Handle ExecuteIndirect by cracking lists apart and patching during read

renderdoc/driver/d3d12/d3d12_command_list.h

@@ -410,8 +410,9 @@ public:
 
   IMPLEMENT_FUNCTION_SERIALISED(virtual void STDMETHODCALLTYPE, EndEvent());
 
-  void PatchedExecuteIndirect(ID3D12GraphicsCommandList *list, ResourceId sig, UINT maxCount,
-                              ResourceId arg, UINT64 argOffs);
+  void ReserveExecuteIndirect(ID3D12GraphicsCommandList *list, ResourceId sig, UINT maxCount);
+  void PatchExecuteIndirect(BakedCmdListInfo &info, uint32_t executeIndex);
+  void ReplayExecuteIndirect(ID3D12GraphicsCommandList *list, BakedCmdListInfo &info);
 
   IMPLEMENT_FUNCTION_SERIALISED(virtual void STDMETHODCALLTYPE,
                                 ExecuteIndirect(ID3D12CommandSignature *pCommandSignature,

renderdoc/driver/d3d12/d3d12_command_queue_wrap.cpp

@@ -123,8 +123,32 @@ bool WrappedID3D12CommandQueue::Serialise_ExecuteCommandLists(UINT NumCommandLis
   {
     for(uint32_t i = 0; i < numCmds; i++)
     {
-      ID3D12CommandList *c = Unwrap(cmds[i]);
-      m_pReal->ExecuteCommandLists(1, &c);
+      if(m_Cmd.m_BakedCmdListInfo[cmdIds[i]].executeEvents.empty() ||
+         m_Cmd.m_BakedCmdListInfo[cmdIds[i]].executeEvents[0].patched)
+      {
+        ID3D12CommandList *list = Unwrap(cmds[i]);
+        m_pReal->ExecuteCommandLists(1, &list);
+      }
+      else
+      {
+        BakedCmdListInfo &info = m_Cmd.m_BakedCmdListInfo[cmdIds[i]];
+
+        // execute the first half of the cracked list
+        ID3D12CommandList *list = Unwrap(info.crackedLists[0]);
+        m_pReal->ExecuteCommandLists(1, &list);
+
+        for(size_t c = 1; c < info.crackedLists.size(); c++)
+        {
+          m_pDevice->GPUSync();
+
+          // readback the patch buffer and perform patching
+          m_ReplayList->PatchExecuteIndirect(info, uint32_t(c - 1));
+
+          // execute next list with this indirect.
+          list = Unwrap(info.crackedLists[c]);
+          m_pReal->ExecuteCommandLists(1, &list);
+        }
+      }
     }
 
     for(uint32_t i = 0; i < numCmds; i++)

renderdoc/driver/d3d12/d3d12_commands.cpp

@@ -695,6 +695,51 @@ HRESULT STDMETHODCALLTYPE WrappedID3D12GraphicsCommandList::QueryInterface(REFII
   return RefCounter12::QueryInterface(riid, ppvObject);
 }
 
+void BakedCmdListInfo::ShiftForRemoved(uint32_t shiftDrawID, uint32_t shiftEID, size_t idx)
+{
+  std::vector<D3D12DrawcallTreeNode> &draws = draw->children;
+
+  drawCount -= shiftDrawID;
+  eventCount -= shiftEID;
+
+  if(idx < draws.size())
+  {
+    for(size_t i = idx; i < draws.size(); i++)
+    {
+      // should have no children as we don't push in for markers since they
+      // can cross command list boundaries.
+      RDCASSERT(draws[i].children.empty());
+
+      draws[i].draw.eventID -= shiftEID;
+      draws[i].draw.drawcallID -= shiftDrawID;
+
+      for(int32_t e = 0; e < draws[i].draw.events.count; e++)
+        draws[i].draw.events[e].eventID -= shiftEID;
+    }
+
+    uint32_t lastEID = draws[idx].draw.eventID;
+
+    // shift any resource usage for drawcalls after the removed section
+    for(size_t i = 0; i < draw->resourceUsage.size(); i++)
+    {
+      if(draw->resourceUsage[i].second.eventID >= lastEID)
+        draw->resourceUsage[i].second.eventID -= shiftEID;
+    }
+
+    // patch any subsequent executes
+    for(size_t i = 0; i < executeEvents.size(); i++)
+    {
+      if(executeEvents[i].baseEvent >= lastEID)
+      {
+        executeEvents[i].baseEvent -= shiftEID;
+
+        if(executeEvents[i].lastEvent > 0)
+          executeEvents[i].lastEvent -= shiftEID;
+      }
+    }
+  }
+}
+
 D3D12CommandData::D3D12CommandData()
 {
   m_CurChunkOffset = 0;

renderdoc/driver/d3d12/d3d12_commands.h

@@ -131,6 +131,7 @@ struct BakedCmdListInfo
     eventCount = parent.curEventID;
     drawCount = parent.drawCount;
     crackedLists.swap(parent.crackedLists);
+    executeEvents.swap(parent.executeEvents);
 
     parentList = parentID;
 
@@ -144,7 +145,37 @@ struct BakedCmdListInfo
     parent.debugMessages.clear();
   }
 
+  void ShiftForRemoved(uint32_t shiftDrawID, uint32_t shiftEID, size_t idx);
+
+  struct ExecuteData
+  {
+    ExecuteData()
+        : baseEvent(0),
+          lastEvent(0),
+          patched(false),
+          argBuf(NULL),
+          countBuf(NULL),
+          argOffs(0),
+          countOffs(0),
+          maxCount(0),
+          realCount(0)
+    {
+    }
+
+    uint32_t baseEvent;
+    uint32_t lastEvent;
+    bool patched;
+    ID3D12Resource *argBuf;
+    ID3D12Resource *countBuf;
+    uint64_t argOffs;
+    uint64_t countOffs;
+    ResourceId sig;
+    UINT maxCount;
+    UINT realCount;
+  };
+
   vector<ID3D12GraphicsCommandList *> crackedLists;
+  vector<ExecuteData> executeEvents;
 
   vector<FetchAPIEvent> curEvents;
   vector<DebugMessage> debugMessages;