early-access version 3527

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 3524.
+This is the source code for early-access 3527.
 
 ## Legal Notice
 

src/core/memory.cpp

@@ -462,7 +462,7 @@ struct Memory::Impl {
         }
 
         if (Settings::IsFastmemEnabled()) {
-            const bool is_read_enable = Settings::IsGPULevelHigh() || !cached;
+            const bool is_read_enable = !Settings::IsGPULevelExtreme() || !cached;
             system.DeviceMemory().buffer.Protect(vaddr, size, is_read_enable, !cached);
         }
 

src/video_core/buffer_cache/buffer_cache.h

@@ -22,7 +22,7 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
     void(slot_buffers.insert(runtime, NullBufferParams{}));
     common_ranges.clear();
 
-    active_async_buffers = IMPLEMENTS_ASYNC_DOWNLOADS && !Settings::IsGPULevelHigh();
+    active_async_buffers = !Settings::IsGPULevelHigh();
 
     if (!runtime.CanReportMemoryUsage()) {
         minimum_memory = DEFAULT_EXPECTED_MEMORY;
@@ -74,7 +74,7 @@ void BufferCache<P>::TickFrame() {
     uniform_cache_hits[0] = 0;
     uniform_cache_shots[0] = 0;
 
-    active_async_buffers = IMPLEMENTS_ASYNC_DOWNLOADS && !Settings::IsGPULevelHigh();
+    active_async_buffers = !Settings::IsGPULevelHigh();
 
     const bool skip_preferred = hits * 256 < shots * 251;
     uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
@@ -88,6 +88,13 @@ void BufferCache<P>::TickFrame() {
     }
     ++frame_tick;
     delayed_destruction_ring.Tick();
+
+    if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
+        for (auto& buffer : async_buffers_death_ring) {
+            runtime.FreeDeferredStagingBuffer(buffer);
+        }
+        async_buffers_death_ring.clear();
+    }
 }
 
 template <class P>
@@ -468,9 +475,11 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
     AccumulateFlushes();
 
     if (committed_ranges.empty()) {
+        if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
             if (active_async_buffers) {
                 async_buffers.emplace_back(std::optional<Async_Buffer>{});
             }
+        }
         return;
     }
     MICROPROFILE_SCOPE(GPU_DownloadMemory);
@@ -529,9 +538,11 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
     }
     committed_ranges.clear();
     if (downloads.empty()) {
+        if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
             if (active_async_buffers) {
                 async_buffers.emplace_back(std::optional<Async_Buffer>{});
             }
+        }
         return;
     }
     if (active_async_buffers) {
@@ -555,6 +566,9 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
             runtime.PostCopyBarrier();
             pending_downloads.emplace_back(std::move(normalized_copies));
             async_buffers.emplace_back(download_staging);
+        } else {
+            committed_ranges.clear();
+            uncommitted_ranges.clear();
         }
     } else {
         if constexpr (USE_MEMORY_MAPS) {
@@ -629,7 +643,7 @@ void BufferCache<P>::PopAsyncBuffers() {
             const IntervalType subtract_interval{cpu_addr, cpu_addr + copy.size};
             RemoveEachInOverlapCounter(async_downloads, subtract_interval, -1);
         }
-        runtime.FreeDeferredStagingBuffer(*async_buffer);
+        async_buffers_death_ring.emplace_back(*async_buffer);
         async_buffers.pop_front();
         pending_downloads.pop_front();
     }

src/video_core/buffer_cache/buffer_cache_base.h

@@ -554,11 +554,7 @@ private:
     std::deque<boost::container::small_vector<BufferCopy, 4>> pending_downloads;
     std::optional<Async_Buffer> current_buffer;
 
-    // queries
+    std::deque<Async_Buffer> async_buffers_death_ring;
-    boost::container::small_vector<std::pair<VAddr, size_t>, 8> pending_queries;
-    std::deque<boost::container::small_vector<BufferCopy, 8>> committed_queries;
-    boost::container::small_vector<u64, 8> flushed_queries;
-    std::deque<std::optional<Async_Buffer>> query_async_buffers;
 
     size_t immediate_buffer_capacity = 0;
     Common::ScratchBuffer<u8> immediate_buffer_alloc;

src/video_core/fence_manager.h

@@ -4,13 +4,20 @@
 #pragma once
 
 #include <algorithm>
+#include <condition_variable>
 #include <cstring>
 #include <deque>
 #include <functional>
 #include <memory>
+#include <mutex>
+#include <thread>
 #include <queue>
 
 #include "common/common_types.h"
+#include "common/microprofile.h"
+#include "common/scope_exit.h"
+#include "common/settings.h"
+#include "common/thread.h"
 #include "video_core/delayed_destruction_ring.h"
 #include "video_core/gpu.h"
 #include "video_core/host1x/host1x.h"
@@ -23,15 +30,26 @@ class FenceBase {
 public:
     explicit FenceBase(bool is_stubbed_) : is_stubbed{is_stubbed_} {}
 
+    bool IsStubbed() const {
+        return is_stubbed;
+    }
+
 protected:
     bool is_stubbed;
 };
 
-template <typename TFence, typename TTextureCache, typename TTBufferCache, typename TQueryCache>
+template <typename Traits>
 class FenceManager {
+    using TFence = typename Traits::FenceType;
+    using TTextureCache = typename Traits::TextureCacheType;
+    using TBufferCache = typename Traits::BufferCacheType;
+    using TQueryCache = typename Traits::QueryCacheType;
+    static constexpr bool can_async_check = Traits::HAS_ASYNC_CHECK;
+
 public:
     /// Notify the fence manager about a new frame
     void TickFrame() {
+        std::unique_lock lock(ring_guard);
         delayed_destruction_ring.Tick();
     }
 
@@ -46,17 +64,32 @@ public:
     }
 
     void SignalFence(std::function<void()>&& func) {
-        TryReleasePendingFences();
+        bool delay_fence = Settings::IsGPULevelHigh();
+        if constexpr (!can_async_check) {
+            TryReleasePendingFences<false>();
+        }
         const bool should_flush = ShouldFlush();
         CommitAsyncFlushes();
-        uncommitted_operations.emplace_back(std::move(func));
-        CommitOperations();
         TFence new_fence = CreateFence(!should_flush);
-        fences.push(new_fence);
+        if constexpr (can_async_check) {
+            guard.lock();
+        }
+        if (delay_fence) {
+            uncommitted_operations.emplace_back(std::move(func));
+        }
+        pending_operations.emplace_back(std::move(uncommitted_operations));
         QueueFence(new_fence);
+        if (!delay_fence) {
+            func();
+        }
+        fences.push(std::move(new_fence));
         if (should_flush) {
             rasterizer.FlushCommands();
         }
+        if constexpr (can_async_check) {
+            guard.unlock();
+            cv.notify_all();
+        }
     }
 
     void SignalSyncPoint(u32 value) {
@@ -66,29 +99,30 @@ public:
     }
 
     void WaitPendingFences() {
-        while (!fences.empty()) {
+        if constexpr (!can_async_check) {
-            TFence& current_fence = fences.front();
+            TryReleasePendingFences<true>();
-            if (ShouldWait()) {
-                WaitFence(current_fence);
-            }
-            PopAsyncFlushes();
-            auto operations = std::move(pending_operations.front());
-            pending_operations.pop_front();
-            for (auto& operation : operations) {
-                operation();
-            }
-            PopFence();
         }
     }
 
 protected:
     explicit FenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
-                          TTextureCache& texture_cache_, TTBufferCache& buffer_cache_,
+                          TTextureCache& texture_cache_, TBufferCache& buffer_cache_,
                           TQueryCache& query_cache_)
         : rasterizer{rasterizer_}, gpu{gpu_}, syncpoint_manager{gpu.Host1x().GetSyncpointManager()},
-          texture_cache{texture_cache_}, buffer_cache{buffer_cache_}, query_cache{query_cache_} {}
+          texture_cache{texture_cache_}, buffer_cache{buffer_cache_}, query_cache{query_cache_} {
+        if constexpr (can_async_check) {
+            fence_thread =
+                std::jthread([this](std::stop_token token) { ReleaseThreadFunc(token); });
+        }
+    }
 
-    virtual ~FenceManager() = default;
+    virtual ~FenceManager() {
+        if constexpr (can_async_check) {
+            fence_thread.request_stop();
+            cv.notify_all();
+            fence_thread.join();
+        }
+    }
 
     /// Creates a Fence Interface, does not create a backend fence if 'is_stubbed' is
     /// true
@@ -104,23 +138,70 @@ protected:
     Tegra::GPU& gpu;
     Tegra::Host1x::SyncpointManager& syncpoint_manager;
     TTextureCache& texture_cache;
-    TTBufferCache& buffer_cache;
+    TBufferCache& buffer_cache;
     TQueryCache& query_cache;
 
 private:
+    template <bool force_wait>
     void TryReleasePendingFences() {
         while (!fences.empty()) {
             TFence& current_fence = fences.front();
             if (ShouldWait() && !IsFenceSignaled(current_fence)) {
+                if constexpr (force_wait) {
+                    WaitFence(current_fence);
+                } else {
                     return;
                 }
+            }
             PopAsyncFlushes();
             auto operations = std::move(pending_operations.front());
             pending_operations.pop_front();
             for (auto& operation : operations) {
                 operation();
             }
-            PopFence();
+            {
+                std::unique_lock lock(ring_guard);
+                delayed_destruction_ring.Push(std::move(current_fence));
+            }
+            fences.pop();
+        }
+    }
+
+    void ReleaseThreadFunc(std::stop_token stop_token) {
+        std::string name = "GPUFencingThread";
+        MicroProfileOnThreadCreate(name.c_str());
+
+        // Cleanup
+        SCOPE_EXIT({ MicroProfileOnThreadExit(); });
+
+        Common::SetCurrentThreadName(name.c_str());
+        Common::SetCurrentThreadPriority(Common::ThreadPriority::High);
+
+        TFence current_fence;
+        std::deque<std::function<void()>> current_operations;
+        while (!stop_token.stop_requested()) {
+            {
+                std::unique_lock lock(guard);
+                cv.wait(lock, [&] { return stop_token.stop_requested() || !fences.empty(); });
+                if (stop_token.stop_requested()) [[unlikely]] {
+                    return;
+                }
+                current_fence = std::move(fences.front());
+                current_operations = std::move(pending_operations.front());
+                fences.pop();
+                pending_operations.pop_front();
+            }
+            if (!current_fence->IsStubbed()) {
+                WaitFence(current_fence);
+            }
+            PopAsyncFlushes();
+            for (auto& operation : current_operations) {
+                operation();
+            }
+            {
+                std::unique_lock lock(ring_guard);
+                delayed_destruction_ring.Push(std::move(current_fence));
+            }
         }
     }
 
@@ -154,19 +235,16 @@ private:
         query_cache.CommitAsyncFlushes();
     }
 
-    void PopFence() {
-        delayed_destruction_ring.Push(std::move(fences.front()));
-        fences.pop();
-    }
-
-    void CommitOperations() {
-        pending_operations.emplace_back(std::move(uncommitted_operations));
-    }
-
     std::queue<TFence> fences;
     std::deque<std::function<void()>> uncommitted_operations;
     std::deque<std::deque<std::function<void()>>> pending_operations;
 
+    std::mutex guard;
+    std::mutex ring_guard;
+    std::condition_variable cv;
+
+    std::jthread fence_thread;
+
     DelayedDestructionRing<TFence, 6> delayed_destruction_ring;
 };
 

src/video_core/query_cache.h

@@ -6,6 +6,7 @@
 #include <algorithm>
 #include <array>
 #include <cstring>
+#include <functional>
 #include <iterator>
 #include <list>
 #include <memory>
@@ -17,13 +18,19 @@
 
 #include "common/assert.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/control/channel_state_cache.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
+#include "video_core/texture_cache/slot_vector.h"
 
 namespace VideoCommon {
 
+using AsyncJobId = SlotId;
+
+static constexpr AsyncJobId NULL_ASYNC_JOB_ID{0};
+
 template <class QueryCache, class HostCounter>
 class CounterStreamBase {
 public:
@@ -93,9 +100,13 @@ private:
 template <class QueryCache, class CachedQuery, class CounterStream, class HostCounter>
 class QueryCacheBase : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo> {
 public:
-    explicit QueryCacheBase(VideoCore::RasterizerInterface& rasterizer_)
+    explicit QueryCacheBase(VideoCore::RasterizerInterface& rasterizer_,
-        : rasterizer{rasterizer_}, streams{{CounterStream{static_cast<QueryCache&>(*this),
+                            Core::Memory::Memory& cpu_memory_)
-                                                          VideoCore::QueryType::SamplesPassed}}} {}
+        : rasterizer{rasterizer_},
+          cpu_memory{cpu_memory_}, streams{{CounterStream{static_cast<QueryCache&>(*this),
+                                                          VideoCore::QueryType::SamplesPassed}}} {
+        (void)slot_async_jobs.insert(); // Null value
+    }
 
     void InvalidateRegion(VAddr addr, std::size_t size) {
         std::unique_lock lock{mutex};
@@ -126,10 +137,15 @@ public:
             query = Register(type, *cpu_addr, host_ptr, timestamp.has_value());
         }
 
-        query->BindCounter(Stream(type).Current(), timestamp);
+        auto result = query->BindCounter(Stream(type).Current());
-        if (Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
+        if (result) {
-            AsyncFlushQuery(*cpu_addr);
+            auto async_job_id = query->GetAsyncJob();
+            auto& async_job = slot_async_jobs[async_job_id];
+            async_job.collected = true;
+            async_job.value = *result;
+            query->SetAsyncJob(NULL_ASYNC_JOB_ID);
         }
+        AsyncFlushQuery(query, timestamp, lock);
     }
 
     /// Updates counters from GPU state. Expected to be called once per draw, clear or dispatch.
@@ -173,15 +189,18 @@ public:
     }
 
     void CommitAsyncFlushes() {
+        std::unique_lock lock{mutex};
         committed_flushes.push_back(uncommitted_flushes);
         uncommitted_flushes.reset();
     }
 
     bool HasUncommittedFlushes() const {
+        std::unique_lock lock{mutex};
         return uncommitted_flushes != nullptr;
     }
 
     bool ShouldWaitAsyncFlushes() const {
+        std::unique_lock lock{mutex};
         if (committed_flushes.empty()) {
             return false;
         }
@@ -189,6 +208,7 @@ public:
     }
 
     void PopAsyncFlushes() {
+        std::unique_lock lock{mutex};
         if (committed_flushes.empty()) {
             return;
         }
@@ -197,15 +217,25 @@ public:
             committed_flushes.pop_front();
             return;
         }
-        for (VAddr query_address : *flush_list) {
+        for (AsyncJobId async_job_id : *flush_list) {
-            FlushAndRemoveRegion(query_address, 4);
+            AsyncJob& async_job = slot_async_jobs[async_job_id];
+            if (!async_job.collected) {
+                FlushAndRemoveRegion(async_job.query_location, 2, true);
+            }
         }
         committed_flushes.pop_front();
     }
 
 private:
+    struct AsyncJob {
+        bool collected = false;
+        u64 value = 0;
+        VAddr query_location = 0;
+        std::optional<u64> timestamp{};
+    };
+
     /// Flushes a memory range to guest memory and removes it from the cache.
-    void FlushAndRemoveRegion(VAddr addr, std::size_t size) {
+    void FlushAndRemoveRegion(VAddr addr, std::size_t size, bool async = false) {
         const u64 addr_begin = addr;
         const u64 addr_end = addr_begin + size;
         const auto in_range = [addr_begin, addr_end](const CachedQuery& query) {
@@ -226,7 +256,16 @@ private:
                     continue;
                 }
                 rasterizer.UpdatePagesCachedCount(query.GetCpuAddr(), query.SizeInBytes(), -1);
-                query.Flush();
+                AsyncJobId async_job_id = query.GetAsyncJob();
+                auto flush_result = query.Flush(async);
+                if (async_job_id == NULL_ASYNC_JOB_ID) {
+                    ASSERT_MSG(false, "This should not be reachable at all");
+                    continue;
+                }
+                AsyncJob& async_job = slot_async_jobs[async_job_id];
+                async_job.collected = true;
+                async_job.value = flush_result;
+                query.SetAsyncJob(NULL_ASYNC_JOB_ID);
             }
             std::erase_if(contents, in_range);
         }
@@ -253,26 +292,54 @@ private:
         return found != std::end(contents) ? &*found : nullptr;
     }
 
-    void AsyncFlushQuery(VAddr addr) {
+    void AsyncFlushQuery(CachedQuery* query, std::optional<u64> timestamp,
+                         std::unique_lock<std::recursive_mutex>& lock) {
+        const AsyncJobId new_async_job_id = slot_async_jobs.insert();
+        {
+            AsyncJob& async_job = slot_async_jobs[new_async_job_id];
+            query->SetAsyncJob(new_async_job_id);
+            async_job.query_location = query->GetCpuAddr();
+            async_job.collected = false;
+
             if (!uncommitted_flushes) {
-            uncommitted_flushes = std::make_shared<std::vector<VAddr>>();
+                uncommitted_flushes = std::make_shared<std::vector<AsyncJobId>>();
             }
-        uncommitted_flushes->push_back(addr);
+            uncommitted_flushes->push_back(new_async_job_id);
+        }
+        lock.unlock();
+        std::function<void()> operation([this, new_async_job_id, timestamp] {
+            std::unique_lock local_lock{mutex};
+            AsyncJob& async_job = slot_async_jobs[new_async_job_id];
+            if (timestamp) {
+                u64 timestamp_value = *timestamp;
+                cpu_memory.WriteBlockUnsafe(async_job.query_location + sizeof(u64),
+                                            &timestamp_value, sizeof(8));
+                cpu_memory.WriteBlockUnsafe(async_job.query_location, &async_job.value, sizeof(8));
+            } else {
+                u32 small_value = static_cast<u32>(async_job.value);
+                cpu_memory.WriteBlockUnsafe(async_job.query_location, &small_value, sizeof(u32));
+            }
+            slot_async_jobs.erase(new_async_job_id);
+        });
+        rasterizer.SyncOperation(std::move(operation));
     }
 
     static constexpr std::uintptr_t YUZU_PAGESIZE = 4096;
     static constexpr unsigned YUZU_PAGEBITS = 12;
 
-    VideoCore::RasterizerInterface& rasterizer;
+    SlotVector<AsyncJob> slot_async_jobs;
 
-    std::recursive_mutex mutex;
+    VideoCore::RasterizerInterface& rasterizer;
+    Core::Memory::Memory& cpu_memory;
+
+    mutable std::recursive_mutex mutex;
 
     std::unordered_map<u64, std::vector<CachedQuery>> cached_queries;
 
     std::array<CounterStream, VideoCore::NumQueryTypes> streams;
 
-    std::shared_ptr<std::vector<VAddr>> uncommitted_flushes{};
+    std::shared_ptr<std::vector<AsyncJobId>> uncommitted_flushes{};
-    std::list<std::shared_ptr<std::vector<VAddr>>> committed_flushes;
+    std::list<std::shared_ptr<std::vector<AsyncJobId>>> committed_flushes;
 };
 
 template <class QueryCache, class HostCounter>
@@ -291,12 +358,12 @@ public:
     virtual ~HostCounterBase() = default;
 
     /// Returns the current value of the query.
-    u64 Query() {
+    u64 Query(bool async = false) {
         if (result) {
             return *result;
         }
 
-        u64 value = BlockingQuery() + base_result;
+        u64 value = BlockingQuery(async) + base_result;
         if (dependency) {
             value += dependency->Query();
             dependency = nullptr;
@@ -317,7 +384,7 @@ public:
 
 protected:
     /// Returns the value of query from the backend API blocking as needed.
-    virtual u64 BlockingQuery() const = 0;
+    virtual u64 BlockingQuery(bool async = false) const = 0;
 
 private:
     std::shared_ptr<HostCounter> dependency; ///< Counter to add to this value.
@@ -340,26 +407,23 @@ public:
     CachedQueryBase& operator=(const CachedQueryBase&) = delete;
 
     /// Flushes the query to guest memory.
-    virtual void Flush() {
+    virtual u64 Flush(bool async = false) {
         // When counter is nullptr it means that it's just been reset. We are supposed to write a
         // zero in these cases.
-        const u64 value = counter ? counter->Query() : 0;
+        const u64 value = counter ? counter->Query(async) : 0;
-        std::memcpy(host_ptr, &value, sizeof(u64));
+        return value;
-
-        if (timestamp) {
-            std::memcpy(host_ptr + TIMESTAMP_OFFSET, &*timestamp, sizeof(u64));
-        }
     }
 
     /// Binds a counter to this query.
-    void BindCounter(std::shared_ptr<HostCounter> counter_, std::optional<u64> timestamp_) {
+    std::optional<u64> BindCounter(std::shared_ptr<HostCounter> counter_) {
+        std::optional<u64> result{};
         if (counter) {
             // If there's an old counter set it means the query is being rewritten by the game.
             // To avoid losing the data forever, flush here.
-            Flush();
+            result = std::make_optional(Flush());
         }
         counter = std::move(counter_);
-        timestamp = timestamp_;
+        return result;
     }
 
     VAddr GetCpuAddr() const noexcept {
@@ -374,6 +438,14 @@ public:
         return with_timestamp ? LARGE_QUERY_SIZE : SMALL_QUERY_SIZE;
     }
 
+    void SetAsyncJob(AsyncJobId assigned_async_job_) {
+        assigned_async_job = assigned_async_job_;
+    }
+
+    AsyncJobId GetAsyncJob() const {
+        return assigned_async_job;
+    }
+
 protected:
     /// Returns true when querying the counter may potentially block.
     bool WaitPending() const noexcept {
@@ -389,6 +461,7 @@ private:
     u8* host_ptr;                         ///< Writable host pointer.
     std::shared_ptr<HostCounter> counter; ///< Host counter to query, owns the dependency tree.
     std::optional<u64> timestamp;         ///< Timestamp to flush to guest memory.
+    AsyncJobId assigned_async_job;
 };
 
 } // namespace VideoCommon

src/video_core/renderer_opengl/gl_fence_manager.h

@@ -30,7 +30,17 @@ private:
 };
 
 using Fence = std::shared_ptr<GLInnerFence>;
-using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCache, BufferCache, QueryCache>;
+
+struct FenceManagerParams {
+    using FenceType = Fence;
+    using BufferCacheType = BufferCache;
+    using TextureCacheType = TextureCache;
+    using QueryCacheType = QueryCache;
+
+    static constexpr bool HAS_ASYNC_CHECK = false;
+};
+
+using GenericFenceManager = VideoCommon::FenceManager<FenceManagerParams>;
 
 class FenceManagerOpenGL final : public GenericFenceManager {
 public:

src/video_core/renderer_opengl/gl_query_cache.cpp

@@ -26,8 +26,8 @@ constexpr GLenum GetTarget(VideoCore::QueryType type) {
 
 } // Anonymous namespace
 
-QueryCache::QueryCache(RasterizerOpenGL& rasterizer_)
+QueryCache::QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_)
-    : QueryCacheBase(rasterizer_), gl_rasterizer{rasterizer_} {}
+    : QueryCacheBase(rasterizer_, cpu_memory_), gl_rasterizer{rasterizer_} {}
 
 QueryCache::~QueryCache() = default;
 
@@ -74,7 +74,7 @@ void HostCounter::EndQuery() {
     glEndQuery(GetTarget(type));
 }
 
-u64 HostCounter::BlockingQuery() const {
+u64 HostCounter::BlockingQuery([[maybe_unused]] bool async) const {
     GLint64 value;
     glGetQueryObjecti64v(query.handle, GL_QUERY_RESULT, &value);
     return static_cast<u64>(value);
@@ -96,7 +96,7 @@ CachedQuery& CachedQuery::operator=(CachedQuery&& rhs) noexcept {
     return *this;
 }
 
-void CachedQuery::Flush() {
+u64 CachedQuery::Flush([[maybe_unused]] bool async) {
     // Waiting for a query while another query of the same target is enabled locks Nvidia's driver.
     // To avoid this disable and re-enable keeping the dependency stream.
     // But we only have to do this if we have pending waits to be done.
@@ -106,11 +106,13 @@ void CachedQuery::Flush() {
         stream.Update(false);
     }
 
-    VideoCommon::CachedQueryBase<HostCounter>::Flush();
+    auto result = VideoCommon::CachedQueryBase<HostCounter>::Flush();
 
     if (slice_counter) {
         stream.Update(true);
     }
+
+    return result;
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_query_cache.h

@@ -28,7 +28,7 @@ using CounterStream = VideoCommon::CounterStreamBase<QueryCache, HostCounter>;
 class QueryCache final
     : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
 public:
-    explicit QueryCache(RasterizerOpenGL& rasterizer_);
+    explicit QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_);
     ~QueryCache();
 
     OGLQuery AllocateQuery(VideoCore::QueryType type);
@@ -51,7 +51,7 @@ public:
     void EndQuery();
 
 private:
-    u64 BlockingQuery() const override;
+    u64 BlockingQuery(bool async = false) const override;
 
     QueryCache& cache;
     const VideoCore::QueryType type;
@@ -70,7 +70,7 @@ public:
     CachedQuery(const CachedQuery&) = delete;
     CachedQuery& operator=(const CachedQuery&) = delete;
 
-    void Flush() override;
+    u64 Flush(bool async = false) override;
 
 private:
     QueryCache* cache;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -63,7 +63,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
       shader_cache(*this, emu_window_, device, texture_cache, buffer_cache, program_manager,
                    state_tracker, gpu.ShaderNotify()),
-      query_cache(*this), accelerate_dma(buffer_cache, texture_cache),
+      query_cache(*this, cpu_memory_), accelerate_dma(buffer_cache, texture_cache),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache),
       blit_image(program_manager_) {}
 

src/video_core/renderer_vulkan/vk_fence_manager.cpp

@@ -5,6 +5,7 @@
 
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_fence_manager.h"
+#include "video_core/renderer_vulkan/vk_query_cache.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_texture_cache.h"
 #include "video_core/vulkan_common/vulkan_device.h"

src/video_core/renderer_vulkan/vk_fence_manager.h

@@ -40,7 +40,16 @@ private:
 };
 using Fence = std::shared_ptr<InnerFence>;
 
-using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCache, BufferCache, QueryCache>;
+struct FenceManagerParams {
+    using FenceType = Fence;
+    using BufferCacheType = BufferCache;
+    using TextureCacheType = TextureCache;
+    using QueryCacheType = QueryCache;
+
+    static constexpr bool HAS_ASYNC_CHECK = true;
+};
+
+using GenericFenceManager = VideoCommon::FenceManager<FenceManagerParams>;
 
 class FenceManager final : public GenericFenceManager {
 public:

src/video_core/renderer_vulkan/vk_query_cache.cpp

@@ -66,9 +66,10 @@ void QueryPool::Reserve(std::pair<VkQueryPool, u32> query) {
     }
 }
 
-QueryCache::QueryCache(VideoCore::RasterizerInterface& rasterizer_, const Device& device_,
+QueryCache::QueryCache(VideoCore::RasterizerInterface& rasterizer_,
+                       Core::Memory::Memory& cpu_memory_, const Device& device_,
                        Scheduler& scheduler_)
-    : QueryCacheBase{rasterizer_}, device{device_}, scheduler{scheduler_},
+    : QueryCacheBase{rasterizer_, cpu_memory_}, device{device_}, scheduler{scheduler_},
       query_pools{
           QueryPool{device_, scheduler_, QueryType::SamplesPassed},
       } {}
@@ -98,8 +99,10 @@ HostCounter::HostCounter(QueryCache& cache_, std::shared_ptr<HostCounter> depend
       query{cache_.AllocateQuery(type_)}, tick{cache_.GetScheduler().CurrentTick()} {
     const vk::Device* logical = &cache.GetDevice().GetLogical();
     cache.GetScheduler().Record([logical, query = query](vk::CommandBuffer cmdbuf) {
+        const bool use_precise = Settings::IsGPULevelHigh();
         logical->ResetQueryPool(query.first, query.second, 1);
-        cmdbuf.BeginQuery(query.first, query.second, VK_QUERY_CONTROL_PRECISE_BIT);
+        cmdbuf.BeginQuery(query.first, query.second,
+                          use_precise ? VK_QUERY_CONTROL_PRECISE_BIT : 0);
     });
 }
 
@@ -112,8 +115,10 @@ void HostCounter::EndQuery() {
         [query = query](vk::CommandBuffer cmdbuf) { cmdbuf.EndQuery(query.first, query.second); });
 }
 
-u64 HostCounter::BlockingQuery() const {
+u64 HostCounter::BlockingQuery(bool async) const {
+    if (!async) {
         cache.GetScheduler().Wait(tick);
+    }
     u64 data;
     const VkResult query_result = cache.GetDevice().GetLogical().GetQueryResults(
         query.first, query.second, 1, sizeof(data), &data, sizeof(data),

src/video_core/renderer_vulkan/vk_query_cache.h

@@ -52,7 +52,8 @@ private:
 class QueryCache final
     : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
 public:
-    explicit QueryCache(VideoCore::RasterizerInterface& rasterizer_, const Device& device_,
+    explicit QueryCache(VideoCore::RasterizerInterface& rasterizer_,
+                        Core::Memory::Memory& cpu_memory_, const Device& device_,
                         Scheduler& scheduler_);
     ~QueryCache();
 
@@ -83,7 +84,7 @@ public:
     void EndQuery();
 
 private:
-    u64 BlockingQuery() const override;
+    u64 BlockingQuery(bool async = false) const override;
 
     QueryCache& cache;
     const VideoCore::QueryType type;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -172,7 +172,8 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
       pipeline_cache(*this, device, scheduler, descriptor_pool, update_descriptor_queue,
                      render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()),
-      query_cache{*this, device, scheduler}, accelerate_dma(buffer_cache, texture_cache, scheduler),
+      query_cache{*this, cpu_memory_, device, scheduler},
+      accelerate_dma(buffer_cache, texture_cache, scheduler),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
     scheduler.SetQueryCache(query_cache);
@@ -675,7 +676,8 @@ bool RasterizerVulkan::AccelerateConditionalRendering() {
     const GPUVAddr condition_address{maxwell3d->regs.render_enable.Address()};
     Maxwell::ReportSemaphore::Compare cmp;
     if (gpu_memory->IsMemoryDirty(condition_address, sizeof(cmp),
-                                  VideoCommon::CacheType::BufferCache)) {
+                                  VideoCommon::CacheType::BufferCache |
+                                      VideoCommon::CacheType::QueryCache)) {
         return true;
     }
     return false;

src/video_core/texture_cache/texture_cache.h

@@ -139,6 +139,13 @@ void TextureCache<P>::TickFrame() {
     runtime.TickFrame();
     critical_gc = 0;
     ++frame_tick;
+
+    if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
+        for (auto& buffer : async_buffers_death_ring) {
+            runtime.FreeDeferredStagingBuffer(buffer);
+        }
+        async_buffers_death_ring.clear();
+    }
 }
 
 template <class P>
@@ -688,10 +695,10 @@ void TextureCache<P>::CommitAsyncFlushes() {
             }
             uncommitted_async_buffers.emplace_back(download_map);
         }
-    }
-    committed_downloads.emplace_back(std::move(uncommitted_downloads));
         async_buffers.emplace_back(std::move(uncommitted_async_buffers));
         uncommitted_async_buffers.clear();
+    }
+    committed_downloads.emplace_back(std::move(uncommitted_downloads));
     uncommitted_downloads.clear();
 }
 
@@ -729,7 +736,7 @@ void TextureCache<P>::PopAsyncFlushes() {
             }
         }
         for (auto& download_buffer : download_map) {
-            runtime.FreeDeferredStagingBuffer(download_buffer);
+            async_buffers_death_ring.emplace_back(download_buffer);
         }
         committed_downloads.pop_front();
         async_buffers.pop_front();

src/video_core/texture_cache/texture_cache_base.h

@@ -449,6 +449,7 @@ private:
     std::deque<std::vector<PendingDownload>> committed_downloads;
     std::vector<AsyncBuffer> uncommitted_async_buffers;
     std::deque<std::vector<AsyncBuffer>> async_buffers;
+    std::deque<AsyncBuffer> async_buffers_death_ring;
 
     struct LRUItemParams {
         using ObjectType = ImageId;