Ops

- [x] I have read and followed the [Contribution Guidelines](https://git.eden-emu.dev/eden-emu/eden/src/branch/master/CONTRIBUTING.md#code-contributions).
- [x] I have read and followed the [AI Policy](https://git.eden-emu.dev/eden-emu/eden/src/branch/master/docs/policies/AI.md)
- [x] I have read and followed the [Coding Guidelines](https://git.eden-emu.dev/eden-emu/eden/src/branch/master/docs/policies/Coding.md) to the best of my ability.

-------------------

Aimed to fix two known UE5 crashes: Kepler uploads and Maxwell macros, both caused by CPU/GPU races due dirty tracking issues.

Kepler ComputeInline: preserved dirty tracking across dma continuations and async readback.
Maxwell macros: preserved gpu owned subranges during page granular cpu uploads.
DiscardWrite: stopped clearing neighboring macro arguments by rounding up ranges.
DMA Step: improved continuation aware dirty sampling.

To the Ender Magnolia crew (maybe 1 or 2 persons): This will fix the dash crash, and the random / shackled beast vaper crashes.

There are some more UE5 issues to go next.

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/4473
Reviewed-by: lizzie <lizzie@eden-emu.dev>
Reviewed-by: MaranBr <maranbr@eden-emu.dev>

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

@@ -549,7 +549,6 @@ 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

@@ -580,8 +580,6 @@
     <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>
 
@@ -1099,7 +1097,6 @@
     <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

@@ -967,7 +967,6 @@ 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

@@ -524,8 +524,6 @@
     <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>
 
@@ -1038,7 +1036,6 @@
     <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

@@ -569,8 +569,6 @@
     <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>
 
@@ -1088,7 +1086,6 @@
     <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

@@ -570,8 +570,6 @@
     <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>
 
@@ -1089,7 +1087,6 @@
     <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

@@ -561,8 +561,6 @@
     <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

@@ -558,7 +558,6 @@
         <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

@@ -586,8 +586,6 @@
     <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>
 
@@ -1138,7 +1136,6 @@
     <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/settings.cpp

@@ -178,10 +178,6 @@ 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::Strict,
+                                                                 GpuFenceBehavior::Accurate,
                                                                  "gpu_fence_behavior",
                                                                  Category::RendererAdvanced,
                                                                  Specialization::Default,
@@ -655,13 +655,6 @@ 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};
@@ -979,7 +972,6 @@ 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, Strict);
+ENUM(GpuFenceBehavior, Default, Immediate, Balanced, Accurate);
 ENUM(CpuBackend, Dynarmic, Nce);
 ENUM(CpuAccuracy, Auto, Accurate, Unsafe, Paranoid, Debugging);
 ENUM(CpuClock, Normal, Boost, Overclock)

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 = this->IsOverlayOpenLocked(*applet) ? Overlay : Background;
+        z_index = Overlay;
     } else if (inherited_foreground) {
         z_index = is_obscured ? Foreground : ForegroundVisible;
     }

src/hid_core/resources/npad/npad.cpp

@@ -45,19 +45,6 @@ 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));
     }
@@ -106,6 +93,16 @@ 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

@@ -226,9 +226,7 @@ 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.\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."));
+           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."));
     INSERT(Settings, use_asynchronous_shaders, tr("Enable asynchronous shader compilation"),
            tr("May reduce shader stutter."));
     INSERT(Settings, gpu_clock, tr("GPU Clocks"),
@@ -442,7 +440,6 @@ 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

@@ -121,7 +121,7 @@ void BufferCache<P>::UnmapGPUMemory(size_t as_id, GPUVAddr gpu_addr, size_t size
 
 template <class P>
 void BufferCache<P>::WriteMemory(DAddr device_addr, u64 size) {
-    if (memory_tracker.IsRegionGpuModified(device_addr, size)) {
+    if (IsRegionGpuModified(device_addr, size)) {
         ClearDownload(device_addr, size);
         gpu_modified_ranges.Subtract(device_addr, size);
     }
@@ -249,6 +249,10 @@ 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();
+        if (should_sync) {
+            runtime.Finish();
+        }
     }
 
     Tegra::Memory::DeviceGuestMemoryScoped<u8, Tegra::Memory::GuestMemoryFlags::UnsafeReadWrite>
@@ -311,11 +315,8 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainCPUBuffer(
         MarkWrittenBuffer(buffer_id, device_addr, size);
         break;
     case ObtainBufferOperation::DiscardWrite: {
-        const DAddr device_addr_start = Common::AlignDown(device_addr, 64);
-        const DAddr device_addr_end = Common::AlignUp(device_addr + size, 64);
-        const size_t new_size = device_addr_end - device_addr_start;
-        ClearDownload(device_addr_start, new_size);
-        gpu_modified_ranges.Subtract(device_addr_start, new_size);
+        ClearDownload(device_addr, size);
+        gpu_modified_ranges.Subtract(device_addr, size);
         break;
     }
     default:
@@ -1233,7 +1234,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
         buffer.MarkUsage(offset, size);
 
         if (is_written) {
-            MarkWrittenBuffer(binding.buffer_id, binding.device_addr, size);
+            MarkWrittenBuffer(binding.buffer_id, binding.device_addr, size, true);
         }
 
         if constexpr (NEEDS_BIND_STORAGE_INDEX) {
@@ -1519,10 +1520,12 @@ void BufferCache<P>::UpdateComputeTextureBuffers() {
 }
 
 template <class P>
-void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size) {
+void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size, bool needs_sync) {
     if constexpr (!IS_OPENGL) {
-        Buffer& buffer = slot_buffers[buffer_id];
-        buffer.setWriteTick(runtime.CurrentTick());
+        if (needs_sync) {
+            Buffer& buffer = slot_buffers[buffer_id];
+            buffer.setWriteTick(runtime.CurrentTick());
+        }
     }
     memory_tracker.MarkRegionAsGpuModified(device_addr, size);
     gpu_modified_ranges.Add(device_addr, size);
@@ -1538,8 +1541,11 @@ 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()) {
@@ -1555,17 +1561,11 @@ BufferId BufferCache<P>::FindBuffer(DAddr device_addr, u32 size, bool sparse_com
 }
 
 template <class P>
-void BufferCache<P>::WaitForGpuFenceIfNeeded(Buffer& buffer) {
+void BufferCache<P>::SynchronizeBufferWrites(Buffer& buffer) {
     if constexpr (!IS_OPENGL) {
-        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);
-            }
+        const u64 buffer_tick = buffer.getWriteTick();
+        if (!runtime.IsFree(buffer_tick)) {
+            runtime.Wait(buffer_tick);
         }
     }
 }
@@ -1742,14 +1742,24 @@ bool BufferCache<P>::SynchronizeBuffer(Buffer& buffer, DAddr device_addr, u32 si
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
     const DAddr buffer_start = buffer.cpu_addr_cached;
-    memory_tracker.ForEachUploadRange(device_addr, size, [&](u64 device_addr_out, u64 range_size) {
+    const auto add_upload = [&](DAddr start, DAddr end) {
+        if (start == end) return;
+        const u64 range_size = end - start;
         upload_copies.push_back(BufferCopy{
             .src_offset = total_size_bytes,
-            .dst_offset = device_addr_out - buffer_start,
+            .dst_offset = start - buffer_start,
             .size = range_size,
         });
         total_size_bytes += range_size;
         largest_copy = (std::max)(largest_copy, range_size);
+    };
+    memory_tracker.ForEachUploadRange(device_addr, size, [&](u64 device_addr_out, u64 range_size) {
+        DAddr upload_start = device_addr_out;
+        gpu_modified_ranges.ForEachInRange(device_addr_out, range_size, [&](DAddr gpu_start, DAddr gpu_end) {
+            add_upload(upload_start, gpu_start);
+            upload_start = gpu_end;
+        });
+        add_upload(upload_start, device_addr_out + range_size);
     });
     if (total_size_bytes == 0) {
         return true;
@@ -1788,9 +1798,6 @@ 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);
             }
@@ -1809,9 +1816,6 @@ 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);
+    void MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size, bool needs_sync = false);
 
     [[nodiscard]] BufferId FindBuffer(DAddr device_addr, u32 size, bool sparse_compatible);
 
-    void WaitForGpuFenceIfNeeded(Buffer& buffer);
+    void SynchronizeBufferWrites(Buffer& buffer);
 
     [[nodiscard]] OverlapResult ResolveOverlaps(DAddr device_addr, u32 wanted_size);
 

src/video_core/dma_pusher.cpp

@@ -16,7 +16,7 @@
 namespace Tegra {
 
 constexpr u32 MacroRegistersStart = 0xE00;
-[[maybe_unused]] constexpr u32 ComputeInline = 0x6D;
+constexpr u32 ComputeInline = 0x6D;
 
 DmaPusher::DmaPusher(Core::System& system_, MemoryManager& memory_manager_, Control::ChannelState& channel_state_)
     : system{system_}
@@ -73,11 +73,16 @@ bool DmaPusher::Step() {
         synced = false;
     }
 
-    if (header.size > 0 && dma_state.method >= MacroRegistersStart && subchannels[dma_state.subchannel]) {
-        subchannels[dma_state.subchannel]->current_dirty = memory_manager.IsMemoryDirty(dma_state.dma_get, header.size * sizeof(u32));
-    }
-
     if (header.size > 0) {
+        if (subchannels[dma_state.subchannel] && dma_state.method_count) {
+            const auto engine = subchannel_type[dma_state.subchannel];
+            const bool kepler_payload = engine == Engines::EngineTypes::KeplerCompute && dma_state.method == ComputeInline && dma_state.non_incrementing;
+            const bool macro_payload = engine == Engines::EngineTypes::Maxwell3D && dma_state.method >= MacroRegistersStart;
+            if (kepler_payload || macro_payload) {
+                const size_t words = std::min<size_t>(dma_state.method_count, header.size);
+                subchannels[dma_state.subchannel]->current_dirty = memory_manager.IsMemoryDirty(dma_state.dma_get, words * sizeof(u32));
+            }
+        }
         const bool use_safe = Settings::IsDMALevelDefault() ? Settings::IsGPULevelHigh() : Settings::IsDMALevelSafe();
         if (use_safe) {
             Tegra::Memory::GpuGuestMemory<Tegra::CommandHeader, Tegra::Memory::GuestMemoryFlags::SafeRead>headers(memory_manager, dma_state.dma_get, header.size, &command_headers);

src/video_core/engines/kepler_compute.cpp

@@ -49,13 +49,16 @@ void KeplerCompute::CallMethod(Core::System& system, u32 method, u32 method_argu
     case KEPLER_COMPUTE_REG_INDEX(exec_upload): {
         UploadInfo info{.upload_address = upload_address,
                         .exec_address = upload_state.ExecTargetAddress(),
-                        .copy_size = upload_state.GetUploadSize()};
+                        .copy_size = upload_state.GetUploadSize(),
+                        .was_dirty = upload_dirty};
         uploads.push_back(info);
         upload_state.ProcessExec(regs.exec_upload.linear != 0);
         break;
     }
     case KEPLER_COMPUTE_REG_INDEX(data_upload): {
         upload_address = current_dma_segment;
+        upload_dirty = current_dirty;
+        current_dirty = false;
         upload_state.ProcessData(method_argument, is_last_call);
         break;
     }
@@ -64,9 +67,11 @@ void KeplerCompute::CallMethod(Core::System& system, u32 method, u32 method_argu
 
         for (auto& data : uploads) {
             const GPUVAddr offset = data.exec_address - launch_desc_loc;
-            if (offset / sizeof(u32) == LAUNCH_REG_INDEX(grid_dim_x) &&
-                memory_manager.IsMemoryDirty(data.upload_address, data.copy_size)) {
-                indirect_compute = {data.upload_address};
+            if (offset / sizeof(u32) == LAUNCH_REG_INDEX(grid_dim_x)) {
+                const bool source_dirty = memory_manager.IsMemoryDirty(data.upload_address, data.copy_size);
+                if (data.was_dirty || source_dirty) {
+                    indirect_compute = {data.upload_address};
+                }
             }
         }
         uploads.clear();
@@ -83,6 +88,8 @@ void KeplerCompute::CallMultiMethod(Core::System& system, u32 method, const u32*
     switch (method) {
     case KEPLER_COMPUTE_REG_INDEX(data_upload):
         upload_address = current_dma_segment;
+        upload_dirty = current_dirty;
+        current_dirty = false;
         upload_state.ProcessData(base_start, amount);
         return;
     default:

src/video_core/engines/kepler_compute.h

@@ -226,11 +226,13 @@ private:
     VideoCore::RasterizerInterface* rasterizer = nullptr;
     Upload::State upload_state;
     GPUVAddr upload_address;
+    bool upload_dirty{};
 
     struct UploadInfo {
         GPUVAddr upload_address;
         GPUVAddr exec_address;
         u32 copy_size;
+        bool was_dirty;
     };
     std::vector<UploadInfo> uploads;
     std::optional<GPUVAddr> indirect_compute{};

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() || Settings::IsGPUFenceBehaviorStrict();
+        const bool delay_fence = Settings::IsGPUFenceBehaviorDefault() ? Settings::IsGPULevelHigh() : Settings::IsGPUFenceBehaviorBalanced() || Settings::IsGPUFenceBehaviorAccurate();
         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() && !Settings::IsGPUFenceBehaviorStrict()) && is_fence;
+    bool is_synced = (Settings::IsGPUFenceBehaviorDefault() ? !Settings::IsGPULevelHigh() : !Settings::IsGPUFenceBehaviorBalanced() && !Settings::IsGPUFenceBehaviorAccurate()) && 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
 
@@ -141,11 +141,8 @@ 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)) {
-        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};
-        }
+        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.GetMasterSemaphore().CurrentTick();
+    return scheduler.CurrentTick();
 }
 
-u64 BufferCacheRuntime::KnownGpuTick() {
-    return scheduler.GetMasterSemaphore().KnownGpuTick();
+bool BufferCacheRuntime::IsFree(u64 tick) {
+    return scheduler.IsFree(tick);
 }
 
-void BufferCacheRuntime::Wait(u64 buffer_tick) {
-    scheduler.Wait(buffer_tick);
+void BufferCacheRuntime::Wait(u64 tick) {
+    scheduler.Wait(tick);
 }
 
 void BufferCacheRuntime::Finish() {

src/video_core/renderer_vulkan/vk_buffer_cache.h

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