[dxvk] Use vkResetQueryPoolEXT to reset individual queries

This is much faster than the fallback path which uses GPU functions.

src/dxvk/dxvk_cmdlist.h

@@ -546,6 +546,10 @@ namespace dxvk {
             VkQueryPool             queryPool,
             uint32_t                queryId,
             VkEvent                 event) {
+      if (event == VK_NULL_HANDLE) {
+        m_vkd->vkResetQueryPoolEXT(
+          m_vkd->device(), queryPool, queryId, 1);
+      } else {
         m_cmdBuffersUsed.set(DxvkCmdBufferFlag::InitBuffer);
 
         m_vkd->vkResetEvent(
@@ -557,6 +561,7 @@ namespace dxvk {
         m_vkd->vkCmdSetEvent(m_initBuffer,
           event, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
       }
+    }
     
     
     void cmdResetQueryPool(

src/dxvk/dxvk_device.cpp

@@ -20,7 +20,7 @@ namespace dxvk {
     m_renderPassPool    (new DxvkRenderPassPool     (vkd)),
     m_pipelineManager   (new DxvkPipelineManager    (this, m_renderPassPool.ptr())),
     m_gpuEventPool      (new DxvkGpuEventPool       (vkd)),
-    m_gpuQueryPool      (new DxvkGpuQueryPool       (vkd)),
+    m_gpuQueryPool      (new DxvkGpuQueryPool       (this)),
     m_metaClearObjects  (new DxvkMetaClearObjects   (vkd)),
     m_metaCopyObjects   (new DxvkMetaCopyObjects    (vkd)),
     m_metaMipGenObjects (new DxvkMetaMipGenObjects  (vkd)),

src/dxvk/dxvk_gpu_query.cpp

@@ -1,6 +1,7 @@
 #include <algorithm>
 
 #include "dxvk_cmdlist.h"
+#include "dxvk_device.h"
 #include "dxvk_gpu_query.h"
 
 namespace dxvk {
@@ -90,13 +91,17 @@ namespace dxvk {
     DxvkQueryData tmpData;
 
     // Wait for the query to be reset first
-    VkResult result = m_vkd->vkGetEventStatus(
+    VkResult result;
+    
+    if (handle.resetEvent) {
+      result = m_vkd->vkGetEventStatus(
         m_vkd->device(), handle.resetEvent);
     
       if (result == VK_EVENT_RESET)
         return DxvkGpuQueryStatus::Pending;
       else if (result != VK_EVENT_SET)
         return DxvkGpuQueryStatus::Failed;
+    }
     
     // Try to copy query data to temporary structure
     result = m_vkd->vkGetQueryPoolResults(m_vkd->device(),
@@ -150,10 +155,11 @@ namespace dxvk {
   
   
   DxvkGpuQueryAllocator::DxvkGpuQueryAllocator(
-    const Rc<vk::DeviceFn>&   vkd,
+    const Rc<DxvkDevice>&     device,
           VkQueryType         queryType,
           uint32_t            queryPoolSize)
-  : m_vkd           (vkd),
+  : m_device        (device),
+    m_vkd           (device->vkd()),
     m_queryType     (queryType),
     m_queryPoolSize (queryPoolSize) {
 
@@ -235,7 +241,8 @@ namespace dxvk {
     for (uint32_t i = 0; i < m_queryPoolSize; i++) {
       VkEvent event = VK_NULL_HANDLE;
 
-      if (m_vkd->vkCreateEvent(m_vkd->device(), &eventInfo, nullptr, &event)) {
+      if (!m_device->features().extHostQueryReset.hostQueryReset
+       && m_vkd->vkCreateEvent(m_vkd->device(), &eventInfo, nullptr, &event) != VK_SUCCESS) {
         Logger::err("DXVK: Failed to create query reset event");
         return;
       }
@@ -247,11 +254,11 @@ namespace dxvk {
 
 
 
-  DxvkGpuQueryPool::DxvkGpuQueryPool(const Rc<vk::DeviceFn>& vkd)
+  DxvkGpuQueryPool::DxvkGpuQueryPool(const Rc<DxvkDevice>& device)
-  : m_occlusion(vkd, VK_QUERY_TYPE_OCCLUSION,                     256),
+  : m_occlusion(device, VK_QUERY_TYPE_OCCLUSION,                     256),
-    m_statistic(vkd, VK_QUERY_TYPE_PIPELINE_STATISTICS,           64),
+    m_statistic(device, VK_QUERY_TYPE_PIPELINE_STATISTICS,           64),
-    m_timestamp(vkd, VK_QUERY_TYPE_TIMESTAMP,                     64),
+    m_timestamp(device, VK_QUERY_TYPE_TIMESTAMP,                     64),
-    m_xfbStream(vkd, VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT, 64) {
+    m_xfbStream(device, VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT, 64) {
     
   }
   

src/dxvk/dxvk_gpu_query.h

@@ -247,7 +247,7 @@ namespace dxvk {
   public:
 
     DxvkGpuQueryAllocator(
-      const Rc<vk::DeviceFn>&   vkd,
+      const Rc<DxvkDevice>&     device,
             VkQueryType         queryType,
             uint32_t            queryPoolSize);
     
@@ -275,6 +275,7 @@ namespace dxvk {
 
   private:
 
+    Rc<DxvkDevice>    m_device;
     Rc<vk::DeviceFn>  m_vkd;
     VkQueryType       m_queryType;
     uint32_t          m_queryPoolSize;
@@ -299,7 +300,7 @@ namespace dxvk {
   public:
 
     DxvkGpuQueryPool(
-      const Rc<vk::DeviceFn>&   vkd);
+      const Rc<DxvkDevice>& device);
     
     ~DxvkGpuQueryPool();