[dxvk] Accumulate query data into query objectquery-rework

And do so when adding additional query handles, in order
to avoid allocating queries indefinitely if End is never
called, which Halo:MCC supposedly does.

Co-authored-by: Sam Edwards <CFSworks@gmail.com>

3 files changed, 77 insertions, 32 deletions

diff --git a/src/dxvk/dxvk_gpu_query.cpp b/src/dxvk/dxvk_gpu_query.cpp
index 7ceb32d1..aaec6a0f 100644
--- a/src/dxvk/dxvk_gpu_query.cpp
+++ b/src/dxvk/dxvk_gpu_query.cpp
@@ -28,7 +28,7 @@ namespace dxvk {
   }
 
 
-  DxvkGpuQueryStatus DxvkGpuQuery::getData(DxvkQueryData& queryData) const {
+  DxvkGpuQueryStatus DxvkGpuQuery::getData(DxvkQueryData& queryData) {
     queryData = DxvkQueryData();
 
     // Callers must ensure that no begin call is pending when
@@ -37,21 +37,21 @@ namespace dxvk {
     if (!m_ended.load(std::memory_order_acquire))
       return DxvkGpuQueryStatus::Invalid;
 
-    // Get query data from all associated handles
-    DxvkGpuQueryStatus status = DxvkGpuQueryStatus::Available;
-
-    for (size_t i = 0; i < m_handles.size()
-        && status == DxvkGpuQueryStatus::Available; i++)
-      status = getDataForHandle(queryData, m_handles[i]);
+    // Accumulate query data from all available queries
+    DxvkGpuQueryStatus status = this->accumulateQueryData();
     
     // Treat non-precise occlusion queries as available
     // if we already know the result will be non-zero
     if ((status == DxvkGpuQueryStatus::Pending)
      && (m_type == VK_QUERY_TYPE_OCCLUSION)
      && !(m_flags & VK_QUERY_CONTROL_PRECISE_BIT)
-     && (queryData.occlusion.samplesPassed))
+     && (m_queryData.occlusion.samplesPassed))
       status = DxvkGpuQueryStatus::Available;
-    
+
+    // Write back accumulated query data if the result is useful
+    if (status == DxvkGpuQueryStatus::Available)
+      queryData = m_queryData;
+
     return status;
   }
 
@@ -61,10 +61,15 @@ namespace dxvk {
     // only the false->true transition is defined.
     m_ended.store(false, std::memory_order_relaxed);
 
+    // Ideally we should have no queries left at this point,
+    // if we do, lifetime-track them with the command list.
     for (size_t i = 0; i < m_handles.size(); i++)
       cmd->trackGpuQuery(m_handles[i]);
 
     m_handles.clear();
+
+    // Reset accumulated query data
+    m_queryData = DxvkQueryData();
   }
 
   
@@ -75,14 +80,18 @@ namespace dxvk {
 
 
   void DxvkGpuQuery::addQueryHandle(const DxvkGpuQueryHandle& handle) {
+    // Already accumulate available queries here in case
+    // we already allocated a large number of queries
+    if (m_handles.size() >= m_handles.MinCapacity)
+      this->accumulateQueryData();
+
     m_handles.push_back(handle);
   }
 
 
-  DxvkGpuQueryStatus DxvkGpuQuery::getDataForHandle(
-          DxvkQueryData&      queryData,
-    const DxvkGpuQueryHandle& handle) const {
-    DxvkQueryData tmpData;
+  DxvkGpuQueryStatus DxvkGpuQuery::accumulateQueryDataForHandle(
+    const DxvkGpuQueryHandle& handle) {
+    DxvkQueryData tmpData = { };
 
     // Try to copy query data to temporary structure
     VkResult result = m_vkd->vkGetQueryPoolResults(m_vkd->device(),
@@ -98,30 +107,30 @@ namespace dxvk {
     // Add numbers to the destination structure
     switch (m_type) {
       case VK_QUERY_TYPE_OCCLUSION:
-        queryData.occlusion.samplesPassed += tmpData.occlusion.samplesPassed;
+        m_queryData.occlusion.samplesPassed += tmpData.occlusion.samplesPassed;
         break;
       
       case VK_QUERY_TYPE_TIMESTAMP:
-        queryData.timestamp.time = tmpData.timestamp.time;
+        m_queryData.timestamp.time = tmpData.timestamp.time;
         break;
       
       case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-        queryData.statistic.iaVertices       += tmpData.statistic.iaVertices;
-        queryData.statistic.iaPrimitives     += tmpData.statistic.iaPrimitives;
-        queryData.statistic.vsInvocations    += tmpData.statistic.vsInvocations;
-        queryData.statistic.gsInvocations    += tmpData.statistic.gsInvocations;
-        queryData.statistic.gsPrimitives     += tmpData.statistic.gsPrimitives;
-        queryData.statistic.clipInvocations  += tmpData.statistic.clipInvocations;
-        queryData.statistic.clipPrimitives   += tmpData.statistic.clipPrimitives;
-        queryData.statistic.fsInvocations    += tmpData.statistic.fsInvocations;
-        queryData.statistic.tcsPatches       += tmpData.statistic.tcsPatches;
-        queryData.statistic.tesInvocations   += tmpData.statistic.tesInvocations;
-        queryData.statistic.csInvocations    += tmpData.statistic.csInvocations;
+        m_queryData.statistic.iaVertices       += tmpData.statistic.iaVertices;
+        m_queryData.statistic.iaPrimitives     += tmpData.statistic.iaPrimitives;
+        m_queryData.statistic.vsInvocations    += tmpData.statistic.vsInvocations;
+        m_queryData.statistic.gsInvocations    += tmpData.statistic.gsInvocations;
+        m_queryData.statistic.gsPrimitives     += tmpData.statistic.gsPrimitives;
+        m_queryData.statistic.clipInvocations  += tmpData.statistic.clipInvocations;
+        m_queryData.statistic.clipPrimitives   += tmpData.statistic.clipPrimitives;
+        m_queryData.statistic.fsInvocations    += tmpData.statistic.fsInvocations;
+        m_queryData.statistic.tcsPatches       += tmpData.statistic.tcsPatches;
+        m_queryData.statistic.tesInvocations   += tmpData.statistic.tesInvocations;
+        m_queryData.statistic.csInvocations    += tmpData.statistic.csInvocations;
         break;
       
       case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-        queryData.xfbStream.primitivesWritten += tmpData.xfbStream.primitivesWritten;
-        queryData.xfbStream.primitivesNeeded  += tmpData.xfbStream.primitivesNeeded;
+        m_queryData.xfbStream.primitivesWritten += tmpData.xfbStream.primitivesWritten;
+        m_queryData.xfbStream.primitivesNeeded  += tmpData.xfbStream.primitivesNeeded;
         break;
       
       default:
@@ -131,6 +140,37 @@ namespace dxvk {
     
     return DxvkGpuQueryStatus::Available;
   }
+
+
+  DxvkGpuQueryStatus DxvkGpuQuery::accumulateQueryData() {
+    DxvkGpuQueryStatus status = DxvkGpuQueryStatus::Available;
+
+    // Process available queries and return them to the
+    // allocator if possible. This may help reduce the
+    // number of Vulkan queries in flight.
+    size_t queriesAvailable = 0;
+
+    while (queriesAvailable < m_handles.size()) {
+      status = this->accumulateQueryDataForHandle(m_handles[queriesAvailable]);
+
+      if (status != DxvkGpuQueryStatus::Available)
+        break;
+
+      queriesAvailable += 1;
+    }
+
+    if (queriesAvailable) {
+      for (size_t i = 0; i < queriesAvailable; i++)
+        m_handles[i].allocator->freeQuery(m_handles[i]);
+
+      for (size_t i = queriesAvailable; i < m_handles.size(); i++)
+        m_handles[i - queriesAvailable] = m_handles[i];
+
+      m_handles.resize(m_handles.size() - queriesAvailable);
+    }
+
+    return status;
+  }
   
   
   
diff --git a/src/dxvk/dxvk_gpu_query.h b/src/dxvk/dxvk_gpu_query.h
index f2318563..919d0e26 100644
--- a/src/dxvk/dxvk_gpu_query.h
+++ b/src/dxvk/dxvk_gpu_query.h
@@ -188,7 +188,7 @@ namespace dxvk {
      * \returns Current query status
      */
     DxvkGpuQueryStatus getData(
-            DxvkQueryData&      queryData) const;
+            DxvkQueryData&      queryData);
 
     /**
      * \brief Begins query
@@ -230,11 +230,14 @@ namespace dxvk {
     uint32_t            m_index;
     std::atomic<bool>   m_ended;
 
+    DxvkQueryData       m_queryData = { };
+
     small_vector<DxvkGpuQueryHandle, 8> m_handles;
     
-    DxvkGpuQueryStatus getDataForHandle(
-            DxvkQueryData&      queryData,
-      const DxvkGpuQueryHandle& handle) const;
+    DxvkGpuQueryStatus accumulateQueryDataForHandle(
+      const DxvkGpuQueryHandle& handle);
+
+    DxvkGpuQueryStatus accumulateQueryData();
 
   };
 
diff --git a/src/util/util_small_vector.h b/src/util/util_small_vector.h
index 13ff1ac7..48fe8bff 100644
--- a/src/util/util_small_vector.h
+++ b/src/util/util_small_vector.h
@@ -9,6 +9,8 @@ namespace dxvk {
     using storage = std::aligned_storage_t<sizeof(T), alignof(T)>;
   public:
 
+    constexpr static size_t MinCapacity = N;
+
     small_vector() { }
 
     small_vector             (const small_vector&) = delete;