Merge pull request #7953 from stenzek/lazy-d3d12-perf-queries

D3D12: Fix case where perf queries weren't flushed

Source/Core/VideoBackends/D3D12/PerfQuery.cpp

@@ -57,6 +57,12 @@ void PerfQuery::EnableQuery(PerfQueryGroup type)
     PartialFlush(do_resolve, blocking);
   }
 
+  // Ensure all state is applied before beginning the query.
+  // This is because we can't leave a query open when submitting a command list, and the draw
+  // call itself may need to execute a command list if we run out of descriptors. Note that
+  // this assumes that the caller has bound all required state prior to enabling the query.
+  Renderer::GetInstance()->ApplyState();
+
   if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
   {
     ActiveQuery& entry = m_query_buffer[m_query_next_pos];
@@ -152,9 +158,9 @@ void PerfQuery::ResolveQueries(u32 query_count)
   m_unresolved_queries -= query_count;
 }
 
-void PerfQuery::ReadbackQueries()
+void PerfQuery::ReadbackQueries(bool blocking)
 {
-  const u64 completed_fence_counter = g_dx_context->GetCompletedFenceValue();
+  u64 completed_fence_counter = g_dx_context->GetCompletedFenceValue();
 
   // Need to save these since ProcessResults will modify them.
   const u32 outstanding_queries = m_query_count;
@@ -163,13 +169,24 @@ void PerfQuery::ReadbackQueries()
   {
     u32 index = (m_query_readback_pos + readback_count) % PERF_QUERY_BUFFER_SIZE;
     const ActiveQuery& entry = m_query_buffer[index];
-    if (!entry.resolved || entry.fence_value > completed_fence_counter)
+    if (!entry.resolved)
       break;
 
+    if (entry.fence_value > completed_fence_counter)
+    {
+      // Query result isn't ready yet. Wait if blocking, otherwise we can't do any more yet.
+      if (!blocking)
+        break;
+
+      ASSERT(entry.fence_value != g_dx_context->GetCurrentFenceValue());
+      g_dx_context->WaitForFence(entry.fence_value);
+      completed_fence_counter = g_dx_context->GetCompletedFenceValue();
+    }
+
     // If this wrapped around, we need to flush the entries before the end of the buffer.
     if (index < m_query_readback_pos)
     {
-      ReadbackQueries(readback_count);
+      AccumulateQueriesFromBuffer(readback_count);
       DEBUG_ASSERT(m_query_readback_pos == 0);
       readback_count = 0;
     }
@@ -178,10 +195,10 @@ void PerfQuery::ReadbackQueries()
   }
 
   if (readback_count > 0)
-    ReadbackQueries(readback_count);
+    AccumulateQueriesFromBuffer(readback_count);
 }
 
-void PerfQuery::ReadbackQueries(u32 query_count)
+void PerfQuery::AccumulateQueriesFromBuffer(u32 query_count)
 {
   // Should be at maximum query_count queries pending.
   ASSERT(query_count <= m_query_count &&
@@ -226,10 +243,10 @@ void PerfQuery::ReadbackQueries(u32 query_count)
 
 void PerfQuery::PartialFlush(bool resolve, bool blocking)
 {
-  // Submit a command buffer in the background if the front query is not bound to one.
-  if ((resolve || blocking) && !m_query_buffer[m_query_resolve_pos].resolved)
-    Renderer::GetInstance()->ExecuteCommandList(blocking);
+  // Submit a command buffer if there are unresolved queries (to write them to the buffer).
+  if (resolve && m_unresolved_queries > 0)
+    Renderer::GetInstance()->ExecuteCommandList(false);
 
-  ReadbackQueries();
+  ReadbackQueries(blocking);
 }
 }  // namespace DX12

Source/Core/VideoBackends/D3D12/PerfQuery.h

@@ -38,8 +38,8 @@ private:
   };
 
   void ResolveQueries(u32 query_count);
-  void ReadbackQueries();
-  void ReadbackQueries(u32 query_count);
+  void ReadbackQueries(bool blocking);
+  void AccumulateQueriesFromBuffer(u32 query_count);
 
   void PartialFlush(bool resolve, bool blocking);
 

Source/Core/VideoBackends/D3D12/Renderer.h

@@ -89,6 +89,9 @@ public:
   void SetVertexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, u32 stride, u32 size);
   void SetIndexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, u32 size, DXGI_FORMAT format);
 
+  // Binds all dirty state
+  bool ApplyState();
+
 protected:
   void OnConfigChanged(u32 bits) override;
 
@@ -131,8 +134,6 @@ private:
 
   void CheckForSwapChainChanges();
 
-  // Binds all dirty state
-  bool ApplyState();
   void BindFramebuffer(DXFramebuffer* fb);
   void SetRootSignatures();
   void SetDescriptorHeaps();

Source/Core/VideoBackends/Vulkan/PerfQuery.cpp

@@ -45,6 +45,10 @@ void PerfQuery::EnableQuery(PerfQueryGroup type)
   if (m_query_count > m_query_buffer.size() / 2)
     PartialFlush(m_query_count == PERF_QUERY_BUFFER_SIZE);
 
+  // Ensure command buffer is ready to go before beginning the query, that way we don't submit
+  // a buffer with open queries.
+  StateTracker::GetInstance()->Bind();
+
   if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
   {
     ActiveQuery& entry = m_query_buffer[m_query_next_pos];

Source/Core/VideoCommon/PerfQueryBase.h

@@ -31,22 +31,31 @@ class PerfQueryBase
 public:
   PerfQueryBase() : m_query_count(0) {}
   virtual ~PerfQueryBase() {}
+
   // Checks if performance queries are enabled in the gameini configuration.
   // NOTE: Called from CPU+GPU thread
   static bool ShouldEmulate();
 
   // Begin querying the specified value for the following host GPU commands
+  // The call to EnableQuery() should be placed immediately before the draw command, otherwise
+  // there is a risk of GPU resets if the query is left open and the buffer is submitted during
+  // resource binding (D3D12/Vulkan).
   virtual void EnableQuery(PerfQueryGroup type) {}
+
   // Stop querying the specified value for the following host GPU commands
   virtual void DisableQuery(PerfQueryGroup type) {}
+
   // Reset query counters to zero and drop any pending queries
   virtual void ResetQuery() {}
+
   // Return the measured value for the specified query type
   // NOTE: Called from CPU thread
   virtual u32 GetQueryResult(PerfQueryType type) { return 0; }
+
   // Request the value of any pending queries - causes a pipeline flush and thus should be used
   // carefully!
   virtual void FlushResults() {}
+
   // True if there are no further pending query results
   // NOTE: Called from CPU thread
   virtual bool IsFlushed() const { return true; }