Fix timer queries reset before use on Vulkan

Source/Engine/GraphicsDevice/Vulkan/Config.h

@@ -28,7 +28,7 @@
 #define VULKAN_RESOURCE_DELETE_SAFE_FRAMES_COUNT 20
 
 #define VULKAN_ENABLE_API_DUMP 0
-#define VULKAN_RESET_QUERY_POOLS 0
+#define VULKAN_RESET_QUERY_POOLS 1
 #define VULKAN_HASH_POOLS_WITH_LAYOUT_TYPES 1
 #define VULKAN_USE_DEBUG_LAYER GPU_ENABLE_DIAGNOSTICS
 #define VULKAN_USE_DEBUG_DATA (GPU_ENABLE_DIAGNOSTICS && COMPILE_WITH_DEV_ENV)

Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp

@@ -594,8 +594,6 @@ RenderPassVulkan::~RenderPassVulkan()
 QueryPoolVulkan::QueryPoolVulkan(GPUDeviceVulkan* device, int32 capacity, VkQueryType type)
     : _device(device)
     , _handle(VK_NULL_HANDLE)
-    , _count(0)
-    , _capacity(capacity)
     , _type(type)
 {
     VkQueryPoolCreateInfo createInfo;
@@ -603,9 +601,11 @@ QueryPoolVulkan::QueryPoolVulkan(GPUDeviceVulkan* device, int32 capacity, VkQuer
     createInfo.queryType = type;
     createInfo.queryCount = capacity;
     VALIDATE_VULKAN_RESULT(vkCreateQueryPool(device->Device, &createInfo, nullptr, &_handle));
+
 #if VULKAN_RESET_QUERY_POOLS
+    // New queries have to be reset before use
+    ResetBeforeUse = true;
     _resetRanges.Add(Range{ 0, static_cast<uint32>(capacity) });
-    device->QueriesToReset.Add(this);
 #endif
 }
 
@@ -626,6 +626,7 @@ void QueryPoolVulkan::Reset(CmdBufferVulkan* cmdBuffer)
         vkCmdResetQueryPool(cmdBuffer->GetHandle(), _handle, range.Start, range.Count);
     }
     _resetRanges.Clear();
+    ResetBeforeUse = false;
 }
 
 #endif
@@ -640,9 +641,9 @@ BufferedQueryPoolVulkan::BufferedQueryPoolVulkan(GPUDeviceVulkan* device, int32
     _readResultsBits.AddZeroed((capacity + 63) / 64);
 }
 
-bool BufferedQueryPoolVulkan::AcquireQuery(uint32& resultIndex)
+bool BufferedQueryPoolVulkan::AcquireQuery(CmdBufferVulkan* cmdBuffer, uint32& resultIndex)
 {
-    const uint64 allUsedMask = (uint64)-1;
+    const uint64 allUsedMask = MAX_uint64;
     for (int32 wordIndex = _lastBeginIndex / 64; wordIndex < _usedQueryBits.Count(); wordIndex++)
     {
         uint64 beginQueryWord = _usedQueryBits[wordIndex];
@@ -659,10 +660,11 @@ bool BufferedQueryPoolVulkan::AcquireQuery(uint32& resultIndex)
             _usedQueryBits[wordIndex] = _usedQueryBits[wordIndex] | bit;
             _readResultsBits[wordIndex] &= ~bit;
             _lastBeginIndex = resultIndex + 1;
+            if (ResetBeforeUse)
+                Reset(cmdBuffer);
             return true;
         }
     }
-
     return false;
 }
 
@@ -675,7 +677,7 @@ void BufferedQueryPoolVulkan::ReleaseQuery(uint32 queryIndex)
     if (queryIndex < (uint32)_lastBeginIndex)
     {
         // Use the lowest word available
-        const uint64 allUsedMask = (uint64)-1;
+        const uint64 allUsedMask = MAX_uint64;
         const int32 lastQueryWord = _lastBeginIndex / 64;
         if (lastQueryWord < _usedQueryBits.Count() && _usedQueryBits[lastQueryWord] == allUsedMask)
         {
@@ -736,7 +738,7 @@ bool BufferedQueryPoolVulkan::GetResults(GPUContextVulkan* context, uint32 index
 
 bool BufferedQueryPoolVulkan::HasRoom() const
 {
-    const uint64 allUsedMask = (uint64)-1;
+    const uint64 allUsedMask = MAX_uint64;
     if (_lastBeginIndex < _usedQueryBits.Count() * 64)
     {
         ASSERT((_usedQueryBits[_lastBeginIndex / 64] & allUsedMask) != allUsedMask);

Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.h

@@ -259,8 +259,6 @@ protected:
     GPUDeviceVulkan* _device;
     VkQueryPool _handle;
 
-    volatile int32 _count;
-    const uint32 _capacity;
     const VkQueryType _type;
 #if VULKAN_RESET_QUERY_POOLS
     Array<Range> _resetRanges;
@@ -277,6 +275,7 @@ public:
     }
 
 #if VULKAN_RESET_QUERY_POOLS
+    bool ResetBeforeUse;
     void Reset(CmdBufferVulkan* cmdBuffer);
 #endif
 };
@@ -294,7 +293,7 @@ private:
 
 public:
     BufferedQueryPoolVulkan(GPUDeviceVulkan* device, int32 capacity, VkQueryType type);
-    bool AcquireQuery(uint32& resultIndex);
+    bool AcquireQuery(CmdBufferVulkan* cmdBuffer, uint32& resultIndex);
     void ReleaseQuery(uint32 queryIndex);
     void MarkQueryAsStarted(uint32 queryIndex);
     bool GetResults(GPUContextVulkan* context, uint32 index, uint64& result);

Source/Engine/GraphicsDevice/Vulkan/GPUTimerQueryVulkan.cpp

@@ -60,13 +60,18 @@ void GPUTimerQueryVulkan::WriteTimestamp(CmdBufferVulkan* cmdBuffer, Query& quer
 {
     auto pool = _device->FindAvailableTimestampQueryPool();
     uint32 index;
-    pool->AcquireQuery(index);
+    if (pool->AcquireQuery(cmdBuffer, index))
-
+    {
-    vkCmdWriteTimestamp(cmdBuffer->GetHandle(), stage, pool->GetHandle(), index);
+        vkCmdWriteTimestamp(cmdBuffer->GetHandle(), stage, pool->GetHandle(), index);
-    pool->MarkQueryAsStarted(index);
+        pool->MarkQueryAsStarted(index);
-
+        query.Pool = pool;
-    query.Pool = pool;
+        query.Index = index;
-    query.Index = index;
+    }
+    else
+    {
+        query.Pool = nullptr;
+        query.Index = 0;
+    }
 }
 
 bool GPUTimerQueryVulkan::TryGetResult()
@@ -104,16 +109,10 @@ bool GPUTimerQueryVulkan::TryGetResult()
     for (int32 i = 0; i < _queries.Count(); i++)
     {
         auto& e = _queries[i];
-
         if (e.Begin.Pool)
-        {
             e.Begin.Pool->ReleaseQuery(e.Begin.Index);
-        }
-
         if (e.End.Pool)
-        {
             e.End.Pool->ReleaseQuery(e.End.Index);
-        }
     }
     _queries.Clear();
 #else
@@ -141,16 +140,10 @@ void GPUTimerQueryVulkan::OnReleaseGPU()
     for (int32 i = 0; i < _queries.Count(); i++)
     {
         auto& e = _queries[i];
-
         if (e.Begin.Pool)
-        {
             e.Begin.Pool->ReleaseQuery(e.Begin.Index);
-        }
-
         if (e.End.Pool)
-        {
             e.End.Pool->ReleaseQuery(e.End.Index);
-        }
     }
     _queries.Clear();
 }
@@ -208,7 +201,6 @@ bool GPUTimerQueryVulkan::HasResult()
         return false;
     if (_hasResult)
         return true;
-
     return TryGetResult();
 }
 
@@ -216,7 +208,6 @@ float GPUTimerQueryVulkan::GetResult()
 {
     if (_hasResult)
         return _timeDelta;
-
     TryGetResult();
     return _timeDelta;
 }