[vulkan] Another approach for queries reset

src/video_core/renderer_vulkan/vk_query_cache.cpp

@@ -41,8 +41,8 @@ class SamplesQueryBank : public VideoCommon::BankBase {
 public:
     static constexpr size_t BANK_SIZE = 256;
     static constexpr size_t QUERY_SIZE = 8;
-    explicit SamplesQueryBank(const Device& device_, size_t index_)
-        : BankBase(BANK_SIZE), device{device_}, index{index_} {
+    explicit SamplesQueryBank(const Device& device_, Scheduler& scheduler_, size_t index_)
+        : BankBase(BANK_SIZE), device{device_}, scheduler{scheduler_}, index{index_} {
         const auto& dev = device.GetLogical();
         query_pool = dev.CreateQueryPool({
             .sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
@@ -60,12 +60,28 @@ public:
     void Reset() override {
         ASSERT(references == 0);
         VideoCommon::BankBase::Reset();
-        const auto& dev = device.GetLogical();
-        dev.ResetQueryPool(*query_pool, 0, BANK_SIZE);
+        if (device.IsHostQueryResetSupported()) {
+            const auto& dev = device.GetLogical();
+            dev.ResetQueryPool(*query_pool, 0, BANK_SIZE);
+        } else {
+            scheduler.RequestOutsideRenderPassOperationContext();
+            scheduler.Record([pool = *query_pool](vk::CommandBuffer cmdbuf) {
+                cmdbuf.ResetQueryPool(pool, 0, BANK_SIZE);
+            });
+        }
         host_results.fill(0ULL);
         next_bank = 0;
     }
 
+    void AddReference(size_t how_many = 1) {
+        BankBase::AddReference(how_many);
+        last_used_tick = scheduler.CurrentTick();
+    }
+
+    [[nodiscard]] bool IsDead() const {
+        return BankBase::IsDead() && scheduler.IsFree(last_used_tick);
+    }
+
     void Sync(size_t start, size_t size) {
         const auto& dev = device.GetLogical();
         const VkResult query_result = dev.GetQueryResults(
@@ -98,9 +114,11 @@ public:
 
 private:
     const Device& device;
+    Scheduler& scheduler;
     const size_t index;
     vk::QueryPool query_pool;
     std::array<u64, BANK_SIZE> host_results;
+    u64 last_used_tick{};
 };
 
 using BaseStreamer = VideoCommon::SimpleStreamer<VideoCommon::HostQueryBase>;
@@ -431,7 +449,7 @@ private:
     void ReserveBank() {
         current_bank_id =
             bank_pool.ReserveBank([this](std::deque<SamplesQueryBank>& queue, size_t index) {
-                queue.emplace_back(device, index);
+                queue.emplace_back(device, scheduler, index);
             });
         if (current_bank) {
             current_bank->next_bank = current_bank_id + 1;
@@ -620,6 +638,15 @@ public:
         VideoCommon::BankBase::Reset();
     }
 
+    void AddReference(size_t how_many = 1) {
+        BankBase::AddReference(how_many);
+        last_used_tick = scheduler.CurrentTick();
+    }
+
+    [[nodiscard]] bool IsDead() const {
+        return BankBase::IsDead() && scheduler.IsFree(last_used_tick);
+    }
+
     void Sync(StagingBufferRef& stagging_buffer, size_t extra_offset, size_t start, size_t size) {
         scheduler.RequestOutsideRenderPassOperationContext();
         scheduler.Record([this, dst_buffer = stagging_buffer.buffer, extra_offset, start,
@@ -645,6 +672,7 @@ private:
     Scheduler& scheduler;
     const size_t index;
     vk::Buffer buffer;
+    u64 last_used_tick{};
 };
 
 class PrimitivesSucceededStreamer;

src/video_core/vulkan_common/vulkan_device.h

@@ -523,7 +523,6 @@ FN_MAX_LIMIT_LIST
     }
 
     /// Returns true if the device supports VK_EXT_shader_stencil_export.
-    /// Note: Most Mali/NVIDIA drivers don't support this. Use hardware blits as fallback.
     bool IsExtShaderStencilExportSupported() const {
         return extensions.shader_stencil_export;
     }
@@ -560,6 +559,11 @@ FN_MAX_LIMIT_LIST
         return extensions.subgroup_size_control;
     }
 
+    /// Returns true if vkResetQueryPool (host-side query reset) is supported.
+    bool IsHostQueryResetSupported() const {
+        return features.host_query_reset.hostQueryReset != VK_FALSE;
+    }
+
     /// Returns true if the device supports VK_EXT_transform_feedback.
     bool IsExtTransformFeedbackSupported() const {
         return extensions.transform_feedback;

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -165,6 +165,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdSetColorWriteEnableEXT);
     X(vkCmdSetColorBlendEnableEXT);
     X(vkCmdSetColorBlendEquationEXT);
+    X(vkCmdResetQueryPool);
     X(vkCmdResolveImage);
     X(vkCreateBuffer);
     X(vkCreateBufferView);

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -241,6 +241,7 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkCmdPipelineBarrier2 vkCmdPipelineBarrier2{};
     PFN_vkCmdPushConstants vkCmdPushConstants{};
     PFN_vkCmdPushDescriptorSetWithTemplateKHR vkCmdPushDescriptorSetWithTemplateKHR{};
+    PFN_vkCmdResetQueryPool vkCmdResetQueryPool{};
     PFN_vkCmdResolveImage vkCmdResolveImage{};
     PFN_vkCmdSetBlendConstants vkCmdSetBlendConstants{};
     PFN_vkCmdSetCullModeEXT vkCmdSetCullModeEXT{};
@@ -1191,6 +1192,10 @@ public:
         dld->vkCmdEndQuery(handle, query_pool, query);
     }
 
+    void ResetQueryPool(VkQueryPool query_pool, u32 first, u32 count) const noexcept {
+        dld->vkCmdResetQueryPool(handle, query_pool, first, count);
+    }
+
     void BindDescriptorSets(VkPipelineBindPoint bind_point, VkPipelineLayout layout, u32 first,
                             Span<VkDescriptorSet> sets, Span<u32> dynamic_offsets) const noexcept {
         dld->vkCmdBindDescriptorSets(handle, bind_point, layout, first, sets.size(), sets.data(),