2026-09-25 13:00:48

Signed-off-by: lizzie <lizzie@eden-emu.dev>

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

@@ -30,6 +30,7 @@ enum class BooleanSetting(override val key: String) : AbstractBooleanSetting {
     RENDERER_REACTIVE_FLUSHING("use_reactive_flushing"),
     ENABLE_BUFFER_HISTORY("enable_buffer_history"),
     USE_OPTIMIZED_VERTEX_BUFFERS("use_optimized_vertex_buffers"),
+    ENABLE_GPU_BUFFER_READBACK("enable_gpu_buffer_readback"),
     SYNC_MEMORY_OPERATIONS("sync_memory_operations"),
     BUFFER_REORDER_DISABLE("disable_buffer_reorder"),
     RENDERER_DEBUG("debug"),

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

@@ -908,6 +908,13 @@ abstract class SettingsItem(
                     descriptionId = R.string.enable_buffer_history_description
                 )
             )
+            put(
+                SwitchSetting(
+                    BooleanSetting.ENABLE_GPU_BUFFER_READBACK,
+                    titleId = R.string.enable_gpu_buffer_readback,
+                    descriptionId = R.string.enable_gpu_buffer_readback_description
+                )
+            )
             put(
                 SwitchSetting(
                     BooleanSetting.USE_OPTIMIZED_VERTEX_BUFFERS,

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

@@ -549,6 +549,7 @@ class SettingsFragmentPresenter(
             add(BooleanSetting.RENDERER_FORCE_MAX_CLOCK.key)
             add(BooleanSetting.RENDERER_REACTIVE_FLUSHING.key)
             add(BooleanSetting.ENABLE_BUFFER_HISTORY.key)
+            add(BooleanSetting.ENABLE_GPU_BUFFER_READBACK.key)
             add(BooleanSetting.USE_OPTIMIZED_VERTEX_BUFFERS.key)
 
             add(HeaderSetting(R.string.hacks))

src/android/app/src/main/res/values-ar/strings.xml

@@ -580,6 +580,8 @@
     <string name="renderer_reactive_flushing_description">يحسن دقة العرض في بعض الألعاب على حساب الأداء.</string>
     <string name="enable_buffer_history">تمكين سجل التخزين المؤقت</string>
     <string name="enable_buffer_history_description">يُتيح هذا الخيار الوصول إلى حالات التخزين المؤقت السابقة. وقد يُحسّن جودة العرض وثبات الأداء في بعض الألعاب.</string>
+    <string name="enable_gpu_buffer_readback">تفعيل قراءة مخزن وحدة معالجة الرسومات</string>
+    <string name="enable_gpu_buffer_readback_description">يحافظ هذا النظام على بيانات المخزن المؤقت المُعدّلة بواسطة وحدة معالجة الرسومات عن طريق قراءتها مرة أخرى قبل التحميل. تتطلب بعض الألعاب ذلك لعرض بعض التأثيرات بشكل صحيح. قد يُسبب ذلك مشاكل إذا لم يتمكن الجهاز من التعامل مع عبء العمل الإضافي.</string>
     <string name="use_optimized_vertex_buffers">مخازن الرؤوس المُحسّنة</string>
     <string name="use_optimized_vertex_buffers_description">يُتيح ربطًا مُحسَّنًا لمخازن الرؤوس لتحسين الأداء. يتطلب برامج تشغيل Mesa 26.0+ Turnip/ برامج تشغيل QCOM. قد يتعطل على برامج تشغيل Turnip القديمة (25.3 وما دون).</string>
 
@@ -1097,6 +1099,7 @@
     <string name="gpu_fence_behavior_immediate">فوري</string>
     <string name="gpu_fence_behavior_balanced">متوازن</string>
     <string name="gpu_fence_behavior_accurate">دقيق</string>
+    <string name="gpu_fence_behavior_strict">صارم</string>
 
     <string name="vram_usage_conservative">محافظ</string>
     <string name="vram_usage_aggressive">عدواني</string>

src/android/app/src/main/res/values-de/strings.xml

@@ -967,6 +967,7 @@ Wirklich fortfahren?</string>
     <string name="gpu_fence_behavior_immediate">Direkt</string>
     <string name="gpu_fence_behavior_balanced">Ausgewogen</string>
     <string name="gpu_fence_behavior_accurate">Genau</string>
+    <string name="gpu_fence_behavior_strict">Strikt</string>
 
     <string name="vram_usage_conservative">Konservativ</string>
     <string name="vram_usage_aggressive">Aggressiv</string>

src/android/app/src/main/res/values-es/strings.xml

@@ -524,6 +524,8 @@
     <string name="renderer_reactive_flushing_description">Mejora la precisión de renderizado en algunos juegos, pero reduce el rendimiento.</string>
     <string name="enable_buffer_history">Activar el historial del búfer</string>
     <string name="enable_buffer_history_description">Permite el acceso al estado del búfer anterior. Esta opción puede mejorar la calidad de renderizado y la consistencia en el rendimiento de algunos juegos.</string>
+    <string name="enable_gpu_buffer_readback">Activar la lectura del buffer de la GPU</string>
+    <string name="enable_gpu_buffer_readback_description">Conserva los datos del búfer modificados por la GPU leyéndolos antes de subirlos.\nAlgunos juegos requieren esto para renderizar correctamente ciertos efectos.\nPuede causar problemas si el hardware no puede soportar la carga de trabajo adicional.</string>
     <string name="use_optimized_vertex_buffers">Búferes de vértices optimizados</string>
     <string name="use_optimized_vertex_buffers_description">Permite la optimización del enlace del búfer de vértices para un mejor rendimiento. Requiere controladores Mesa 26.0+ Turnip/ controladores QCOM. Fallará con controladores Turnip más antiguos (versión 25.3 o inferior).</string>
 
@@ -1036,6 +1038,7 @@
     <string name="gpu_fence_behavior_immediate">Inmediato</string>
     <string name="gpu_fence_behavior_balanced">Equilibrado</string>
     <string name="gpu_fence_behavior_accurate">Preciso</string>
+    <string name="gpu_fence_behavior_strict">Estricto</string>
 
     <string name="vram_usage_conservative">Conservador</string>
     <string name="vram_usage_aggressive">Agresivo</string>

src/android/app/src/main/res/values-ru/strings.xml

@@ -569,6 +569,8 @@
     <string name="renderer_reactive_flushing_description">Повышение точности рендеринга в некоторых играх за счет снижения производительности.</string>
     <string name="enable_buffer_history">Включить историю буфера</string>
     <string name="enable_buffer_history_description">Позволяет обращаться к предыдущим состояниям буфера. Эта опция может повысить качество рендеринга и стабильность производительности в некоторых играх.</string>
+    <string name="enable_gpu_buffer_readback">Включить обратное чтение буфера ГПУ</string>
+    <string name="enable_gpu_buffer_readback_description">Сохраняет измененные ГПУ данные буфера путем чтения их обратно перед выгрузками. Некоторые игры требуют этого, чтобы рендерить определенные эффекты правильно. Может вызывать проблемы если оборудование не может обработать дополнительную рабочую нагрузку.</string>
     <string name="use_optimized_vertex_buffers">Оптимизированные вершинные буферы</string>
     <string name="use_optimized_vertex_buffers_description">Включает оптимизированную привязку вершинного буфера для повышения производительности. Требует Mesa Turnip 26.0+ / QCOM. Приводит к вылету на старых версиях драйверов Turnip (25.3 и ниже).</string>
 
@@ -1086,6 +1088,7 @@
     <string name="gpu_fence_behavior_immediate">Мгновенный</string>
     <string name="gpu_fence_behavior_balanced">Сбалансированный</string>
     <string name="gpu_fence_behavior_accurate">Точный</string>
+    <string name="gpu_fence_behavior_strict">Строгий</string>
 
     <string name="vram_usage_conservative">Консервативный</string>
     <string name="vram_usage_aggressive">Агрессивный</string>

src/android/app/src/main/res/values-zh-rCN/strings.xml

@@ -570,6 +570,8 @@
     <string name="renderer_reactive_flushing_description">通过牺牲性能来提升某些游戏的渲染精度。</string>
     <string name="enable_buffer_history">启用缓冲区历史</string>
     <string name="enable_buffer_history_description">启用对先前缓冲区状态的访问。此选项可在某些游戏中提升渲染质量并保持性能的一致性。</string>
+    <string name="enable_gpu_buffer_readback">启用 GPU 缓冲区回读</string>
+    <string name="enable_gpu_buffer_readback_description">在上传前回读经由 GPU 修改过的缓冲区数据，以将其保留。一些游戏会用到这项设定以正确渲染某些效果。如果硬件无法处理额外的工作负载，则可能会导致问题。</string>
     <string name="use_optimized_vertex_buffers">优化顶点缓冲区</string>
     <string name="use_optimized_vertex_buffers_description">启用经过优化的顶点缓冲区绑定以提升性能。需要 Mesa 26.0 及以上版本的 Turnip 或 QCOM 驱动程序。若使用较旧版本的 Turnip 驱动 (25.3 及以下版本) 则会导致崩溃。</string>
 
@@ -1087,6 +1089,7 @@
     <string name="gpu_fence_behavior_immediate">即时</string>
     <string name="gpu_fence_behavior_balanced">均衡</string>
     <string name="gpu_fence_behavior_accurate">精确</string>
+    <string name="gpu_fence_behavior_strict">严格</string>
 
     <string name="vram_usage_conservative">保守式</string>
     <string name="vram_usage_aggressive">主动式</string>

src/android/app/src/main/res/values-zh-rTW/strings.xml

@@ -561,6 +561,8 @@
     <string name="renderer_reactive_flushing_description">犧牲效能，以改善部分遊戲的轉譯準確度</string>
     <string name="enable_buffer_history">啟用緩衝區歷史</string>
     <string name="enable_buffer_history_description">允許存取先前的緩衝區狀態。此選項可能會改善部分遊戲的渲染品質與效能穩定性</string>
+    <string name="enable_gpu_buffer_readback">啟用 GPU 緩衝區讀回</string>
+    <string name="enable_gpu_buffer_readback_description">透過在上傳之前先將 GPU 修改過的緩衝區資料讀回來保存資料，部分遊戲需要啟用此功能才能正常渲染遊戲特效。如果硬體無法負荷可能會導致錯誤</string>
     <string name="use_optimized_vertex_buffers">最佳化頂點緩衝區</string>
     <string name="use_optimized_vertex_buffers_description">啟用最佳化的頂點緩衝區綁定。需要安裝 Mesa 26.0+ Turnip drivers/Qualcomm drivers。使用舊版 Turnip drivers 會導致當機 (25.3版和更低的版本)</string>
 

src/android/app/src/main/res/values/arrays.xml

@@ -558,6 +558,7 @@
         <item>@string/gpu_fence_behavior_immediate</item>
         <item>@string/gpu_fence_behavior_balanced</item>
         <item>@string/gpu_fence_behavior_accurate</item>
+        <item>@string/gpu_fence_behavior_strict</item>
     </string-array>
     <integer-array name="gpuFenceBehaviorValues">
         <item>0</item>

src/android/app/src/main/res/values/strings.xml

@@ -586,6 +586,8 @@
     <string name="renderer_reactive_flushing_description">Improves rendering accuracy in some games at the cost of performance.</string>
     <string name="enable_buffer_history">Enable buffer history</string>
     <string name="enable_buffer_history_description">Enables access to previous buffer states. This option may improve rendering quality and performance consistency in some games.</string>
+    <string name="enable_gpu_buffer_readback">Enable GPU Buffer Readback</string>
+    <string name="enable_gpu_buffer_readback_description">Preserves GPU-modified buffer data by reading it back before uploads. Some games require this to render certain effects properly. May cause issues if the hardware cannot handle the additional workload.</string>
     <string name="use_optimized_vertex_buffers">Optimized Vertex Buffers</string>
     <string name="use_optimized_vertex_buffers_description">Enables optimized vertex buffer binding for improved performance. Requires Mesa 26.0+ Turnip drivers/ QCOM drivers. Will crash on older Turnip drivers (25.3 and below).</string>
 
@@ -1136,6 +1138,7 @@
     <string name="gpu_fence_behavior_immediate">Immediate</string>
     <string name="gpu_fence_behavior_balanced">Balanced</string>
     <string name="gpu_fence_behavior_accurate">Accurate</string>
+    <string name="gpu_fence_behavior_strict">Strict</string>
 
     <!-- ASTC Decoding Method Choices -->
     <string name="accelerate_astc_cpu" translatable="false">CPU</string>

src/common/CMakeLists.txt

@@ -76,7 +76,6 @@ add_library(
     logging.h
     lz4_compression.cpp
     lz4_compression.h
-    make_unique_for_overwrite.h
     math_util.h
     memory_detect.cpp
     memory_detect.h

src/common/make_unique_for_overwrite.h

@@ -1,27 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
-
-#pragma once
-
-#include <memory>
-#include <type_traits>
-
-namespace Common {
-
-template <class T>
-    requires(!std::is_array_v<T>)
-std::unique_ptr<T> make_unique_for_overwrite() {
-    return std::unique_ptr<T>(new T);
-}
-
-template <class T>
-    requires std::is_unbounded_array_v<T>
-std::unique_ptr<T> make_unique_for_overwrite(std::size_t n) {
-    return std::unique_ptr<T>(new std::remove_extent_t<T>[n]);
-}
-
-template <class T, class... Args>
-    requires std::is_bounded_array_v<T>
-void make_unique_for_overwrite(Args&&...) = delete;
-
-} // namespace Common

src/common/scratch_buffer.h

@@ -8,8 +8,7 @@
 
 #include <iterator>
 #include <cstring>
-
-#include "common/make_unique_for_overwrite.h"
+#include <memory>
 
 namespace Common {
 
@@ -38,8 +37,9 @@ public:
     ScratchBuffer() = default;
 
     explicit ScratchBuffer(size_type initial_capacity)
-        : last_requested_size{initial_capacity}, buffer_capacity{initial_capacity},
-          buffer{Common::make_unique_for_overwrite<T[]>(initial_capacity)} {}
+        : last_requested_size{initial_capacity}
+        , buffer_capacity{initial_capacity}
+        , buffer{std::make_unique_for_overwrite<T[]>(initial_capacity)} {}
 
     ~ScratchBuffer() = default;
     ScratchBuffer(const ScratchBuffer&) = delete;
@@ -64,7 +64,7 @@ public:
     /// The previously held data will remain intact.
     void resize(size_type size) {
         if (size > buffer_capacity) {
-            auto new_buffer = Common::make_unique_for_overwrite<T[]>(size);
+            auto new_buffer = std::make_unique_for_overwrite<T[]>(size);
             std::memcpy(new_buffer.get(), buffer.get(), buffer_capacity * sizeof(T));
             buffer = std::move(new_buffer);
             buffer_capacity = size;
@@ -77,7 +77,7 @@ public:
     void resize_destructive(size_type size) {
         if (size > buffer_capacity) {
             buffer_capacity = size;
-            buffer = Common::make_unique_for_overwrite<T[]>(buffer_capacity);
+            buffer = std::make_unique_for_overwrite<T[]>(buffer_capacity);
         }
         last_requested_size = size;
     }

src/common/settings.cpp

@@ -178,6 +178,10 @@ bool IsGPUFenceBehaviorAccurate() {
     return values.gpu_fence_behavior.GetValue() == GpuFenceBehavior::Accurate;
 }
 
+bool IsGPUFenceBehaviorStrict() {
+    return values.gpu_fence_behavior.GetValue() == GpuFenceBehavior::Strict;
+}
+
 bool IsFastmemEnabled() {
     if (values.cpu_accuracy.GetValue() == Settings::CpuAccuracy::Debugging)
         return bool(values.cpuopt_fastmem);

src/common/settings.h

@@ -525,7 +525,7 @@ struct Values {
     SwitchableSetting<GpuFenceBehavior, true> gpu_fence_behavior{linkage,
                                                                  GpuFenceBehavior::Default,
                                                                  GpuFenceBehavior::Default,
-                                                                 GpuFenceBehavior::Accurate,
+                                                                 GpuFenceBehavior::Strict,
                                                                  "gpu_fence_behavior",
                                                                  Category::RendererAdvanced,
                                                                  Specialization::Default,
@@ -655,6 +655,13 @@ struct Values {
 
     SwitchableSetting<bool> rescale_hack{linkage, false, "rescale_hack",
                                          Category::RendererHacks};
+    SwitchableSetting<bool> enable_gpu_buffer_readback{linkage,
+                                                       false,
+                                                       "enable_gpu_buffer_readback",
+                                                       Category::RendererAdvanced,
+                                                       Specialization::Default,
+                                                       true,
+                                                       true};
 
     SwitchableSetting<bool> use_asynchronous_shaders{linkage, false, "use_asynchronous_shaders",
                                                      Category::RendererHacks};
@@ -972,6 +979,7 @@ bool IsDMALevelSafe();
 bool IsGPUFenceBehaviorDefault();
 bool IsGPUFenceBehaviorBalanced();
 bool IsGPUFenceBehaviorAccurate();
+bool IsGPUFenceBehaviorStrict();
 
 bool IsFastmemEnabled();
 void SetNceEnabled(bool is_64bit);

src/common/settings_enums.h

@@ -137,7 +137,7 @@ ENUM(VramUsageMode, Conservative, Aggressive);
 ENUM(RendererBackend, OpenGL_GLSL, Vulkan, Null, OpenGL_GLASM, OpenGL_SPIRV);
 ENUM(GpuAccuracy, Low, High);
 ENUM(DmaAccuracy, Default, Unsafe, Safe);
-ENUM(GpuFenceBehavior, Default, Immediate, Balanced, Accurate);
+ENUM(GpuFenceBehavior, Default, Immediate, Balanced, Accurate, Strict);
 ENUM(CpuBackend, Dynarmic, Nce);
 ENUM(CpuAccuracy, Auto, Accurate, Unsafe, Paranoid, Debugging);
 ENUM(CpuClock, Normal, Boost, Overclock)

src/common/sparse_large_vector.cpp

@@ -8,7 +8,14 @@
 #ifdef _WIN32
 #include <windows.h>
 #include <mutex>
-#else
+#include <algorithm>
+#include <vector>
+#endif
+
+#include <cerrno>
+#include <cstring>
+
+#ifndef _WIN32
 #include <sys/mman.h>
 #endif
 
@@ -19,54 +26,60 @@
 namespace Common {
 
 #ifdef _WIN32
-static std::vector<std::pair<u64, u64>> vector_regions {};
 
-// Workaround for handling non-commited memory accessed by Dynarmic; usually result of an error
+struct VectorRegion {
+    u64 start_page;
+    u64 end_page;
+};
+
+static std::mutex& GetVectorRegionsMutex() {
+    static std::mutex* m = new std::mutex();
+    return *m;
+}
+
+static std::vector<VectorRegion>& GetVectorRegions() {
+    static std::vector<VectorRegion>* v = new std::vector<VectorRegion>();
+    return *v;
+}
+
 static LONG WINAPI FakePageFaultHandler(PEXCEPTION_POINTERS info) {
-    DWORD code = info->ExceptionRecord->ExceptionCode;
-    u64 exception_addr = reinterpret_cast<u64>(info->ExceptionRecord->ExceptionAddress);
-
-    if (code != EXCEPTION_ACCESS_VIOLATION) {
-        // Not our problem
+    if (info->ExceptionRecord->ExceptionCode != EXCEPTION_ACCESS_VIOLATION) {
         return EXCEPTION_CONTINUE_SEARCH;
     }
 
-    u64 addr = 0, addr2 = 0;
+    const u64 fault_addr = info->ExceptionRecord->ExceptionInformation[1];
+    const u64 access_type = info->ExceptionRecord->ExceptionInformation[0];
+    const bool is_write = (access_type == 1);
+    const u64 fault_page = fault_addr >> HostPageBits;
 
-    for (auto region: vector_regions) {
-        auto addr_shifted = exception_addr >> HostPageBits;
-        if (region.first <= addr_shifted && addr_shifted <= region.second) {
-            addr = addr_shifted;
-        }
+    u64 addr = 0;
+    u64 addr2 = 0;
 
-        // Page-boundary accesses
-        if (auto addr_ = (exception_addr + 0x40) >> HostPageBits; addr_ != addr_shifted && region.first <= addr_ && addr_ <= region.second) {
-            addr2 = addr_;
-        }
-
-        if (addr != 0 || addr2 != 0) {
-            break;
+    {
+        std::lock_guard lock(GetVectorRegionsMutex());
+        for (const auto& region : GetVectorRegions()) {
+            if (fault_page >= region.start_page && fault_page < region.end_page) {
+                addr = fault_page;
+            }
+            const u64 page2 = (fault_addr + 0x3F) >> HostPageBits;
+            if (page2 != fault_page && page2 >= region.start_page && page2 < region.end_page) {
+                addr2 = page2;
+            }
+            if (addr != 0 || addr2 != 0) break;
         }
     }
 
     if (addr == 0 && addr2 == 0) {
-        // Not our problem
         return EXCEPTION_CONTINUE_SEARCH;
     }
 
-    LOG_ERROR(HW_Memory, "Accessing an unallocated region of a SparseLargeVector at {:#x}; this shouldn't happen and is likely a Dynarmic error!", exception_addr);
+    LOG_ERROR(HW_Memory, "Accessing an unallocated region of a SparseLargeVector at {:#x}; this shouldn't happen and is likely a Dynarmic error!", fault_addr);
 
-    // Commit this region
-    if (addr != 0) {
-        if (!CommitVectorPage(addr << HostPageBits, false)) {
-            return EXCEPTION_CONTINUE_SEARCH;
-        }
+    if (addr != 0 && !CommitVectorPage(addr << HostPageBits, is_write)) {
+        return EXCEPTION_CONTINUE_SEARCH;
     }
-    // Commit next region if needed
-    if (addr2 != 0) {
-        if (!CommitVectorPage(addr2 << HostPageBits, false)) {
-            return EXCEPTION_CONTINUE_SEARCH;
-        }
+    if (addr2 != 0 && !CommitVectorPage(addr2 << HostPageBits, is_write)) {
+        return EXCEPTION_CONTINUE_SEARCH;
     }
 
     return EXCEPTION_CONTINUE_EXECUTION;
@@ -74,23 +87,31 @@ static LONG WINAPI FakePageFaultHandler(PEXCEPTION_POINTERS info) {
 
 bool CommitVectorPage(uintptr_t addr, bool write) noexcept {
     MEMORY_BASIC_INFORMATION info {};
-    auto res = VirtualQuery(reinterpret_cast<void*>(addr), &info, sizeof(info));
+    const auto res = VirtualQuery(reinterpret_cast<void*>(addr), &info, sizeof(info));
+    const DWORD perm = write ? PAGE_READWRITE : PAGE_READONLY;
+
     if (res == 0) {
         LOG_CRITICAL(HW_Memory, "Failed to query large buffer region at {:#x} with error {}, will try committing anyway", addr, GetLastError());
+    } else if (info.State == MEM_COMMIT) {
+        DWORD old_protect {};
+        if (!VirtualProtect(reinterpret_cast<void*>(addr), HostPageSize, perm, &old_protect)) {
+            LOG_ERROR(HW_Memory, "VirtualProtect failed at {:#x}, error {}", addr, GetLastError());
+            return false;
+        }
+        return true;
     } else if (info.State != MEM_RESERVE) {
-        LOG_ERROR(HW_Memory, "Tried to commit an unreserved large buffer region at {:#x} that is not mapped or is already committed (state {:#x})", addr, info.State);
+        LOG_ERROR(HW_Memory, "Tried to commit an unreserved large buffer region at {:#x} (state {:#x})", addr, info.State);
         return false;
     }
 
-    auto perm = write ? PAGE_READWRITE : PAGE_READONLY;
-    void* res2 = VirtualAlloc(reinterpret_cast<LPVOID>(addr), HostPageSize, MEM_COMMIT, perm);
-    if (res2 == nullptr) {
+    if (VirtualAlloc(reinterpret_cast<LPVOID>(addr), HostPageSize, MEM_COMMIT, perm) == nullptr) {
         LOG_ERROR(HW_Memory, "Failed to commit large buffer region at {:#x}, error {}", addr, GetLastError());
         return false;
     }
 
     return true;
 }
+
 #endif
 
 #ifndef MAP_NOCORE
@@ -102,57 +123,103 @@ bool CommitVectorPage(uintptr_t addr, bool write) noexcept {
 
 void DecommitVectorPage(uintptr_t base) noexcept {
 #if defined(_WIN32)
-    VirtualFree(reinterpret_cast<LPVOID>(base), HostPageSize, MEM_DECOMMIT);
+    if (!VirtualFree(reinterpret_cast<LPVOID>(base), HostPageSize, MEM_DECOMMIT)) {
+        LOG_WARNING(HW_Memory, "VirtualFree(MEM_DECOMMIT) failed at {:#x}, error {}", base, GetLastError());
+    }
 #elif defined(__linux__)
-    // Linux's MADV_DONTNEED zeros out pages for us
-    madvise(reinterpret_cast<void*>(base), HostPageSize, MADV_DONTNEED);
+    if (madvise(reinterpret_cast<void*>(base), HostPageSize, MADV_DONTNEED) != 0) {
+        LOG_WARNING(HW_Memory, "madvise(MADV_DONTNEED) failed at {:#x}: {}", base, std::strerror(errno));
+    }
 #else
-    madvise(reinterpret_cast<void*>(base), HostPageSize, MADV_FREE);
+    if (madvise(reinterpret_cast<void*>(base), HostPageSize, MADV_FREE) != 0) {
+        LOG_WARNING(HW_Memory, "madvise(MADV_FREE) failed at {:#x}: {}", base, std::strerror(errno));
+    }
     std::memset(reinterpret_cast<void*>(base), 0, HostPageSize);
 #endif
 }
 
 void* AllocateMemoryPages(std::size_t size) noexcept {
-    if (auto page = HostPageSize; size % page != 0) {
+    if (size == 0) {
+        return nullptr;
+    }
+
+    const auto page = HostPageSize;
+    if (size % page != 0) {
         LOG_WARNING(HW_Memory, "Allocating unaligned large vector with size {:#x}; aligning to {} page size", size, page);
+        if (size > SIZE_MAX - (page - 1)) {
+            LOG_CRITICAL(HW_Memory, "Size {:#x} would overflow page alignment", size);
+            return nullptr;
+        }
         size = AlignUp(size, page);
     }
 
 #ifdef _WIN32
-    // We will never use this memory entirely so instead of committing it up front let's just reserve it and commit each page individually
     void* base = VirtualAlloc(nullptr, size, MEM_RESERVE, PAGE_READWRITE);
 
     if (base != nullptr) {
-        vector_regions.emplace_back(reinterpret_cast<u64>(base), reinterpret_cast<u64>(base) + size);
+        {
+            std::lock_guard lock(GetVectorRegionsMutex());
+            GetVectorRegions().push_back({
+                reinterpret_cast<u64>(base) >> HostPageBits,
+                (reinterpret_cast<u64>(base) + size) >> HostPageBits,
+            });
+        }
 
         static std::once_flag flag;
         std::call_once(flag, []() { AddVectoredExceptionHandler(1, FakePageFaultHandler); });
     } else {
-        // Try committing everything instead??
         LOG_WARNING(HW_Memory, "Failed to reserve large vector region with error {}, trying to commit instead..", GetLastError());
         base = VirtualAlloc(nullptr, size, MEM_COMMIT, PAGE_READWRITE);
     }
     ASSERT_MSG(base, "Failed to reserve {:#x} sized region with error {}", size, GetLastError());
 #else
-    void* base = mmap(nullptr, size, PROT_READ, MAP_ANON | MAP_PRIVATE | MAP_NOCORE, -1, 0);
-    if (base == MAP_FAILED)
-        base = nullptr;
-    ASSERT_MSG(base, "Failed to allocate {:#x} sized region with error {}", size, strerror(errno));
+    int flags = MAP_ANON | MAP_PRIVATE;
+#ifdef MAP_NORESERVE
+    flags |= MAP_NORESERVE;
 #endif
+#if defined(MAP_NOCORE)
+    flags |= MAP_NOCORE;
+#endif
+    void* base = mmap(nullptr, size, PROT_READ, flags, -1, 0);
+    if (base == MAP_FAILED) {
+        base = nullptr;
+    }
+#ifdef MADV_HUGEPAGE
+    if (base != nullptr) {
+        madvise(base, size, MADV_HUGEPAGE);
+    }
+#endif
+    ASSERT_MSG(base, "Failed to allocate {:#x} sized region with error {}", size, std::strerror(errno));
+#endif
+
     return base;
 }
 
 void FreeMemoryPages(void* base, [[maybe_unused]] std::size_t size) noexcept {
-    if (auto page = HostPageSize; size % page != 0) {
+    if (base == nullptr) {
+        return;
+    }
+
+    if (const auto page = HostPageSize; size % page != 0) {
         size = AlignUp(size, page);
     }
-    if (!base)
-        return;
+
 #ifdef _WIN32
-    ASSERT(VirtualFree(base, 0, MEM_RELEASE));
+    {
+        std::lock_guard lock(GetVectorRegionsMutex());
+        auto& regions = GetVectorRegions();
+        const u64 base_page = reinterpret_cast<u64>(base) >> HostPageBits;
+        regions.erase(std::remove_if(regions.begin(), regions.end(),
+            [base_page](const VectorRegion& r) { return r.start_page == base_page; }), regions.end());
+    }
+    if (!VirtualFree(base, 0, MEM_RELEASE)) {
+        LOG_ERROR(HW_Memory, "VirtualFree failed, error {}", GetLastError());
+    }
 #else
-    ASSERT(munmap(base, size) == 0);
+    if (munmap(base, size) != 0) {
+        LOG_ERROR(HW_Memory, "munmap failed: {}", std::strerror(errno));
+    }
 #endif
 }
 
-} // namespace Common
+} // namespace Common

src/common/sparse_large_vector.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 /* virtual_buffer.h */
@@ -7,10 +7,14 @@
 
 #pragma once
 
+#include <array>
 #include <atomic>
 #include <bit>
-#include <utility>
-#include <vector>
+#include <cerrno>
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+#include <type_traits>
 
 #ifndef _WIN32
 #include <unistd.h>
@@ -28,9 +32,9 @@ constexpr u64 HostPageBits = 12;
 constexpr u64 HostPageMask = ~(HostPageSize - 1);
 bool CommitVectorPage(uintptr_t addr, bool write) noexcept;
 #else
-const u64 HostPageSize = sysconf(_SC_PAGESIZE);
-const u64 HostPageBits = std::countr_zero(HostPageSize);
-const u64 HostPageMask = ~(HostPageSize - 1);
+inline const u64 HostPageSize = static_cast<u64>(sysconf(_SC_PAGESIZE));
+inline const u64 HostPageBits = std::countr_zero(HostPageSize);
+inline const u64 HostPageMask = ~(HostPageSize - 1);
 #endif
 
 void* AllocateMemoryPages(std::size_t size) noexcept;
@@ -43,20 +47,18 @@ template <typename T>
     // requires std::is_trivially_copyable_v<T>
 class SparseLargeVector final {
 public:
-    constexpr SparseLargeVector() = default;
+    SparseLargeVector() = default;
 
-    explicit SparseLargeVector(std::size_t count) noexcept
-        : alloc_size{count * sizeof(T)}
-    {
-        base_ptr = static_cast<T*>(AllocateMemoryPages(alloc_size));
-
-        // each item in vector holds information for 64 pages
-        auto denom = HostPageSize * 64;
-        committed_pages = std::vector<std::atomic<u64>>((alloc_size + denom - 1) / denom);
+    explicit SparseLargeVector(std::size_t count) noexcept {
+        if (count > SIZE_MAX / sizeof(T)) {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector size overflow: {} elements", count);
+            return;
+        }
+        Allocate(count * sizeof(T));
     }
 
     ~SparseLargeVector() noexcept {
-        FreeMemoryPages(base_ptr, alloc_size);
+        Release();
     }
 
     SparseLargeVector(const SparseLargeVector&) = delete;
@@ -65,145 +67,181 @@ public:
     SparseLargeVector& operator=(SparseLargeVector&& other) = delete;
 
     void ResizeAndClear(std::size_t count) noexcept {
-        if (auto const new_size = count * sizeof(T); new_size != alloc_size) {
-            FreeMemoryPages(base_ptr, alloc_size);
-            alloc_size = new_size;
-            base_ptr = static_cast<T*>(AllocateMemoryPages(alloc_size));
-
-            auto denom = HostPageSize * 64;
-            committed_pages = std::vector<std::atomic<u64>>((alloc_size + denom - 1) / denom);
+        if (count > SIZE_MAX / sizeof(T)) {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector resize overflow: {} elements", count);
+            return;
         }
+        const std::size_t new_size = count * sizeof(T);
+        if (new_size == alloc_size) {
+            ZeroRegion(0, alloc_size / sizeof(T));
+            return;
+        }
+        Release();
+        Allocate(new_size);
     }
 
-    /// Returns a reference to the value of the requested index and allocates memory if needed.
     T& GetAndFault(std::size_t index) noexcept {
-        if (index > alloc_size / sizeof(T)) {
-            UNREACHABLE_MSG("Out of bounds RW access on SparseLargeVector @ {}", index);
+        if (base_ptr == nullptr || index >= size()) [[unlikely]] {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector RW access out of bounds @ {} (size {})", index, size());
+            std::abort();
         }
-
-        if (!IsCommittedPage(index)) {
-            CommitPage(index);
+        const u64 byte_offset = static_cast<u64>(index) * sizeof(T);
+        if (!CommitPage(byte_offset)) [[unlikely]] {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector commit failed @ {} (offset {:#x})", index, byte_offset);
+            std::abort();
         }
         return base_ptr[index];
     }
 
-    /// Returns a reference to the value of the requested index if initialized, or will otherwise return a zero-initialized object.
-    const T& GetOrDefault(std::size_t index) const {
+    const T& GetOrDefault(std::size_t index) const noexcept {
+        if (base_ptr == nullptr || index >= size()) [[unlikely]] {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector RO access out of bounds @ {}", index);
+            return DefaultValue();
+        }
 #ifdef _WIN32
-        if (!IsCommittedPage(index)) {
-            return *reinterpret_cast<const T*>(&default_val);
+        if (!IsPageCommitted(static_cast<u64>(index) * sizeof(T))) {
+            return DefaultValue();
         }
 #endif
-        // On non-Windows, OS page table should optimize this by pointing to a zero page if unallocated.
         return base_ptr[index];
     }
 
     void Set(std::size_t index, const T& value) noexcept {
-        if (index > alloc_size / sizeof(T)) {
-            LOG_CRITICAL(Common_Memory, "Out of bounds write on SparseLargeVector @ {}", index);
+        if (base_ptr == nullptr || index >= size()) [[unlikely]] {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector write out of bounds @ {}", index);
+            return;
+        }
+        const u64 byte_offset = static_cast<u64>(index) * sizeof(T);
+        if (!CommitPage(byte_offset)) [[unlikely]] {
+            LOG_CRITICAL(Common_Memory, "SparseLargeVector commit failed for write @ {}", index);
             return;
         }
-        if (!IsCommittedPage(index))
-            CommitPage(index);
         base_ptr[index] = value;
     }
 
     void ZeroRegion(std::size_t start, std::size_t end_) noexcept {
-        u64 base = reinterpret_cast<u64>(&base_ptr[start]);
-        const u64 end = reinterpret_cast<u64>(&base_ptr[end_]);
+        if (base_ptr == nullptr || start >= end_) return;
 
-        const u64 end_page = AlignUp(base, HostPageSize);
-        const u64 first_size = (std::min)(end_page, end) - base;
+        const u64 start_off = static_cast<u64>(start) * sizeof(T);
+        const u64 end_off = static_cast<u64>(end_) * sizeof(T);
+        const u64 first_page_end = (start_off + HostPageSize - 1) & HostPageMask;
 
-        if (IsCommittedPage(start)) {
-            std::memset(reinterpret_cast<void*>(base), 0, first_size);
+        if (start_off < first_page_end) {
+            const u64 chunk_end = (std::min)(first_page_end, end_off);
+            const u64 chunk_size = chunk_end - start_off;
+            if (chunk_size != 0 && IsPageCommitted(start_off)) {
+                std::memset(reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(base_ptr) + start_off), 0, chunk_size);
+            }
+            if (end_off <= first_page_end) return;
         }
 
-        if (end <= end_page)
-            return;
-
-        base = end_page;
-
-        for (u64 page = base; page < end; page += HostPageSize) {
-            auto index = (page - reinterpret_cast<u64>(base_ptr)) / sizeof(T);
-            if (!IsCommittedPage(index)) {
-                continue;
-            }
-
-            if (end - page >= HostPageSize) {
-                DecommitPage(index);
+        for (u64 off = first_page_end; off < end_off; off += HostPageSize) {
+            if (!IsPageCommitted(off)) continue;
+            const u64 remaining = end_off - off;
+            if (remaining >= HostPageSize) {
+                DecommitPage(off);
             } else {
-                std::memset(reinterpret_cast<void*>(page), 0, end - page);
+                std::memset(reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(base_ptr) + off), 0, remaining);
             }
         }
     }
 
-    constexpr void CommitRegion(size_t index, size_t end_) {
-        const u64 base = static_cast<u64>(index) * sizeof(T);
-        const u64 end = static_cast<u64>(end_) * sizeof(T);
-
-        for (u64 page = AlignDown(base, HostPageSize); page < end; page += HostPageSize) {
-            if (!IsCommittedPage(page / sizeof(T))) {
-                CommitPage(page / sizeof(T));
+    void CommitRegion(std::size_t index, std::size_t end_) noexcept {
+        if (base_ptr == nullptr || index >= end_) return;
+        const u64 start_off = static_cast<u64>(index) * sizeof(T);
+        const u64 end_off = static_cast<u64>(end_) * sizeof(T);
+        const u64 start_page = start_off & HostPageMask;
+        for (u64 off = start_page; off < end_off; off += HostPageSize) {
+            if (!IsPageCommitted(off)) {
+                (void)CommitPage(off);
             }
         }
     }
 
-    constexpr T& GetUnchecked(size_t index) {
-        return base_ptr[index];
-    }
+    T& GetUnchecked(std::size_t index) noexcept { return base_ptr[index]; }
 
-    [[nodiscard]] constexpr const T& operator[](std::size_t index) const noexcept {
-        return GetOrDefault(index);
-    }
-
-    [[nodiscard]] constexpr const T* data() const noexcept {
-        return base_ptr;
-    }
-
-    [[nodiscard]] constexpr std::size_t size() const noexcept {
-        return alloc_size / sizeof(T);
-    }
+    [[nodiscard]] const T& operator[](std::size_t index) const noexcept { return GetOrDefault(index); }
+    [[nodiscard]] const T* data() const noexcept { return base_ptr; }
+    [[nodiscard]] std::size_t size() const noexcept { return alloc_size / sizeof(T); }
 
 private:
-    [[nodiscard]] constexpr bool IsCommittedPage(std::size_t index) const noexcept {
-        if (index > alloc_size / sizeof(T)) {
-            LOG_CRITICAL(Common_Memory, "Out of bounds access on large vector @ {}", index);
+    void Allocate(std::size_t new_size) noexcept {
+        alloc_size = new_size;
+        if (alloc_size == 0) {
+            base_ptr = nullptr;
+            committed_pages.reset();
+            return;
+        }
+        base_ptr = static_cast<T*>(AllocateMemoryPages(alloc_size));
+        const std::size_t num_pages = NumPages();
+        const std::size_t num_words = (num_pages + 63) / 64;
+        committed_pages = std::make_unique<std::atomic<u64>[]>(num_words);
+    }
+
+    void Release() noexcept {
+        if (base_ptr != nullptr) {
+            FreeMemoryPages(base_ptr, alloc_size);
+            base_ptr = nullptr;
+        }
+        committed_pages.reset();
+        alloc_size = 0;
+    }
+
+    [[nodiscard]] u64 NumPages() const noexcept {
+        return (alloc_size + HostPageSize - 1) >> HostPageBits;
+    }
+
+    [[nodiscard]] bool IsPageCommitted(u64 byte_offset) const noexcept {
+        const u64 page_index = byte_offset >> HostPageBits;
+        if (committed_pages == nullptr || page_index >= NumPages()) return false;
+        const auto val = committed_pages[page_index >> 6].load(std::memory_order_acquire);
+        return (val >> (page_index & 63)) & 1;
+    }
+
+    void SetPageBit(u64 page_index, bool value) noexcept {
+        if (committed_pages == nullptr) return;
+        const u64 bit = 1ULL << (page_index & 63);
+        auto& atom = committed_pages[page_index >> 6];
+        if (value) {
+            atom.fetch_or(bit, std::memory_order_release);
+        } else {
+            atom.fetch_and(~bit, std::memory_order_release);
+        }
+    }
+
+    bool CommitPage(u64 byte_offset) noexcept {
+        const u64 page_index = byte_offset >> HostPageBits;
+        const uintptr_t page_addr = (reinterpret_cast<uintptr_t>(base_ptr) + byte_offset) & HostPageMask;
+
+        if (IsPageCommitted(byte_offset)) return true;
+
+#if defined(_WIN32)
+        if (!CommitVectorPage(page_addr, true)) return false;
+#else
+        if (mprotect(reinterpret_cast<void*>(page_addr), HostPageSize, PROT_READ | PROT_WRITE) != 0) {
+            LOG_ERROR(Common_Memory, "mprotect failed at {:#x}: {}", page_addr, std::strerror(errno));
             return false;
         }
-
-        auto page = (index * sizeof(T)) >> HostPageBits;
-        auto val = committed_pages[page >> 6].load(std::memory_order_acquire);
-        return (val >> (page & 63)) & 1;
-    }
-
-    constexpr void CommitPage(std::size_t index) noexcept {
-        auto page_index = (index * sizeof(T)) >> HostPageBits;
-        auto page = reinterpret_cast<uintptr_t>(base_ptr + index) & HostPageMask;
-#if defined(_WIN32)
-        CommitVectorPage(page, true);
-#else
-        mprotect(reinterpret_cast<void*>(page), HostPageSize, PROT_READ | PROT_WRITE);
 #endif
-
-        committed_pages[page_index >> 6].fetch_or(1ULL << (page_index & 63), std::memory_order_release);
+        SetPageBit(page_index, true);
+        return true;
     }
 
-    constexpr void DecommitPage(std::size_t index) noexcept {
-        auto page_index = (index * sizeof(T)) >> HostPageBits;
-        auto page = reinterpret_cast<uintptr_t>(base_ptr + index) & HostPageMask;
+    void DecommitPage(u64 byte_offset) noexcept {
+        const u64 page_index = byte_offset >> HostPageBits;
+        const uintptr_t page_addr = (reinterpret_cast<uintptr_t>(base_ptr) + byte_offset) & HostPageMask;
+        DecommitVectorPage(page_addr);
+        SetPageBit(page_index, false);
+    }
 
-        committed_pages[page_index >> 6].fetch_and(~(1ULL << (page_index & 63)), std::memory_order_release);
-        DecommitVectorPage(page);
+    [[nodiscard]] const T& DefaultValue() const noexcept {
+        return *reinterpret_cast<const T*>(&default_val);
     }
 
     std::size_t alloc_size{};
     T* base_ptr{};
-
-    std::vector<std::atomic<u64>> committed_pages{};
-#ifdef _WIN32
-    const std::array<u8, sizeof(T)> default_val{};
-#endif
+    std::unique_ptr<std::atomic<u64>[]> committed_pages{};
+    alignas(T) const std::array<u8, sizeof(T)> default_val{};
 };
 
-} // namespace Common
+} // namespace Common

src/core/file_sys/vfs/vfs_real.cpp

@@ -109,8 +109,7 @@ VirtualFile RealVfsFilesystem::OpenFileFromEntry(std::string_view path_, std::op
     auto reference = std::make_unique<FileReference>();
     this->InsertReferenceIntoListLocked(*reference);
 
-    auto file = std::shared_ptr<RealVfsFile>(
-        new RealVfsFile(*this, std::move(reference), path, perms, size, std::move(parent_path)));
+    auto file = std::make_shared<RealVfsFile>(*this, std::move(reference), path, perms, size, std::move(parent_path));
     cache[path] = file;
 
     return file;
@@ -177,7 +176,7 @@ bool RealVfsFilesystem::DeleteFile(std::string_view path_) {
 
 VirtualDir RealVfsFilesystem::OpenDirectory(std::string_view path_, OpenMode perms) {
     const auto path = FS::SanitizePath(path_, FS::DirectorySeparator::PlatformDefault);
-    return std::shared_ptr<RealVfsDirectory>(new RealVfsDirectory(*this, path, perms));
+    return std::make_shared<RealVfsDirectory>(*this, path, perms);
 }
 
 VirtualDir RealVfsFilesystem::CreateDirectory(std::string_view path_, OpenMode perms) {
@@ -185,7 +184,7 @@ VirtualDir RealVfsFilesystem::CreateDirectory(std::string_view path_, OpenMode p
     if (!FS::CreateDirs(path)) {
         return nullptr;
     }
-    return std::shared_ptr<RealVfsDirectory>(new RealVfsDirectory(*this, path, perms));
+    return std::make_shared<RealVfsDirectory>(*this, path, perms);
 }
 
 VirtualDir RealVfsFilesystem::CopyDirectory(std::string_view old_path_,

src/core/file_sys/vfs/vfs_real.h

@@ -82,6 +82,9 @@ class RealVfsFile : public VfsFile {
     friend class RealVfsFilesystem;
 
 public:
+    RealVfsFile(RealVfsFilesystem& base, std::unique_ptr<FileReference> reference,
+                const std::string& path, OpenMode perms = OpenMode::Read,
+                std::optional<u64> size = {}, std::optional<std::string> parent_path = {});
     ~RealVfsFile() override;
 
     std::string GetName() const override;
@@ -95,9 +98,6 @@ public:
     bool Rename(std::string_view name) override;
 
 private:
-    RealVfsFile(RealVfsFilesystem& base, std::unique_ptr<FileReference> reference,
-                const std::string& path, OpenMode perms = OpenMode::Read,
-                std::optional<u64> size = {}, std::optional<std::string> parent_path = {});
 
     RealVfsFilesystem& base;
     std::unique_ptr<FileReference> reference;
@@ -113,6 +113,8 @@ class RealVfsDirectory : public VfsDirectory {
     friend class RealVfsFilesystem;
 
 public:
+    RealVfsDirectory(RealVfsFilesystem& base, const std::string& path,
+                     OpenMode perms = OpenMode::Read);
     ~RealVfsDirectory() override;
 
     VirtualFile GetFileRelative(std::string_view relative_path) const override;
@@ -138,9 +140,6 @@ public:
     std::map<std::string, VfsEntryType, std::less<>> GetEntries() const override;
 
 private:
-    RealVfsDirectory(RealVfsFilesystem& base, const std::string& path,
-                     OpenMode perms = OpenMode::Read);
-
     template <typename T, typename R>
     std::vector<std::shared_ptr<R>> IterateEntries() const;
 

src/core/hle/service/am/window_system.cpp

@@ -517,7 +517,7 @@ void WindowSystem::UpdateAppletStateLocked(Applet* applet, bool is_foreground, b
     // Layer ordering. Composition sorts back-to-front. Now with enums for calrity.
     s32 z_index = Background;
     if (is_overlay) {
-        z_index = Overlay;
+        z_index = this->IsOverlayOpenLocked(*applet) ? Overlay : Background;
     } else if (inherited_foreground) {
         z_index = is_obscured ? Foreground : ForegroundVisible;
     }

src/core/hle/service/psc/ovln/receiver_service.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 2024 yuzu Emulator Project
@@ -24,7 +24,7 @@ IReceiverService::~IReceiverService() = default;
 
 Result IReceiverService::OpenReceiver(Out<SharedPointer<IReceiver>> out_receiver) {
     LOG_DEBUG(Service_PSC, "called");
-    *out_receiver = std::shared_ptr<IReceiver>(new IReceiver(system));
+    *out_receiver = std::make_shared<IReceiver>(system);
     R_SUCCEED();
 }
 

src/hid_core/resources/npad/npad.cpp

@@ -45,6 +45,19 @@ NPad::NPad(Core::HID::HIDCore& hid_core_, KernelHelpers::ServiceContext& service
         AbstractPad{hid_core_.kernel},
     }}
 {
+    for (std::size_t aruid_index = 0; aruid_index < AruidIndexMax; ++aruid_index) {
+        for (std::size_t i = 0; i < controller_data[aruid_index].size(); ++i) {
+            auto& controller = controller_data[aruid_index][i];
+            controller.device = hid_core.GetEmulatedControllerByIndex(i);
+            Core::HID::ControllerUpdateCallback engine_callback{
+                .on_change = [this, i, kernel = &hid_core.kernel](Core::HID::ControllerTriggerType type) {
+                    ControllerUpdate(*kernel, type, i);
+                },
+                .is_npad_service = true,
+            };
+            controller.callback_key = controller.device->SetCallback(engine_callback);
+        }
+    }
     for (std::size_t i = 0; i < abstracted_pads.size(); ++i) {
         abstracted_pads[i].SetNpadId(IndexToNpadIdType(i));
     }
@@ -93,16 +106,6 @@ Result NPad::Activate(u64 aruid) {
     for (std::size_t i = 0; i < controller_data[aruid_index].size(); ++i) {
         auto& controller = controller_data[aruid_index][i];
         controller.shared_memory = &data->shared_memory_format->npad.npad_entry[i].internal_state;
-        controller.device = hid_core.GetEmulatedControllerByIndex(i);
-        if (!controller.callback_key) {
-            Core::HID::ControllerUpdateCallback engine_callback{
-                .on_change = [this, i](Core::HID::ControllerTriggerType type) {
-                    ControllerUpdate(hid_core.kernel, type, i);
-                },
-                .is_npad_service = true,
-            };
-            controller.callback_key = controller.device->SetCallback(engine_callback);
-        }
     }
 
     // Prefill controller buffers

src/qt_common/config/shared_translation.cpp

@@ -226,7 +226,9 @@ std::unique_ptr<TranslationMap> InitializeTranslations(QObject* parent) {
     INSERT(Settings, dma_accuracy, tr("DMA Accuracy:"),
            tr("Controls the DMA read mode.\nUnsafe is faster, while Safe is more stable and can fix issues in some games.\nDefault follows the GPU Accuracy setting."));
     INSERT(Settings, gpu_fence_behavior, tr("GPU Fence Behavior:"),
-           tr("Controls the GPU fence synchronization behavior.\nImmediate is the fastest option, but can introduce some issues.\nBalanced offers better compatibility and may fix issues in some games.\nAccurate further improves compatibility at the cost of some performance.\nDefault follows the GPU Mode setting."));
+           tr("Controls the GPU fence synchronization behavior.\nImmediate is the fastest option, but can introduce some issues.\nBalanced offers better compatibility and may fix issues in some games.\nAccurate further improves compatibility at the cost of some performance.\nStrict is the slowest option, but can fix issues that require stricter synchronization.\nDefault follows the GPU Accuracy setting."));
+    INSERT(Settings, enable_gpu_buffer_readback, tr("Enable GPU buffer readback"),
+           tr("Preserves GPU-modified data by reading it back before uploading.\nSome games require this to render certain effects properly."));
     INSERT(Settings, use_asynchronous_shaders, tr("Enable asynchronous shader compilation"),
            tr("May reduce shader stutter."));
     INSERT(Settings, gpu_clock, tr("GPU Clocks"),
@@ -440,6 +442,7 @@ std::unique_ptr<ComboboxTranslationMap> ComboboxEnumeration(QObject* parent) {
                               PAIR(GpuFenceBehavior, Immediate, tr("Immediate")),
                               PAIR(GpuFenceBehavior, Balanced, tr("Balanced")),
                               PAIR(GpuFenceBehavior, Accurate, tr("Accurate")),
+                              PAIR(GpuFenceBehavior, Strict, tr("Strict")),
                           }});
     translations->insert(
         {Settings::EnumMetadata<Settings::CpuAccuracy>::Index(),

src/video_core/buffer_cache/buffer_cache.h

@@ -249,10 +249,6 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     runtime.CopyBuffer(dest_buffer, src_buffer, copies, true);
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
-        const bool should_sync = Settings::IsGPUFenceBehaviorBalanced() || Settings::IsGPUFenceBehaviorAccurate();
-        if (should_sync) {
-            runtime.Finish();
-        }
     }
 
     Tegra::Memory::DeviceGuestMemoryScoped<u8, Tegra::Memory::GuestMemoryFlags::UnsafeReadWrite>
@@ -1234,7 +1230,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
         buffer.MarkUsage(offset, size);
 
         if (is_written) {
-            MarkWrittenBuffer(binding.buffer_id, binding.device_addr, size, true);
+            MarkWrittenBuffer(binding.buffer_id, binding.device_addr, size);
         }
 
         if constexpr (NEEDS_BIND_STORAGE_INDEX) {
@@ -1520,12 +1516,10 @@ void BufferCache<P>::UpdateComputeTextureBuffers() {
 }
 
 template <class P>
-void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size, bool needs_sync) {
+void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size) {
     if constexpr (!IS_OPENGL) {
-        if (needs_sync) {
-            Buffer& buffer = slot_buffers[buffer_id];
-            buffer.setWriteTick(runtime.CurrentTick());
-        }
+        Buffer& buffer = slot_buffers[buffer_id];
+        buffer.setWriteTick(runtime.CurrentTick());
     }
     memory_tracker.MarkRegionAsGpuModified(device_addr, size);
     gpu_modified_ranges.Add(device_addr, size);
@@ -1541,11 +1535,8 @@ BufferId BufferCache<P>::FindBuffer(DAddr device_addr, u32 size, bool sparse_com
     const BufferId buffer_id = page_table[page];
     if (buffer_id) {
         Buffer& buffer = slot_buffers[buffer_id];
+        WaitForGpuFenceIfNeeded(buffer);
         if (buffer.IsInBounds(device_addr, size)) {
-            const bool should_sync = Settings::IsGPUFenceBehaviorAccurate();
-            if (should_sync) {
-                SynchronizeBufferWrites(buffer);
-            }
             bool usable = true;
             if constexpr (requires { buffer.IsSparseCompatible(); }) {
                 if (sparse_compatible && !buffer.IsSparseCompatible()) {
@@ -1561,11 +1552,17 @@ BufferId BufferCache<P>::FindBuffer(DAddr device_addr, u32 size, bool sparse_com
 }
 
 template <class P>
-void BufferCache<P>::SynchronizeBufferWrites(Buffer& buffer) {
+void BufferCache<P>::WaitForGpuFenceIfNeeded(Buffer& buffer) {
     if constexpr (!IS_OPENGL) {
-        const u64 buffer_tick = buffer.getWriteTick();
-        if (!runtime.IsFree(buffer_tick)) {
-            runtime.Wait(buffer_tick);
+        const bool gpu_fence_accurate = Settings::IsGPUFenceBehaviorAccurate();
+        const bool gpu_fence_strict = Settings::IsGPUFenceBehaviorStrict();
+        if (gpu_fence_accurate || gpu_fence_strict) {
+            const u64 gpu_tick_delay = gpu_fence_strict ? 0 : 3;
+            const u64 buffer_tick = buffer.getWriteTick();
+            const u64 gpu_tick = runtime.KnownGpuTick();
+            if (buffer_tick > gpu_tick + gpu_tick_delay) {
+                runtime.Wait(buffer_tick);
+            }
         }
     }
 }
@@ -1798,6 +1795,9 @@ void BufferCache<P>::ImmediateUploadMemory([[maybe_unused]] Buffer& buffer,
                 if (immediate_buffer.empty()) {
                     immediate_buffer = ImmediateBuffer(largest_copy);
                 }
+                if (Settings::values.enable_gpu_buffer_readback.GetValue()) {
+                    DownloadBufferMemory(buffer, device_addr, copy.size);
+                }
                 device_memory.ReadBlockUnsafe(device_addr, immediate_buffer.data(), copy.size);
                 upload_span = immediate_buffer.subspan(0, copy.size);
             }
@@ -1816,6 +1816,9 @@ void BufferCache<P>::MappedUploadMemory([[maybe_unused]] Buffer& buffer,
         for (BufferCopy& copy : copies) {
             u8* const src_pointer = staging_pointer.data() + copy.src_offset;
             const DAddr device_addr = buffer.CpuAddr() + copy.dst_offset;
+            if (Settings::values.enable_gpu_buffer_readback.GetValue()) {
+                DownloadBufferMemory(buffer, device_addr, copy.size);
+            }
             device_memory.ReadBlockUnsafe(device_addr, src_pointer, copy.size);
             // Apply the staging offset
             copy.src_offset += upload_staging.offset;

src/video_core/buffer_cache/buffer_cache_base.h

@@ -430,11 +430,11 @@ private:
 
     void UpdateComputeTextureBuffers();
 
-    void MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size, bool needs_sync = false);
+    void MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size);
 
     [[nodiscard]] BufferId FindBuffer(DAddr device_addr, u32 size, bool sparse_compatible);
 
-    void SynchronizeBufferWrites(Buffer& buffer);
+    void WaitForGpuFenceIfNeeded(Buffer& buffer);
 
     [[nodiscard]] OverlapResult ResolveOverlaps(DAddr device_addr, u32 wanted_size);
 

src/video_core/fence_manager.h

@@ -72,7 +72,7 @@ public:
     }
 
     void SignalFence(std::function<void()>&& func) {
-        const bool delay_fence = Settings::IsGPUFenceBehaviorDefault() ? Settings::IsGPULevelHigh() : Settings::IsGPUFenceBehaviorBalanced() || Settings::IsGPUFenceBehaviorAccurate();
+        const bool delay_fence = Settings::IsGPUFenceBehaviorDefault() ? Settings::IsGPULevelHigh() : Settings::IsGPUFenceBehaviorBalanced() || Settings::IsGPUFenceBehaviorAccurate() || Settings::IsGPUFenceBehaviorStrict();
         const bool should_flush = ShouldFlush();
         if constexpr (!can_async_check) {
             TryReleasePendingFences<false>();

src/video_core/query_cache/query_cache.h

@@ -260,7 +260,7 @@ void QueryCacheBase<Traits>::CounterReport(GPUVAddr addr, QueryType counter_type
     };
     u8* pointer = impl->device_memory.template GetPointer<u8>(cpu_addr);
     u8* pointer_timestamp = impl->device_memory.template GetPointer<u8>(cpu_addr + 8);
-    bool is_synced = (Settings::IsGPUFenceBehaviorDefault() ? !Settings::IsGPULevelHigh() : !Settings::IsGPUFenceBehaviorBalanced() && !Settings::IsGPUFenceBehaviorAccurate()) && is_fence;
+    bool is_synced = (Settings::IsGPUFenceBehaviorDefault() ? !Settings::IsGPULevelHigh() : !Settings::IsGPUFenceBehaviorBalanced() && !Settings::IsGPUFenceBehaviorAccurate() && !Settings::IsGPUFenceBehaviorStrict()) && is_fence;
     std::function<void()> operation([this, is_synced, streamer, query_base = query, query_location,
                                      pointer, pointer_timestamp] {
         if (True(query_base->flags & QueryFlagBits::IsInvalidated)) {

src/video_core/renderer_vulkan/present/layer.cpp

@@ -94,13 +94,13 @@ void Layer::ConfigureDraw(const Device& device, PresentPushConstants* out_push_c
     const u32 scaled_width = texture_info ? texture_info->scaled_width : texture_width;
     const u32 scaled_height = texture_info ? texture_info->scaled_height : texture_height;
     const bool use_accelerated = texture_info.has_value();
+    const bool is_applet =
+        (framebuffer.layer_stack_mask & Service::Nvnflinger::LayerStackBit(
+                                            Service::Nvnflinger::LayerStackId::Recording)) == 0;
 
     RefreshResources(device, framebuffer);
     SetAntiAliasPass(device);
 #ifdef HAS_RESHADE
-    const bool is_applet =
-        (framebuffer.layer_stack_mask & Service::Nvnflinger::LayerStackBit(
-                                            Service::Nvnflinger::LayerStackId::Recording)) == 0;
     SetPostProcessPass(device, is_applet);
 #endif
 
@@ -141,8 +141,11 @@ void Layer::ConfigureDraw(const Device& device, PresentPushConstants* out_push_c
         source_image_view = fsr->Draw(device, scheduler, image_index, source_image, source_image_view, render_extent, crop_rect);
         crop_rect = {0, 0, 1, 1};
     } else if (auto* sgsr = std::get_if<SGSR>(&sr_filter)) {
-        source_image_view = sgsr->Draw(device, scheduler, image_index, source_image, source_image_view, render_extent, crop_rect);
-        crop_rect = {0, 0, 1, 1};
+        if (!is_applet) {
+            source_image_view = sgsr->Draw(device, scheduler, image_index, source_image,
+                                           source_image_view, render_extent, crop_rect);
+            crop_rect = {0, 0, 1, 1};
+        }
     }
 
     SetMatrixData(device, *out_push_constants, layout);

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -419,15 +419,15 @@ void BufferCacheRuntime::TickFrame(Common::SlotVector<Buffer>& slot_buffers) noe
 }
 
 u64 BufferCacheRuntime::CurrentTick() {
-    return scheduler.CurrentTick();
+    return scheduler.GetMasterSemaphore().CurrentTick();
 }
 
-bool BufferCacheRuntime::IsFree(u64 tick) {
-    return scheduler.IsFree(tick);
+u64 BufferCacheRuntime::KnownGpuTick() {
+    return scheduler.GetMasterSemaphore().KnownGpuTick();
 }
 
-void BufferCacheRuntime::Wait(u64 tick) {
-    scheduler.Wait(tick);
+void BufferCacheRuntime::Wait(u64 buffer_tick) {
+    scheduler.Wait(buffer_tick);
 }
 
 void BufferCacheRuntime::Finish() {

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -112,9 +112,9 @@ public:
 
     u64 CurrentTick();
 
-    bool IsFree(u64 tick);
+    u64 KnownGpuTick();
 
-    void Wait(u64 tick);
+    void Wait(u64 buffer_tick);
 
     void Finish();