Fix access to render state in D3D12

* We need to copy the external-facing m_RenderState into the partial command
  buffer's state for partial replay, in case it was modified externally.
* Also when accessing the render state inside a drawcall callback we need to use
  the command buffer's local state, not m_RenderState which isn't updated until
  the replay completes.

renderdoc/driver/d3d12/d3d12_command_list1_wrap.cpp

@@ -543,14 +543,19 @@ bool WrappedID3D12GraphicsCommandList::Serialise_SetViewInstanceMask(SerialiserT
 
     m_Cmd->m_LastCmdListID = GetResourceManager()->GetOriginalID(GetResID(pCommandList));
 
+    bool stateUpdate = false;
+
     if(IsActiveReplaying(m_State))
     {
       if(m_Cmd->InRerecordRange(m_Cmd->m_LastCmdListID))
       {
         Unwrap1(m_Cmd->RerecordCmdList(m_Cmd->m_LastCmdListID))->SetViewInstanceMask(Mask);
 
-        if(m_Cmd->IsPartialCmdList(m_Cmd->m_LastCmdListID))
-          m_Cmd->m_RenderState.viewInstMask = Mask;
+        stateUpdate = true;
+      }
+      else if(!m_Cmd->IsPartialCmdList(m_Cmd->m_LastCmdListID))
+      {
+        stateUpdate = true;
       }
     }
     else
@@ -558,8 +563,11 @@ bool WrappedID3D12GraphicsCommandList::Serialise_SetViewInstanceMask(SerialiserT
       Unwrap1(pCommandList)->SetViewInstanceMask(Mask);
       GetCrackedList1()->SetViewInstanceMask(Mask);
 
-      m_Cmd->m_BakedCmdListInfo[m_Cmd->m_LastCmdListID].state.viewInstMask = Mask;
+      stateUpdate = true;
     }
+
+    if(stateUpdate)
+      m_Cmd->m_BakedCmdListInfo[m_Cmd->m_LastCmdListID].state.viewInstMask = Mask;
   }
 
   return true;

renderdoc/driver/d3d12/d3d12_commands.h

@@ -241,6 +241,11 @@ struct D3D12CommandData
 
   D3D12RenderState m_RenderState;
 
+  D3D12RenderState &GetCurRenderState()
+  {
+    return m_LastCmdListID == ResourceId() ? m_RenderState
+                                           : m_BakedCmdListInfo[m_LastCmdListID].state;
+  }
   enum PartialReplayIndex
   {
     Primary,

renderdoc/driver/d3d12/d3d12_device.cpp

@@ -3470,6 +3470,12 @@ void WrappedID3D12Device::ReplayLog(uint32_t startEventID, uint32_t endEventID,
       cmd.m_RenderState.m_ResourceManager = GetResourceManager();
       cmd.m_RenderState.m_DebugManager = m_Replay->GetDebugManager();
     }
+    else
+    {
+      // Copy the state in case m_RenderState was modified externally for the partial replay.
+      cmd.m_BakedCmdListInfo[cmd.m_Partial[D3D12CommandData::Primary].partialParent].state =
+          cmd.m_RenderState;
+    }
 
     // we'll need our own command list if we're replaying just a subsection
     // of events within a single command list record - always if it's only
@@ -3506,9 +3512,8 @@ void WrappedID3D12Device::ReplayLog(uint32_t startEventID, uint32_t endEventID,
       cmd.m_OutsideCmdList = NULL;
     }
 
-    if(!partial)
-      cmd.m_RenderState =
-          cmd.m_BakedCmdListInfo[cmd.m_Partial[D3D12CommandData::Primary].partialParent].state;
+    cmd.m_RenderState =
+        cmd.m_BakedCmdListInfo[cmd.m_Partial[D3D12CommandData::Primary].partialParent].state;
 
 #if ENABLED(SINGLE_FLUSH_VALIDATE)
     FlushLists(true);

renderdoc/driver/d3d12/d3d12_overlay.cpp

@@ -66,7 +66,7 @@ struct D3D12QuadOverdrawCallback : public D3D12DrawcallCallback
     // and substitute our quad calculation fragment shader that writes to a storage image
     // that is bound in a new root signature element.
 
-    D3D12RenderState &rs = m_pDevice->GetQueue()->GetCommandData()->m_RenderState;
+    D3D12RenderState &rs = m_pDevice->GetQueue()->GetCommandData()->GetCurRenderState();
     m_PrevState = rs;
 
     // check cache first
@@ -181,10 +181,10 @@ struct D3D12QuadOverdrawCallback : public D3D12DrawcallCallback
       return false;
 
     // restore the render state and go ahead with the real draw
-    m_pDevice->GetQueue()->GetCommandData()->m_RenderState = m_PrevState;
+    m_pDevice->GetQueue()->GetCommandData()->GetCurRenderState() = m_PrevState;
 
     RDCASSERT(cmd);
-    m_pDevice->GetQueue()->GetCommandData()->m_RenderState.ApplyState(m_pDevice, cmd);
+    m_pDevice->GetQueue()->GetCommandData()->GetCurRenderState().ApplyState(m_pDevice, cmd);
 
     return true;
   }

renderdoc/driver/d3d12/d3d12_postvs.cpp

@@ -173,7 +173,7 @@ void D3D12Replay::InitPostVSBuffers(uint32_t eventId)
   D3D12MarkerRegion postvs(m_pDevice->GetQueue(), StringFormat::Fmt("PostVS for %u", eventId));
 
   D3D12CommandData *cmd = m_pDevice->GetQueue()->GetCommandData();
-  const D3D12RenderState &rs = cmd->m_RenderState;
+  const D3D12RenderState &rs = cmd->GetCurRenderState();
 
   if(rs.pipe == ResourceId())
     return;

util/test/rdtest/shared/Overlay_Test.py

@@ -362,6 +362,8 @@ class Overlay_Test(rdtest.TestCase):
 
             rdtest.log.success("Picked pixels are as expected for {}".format(str(overlay)))
 
+        rdtest.log.success("All normal overlays are as expected")
+
         # Now check clear-before-X by hand, for colour and for depth
 
         depth_tex: rd.ResourceId = pipe.GetDepthTarget().resourceId
@@ -387,6 +389,8 @@ class Overlay_Test(rdtest.TestCase):
         self.check_pixel_value(depth_tex, 250, 250, [0.95, 0.0/255.0, 0.0, 1.0], eps=eps)
         self.check_pixel_value(depth_tex, 50, 50, [1.0, 0.0/255.0, 0.0, 1.0], eps=eps)
 
+        rdtest.log.success("Colour and depth at end are correct")
+
         # Check clear before pass
         tex.resourceId = col_tex
         tex.overlay = rd.DebugOverlay.ClearBeforePass