[d3d11] Consider flushing after each CS chunk

This way we will never end up with overly long command lists.

src/d3d11/d3d11_context.cpp

@@ -4918,8 +4918,10 @@ namespace dxvk {
     if (needsUpdate) {
       BindFramebuffer();
 
-      if constexpr (!IsDeferred)
+      if constexpr (!IsDeferred) {
+        // Doing this makes it less likely to flush during render passes
         GetTypedContext()->ConsiderFlush(GpuFlushType::ImplicitWeakHint);
+      }
     }
   }
 

src/d3d11/d3d11_context.h

@@ -10,6 +10,8 @@
 
 #include "../d3d10/d3d10_multithread.h"
 
+#include "../util/util_flush.h"
+
 #include "d3d11_annotation.h"
 #include "d3d11_buffer.h"
 #include "d3d11_cmd.h"
@@ -1064,27 +1066,35 @@ namespace dxvk {
             DxvkMultisampleState*             pMsState,
             UINT                              SampleMask);
 
-    template<typename Cmd>
+    template<bool AllowFlush = !IsDeferred, typename Cmd>
     void EmitCs(Cmd&& command) {
       m_cmdData = nullptr;
 
       if (unlikely(!m_csChunk->push(command))) {
         GetTypedContext()->EmitCsChunk(std::move(m_csChunk));
-        
         m_csChunk = AllocCsChunk();
+
+        if constexpr (AllowFlush)
+          GetTypedContext()->ConsiderFlush(GpuFlushType::ImplicitWeakHint);
+
         m_csChunk->push(command);
       }
     }
 
-    template<typename M, typename Cmd, typename... Args>
+    template<typename M, bool AllowFlush = !IsDeferred, typename Cmd, typename... Args>
     M* EmitCsCmd(Cmd&& command, Args&&... args) {
       M* data = m_csChunk->pushCmd<M, Cmd, Args...>(
         command, std::forward<Args>(args)...);
 
       if (unlikely(!data)) {
         GetTypedContext()->EmitCsChunk(std::move(m_csChunk));
-        
         m_csChunk = AllocCsChunk();
+
+        if constexpr (AllowFlush)
+          GetTypedContext()->ConsiderFlush(GpuFlushType::ImplicitWeakHint);
+
+        // We must record this command after the potential
+        // flush since the caller may still access the data
         data = m_csChunk->pushCmd<M, Cmd, Args...>(
           command, std::forward<Args>(args)...);
       }

src/d3d11/d3d11_context_imm.cpp

@@ -236,7 +236,7 @@ namespace dxvk {
     // As an optimization, flush everything if the
     // 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);
@@ -247,7 +247,7 @@ namespace dxvk {
     else
       ResetContextState();
 
-    // Flush after if the command list was sufficiently long
+    // Flush again if the command list was sufficiently long
     ConsiderFlush(GpuFlushType::ImplicitWeakHint);
   }
   
@@ -378,8 +378,6 @@ namespace dxvk {
       }
 
       if (doInvalidatePreserve) {
-        ConsiderFlush(GpuFlushType::ImplicitWeakHint);
-
         auto prevSlice = pResource->GetMappedSlice();
         auto physSlice = pResource->DiscardSlice();
 
@@ -525,8 +523,6 @@ namespace dxvk {
       }
 
       if (doFlags & DoInvalidate) {
-        ConsiderFlush(GpuFlushType::ImplicitWeakHint);
-
         DxvkBufferSliceHandle prevSlice = pResource->GetMappedSlice(Subresource);
         DxvkBufferSliceHandle physSlice = pResource->DiscardSlice(Subresource);
 
@@ -770,7 +766,7 @@ namespace dxvk {
   void D3D11ImmediateContext::EndFrame() {
     D3D10DeviceLock lock = LockContext();
 
-    EmitCs([] (DxvkContext* ctx) {
+    EmitCs<false>([] (DxvkContext* ctx) {
       ctx->endFrame();
     });
   }
@@ -887,7 +883,7 @@ namespace dxvk {
       });
     }
 
-    EmitCs([
+    EmitCs<false>([
       cSubmissionFence  = m_submissionFence,
       cSubmissionId     = submissionId
     ] (DxvkContext* ctx) {

src/d3d11/d3d11_context_imm.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include "../util/util_flush.h"
 #include "../util/util_time.h"
 
 #include "../util/sync/sync_signal.h"