Query Cache: Fix guest side sample counting

5 files changed, 97 insertions, 46 deletions

diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 46b9c548a..32d767d85 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -586,12 +586,6 @@ void Maxwell3D::ProcessQueryCondition() {
 }
 
 void Maxwell3D::ProcessCounterReset() {
-#if ANDROID
-    if (!Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        return;
-    }
-#endif
     switch (regs.clear_report_value) {
     case Regs::ClearReport::ZPassPixelCount:
         rasterizer->ResetCounter(VideoCommon::QueryType::ZPassPixelCount64);
diff --git a/src/video_core/query_cache/query_base.h b/src/video_core/query_cache/query_base.h
index 993a13eac..1d786b3a7 100644
--- a/src/video_core/query_cache/query_base.h
+++ b/src/video_core/query_cache/query_base.h
@@ -9,16 +9,15 @@
 namespace VideoCommon {
 
 enum class QueryFlagBits : u32 {
-    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
+    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.
 };
 DECLARE_ENUM_FLAG_OPERATORS(QueryFlagBits)
 
diff --git a/src/video_core/query_cache/query_cache.h b/src/video_core/query_cache/query_cache.h
index 042af053c..4b89b5bf6 100644
--- a/src/video_core/query_cache/query_cache.h
+++ b/src/video_core/query_cache/query_cache.h
@@ -256,30 +256,32 @@ void QueryCacheBase<Traits>::CounterReport(GPUVAddr addr, QueryType counter_type
     u8* pointer = impl->cpu_memory.GetPointer(cpu_addr);
     u8* pointer_timestamp = impl->cpu_memory.GetPointer(cpu_addr + 8);
     bool is_synced = !Settings::IsGPULevelHigh() && is_fence;
-    std::function<void()> operation(
-        [this, is_synced, query_base = query, query_location, pointer, pointer_timestamp] {
-            if (True(query_base->flags & QueryFlagBits::IsInvalidated)) {
-                if (!is_synced) [[likely]] {
-                    impl->pending_unregister.push_back(query_location);
-                }
-                return;
-            }
-            if (False(query_base->flags & QueryFlagBits::IsFinalValueSynced)) [[unlikely]] {
-                UNREACHABLE();
-                return;
-            }
-            if (True(query_base->flags & QueryFlagBits::HasTimestamp)) {
-                u64 timestamp = impl->gpu.GetTicks();
-                std::memcpy(pointer_timestamp, &timestamp, sizeof(timestamp));
-                std::memcpy(pointer, &query_base->value, sizeof(query_base->value));
-            } else {
-                u32 value = static_cast<u32>(query_base->value);
-                std::memcpy(pointer, &value, sizeof(value));
-            }
+    std::function<void()> operation([this, is_synced, streamer, query_base = query, query_location,
+                                     pointer, pointer_timestamp] {
+        if (True(query_base->flags & QueryFlagBits::IsInvalidated)) {
             if (!is_synced) [[likely]] {
                 impl->pending_unregister.push_back(query_location);
             }
-        });
+            return;
+        }
+        if (False(query_base->flags & QueryFlagBits::IsFinalValueSynced)) [[unlikely]] {
+            UNREACHABLE();
+            return;
+        }
+        query_base->value += streamer->GetAmmendValue();
+        streamer->SetAccumulationValue(query_base->value);
+        if (True(query_base->flags & QueryFlagBits::HasTimestamp)) {
+            u64 timestamp = impl->gpu.GetTicks();
+            std::memcpy(pointer_timestamp, &timestamp, sizeof(timestamp));
+            std::memcpy(pointer, &query_base->value, sizeof(query_base->value));
+        } else {
+            u32 value = static_cast<u32>(query_base->value);
+            std::memcpy(pointer, &value, sizeof(value));
+        }
+        if (!is_synced) [[likely]] {
+            impl->pending_unregister.push_back(query_location);
+        }
+    });
     if (is_fence) {
         impl->rasterizer.SignalFence(std::move(operation));
     } else {
@@ -354,9 +356,9 @@ void QueryCacheBase<Traits>::NotifySegment(bool resume) {
     if (resume) {
         impl->runtime.ResumeHostConditionalRendering();
     } else {
-        impl->runtime.PauseHostConditionalRendering();
         CounterClose(VideoCommon::QueryType::ZPassPixelCount64);
         CounterClose(VideoCommon::QueryType::StreamingByteCount);
+        impl->runtime.PauseHostConditionalRendering();
     }
 }
 
diff --git a/src/video_core/query_cache/query_stream.h b/src/video_core/query_cache/query_stream.h
index e7aac955b..39da6ac07 100644
--- a/src/video_core/query_cache/query_stream.h
+++ b/src/video_core/query_cache/query_stream.h
@@ -78,6 +78,14 @@ public:
         return dependence_mask;
     }
 
+    u64 GetAmmendValue() const {
+        return ammend_value;
+    }
+
+    void SetAccumulationValue(u64 new_value) {
+        acumulation_value = new_value;
+    }
+
 protected:
     void MakeDependent(StreamerInterface* depend_on) {
         dependence_mask |= 1ULL << depend_on->id;
@@ -87,6 +95,8 @@ protected:
     const size_t id;
     u64 dependence_mask;
     u64 dependent_mask;
+    u64 ammend_value{};
+    u64 acumulation_value{};
 };
 
 template <typename QueryType>
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.cpp b/src/video_core/renderer_vulkan/vk_query_cache.cpp
index add0c6fb3..2147776f8 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_query_cache.cpp
@@ -110,13 +110,16 @@ struct HostSyncValues {
 
 class SamplesStreamer : public BaseStreamer {
 public:
-    explicit SamplesStreamer(size_t id_, QueryCacheRuntime& runtime_, const Device& device_,
+    explicit SamplesStreamer(size_t id_, QueryCacheRuntime& runtime_,
+                             VideoCore::RasterizerInterface* rasterizer_, const Device& device_,
                              Scheduler& scheduler_, const MemoryAllocator& memory_allocator_)
-        : BaseStreamer(id_), runtime{runtime_}, device{device_}, scheduler{scheduler_},
-          memory_allocator{memory_allocator_} {
+        : BaseStreamer(id_), runtime{runtime_}, rasterizer{rasterizer_}, device{device_},
+          scheduler{scheduler_}, memory_allocator{memory_allocator_} {
         BuildResolveBuffer();
         current_bank = nullptr;
         current_query = nullptr;
+        ammend_value = 0;
+        acumulation_value = 0;
     }
 
     ~SamplesStreamer() = default;
@@ -151,6 +154,11 @@ public:
             PauseCounter();
         }
         AbandonCurrentQuery();
+        std::function<void()> func([this, counts = pending_flush_queries.size()] {
+            ammend_value = 0;
+            acumulation_value = 0;
+        });
+        rasterizer->SyncOperation(std::move(func));
     }
 
     void CloseCounter() override {
@@ -244,7 +252,7 @@ public:
             }
             if (query->size_slots > 1) {
                 // This is problematic.
-                UNIMPLEMENTED();
+                // UNIMPLEMENTED();
             }
             query->flags |= VideoCommon::QueryFlagBits::IsHostSynced;
             auto loc_data = offsets[query->start_bank_id];
@@ -255,16 +263,20 @@ public:
             });
         }
 
+        ReplicateCurrentQueryIfNeeded();
+        std::function<void()> func([this] { ammend_value = acumulation_value; });
+        rasterizer->SyncOperation(std::move(func));
         AbandonCurrentQuery();
         pending_sync.clear();
     }
 
     size_t WriteCounter(VAddr address, bool has_timestamp, u32 value,
                         [[maybe_unused]] std::optional<u32> subreport) override {
+        PauseCounter();
         auto index = BuildQuery();
         auto* new_query = GetQuery(index);
         new_query->guest_address = address;
-        new_query->value = 100;
+        new_query->value = 0;
         new_query->flags &= ~VideoCommon::QueryFlagBits::IsOrphan;
         if (has_timestamp) {
             new_query->flags |= VideoCommon::QueryFlagBits::HasTimestamp;
@@ -291,6 +303,7 @@ public:
 
     void PushUnsyncedQueries() override {
         PauseCounter();
+        current_bank->Close();
         {
             std::scoped_lock lk(flush_guard);
             pending_flush_sets.emplace_back(std::move(pending_flush_queries));
@@ -429,6 +442,34 @@ private:
         current_query_id = 0;
     }
 
+    void ReplicateCurrentQueryIfNeeded() {
+        if (pending_sync.empty()) {
+            return;
+        }
+        if (!current_query) {
+            return;
+        }
+        auto index = BuildQuery();
+        auto* new_query = GetQuery(index);
+        new_query->guest_address = 0;
+        new_query->value = 0;
+        new_query->flags &= ~VideoCommon::QueryFlagBits::IsOrphan;
+        new_query->start_bank_id = current_query->start_bank_id;
+        new_query->size_banks = current_query->size_banks;
+        new_query->start_slot = current_query->start_slot;
+        new_query->size_slots = current_query->size_slots;
+        ApplyBankOp(new_query, [](SamplesQueryBank* bank, size_t start, size_t amount) {
+            bank->AddReference(amount);
+        });
+        pending_flush_queries.push_back(index);
+        std::function<void()> func([this, index] {
+            auto* query = GetQuery(index);
+            query->value += GetAmmendValue();
+            SetAccumulationValue(query->value);
+            Free(index);
+        });
+    }
+
     void BuildResolveBuffer() {
         const VkBufferCreateInfo buffer_ci = {
             .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
@@ -448,6 +489,7 @@ private:
     static constexpr size_t resolve_slots = 8;
 
     QueryCacheRuntime& runtime;
+    VideoCore::RasterizerInterface* rasterizer;
     const Device& device;
     Scheduler& scheduler;
     const MemoryAllocator& memory_allocator;
@@ -470,6 +512,7 @@ private:
     size_t current_query_id;
     VideoCommon::HostQueryBase* current_query;
     bool has_started{};
+    bool current_unset{};
     std::mutex flush_guard;
 };
 
@@ -677,7 +720,6 @@ public:
         size_t offset_base = staging_ref.offset;
         for (auto q : pending_flush_queries) {
             auto* query = GetQuery(q);
-            query->flags |= VideoCommon::QueryFlagBits::IsQueuedForAsyncFlush;
             auto& bank = bank_pool.GetBank(query->start_bank_id);
             bank.Sync(staging_ref, offset_base, query->start_slot, 1);
             offset_base += TFBQueryBank::QUERY_SIZE;
@@ -1047,8 +1089,8 @@ struct QueryCacheRuntimeImpl {
           buffer_cache{buffer_cache_}, device{device_},
           memory_allocator{memory_allocator_}, scheduler{scheduler_}, staging_pool{staging_pool_},
           guest_streamer(0, runtime),
-          sample_streamer(static_cast<size_t>(QueryType::ZPassPixelCount64), runtime, device,
-                          scheduler, memory_allocator),
+          sample_streamer(static_cast<size_t>(QueryType::ZPassPixelCount64), runtime, rasterizer,
+                          device, scheduler, memory_allocator),
           tfb_streamer(static_cast<size_t>(QueryType::StreamingByteCount), runtime, device,
                        scheduler, memory_allocator, staging_pool),
           primitives_succeeded_streamer(
@@ -1277,6 +1319,10 @@ bool QueryCacheRuntime::HostConditionalRenderingCompareValues(VideoCommon::Looku
             return true;
         }
     }
+    if (!is_in_bc[0] && !is_in_bc[1]) {
+        // Both queries are in query cache, it's best to just flush.
+        return false;
+    }
     HostConditionalRenderingCompareBCImpl(object_1.address, equal_check);
     return true;
 }