[VK]Improve query reset before use so that now queries are reset outside render pass and are batched

Note to testers, please enable validation layers and see if there are any errors related to queries not being reset before every use, Thanks.

src/video_core/query_cache/query_cache.h

@@ -27,7 +27,7 @@
 #include "video_core/query_cache/query_cache_base.h"
 #include "video_core/query_cache/query_stream.h"
 #include "video_core/query_cache/types.h"
-
+namespace Vulkan { class Scheduler; }
 namespace VideoCommon {
 
 using Maxwell = Tegra::Engines::Maxwell3D;
@@ -222,6 +222,12 @@ void QueryCacheBase<Traits>::CounterReset(QueryType counter_type) {
     streamer->ResetCounter();
 }
 
+// Called at frame start to batch vkCmdResetQueryPool outside render passes.
+template <typename Traits>
+void QueryCacheBase<Traits>::FramePrologueResets(Vulkan::Scheduler& scheduler) {
+        impl->runtime.FramePrologueResets(scheduler);
+}
+
 template <typename Traits>
 void QueryCacheBase<Traits>::BindToChannel(s32 id) {
     VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo>::BindToChannel(id);

src/video_core/query_cache/query_cache_base.h

@@ -21,7 +21,9 @@
 namespace VideoCore {
 class RasterizerInterface;
 }
-
+namespace Vulkan {
+    class Scheduler;
+}
 namespace Tegra {
 class GPU;
 }
@@ -53,7 +55,7 @@ public:
                             RuntimeType& runtime_);
 
     ~QueryCacheBase();
-
+    void FramePrologueResets(Vulkan::Scheduler& scheduler);
     void InvalidateRegion(VAddr addr, std::size_t size) {
         IterateCache<true>(addr, size,
                            [this](QueryLocation location) { InvalidateQuery(location); });

src/video_core/renderer_vulkan/vk_query_cache.cpp

@@ -156,7 +156,7 @@ public:
 
         ReserveHostQuery();
 
-        // Ensure outside render pass
+        /* Ensure outside render pass
         scheduler.RequestOutsideRenderPassOperationContext();
 
         // Reset query pool outside render pass
@@ -167,7 +167,7 @@ public:
 
         // Manually restart the render pass (required for vkCmdClearAttachments, etc.)
         scheduler.RequestRenderpass(texture_cache.GetFramebuffer());
-
+        */
         // Begin query inside the newly started render pass
         scheduler.Record([query_pool = current_query_pool,
                                  query_index = current_bank_slot](vk::CommandBuffer cmdbuf) {
@@ -402,6 +402,13 @@ public:
         }
     }
 
+    VkQueryPool GetOrCreateCurrentPoolForPrologue() {
+               if (!current_bank || current_bank->IsClosed()) {
+                        ReserveBank();
+                    }
+                return current_query_pool;
+    }
+
 private:
     template <typename Func>
     void ApplyBankOp(VideoCommon::HostQueryBase* query, Func&& func) {
@@ -1305,6 +1312,19 @@ void QueryCacheRuntime::ResumeHostConditionalRendering() {
     impl->is_hcr_running = true;
 }
 
+void QueryCacheRuntime::FramePrologueResets(Scheduler& scheduler) {
+            // Reset the occlusion queries we plan to use this frame in one go.
+                    // Ensure this is recorded OUTSIDE any render pass.
+                            const VkQueryPool pool = impl->sample_streamer.GetOrCreateCurrentPoolForPrologue();
+           scheduler.RequestOutsideRenderPassOperationContext();
+           scheduler.Record([pool](vk::CommandBuffer cmdbuf) {
+            // Reset the whole bank so subsequent BeginQuery calls don't need per-slot resets.
+            cmdbuf.ResetQueryPool(pool,
+                                  /*first*/ 0,
+                                  /*count*/ SamplesQueryBank::BANK_SIZE);
+            });
+}
+
 void QueryCacheRuntime::HostConditionalRenderingCompareValueImpl(VideoCommon::LookupData object,
                                                                  bool is_equal) {
     {

src/video_core/renderer_vulkan/vk_query_cache.h

@@ -20,7 +20,7 @@ namespace Vulkan {
 class Device;
 class Scheduler;
 class StagingBufferPool;
-
+class Scheduler;
 struct QueryCacheRuntimeImpl;
 
 class QueryCacheRuntime {
@@ -33,7 +33,7 @@ public:
                                ComputePassDescriptorQueue& compute_pass_descriptor_queue,
                                DescriptorPool& descriptor_pool, TextureCache& texture_cache_);
     ~QueryCacheRuntime();
-
+    void FramePrologueResets(Scheduler& scheduler);
     template <typename SyncValuesType>
     void SyncValues(std::span<SyncValuesType> values, VkBuffer base_src_buffer = nullptr);
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -257,6 +257,11 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
 
 void Scheduler::AllocateNewContext() {
     // Enable counters once again. These are disabled when a command buffer is finished.
+    // Record per frame query resets outside any render pass, before the first draw.
+    if (query_cache) {
+        query_cache->FramePrologueResets(*this);
+    }
+
 }
 
 void Scheduler::InvalidateState() {