[d3d11] Do not synchronize with CS thread in GetData

Instead, use a counter to determine whether there are any
pending operation for the query on the CS thread.

src/d3d11/d3d11_context_imm.cpp

@@ -77,10 +77,6 @@ namespace dxvk {
     // DataSize is 0, but we should ignore that pointer
     pData = DataSize ? pData : nullptr;
 
-    // Ensure that all query commands actually get
-    // executed before trying to access the query
-    SynchronizeCsThread();
-
     // Get query status directly from the query object
     auto query = static_cast<D3D11Query*>(pAsync);
     HRESULT hr = query->GetData(pData, GetDataFlags);

src/d3d11/d3d11_query.cpp

@@ -214,6 +214,8 @@ namespace dxvk {
 
     if (unlikely(m_predicate != nullptr))
       ctx->writePredicate(DxvkBufferSlice(m_predicate), m_query[0]);
+
+    m_resetCtr -= 1;
   }
   
   
@@ -230,6 +232,7 @@ namespace dxvk {
       return false;
 
     m_state = D3D11_VK_QUERY_ENDED;
+    m_resetCtr += 1;
     return true;
   }
 
@@ -237,6 +240,12 @@ namespace dxvk {
   HRESULT STDMETHODCALLTYPE D3D11Query::GetData(
           void*                             pData,
           UINT                              GetDataFlags) {
+    if (m_state != D3D11_VK_QUERY_ENDED)
+      return DXGI_ERROR_INVALID_CALL;
+
+    if (m_resetCtr != 0u)
+      return S_FALSE;
+
     if (m_desc.Query == D3D11_QUERY_EVENT) {
       DxvkGpuEventStatus status = m_event[0]->test();
 

src/d3d11/d3d11_query.h

@@ -114,6 +114,8 @@ namespace dxvk {
     uint32_t m_stallMask = 0;
     bool     m_stallFlag = false;
 
+    std::atomic<uint32_t> m_resetCtr = { 0u };
+
     UINT64 GetTimestampQueryFrequency() const;
 
     Rc<DxvkBuffer> CreatePredicateBuffer();