[TEST] Line/ conservative rasterization adjustment

src/android/app/build.gradle.kts

@@ -65,7 +65,7 @@ android {
 
     defaultConfig {
         applicationId = "dev.eden.eden_emulator"
-        minSdk = 24
+        minSdk = 33
         targetSdk = 36
         versionName = getGitVersion()
         versionCode = autoVersion

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/model/view/SettingsItem.kt

@@ -594,7 +594,7 @@ abstract class SettingsItem(
                     IntSetting.ANDROID_PIPELINE_WORKERS,
                     titleId = R.string.pipeline_worker_cores,
                     descriptionId = R.string.pipeline_worker_cores_description,
-                    min = 4,
+                    min = 2,
                     max = 8,
                     units = "cores"
                 )

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/ui/InputDialogFragment.kt

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: 2024 yuzu Emulator Project
@@ -169,7 +169,7 @@ class InputDialogFragment : DialogFragment() {
         NativeInput.onGamePadButtonEvent(
             controllerData.getGUID(),
             controllerData.getPort(),
-            InputHandler.getButtonIdFromEvent(event),
+            event.keyCode,
             action
         )
         onInputReceived(event.device)

src/android/app/src/main/java/org/yuzu/yuzu_emu/utils/InputHandler.kt

@@ -49,12 +49,6 @@ object InputHandler {
         MotionEvent.AXIS_RTRIGGER
     )
 
-    // Currently, Android doesn't support Joy-Con D-pad buttons. We fall back to the scan code
-    private const val LINUX_BUTTON_DPAD_UP = 0x220
-    private const val LINUX_BUTTON_DPAD_DOWN = 0x221
-    private const val LINUX_BUTTON_DPAD_LEFT = 0x222
-    private const val LINUX_BUTTON_DPAD_RIGHT = 0x223
-
     fun isPhysicalGameController(device: InputDevice?): Boolean {
         device ?: return false
 
@@ -93,25 +87,12 @@ object InputHandler {
         NativeInput.onGamePadButtonEvent(
             controllerData.getGUID(),
             controllerData.getPort(),
-            getButtonIdFromEvent(event),
+            event.keyCode,
             action
         )
         return true
     }
 
-    fun getButtonIdFromEvent(event: KeyEvent): Int {
-        if (event.keyCode == 0) {
-            return when (event.scanCode) {
-                LINUX_BUTTON_DPAD_UP -> KeyEvent.KEYCODE_DPAD_UP
-                LINUX_BUTTON_DPAD_DOWN -> KeyEvent.KEYCODE_DPAD_DOWN
-                LINUX_BUTTON_DPAD_LEFT -> KeyEvent.KEYCODE_DPAD_LEFT
-                LINUX_BUTTON_DPAD_RIGHT -> KeyEvent.KEYCODE_DPAD_RIGHT
-                else -> return 0
-            }
-        }
-        return event.keyCode
-    }
-
     fun dispatchGenericMotionEvent(event: MotionEvent): Boolean {
         val controllerData =
             androidControllers[event.device.controllerNumber] ?: return false

src/common/thread.cpp

@@ -1,5 +1,6 @@
 // SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: 2013 Dolphin Emulator Project
 // SPDX-FileCopyrightText: 2014 Citra Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
@@ -39,6 +40,110 @@
 #include <unistd.h>
 #endif
 
+#ifdef __ANDROID__
+#include <sys/resource.h>
+#include <algorithm>
+#include <fstream>
+#include <utility>
+#include <vector>
+
+namespace {
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_URGENT_AUDIO = -19;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_AUDIO = -16;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_URGENT_DISPLAY = -8;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_DISPLAY = -4;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_FOREGROUND = -2;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_MORE_FAVORABLE = -1;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_DEFAULT = 0;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_LESS_FAVORABLE = 1;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_BACKGROUND = 10;
+[[maybe_unused]] constexpr int ANDROID_THREAD_PRIORITY_LOWEST = 19;
+
+constexpr size_t ANDROID_MINIMUM_PERFORMANCE_CORES = 4;
+
+cpu_set_t ComputePerformanceCoreMask() {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+
+    cpu_set_t allowed;
+    CPU_ZERO(&allowed);
+    if (sched_getaffinity(gettid(), sizeof(allowed), &allowed) != 0) {
+        return mask;
+    }
+
+    std::vector<std::pair<long, int>> cores;
+    const int total = static_cast<int>(std::thread::hardware_concurrency());
+    for (int cpu = 0; cpu < total; ++cpu) {
+        if (!CPU_ISSET(cpu, &allowed)) {
+            continue;
+        }
+        long max_frequency = 0;
+        std::ifstream file("/sys/devices/system/cpu/cpu" + std::to_string(cpu) +
+                           "/cpufreq/cpuinfo_max_freq");
+        if (!file || !(file >> max_frequency) || max_frequency <= 0) {
+            CPU_ZERO(&mask);
+            return mask;
+        }
+        cores.emplace_back(max_frequency, cpu);
+    }
+    if (cores.empty()) {
+        return mask;
+    }
+
+    std::sort(cores.begin(), cores.end(),
+              [](const auto& lhs, const auto& rhs) { return lhs.first > rhs.first; });
+
+    size_t taken = 0;
+    long cluster_frequency = cores.front().first;
+    for (const auto& [frequency, cpu] : cores) {
+        if (frequency != cluster_frequency) {
+            if (taken >= ANDROID_MINIMUM_PERFORMANCE_CORES) {
+                break;
+            }
+            cluster_frequency = frequency;
+        }
+        CPU_SET(cpu, &mask);
+        ++taken;
+    }
+    return mask;
+}
+
+const cpu_set_t& PerformanceCoreMask() {
+    static const cpu_set_t mask = ComputePerformanceCoreMask();
+    return mask;
+}
+
+cpu_set_t ComputeEfficiencyCoreMask() {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+
+    const cpu_set_t& performance = PerformanceCoreMask();
+    if (CPU_COUNT(&performance) == 0) {
+        return mask;
+    }
+
+    cpu_set_t allowed;
+    CPU_ZERO(&allowed);
+    if (sched_getaffinity(gettid(), sizeof(allowed), &allowed) != 0) {
+        return mask;
+    }
+
+    const int total = static_cast<int>(std::thread::hardware_concurrency());
+    for (int cpu = 0; cpu < total; ++cpu) {
+        if (CPU_ISSET(cpu, &allowed) && !CPU_ISSET(cpu, &performance)) {
+            CPU_SET(cpu, &mask);
+        }
+    }
+    return mask;
+}
+
+const cpu_set_t& EfficiencyCoreMask() {
+    static const cpu_set_t mask = ComputeEfficiencyCoreMask();
+    return mask;
+}
+} // Anonymous namespace
+#endif
+
 #include "common/cpu_features.h"
 #ifdef ARCHITECTURE_x86_64
 #ifdef _MSC_VER
@@ -78,6 +183,21 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
         }
     }();
     set_thread_priority(find_thread(NULL), priority);
+#elif defined(__ANDROID__)
+    const int nice_value = [&]() {
+        switch (new_priority) {
+        case ThreadPriority::Low: return ANDROID_THREAD_PRIORITY_BACKGROUND;
+        case ThreadPriority::Normal: return ANDROID_THREAD_PRIORITY_DEFAULT;
+        case ThreadPriority::High: return ANDROID_THREAD_PRIORITY_DISPLAY;
+        case ThreadPriority::VeryHigh: return ANDROID_THREAD_PRIORITY_URGENT_DISPLAY;
+        case ThreadPriority::Critical: return ANDROID_THREAD_PRIORITY_AUDIO;
+        default: return ANDROID_THREAD_PRIORITY_DEFAULT;
+        }
+    }();
+    if (setpriority(PRIO_PROCESS, static_cast<id_t>(gettid()), nice_value) != 0) {
+        LOG_DEBUG(Common, "Could not set thread nice value to {}: {}", nice_value,
+                  GetLastErrorMsg());
+    }
 #else
     pthread_t this_thread = pthread_self();
     const auto scheduling_type = SCHED_OTHER;
@@ -132,29 +252,28 @@ void SetCurrentThreadName(const char* name) {
 #endif
 }
 
-void PinCurrentThreadToPerformanceCore(size_t core_id) {
-    ASSERT(core_id < 4);
-    // If we set a flag for a CPU that doesn't exist, the thread may not be allowed to
-    // run in ANY processor!
-    auto const total_cores = std::thread::hardware_concurrency();
-    if (core_id < total_cores) {
+void SetCurrentThreadToPerformanceCores() {
 #if defined(__ANDROID__)
-        cpu_set_t set;
-        CPU_ZERO(&set);
-        CPU_SET(core_id, &set);
-        sched_setaffinity(pthread_self(), sizeof(set), &set);
-#elif defined(__linux__) || defined(__FreeBSD__)
-        cpu_set_t set;
-        CPU_ZERO(&set);
-        CPU_SET(core_id, &set);
-        pthread_setaffinity_np(pthread_self(), sizeof(set), &set);
-#elif defined(_WIN32)
-        DWORD set = 1UL << core_id;
-        SetThreadAffinityMask(GetCurrentThread(), set);
-#else
-        // No pin functionality implemented
-#endif
+    const cpu_set_t& mask = PerformanceCoreMask();
+    if (CPU_COUNT(&mask) == 0) {
+        return;
     }
+    if (sched_setaffinity(gettid(), sizeof(mask), &mask) != 0) {
+        LOG_DEBUG(Common, "Could not restrict thread to performance cores: {}", GetLastErrorMsg());
+    }
+#endif
+}
+
+void SetCurrentThreadToEfficiencyCores() {
+#if defined(__ANDROID__)
+    const cpu_set_t& mask = EfficiencyCoreMask();
+    if (CPU_COUNT(&mask) == 0) {
+        return;
+    }
+    if (sched_setaffinity(gettid(), sizeof(mask), &mask) != 0) {
+        LOG_DEBUG(Common, "Could not restrict thread to efficiency cores: {}", GetLastErrorMsg());
+    }
+#endif
 }
 
 #ifdef ARCHITECTURE_x86_64

src/common/thread.h

@@ -99,8 +99,14 @@ enum class ThreadPriority : u32 {
     Critical = 4,
 };
 
+enum class ThreadPlacement : u32 {
+    Default = 0,
+    Background = 1,
+};
+
 void SetCurrentThreadPriority(ThreadPriority new_priority);
 void SetCurrentThreadName(const char* name);
-void PinCurrentThreadToPerformanceCore(size_t core_id);
+void SetCurrentThreadToPerformanceCores();
+void SetCurrentThreadToEfficiencyCores();
 
 } // namespace Common

src/common/thread_worker.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
@@ -37,10 +37,15 @@ class StatefulThreadWorker {
     using StateMaker = std::conditional_t<with_state, std::function<StateType()>, DummyCallable>;
 
 public:
-    explicit StatefulThreadWorker(size_t num_workers, std::string name, StateMaker func = {})
+    explicit StatefulThreadWorker(size_t num_workers, std::string name, StateMaker func = {},
+                                  ThreadPlacement placement = ThreadPlacement::Default)
         : workers_queued{num_workers}, thread_name{std::move(name)} {
-        const auto lambda = [this, func](std::stop_token stop_token) {
+        const auto lambda = [this, func, placement](std::stop_token stop_token) {
             Common::SetCurrentThreadName(thread_name.c_str());
+            if (placement == ThreadPlacement::Background) {
+                Common::SetCurrentThreadPriority(ThreadPriority::Low);
+                Common::SetCurrentThreadToEfficiencyCores();
+            }
             {
                 [[maybe_unused]] std::conditional_t<with_state, StateType, int> state{func()};
                 while (!stop_token.stop_requested()) {

src/core/cpu_manager.cpp

@@ -174,12 +174,7 @@ void CpuManager::RunThread(std::stop_token token, std::size_t core) {
     std::string name = is_multicore ? ("CPUCore_" + std::to_string(core)) : std::string{"CPUThread"};
     Common::SetCurrentThreadName(name.c_str());
     Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
-#ifdef __ANDROID__
-    // Aimed specifically for Snapdragon 8 Elite devices
-    // This kills performance on desktop, but boosts perf for UMA devices
-    // like the S8E. Mediatek and Mali likely won't suffer.
-    Common::PinCurrentThreadToPerformanceCore(core);
-#endif
+    Common::SetCurrentThreadToPerformanceCores();
     auto& data = core_data[core];
     data.host_context = Common::Fiber::ThreadToFiber();
 

src/video_core/CMakeLists.txt

@@ -33,7 +33,7 @@ add_library(video_core STATIC
     control/channel_state_cache.h
     control/scheduler.cpp
     control/scheduler.h
-    delayed_destruction_ring.h
+    deferred_destruction_queue.h
     dirty_flags.cpp
     dirty_flags.h
     dma_pusher.cpp

src/video_core/buffer_cache/buffer_cache.h

@@ -31,44 +31,77 @@ BufferCache<P>::BufferCache(Tegra::MaxwellDeviceMemoryManager& device_memory_, R
     immediately_free = (Settings::values.vram_usage_mode.GetValue() == Settings::VramUsageMode::Aggressive);
 #endif
     if (!runtime.CanReportMemoryUsage()) {
-        minimum_memory = DEFAULT_EXPECTED_MEMORY;
-        critical_memory = DEFAULT_CRITICAL_MEMORY;
+        memory_budget = FALLBACK_MEMORY_BUDGET;
         return;
     }
 
-    const s64 device_local_memory = static_cast<s64>(runtime.GetDeviceLocalMemory());
-    const s64 min_spacing_expected = device_local_memory - 1_GiB;
-    const s64 min_spacing_critical = device_local_memory - 512_MiB;
-    const s64 mem_threshold = (std::min)(device_local_memory, TARGET_THRESHOLD);
-    const s64 min_vacancy_expected = (6 * mem_threshold) / 10;
-    const s64 min_vacancy_critical = (2 * mem_threshold) / 10;
-    minimum_memory = static_cast<u64>(
-        (std::max)((std::min)(device_local_memory - min_vacancy_expected, min_spacing_expected),
-                 DEFAULT_EXPECTED_MEMORY));
-    critical_memory = static_cast<u64>(
-        (std::max)((std::min)(device_local_memory - min_vacancy_critical, min_spacing_critical),
-                 DEFAULT_CRITICAL_MEMORY));
+    memory_budget = runtime.GetDeviceLocalMemory();
 }
 
 template <class P>
 BufferCache<P>::~BufferCache() = default;
 
 template <class P>
-void BufferCache<P>::RunGarbageCollector() {
-    const bool aggressive_gc = total_used_memory >= critical_memory;
-    const u64 ticks_to_destroy = aggressive_gc ? 60 : 120;
-    int num_iterations = aggressive_gc ? 64 : 32;
-    const auto clean_up = [this, &num_iterations](BufferId buffer_id) {
-        if (num_iterations == 0) {
+u64 BufferCache<P>::DeviceUsage(bool force_refresh) {
+    if (!runtime.CanReportAllocationUsage()) {
+        return total_used_memory;
+    }
+    if (force_refresh || usage_refresh_countdown == 0) {
+        cached_device_usage = runtime.GetDeviceAllocationUsage();
+        usage_refresh_countdown = USAGE_REFRESH_INTERVAL;
+    } else {
+        --usage_refresh_countdown;
+    }
+    return cached_device_usage;
+}
+
+template <class P>
+u64 BufferCache<P>::ReclaimMemory(u64 target_bytes, bool allow_download) {
+    if (target_bytes == 0 || in_reclaim) {
+        return 0;
+    }
+    in_reclaim = true;
+    sentenced_buffers.Reclaim(runtime.CompletedSyncPoint());
+    u64 freed = 0;
+    const auto clean_up = [&](BufferId buffer_id) {
+        if (freed >= target_bytes) {
             return true;
         }
-        --num_iterations;
         auto& buffer = slot_buffers[buffer_id];
+        if (!allow_download && IsRegionGpuModified(buffer.CpuAddr(), buffer.SizeBytes())) {
+            return false;
+        }
+        const u64 buffer_bytes = Common::AlignUp(buffer.SizeBytes(), 1024);
         DownloadBufferMemory(buffer);
         DeleteBuffer(buffer_id);
+        freed += buffer_bytes;
         return false;
     };
-    lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, clean_up);
+    const u64 cold_tick =
+        frame_tick > RECLAIM_GUARD_FRAMES ? frame_tick - RECLAIM_GUARD_FRAMES : 0;
+    lru_cache.ForEachItemBelow(cold_tick, clean_up);
+    if (freed < target_bytes) {
+        lru_cache.ForEachItemBelow(frame_tick > 0 ? frame_tick - 1 : 0, clean_up);
+    }
+    sentenced_buffers.Reclaim(runtime.CompletedSyncPoint());
+    in_reclaim = false;
+    usage_refresh_countdown = 0;
+    reclaim_stalled = freed == 0;
+    return freed;
+}
+
+template <class P>
+void BufferCache<P>::EnsureHeadroom(bool allow_download) {
+    if (reclaim_stalled) {
+        return;
+    }
+    const u64 limit = memory_budget > RECLAIM_HEADROOM ? memory_budget - RECLAIM_HEADROOM : 0;
+    const u64 usage = DeviceUsage(false);
+    if (usage <= limit) {
+        return;
+    }
+    const u64 target = (limit / 100) * RECLAIM_TARGET_PERCENT;
+    ReclaimMemory((std::min)(usage - target, total_used_memory), allow_download);
 }
 
 template <class P>
@@ -96,15 +129,11 @@ void BufferCache<P>::TickFrame() {
     const bool skip_preferred = hits * 256 < shots * 251;
     channel_state->uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
 
-    // If we can obtain the memory info, use it instead of the estimate.
-    if (runtime.CanReportMemoryUsage()) {
-        total_used_memory = runtime.GetDeviceMemoryUsage();
-    }
-    if (total_used_memory >= minimum_memory) {
-        RunGarbageCollector();
-    }
+    usage_refresh_countdown = 0;
+    reclaim_stalled = false;
+    EnsureHeadroom(true);
     ++frame_tick;
-    delayed_destruction_ring.Tick();
+    sentenced_buffers.Reclaim(runtime.CompletedSyncPoint());
 
     for (auto& buffer : async_buffers_death_ring) {
         runtime.FreeDeferredStagingBuffer(buffer);
@@ -1576,6 +1605,7 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
 
 template <class P>
 BufferId BufferCache<P>::CreateBuffer(DAddr device_addr, u32 wanted_size) {
+    EnsureHeadroom(false);
     DAddr device_addr_end = Common::AlignUp(device_addr + wanted_size, CACHING_PAGESIZE);
     device_addr = Common::AlignDown(device_addr, CACHING_PAGESIZE);
     wanted_size = static_cast<u32>(device_addr_end - device_addr);
@@ -1613,7 +1643,7 @@ void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
         total_used_memory += Common::AlignUp(size, 1024);
         buffer.setLRUID(lru_cache.Insert(buffer_id, frame_tick));
     } else {
-        total_used_memory -= Common::AlignUp(size, 1024);
+        total_used_memory -= std::min<u64>(total_used_memory, Common::AlignUp(size, 1024));
         lru_cache.Free(buffer.getLRUID());
     }
     const DAddr device_addr_begin = buffer.CpuAddr();
@@ -1872,7 +1902,7 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
 #ifdef YUZU_LEGACY
     if (!do_not_mark || !immediately_free)
 #endif
-        delayed_destruction_ring.Push(std::move(slot_buffers[buffer_id]));
+        sentenced_buffers.Push(std::move(slot_buffers[buffer_id]), runtime.CurrentSyncPoint());
 
     slot_buffers.erase(buffer_id);
 

src/video_core/buffer_cache/buffer_cache_base.h

@@ -9,6 +9,7 @@
 #include <algorithm>
 #include <array>
 #include <bit>
+#include <deque>
 #include <functional>
 #include <memory>
 #include <mutex>
@@ -30,7 +31,7 @@
 #include "common/slot_vector.h"
 #include "video_core/buffer_cache/buffer_base.h"
 #include "video_core/control/channel_state_cache.h"
-#include "video_core/delayed_destruction_ring.h"
+#include "video_core/deferred_destruction_queue.h"
 #include "video_core/dirty_flags.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/kepler_compute.h"
@@ -182,13 +183,15 @@ class BufferCache : public VideoCommon::ChannelSetupCaches<BufferCacheChannelInf
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = P::USE_MEMORY_MAPS_FOR_UPLOADS;
 
 #ifdef YUZU_LEGACY
-    static constexpr s64 TARGET_THRESHOLD = 3_GiB;
+    static constexpr u64 RECLAIM_HEADROOM = 384_MiB;
 #else
-    static constexpr s64 TARGET_THRESHOLD = 4_GiB;
+    static constexpr u64 RECLAIM_HEADROOM = 512_MiB;
 #endif
 
-    static constexpr s64 DEFAULT_EXPECTED_MEMORY = 512_MiB;
-    static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB;
+    static constexpr u64 FALLBACK_MEMORY_BUDGET = 2_GiB;
+    static constexpr u32 USAGE_REFRESH_INTERVAL = 16;
+    static constexpr u64 RECLAIM_GUARD_FRAMES = 8;
+    static constexpr u64 RECLAIM_TARGET_PERCENT = 88;
 
     // Debug Flags.
 
@@ -215,6 +218,8 @@ public:
 
     void TickFrame();
 
+    u64 ReclaimMemory(u64 target_bytes, bool allow_download);
+
     void WriteMemory(DAddr device_addr, u64 size);
 
     void CachedWriteMemory(DAddr device_addr, u64 size);
@@ -358,7 +363,9 @@ private:
                ((device_addr + size) & ~Core::DEVICE_PAGEMASK);
     }
 
-    void RunGarbageCollector();
+    u64 DeviceUsage(bool force_refresh);
+
+    void EnsureHeadroom(bool allow_download);
 
     void BindHostIndexBuffer();
 
@@ -475,12 +482,7 @@ private:
     Tegra::MaxwellDeviceMemoryManager& device_memory;
 
     Common::SlotVector<Buffer> slot_buffers;
-#ifdef YUZU_LEGACY
-    static constexpr size_t TICKS_TO_DESTROY = 6;
-#else
-    static constexpr size_t TICKS_TO_DESTROY = 8;
-#endif
-    DelayedDestructionRing<Buffer, TICKS_TO_DESTROY> delayed_destruction_ring;
+    DeferredDestructionQueue<Buffer> sentenced_buffers;
 
     const Tegra::Engines::Maxwell3D::DrawManager::IndirectParams* current_draw_indirect{};
 
@@ -515,8 +517,11 @@ private:
     Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
     u64 frame_tick = 0;
     u64 total_used_memory = 0;
-    u64 minimum_memory = 0;
-    u64 critical_memory = 0;
+    u64 memory_budget = 0;
+    u64 cached_device_usage = 0;
+    u32 usage_refresh_countdown = 0;
+    bool in_reclaim = false;
+    bool reclaim_stalled = false;
     BufferId inline_buffer_id;
 #ifdef YUZU_LEGACY
     bool immediately_free = false;

src/video_core/deferred_destruction_queue.h

@@ -0,0 +1,56 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <cstddef>
+#include <utility>
+
+#include <boost/container/deque.hpp>
+#include <boost/container/options.hpp>
+
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+template <typename T>
+class DeferredDestructionQueue {
+public:
+    void Push(T&& object, u64 sync_point) {
+        entries.emplace_back(std::move(object), sync_point);
+    }
+
+    void Reclaim(u64 completed_sync_point) {
+        while (!entries.empty() && entries.front().sync_point <= completed_sync_point) {
+            entries.pop_front();
+        }
+    }
+
+    void Clear() {
+        entries.clear();
+    }
+
+    [[nodiscard]] size_t Size() const noexcept {
+        return entries.size();
+    }
+
+    [[nodiscard]] bool Empty() const noexcept {
+        return entries.empty();
+    }
+
+private:
+    struct Entry {
+        Entry(T&& object_, u64 sync_point_) noexcept
+            : object{std::move(object_)}, sync_point{sync_point_} {}
+
+        T object;
+        u64 sync_point;
+    };
+
+    using EntryDequeOptions =
+        boost::container::deque_options<boost::container::block_size<8u>>::type;
+
+    boost::container::deque<Entry, void, EntryDequeOptions> entries;
+};
+
+} // namespace VideoCommon

src/video_core/delayed_destruction_ring.h

@@ -1,34 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
-// SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
-
-#pragma once
-
-#include <array>
-#include <cstddef>
-#include <utility>
-#include <vector>
-
-namespace VideoCommon {
-
-/// Container to push objects to be destroyed a few ticks in the future
-template <typename T, size_t TICKS_TO_DESTROY>
-class DelayedDestructionRing {
-public:
-    void Tick() {
-        index = (index + 1) % TICKS_TO_DESTROY;
-        elements[index].clear();
-    }
-
-    void Push(T&& object) {
-        elements[index].push_back(std::move(object));
-    }
-
-private:
-    size_t index = 0;
-    std::array<std::vector<T>, TICKS_TO_DESTROY> elements;
-};
-
-} // namespace VideoCommon

src/video_core/fence_manager.h

@@ -18,7 +18,7 @@
 #include "common/common_types.h"
 #include "common/settings.h"
 #include "common/thread.h"
-#include "video_core/delayed_destruction_ring.h"
+#include "video_core/deferred_destruction_queue.h"
 #include "video_core/gpu.h"
 #include "video_core/host1x/host1x.h"
 #include "video_core/host1x/syncpoint_manager.h"
@@ -50,7 +50,8 @@ public:
     /// Notify the fence manager about a new frame
     void TickFrame() {
         std::unique_lock lock(ring_guard);
-        delayed_destruction_ring.Tick();
+        ++retire_tick;
+        sentenced_fences.Reclaim(retire_tick > RETIRE_DELAY ? retire_tick - RETIRE_DELAY : 0);
     }
 
     // Unlike other fences, this one doesn't
@@ -186,7 +187,7 @@ private:
             }
             {
                 std::unique_lock lock(ring_guard);
-                delayed_destruction_ring.Push(std::move(current_fence));
+                sentenced_fences.Push(std::move(current_fence), retire_tick);
             }
             fences.pop();
         }
@@ -219,7 +220,7 @@ private:
             }
             {
                 std::unique_lock lock(ring_guard);
-                delayed_destruction_ring.Push(std::move(current_fence));
+                sentenced_fences.Push(std::move(current_fence), retire_tick);
             }
         }
     }
@@ -264,7 +265,9 @@ private:
 
     std::jthread fence_thread;
 
-    DelayedDestructionRing<TFence, 8> delayed_destruction_ring;
+    static constexpr u64 RETIRE_DELAY = 8;
+    u64 retire_tick = 1;
+    DeferredDestructionQueue<TFence> sentenced_fences;
 };
 
 } // namespace VideoCommon

src/video_core/gpu_thread.cpp

@@ -30,6 +30,7 @@ void ThreadManager::StartThread(VideoCore::RendererBase& renderer, Core::Fronten
     thread = std::jthread([&](std::stop_token stop_token) {
         Common::SetCurrentThreadName("GPU");
         Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
+        Common::SetCurrentThreadToPerformanceCores();
         system.RegisterHostThread();
 
         auto current_context = context.Acquire();

src/video_core/host_shaders/CMakeLists.txt

@@ -32,6 +32,7 @@ set(SHADER_FILES
     ${CMAKE_CURRENT_SOURCE_DIR}/convert_msaa_to_non_msaa.frag
     ${CMAKE_CURRENT_SOURCE_DIR}/convert_non_msaa_to_msaa.comp
     ${CMAKE_CURRENT_SOURCE_DIR}/convert_non_msaa_to_msaa.frag
+    ${CMAKE_CURRENT_SOURCE_DIR}/convert_non_msaa_to_msaa_depth.frag
     ${CMAKE_CURRENT_SOURCE_DIR}/convert_s8d24_to_abgr8.frag
     ${CMAKE_CURRENT_SOURCE_DIR}/full_screen_triangle.vert
     ${CMAKE_CURRENT_SOURCE_DIR}/fxaa.frag

src/video_core/host_shaders/astc_decoder.comp

@@ -1384,11 +1384,7 @@ void DecompressBlock(ivec3 coord) {
                 p = Cf / 65535.0f;
             }
 
-#ifdef VULKAN
-            imageStore(dest_image, coord + ivec3(i, j, 0), p.gbar);
-#else
             imageStore(dest_image, coord + ivec3(i, j, 0), clamp(p, 0.0f, 1.0f).gbar);
-#endif
         }
     }
 }

src/video_core/host_shaders/convert_non_msaa_to_msaa_depth.frag

@@ -0,0 +1,19 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#version 450 core
+
+layout(binding = 0) uniform sampler2D img_in;
+
+layout(push_constant) uniform PushConstants {
+    ivec2 dst_offset;
+    ivec2 src_offset;
+    ivec2 scale;
+};
+
+void main() {
+    const ivec2 msaa_coord = ivec2(gl_FragCoord.xy) - dst_offset;
+    const ivec2 sample_offset = ivec2(gl_SampleID % scale.x, gl_SampleID / scale.x);
+    const ivec2 coord = msaa_coord * scale + sample_offset + src_offset;
+    gl_FragDepth = texelFetch(img_in, coord, 0).r;
+}

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -93,7 +93,17 @@ public:
     void PostCopyBarrier();
     void Finish();
 
-    void TickFrame(Common::SlotVector<Buffer>&) noexcept {}
+    void TickFrame(Common::SlotVector<Buffer>&) noexcept {
+        ++sync_point;
+    }
+
+    u64 CurrentSyncPoint() const noexcept {
+        return sync_point;
+    }
+
+    u64 CompletedSyncPoint() const noexcept {
+        return sync_point > SYNC_POINT_DELAY ? sync_point - SYNC_POINT_DELAY : 0;
+    }
 
     void ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value);
 
@@ -128,6 +138,14 @@ public:
 
     u64 GetDeviceMemoryUsage() const;
 
+    u64 GetDeviceAllocationUsage() const {
+        return GetDeviceMemoryUsage();
+    }
+
+    bool CanReportAllocationUsage() const {
+        return device.CanReportMemoryUsage();
+    }
+
     void BindFastUniformBuffer(size_t stage, u32 binding_index, u32 size) {
         const GLuint handle = fast_uniforms[stage][binding_index].handle;
         const GLsizeiptr gl_size = static_cast<GLsizeiptr>(size);
@@ -213,9 +231,13 @@ private:
         GL_FRAGMENT_PROGRAM_PARAMETER_BUFFER_NV,
     };
 
+    static constexpr u64 SYNC_POINT_DELAY = 8;
+
     const Device& device;
     StagingBufferPool& staging_buffer_pool;
 
+    u64 sync_point = 1;
+
     bool has_fast_buffer_sub_data = false;
     bool use_assembly_shaders = false;
     bool has_unified_vertex_buffers = false;

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -87,6 +87,14 @@ public:
 
     u64 GetDeviceMemoryUsage() const;
 
+    u64 GetDeviceAllocationUsage() const {
+        return GetDeviceMemoryUsage();
+    }
+
+    bool CanReportAllocationUsage() const {
+        return device.CanReportMemoryUsage();
+    }
+
     bool CanReportMemoryUsage() const {
         return device.CanReportMemoryUsage();
     }
@@ -139,7 +147,19 @@ public:
 
     bool HasNativeASTC() const noexcept;
 
-    void TickFrame() {}
+    void TickFrame() {
+        ++sync_point;
+    }
+
+    u64 CurrentSyncPoint() const noexcept {
+        return sync_point;
+    }
+
+    u64 CompletedSyncPoint() const noexcept {
+        return sync_point > SYNC_POINT_DELAY ? sync_point - SYNC_POINT_DELAY : 0;
+    }
+
+    void WaitSyncPoint(u64) {}
 
     StateTracker& GetStateTracker() {
         return state_tracker;
@@ -174,6 +194,9 @@ private:
     std::array<OGLFramebuffer, 4> rescale_read_fbos;
     const Settings::ResolutionScalingInfo& resolution;
     u64 device_access_memory;
+
+    static constexpr u64 SYNC_POINT_DELAY = 8;
+    u64 sync_point = 1;
 };
 
 class Image : public VideoCommon::ImageBase {
@@ -370,6 +393,7 @@ struct TextureCacheParams {
     static constexpr bool HAS_EMULATED_COPIES = true;
     static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
     static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = true;
+    static constexpr bool HAS_TIMELINE_SYNC_POINTS = false;
 
     using Runtime = OpenGL::TextureCacheRuntime;
     using Image = OpenGL::Image;

src/video_core/renderer_vulkan/blit_image.cpp

@@ -21,6 +21,7 @@
 #include "video_core/host_shaders/convert_depth_to_float_frag_spv.h"
 #include "video_core/host_shaders/convert_float_to_depth_frag_spv.h"
 #include "video_core/host_shaders/convert_msaa_to_non_msaa_frag_spv.h"
+#include "video_core/host_shaders/convert_non_msaa_to_msaa_depth_frag_spv.h"
 #include "video_core/host_shaders/convert_non_msaa_to_msaa_frag_spv.h"
 #include "video_core/host_shaders/convert_s8d24_to_abgr8_frag_spv.h"
 #include "video_core/host_shaders/full_screen_triangle_vert_spv.h"
@@ -519,7 +520,8 @@ void RecordShaderReadBarrier(Scheduler& scheduler, const ImageView& image_view)
 }
 
 [[nodiscard]] vk::ImageView MakeMSAACopyView(const vk::Device& device, VkImage image,
-                                             VkFormat format, u32 base_level) {
+                                             VkFormat format, u32 base_level,
+                                             VkImageAspectFlags aspect_mask) {
     return device.CreateImageView(VkImageViewCreateInfo{
         .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
         .pNext = nullptr,
@@ -534,7 +536,7 @@ void RecordShaderReadBarrier(Scheduler& scheduler, const ImageView& image_view)
             .a = VK_COMPONENT_SWIZZLE_IDENTITY,
         },
         .subresourceRange{
-            .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            .aspectMask = aspect_mask,
             .baseMipLevel = base_level,
             .levelCount = 1,
             .baseArrayLayer = 0,
@@ -610,6 +612,8 @@ BlitImageHelper::BlitImageHelper(const Device& device_, Scheduler& scheduler_,
       convert_s8d24_to_abgr8_frag(BuildShader(device, CONVERT_S8D24_TO_ABGR8_FRAG_SPV)),
       convert_msaa_to_non_msaa_frag(BuildShader(device, CONVERT_MSAA_TO_NON_MSAA_FRAG_SPV)),
       convert_non_msaa_to_msaa_frag(BuildShader(device, CONVERT_NON_MSAA_TO_MSAA_FRAG_SPV)),
+      convert_non_msaa_to_msaa_depth_frag(
+          BuildShader(device, CONVERT_NON_MSAA_TO_MSAA_DEPTH_FRAG_SPV)),
       linear_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_LINEAR>)),
       nearest_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_NEAREST>)) {}
 
@@ -895,16 +899,34 @@ void BlitImageHelper::CopyMSAA(RenderPassCache& render_pass_cache, VkImage dst_i
     const s32 scale_y = 1 << samples_y;
     const VkSampleCountFlagBits samples =
         msaa_to_non_msaa ? VK_SAMPLE_COUNT_1_BIT : SampleCountFlag(num_samples);
+    const auto dst_surface_type = VideoCore::Surface::GetFormatType(dst_format);
+    const bool is_depth = dst_surface_type == VideoCore::Surface::SurfaceType::Depth ||
+                          dst_surface_type == VideoCore::Surface::SurfaceType::DepthStencil;
+    const bool has_stencil = dst_surface_type == VideoCore::Surface::SurfaceType::DepthStencil;
+    const VkImageAspectFlags view_aspect =
+        is_depth ? VK_IMAGE_ASPECT_DEPTH_BIT : VK_IMAGE_ASPECT_COLOR_BIT;
+    VkImageAspectFlags barrier_aspect = VK_IMAGE_ASPECT_COLOR_BIT;
+    if (is_depth) {
+        barrier_aspect = VK_IMAGE_ASPECT_DEPTH_BIT;
+        if (has_stencil) {
+            barrier_aspect |= VK_IMAGE_ASPECT_STENCIL_BIT;
+        }
+    }
     RenderPassKey renderpass_key{};
     renderpass_key.color_formats.fill(VideoCore::Surface::PixelFormat::Invalid);
-    renderpass_key.color_formats[0] = dst_format;
-    renderpass_key.depth_format = VideoCore::Surface::PixelFormat::Invalid;
+    if (is_depth) {
+        renderpass_key.depth_format = dst_format;
+    } else {
+        renderpass_key.color_formats[0] = dst_format;
+        renderpass_key.depth_format = VideoCore::Surface::PixelFormat::Invalid;
+    }
     renderpass_key.samples = samples;
     const VkRenderPass renderpass = render_pass_cache.Get(renderpass_key);
     const MSAACopyPipelineKey key{
         .renderpass = renderpass,
         .samples = samples,
         .msaa_to_non_msaa = msaa_to_non_msaa,
+        .is_depth = is_depth,
     };
     const VkPipeline pipeline = FindOrEmplaceMSAACopyPipeline(key);
     const VkPipelineLayout layout = *msaa_copy_pipeline_layout;
@@ -920,10 +942,10 @@ void BlitImageHelper::CopyMSAA(RenderPassCache& render_pass_cache, VkImage dst_i
         ASSERT(copy.dst_subresource.num_layers == 1);
         vk::ImageView src_view =
             MakeMSAACopyView(device.GetLogical(), src_image, src_vk_format,
-                             static_cast<u32>(copy.src_subresource.base_level));
+                             static_cast<u32>(copy.src_subresource.base_level), view_aspect);
         vk::ImageView dst_view =
             MakeMSAACopyView(device.GetLogical(), dst_image, dst_vk_format,
-                             static_cast<u32>(copy.dst_subresource.base_level));
+                             static_cast<u32>(copy.dst_subresource.base_level), view_aspect);
         const VkOffset2D dst_offset{copy.dst_offset.x, copy.dst_offset.y};
         const VkExtent2D dst_extent{copy.extent.width, copy.extent.height};
         const VkRect2D render_area{
@@ -949,50 +971,64 @@ void BlitImageHelper::CopyMSAA(RenderPassCache& render_pass_cache, VkImage dst_i
         scheduler.RequestOutsideRenderPassOperationContext();
         scheduler.Record([this, pipeline, layout, sampler, renderpass,
                           framebuffer_handle = *framebuffer, src_view_handle = *src_view,
-                          src = src_image, dst = dst_image, render_area,
+                          src = src_image, dst = dst_image, render_area, is_depth, barrier_aspect,
                           push_constants](vk::CommandBuffer cmdbuf) {
-            constexpr VkImageSubresourceRange color_range{
-                .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            const VkImageSubresourceRange src_range{
+                .aspectMask = barrier_aspect,
                 .baseMipLevel = 0,
                 .levelCount = VK_REMAINING_MIP_LEVELS,
                 .baseArrayLayer = 0,
                 .layerCount = VK_REMAINING_ARRAY_LAYERS,
             };
+            const VkImageSubresourceRange dst_range = src_range;
+            const VkAccessFlags attachment_read =
+                is_depth ? VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT
+                         : VK_ACCESS_COLOR_ATTACHMENT_READ_BIT;
+            const VkAccessFlags attachment_write =
+                is_depth ? VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
+                         : VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+            const VkPipelineStageFlags depth_stage =
+                VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
+                VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
+            const VkPipelineStageFlags attachment_stage =
+                is_depth ? depth_stage
+                         : static_cast<VkPipelineStageFlags>(
+                               VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT);
             const std::array pre_barriers{
                 VkImageMemoryBarrier{
                     .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
                     .pNext = nullptr,
-                    .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-                                     VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
+                    .srcAccessMask = attachment_write | VK_ACCESS_SHADER_WRITE_BIT |
+                                     VK_ACCESS_TRANSFER_WRITE_BIT,
                     .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
                     .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
                     .newLayout = VK_IMAGE_LAYOUT_GENERAL,
                     .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                     .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                     .image = src,
-                    .subresourceRange = color_range,
+                    .subresourceRange = src_range,
                 },
                 VkImageMemoryBarrier{
                     .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
                     .pNext = nullptr,
-                    .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-                                     VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
-                    .dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
-                                     VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+                    .srcAccessMask = attachment_write | VK_ACCESS_SHADER_WRITE_BIT |
+                                     VK_ACCESS_TRANSFER_WRITE_BIT,
+                    .dstAccessMask = attachment_read | attachment_write,
                     .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
                     .newLayout = VK_IMAGE_LAYOUT_GENERAL,
                     .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                     .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                     .image = dst,
-                    .subresourceRange = color_range,
+                    .subresourceRange = dst_range,
                 },
             };
             cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
+                                       VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
+                                       VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
                                        VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
                                        VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
                                        VK_PIPELINE_STAGE_TRANSFER_BIT,
-                                   VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-                                       VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+                                   VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | attachment_stage,
                                    0, nullptr, nullptr, pre_barriers);
             const VkRenderPassBeginInfo renderpass_bi{
                 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
@@ -1025,16 +1061,16 @@ void BlitImageHelper::CopyMSAA(RenderPassCache& render_pass_cache, VkImage dst_i
             const VkImageMemoryBarrier post_barrier{
                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
                 .pNext = nullptr,
-                .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+                .srcAccessMask = attachment_write,
                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT,
                 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
                 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                 .image = dst,
-                .subresourceRange = color_range,
+                .subresourceRange = dst_range,
             };
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+            cmdbuf.PipelineBarrier(attachment_stage,
                                    VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
                                        VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
                                        VK_PIPELINE_STAGE_TRANSFER_BIT,
@@ -1423,9 +1459,36 @@ VkPipeline BlitImageHelper::FindOrEmplaceMSAACopyPipeline(const MSAACopyPipeline
         return *msaa_copy_pipelines[std::distance(msaa_copy_keys.begin(), it)];
     }
     msaa_copy_keys.push_back(key);
-    const std::array stages = MakeStages(*clear_color_vert, key.msaa_to_non_msaa
-                                                                ? *convert_msaa_to_non_msaa_frag
-                                                                : *convert_non_msaa_to_msaa_frag);
+    const VkShaderModule frag_module =
+        key.msaa_to_non_msaa
+            ? *convert_msaa_to_non_msaa_frag
+            : (key.is_depth ? *convert_non_msaa_to_msaa_depth_frag
+                            : *convert_non_msaa_to_msaa_frag);
+    const std::array stages = MakeStages(*clear_color_vert, frag_module);
+    const VkPipelineDepthStencilStateCreateInfo depth_stencil_ci{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .depthTestEnable = VK_TRUE,
+        .depthWriteEnable = VK_TRUE,
+        .depthCompareOp = VK_COMPARE_OP_ALWAYS,
+        .depthBoundsTestEnable = VK_FALSE,
+        .stencilTestEnable = VK_FALSE,
+        .front = {},
+        .back = {},
+        .minDepthBounds = 0.0f,
+        .maxDepthBounds = 0.0f,
+    };
+    static constexpr VkPipelineColorBlendStateCreateInfo no_color_blend_ci{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .logicOpEnable = VK_FALSE,
+        .logicOp = VK_LOGIC_OP_CLEAR,
+        .attachmentCount = 0,
+        .pAttachments = nullptr,
+        .blendConstants = {0.0f, 0.0f, 0.0f, 0.0f},
+    };
     const VkPipelineMultisampleStateCreateInfo multisample_ci{
         .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
         .pNext = nullptr,
@@ -1450,8 +1513,9 @@ VkPipeline BlitImageHelper::FindOrEmplaceMSAACopyPipeline(const MSAACopyPipeline
         .pViewportState = &PIPELINE_VIEWPORT_STATE_CREATE_INFO,
         .pRasterizationState = &PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
         .pMultisampleState = &multisample_ci,
-        .pDepthStencilState = nullptr,
-        .pColorBlendState = &PIPELINE_COLOR_BLEND_STATE_GENERIC_CREATE_INFO,
+        .pDepthStencilState = key.is_depth ? &depth_stencil_ci : nullptr,
+        .pColorBlendState = key.is_depth ? &no_color_blend_ci
+                                         : &PIPELINE_COLOR_BLEND_STATE_GENERIC_CREATE_INFO,
         .pDynamicState = &PIPELINE_DYNAMIC_STATE_CREATE_INFO,
         .layout = *msaa_copy_pipeline_layout,
         .renderPass = key.renderpass,

src/video_core/renderer_vulkan/blit_image.h

@@ -51,6 +51,7 @@ struct MSAACopyPipelineKey {
     VkRenderPass renderpass;
     VkSampleCountFlagBits samples;
     bool msaa_to_non_msaa;
+    bool is_depth;
 };
 
 struct BlitMSAAPipelineKey {
@@ -180,6 +181,7 @@ private:
     vk::ShaderModule convert_s8d24_to_abgr8_frag;
     vk::ShaderModule convert_msaa_to_non_msaa_frag;
     vk::ShaderModule convert_non_msaa_to_msaa_frag;
+    vk::ShaderModule convert_non_msaa_to_msaa_depth_frag;
     vk::Sampler linear_sampler;
     vk::Sampler nearest_sampler;
 

src/video_core/renderer_vulkan/fixed_pipeline_state.cpp

@@ -164,7 +164,9 @@ void FixedPipelineState::Refresh(Tegra::Engines::Maxwell3D& maxwell3d, DynamicFe
     }
 
     provoking_vertex_last.Assign(use_last_provoking_vertex ? 1 : 0);
-    conservative_raster_enable.Assign(regs.conservative_raster_enable != 0 ? 1 : 0);
+    if (!features.has_dynamic_state3_conservative_raster_mode) {
+        conservative_raster_enable.Assign(regs.conservative_raster_enable != 0 ? 1 : 0);
+    }
     smooth_lines.Assign(regs.line_anti_alias_enable != 0 ? 1 : 0);
     alpha_to_coverage_enabled.Assign(regs.anti_alias_alpha_control.alpha_to_coverage != 0 ? 1 : 0);
     alpha_to_one_enabled.Assign(regs.anti_alias_alpha_control.alpha_to_one != 0 ? 1 : 0);
@@ -360,18 +362,35 @@ void FixedPipelineState::DynamicState::Refresh2(const Maxwell& regs,
     depth_bias_enable.Assign(enabled_lut[POLYGON_OFFSET_ENABLE_LUT[topology_index]] != 0 ? 1 : 0);
 }
 
+bool IsDepthClipEnabled(const Maxwell& regs) {
+    const auto clip = regs.viewport_clip_control.geometry_clip.Value();
+    return clip == Maxwell::ViewportClipControl::GeometryClip::Passthrough ||
+           clip == Maxwell::ViewportClipControl::GeometryClip::FrustumXYZ ||
+           clip == Maxwell::ViewportClipControl::GeometryClip::FrustumZ;
+}
+
+bool IsDepthClampEnabled(const Maxwell& regs, bool has_depth_clip_enable) {
+    if (!IsDepthClipEnabled(regs)) {
+        return true;
+    }
+    if (!has_depth_clip_enable) {
+        return false;
+    }
+    return regs.viewport_clip_control.pixel_min_z.Value() != 0 ||
+           regs.viewport_clip_control.pixel_max_z.Value() != 0;
+}
+
 void FixedPipelineState::DynamicState::Refresh3(const Maxwell& regs,
                                                 const DynamicFeatures& features) {
     if (!features.has_dynamic_state3_logic_op_enable) {
         logic_op_enable.Assign(regs.logic_op.enable != 0 ? 1 : 0);
     }
+    if (features.has_depth_clip_enable) {
+        depth_clip_disabled.Assign(IsDepthClipEnabled(regs) ? 0 : 1);
+    }
     if (!features.has_dynamic_state3_depth_clamp_enable) {
-        depth_clamp_disabled.Assign(regs.viewport_clip_control.geometry_clip ==
-                                        Maxwell::ViewportClipControl::GeometryClip::Passthrough ||
-                                    regs.viewport_clip_control.geometry_clip ==
-                                        Maxwell::ViewportClipControl::GeometryClip::FrustumXYZ ||
-                                    regs.viewport_clip_control.geometry_clip ==
-                                        Maxwell::ViewportClipControl::GeometryClip::FrustumZ);
+        depth_clamp_disabled.Assign(
+            IsDepthClampEnabled(regs, features.has_depth_clip_enable) ? 0 : 1);
     }
     if (!features.has_dynamic_state3_line_stipple_enable) {
         line_stipple_enable.Assign(regs.line_stipple_enable);

src/video_core/renderer_vulkan/fixed_pipeline_state.h

@@ -30,6 +30,8 @@ struct DynamicFeatures {
     bool has_extended_dynamic_state_3_blend;
     bool has_extended_dynamic_state_3_enables;
     bool has_dynamic_state3_depth_clamp_enable;
+    bool has_dynamic_state3_conservative_raster_mode;
+    bool has_depth_clip_enable;
     bool has_dynamic_state3_logic_op_enable;
     bool has_dynamic_state3_line_stipple_enable;
     bool has_dynamic_vertex_input;
@@ -165,6 +167,7 @@ struct FixedPipelineState {
             BitField<10, 1, u32> logic_op_enable;
             BitField<11, 1, u32> depth_clamp_disabled;
             BitField<12, 1, u32> line_stipple_enable;
+            BitField<13, 1, u32> depth_clip_disabled;
         };
         union {
             u32 raw2;
@@ -298,6 +301,9 @@ static_assert(std::has_unique_object_representations_v<FixedPipelineState>);
 static_assert(std::is_trivially_copyable_v<FixedPipelineState>);
 static_assert(std::is_trivially_constructible_v<FixedPipelineState>);
 
+bool IsDepthClipEnabled(const Maxwell& regs);
+bool IsDepthClampEnabled(const Maxwell& regs, bool has_depth_clip_enable);
+
 } // namespace Vulkan
 
 namespace std {

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -246,7 +246,6 @@ protected:
     StagingBufferPool& staging_pool;
 
     vk::Buffer buffer{};
-    MemoryCommit memory_commit{};
     VkIndexType index_type{};
     u32 num_indices = 0;
 };
@@ -376,6 +375,10 @@ u64 BufferCacheRuntime::GetDeviceMemoryUsage() const {
     return device.GetDeviceMemoryUsage();
 }
 
+u64 BufferCacheRuntime::GetDeviceAllocationUsage() const {
+    return device.GetMemoryBudgetInfo().allocation_bytes;
+}
+
 bool BufferCacheRuntime::CanReportMemoryUsage() const {
     return device.CanReportMemoryUsage();
 }
@@ -404,6 +407,16 @@ u64 BufferCacheRuntime::KnownGpuTick() {
     return scheduler.GetMasterSemaphore().KnownGpuTick();
 }
 
+u64 BufferCacheRuntime::CurrentSyncPoint() const noexcept {
+    return scheduler.GetMasterSemaphore().CurrentTick();
+}
+
+u64 BufferCacheRuntime::CompletedSyncPoint() const {
+    auto& master_semaphore = scheduler.GetMasterSemaphore();
+    master_semaphore.Refresh();
+    return master_semaphore.KnownGpuTick();
+}
+
 void BufferCacheRuntime::Wait(u64 buffer_tick) {
     scheduler.Wait(buffer_tick);
 }
@@ -641,6 +654,7 @@ void BufferCacheRuntime::BindTransformFeedbackBuffer(u32 index, VkBuffer buffer,
         offset = 0;
         size = 0;
     }
+    scheduler.MarkTransformFeedbackUsed();
     scheduler.Record([index, buffer, offset, size](vk::CommandBuffer cmdbuf) {
         const VkDeviceSize vk_offset = offset;
         const VkDeviceSize vk_size = size;
@@ -653,19 +667,26 @@ void BufferCacheRuntime::BindTransformFeedbackBuffers(VideoCommon::HostBindings<
         // Already logged in the rasterizer
         return;
     }
-    boost::container::static_vector<VkBuffer, VideoCommon::NUM_VERTEX_BUFFERS> buffer_handles(bindings.buffers.size());
-    for (u32 i = 0; i < bindings.buffers.size(); ++i) {
+    const u32 count = std::min<u32>(static_cast<u32>(bindings.buffers.size()),
+                                    VideoCommon::NUM_TRANSFORM_FEEDBACK_BUFFERS);
+    std::array<VkBuffer, VideoCommon::NUM_TRANSFORM_FEEDBACK_BUFFERS> handles{};
+    std::array<VkDeviceSize, VideoCommon::NUM_TRANSFORM_FEEDBACK_BUFFERS> offsets{};
+    std::array<VkDeviceSize, VideoCommon::NUM_TRANSFORM_FEEDBACK_BUFFERS> sizes{};
+    for (u32 i = 0; i < count; ++i) {
         auto handle = bindings.buffers[i]->Handle();
         if (handle == VK_NULL_HANDLE) {
             ReserveNullBuffer();
             handle = *null_buffer;
-            bindings.offsets[i] = 0;
-            bindings.sizes[i] = 0;
+        } else {
+            offsets[i] = bindings.offsets[i];
+            sizes[i] = bindings.sizes[i];
         }
-        buffer_handles[i] = handle;
+        handles[i] = handle;
     }
-    scheduler.Record([bindings_ = std::move(bindings), buffer_handles_ = std::move(buffer_handles)](vk::CommandBuffer cmdbuf) {
-        cmdbuf.BindTransformFeedbackBuffersEXT(0, u32(buffer_handles_.size()), buffer_handles_.data(), bindings_.offsets.data(), bindings_.sizes.data());
+    scheduler.MarkTransformFeedbackUsed();
+    scheduler.Record([count, handles, offsets, sizes](vk::CommandBuffer cmdbuf) {
+        cmdbuf.BindTransformFeedbackBuffersEXT(0, count, handles.data(), offsets.data(),
+                                               sizes.data());
     });
 }
 

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -100,10 +100,20 @@ public:
 
     void Finish();
 
+    u64 CurrentSyncPoint() const noexcept;
+
+    u64 CompletedSyncPoint() const;
+
     u64 GetDeviceLocalMemory() const;
 
     u64 GetDeviceMemoryUsage() const;
 
+    u64 GetDeviceAllocationUsage() const;
+
+    bool CanReportAllocationUsage() const noexcept {
+        return true;
+    }
+
     bool CanReportMemoryUsage() const;
 
     u32 GetUniformBufferAlignment() const;

src/video_core/renderer_vulkan/vk_fence_manager.cpp

@@ -37,7 +37,7 @@ void InnerFence::Wait() {
     if (is_stubbed) {
         return;
     }
-    scheduler.Wait(wait_tick);
+    scheduler.WaitSubmitted(wait_tick);
 }
 
 FenceManager::FenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -757,16 +757,13 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
         .lineWidth = 1.0f,
         // TODO(alekpop): Transfer from regs
     };
-    const bool smooth_lines_supported =
-        device.IsExtLineRasterizationSupported() && device.SupportsSmoothLines();
-    const bool stippled_lines_supported =
-        device.IsExtLineRasterizationSupported() && device.SupportsStippledRectangularLines();
+    const VkLineRasterizationModeEXT line_raster_mode =
+        device.GetLineRasterizationMode(key.state.smooth_lines != 0);
+    const bool stippled_lines_supported = device.SupportsStippleForMode(line_raster_mode);
     VkPipelineRasterizationLineStateCreateInfoEXT line_state{
         .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT,
         .pNext = nullptr,
-        .lineRasterizationMode = key.state.smooth_lines != 0 && smooth_lines_supported
-                                     ? VK_LINE_RASTERIZATION_MODE_RECTANGULAR_SMOOTH_EXT
-                                     : VK_LINE_RASTERIZATION_MODE_RECTANGULAR_EXT,
+        .lineRasterizationMode = line_raster_mode,
         .stippledLineEnable =
             (dynamic.line_stipple_enable && stippled_lines_supported) ? VK_TRUE : VK_FALSE,
         .lineStippleFactor = key.state.line_stipple_factor,
@@ -805,6 +802,16 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
     if (device.IsExtProvokingVertexSupported()) {
         provoking_vertex.pNext = std::exchange(rasterization_ci.pNext, &provoking_vertex);
     }
+    VkPipelineRasterizationDepthClipStateCreateInfoEXT depth_clip_state{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT,
+        .pNext = nullptr,
+        .flags = 0,
+        .depthClipEnable = static_cast<VkBool32>(dynamic.depth_clip_disabled == 0 ? VK_TRUE
+                                                                                  : VK_FALSE),
+    };
+    if (device.IsExtDepthClipEnableSupported()) {
+        depth_clip_state.pNext = std::exchange(rasterization_ci.pNext, &depth_clip_state);
+    }
 
     const bool supports_alpha_output = fragment_has_color0_output;
     const bool alpha_to_one_supported = device.SupportsAlphaToOne();

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -305,7 +305,7 @@ size_t GetTotalPipelineWorkers() {
         std::max<size_t>(static_cast<size_t>(std::thread::hardware_concurrency()), 2ULL) - 1ULL;
 #ifdef __ANDROID__
     const int configured = AndroidSettings::values.pipeline_worker_count.GetValue();
-    const int clamped = std::clamp(configured, 4, 8);
+    const int clamped = std::clamp(configured, 2, 8);
     const size_t desired = static_cast<size_t>(clamped);
     if (desired == 0) {
         return 1ULL;
@@ -349,8 +349,9 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
       use_asynchronous_shaders{Settings::values.use_asynchronous_shaders.GetValue()},
       use_vulkan_pipeline_cache{Settings::values.use_vulkan_driver_pipeline_cache.GetValue()},
       workers(device.HasBrokenParallelShaderCompiling() ? 1ULL : GetTotalPipelineWorkers(),
-              "VkPipelineBuilder"),
-      serialization_thread(1, "VkPipelineSerialization") {
+              "VkPipelineBuilder", {}, Common::ThreadPlacement::Background),
+      serialization_thread(1, "VkPipelineSerialization", {},
+                           Common::ThreadPlacement::Background) {
     const auto& float_control{device.FloatControlProperties()};
     const VkDriverId driver_id{device.GetDriverID()};
     const VkShaderStageFlags subgroup_stages{device.GetSubgroupSupportedStages()};
@@ -514,6 +515,11 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
     dynamic_features.has_dynamic_state3_depth_clamp_enable =
         dynamic_features.has_extended_dynamic_state_3_enables &&
         device.SupportsDynamicState3DepthClampEnable();
+    dynamic_features.has_dynamic_state3_conservative_raster_mode =
+        dynamic_features.has_extended_dynamic_state_3_enables &&
+        device.SupportsDynamicState3ConservativeRasterizationMode();
+    dynamic_features.has_depth_clip_enable = 
+        device.IsExtDepthClipEnableSupported();
     dynamic_features.has_dynamic_state3_logic_op_enable =
         dynamic_features.has_extended_dynamic_state_3_enables &&
         device.SupportsDynamicState3LogicOpEnable();
@@ -526,7 +532,8 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         device.IsExtVertexInputDynamicStateSupported() &&
         Settings::values.vertex_input_dynamic_state.GetValue();
 
-    dynamic_features.has_provoking_vertex = device.IsExtProvokingVertexSupported();
+    dynamic_features.has_provoking_vertex =
+        device.IsExtProvokingVertexSupported();
     dynamic_features.has_provoking_vertex_first_mode =
         device.SupportsProvokingVertexFirstMode();
     dynamic_features.has_provoking_vertex_last_mode =

src/video_core/renderer_vulkan/vk_present_manager.cpp

@@ -296,9 +296,6 @@ void PresentManager::RecreateSwapchain(Frame* frame) {
 }
 
 void PresentManager::SetImageCount() {
-    // We cannot have more than 7 images in flight at any given time.
-    // FRAMES_IN_FLIGHT is 8, and the cache TICKS_TO_DESTROY is 8.
-    // Mali drivers will give us 6.
     image_count = std::min<size_t>(swapchain.GetImageCount(), 7);
 }
 

src/video_core/renderer_vulkan/vk_query_cache.cpp

@@ -919,7 +919,7 @@ private:
             return;
         }
         has_flushed_end_pending = true;
-        // Refresh buffers state before beginning transform feedback so counters are up-to-date
+        scheduler.MarkTransformFeedbackUsed();
         UpdateBuffers();
         if (!has_started || buffers_count == 0) {
             // No counter buffers available: begin without counters

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -6,6 +6,7 @@
 
 #include <algorithm>
 #include <array>
+#include <limits>
 #include <memory>
 #include <mutex>
 
@@ -203,7 +204,8 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
     : gpu{gpu_}, device_memory{device_memory_}, device{device_},
       memory_allocator{memory_allocator_}, state_tracker{state_tracker_}, scheduler{scheduler_},
       staging_pool(device, memory_allocator, scheduler), descriptor_pool(device, scheduler),
-      guest_descriptor_queue(device), compute_pass_descriptor_queue(device),
+      guest_descriptor_queue(device, UpdateDescriptorQueue::GUEST_FRAME_PAYLOAD_SIZE),
+      compute_pass_descriptor_queue(device, UpdateDescriptorQueue::COMPUTE_FRAME_PAYLOAD_SIZE),
       blit_image(device, scheduler, state_tracker, descriptor_pool), render_pass_cache(device),
       texture_cache_runtime{
           device,     scheduler,         memory_allocator, staging_pool,
@@ -221,9 +223,23 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
     scheduler.SetQueryCache(query_cache);
+    memory_allocator.SetReclaimCallback([this](u64 bytes) -> u64 {
+        u64 freed = staging_pool.ReclaimMemory(bytes);
+        if (freed < bytes) {
+            freed += texture_cache.ReclaimMemory(bytes - freed, false);
+        }
+        if (freed < bytes) {
+            freed += buffer_cache.ReclaimMemory(bytes - freed, false);
+        }
+        auto& master_semaphore = scheduler.GetMasterSemaphore();
+        master_semaphore.Refresh();
+        vk::TickDeletionQueue(master_semaphore.KnownGpuTick());
+        return freed;
+    });
 }
 
 RasterizerVulkan::~RasterizerVulkan() {
+    memory_allocator.SetReclaimCallback(nullptr);
     scheduler.WaitWorker();
     scheduler.Finish();
 }
@@ -880,6 +896,9 @@ void RasterizerVulkan::FlushCommands() {
 
 void RasterizerVulkan::TickFrame() {
     draw_counter = 0;
+    auto& master_semaphore = scheduler.GetMasterSemaphore();
+    master_semaphore.Refresh();
+    vk::TickDeletionQueue(master_semaphore.KnownGpuTick());
     guest_descriptor_queue.TickFrame();
     compute_pass_descriptor_queue.TickFrame();
     fence_manager.TickFrame();
@@ -1451,7 +1470,10 @@ void RasterizerVulkan::UpdateLineWidth(Tegra::Engines::Maxwell3D::Regs& regs) {
     }
     const float width =
         regs.line_anti_alias_enable ? regs.line_width_smooth : regs.line_width_aliased;
-    scheduler.Record([width](vk::CommandBuffer cmdbuf) { cmdbuf.SetLineWidth(width); });
+    const float clamped_width = device.ClampLineWidth(width);
+    scheduler.Record([clamped_width](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetLineWidth(clamped_width);
+    });
 }
 
 void RasterizerVulkan::UpdateCullMode(Tegra::Engines::Maxwell3D::Regs& regs) {
@@ -1548,7 +1570,10 @@ void RasterizerVulkan::UpdateLineStippleEnable(Tegra::Engines::Maxwell3D::Regs&
         return;
     }
 
-    scheduler.Record([enable = regs.line_stipple_enable](vk::CommandBuffer cmdbuf) {
+    const VkLineRasterizationModeEXT mode =
+        device.GetLineRasterizationMode(regs.line_anti_alias_enable != 0);
+    const bool enable = regs.line_stipple_enable != 0 && device.SupportsStippleForMode(mode);
+    scheduler.Record([enable](vk::CommandBuffer cmdbuf) {
         cmdbuf.SetLineStippleEnableEXT(enable);
     });
 }
@@ -1562,28 +1587,24 @@ void RasterizerVulkan::UpdateLineRasterizationMode(Tegra::Engines::Maxwell3D::Re
     }
 
     if (!device.SupportsDynamicState3LineRasterizationMode()) {
-        static std::once_flag warn_missing_rect;
-        std::call_once(warn_missing_rect, [] {
+        static std::once_flag warn_missing_dynamic_state;
+        std::call_once(warn_missing_dynamic_state, [] {
             LOG_WARNING(Render_Vulkan,
-                        "Driver lacks rectangular line rasterization support; skipping dynamic "
-                        "line state updates");
+                        "Driver lacks dynamic line rasterization mode; the pipeline static value "
+                        "is used instead");
         });
         return;
     }
 
     const bool wants_smooth = regs.line_anti_alias_enable != 0;
-    VkLineRasterizationModeEXT mode = VK_LINE_RASTERIZATION_MODE_RECTANGULAR_EXT;
-    if (wants_smooth) {
-        if (device.SupportsSmoothLines()) {
-            mode = VK_LINE_RASTERIZATION_MODE_RECTANGULAR_SMOOTH_EXT;
-        } else {
-            static std::once_flag warn_missing_smooth;
-            std::call_once(warn_missing_smooth, [] {
-                LOG_WARNING(Render_Vulkan,
-                            "Line anti-aliasing requested but smoothLines feature unavailable; "
-                            "using rectangular rasterization");
-            });
-        }
+    const VkLineRasterizationModeEXT mode = device.GetLineRasterizationMode(wants_smooth);
+    if (wants_smooth && mode != VK_LINE_RASTERIZATION_MODE_RECTANGULAR_SMOOTH_EXT) {
+        static std::once_flag warn_missing_smooth;
+        std::call_once(warn_missing_smooth, [] {
+            LOG_WARNING(Render_Vulkan,
+                        "Line anti-aliasing requested but smoothLines feature unavailable; "
+                        "falling back to the closest supported mode");
+        });
     }
     scheduler.Record([mode](vk::CommandBuffer cmdbuf) {
         cmdbuf.SetLineRasterizationModeEXT(mode);
@@ -1643,12 +1664,7 @@ void RasterizerVulkan::UpdateDepthClampEnable(Tegra::Engines::Maxwell3D::Regs& r
     if (!device.SupportsDynamicState3DepthClampEnable()) {
         return;
     }
-    bool is_enabled = !(regs.viewport_clip_control.geometry_clip ==
-                            Maxwell::ViewportClipControl::GeometryClip::Passthrough ||
-                        regs.viewport_clip_control.geometry_clip ==
-                            Maxwell::ViewportClipControl::GeometryClip::FrustumXYZ ||
-                        regs.viewport_clip_control.geometry_clip ==
-                            Maxwell::ViewportClipControl::GeometryClip::FrustumZ);
+    const bool is_enabled = IsDepthClampEnabled(regs, device.IsExtDepthClipEnableSupported());
     scheduler.Record(
         [is_enabled](vk::CommandBuffer cmdbuf) { cmdbuf.SetDepthClampEnableEXT(is_enabled); });
 }

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -179,7 +179,6 @@ private:
     void UpdateRasterizerDiscardEnable(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateConservativeRasterizationMode(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateLineStippleEnable(Tegra::Engines::Maxwell3D::Regs& regs);
-    void UpdateLineStipple(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateLineRasterizationMode(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateDepthBiasEnable(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateLogicOpEnable(Tegra::Engines::Maxwell3D::Regs& regs);

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -47,6 +47,7 @@ Scheduler::Scheduler(const Device& device_, StateTracker& state_tracker_)
       master_semaphore{std::make_unique<MasterSemaphore>(device)},
       command_pool{std::make_unique<CommandPool>(*master_semaphore, device)} {
 
+    vk::SetDeletionTimeline(master_semaphore->CurrentTick());
     AcquireNewChunk();
     AllocateWorkerCommandBuffer();
     worker_thread = std::jthread([this](std::stop_token token) { WorkerThread(token); });
@@ -322,6 +323,7 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
     InvalidateState();
 
     const u64 signal_value = master_semaphore->NextTick();
+    vk::SetDeletionTimeline(master_semaphore->CurrentTick());
     RecordWithUploadBuffer([signal_semaphore, wait_semaphore, signal_value,
                             this](vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) {
         static constexpr VkMemoryBarrier WRITE_BARRIER{
@@ -398,7 +400,7 @@ void Scheduler::EndRenderPass()
         Record([num_images = num_renderpass_images,
                        images = renderpass_images,
                        ranges = renderpass_image_ranges,
-                       has_transform_feedback = device.IsExtTransformFeedbackSupported()](
+                       has_transform_feedback = state.uses_transform_feedback](
                           vk::CommandBuffer cmdbuf) {
             std::array<VkImageMemoryBarrier, 9> barriers;
             for (size_t i = 0; i < num_images; ++i) {
@@ -453,6 +455,7 @@ void Scheduler::EndRenderPass()
         });
 
         state.renderpass = VkRenderPass{};
+        state.uses_transform_feedback = false;
         num_renderpass_images = 0;
     }
 

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -74,6 +74,11 @@ public:
         return state.renderpass != VK_NULL_HANDLE;
     }
 
+    /// Flags that transform feedback writes have been recorded since the last render pass end.
+    void MarkTransformFeedbackUsed() noexcept {
+        state.uses_transform_feedback = true;
+    }
+
     /// Update the pipeline to the current execution context.
     bool UpdateGraphicsPipeline(GraphicsPipeline* pipeline);
 
@@ -131,33 +136,16 @@ public:
             }
             master_semaphore->Wait(tick);
         }
-        if (Settings::values.use_speed_limit.GetValue() && target_fps > 0.0) {
-            auto now = std::chrono::steady_clock::now();
-            if (last_target_fps != target_fps) {
-                frame_interval = std::chrono::duration_cast<std::chrono::steady_clock::duration>(std::chrono::duration<double>(1.0 / target_fps));
-                max_frame_count = static_cast<int>(0.1 * target_fps);
-                last_target_fps = target_fps;
-                frame_counter = 0;
-                start_time = now;
-            }
-            frame_counter++;
-            auto target_time = start_time + frame_interval * frame_counter;
-            if (target_time >= now) {
-                auto sleep_time = target_time - now;
-                if (sleep_time > std::chrono::milliseconds(15)) {
-                    std::this_thread::sleep_for(sleep_time - std::chrono::milliseconds(1));
-                }
-                while (std::chrono::steady_clock::now() < target_time) {
-                    std::this_thread::yield();
-                }
-            } else if (frame_counter > max_frame_count) {
-                frame_counter = 0;
-                start_time = now;
-            }
-        }
+        ApplyFramePacing(target_fps);
+    }
+
+    void WaitSubmitted(u64 tick, double target_fps = 0.0) {
+        if (tick > 0 && tick < master_semaphore->CurrentTick()) {
+            master_semaphore->Wait(tick);
+        }
+        ApplyFramePacing(target_fps);
     }
 
-    /// Returns the master timeline semaphore.
     [[nodiscard]] MasterSemaphore& GetMasterSemaphore() const noexcept {
         return *master_semaphore;
     }
@@ -165,6 +153,35 @@ public:
     std::mutex submit_mutex;
 
 private:
+    void ApplyFramePacing(double target_fps) {
+        if (!Settings::values.use_speed_limit.GetValue() || target_fps <= 0.0) {
+            return;
+        }
+        auto now = std::chrono::steady_clock::now();
+        if (last_target_fps != target_fps) {
+            frame_interval = std::chrono::duration_cast<std::chrono::steady_clock::duration>(
+                std::chrono::duration<double>(1.0 / target_fps));
+            max_frame_count = static_cast<int>(0.1 * target_fps);
+            last_target_fps = target_fps;
+            frame_counter = 0;
+            start_time = now;
+        }
+        frame_counter++;
+        auto target_time = start_time + frame_interval * frame_counter;
+        if (target_time >= now) {
+            auto sleep_time = target_time - now;
+            if (sleep_time > std::chrono::milliseconds(15)) {
+                std::this_thread::sleep_for(sleep_time - std::chrono::milliseconds(1));
+            }
+            while (std::chrono::steady_clock::now() < target_time) {
+                std::this_thread::yield();
+            }
+        } else if (frame_counter > max_frame_count) {
+            frame_counter = 0;
+            start_time = now;
+        }
+    }
+
     class Command {
     public:
         virtual ~Command() = default;
@@ -255,6 +272,7 @@ private:
         bool is_rescaling = false;
         bool rescaling_defined = false;
         bool needs_state_enable_refresh = false;
+        bool uses_transform_feedback = false;
     };
 
     struct DeferredClear {

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -252,25 +252,62 @@ void StagingBufferPool::ReleaseLevel(StagingBuffersCache& cache, size_t log2) {
     constexpr size_t deletions_per_tick = 16;
     auto& staging = cache[log2];
     auto& entries = staging.entries;
-    const size_t old_size = entries.size();
+    if (entries.empty()) {
+        staging.delete_index = 0;
+        staging.iterate_index = 0;
+        return;
+    }
 
     const auto is_deletable = [this](const StagingBuffer& entry) {
         return scheduler.IsFree(entry.tick);
     };
-    const size_t begin_offset = staging.delete_index;
-    const size_t end_offset = (std::min)(begin_offset + deletions_per_tick, old_size);
+    const size_t begin_offset = (std::min)(staging.delete_index, entries.size());
+    const size_t end_offset = (std::min)(begin_offset + deletions_per_tick, entries.size());
     const auto begin = entries.begin() + begin_offset;
     const auto end = entries.begin() + end_offset;
-    entries.erase(std::remove_if(begin, end, is_deletable), end);
+    const auto surviving_end = std::remove_if(begin, end, is_deletable);
+    const size_t removed = static_cast<size_t>(std::distance(surviving_end, end));
+    entries.erase(surviving_end, end);
 
-    const size_t new_size = entries.size();
-    staging.delete_index += deletions_per_tick;
-    if (staging.delete_index >= new_size) {
+    staging.delete_index = end_offset - removed;
+    if (staging.delete_index >= entries.size()) {
         staging.delete_index = 0;
     }
-    if (staging.iterate_index > new_size) {
+    if (staging.iterate_index > entries.size()) {
         staging.iterate_index = 0;
     }
 }
 
+u64 StagingBufferPool::ReclaimMemory(u64 target_bytes) {
+    u64 freed = 0;
+    const auto is_deletable = [this](const StagingBuffer& entry) {
+        return scheduler.IsFree(entry.tick);
+    };
+    const auto reclaim_cache = [&](StagingBuffersCache& cache) {
+        for (size_t level = NUM_LEVELS; level-- > 0 && freed < target_bytes;) {
+            auto& staging = cache[level];
+            auto& entries = staging.entries;
+            if (entries.empty()) {
+                continue;
+            }
+            const u64 entry_bytes = 1ULL << level;
+            auto it = entries.begin();
+            while (it != entries.end() && freed < target_bytes) {
+                if (is_deletable(*it)) {
+                    it = entries.erase(it);
+                    freed += entry_bytes;
+                } else {
+                    ++it;
+                }
+            }
+            staging.delete_index = 0;
+            staging.iterate_index = 0;
+        }
+    };
+    reclaim_cache(device_local_cache);
+    reclaim_cache(upload_cache);
+    reclaim_cache(download_cache);
+    return freed;
+}
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
@@ -42,6 +45,8 @@ public:
 
     void TickFrame();
 
+    u64 ReclaimMemory(u64 target_bytes);
+
 private:
     struct StreamBufferCommit {
         size_t upper_bound;

src/video_core/renderer_vulkan/vk_swapchain.cpp

@@ -9,10 +9,6 @@
 #include <limits>
 #include <vector>
 
-#ifdef __ANDROID__
-#include <android/api-level.h>
-#endif
-
 #include "common/logging.h"
 #include "common/settings.h"
 #include "common/settings_enums.h"
@@ -176,34 +172,26 @@ bool Swapchain::AcquireNextImage() {
         break;
     }
 
-    const auto wait_with_frame_pacing = [this] {
+#ifdef __ANDROID__
+    scheduler.WaitSubmitted(resource_ticks[image_index]);
+#else
     switch (Settings::values.frame_pacing_mode.GetValue()) {
     case Settings::FramePacingMode::Target_Auto:
-        scheduler.Wait(resource_ticks[image_index]);
+        scheduler.WaitSubmitted(resource_ticks[image_index]);
         break;
     case Settings::FramePacingMode::Target_30:
-        scheduler.Wait(resource_ticks[image_index], 30.0);
+        scheduler.WaitSubmitted(resource_ticks[image_index], 30.0);
         break;
     case Settings::FramePacingMode::Target_60:
-        scheduler.Wait(resource_ticks[image_index], 60.0);
+        scheduler.WaitSubmitted(resource_ticks[image_index], 60.0);
         break;
     case Settings::FramePacingMode::Target_90:
-        scheduler.Wait(resource_ticks[image_index], 90.0);
+        scheduler.WaitSubmitted(resource_ticks[image_index], 90.0);
         break;
     case Settings::FramePacingMode::Target_120:
-        scheduler.Wait(resource_ticks[image_index], 120.0);
+        scheduler.WaitSubmitted(resource_ticks[image_index], 120.0);
         break;
     }
-    };
-
-#ifdef __ANDROID__
-    if (android_get_device_api_level() >= 30) {
-        scheduler.Wait(resource_ticks[image_index]);
-    } else {
-        wait_with_frame_pacing();
-    }
-#else
-    wait_with_frame_pacing();
 #endif
 
     resource_ticks[image_index] = scheduler.CurrentTick();

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -144,11 +144,6 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
           info.size.depth == 1;
 }
 
-[[nodiscard]] bool WillUseWidenedAstcFormat(const Device& device, const ImageInfo& info) {
-    return WillUseAcceleratedAstcDecode(device, info) &&
-           !VideoCore::Surface::IsPixelFormatSRGB(info.format);
-}
-
 [[nodiscard]] VkImageCreateInfo MakeImageCreateInfo(const Device& device, const ImageInfo& info,
                                                     std::optional<VkFormat> format_override = {}) {
     auto format_info =
@@ -212,7 +207,11 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
                 return device.IsFormatSupported(view_format, VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT,
                                                 FormatType::Optimal);
             });
-        if (has_storage_compatible_view) {
+        const bool storage_allowed_for_samples =
+            image_ci.samples == VK_SAMPLE_COUNT_1_BIT ||
+            (device.GetStorageImageSampleCounts() &
+             static_cast<VkSampleCountFlags>(image_ci.samples)) != 0;
+        if (has_storage_compatible_view && storage_allowed_for_samples) {
             image_ci.usage |= VK_IMAGE_USAGE_STORAGE_BIT;
         }
 
@@ -896,6 +895,15 @@ void BlitScale(Scheduler& scheduler, VkImage src_image, VkImage dst_image, const
                        0, nullptr, nullptr, write_barriers);
     });
 }
+
+[[nodiscard]] bool CanBlitNatively(const Device& device, PixelFormat format) {
+    static constexpr auto OPTIMAL_FORMAT = FormatType::Optimal;
+    static constexpr VkFormatFeatureFlags BLIT_USAGE =
+        VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
+    const VkFormat vk_format =
+        MaxwellToVK::SurfaceFormat(device, OPTIMAL_FORMAT, false, format).format;
+    return device.IsFormatSupported(vk_format, BLIT_USAGE, OPTIMAL_FORMAT);
+}
 } // Anonymous namespace
 
 TextureCacheRuntime::TextureCacheRuntime(const Device& device_, Scheduler& scheduler_,
@@ -1228,27 +1236,19 @@ void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst
         blit_image_helper.ResolveDepthStencil(dst_framebuffer, src, dst_region, src_region);
         return;
     }
-    if (aspect_mask == (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
-        const auto format = src.format;
-        const auto can_blit_depth_stencil = [this, format] {
-            switch (format) {
-            case VideoCore::Surface::PixelFormat::D24_UNORM_S8_UINT:
-            case VideoCore::Surface::PixelFormat::S8_UINT_D24_UNORM:
-                return device.IsBlitDepth24Stencil8Supported();
-            case VideoCore::Surface::PixelFormat::D32_FLOAT_S8_UINT:
-                return device.IsBlitDepth32Stencil8Supported();
-            default:
-                UNREACHABLE();
-            }
-        }();
-        // Use shader-based depth/stencil blits if hardware doesn't support the format
-        // Note: MSAA resolves (MSAA->single) use vkCmdResolveImage which works fine
-        if (!can_blit_depth_stencil) {
-            UNIMPLEMENTED_IF(is_src_msaa || is_dst_msaa);
-            blit_image_helper.BlitDepthStencil(dst_framebuffer, src, dst_region, src_region,
-                                               filter, operation);
+    static constexpr VkImageAspectFlags DEPTH_STENCIL_ASPECTS =
+        VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
+    if ((aspect_mask & DEPTH_STENCIL_ASPECTS) != 0 && !CanBlitNatively(device, src.format)) {
+        if (aspect_mask != DEPTH_STENCIL_ASPECTS) {
+            UNIMPLEMENTED_MSG("Host cannot blit format {} and no helper path exists for aspect "
+                              "mask 0x{:x}",
+                              src.format, aspect_mask);
             return;
         }
+        UNIMPLEMENTED_IF(is_src_msaa || is_dst_msaa);
+        blit_image_helper.BlitDepthStencil(dst_framebuffer, src, dst_region, src_region, filter,
+                                           operation);
+        return;
     }
     ASSERT(!(is_dst_msaa && !is_src_msaa));
     ASSERT(operation == Fermi2D::Operation::SrcCopy);
@@ -1643,7 +1643,14 @@ void TextureCacheRuntime::CopyImageMSAA(Image& dst, Image& src,
     const u32 num_samples = msaa_to_non_msaa ? src.info.num_samples : dst.info.num_samples;
     if (dst.AspectMask() != VK_IMAGE_ASPECT_COLOR_BIT ||
         VideoCore::Surface::IsPixelFormatInteger(dst.info.format)) {
-        UNIMPLEMENTED_MSG("Copying images with different samples is not supported.");
+        const u64 key{(static_cast<u64>(dst.AspectMask()) << 32) |
+                      static_cast<u64>(dst.info.format)};
+        if (unsupported_msaa_resolves.insert(key).second) {
+            LOG_WARNING(Render_Vulkan,
+                        "MSAA resolve unsupported: format={}, aspect={:#x}, samples {}->{}",
+                        dst.info.format, dst.AspectMask(), src.info.num_samples,
+                        dst.info.num_samples);
+        }
         return;
     }
     if (ENABLE_MSAA_RESOLVE_CONSUME && msaa_to_non_msaa && copies.size() == 1 &&
@@ -1753,6 +1760,20 @@ void TextureCacheRuntime::CopyImageMSAA(Image& dst, Image& src,
                                src.info.format, num_samples, copies, msaa_to_non_msaa);
 }
 
+u64 TextureCacheRuntime::CurrentSyncPoint() const noexcept {
+    return scheduler.CurrentTick();
+}
+
+u64 TextureCacheRuntime::CompletedSyncPoint() const {
+    auto& master_semaphore = scheduler.GetMasterSemaphore();
+    master_semaphore.Refresh();
+    return master_semaphore.KnownGpuTick();
+}
+
+void TextureCacheRuntime::WaitSyncPoint(u64 sync_point) {
+    scheduler.Wait(sync_point);
+}
+
 u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
     return device.GetDeviceLocalMemory();
 }
@@ -1761,6 +1782,10 @@ u64 TextureCacheRuntime::GetDeviceMemoryUsage() const {
     return device.GetDeviceMemoryUsage();
 }
 
+u64 TextureCacheRuntime::GetDeviceAllocationUsage() const {
+    return device.GetMemoryBudgetInfo().allocation_bytes;
+}
+
 bool TextureCacheRuntime::CanReportMemoryUsage() const {
     return device.CanReportMemoryUsage();
 }
@@ -1770,6 +1795,7 @@ std::optional<size_t> TextureCacheRuntime::GetSamplerHeapBudget() const {
 }
 
 void TextureCacheRuntime::TickFrame() {
+    device.TickAllocatorFrame();
     std::erase_if(pending_msaa_images, [this](const auto& pending) {
         return scheduler.IsFree(pending.first);
     });
@@ -1780,12 +1806,7 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
     : VideoCommon::ImageBase(info_, gpu_addr_, cpu_addr_), scheduler{&runtime_.scheduler},
       runtime{&runtime_},
       original_image(MakeImage(runtime_.device, runtime_.memory_allocator, info,
-                               WillUseWidenedAstcFormat(runtime_.device, info)
-                                   ? std::span<const VkFormat>{}
-                                   : runtime->ViewFormats(info.format),
-                               WillUseWidenedAstcFormat(runtime_.device, info)
-                                   ? std::make_optional(VK_FORMAT_R32G32B32A32_SFLOAT)
-                                   : std::nullopt)),
+                               runtime->ViewFormats(info.format))),
       aspect_mask(ImageAspectMask(info.format)) {
     if (IsPixelFormatASTC(info.format) && !runtime->device.IsOptimalAstcSupported()) {
         switch (Settings::values.accelerate_astc.GetValue()) {
@@ -1816,9 +1837,7 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
         Settings::values.astc_recompression.GetValue() ==
             Settings::AstcRecompression::Uncompressed) {
         const auto& device = runtime->device.GetLogical();
-        const VkFormat storage_format = WillUseWidenedAstcFormat(runtime->device, info)
-                                            ? VK_FORMAT_R32G32B32A32_SFLOAT
-                                            : VK_FORMAT_A8B8G8R8_UNORM_PACK32;
+        const VkFormat storage_format = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
         for (s32 level = 0; level < info.resources.levels; ++level) {
             storage_image_views[level] =
                 MakeStorageView(device, level, *original_image, storage_format);
@@ -1892,9 +1911,11 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
         ScaleDown(true);
     }
 
-    const bool wants_msaa_upload = info.num_samples > 1
-        && (aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT) != 0
+    const bool is_color_upload = (aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT) != 0
         && !VideoCore::Surface::IsPixelFormatInteger(info.format);
+    const bool is_depth_upload = (aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT) != 0;
+    const bool wants_msaa_upload =
+        info.num_samples > 1 && (is_color_upload || is_depth_upload);
 
     if (wants_msaa_upload) {
         ImageInfo temp_info = info;
@@ -1933,10 +1954,10 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
             image_copies.push_back(image_copy);
         }
 
+        runtime->TransitionImageLayout(*this);
         runtime->blit_image_helper.CopyMSAA(runtime->render_pass_cache, Handle(), info.format,
                                             temp_vk_image, info.format, info.num_samples,
                                             image_copies, false);
-        initialized = true;
         runtime->pending_msaa_images.emplace_back(scheduler->CurrentTick(), std::move(temp_image));
 
         if (is_rescaled) {
@@ -1947,6 +1968,9 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
 
     if (info.num_samples > 1) {
         LOG_WARNING(Render_Vulkan, "MSAA upload not implemented for format {}", info.format);
+        if (runtime != nullptr) {
+            runtime->TransitionImageLayout(*this);
+        }
         if (is_rescaled) {
             ScaleUp();
         }
@@ -2204,9 +2228,7 @@ VkImageView Image::StorageImageView(s32 level) noexcept {
         auto format_info =
             MaxwellToVK::SurfaceFormat(runtime->device, FormatType::Optimal, true, info.format);
         if (WillUseAcceleratedAstcDecode(runtime->device, info)) {
-            format_info.format = WillUseWidenedAstcFormat(runtime->device, info)
-                                     ? VK_FORMAT_R32G32B32A32_SFLOAT
-                                     : VK_FORMAT_A8B8G8R8_UNORM_PACK32;
+            format_info.format = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
         }
         view = MakeStorageView(runtime->device.GetLogical(), level, *(this->*current_image),
                                format_info.format);
@@ -2382,11 +2404,7 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
             SanitizeDepthStencilSwizzle(swizzle, device->SupportsDepthStencilSwizzleOne());
         }
     }
-    uses_widened_astc_format = WillUseWidenedAstcFormat(*device, image.info);
     auto format_info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
-    if (uses_widened_astc_format) {
-        format_info.format = VK_FORMAT_R32G32B32A32_SFLOAT;
-    }
     if (device->ApiVersion() >= VK_API_VERSION_1_3) {
         const VkFormatProperties3 properties3 =
             device->GetPhysical().GetFormatProperties3(format_info.format);
@@ -2529,14 +2547,12 @@ VkImageView ImageView::StorageView(Shader::TextureType texture_type,
                                    Shader::ImageFormat image_format) {
     if (image_handle) {
         if (image_format == Shader::ImageFormat::Typeless) {
-            if (!typeless_storage_view) {
+            auto& view{typeless_storage_views[static_cast<size_t>(texture_type)]};
+            if (!view) {
                 auto info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
-                if (uses_widened_astc_format) {
-                    info.format = VK_FORMAT_R32G32B32A32_SFLOAT;
-                }
-                typeless_storage_view = MakeView(info.format, VK_IMAGE_ASPECT_COLOR_BIT, texture_type);
+                view = MakeView(info.format, VK_IMAGE_ASPECT_COLOR_BIT, texture_type);
             }
-            return *typeless_storage_view;
+            return *view;
         }
         const bool is_signed = image_format == Shader::ImageFormat::R8_SINT
             || image_format == Shader::ImageFormat::R16_SINT;

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -60,10 +60,22 @@ public:
 
     void TickFrame();
 
+    u64 CurrentSyncPoint() const noexcept;
+
+    u64 CompletedSyncPoint() const;
+
+    void WaitSyncPoint(u64 sync_point);
+
     u64 GetDeviceLocalMemory() const;
 
     u64 GetDeviceMemoryUsage() const;
 
+    u64 GetDeviceAllocationUsage() const;
+
+    bool CanReportAllocationUsage() const noexcept {
+        return true;
+    }
+
     bool CanReportMemoryUsage() const;
 
     std::optional<size_t> GetSamplerHeapBudget() const;
@@ -156,6 +168,7 @@ public:
     std::array<vk::Buffer, indexing_slots> buffers{};
     std::vector<std::pair<u64, vk::Image>> pending_msaa_images;
     ankerl::unordered_dense::map<VkImage, ResolveShadow> resolve_shadows;
+    ankerl::unordered_dense::set<u64> unsupported_msaa_resolves;
 };
 
 class Framebuffer {
@@ -426,7 +439,7 @@ private:
 
     std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> image_views;
     std::optional<StorageViews> storage_views;
-    vk::ImageView typeless_storage_view;
+    std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> typeless_storage_views;
     vk::ImageView depth_view;
     vk::ImageView stencil_view;
     vk::ImageView color_view;
@@ -436,7 +449,6 @@ private:
     VkSampleCountFlagBits samples = VK_SAMPLE_COUNT_1_BIT;
     u32 buffer_size = 0;
 
-    bool uses_widened_astc_format = false;
     bool supports_depth_comparison = false;
 };
 
@@ -487,6 +499,7 @@ struct TextureCacheParams {
     static constexpr bool HAS_EMULATED_COPIES = false;
     static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
     static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = true;
+    static constexpr bool HAS_TIMELINE_SYNC_POINTS = true;
 
     using Runtime = Vulkan::TextureCacheRuntime;
     using Image = Vulkan::Image;

src/video_core/renderer_vulkan/vk_update_descriptor.cpp

@@ -16,8 +16,9 @@
 
 namespace Vulkan {
 
-UpdateDescriptorQueue::UpdateDescriptorQueue(const Device& device_)
-    : device{device_}
+UpdateDescriptorQueue::UpdateDescriptorQueue(const Device& device_, size_t frame_payload_size_)
+    : device{device_}, frame_payload_size{frame_payload_size_},
+      payload(frame_payload_size_ * FRAMES_IN_FLIGHT)
 {
     payload_start = payload.data();
     payload_cursor = payload.data();
@@ -29,19 +30,19 @@ void UpdateDescriptorQueue::TickFrame() {
     if (++frame_index >= FRAMES_IN_FLIGHT) {
         frame_index = 0;
     }
-    payload_start = payload.data() + frame_index * FRAME_PAYLOAD_SIZE;
+    payload_start = payload.data() + frame_index * frame_payload_size;
     payload_cursor = payload_start;
 }
 
 void UpdateDescriptorQueue::Acquire(Scheduler& scheduler, size_t required_entries) {
     static constexpr size_t DEFAULT_REQUIRED_ENTRIES = 0x400;
     const size_t reserve = required_entries > 0 ? required_entries : DEFAULT_REQUIRED_ENTRIES;
-    ASSERT_MSG(reserve < FRAME_PAYLOAD_SIZE, "Descriptor reservation {} >= frame capacity {}",
-               reserve, FRAME_PAYLOAD_SIZE);
+    ASSERT_MSG(reserve < frame_payload_size, "Descriptor reservation {} >= frame capacity {}",
+               reserve, frame_payload_size);
     const size_t used = static_cast<size_t>(std::distance(payload_start, payload_cursor));
-    if (used + reserve >= FRAME_PAYLOAD_SIZE) {
+    if (used + reserve >= frame_payload_size) {
         LOG_WARNING(Render_Vulkan, "Payload overflow (used={}, reserve={}, capacity={})",
-                    used, reserve, FRAME_PAYLOAD_SIZE);
+                    used, reserve, frame_payload_size);
         scheduler.WaitWorker();
         payload_cursor = payload_start;
     }

src/video_core/renderer_vulkan/vk_update_descriptor.h

@@ -6,8 +6,8 @@
 
 #pragma once
 
-#include <array>
 #include <variant>
+#include <vector>
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
@@ -30,11 +30,12 @@ class UpdateDescriptorQueue final {
     // This should be plenty for the vast majority of cases. Most desktop platforms only
     // provide up to 3 swapchain images.
     static constexpr size_t FRAMES_IN_FLIGHT = 8;
-    static constexpr size_t FRAME_PAYLOAD_SIZE = 0x20000;
-    static constexpr size_t PAYLOAD_SIZE = FRAME_PAYLOAD_SIZE * FRAMES_IN_FLIGHT;
 
 public:
-    explicit UpdateDescriptorQueue(const Device& device_);
+    static constexpr size_t GUEST_FRAME_PAYLOAD_SIZE = 0x80000;
+    static constexpr size_t COMPUTE_FRAME_PAYLOAD_SIZE = 0x20000;
+
+    explicit UpdateDescriptorQueue(const Device& device_, size_t frame_payload_size_);
     ~UpdateDescriptorQueue();
 
     void TickFrame();
@@ -74,11 +75,12 @@ public:
 
 private:
     const Device& device;
+    const size_t frame_payload_size;
     size_t frame_index{0};
     DescriptorUpdateEntry* payload_cursor = nullptr;
     DescriptorUpdateEntry* payload_start = nullptr;
     const DescriptorUpdateEntry* upload_start = nullptr;
-    std::array<DescriptorUpdateEntry, PAYLOAD_SIZE> payload;
+    std::vector<DescriptorUpdateEntry> payload;
 };
 
 // TODO: should these be separate classes instead?

src/video_core/shader_environment.cpp

@@ -284,7 +284,24 @@ std::optional<u64> GenericEnvironment::TryFindSize() {
 Tegra::Texture::TICEntry GenericEnvironment::ReadTextureInfo(GPUVAddr tic_addr, u32 tic_limit,
                                                              bool via_header_index, u32 raw) {
     const auto handle{Tegra::Texture::TexturePair(raw, via_header_index)};
-    ASSERT(handle.first <= tic_limit);
+    if (handle.first > tic_limit) {
+        LOG_CRITICAL(Shader,
+                    "TIC index out of range: raw=0x{:08x} tic_index={} tsc_index={} tic_limit={} "
+                    "tic_addr=0x{:x} via_header_index={} stage={} program_base=0x{:x} "
+                    "start_address=0x{:x}",
+                    raw, handle.first, handle.second, tic_limit, tic_addr, via_header_index,
+                    static_cast<u32>(stage), program_base, start_address);
+        ASSERT(handle.first <= tic_limit);
+        Tegra::Texture::TICEntry fallback{};
+        fallback.format.Assign(Tegra::Texture::TextureFormat::A8B8G8R8);
+        fallback.r_type.Assign(Tegra::Texture::ComponentType::UNORM);
+        fallback.g_type.Assign(Tegra::Texture::ComponentType::UNORM);
+        fallback.b_type.Assign(Tegra::Texture::ComponentType::UNORM);
+        fallback.a_type.Assign(Tegra::Texture::ComponentType::UNORM);
+        fallback.texture_type.Assign(Tegra::Texture::TextureType::Texture2D);
+        fallback.normalized_coords.Assign(1);
+        return fallback;
+    }
     const GPUVAddr descriptor_addr{tic_addr + handle.first * sizeof(Tegra::Texture::TICEntry)};
     Tegra::Texture::TICEntry entry;
     gpu_memory->ReadBlock(descriptor_addr, &entry, sizeof(entry));

src/video_core/texture_cache/texture_cache.h

@@ -58,23 +58,9 @@ TextureCache<P>::TextureCache(Runtime& runtime_, Tegra::MaxwellDeviceMemoryManag
     void(slot_samplers.insert(runtime, sampler_descriptor));
 
     if constexpr (HAS_DEVICE_MEMORY_INFO) {
-        const s64 device_local_memory = static_cast<s64>(runtime.GetDeviceLocalMemory());
-        const s64 min_spacing_expected = device_local_memory - 1_GiB;
-        const s64 min_spacing_critical = device_local_memory - 512_MiB;
-        const s64 mem_threshold = (std::min)(device_local_memory, TARGET_THRESHOLD);
-        const s64 min_vacancy_expected = (6 * mem_threshold) / 10;
-        const s64 min_vacancy_critical = (2 * mem_threshold) / 10;
-        expected_memory = static_cast<u64>(
-            (std::max)((std::min)(device_local_memory - min_vacancy_expected, min_spacing_expected),
-                     DEFAULT_EXPECTED_MEMORY));
-        critical_memory = static_cast<u64>(
-            (std::max)((std::min)(device_local_memory - min_vacancy_critical, min_spacing_critical),
-                     DEFAULT_CRITICAL_MEMORY));
-        minimum_memory = static_cast<u64>((device_local_memory - mem_threshold) / 2);
+        memory_budget = runtime.GetDeviceLocalMemory();
     } else {
-        expected_memory = DEFAULT_EXPECTED_MEMORY + 512_MiB;
-        critical_memory = DEFAULT_CRITICAL_MEMORY + 1_GiB;
-        minimum_memory = 0;
+        memory_budget = FALLBACK_MEMORY_BUDGET;
     }
 
     const bool gpu_unswizzle_enabled = Settings::values.gpu_unswizzle_enabled.GetValue();
@@ -114,71 +100,154 @@ TextureCache<P>::TextureCache(Runtime& runtime_, Tegra::MaxwellDeviceMemoryManag
 }
 
 template <class P>
-void TextureCache<P>::RunGarbageCollector() {
-    bool high_priority_mode = false;
-    bool aggressive_mode = false;
-    u64 ticks_to_destroy = 0;
-    size_t num_iterations = 0;
-    const auto Configure = [&](bool allow_aggressive) {
-        high_priority_mode = total_used_memory >= expected_memory;
-        aggressive_mode = allow_aggressive && total_used_memory >= critical_memory;
-        ticks_to_destroy = aggressive_mode ? 10ULL : high_priority_mode ? 25ULL : 50ULL;
-        num_iterations = aggressive_mode ? 40 : (high_priority_mode ? 20 : 10);
-    };
-    const auto Cleanup = [this, &num_iterations, &high_priority_mode, &aggressive_mode](ImageId image_id) {
-        if (num_iterations == 0) {
-            return true;
-        }
-        --num_iterations;
-        auto& image = slot_images[image_id];
-        if (True(image.flags & ImageFlagBits::IsDecoding)) {
-            return false;
-        }
-        const bool must_download = image.IsSafeDownload() && False(image.flags & ImageFlagBits::BadOverlap);
-        if ((!aggressive_mode && True(image.flags & ImageFlagBits::CostlyLoad)) || (!high_priority_mode && must_download)) {
-            return false;
-        }
-        if (must_download) {
-            auto map = runtime.DownloadStagingBuffer(image.unswizzled_size_bytes);
-            const auto copies = FixSmallVectorADL(FullDownloadCopies(image.info));
-            image.DownloadMemory(map, copies);
-            runtime.Finish();
-            SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span, swizzle_data_buffer);
-        }
-        if (True(image.flags & ImageFlagBits::Tracked)) {
-            UntrackImage(image, image_id);
-        }
-        UnregisterImage(image_id);
-        DeleteImage(image_id, image.scale_tick > frame_tick + 5);
-        if (aggressive_mode && total_used_memory < critical_memory) {
-            num_iterations >>= 2;
-            aggressive_mode = false;
-        } else if (high_priority_mode && total_used_memory < expected_memory) {
-            num_iterations >>= 1;
-            high_priority_mode = false;
-        }
-        return false;
-    };
-    Configure(false);
-    lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, Cleanup);
-    if (total_used_memory >= critical_memory) {
-        Configure(true);
-        lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, Cleanup);
+void TextureCache<P>::QueueEvictionDownload(Image& image) {
+    auto copies = FullDownloadCopies(image.info);
+    auto staging = runtime.DownloadStagingBuffer(image.unswizzled_size_bytes, true);
+    image.DownloadMemory(staging, FixSmallVectorADL(copies));
+    pending_eviction_downloads.push_back(PendingEvictionDownload{
+        .staging = staging,
+        .gpu_memory = gpu_memory,
+        .copies = std::move(copies),
+        .info = image.info,
+        .gpu_addr = image.gpu_addr,
+        .sync_point = runtime.CurrentSyncPoint(),
+    });
+}
+
+template <class P>
+void TextureCache<P>::TickEvictionDownloads(u64 completed_sync_point) {
+    while (!pending_eviction_downloads.empty() &&
+           pending_eviction_downloads.front().sync_point <= completed_sync_point) {
+        auto& entry = pending_eviction_downloads.front();
+        SwizzleImage(*entry.gpu_memory, entry.gpu_addr, entry.info, FixSmallVectorADL(entry.copies),
+                     entry.staging.mapped_span.subspan(entry.staging.offset), swizzle_data_buffer);
+        runtime.FreeDeferredStagingBuffer(entry.staging);
+        pending_eviction_downloads.pop_front();
     }
 }
 
+template <class P>
+void TextureCache<P>::FlushEvictionDownloads() {
+    if (pending_eviction_downloads.empty()) {
+        return;
+    }
+    const u64 last_sync_point = pending_eviction_downloads.back().sync_point;
+    runtime.WaitSyncPoint(last_sync_point);
+    TickEvictionDownloads(last_sync_point);
+}
+
+template <class P>
+u64 TextureCache<P>::ImageSizeBytes(const ImageBase& image) {
+    u64 tentative_size = (std::max)(image.guest_size_bytes, image.unswizzled_size_bytes);
+    if ((IsPixelFormatASTC(image.info.format) &&
+         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
+        True(image.flags & ImageFlagBits::Converted)) {
+        tentative_size = TranscodedAstcSize(tentative_size, image.info.format);
+    }
+    u64 size = Common::AlignUp(tentative_size, 1024);
+    if (image.HasScaled()) {
+        size += GetScaledImageSizeBytes(image);
+    }
+    return size;
+}
+
+template <class P>
+u64 TextureCache<P>::DeviceUsage(bool force_refresh) {
+    if (!runtime.CanReportAllocationUsage()) {
+        return total_used_memory;
+    }
+    if (force_refresh || usage_refresh_countdown == 0) {
+        cached_device_usage = runtime.GetDeviceAllocationUsage();
+        usage_refresh_countdown = USAGE_REFRESH_INTERVAL;
+    } else {
+        --usage_refresh_countdown;
+    }
+    return cached_device_usage;
+}
+
+template <class P>
+u64 TextureCache<P>::ReclaimMemory(u64 target_bytes, bool allow_download) {
+    if (target_bytes == 0 || in_reclaim) {
+        return 0;
+    }
+    in_reclaim = true;
+    const u64 drain_point = runtime.CompletedSyncPoint();
+    TickEvictionDownloads(drain_point);
+    sentenced_images.Reclaim(drain_point);
+    sentenced_image_view.Reclaim(drain_point);
+    sentenced_framebuffers.Reclaim(drain_point);
+    u64 freed = 0;
+    const auto evict = [&](ImageId image_id) {
+        if (freed >= target_bytes) {
+            return true;
+        }
+        auto& image = slot_images[image_id];
+        if (True(image.flags & ImageFlagBits::IsDecoding)) {
+            return false;
+        }
+        const bool must_download =
+            image.IsSafeDownload() && False(image.flags & ImageFlagBits::BadOverlap);
+        bool queued_download = false;
+        if (must_download) {
+            if constexpr (HAS_TIMELINE_SYNC_POINTS) {
+                if (!allow_download) {
+                    return false;
+                }
+                QueueEvictionDownload(image);
+                queued_download = true;
+            } else {
+                return false;
+            }
+        }
+        const u64 image_bytes = ImageSizeBytes(image);
+        if (True(image.flags & ImageFlagBits::Tracked)) {
+            UntrackImage(image, image_id);
+        }
+        UnregisterImage(image_id);
+        DeleteImage(image_id, !queued_download && image.scale_tick > frame_tick + 5);
+        freed += image_bytes;
+        return false;
+    };
+    const u64 cold_tick =
+        frame_tick > RECLAIM_GUARD_FRAMES ? frame_tick - RECLAIM_GUARD_FRAMES : 0;
+    lru_cache.ForEachItemBelow(cold_tick, evict);
+    if (freed < target_bytes) {
+        lru_cache.ForEachItemBelow(frame_tick > 0 ? frame_tick - 1 : 0, evict);
+    }
+    const u64 exit_point = runtime.CompletedSyncPoint();
+    sentenced_images.Reclaim(exit_point);
+    sentenced_image_view.Reclaim(exit_point);
+    sentenced_framebuffers.Reclaim(exit_point);
+    in_reclaim = false;
+    usage_refresh_countdown = 0;
+    reclaim_stalled = freed == 0;
+    return freed;
+}
+
+template <class P>
+void TextureCache<P>::EnsureHeadroom(bool allow_download) {
+    if (reclaim_stalled) {
+        return;
+    }
+    const u64 limit = memory_budget > RECLAIM_HEADROOM ? memory_budget - RECLAIM_HEADROOM : 0;
+    const u64 usage = DeviceUsage(false);
+    if (usage <= limit) {
+        return;
+    }
+    const u64 target = (limit / 100) * RECLAIM_TARGET_PERCENT;
+    ReclaimMemory((std::min)(usage - target, total_used_memory), allow_download);
+}
+
 template <class P>
 void TextureCache<P>::TickFrame() {
-    // If we can obtain the memory info, use it instead of the estimate.
-    if (runtime.CanReportMemoryUsage()) {
-        total_used_memory = runtime.GetDeviceMemoryUsage();
-    }
-    if (total_used_memory > minimum_memory) {
-        RunGarbageCollector();
-    }
-    sentenced_images.Tick();
-    sentenced_framebuffers.Tick();
-    sentenced_image_view.Tick();
+    usage_refresh_countdown = 0;
+    reclaim_stalled = false;
+    EnsureHeadroom(true);
+    const u64 completed_sync_point = runtime.CompletedSyncPoint();
+    TickEvictionDownloads(completed_sync_point);
+    sentenced_images.Reclaim(completed_sync_point);
+    sentenced_framebuffers.Reclaim(completed_sync_point);
+    sentenced_image_view.Reclaim(completed_sync_point);
     TickAsyncDecode();
     TickAsyncUnswizzle();
 
@@ -596,6 +665,7 @@ void TextureCache<P>::WriteMemory(DAddr cpu_addr, size_t size) {
 
 template <class P>
 void TextureCache<P>::DownloadMemory(DAddr cpu_addr, size_t size) {
+    FlushEvictionDownloads();
     boost::container::small_vector<ImageId, 16> images;
     ForEachImageInRegion(cpu_addr, size, [&images](ImageId image_id, ImageBase& image) {
         if (!image.IsSafeDownload()) {
@@ -894,6 +964,7 @@ void TextureCache<P>::CommitAsyncFlushes() {
 
 template <class P>
 void TextureCache<P>::PopAsyncFlushes() {
+    TickEvictionDownloads(runtime.CompletedSyncPoint());
     if (committed_downloads.empty()) {
         return;
     }
@@ -1294,8 +1365,9 @@ void TextureCache<P>::InvalidateScale(Image& image) {
     }
     RemoveImageViewReferences(image_view_ids);
     RemoveFramebuffers(image_view_ids);
+    const u64 sync_point = runtime.CurrentSyncPoint();
     for (const ImageViewId image_view_id : image_view_ids) {
-        sentenced_image_view.Push(std::move(slot_image_views[image_view_id]));
+        sentenced_image_view.Push(std::move(slot_image_views[image_view_id]), sync_point);
         slot_image_views.erase(image_view_id);
     }
     image.image_view_ids.clear();
@@ -1331,6 +1403,7 @@ void TextureCache<P>::QueueAsyncDecode(Image& image, ImageId image_id) {
     LOG_INFO(HW_GPU, "Queuing async texture decode");
 
     image.flags |= ImageFlagBits::IsDecoding;
+    runtime.TransitionImageLayout(image);
     auto decode = std::make_unique<AsyncDecodeContext>();
     auto* decode_ptr = decode.get();
     decode->image_id = image_id;
@@ -1363,6 +1436,7 @@ void TextureCache<P>::QueueAsyncUnswizzle(Image& image, ImageId image_id) {
     }
 
     image.flags |= ImageFlagBits::IsDecoding;
+    runtime.TransitionImageLayout(image);
 
     unswizzle_queue.push_back({
         .image_id = image_id,
@@ -1523,6 +1597,7 @@ ImageId TextureCache<P>::InsertImage(const ImageInfo& info, GPUVAddr gpu_addr,
 
 template <class P>
 ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, DAddr cpu_addr) {
+    EnsureHeadroom(false);
     ImageInfo new_info = info;
     const size_t size_bytes = CalculateGuestSizeInBytes(new_info);
     const bool broken_views = runtime.HasBrokenTextureViewFormats();
@@ -1631,7 +1706,28 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, DA
     for (const ImageId overlap_id : join_ignore_textures) {
         Image& overlap = slot_images[overlap_id];
         if (True(overlap.flags & ImageFlagBits::GpuModified)) {
-            UNIMPLEMENTED();
+            if (new_image.TryFindBase(overlap.gpu_addr) &&
+                (!can_rescale || ImageCanRescale(overlap))) {
+                if (can_rescale) {
+                    ScaleUp(overlap);
+                } else {
+                    ScaleDown(overlap);
+                }
+                join_copies_to_do.emplace_back(JoinCopy{false, overlap_id});
+                continue;
+            }
+            if (overlap.IsSafeDownload() && False(overlap.flags & ImageFlagBits::BadOverlap) &&
+                gpu_memory->GpuToCpuAddress(overlap.gpu_addr).has_value()) {
+                QueueEvictionDownload(overlap);
+            } else {
+                LOG_WARNING(HW_GPU,
+                            "Dropping GPU modified overlap, contents are not recoverable: "
+                            "gpu_addr=0x{:x} format={} size={}x{}x{} levels={} layers={}",
+                            overlap.gpu_addr, static_cast<int>(overlap.info.format),
+                            overlap.info.size.width, overlap.info.size.height,
+                            overlap.info.size.depth, overlap.info.resources.levels,
+                            overlap.info.resources.layers);
+            }
         }
         if (True(overlap.flags & ImageFlagBits::Tracked)) {
             UntrackImage(overlap, overlap_id);
@@ -2185,13 +2281,7 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
     ASSERT_MSG(False(image.flags & ImageFlagBits::Registered),
                "Trying to register an already registered image");
     image.flags |= ImageFlagBits::Registered;
-    u64 tentative_size = (std::max)(image.guest_size_bytes, image.unswizzled_size_bytes);
-    if ((IsPixelFormatASTC(image.info.format) &&
-         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
-        True(image.flags & ImageFlagBits::Converted)) {
-        tentative_size = TranscodedAstcSize(tentative_size, image.info.format);
-    }
-    total_used_memory += Common::AlignUp(tentative_size, 1024);
+    total_used_memory += ImageSizeBytes(image);
     image.lru_index = lru_cache.Insert(image_id, frame_tick);
 
     ForEachGPUPage(image.gpu_addr, image.guest_size_bytes, [this, image_id](u64 page) {
@@ -2354,16 +2444,7 @@ void TextureCache<P>::UntrackImage(ImageBase& image, ImageId image_id) {
 template <class P>
 void TextureCache<P>::DeleteImage(ImageId image_id, bool immediate_delete) {
     ImageBase& image = slot_images[image_id];
-    if (image.HasScaled()) {
-        total_used_memory -= GetScaledImageSizeBytes(image);
-    }
-    u64 tentative_size = (std::max)(image.guest_size_bytes, image.unswizzled_size_bytes);
-    if ((IsPixelFormatASTC(image.info.format) &&
-         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
-        True(image.flags & ImageFlagBits::Converted)) {
-        tentative_size = TranscodedAstcSize(tentative_size, image.info.format);
-    }
-    total_used_memory -= Common::AlignUp(tentative_size, 1024);
+    total_used_memory -= std::min<u64>(total_used_memory, ImageSizeBytes(image));
     const GPUVAddr gpu_addr = image.gpu_addr;
     const auto alloc_it = image_allocs_table.find(gpu_addr);
     if (alloc_it == image_allocs_table.end()) {
@@ -2417,14 +2498,15 @@ void TextureCache<P>::DeleteImage(ImageId image_id, bool immediate_delete) {
         ASSERT_MSG(num_removed_overlaps == 1, "Invalid number of removed overlapps: {}",
                    num_removed_overlaps);
     }
+    const u64 sync_point = runtime.CurrentSyncPoint();
     for (const ImageViewId image_view_id : image_view_ids) {
         if (!immediate_delete) {
-            sentenced_image_view.Push(std::move(slot_image_views[image_view_id]));
+            sentenced_image_view.Push(std::move(slot_image_views[image_view_id]), sync_point);
         }
         slot_image_views.erase(image_view_id);
     }
     if (!immediate_delete) {
-        sentenced_images.Push(std::move(slot_images[image_id]));
+        sentenced_images.Push(std::move(slot_images[image_id]), sync_point);
     }
     slot_images.erase(image_id);
 
@@ -2470,7 +2552,8 @@ void TextureCache<P>::RemoveFramebuffers(std::span<const ImageViewId> removed_vi
                 last_framebuffer_id = {};
                 last_framebuffer_serial = 0;
             }
-            sentenced_framebuffers.Push(std::move(slot_framebuffers[framebuffer_id]));
+            sentenced_framebuffers.Push(std::move(slot_framebuffers[framebuffer_id]),
+                                        runtime.CurrentSyncPoint());
             it = framebuffers.erase(it);
         } else {
             ++it;

src/video_core/texture_cache/texture_cache_base.h

@@ -30,7 +30,7 @@
 #include "common/thread_worker.h"
 #include "video_core/compatible_formats.h"
 #include "video_core/control/channel_state_cache.h"
-#include "video_core/delayed_destruction_ring.h"
+#include "video_core/deferred_destruction_queue.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/surface.h"
 #include "video_core/texture_cache/descriptor_table.h"
@@ -108,18 +108,20 @@ class TextureCache : public VideoCommon::ChannelSetupCaches<TextureCacheChannelI
     static constexpr bool HAS_DEVICE_MEMORY_INFO = P::HAS_DEVICE_MEMORY_INFO;
     /// True when the API can do asynchronous texture downloads.
     static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = P::IMPLEMENTS_ASYNC_DOWNLOADS;
+    static constexpr bool HAS_TIMELINE_SYNC_POINTS = P::HAS_TIMELINE_SYNC_POINTS;
 
     static constexpr size_t UNSET_CHANNEL{(std::numeric_limits<size_t>::max)()};
 
 #ifdef YUZU_LEGACY
-    static constexpr s64 TARGET_THRESHOLD = 3_GiB;
+    static constexpr u64 RECLAIM_HEADROOM = 384_MiB;
 #else
-    static constexpr s64 TARGET_THRESHOLD = 4_GiB;
+    static constexpr u64 RECLAIM_HEADROOM = 512_MiB;
 #endif
 
-    static constexpr s64 DEFAULT_EXPECTED_MEMORY = 1_GiB + 125_MiB;
-    static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB + 625_MiB;
-    static constexpr size_t GC_EMERGENCY_COUNTS = 2;
+    static constexpr u64 FALLBACK_MEMORY_BUDGET = 2_GiB;
+    static constexpr u32 USAGE_REFRESH_INTERVAL = 16;
+    static constexpr u64 RECLAIM_GUARD_FRAMES = 8;
+    static constexpr u64 RECLAIM_TARGET_PERCENT = 88;
 
     using Runtime = typename P::Runtime;
     using Image = typename P::Image;
@@ -154,6 +156,8 @@ public:
     /// Notify the cache that a new frame has been queued
     void TickFrame();
 
+    u64 ReclaimMemory(u64 target_bytes, bool allow_download);
+
     /// Return a constant reference to the given image view id
     [[nodiscard]] const ImageView& GetImageView(ImageViewId id) const noexcept;
 
@@ -293,8 +297,17 @@ private:
 
     void OnGPUASRegister(size_t map_id) final override;
 
-    /// Runs the Garbage Collector.
-    void RunGarbageCollector();
+    u64 ImageSizeBytes(const ImageBase& image);
+
+    u64 DeviceUsage(bool force_refresh);
+
+    void EnsureHeadroom(bool allow_download);
+
+    void QueueEvictionDownload(Image& image);
+
+    void TickEvictionDownloads(u64 completed_sync_point);
+
+    void FlushEvictionDownloads();
 
     /// Find or create an image view in the guest descriptor table
     ImageViewId VisitImageView(u32 index, bool compute);
@@ -451,9 +464,11 @@ private:
     bool has_deleted_images = false;
     bool is_rescaling = false;
     u64 total_used_memory = 0;
-    u64 minimum_memory;
-    u64 expected_memory;
-    u64 critical_memory;
+    u64 memory_budget = 0;
+    u64 cached_device_usage = 0;
+    u32 usage_refresh_countdown = 0;
+    bool in_reclaim = false;
+    bool reclaim_stalled = false;
     size_t gpu_unswizzle_maxsize = 0;
     size_t swizzle_chunk_size = 0;
     u32 swizzle_slices_per_batch = 0;
@@ -491,14 +506,19 @@ private:
     };
     Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
 
- #ifdef YUZU_LEGACY
-    static constexpr size_t TICKS_TO_DESTROY = 6;
- #else
-    static constexpr size_t TICKS_TO_DESTROY = 8;
-#endif
-    DelayedDestructionRing<Image, TICKS_TO_DESTROY> sentenced_images;
-    DelayedDestructionRing<ImageView, TICKS_TO_DESTROY> sentenced_image_view;
-    DelayedDestructionRing<Framebuffer, TICKS_TO_DESTROY> sentenced_framebuffers;
+    DeferredDestructionQueue<Image> sentenced_images;
+    DeferredDestructionQueue<ImageView> sentenced_image_view;
+    DeferredDestructionQueue<Framebuffer> sentenced_framebuffers;
+
+    struct PendingEvictionDownload {
+        AsyncBuffer staging;
+        Tegra::MemoryManager* gpu_memory;
+        boost::container::small_vector<VideoCommon::BufferImageCopy, 16> copies;
+        VideoCommon::ImageInfo info;
+        GPUVAddr gpu_addr;
+        u64 sync_point;
+    };
+    std::deque<PendingEvictionDownload> pending_eviction_downloads;
 
     ankerl::unordered_dense::map<GPUVAddr, ImageAllocId> image_allocs_table;
 
@@ -509,7 +529,8 @@ private:
     u64 frame_tick = 0;
     u64 last_sampler_gc_frame = (std::numeric_limits<u64>::max)();
 
-    Common::ThreadWorker texture_decode_worker{1, "TextureDecoder"};
+    Common::ThreadWorker texture_decode_worker{1, "TextureDecoder", {},
+                                               Common::ThreadPlacement::Background};
     std::vector<std::unique_ptr<AsyncDecodeContext>> async_decodes;
 
     std::deque<PendingUnswizzle> unswizzle_queue;

src/video_core/texture_cache/util.cpp

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
@@ -747,7 +747,7 @@ boost::container::small_vector<ImageCopy, 16> MakeShrinkImageCopies(const ImageI
 
     const bool is_dst_3d = dst.type == ImageType::e3D;
     if (is_dst_3d) {
-        ASSERT(src.type == ImageType::e3D);
+        ASSERT(src.type == ImageType::e3D || src.resources.layers == 1);
         ASSERT(src.resources.levels == 1);
     }
     const bool both_2d{src.type == ImageType::e2D && dst.type == ImageType::e2D};

src/video_core/textures/workers.cpp

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
@@ -10,7 +10,8 @@ namespace Tegra::Texture {
 
 Common::ThreadWorker& GetThreadWorkers() {
     static Common::ThreadWorker workers{(std::max)(std::thread::hardware_concurrency(), 2U) / 2,
-                                        "ImageTranscode"};
+                                        "ImageTranscode", {},
+                                        Common::ThreadPlacement::Background};
 
     return workers;
 }

src/video_core/vulkan_common/vulkan_device.cpp

@@ -5,6 +5,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <algorithm>
+#include <array>
 #include <bitset>
 #include <chrono>
 #include <optional>
@@ -732,7 +733,7 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
             .device = *logical,
             .preferredLargeHeapBlockSize = is_integrated
                                            ? (64u * 1024u * 1024u)
-                                           : (256u * 1024u * 1024u),
+                                           : (128u * 1024u * 1024u),
             .pAllocationCallbacks = nullptr,
             .pDeviceMemoryCallbacks = nullptr,
             .pHeapSizeLimit = nullptr,
@@ -744,12 +745,32 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
 
     vk::Check(vmaCreateAllocator(&allocator_info, &allocator));
 
+    {
+        const auto& limits = properties.properties.limits;
+        LOG_INFO(Render_Vulkan, "MSAA sample count support:");
+        LOG_INFO(Render_Vulkan, "  framebufferColorSampleCounts:   {:#x}",
+                 limits.framebufferColorSampleCounts);
+        LOG_INFO(Render_Vulkan, "  framebufferDepthSampleCounts:   {:#x}",
+                 limits.framebufferDepthSampleCounts);
+        LOG_INFO(Render_Vulkan, "  framebufferStencilSampleCounts: {:#x}",
+                 limits.framebufferStencilSampleCounts);
+        LOG_INFO(Render_Vulkan, "  sampledImageColorSampleCounts:  {:#x}",
+                 limits.sampledImageColorSampleCounts);
+        LOG_INFO(Render_Vulkan, "  sampledImageDepthSampleCounts:  {:#x}",
+                 limits.sampledImageDepthSampleCounts);
+        LOG_INFO(Render_Vulkan, "  sampledImageIntegerSampleCounts:{:#x}",
+                 limits.sampledImageIntegerSampleCounts);
+        LOG_INFO(Render_Vulkan, "  storageImageSampleCounts:       {:#x}",
+                 limits.storageImageSampleCounts);
+    }
+
     // Initialize GPU logging if enabled
     InitializeGPULogging();
 }
 
 Device::~Device() {
     ShutdownGPULogging();
+    vk::FlushDeletionQueue();
     vmaDestroyAllocator(allocator);
 }
 
@@ -817,7 +838,8 @@ bool Device::ComputeIsOptimalAstcSupported() const {
         VK_FORMAT_ASTC_12x10_UNORM_BLOCK, VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
         VK_FORMAT_ASTC_12x12_UNORM_BLOCK, VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
     };
-    if (!features.features.textureCompressionASTC_LDR) {
+    if (!features.features.textureCompressionASTC_LDR ||
+        !features.texture_compression_astc_hdr.textureCompressionASTC_HDR) {
         return false;
     }
     const auto format_feature_usage{VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
@@ -1209,10 +1231,20 @@ void Device::RemoveUnsuitableExtensions() {
     RemoveExtensionFeatureIfUnsuitable(extensions.depth_bias_control, features.depth_bias_control,
                                        VK_EXT_DEPTH_BIAS_CONTROL_EXTENSION_NAME);
 
+    // VK_EXT_depth_clamp_zero_one
+    extensions.depth_clamp_zero_one = features.depth_clamp_zero_one.depthClampZeroOne;
+    RemoveExtensionFeatureIfUnsuitable(extensions.depth_clamp_zero_one,
+                                       features.depth_clamp_zero_one,
+                                       VK_EXT_DEPTH_CLAMP_ZERO_ONE_EXTENSION_NAME);
+
     // VK_EXT_depth_clip_control
     extensions.depth_clip_control = features.depth_clip_control.depthClipControl;
     RemoveExtensionFeatureIfUnsuitable(extensions.depth_clip_control, features.depth_clip_control,
                                        VK_EXT_DEPTH_CLIP_CONTROL_EXTENSION_NAME);
+    // VK_EXT_depth_clip_enable
+    extensions.depth_clip_enable = features.depth_clip_enable.depthClipEnable;
+    RemoveExtensionFeatureIfUnsuitable(extensions.depth_clip_enable, features.depth_clip_enable,
+                                       VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME);
 
     // VK_EXT_extended_dynamic_state
     extensions.extended_dynamic_state = features.extended_dynamic_state.extendedDynamicState;
@@ -1444,6 +1476,23 @@ std::optional<size_t> Device::GetSamplerHeapBudget() const {
     return sampler_heap_budget;
 }
 
+Device::MemoryBudgetInfo Device::GetMemoryBudgetInfo() const {
+    std::array<VmaBudget, VK_MAX_MEMORY_HEAPS> budgets{};
+    vmaGetHeapBudgets(allocator, budgets.data());
+    MemoryBudgetInfo info{};
+    for (const size_t heap : valid_heap_memory) {
+        info.usage += budgets[heap].usage;
+        info.budget += budgets[heap].budget;
+        info.block_bytes += budgets[heap].statistics.blockBytes;
+        info.allocation_bytes += budgets[heap].statistics.allocationBytes;
+    }
+    return info;
+}
+
+void Device::TickAllocatorFrame() const {
+    vmaSetCurrentFrameIndex(allocator, ++allocator_frame_index);
+}
+
 u64 Device::GetDeviceMemoryUsage() const {
     VkPhysicalDeviceMemoryBudgetPropertiesEXT budget;
     budget.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT;
@@ -1493,9 +1542,12 @@ void Device::CollectPhysicalMemoryInfo() {
         device_access_memory -= reserve_memory;
         if (Settings::values.vram_usage_mode.GetValue() != Settings::VramUsageMode::Aggressive) {
             // Account for resolution scaling in memory limits
-            const size_t normal_memory = 6_GiB;
-            const size_t scaler_memory = 1_GiB * Settings::values.resolution_info.ScaleUp(1);
-            device_access_memory = std::min<u64>(device_access_memory, normal_memory + scaler_memory);
+            const u64 normal_memory = 6_GiB;
+            const u64 scaler_memory = 1_GiB * Settings::values.resolution_info.ScaleUp(1);
+            const u64 baseline = normal_memory + scaler_memory;
+            const u64 proportional = (device_access_memory / 4) * 3;
+            device_access_memory =
+                std::min<u64>(device_access_memory, std::max<u64>(baseline, proportional));
         }
     }
 }

src/video_core/vulkan_common/vulkan_device.h

@@ -54,7 +54,9 @@ VK_DEFINE_HANDLE(VmaAllocator)
     FEATURE(EXT, ColorWriteEnable, COLOR_WRITE_ENABLE, color_write_enable)                         \
     FEATURE(EXT, CustomBorderColor, CUSTOM_BORDER_COLOR, custom_border_color)                      \
     FEATURE(EXT, DepthBiasControl, DEPTH_BIAS_CONTROL, depth_bias_control)                         \
+    FEATURE(EXT, DepthClampZeroOne, DEPTH_CLAMP_ZERO_ONE, depth_clamp_zero_one)                   \
     FEATURE(EXT, DepthClipControl, DEPTH_CLIP_CONTROL, depth_clip_control)                         \
+    FEATURE(EXT, DepthClipEnable, DEPTH_CLIP_ENABLE, depth_clip_enable)                            \
     FEATURE(EXT, ExtendedDynamicState, EXTENDED_DYNAMIC_STATE, extended_dynamic_state)             \
     FEATURE(EXT, ExtendedDynamicState2, EXTENDED_DYNAMIC_STATE_2, extended_dynamic_state2)         \
     FEATURE(EXT, ExtendedDynamicState3, EXTENDED_DYNAMIC_STATE_3, extended_dynamic_state3)         \
@@ -255,6 +257,17 @@ public:
         return allocator;
     }
 
+    struct MemoryBudgetInfo {
+        u64 usage;
+        u64 budget;
+        u64 block_bytes;
+        u64 allocation_bytes;
+    };
+
+    MemoryBudgetInfo GetMemoryBudgetInfo() const;
+
+    void TickAllocatorFrame() const;
+
     /// Returns the logical device.
     const vk::Device& GetLogical() const {
         return logical;
@@ -369,8 +382,7 @@ FN_MAX_LIMIT_LIST
     }
 
     bool IsOptimalAstcSupported() const {
-        return features.features.textureCompressionASTC_LDR &&
-               features.texture_compression_astc_hdr.textureCompressionASTC_HDR;
+        return is_optimal_astc_supported;
     }
 
     /// Returns true if BCn is natively supported.
@@ -602,6 +614,16 @@ FN_MAX_LIMIT_LIST
         return extensions.depth_clip_control;
     }
 
+    /// Returns true if the device supports VK_EXT_depth_clamp_zero_one.
+    bool IsExtDepthClampZeroOneSupported() const {
+        return extensions.depth_clamp_zero_one;
+    }
+
+    /// Returns true if the device supports VK_EXT_depth_clip_enable.
+    bool IsExtDepthClipEnableSupported() const {
+        return extensions.depth_clip_enable;
+    }
+
     /// Returns true if the device supports VK_EXT_depth_bias_control.
     bool IsExtDepthBiasControlSupported() const {
         return extensions.depth_bias_control;
@@ -734,6 +756,38 @@ FN_MAX_LIMIT_LIST
         return features.line_rasterization.stippledRectangularLines != VK_FALSE;
     }
 
+    VkLineRasterizationModeEXT GetLineRasterizationMode(bool wants_smooth) const {
+        if (wants_smooth && SupportsSmoothLines()) {
+            return VK_LINE_RASTERIZATION_MODE_RECTANGULAR_SMOOTH_EXT;
+        }
+        if (SupportsRectangularLines()) {
+            return VK_LINE_RASTERIZATION_MODE_RECTANGULAR_EXT;
+        }
+        return VK_LINE_RASTERIZATION_MODE_DEFAULT_EXT;
+    }
+
+    bool SupportsStippleForMode(VkLineRasterizationModeEXT mode) const {
+        switch (mode) {
+        case VK_LINE_RASTERIZATION_MODE_RECTANGULAR_SMOOTH_EXT:
+            return features.line_rasterization.stippledSmoothLines != VK_FALSE;
+        case VK_LINE_RASTERIZATION_MODE_BRESENHAM_EXT:
+            return features.line_rasterization.stippledBresenhamLines != VK_FALSE;
+        default:
+            return features.line_rasterization.stippledRectangularLines != VK_FALSE;
+        }
+    }
+
+    float ClampLineWidth(float width) const {
+        if (!features.features.wideLines) {
+            return 1.0f;
+        }
+        const auto& range = properties.properties.limits.lineWidthRange;
+        if (!(width >= range[0])) {
+            return range[0];
+        }
+        return width > range[1] ? range[1] : width;
+    }
+
     bool SupportsAlphaToOne() const {
         return features.features.alphaToOne != VK_FALSE;
     }
@@ -876,6 +930,10 @@ FN_MAX_LIMIT_LIST
 
     u64 GetDeviceMemoryUsage() const;
 
+    VkSampleCountFlags GetStorageImageSampleCounts() const {
+        return properties.properties.limits.storageImageSampleCounts;
+    }
+
     u32 GetSetsPerPool() const {
         return sets_per_pool;
     }
@@ -1060,6 +1118,7 @@ private:
 private:
     VkInstance instance;         ///< Vulkan instance.
     VmaAllocator allocator;      ///< VMA allocator.
+    mutable u32 allocator_frame_index{};
     vk::DeviceDispatch dld;      ///< Device function pointers.
     vk::PhysicalDevice physical; ///< Physical device.
     vk::Device logical;          ///< Logical device.

src/video_core/vulkan_common/vulkan_memory_allocator.cpp

@@ -30,26 +30,6 @@ namespace Vulkan {
 
 // Helpers translating MemoryUsage to flags/usage
 
-        [[maybe_unused]] VkMemoryPropertyFlags MemoryUsagePropertyFlags(MemoryUsage usage) {
-            switch (usage) {
-                case MemoryUsage::DeviceLocal:
-                    return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
-                case MemoryUsage::Upload:
-                    return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-                case MemoryUsage::Download:
-                    return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
-                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
-                case MemoryUsage::Stream:
-                    return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
-                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-            }
-            ASSERT_MSG(false, "Invalid memory usage={}", usage);
-            return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-        }
-
         [[nodiscard]] VkMemoryPropertyFlags MemoryUsagePreferredVmaFlags(MemoryUsage usage) {
             if (usage == MemoryUsage::Download) {
                 return VK_MEMORY_PROPERTY_HOST_CACHED_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
@@ -86,125 +66,11 @@ namespace Vulkan {
         }
 
 
-// This avoids calling vkGetBufferMemoryRequirements* directly.
-        template<typename T>
-        static VkBuffer GetVkHandleFromBuffer(const T &buf) {
-            if constexpr (requires { static_cast<VkBuffer>(buf); }) {
-                return static_cast<VkBuffer>(buf);
-            } else if constexpr (requires {{ buf.GetHandle() } -> std::convertible_to<VkBuffer>; }) {
-                return buf.GetHandle();
-            } else if constexpr (requires {{ buf.Handle() } -> std::convertible_to<VkBuffer>; }) {
-                return buf.Handle();
-            } else if constexpr (requires {{ buf.vk_handle() } -> std::convertible_to<VkBuffer>; }) {
-                return buf.vk_handle();
-            } else {
-                static_assert(sizeof(T) == 0, "Cannot extract VkBuffer handle from vk::Buffer");
-                return VK_NULL_HANDLE;
-            }
-        }
-
     } // namespace
 
-//MemoryCommit is now VMA-backed
-    MemoryCommit::MemoryCommit(VmaAllocator alloc, VmaAllocation a,
-                               const VmaAllocationInfo &info) noexcept
-            : allocator{alloc}, allocation{a}, memory{info.deviceMemory},
-              offset{info.offset}, size{info.size}, mapped_ptr{info.pMappedData} {
-        // Log GPU memory allocation
-        if (GPU::Logging::IsActive() &&
-            Settings::values.gpu_log_memory_tracking.GetValue()) {
-            GPU::Logging::GPULogger::GetInstance().LogMemoryAllocation(
-                reinterpret_cast<uintptr_t>(memory),
-                static_cast<u64>(size),
-                0  // Memory property flags (not easily available from VMA)
-            );
-        }
-    }
-
-    MemoryCommit::~MemoryCommit() { Release(); }
-
-    MemoryCommit::MemoryCommit(MemoryCommit &&rhs) noexcept
-            : allocator{std::exchange(rhs.allocator, nullptr)},
-              allocation{std::exchange(rhs.allocation, nullptr)},
-              memory{std::exchange(rhs.memory, VK_NULL_HANDLE)},
-              offset{std::exchange(rhs.offset, 0)},
-              size{std::exchange(rhs.size, 0)},
-              mapped_ptr{std::exchange(rhs.mapped_ptr, nullptr)} {}
-
-    MemoryCommit &MemoryCommit::operator=(MemoryCommit &&rhs) noexcept {
-        if (this != &rhs) {
-            Release();
-            allocator = std::exchange(rhs.allocator, nullptr);
-            allocation = std::exchange(rhs.allocation, nullptr);
-            memory = std::exchange(rhs.memory, VK_NULL_HANDLE);
-            offset = std::exchange(rhs.offset, 0);
-            size = std::exchange(rhs.size, 0);
-            mapped_ptr = std::exchange(rhs.mapped_ptr, nullptr);
-        }
-        return *this;
-    }
-
-    std::span<u8> MemoryCommit::Map()
-    {
-        if (!allocation) return {};
-        if (!mapped_ptr) {
-            if (vmaMapMemory(allocator, allocation, &mapped_ptr) != VK_SUCCESS) return {};
-        }
-        const size_t n = static_cast<size_t>(std::min<VkDeviceSize>(size,
-                                                                    (std::numeric_limits<size_t>::max)()));
-        return std::span<u8>{static_cast<u8 *>(mapped_ptr), n};
-    }
-
-    std::span<const u8> MemoryCommit::Map() const
-    {
-        if (!allocation) return {};
-        if (!mapped_ptr) {
-            void *p = nullptr;
-            if (vmaMapMemory(allocator, allocation, &p) != VK_SUCCESS) return {};
-            const_cast<MemoryCommit *>(this)->mapped_ptr = p;
-        }
-        const size_t n = static_cast<size_t>(std::min<VkDeviceSize>(size,
-                                                                    (std::numeric_limits<size_t>::max)()));
-        return std::span<const u8>{static_cast<const u8 *>(mapped_ptr), n};
-    }
-
-    void MemoryCommit::Unmap()
-    {
-        if (allocation && mapped_ptr) {
-            vmaUnmapMemory(allocator, allocation);
-            mapped_ptr = nullptr;
-        }
-    }
-
-    void MemoryCommit::Release() {
-        if (allocation && allocator) {
-            // Log GPU memory deallocation
-            if (GPU::Logging::IsActive() &&
-                Settings::values.gpu_log_memory_tracking.GetValue() &&
-                memory != VK_NULL_HANDLE) {
-                GPU::Logging::GPULogger::GetInstance().LogMemoryDeallocation(
-                    reinterpret_cast<uintptr_t>(memory)
-                );
-            }
-
-            if (mapped_ptr) {
-                vmaUnmapMemory(allocator, allocation);
-                mapped_ptr = nullptr;
-            }
-            vmaFreeMemory(allocator, allocation);
-        }
-        allocation = nullptr;
-        allocator = nullptr;
-        memory = VK_NULL_HANDLE;
-        offset = 0;
-        size = 0;
-    }
-
     MemoryAllocator::MemoryAllocator(const Device &device_)
             : device{device_}, allocator{device.GetAllocator()},
-              properties{device_.GetPhysical().GetMemoryProperties().memoryProperties},
-              buffer_image_granularity{
-                      device_.GetPhysical().GetProperties().limits.bufferImageGranularity} {
+              properties{device_.GetPhysical().GetMemoryProperties().memoryProperties} {
 
         // Preserve the previous "RenderDoc small heap" trimming behavior that we had in original vma minus the heap bug
         if (device.HasDebuggingToolAttached())
@@ -224,6 +90,21 @@ namespace Vulkan {
 
     MemoryAllocator::~MemoryAllocator() = default;
 
+    void MemoryAllocator::SetReclaimCallback(ReclaimCallback callback) {
+        reclaim_callback = std::move(callback);
+        vk::SetAllocatorOwnerThread();
+    }
+
+    bool MemoryAllocator::ReclaimAtLeast(u64 hint_bytes) const {
+        if (!reclaim_callback || in_reclaim) {
+            return false;
+        }
+        in_reclaim = true;
+        const u64 freed = reclaim_callback(hint_bytes);
+        in_reclaim = false;
+        return freed > 0;
+    }
+
     vk::Image MemoryAllocator::CreateImage(const VkImageCreateInfo &ci) const
     {
         const VmaAllocationCreateInfo alloc_ci = {
@@ -240,7 +121,26 @@ namespace Vulkan {
         VkImage handle{};
         VmaAllocation allocation{};
         VmaAllocationInfo alloc_info{};
-        vk::Check(vmaCreateImage(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info));
+        DEBUG_ASSERT(vk::OnAllocatorOwnerThread());
+
+        VkResult res = vmaCreateImage(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info);
+
+        if (res != VK_SUCCESS && ReclaimAtLeast(IMAGE_RECLAIM_HINT)) {
+            res = vmaCreateImage(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info);
+        }
+
+        if (res != VK_SUCCESS) {
+            auto relaxed_ci = alloc_ci;
+            relaxed_ci.flags &= ~VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT;
+            res = vmaCreateImage(allocator, &ci, &relaxed_ci, &handle, &allocation, &alloc_info);
+
+            if (res != VK_SUCCESS) {
+                relaxed_ci.preferredFlags &= ~VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+                res = vmaCreateImage(allocator, &ci, &relaxed_ci, &handle, &allocation, &alloc_info);
+            }
+        }
+
+        vk::Check(res);
 
         // Log GPU memory allocation for images
         if (GPU::Logging::IsActive() &&
@@ -277,7 +177,28 @@ namespace Vulkan {
         VmaAllocation allocation{};
         VkMemoryPropertyFlags property_flags{};
 
-        vk::Check(vmaCreateBuffer(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info));
+        DEBUG_ASSERT(vk::OnAllocatorOwnerThread());
+
+        VkResult res = vmaCreateBuffer(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info);
+
+        if (res != VK_SUCCESS && ReclaimAtLeast(ci.size)) {
+            res = vmaCreateBuffer(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info);
+        }
+
+        if (res != VK_SUCCESS) {
+            auto relaxed_ci = alloc_ci;
+            relaxed_ci.flags &= ~VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT;
+            res = vmaCreateBuffer(allocator, &ci, &relaxed_ci, &handle, &allocation, &alloc_info);
+
+            if (res != VK_SUCCESS &&
+                (relaxed_ci.preferredFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT)) {
+                relaxed_ci.preferredFlags &= ~VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+                res = vmaCreateBuffer(allocator, &ci, &relaxed_ci, &handle, &allocation,
+                                      &alloc_info);
+            }
+        }
+
+        vk::Check(res);
         vmaGetAllocationMemoryProperties(allocator, allocation, &property_flags);
 
         // Log GPU memory allocation for buffers
@@ -299,77 +220,4 @@ namespace Vulkan {
                           device.GetDispatchLoader());
     }
 
-    MemoryCommit MemoryAllocator::Commit(const VkMemoryRequirements &reqs, MemoryUsage usage)
-    {
-        const auto vma_usage = MemoryUsageVma(usage);
-        VmaAllocationCreateInfo ci{};
-        ci.flags = VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | MemoryUsageVmaFlags(usage);
-        ci.usage = vma_usage;
-        ci.memoryTypeBits = reqs.memoryTypeBits & valid_memory_types;
-        ci.requiredFlags = 0;
-        ci.preferredFlags = MemoryUsagePreferredVmaFlags(usage);
-
-        VmaAllocation a{};
-        VmaAllocationInfo info{};
-
-        VkResult res = vmaAllocateMemory(allocator, &reqs, &ci, &a, &info);
-
-        if (res != VK_SUCCESS) {
-            // Relax 1: drop budget constraint
-            auto ci2 = ci;
-            ci2.flags &= ~VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT;
-            res = vmaAllocateMemory(allocator, &reqs, &ci2, &a, &info);
-
-            // Relax 2: if we preferred DEVICE_LOCAL, drop that preference
-            if (res != VK_SUCCESS && (ci.preferredFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT)) {
-                auto ci3 = ci2;
-                ci3.preferredFlags &= ~VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
-                res = vmaAllocateMemory(allocator, &reqs, &ci3, &a, &info);
-            }
-        }
-
-        vk::Check(res);
-        return MemoryCommit(allocator, a, info);
-    }
-
-    MemoryCommit MemoryAllocator::Commit(const vk::Buffer &buffer, MemoryUsage usage) {
-        // Allocate memory appropriate for this buffer automatically
-        const auto vma_usage = MemoryUsageVma(usage);
-
-        VmaAllocationCreateInfo ci{};
-        ci.flags = VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | MemoryUsageVmaFlags(usage);
-        ci.usage = vma_usage;
-        ci.requiredFlags = 0;
-        ci.preferredFlags = MemoryUsagePreferredVmaFlags(usage);
-        ci.pool = VK_NULL_HANDLE;
-        ci.pUserData = nullptr;
-        ci.priority = 0.0f;
-
-        const VkBuffer raw = *buffer;
-
-        VmaAllocation a{};
-        VmaAllocationInfo info{};
-
-        // Let VMA infer memory requirements from the buffer
-        VkResult res = vmaAllocateMemoryForBuffer(allocator, raw, &ci, &a, &info);
-
-        if (res != VK_SUCCESS) {
-            auto ci2 = ci;
-            ci2.flags &= ~VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT;
-            res = vmaAllocateMemoryForBuffer(allocator, raw, &ci2, &a, &info);
-
-            if (res != VK_SUCCESS && (ci.preferredFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT)) {
-                auto ci3 = ci2;
-                ci3.preferredFlags &= ~VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
-                res = vmaAllocateMemoryForBuffer(allocator, raw, &ci3, &a, &info);
-            }
-        }
-
-        vk::Check(res);
-        vk::Check(vmaBindBufferMemory2(allocator, a, 0, raw, nullptr));
-        return MemoryCommit(allocator, a, info);
-    }
-
-
-
 } // namespace Vulkan

src/video_core/vulkan_common/vulkan_memory_allocator.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project
@@ -6,6 +6,7 @@
 
 #pragma once
 
+#include <functional>
 #include <memory>
 #include <span>
 #include <vector>
@@ -39,51 +40,6 @@ namespace Vulkan {
         }
     }
 
-/// Ownership handle of a memory commitment (real VMA allocation).
-    class MemoryCommit {
-    public:
-        MemoryCommit() noexcept = default;
-
-        MemoryCommit(VmaAllocator allocator, VmaAllocation allocation,
-                     const VmaAllocationInfo &info) noexcept;
-
-        ~MemoryCommit();
-
-        MemoryCommit(const MemoryCommit &) = delete;
-
-        MemoryCommit &operator=(const MemoryCommit &) = delete;
-
-        MemoryCommit(MemoryCommit &&) noexcept;
-
-        MemoryCommit &operator=(MemoryCommit &&) noexcept;
-
-        [[nodiscard]] std::span<u8> Map();
-
-        [[nodiscard]] std::span<const u8> Map() const;
-
-        void Unmap();
-
-        explicit operator bool() const noexcept { return allocation != nullptr; }
-
-        VkDeviceMemory Memory() const noexcept { return memory; }
-
-        VkDeviceSize Offset() const noexcept { return offset; }
-
-        VkDeviceSize Size() const noexcept { return size; }
-
-        VmaAllocation Allocation() const noexcept { return allocation; }
-
-    private:
-        void Release();
-
-        VmaAllocator allocator{};   ///< VMA allocator
-        VmaAllocation allocation{};  ///< VMA allocation handle
-        VkDeviceMemory memory{};      ///< Underlying VkDeviceMemory chosen by VMA
-        VkDeviceSize offset{};      ///< Offset of this allocation inside VkDeviceMemory
-        VkDeviceSize size{};        ///< Size of the allocation
-        void *mapped_ptr{};  ///< Optional persistent mapped pointer
-    };
-
 /// Memory allocator container.
 /// Allocates and releases memory allocations on demand.
     class MemoryAllocator {
@@ -107,36 +63,21 @@ namespace Vulkan {
 
         vk::Buffer CreateBuffer(const VkBufferCreateInfo &ci, MemoryUsage usage) const;
 
-        /**
-         * Commits a memory with the specified requirements.
-         *
-         * @param requirements Requirements returned from a Vulkan call.
-         * @param usage        Indicates how the memory will be used.
-         *
-         * @returns A memory commit.
-         */
-        MemoryCommit Commit(const VkMemoryRequirements &requirements, MemoryUsage usage);
+        using ReclaimCallback = std::function<u64(u64)>;
 
-        /// Commits memory required by the buffer and binds it (for buffers created outside VMA).
-        MemoryCommit Commit(const vk::Buffer &buffer, MemoryUsage usage);
+        void SetReclaimCallback(ReclaimCallback callback);
 
     private:
-        static bool IsAutoUsage(VmaMemoryUsage u) noexcept {
-            switch (u) {
-                case VMA_MEMORY_USAGE_AUTO:
-                case VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE:
-                case VMA_MEMORY_USAGE_AUTO_PREFER_HOST:
-                    return true;
-                default:
-                    return false;
-            }
-        }
+        bool ReclaimAtLeast(u64 hint_bytes) const;
+
+        static constexpr u64 IMAGE_RECLAIM_HINT = 64ULL * 1024 * 1024;
 
         const Device &device;                              ///< Device handle.
         VmaAllocator allocator;                           ///< VMA allocator.
         const VkPhysicalDeviceMemoryProperties properties; ///< Physical device memory properties.
-        VkDeviceSize buffer_image_granularity;            ///< Adjacent buffer/image granularity
         u32 valid_memory_types{~0u};
+        ReclaimCallback reclaim_callback;
+        mutable bool in_reclaim{false};
     };
 
 } // namespace Vulkan

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -5,11 +5,16 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <algorithm>
+#include <atomic>
+#include <limits>
 #include <memory>
+#include <mutex>
 #include <optional>
+#include <thread>
 #include <utility>
 #include <vector>
 
+#include "common/assert.h"
 #include "common/common_types.h"
 #include "common/logging.h"
 #include "video_core/vulkan_common/vk_enum_string_helper.h"
@@ -20,6 +25,60 @@ namespace Vulkan::vk {
 
 namespace {
 
+std::thread::id allocator_owner_thread;
+
+template <typename HandleType>
+struct PendingRelease {
+    VmaAllocator allocator;
+    HandleType handle;
+    VmaAllocation allocation;
+    u64 timeline;
+};
+
+std::mutex deletion_mutex;
+std::atomic<u64> deletion_timeline{1};
+std::vector<PendingRelease<VkImage>> pending_images;
+std::vector<PendingRelease<VkBuffer>> pending_buffers;
+
+template <typename HandleType>
+void PushPendingRelease(std::vector<PendingRelease<HandleType>>& pending, VmaAllocator allocator,
+                        HandleType handle, VmaAllocation allocation) noexcept {
+    std::scoped_lock lock{deletion_mutex};
+    pending.push_back(PendingRelease<HandleType>{
+        .allocator = allocator,
+        .handle = handle,
+        .allocation = allocation,
+        .timeline = deletion_timeline.load(std::memory_order_acquire),
+    });
+}
+
+template <typename HandleType>
+void ExtractReleased(std::vector<PendingRelease<HandleType>>& pending,
+                     std::vector<PendingRelease<HandleType>>& released, u64 completed_value) {
+    const auto split = std::partition(pending.begin(), pending.end(),
+                                      [completed_value](const PendingRelease<HandleType>& entry) {
+                                          return entry.timeline > completed_value;
+                                      });
+    released.assign(split, pending.end());
+    pending.erase(split, pending.end());
+}
+
+void DrainDeletionQueue(u64 completed_value) noexcept {
+    std::vector<PendingRelease<VkImage>> images;
+    std::vector<PendingRelease<VkBuffer>> buffers;
+    {
+        std::scoped_lock lock{deletion_mutex};
+        ExtractReleased(pending_images, images, completed_value);
+        ExtractReleased(pending_buffers, buffers, completed_value);
+    }
+    for (const auto& entry : images) {
+        vmaDestroyImage(entry.allocator, entry.handle, entry.allocation);
+    }
+    for (const auto& entry : buffers) {
+        vmaDestroyBuffer(entry.allocator, entry.handle, entry.allocation);
+    }
+}
+
 template <typename Func>
 void SortPhysicalDevices(std::vector<VkPhysicalDevice>& devices, const InstanceDispatch& dld,
                          Func&& func) {
@@ -502,13 +561,35 @@ DebugReportCallback Instance::CreateDebugReportCallback(
     return DebugReportCallback(object, handle, *dld);
 }
 
+void SetAllocatorOwnerThread() {
+    allocator_owner_thread = std::this_thread::get_id();
+}
+
+bool OnAllocatorOwnerThread() noexcept {
+    return allocator_owner_thread == std::thread::id{} ||
+           allocator_owner_thread == std::this_thread::get_id();
+}
+
+void SetDeletionTimeline(u64 value) noexcept {
+    deletion_timeline.store(value, std::memory_order_release);
+}
+
+void TickDeletionQueue(u64 completed_value) noexcept {
+    DEBUG_ASSERT(OnAllocatorOwnerThread());
+    DrainDeletionQueue(completed_value);
+}
+
+void FlushDeletionQueue() noexcept {
+    DrainDeletionQueue((std::numeric_limits<u64>::max)());
+}
+
 void Image::SetObjectNameEXT(const char* name) const {
     SetObjectName(dld, owner, handle, VK_OBJECT_TYPE_IMAGE, name);
 }
 
 void Image::Release() const noexcept {
     if (handle) {
-        vmaDestroyImage(allocator, handle, allocation);
+        PushPendingRelease(pending_images, allocator, handle, allocation);
     }
 }
 
@@ -530,7 +611,7 @@ void Buffer::SetObjectNameEXT(const char* name) const {
 
 void Buffer::Release() const noexcept {
     if (handle) {
-        vmaDestroyBuffer(allocator, handle, allocation);
+        PushPendingRelease(pending_buffers, allocator, handle, allocation);
     }
 }
 

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -131,6 +131,16 @@ private:
     VkResult result;
 };
 
+void SetAllocatorOwnerThread();
+
+[[nodiscard]] bool OnAllocatorOwnerThread() noexcept;
+
+void SetDeletionTimeline(u64 value) noexcept;
+
+void TickDeletionQueue(u64 completed_value) noexcept;
+
+void FlushDeletionQueue() noexcept;
+
 /// Throws a Vulkan exception if result is not success.
 inline void Check(VkResult result) {
     if (result != VK_SUCCESS) {