3D12Device::CreateDescriptorHeap(const D3D12_DESCRIPTOR_HEAP_DE
       record->AddChunk(scope.Get());
 
       {
-        SCOPED_LOCK(m_CapTransitionLock);
+        SCOPED_READLOCK(m_CapTransitionLock);
         if(IsBackgroundCapturing(m_State))
           GetResourceManager()->MarkDirtyResource(wrapped->GetResourceID());
         else
@@ -912,7 +912,7 @@ void WrappedID3D12Device::CreateConstantBufferView(const D3D12_CONSTANT_BUFFER_V
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -955,7 +955,7 @@ void WrappedID3D12Device::CreateShaderResourceView(ID3D12Resource *pResource,
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1007,7 +1007,7 @@ void WrappedID3D12Device::CreateUnorderedAccessView(ID3D12Resource *pResource,
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1053,7 +1053,7 @@ void WrappedID3D12Device::CreateRenderTargetView(ID3D12Resource *pResource,
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1097,7 +1097,7 @@ void WrappedID3D12Device::CreateDepthStencilView(ID3D12Resource *pResource,
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1138,7 +1138,7 @@ void WrappedID3D12Device::CreateSampler(const D3D12_SAMPLER_DESC *pDesc,
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -1329,7 +1329,7 @@ HRESULT WrappedID3D12Device::CreateCommittedResource(const D3D12_HEAP_PROPERTIES
       record->AddChunk(scope.Get());
 
       {
-        SCOPED_LOCK(m_CapTransitionLock);
+        SCOPED_READLOCK(m_CapTransitionLock);
         if(IsBackgroundCapturing(m_State))
           GetResourceManager()->MarkDirtyResource(wrapped->GetResourceID());
         else
@@ -1576,7 +1576,7 @@ HRESULT WrappedID3D12Device::CreatePlacedResource(ID3D12Heap *pHeap, UINT64 Heap
       record->AddChunk(scope.Get());
 
       {
-        SCOPED_LOCK(m_CapTransitionLock);
+        SCOPED_READLOCK(m_CapTransitionLock);
         if(IsBackgroundCapturing(m_State))
           GetResourceManager()->MarkDirtyResource(wrapped->GetResourceID());
         else
@@ -1969,7 +1969,7 @@ void WrappedID3D12Device::CopyDescriptors(
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -2076,7 +2076,7 @@ void WrappedID3D12Device::CopyDescriptorsSimple(UINT NumDescriptors,
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 
@@ -2172,18 +2172,6 @@ HRESULT WrappedID3D12Device::MakeResident(UINT NumObjects, ID3D12Pageable *const
     }
   }
 
-  bool capframe = false;
-
-  {
-    SCOPED_LOCK(m_CapTransitionLock);
-    capframe = IsActiveCapturing(m_State);
-  }
-
-  if(capframe)
-  {
-    // serialise
-  }
-
   return m_pDevice->MakeResident(NumObjects, unwrapped);
 }
 
@@ -2211,18 +2199,6 @@ HRESULT WrappedID3D12Device::Evict(UINT NumObjects, ID3D12Pageable *const *ppObj
     }
   }
 
-  bool capframe = false;
-
-  {
-    SCOPED_LOCK(m_CapTransitionLock);
-    capframe = IsActiveCapturing(m_State);
-  }
-
-  if(capframe)
-  {
-    // serialise
-  }
-
   return m_pDevice->Evict(NumObjects, unwrapped);
 }
 
@@ -2742,7 +2718,7 @@ HRESULT WrappedID3D12Device::EnqueueMakeResident(D3D12_RESIDENCY_FLAGS Flags, UI
   bool capframe = false;
 
   {
-    SCOPED_LOCK(m_CapTransitionLock);
+    SCOPED_READLOCK(m_CapTransitionLock);
     capframe = IsActiveCapturing(m_State);
   }
 

renderdoc/driver/d3d12/d3d12_initstate.cpp

@@ -117,7 +117,7 @@ bool D3D12ResourceManager::Prepare_InitialState(ID3D12DeviceChild *res)
       {
         m_Device->CloseInitialStateList();
 
-        m_Device->ExecuteLists();
+        m_Device->ExecuteLists(NULL, true);
         m_Device->FlushLists();
 
         m_Device->Evict(1, &pageable);
@@ -228,7 +228,7 @@ bool D3D12ResourceManager::Prepare_InitialState(ID3D12DeviceChild *res)
       {
         m_Device->CloseInitialStateList();
 
-        m_Device->ExecuteLists();
+        m_Device->ExecuteLists(NULL, true);
         m_Device->FlushLists();
 
         m_Device->Evict(1, &pageable);
@@ -374,7 +374,7 @@ bool D3D12ResourceManager::Serialise_InitialState(SerialiserType &ser, ResourceI
 
     if(ser.IsWriting())
     {
-      m_Device->ExecuteLists();
+      m_Device->ExecuteLists(NULL, true);
       m_Device->FlushLists();
 
       mappedBuffer = (ID3D12Resource *)initContents.resource;
@@ -738,7 +738,7 @@ void D3D12ResourceManager::Apply_InitialState(ID3D12DeviceChild *live, D3D12Init
 
 #if ENABLED(SINGLE_FLUSH_VALIDATE)
         m_Device->CloseInitialStateList();
-        m_Device->ExecuteLists();
+        m_Device->ExecuteLists(NULL, true);
         m_Device->FlushLists(true);
 #endif
       }

renderdoc/driver/d3d12/d3d12_manager.h

@@ -320,11 +320,11 @@ struct GPUAddressRangeTracker
   GPUAddressRangeTracker &operator=(const GPUAddressRangeTracker &);
 
   std::vector<GPUAddressRange> addresses;
-  Threading::CriticalSection addressLock;
+  Threading::RWLock addressLock;
 
   void AddTo(GPUAddressRange range)
   {
-    SCOPED_LOCK(addressLock);
+    SCOPED_WRITELOCK(addressLock);
     auto it = std::lower_bound(addresses.begin(), addresses.end(), range.start);
     RDCASSERT(it == addresses.begin() || it == addresses.end() || range.start < it->start ||
               range.start >= it->end);
@@ -334,7 +334,7 @@ struct GPUAddressRangeTracker
 
   void RemoveFrom(D3D12_GPU_VIRTUAL_ADDRESS baseAddr)
   {
-    SCOPED_LOCK(addressLock);
+    SCOPED_WRITELOCK(addressLock);
     auto it = std::lower_bound(addresses.begin(), addresses.end(), baseAddr);
     RDCASSERT(it != addresses.end() && baseAddr >= it->start && baseAddr < it->end);
 
@@ -353,7 +353,7 @@ struct GPUAddressRangeTracker
 
     // this should really be a read-write lock
     {
-      SCOPED_LOCK(addressLock);
+      SCOPED_READLOCK(addressLock);
 
       auto it = std::lower_bound(addresses.begin(), addresses.end(), addr);
       if(it == addresses.end())

renderdoc/driver/d3d12/d3d12_resources.cpp

@@ -446,7 +446,7 @@ HRESULT STDMETHODCALLTYPE WrappedID3D12Resource::WriteToSubresource(UINT DstSubr
 void WrappedID3D12Resource::RefBuffers(D3D12ResourceManager *rm)
 {
   // only buffers go into m_Addresses
-  SCOPED_LOCK(m_Addresses.addressLock);
+  SCOPED_READLOCK(m_Addresses.addressLock);
   for(size_t i = 0; i < m_Addresses.addresses.size(); i++)
     rm->MarkResourceFrameReferenced(m_Addresses.addresses[i].id, eFrameRef_Read);
 }

renderdoc/driver/gl/gl_resources.h

@@ -238,24 +238,6 @@ struct GLResourceRecord : public ResourceRecord
     int64_t persistentMaps;    // counter indicating how many coherent maps are 'live'
   } Map;
 
-  template <typename ChunkFilter>
-  void FilterChunks(const ChunkFilter &filter)
-  {
-    LockChunks();
-    std::vector<std::map<int32_t, Chunk *>::iterator> deletions;
-    for(auto it = m_Chunks.begin(); it != m_Chunks.end(); ++it)
-    {
-      if(filter(it->second))
-        deletions.push_back(it);
-    }
-    for(size_t i = 0; i < deletions.size(); i++)
-    {
-      SAFE_DELETE(deletions[i]->second);
-      m_Chunks.erase(deletions[i]);
-    }
-    UnlockChunks();
-  }
-
   void VerifyDataType(GLenum target)
   {
 #if ENABLED(RDOC_DEVEL)