Query Cache: address issues

21 files changed, 270 insertions, 214 deletions

diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index f91b7d1e4..9e90c587c 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -276,9 +276,8 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_ad
 }
 
 template <class P>
-std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainCPUBuffer(VAddr cpu_addr, u32 size,
-                                                                 ObtainBufferSynchronize sync_info,
-                                                                 ObtainBufferOperation post_op) {
+std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainCPUBuffer(
+    VAddr cpu_addr, u32 size, ObtainBufferSynchronize sync_info, ObtainBufferOperation post_op) {
     const BufferId buffer_id = FindBuffer(cpu_addr, size);
     Buffer& buffer = slot_buffers[buffer_id];
 
diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h
index 9507071e5..c4f6e8d12 100644
--- a/src/video_core/buffer_cache/buffer_cache_base.h
+++ b/src/video_core/buffer_cache/buffer_cache_base.h
@@ -297,8 +297,8 @@ public:
                                                        ObtainBufferOperation post_op);
 
     [[nodiscard]] std::pair<Buffer*, u32> ObtainCPUBuffer(VAddr gpu_addr, u32 size,
-                                                       ObtainBufferSynchronize sync_info,
-                                                       ObtainBufferOperation post_op);
+                                                          ObtainBufferSynchronize sync_info,
+                                                          ObtainBufferOperation post_op);
     void FlushCachedWrites();
 
     /// Return true when there are uncommitted buffers to be downloaded
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 922c399e6..46b9c548a 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -596,12 +596,6 @@ void Maxwell3D::ProcessCounterReset() {
     case Regs::ClearReport::ZPassPixelCount:
         rasterizer->ResetCounter(VideoCommon::QueryType::ZPassPixelCount64);
         break;
-    case Regs::ClearReport::PrimitivesGenerated:
-        rasterizer->ResetCounter(VideoCommon::QueryType::StreamingByteCount);
-        break;
-    case Regs::ClearReport::VtgPrimitivesOut:
-        rasterizer->ResetCounter(VideoCommon::QueryType::StreamingByteCount);
-        break;
     default:
         LOG_DEBUG(Render_OpenGL, "Unimplemented counter reset={}", regs.clear_report_value);
         break;
diff --git a/src/video_core/engines/puller.cpp b/src/video_core/engines/puller.cpp
index 582738234..8dd34c04a 100644
--- a/src/video_core/engines/puller.cpp
+++ b/src/video_core/engines/puller.cpp
@@ -82,7 +82,8 @@ void Puller::ProcessSemaphoreTriggerMethod() {
     if (op == GpuSemaphoreOperation::WriteLong) {
         const GPUVAddr sequence_address{regs.semaphore_address.SemaphoreAddress()};
         const u32 payload = regs.semaphore_sequence;
-        rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload, VideoCommon::QueryPropertiesFlags::HasTimeout, payload, 0);
+        rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload,
+                          VideoCommon::QueryPropertiesFlags::HasTimeout, payload, 0);
     } else {
         do {
             const u32 word{memory_manager.Read<u32>(regs.semaphore_address.SemaphoreAddress())};
@@ -117,7 +118,8 @@ void Puller::ProcessSemaphoreTriggerMethod() {
 void Puller::ProcessSemaphoreRelease() {
     const GPUVAddr sequence_address{regs.semaphore_address.SemaphoreAddress()};
     const u32 payload = regs.semaphore_release;
-    rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload, VideoCommon::QueryPropertiesFlags::IsAFence, payload, 0);
+    rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload,
+                      VideoCommon::QueryPropertiesFlags::IsAFence, payload, 0);
 }
 
 void Puller::ProcessSemaphoreAcquire() {
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 8459a3092..805a89900 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -55,6 +55,9 @@ public:
 
     // Unlike other fences, this one doesn't
     void SignalOrdering() {
+        if constexpr (!can_async_check) {
+            TryReleasePendingFences<false>();
+        }
         std::scoped_lock lock{buffer_cache.mutex};
         buffer_cache.AccumulateFlushes();
     }
@@ -104,13 +107,9 @@ public:
         SignalFence(std::move(func));
     }
 
-    void WaitPendingFences(bool force) {
+    void WaitPendingFences([[maybe_unused]] bool force) {
         if constexpr (!can_async_check) {
-            if (force) {
-                TryReleasePendingFences<true>();
-            } else {
-                TryReleasePendingFences<false>();
-            }
+            TryReleasePendingFences<true>();
         } else {
             if (!force) {
                 return;
@@ -125,7 +124,8 @@ public:
             });
             SignalFence(std::move(func));
             std::unique_lock lk(wait_mutex);
-            wait_cv.wait(lk, [&wait_finished] { return wait_finished.load(std::memory_order_relaxed); });
+            wait_cv.wait(
+                lk, [&wait_finished] { return wait_finished.load(std::memory_order_relaxed); });
         }
     }
 
diff --git a/src/video_core/query_cache/bank_base.h b/src/video_core/query_cache/bank_base.h
index 4246a609d..420927091 100644
--- a/src/video_core/query_cache/bank_base.h
+++ b/src/video_core/query_cache/bank_base.h
@@ -7,21 +7,19 @@
 #include <deque>
 #include <utility>
 
-
 #include "common/common_types.h"
 
 namespace VideoCommon {
 
 class BankBase {
 protected:
-    const size_t base_bank_size;
-    size_t bank_size;
-    std::atomic<size_t> references;
-    size_t current_slot;
+    const size_t base_bank_size{};
+    size_t bank_size{};
+    std::atomic<size_t> references{};
+    size_t current_slot{};
 
 public:
-    BankBase(size_t bank_size_)
-        : base_bank_size{bank_size_}, bank_size(bank_size_), references(0), current_slot(0) {}
+    explicit BankBase(size_t bank_size_) : base_bank_size{bank_size_}, bank_size(bank_size_) {}
 
     virtual ~BankBase() = default;
 
@@ -58,11 +56,11 @@ public:
         bank_size = current_slot;
     }
 
-    constexpr bool IsClosed() {
+    bool IsClosed() const {
         return current_slot >= bank_size;
     }
 
-    bool IsDead() {
+    bool IsDead() const {
         return IsClosed() && references == 0;
     }
 };
diff --git a/src/video_core/query_cache/query_base.h b/src/video_core/query_cache/query_base.h
index 0ae23af9f..993a13eac 100644
--- a/src/video_core/query_cache/query_base.h
+++ b/src/video_core/query_cache/query_base.h
@@ -9,28 +9,28 @@
 namespace VideoCommon {
 
 enum class QueryFlagBits : u32 {
-    HasTimestamp = 1 << 0,       ///< Indicates if this query has a tiemstamp.
-    IsFinalValueSynced = 1 << 1, ///< Indicates if the query has been synced in the host
-    IsHostSynced = 1 << 2,       ///< Indicates if the query has been synced in the host
-    IsGuestSynced = 1 << 3,      ///< Indicates if the query has been synced with the guest.
-    IsHostManaged = 1 << 4,      ///< Indicates if this query points to a host query
-    IsRewritten = 1 << 5,        ///< Indicates if this query was rewritten by another query
-    IsInvalidated = 1 << 6,      ///< Indicates the value of th query has been nullified.
-    IsOrphan = 1 << 7,           ///< Indicates the query has not been set by a guest query.
-    IsFence = 1 << 8,            ///< Indicates the query is a fence.
-    IsQueuedForAsyncFlush = 1 <<9,///< Indicates that the query can be flushed at any moment
+    HasTimestamp = 1 << 0,          ///< Indicates if this query has a timestamp.
+    IsFinalValueSynced = 1 << 1,    ///< Indicates if the query has been synced in the host
+    IsHostSynced = 1 << 2,          ///< Indicates if the query has been synced in the host
+    IsGuestSynced = 1 << 3,         ///< Indicates if the query has been synced with the guest.
+    IsHostManaged = 1 << 4,         ///< Indicates if this query points to a host query
+    IsRewritten = 1 << 5,           ///< Indicates if this query was rewritten by another query
+    IsInvalidated = 1 << 6,         ///< Indicates the value of th query has been nullified.
+    IsOrphan = 1 << 7,              ///< Indicates the query has not been set by a guest query.
+    IsFence = 1 << 8,               ///< Indicates the query is a fence.
+    IsQueuedForAsyncFlush = 1 << 9, ///< Indicates that the query can be flushed at any moment
 };
 DECLARE_ENUM_FLAG_OPERATORS(QueryFlagBits)
 
 class QueryBase {
 public:
-    VAddr guest_address;
-    QueryFlagBits flags;
-    u64 value;
+    VAddr guest_address{};
+    QueryFlagBits flags{};
+    u64 value{};
 
 protected:
     // Default constructor
-    QueryBase() : guest_address(0), flags{}, value{} {}
+    QueryBase() = default;
 
     // Parameterized constructor
     QueryBase(VAddr address, QueryFlagBits flags_, u64 value_)
@@ -51,23 +51,21 @@ public:
 class HostQueryBase : public QueryBase {
 public:
     // Default constructor
-    HostQueryBase()
-        : QueryBase(0, QueryFlagBits::IsHostManaged | QueryFlagBits::IsOrphan, 0), start_bank_id{},
-          size_banks{}, start_slot{}, size_slots{} {}
+    HostQueryBase() : QueryBase(0, QueryFlagBits::IsHostManaged | QueryFlagBits::IsOrphan, 0) {}
 
     // Parameterized constructor
-    HostQueryBase(bool isLong, VAddr address)
+    HostQueryBase(bool has_timestamp, VAddr address)
         : QueryBase(address, QueryFlagBits::IsHostManaged, 0), start_bank_id{}, size_banks{},
           start_slot{}, size_slots{} {
-        if (isLong) {
+        if (has_timestamp) {
             flags |= QueryFlagBits::HasTimestamp;
         }
     }
 
-    u32 start_bank_id;
-    u32 size_banks;
-    size_t start_slot;
-    size_t size_slots;
+    u32 start_bank_id{};
+    u32 size_banks{};
+    size_t start_slot{};
+    size_t size_slots{};
 };
 
 } // namespace VideoCommon
\ No newline at end of file
diff --git a/src/video_core/query_cache/query_cache.h b/src/video_core/query_cache/query_cache.h
index f1393d5c7..042af053c 100644
--- a/src/video_core/query_cache/query_cache.h
+++ b/src/video_core/query_cache/query_cache.h
@@ -54,7 +54,7 @@ public:
         return new_id;
     }
 
-    bool HasPendingSync() override {
+    bool HasPendingSync() const override {
         return !pending_sync.empty();
     }
 
@@ -71,8 +71,10 @@ public:
                 continue;
             }
             query.flags |= QueryFlagBits::IsHostSynced;
-            sync_values.emplace_back(query.guest_address, query.value,
-                                     True(query.flags & QueryFlagBits::HasTimestamp) ? 8 : 4);
+            sync_values.emplace_back(SyncValuesStruct{
+                .address = query.guest_address,
+                .value = query.value,
+                .size = static_cast<u64>(True(query.flags & QueryFlagBits::HasTimestamp) ? 8 : 4)});
         }
         pending_sync.clear();
         if (sync_values.size() > 0) {
@@ -90,15 +92,20 @@ class StubStreamer : public GuestStreamer<Traits> {
 public:
     using RuntimeType = typename Traits::RuntimeType;
 
-    StubStreamer(size_t id_, RuntimeType& runtime_) : GuestStreamer<Traits>(id_, runtime_) {}
+    StubStreamer(size_t id_, RuntimeType& runtime_, u32 stub_value_)
+        : GuestStreamer<Traits>(id_, runtime_), stub_value{stub_value_} {}
 
     ~StubStreamer() override = default;
 
     size_t WriteCounter(VAddr address, bool has_timestamp, [[maybe_unused]] u32 value,
                         std::optional<u32> subreport = std::nullopt) override {
-        size_t new_id = GuestStreamer<Traits>::WriteCounter(address, has_timestamp, 1U, subreport);
+        size_t new_id =
+            GuestStreamer<Traits>::WriteCounter(address, has_timestamp, stub_value, subreport);
         return new_id;
     }
+
+private:
+    u32 stub_value;
 };
 
 template <typename Traits>
@@ -113,7 +120,7 @@ struct QueryCacheBase<Traits>::QueryCacheBaseImpl {
         for (size_t i = 0; i < static_cast<size_t>(QueryType::MaxQueryTypes); i++) {
             streamers[i] = runtime.GetStreamerInterface(static_cast<QueryType>(i));
             if (streamers[i]) {
-                streamer_mask |= 1ULL << i;
+                streamer_mask |= 1ULL << streamers[i]->GetId();
             }
         }
     }
@@ -152,7 +159,7 @@ struct QueryCacheBase<Traits>::QueryCacheBaseImpl {
     QueryCacheBase<Traits>* owner;
     VideoCore::RasterizerInterface& rasterizer;
     Core::Memory::Memory& cpu_memory;
-    Traits::RuntimeType& runtime;
+    RuntimeType& runtime;
     Tegra::GPU& gpu;
     std::array<StreamerInterface*, static_cast<size_t>(QueryType::MaxQueryTypes)> streamers;
     u64 streamer_mask;
@@ -223,15 +230,11 @@ void QueryCacheBase<Traits>::CounterReport(GPUVAddr addr, QueryType counter_type
     const bool is_fence = True(flags & QueryPropertiesFlags::IsAFence);
     size_t streamer_id = static_cast<size_t>(counter_type);
     auto* streamer = impl->streamers[streamer_id];
-    if (!streamer) [[unlikely]] {
-        if (has_timestamp) {
-            u64 timestamp = impl->gpu.GetTicks();
-            gpu_memory->Write<u64>(addr + 8, timestamp);
-            gpu_memory->Write<u64>(addr, 1ULL);
-        } else {
-            gpu_memory->Write<u32>(addr, 1U);
-        }
-        return;
+    if (streamer == nullptr) [[unlikely]] {
+        counter_type = QueryType::Payload;
+        payload = 1U;
+        streamer_id = static_cast<size_t>(counter_type);
+        streamer = impl->streamers[streamer_id];
     }
     auto cpu_addr_opt = gpu_memory->GpuToCpuAddress(addr);
     if (!cpu_addr_opt) [[unlikely]] {
@@ -403,12 +406,6 @@ bool QueryCacheBase<Traits>::AccelerateHostConditionalRendering() {
         impl->runtime.EndHostConditionalRendering();
         return false;
     }
-    /*if (!Settings::IsGPULevelHigh()) {
-        impl->runtime.EndHostConditionalRendering();
-        return gpu_memory->IsMemoryDirty(regs.render_enable.Address(), 24,
-                                         VideoCommon::CacheType::BufferCache |
-                                             VideoCommon::CacheType::QueryCache);
-    }*/
     const ComparisonMode mode = static_cast<ComparisonMode>(regs.render_enable.mode);
     const GPUVAddr address = regs.render_enable.Address();
     switch (mode) {
@@ -442,6 +439,9 @@ bool QueryCacheBase<Traits>::AccelerateHostConditionalRendering() {
 // Async downloads
 template <typename Traits>
 void QueryCacheBase<Traits>::CommitAsyncFlushes() {
+    // Make sure to have the results synced in Host.
+    NotifyWFI();
+
     u64 mask{};
     {
         std::scoped_lock lk(impl->flush_guard);
@@ -458,8 +458,19 @@ void QueryCacheBase<Traits>::CommitAsyncFlushes() {
     if (mask == 0) {
         return;
     }
-    impl->ForEachStreamerIn(mask,
-                            [](StreamerInterface* streamer) { streamer->PushUnsyncedQueries(); });
+    u64 ran_mask = ~mask;
+    while (mask) {
+        impl->ForEachStreamerIn(mask, [&mask, &ran_mask](StreamerInterface* streamer) {
+            u64 dep_mask = streamer->GetDependentMask();
+            if ((dep_mask & ~ran_mask) != 0) {
+                return;
+            }
+            u64 index = streamer->GetId();
+            ran_mask |= (1ULL << index);
+            mask &= ~(1ULL << index);
+            streamer->PushUnsyncedQueries();
+        });
+    }
 }
 
 template <typename Traits>
@@ -489,13 +500,11 @@ void QueryCacheBase<Traits>::PopAsyncFlushes() {
     if (mask == 0) {
         return;
     }
-    u64 ran_mask = 0;
-    u64 next_phase = 0;
+    u64 ran_mask = ~mask;
     while (mask) {
-        impl->ForEachStreamerIn(mask, [&mask, &ran_mask, &next_phase](StreamerInterface* streamer) {
+        impl->ForEachStreamerIn(mask, [&mask, &ran_mask](StreamerInterface* streamer) {
             u64 dep_mask = streamer->GetDependenceMask();
             if ((dep_mask & ~ran_mask) != 0) {
-                next_phase |= dep_mask;
                 return;
             }
             u64 index = streamer->GetId();
@@ -503,7 +512,6 @@ void QueryCacheBase<Traits>::PopAsyncFlushes() {
             mask &= ~(1ULL << index);
             streamer->PopUnsyncedQueries();
         });
-        ran_mask |= next_phase;
     }
 }
 
diff --git a/src/video_core/query_cache/query_cache_base.h b/src/video_core/query_cache/query_cache_base.h
index 55f508dd1..07be421c6 100644
--- a/src/video_core/query_cache/query_cache_base.h
+++ b/src/video_core/query_cache/query_cache_base.h
@@ -47,7 +47,7 @@ public:
         BitField<0, 27, u32> query_id;
         u32 raw;
 
-        std::pair<size_t, size_t> unpack() {
+        std::pair<size_t, size_t> unpack() const {
             return {static_cast<size_t>(stream_id.Value()), static_cast<size_t>(query_id.Value())};
         }
     };
@@ -73,7 +73,7 @@ public:
         }
     }
 
-    static u64 BuildMask(std::span<QueryType> types) {
+    static u64 BuildMask(std::span<const QueryType> types) {
         u64 mask = 0;
         for (auto query_type : types) {
             mask |= 1ULL << (static_cast<u64>(query_type));
@@ -160,7 +160,7 @@ protected:
         }
     }
 
-    using ContentCache = typename std::unordered_map<u64, std::unordered_map<u32, QueryLocation>>;
+    using ContentCache = std::unordered_map<u64, std::unordered_map<u32, QueryLocation>>;
 
     void InvalidateQuery(QueryLocation location);
     bool IsQueryDirty(QueryLocation location);
@@ -175,7 +175,7 @@ protected:
     friend struct QueryCacheBaseImpl;
     friend RuntimeType;
 
-    std::unique_ptr<struct QueryCacheBaseImpl> impl;
+    std::unique_ptr<QueryCacheBaseImpl> impl;
 };
 
 } // namespace VideoCommon
\ No newline at end of file
diff --git a/src/video_core/query_cache/query_stream.h b/src/video_core/query_cache/query_stream.h
index 0e9275565..e7aac955b 100644
--- a/src/video_core/query_cache/query_stream.h
+++ b/src/video_core/query_cache/query_stream.h
@@ -16,7 +16,7 @@ namespace VideoCommon {
 
 class StreamerInterface {
 public:
-    StreamerInterface(size_t id_, u64 dependance_mask_ = 0) : id{id_}, dependance_mask{dependance_mask_} {}
+    explicit StreamerInterface(size_t id_) : id{id_}, dependence_mask{}, dependent_mask{} {}
     virtual ~StreamerInterface() = default;
 
     virtual QueryBase* GetQuery(size_t id) = 0;
@@ -37,7 +37,7 @@ public:
         /* Do Nothing */
     }
 
-    virtual bool HasPendingSync() {
+    virtual bool HasPendingSync() const {
         return false;
     }
 
@@ -52,7 +52,7 @@ public:
     virtual size_t WriteCounter(VAddr address, bool has_timestamp, u32 value,
                                 std::optional<u32> subreport = std::nullopt) = 0;
 
-    virtual bool HasUnsyncedQueries() {
+    virtual bool HasUnsyncedQueries() const {
         return false;
     }
 
@@ -71,18 +71,28 @@ public:
     }
 
     u64 GetDependenceMask() const {
-        return dependance_mask;
+        return dependence_mask;
+    }
+
+    u64 GetDependentMask() const {
+        return dependence_mask;
     }
 
 protected:
+    void MakeDependent(StreamerInterface* depend_on) {
+        dependence_mask |= 1ULL << depend_on->id;
+        depend_on->dependent_mask |= 1ULL << id;
+    }
+
     const size_t id;
-    const u64 dependance_mask;
+    u64 dependence_mask;
+    u64 dependent_mask;
 };
 
 template <typename QueryType>
 class SimpleStreamer : public StreamerInterface {
 public:
-    SimpleStreamer(size_t id_, u64 dependance_mask_ = 0) : StreamerInterface{id_, dependance_mask_} {}
+    explicit SimpleStreamer(size_t id_) : StreamerInterface{id_} {}
     virtual ~SimpleStreamer() = default;
 
 protected:
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 2ba7cbb0d..af1469147 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -9,10 +9,10 @@
 #include <utility>
 #include "common/common_types.h"
 #include "common/polyfill_thread.h"
-#include "video_core/query_cache/types.h"
 #include "video_core/cache_types.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/gpu.h"
+#include "video_core/query_cache/types.h"
 #include "video_core/rasterizer_download_area.h"
 
 namespace Tegra {
@@ -57,7 +57,8 @@ public:
     virtual void ResetCounter(VideoCommon::QueryType type) = 0;
 
     /// Records a GPU query and caches it
-    virtual void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type, VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) = 0;
+    virtual void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+                       VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) = 0;
 
     /// Signal an uniform buffer binding
     virtual void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
diff --git a/src/video_core/renderer_null/null_rasterizer.h b/src/video_core/renderer_null/null_rasterizer.h
index 57a8c4c85..23001eeb8 100644
--- a/src/video_core/renderer_null/null_rasterizer.h
+++ b/src/video_core/renderer_null/null_rasterizer.h
@@ -43,7 +43,8 @@ public:
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
     void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type, VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index a975bbe75..27e2de1bf 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -405,8 +405,6 @@ void RasterizerOpenGL::ResetCounter(VideoCommon::QueryType type) {
 void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
                              VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
     if (type == VideoCommon::QueryType::ZPassPixelCount64) {
-        std::optional<u64> timestamp{True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)
-                                     ? std::make_optional<u64>(gpu.GetTicks()) : std:: nullopt };
         if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
             query_cache.Query(gpu_addr, VideoCore::QueryType::SamplesPassed, {gpu.GetTicks()});
         } else {
@@ -414,13 +412,23 @@ void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
         }
         return;
     }
-    if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
-        u64 ticks = gpu.GetTicks();
-        gpu_memory->Write<u64>(gpu_addr + 8, ticks);
-        gpu_memory->Write<u64>(gpu_addr, static_cast<u64>(payload));
-    } else {
-        gpu_memory->Write<u32>(gpu_addr, payload);
+    if (type != VideoCommon::QueryType::Payload) {
+        payload = 1u;
+    }
+    std::function<void()> func([this, gpu_addr, flags, memory_manager = gpu_memory, payload]() {
+        if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
+            u64 ticks = gpu.GetTicks();
+            memory_manager->Write<u64>(gpu_addr + 8, ticks);
+            memory_manager->Write<u64>(gpu_addr, static_cast<u64>(payload));
+        } else {
+            memory_manager->Write<u32>(gpu_addr, payload);
+        }
+    });
+    if (True(flags & VideoCommon::QueryPropertiesFlags::IsAFence)) {
+        SignalFence(std::move(func));
+        return;
     }
+    func();
 }
 
 void RasterizerOpenGL::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 05e048e15..ceffe1f1e 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -87,7 +87,8 @@ public:
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
     void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type, VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.cpp b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
index 97cd4521d..039dc95e1 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
@@ -303,9 +303,9 @@ std::pair<VkBuffer, VkDeviceSize> QuadIndexedPass::Assemble(
     return {staging.buffer, staging.offset};
 }
 
-ConditionalRenderingResolvePass::ConditionalRenderingResolvePass(const Device& device_,
-                                                                 Scheduler& scheduler_,
-                                                                 DescriptorPool& descriptor_pool_, ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+ConditionalRenderingResolvePass::ConditionalRenderingResolvePass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
     : ComputePass(device_, descriptor_pool_, INPUT_OUTPUT_DESCRIPTOR_SET_BINDINGS,
                   INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE, INPUT_OUTPUT_BANK_INFO, nullptr,
                   RESOLVE_CONDITIONAL_RENDER_COMP_SPV),
diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.h b/src/video_core/renderer_vulkan/vk_fence_manager.h
index 14fc5ad71..336573574 100644
--- a/src/video_core/renderer_vulkan/vk_fence_manager.h
+++ b/src/video_core/renderer_vulkan/vk_fence_manager.h
@@ -7,8 +7,8 @@
 
 #include "video_core/fence_manager.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
-#include "video_core/renderer_vulkan/vk_texture_cache.h"
 #include "video_core/renderer_vulkan/vk_query_cache.h"
+#include "video_core/renderer_vulkan/vk_texture_cache.h"
 
 namespace Core {
 class System;
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.cpp b/src/video_core/renderer_vulkan/vk_query_cache.cpp
index ef891e26b..add0c6fb3 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_query_cache.cpp
@@ -11,11 +11,9 @@
 #include <utility>
 #include <vector>
 
-#include <boost/container/small_vector.hpp>
-#include <boost/icl/interval_set.hpp>
-
 #include "common/common_types.h"
 #include "core/memory.h"
+#include "video_core/engines/draw_manager.h"
 #include "video_core/query_cache/query_cache.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
@@ -30,6 +28,7 @@
 
 namespace Vulkan {
 
+using Tegra::Engines::Maxwell3D;
 using VideoCommon::QueryType;
 
 namespace {
@@ -37,7 +36,7 @@ class SamplesQueryBank : public VideoCommon::BankBase {
 public:
     static constexpr size_t BANK_SIZE = 256;
     static constexpr size_t QUERY_SIZE = 8;
-    SamplesQueryBank(const Device& device_, size_t index_)
+    explicit SamplesQueryBank(const Device& device_, size_t index_)
         : BankBase(BANK_SIZE), device{device_}, index{index_} {
         const auto& dev = device.GetLogical();
         query_pool = dev.CreateQueryPool({
@@ -109,18 +108,19 @@ struct HostSyncValues {
     static constexpr bool GeneratesBaseBuffer = false;
 };
 
-template <typename Traits>
 class SamplesStreamer : public BaseStreamer {
 public:
-    SamplesStreamer(size_t id, QueryCacheRuntime& runtime_, const Device& device_,
-                    Scheduler& scheduler_, const MemoryAllocator& memory_allocator_)
-        : BaseStreamer(id), runtime{runtime_}, device{device_}, scheduler{scheduler_},
+    explicit SamplesStreamer(size_t id_, QueryCacheRuntime& runtime_, const Device& device_,
+                             Scheduler& scheduler_, const MemoryAllocator& memory_allocator_)
+        : BaseStreamer(id_), runtime{runtime_}, device{device_}, scheduler{scheduler_},
           memory_allocator{memory_allocator_} {
         BuildResolveBuffer();
         current_bank = nullptr;
         current_query = nullptr;
     }
 
+    ~SamplesStreamer() = default;
+
     void StartCounter() override {
         if (has_started) {
             return;
@@ -157,7 +157,7 @@ public:
         PauseCounter();
     }
 
-    bool HasPendingSync() override {
+    bool HasPendingSync() const override {
         return !pending_sync.empty();
     }
 
@@ -198,7 +198,7 @@ public:
                     }
                     resolve_slots_remaining = resolve_slots;
                     sync_values_stash.emplace_back();
-                    sync_values = sync_values = &sync_values_stash.back();
+                    sync_values = &sync_values_stash.back();
                     sync_values->reserve(resolve_slots * SamplesQueryBank::BANK_SIZE);
                 }
                 resolve_slots_remaining--;
@@ -207,6 +207,7 @@ public:
             const size_t base_offset = SamplesQueryBank::QUERY_SIZE * SamplesQueryBank::BANK_SIZE *
                                        (resolve_slots - resolve_slots_remaining - 1);
             VkQueryPool query_pool = bank->GetInnerPool();
+            scheduler.RequestOutsideRenderPassOperationContext();
             scheduler.Record([start, amount, base_offset, query_pool,
                               buffer = *resolve_buffer](vk::CommandBuffer cmdbuf) {
                 size_t final_offset = base_offset + start * SamplesQueryBank::QUERY_SIZE;
@@ -284,7 +285,7 @@ public:
         return index;
     }
 
-    bool HasUnsyncedQueries() override {
+    bool HasUnsyncedQueries() const override {
         return !pending_flush_queries.empty();
     }
 
@@ -348,8 +349,8 @@ private:
         for (auto q : queries) {
             auto* query = GetQuery(q);
             ApplyBankOp(query, [&indexer](SamplesQueryBank* bank, size_t start, size_t amount) {
-                auto id = bank->GetIndex();
-                auto pair = indexer.try_emplace(id, std::numeric_limits<size_t>::max(),
+                auto id_ = bank->GetIndex();
+                auto pair = indexer.try_emplace(id_, std::numeric_limits<size_t>::max(),
                                                 std::numeric_limits<size_t>::min());
                 auto& current_pair = pair.first->second;
                 current_pair.first = std::min(current_pair.first, start);
@@ -434,13 +435,14 @@ private:
             .pNext = nullptr,
             .flags = 0,
             .size = SamplesQueryBank::QUERY_SIZE * SamplesQueryBank::BANK_SIZE * resolve_slots,
-            .usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
+            .usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
+                     VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
             .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
             .queueFamilyIndexCount = 0,
             .pQueueFamilyIndices = nullptr,
         };
         resolve_buffers.emplace_back(
-            std::move(memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal)));
+            memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal));
     }
 
     static constexpr size_t resolve_slots = 8;
@@ -476,7 +478,8 @@ class TFBQueryBank : public VideoCommon::BankBase {
 public:
     static constexpr size_t BANK_SIZE = 1024;
     static constexpr size_t QUERY_SIZE = 4;
-    TFBQueryBank(Scheduler& scheduler_, const MemoryAllocator& memory_allocator, size_t index_)
+    explicit TFBQueryBank(Scheduler& scheduler_, const MemoryAllocator& memory_allocator,
+                          size_t index_)
         : BankBase(BANK_SIZE), scheduler{scheduler_}, index{index_} {
         const VkBufferCreateInfo buffer_ci = {
             .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
@@ -525,22 +528,21 @@ private:
     vk::Buffer buffer;
 };
 
-template <typename Traits>
 class PrimitivesSucceededStreamer;
 
-template <typename Traits>
 class TFBCounterStreamer : public BaseStreamer {
 public:
-    TFBCounterStreamer(size_t id, QueryCacheRuntime& runtime_, const Device& device_,
-                       Scheduler& scheduler_, const MemoryAllocator& memory_allocator_,
-                       StagingBufferPool& staging_pool_)
-        : BaseStreamer(id), runtime{runtime_}, device{device_}, scheduler{scheduler_},
+    explicit TFBCounterStreamer(size_t id_, QueryCacheRuntime& runtime_, const Device& device_,
+                                Scheduler& scheduler_, const MemoryAllocator& memory_allocator_,
+                                StagingBufferPool& staging_pool_)
+        : BaseStreamer(id_), runtime{runtime_}, device{device_}, scheduler{scheduler_},
           memory_allocator{memory_allocator_}, staging_pool{staging_pool_} {
         buffers_count = 0;
         current_bank = nullptr;
         counter_buffers.fill(VK_NULL_HANDLE);
         offsets.fill(0);
         last_queries.fill(0);
+        last_queries_stride.fill(1);
         const VkBufferCreateInfo buffer_ci = {
             .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
             .pNext = nullptr,
@@ -564,6 +566,8 @@ public:
         }
     }
 
+    ~TFBCounterStreamer() = default;
+
     void StartCounter() override {
         FlushBeginTFB();
         has_started = true;
@@ -581,15 +585,15 @@ public:
         if (has_flushed_end_pending) {
             FlushEndTFB();
         }
-        runtime.View3DRegs([this](Tegra::Engines::Maxwell3D::Regs& regs) {
-            if (regs.transform_feedback_enabled == 0) {
+        runtime.View3DRegs([this](Maxwell3D& maxwell3d) {
+            if (maxwell3d.regs.transform_feedback_enabled == 0) {
                 streams_mask = 0;
                 has_started = false;
             }
         });
     }
 
-    bool HasPendingSync() override {
+    bool HasPendingSync() const override {
         return !pending_sync.empty();
     }
 
@@ -650,14 +654,19 @@ public:
         return index;
     }
 
-    std::optional<VAddr> GetLastQueryStream(size_t stream) {
+    std::optional<std::pair<VAddr, size_t>> GetLastQueryStream(size_t stream) {
         if (last_queries[stream] != 0) {
-            return {last_queries[stream]};
+            std::pair<VAddr, size_t> result(last_queries[stream], last_queries_stride[stream]);
+            return result;
         }
         return std::nullopt;
     }
 
-    bool HasUnsyncedQueries() override {
+    Maxwell3D::Regs::PrimitiveTopology GetOutputTopology() const {
+        return out_topology;
+    }
+
+    bool HasUnsyncedQueries() const override {
         return !pending_flush_queries.empty();
     }
 
@@ -762,15 +771,17 @@ private:
 
     void UpdateBuffers() {
         last_queries.fill(0);
-        runtime.View3DRegs([this](Tegra::Engines::Maxwell3D::Regs& regs) {
+        last_queries_stride.fill(1);
+        runtime.View3DRegs([this](Maxwell3D& maxwell3d) {
             buffers_count = 0;
-            for (size_t i = 0; i < Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers;
-                 i++) {
-                const auto& tf = regs.transform_feedback;
+            out_topology = maxwell3d.draw_manager->GetDrawState().topology;
+            for (size_t i = 0; i < Maxwell3D::Regs::NumTransformFeedbackBuffers; i++) {
+                const auto& tf = maxwell3d.regs.transform_feedback;
                 if (tf.buffers[i].enable == 0) {
                     continue;
                 }
                 const size_t stream = tf.controls[i].stream;
+                last_queries_stride[stream] = tf.controls[i].stride;
                 streams_mask |= 1ULL << stream;
                 buffers_count = std::max<size_t>(buffers_count, stream + 1);
             }
@@ -785,7 +796,8 @@ private:
                 });
             current_bank = &bank_pool.GetBank(current_bank_id);
         }
-        auto [dont_care, slot] = current_bank->Reserve();
+        auto [dont_care, other] = current_bank->Reserve();
+        const size_t slot = other; // workaround to compile bug.
         current_bank->AddReference();
 
         static constexpr VkMemoryBarrier READ_BARRIER{
@@ -818,11 +830,9 @@ private:
         return {current_bank_id, slot};
     }
 
-    template <typename Traits>
     friend class PrimitivesSucceededStreamer;
 
     static constexpr size_t NUM_STREAMS = 4;
-    static constexpr size_t STREAMS_MASK = (1ULL << NUM_STREAMS) - 1ULL;
 
     QueryCacheRuntime& runtime;
     const Device& device;
@@ -851,6 +861,8 @@ private:
     std::array<VkBuffer, NUM_STREAMS> counter_buffers{};
     std::array<VkDeviceSize, NUM_STREAMS> offsets{};
     std::array<VAddr, NUM_STREAMS> last_queries;
+    std::array<size_t, NUM_STREAMS> last_queries_stride;
+    Maxwell3D::Regs::PrimitiveTopology out_topology;
     u64 streams_mask;
 };
 
@@ -858,32 +870,34 @@ class PrimitivesQueryBase : public VideoCommon::QueryBase {
 public:
     // Default constructor
     PrimitivesQueryBase()
-        : VideoCommon::QueryBase(0, VideoCommon::QueryFlagBits::IsHostManaged, 0), stride{},
-          dependant_index{}, dependant_manage{} {}
+        : VideoCommon::QueryBase(0, VideoCommon::QueryFlagBits::IsHostManaged, 0) {}
 
     // Parameterized constructor
-    PrimitivesQueryBase(bool is_long, VAddr address)
-        : VideoCommon::QueryBase(address, VideoCommon::QueryFlagBits::IsHostManaged, 0), stride{},
-          dependant_index{}, dependant_manage{} {
-        if (is_long) {
+    PrimitivesQueryBase(bool has_timestamp, VAddr address)
+        : VideoCommon::QueryBase(address, VideoCommon::QueryFlagBits::IsHostManaged, 0) {
+        if (has_timestamp) {
             flags |= VideoCommon::QueryFlagBits::HasTimestamp;
         }
     }
 
-    u64 stride;
-    VAddr dependant_address;
-    size_t dependant_index;
-    bool dependant_manage;
+    u64 stride{};
+    VAddr dependant_address{};
+    Maxwell3D::Regs::PrimitiveTopology topology{Maxwell3D::Regs::PrimitiveTopology::Points};
+    size_t dependant_index{};
+    bool dependant_manage{};
 };
 
-template <typename Traits>
 class PrimitivesSucceededStreamer : public VideoCommon::SimpleStreamer<PrimitivesQueryBase> {
 public:
-    PrimitivesSucceededStreamer(size_t id, QueryCacheRuntime& runtime_,
-                                TFBCounterStreamer<QueryCacheParams>& tfb_streamer_, Core::Memory::Memory& cpu_memory_)
-        : VideoCommon::SimpleStreamer<PrimitivesQueryBase>(
-              id, 1ULL << static_cast<u64>(VideoCommon::QueryType::StreamingByteCount)),
-          runtime{runtime_}, tfb_streamer{tfb_streamer_}, cpu_memory{cpu_memory_} {}
+    explicit PrimitivesSucceededStreamer(size_t id_, QueryCacheRuntime& runtime_,
+                                         TFBCounterStreamer& tfb_streamer_,
+                                         Core::Memory::Memory& cpu_memory_)
+        : VideoCommon::SimpleStreamer<PrimitivesQueryBase>(id_), runtime{runtime_},
+          tfb_streamer{tfb_streamer_}, cpu_memory{cpu_memory_} {
+        MakeDependent(&tfb_streamer);
+    }
+
+    ~PrimitivesSucceededStreamer() = default;
 
     size_t WriteCounter(VAddr address, bool has_timestamp, u32 value,
                         std::optional<u32> subreport_) override {
@@ -901,8 +915,11 @@ public:
         const size_t subreport = static_cast<size_t>(*subreport_);
         auto dependant_address_opt = tfb_streamer.GetLastQueryStream(subreport);
         bool must_manage_dependance = false;
+        new_query->topology = tfb_streamer.GetOutputTopology();
         if (dependant_address_opt) {
-            new_query->dependant_address = *dependant_address_opt;
+            auto [dep_address, stride] = *dependant_address_opt;
+            new_query->dependant_address = dep_address;
+            new_query->stride = stride;
         } else {
             new_query->dependant_index =
                 tfb_streamer.WriteCounter(address, has_timestamp, value, subreport_);
@@ -917,25 +934,28 @@ public:
                 }
                 return index;
             }
+            new_query->stride = 1;
+            runtime.View3DRegs([new_query, subreport](Maxwell3D& maxwell3d) {
+                for (size_t i = 0; i < Maxwell3D::Regs::NumTransformFeedbackBuffers; i++) {
+                    const auto& tf = maxwell3d.regs.transform_feedback;
+                    if (tf.buffers[i].enable == 0) {
+                        continue;
+                    }
+                    if (tf.controls[i].stream != subreport) {
+                        continue;
+                    }
+                    new_query->stride = tf.controls[i].stride;
+                    break;
+                }
+            });
         }
 
         new_query->dependant_manage = must_manage_dependance;
-        runtime.View3DRegs([new_query, subreport](Tegra::Engines::Maxwell3D::Regs& regs) {
-            for (size_t i = 0; i < Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers;
-                 i++) {
-                const auto& tf = regs.transform_feedback;
-                if (tf.controls[i].stream != subreport) {
-                    continue;
-                }
-                new_query->stride = tf.controls[i].stride;
-                break;
-            }
-        });
         pending_flush_queries.push_back(index);
         return index;
     }
 
-    bool HasUnsyncedQueries() override {
+    bool HasUnsyncedQueries() const override {
         return !pending_flush_queries.empty();
     }
 
@@ -960,22 +980,49 @@ public:
             }
 
             query->flags |= VideoCommon::QueryFlagBits::IsFinalValueSynced;
+            u64 num_vertices = 0;
             if (query->dependant_manage) {
                 auto* dependant_query = tfb_streamer.GetQuery(query->dependant_index);
-                query->value = dependant_query->value / query->stride;
+                num_vertices = dependant_query->value / query->stride;
                 tfb_streamer.Free(query->dependant_index);
             } else {
                 u8* pointer = cpu_memory.GetPointer(query->dependant_address);
                 u32 result;
                 std::memcpy(&result, pointer, sizeof(u32));
-                query->value = static_cast<u64>(result) / query->stride;
+                num_vertices = static_cast<u64>(result) / query->stride;
             }
+            query->value = [&]() -> u64 {
+                switch (query->topology) {
+                case Maxwell3D::Regs::PrimitiveTopology::Points:
+                    return num_vertices;
+                case Maxwell3D::Regs::PrimitiveTopology::Lines:
+                    return num_vertices / 2;
+                case Maxwell3D::Regs::PrimitiveTopology::LineLoop:
+                    return (num_vertices / 2) + 1;
+                case Maxwell3D::Regs::PrimitiveTopology::LineStrip:
+                    return num_vertices - 1;
+                case Maxwell3D::Regs::PrimitiveTopology::Patches:
+                case Maxwell3D::Regs::PrimitiveTopology::Triangles:
+                case Maxwell3D::Regs::PrimitiveTopology::TrianglesAdjacency:
+                    return num_vertices / 3;
+                case Maxwell3D::Regs::PrimitiveTopology::TriangleFan:
+                case Maxwell3D::Regs::PrimitiveTopology::TriangleStrip:
+                case Maxwell3D::Regs::PrimitiveTopology::TriangleStripAdjacency:
+                    return num_vertices - 2;
+                case Maxwell3D::Regs::PrimitiveTopology::Quads:
+                    return num_vertices / 4;
+                case Maxwell3D::Regs::PrimitiveTopology::Polygon:
+                    return 1U;
+                default:
+                    return num_vertices;
+                }
+            }();
         }
     }
 
 private:
     QueryCacheRuntime& runtime;
-    TFBCounterStreamer<QueryCacheParams>& tfb_streamer;
+    TFBCounterStreamer& tfb_streamer;
     Core::Memory::Memory& cpu_memory;
 
     // syncing queue
@@ -1005,7 +1052,10 @@ struct QueryCacheRuntimeImpl {
           tfb_streamer(static_cast<size_t>(QueryType::StreamingByteCount), runtime, device,
                        scheduler, memory_allocator, staging_pool),
           primitives_succeeded_streamer(
-              static_cast<size_t>(QueryType::StreamingPrimitivesSucceeded), runtime, tfb_streamer, cpu_memory_),
+              static_cast<size_t>(QueryType::StreamingPrimitivesSucceeded), runtime, tfb_streamer,
+              cpu_memory_),
+          primitives_needed_minus_suceeded_streamer(
+              static_cast<size_t>(QueryType::StreamingPrimitivesNeededMinusSucceeded), runtime, 0u),
           hcr_setup{}, hcr_is_set{}, is_hcr_running{} {
 
         hcr_setup.sType = VK_STRUCTURE_TYPE_CONDITIONAL_RENDERING_BEGIN_INFO_EXT;
@@ -1040,9 +1090,10 @@ struct QueryCacheRuntimeImpl {
 
     // Streamers
     VideoCommon::GuestStreamer<QueryCacheParams> guest_streamer;
-    SamplesStreamer<QueryCacheParams> sample_streamer;
-    TFBCounterStreamer<QueryCacheParams> tfb_streamer;
-    PrimitivesSucceededStreamer<QueryCacheParams> primitives_succeeded_streamer;
+    SamplesStreamer sample_streamer;
+    TFBCounterStreamer tfb_streamer;
+    PrimitivesSucceededStreamer primitives_succeeded_streamer;
+    VideoCommon::StubStreamer<QueryCacheParams> primitives_needed_minus_suceeded_streamer;
 
     std::vector<std::pair<VAddr, VAddr>> little_cache;
     std::vector<std::pair<VkBuffer, VkDeviceSize>> buffers_to_upload_to;
@@ -1059,7 +1110,7 @@ struct QueryCacheRuntimeImpl {
     bool is_hcr_running;
 
     // maxwell3d
-    Tegra::Engines::Maxwell3D* maxwell3d;
+    Maxwell3D* maxwell3d;
 };
 
 QueryCacheRuntime::QueryCacheRuntime(VideoCore::RasterizerInterface* rasterizer,
@@ -1074,13 +1125,13 @@ QueryCacheRuntime::QueryCacheRuntime(VideoCore::RasterizerInterface* rasterizer,
         staging_pool_, compute_pass_descriptor_queue, descriptor_pool);
 }
 
-void QueryCacheRuntime::Bind3DEngine(Tegra::Engines::Maxwell3D* maxwell3d) {
+void QueryCacheRuntime::Bind3DEngine(Maxwell3D* maxwell3d) {
     impl->maxwell3d = maxwell3d;
 }
 
 template <typename Func>
 void QueryCacheRuntime::View3DRegs(Func&& func) {
-    func(impl->maxwell3d->regs);
+    func(*impl->maxwell3d);
 }
 
 void QueryCacheRuntime::EndHostConditionalRendering() {
@@ -1240,8 +1291,12 @@ VideoCommon::StreamerInterface* QueryCacheRuntime::GetStreamerInterface(QueryTyp
         return &impl->sample_streamer;
     case QueryType::StreamingByteCount:
         return &impl->tfb_streamer;
+    case QueryType::StreamingPrimitivesNeeded:
+    case QueryType::VtgPrimitivesOut:
     case QueryType::StreamingPrimitivesSucceeded:
         return &impl->primitives_succeeded_streamer;
+    case QueryType::StreamingPrimitivesNeededMinusSucceeded:
+        return &impl->primitives_needed_minus_suceeded_streamer;
     default:
         return nullptr;
     }
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.h b/src/video_core/renderer_vulkan/vk_query_cache.h
index 9ad2929d7..e9a1ea169 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.h
+++ b/src/video_core/renderer_vulkan/vk_query_cache.h
@@ -49,7 +49,8 @@ public:
     bool HostConditionalRenderingCompareValue(VideoCommon::LookupData object_1, bool qc_dirty);
 
     bool HostConditionalRenderingCompareValues(VideoCommon::LookupData object_1,
-                                               VideoCommon::LookupData object_2, bool qc_dirty, bool equal_check);
+                                               VideoCommon::LookupData object_2, bool qc_dirty,
+                                               bool equal_check);
 
     VideoCommon::StreamerInterface* GetStreamerInterface(VideoCommon::QueryType query_type);
 
@@ -66,7 +67,7 @@ private:
 };
 
 struct QueryCacheParams {
-    using RuntimeType = Vulkan::QueryCacheRuntime;
+    using RuntimeType = typename Vulkan::QueryCacheRuntime;
 };
 
 using QueryCache = VideoCommon::QueryCacheBase<QueryCacheParams>;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index e8862ba04..c7ce7c312 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -194,15 +194,6 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
 
     query_cache.NotifySegment(true);
 
-#if ANDROID
-    if (Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        // query_cache.UpdateCounters();
-    }
-#else
-    // query_cache.UpdateCounters();
-#endif
-
     GraphicsPipeline* const pipeline{pipeline_cache.CurrentGraphicsPipeline()};
     if (!pipeline) {
         return;
@@ -294,15 +285,6 @@ void RasterizerVulkan::DrawTexture() {
 
     query_cache.NotifySegment(true);
 
-#if ANDROID
-    if (Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        // query_cache.UpdateCounters();
-    }
-#else
-    // query_cache.UpdateCounters();
-#endif
-
     texture_cache.SynchronizeGraphicsDescriptors();
     texture_cache.UpdateRenderTargets(false);
 
@@ -332,15 +314,6 @@ void RasterizerVulkan::Clear(u32 layer_count) {
     FlushWork();
     gpu_memory->FlushCaching();
 
-#if ANDROID
-    if (Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        // query_cache.UpdateCounters();
-    }
-#else
-    // query_cache.UpdateCounters();
-#endif
-
     query_cache.NotifySegment(true);
     query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
                               maxwell3d->regs.zpass_pixel_count_enable);
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index ffd44c68d..ad069556c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -85,7 +85,8 @@ public:
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
     void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type, VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.h b/src/video_core/renderer_vulkan/vk_scheduler.h
index c87e5fb07..da03803aa 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.h
+++ b/src/video_core/renderer_vulkan/vk_scheduler.h
@@ -15,9 +15,13 @@
 #include "common/common_types.h"
 #include "common/polyfill_thread.h"
 #include "video_core/renderer_vulkan/vk_master_semaphore.h"
-#include "video_core/renderer_vulkan/vk_query_cache.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
+namespace VideoCommon {
+template <typename Trait>
+class QueryCacheBase;
+}
+
 namespace Vulkan {
 
 class CommandPool;
@@ -26,6 +30,8 @@ class Framebuffer;
 class GraphicsPipeline;
 class StateTracker;
 
+struct QueryCacheParams;
+
 /// The scheduler abstracts command buffer and fence management with an interface that's able to do
 /// OpenGL-like operations on Vulkan command buffers.
 class Scheduler {
@@ -63,7 +69,7 @@ public:
     void InvalidateState();
 
     /// Assigns the query cache.
-    void SetQueryCache(QueryCache& query_cache_) {
+    void SetQueryCache(VideoCommon::QueryCacheBase<QueryCacheParams>& query_cache_) {
         query_cache = &query_cache_;
     }
 
@@ -219,7 +225,7 @@ private:
     std::unique_ptr<MasterSemaphore> master_semaphore;
     std::unique_ptr<CommandPool> command_pool;
 
-    QueryCache* query_cache = nullptr;
+    VideoCommon::QueryCacheBase<QueryCacheParams>* query_cache = nullptr;
 
     vk::CommandBuffer current_cmdbuf;