[d3d11] Improve flushing around deferred context submissions

src/d3d11/d3d11_cmdlist.cpp

@@ -67,20 +67,20 @@ namespace dxvk {
   }
   
   
-  uint64_t D3D11CommandList::EmitToCsThread(DxvkCsThread* CsThread) {
+  void D3D11CommandList::EmitToCsThread(
+    const D3D11ChunkDispatchProc& DispatchProc) {
     uint64_t seq = 0;
 
     for (const auto& query : m_queries)
       query->DoDeferredEnd();
 
     for (const auto& chunk : m_chunks)
-      seq = CsThread->dispatchChunk(DxvkCsChunkRef(chunk));
+      seq = DispatchProc(DxvkCsChunkRef(chunk));
     
     for (const auto& resource : m_resources)
       TrackResourceSequenceNumber(resource, seq);
 
     MarkSubmitted();
-    return seq;
   }
   
   

src/d3d11/d3d11_cmdlist.h

@@ -1,9 +1,13 @@
 #pragma once
 
+#include <functional>
+
 #include "d3d11_context.h"
 
 namespace dxvk {
   
+  using D3D11ChunkDispatchProc = std::function<uint64_t (DxvkCsChunkRef&&)>;
+
   class D3D11CommandList : public D3D11DeviceChild<ID3D11CommandList> {
     
   public:
@@ -29,14 +33,18 @@ namespace dxvk {
     void EmitToCommandList(
             ID3D11CommandList*  pCommandList);
     
-    uint64_t EmitToCsThread(
-            DxvkCsThread*       CsThread);
+    void EmitToCsThread(
+      const D3D11ChunkDispatchProc& DispatchProc);
 
     void TrackResourceUsage(
             ID3D11Resource*     pResource,
             D3D11_RESOURCE_DIMENSION ResourceType,
             UINT                Subresource);
 
+    bool HasTrackedResources() const {
+      return !m_resources.empty();
+    }
+
   private:
 
     UINT         const m_contextFlags;

src/d3d11/d3d11_context_imm.cpp

@@ -237,18 +237,29 @@ namespace dxvk {
     // number of pending draw calls is high enough.
     ConsiderFlush(GpuFlushType::ImplicitWeakHint);
 
-    // Dispatch command list to the CS thread and
-    // restore the immediate context's state
-    uint64_t csSeqNum = commandList->EmitToCsThread(&m_csThread);
-    m_csSeqNum = std::max(m_csSeqNum, csSeqNum);
-    
+    // Dispatch command list to the CS thread
+    commandList->EmitToCsThread([this] (DxvkCsChunkRef&& chunk) {
+      EmitCsChunk(std::move(chunk));
+
+      // Return the sequence number from before the flush since
+      // that is actually going to be needed for resource tracking
+      uint64_t csSeqNum = m_csSeqNum;
+
+      // Consider a flush after every chunk in case the app
+      // submits a very large command list or the GPU is idle
+      ConsiderFlush(GpuFlushType::ImplicitWeakHint);
+      return csSeqNum;
+    });
+
+    // If any resource tracking took place, flush with a strong hint
+    if (commandList->HasTrackedResources())
+      ConsiderFlush(GpuFlushType::ImplicitStrongHint);
+
+    // Restore the immediate context's state
     if (RestoreContextState)
       RestoreCommandListState();
     else
       ResetContextState();
-
-    // Flush again if the command list was sufficiently long
-    ConsiderFlush(GpuFlushType::ImplicitWeakHint);
   }