fix totk purple

Begin and end were landing in different scopes: the clear and draw
texture paths opened it with no pass bound, the compare helpers reopened
it after their resolve pass had closed one, and the pass teardown could
return before pausing, leaving it active at submit.

src/android/app/build.gradle.kts

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

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

@@ -27,6 +27,7 @@ enum class BooleanSetting(override val key: String) : AbstractBooleanSetting {
     RENDERER_ASYNCHRONOUS_GPU_EMULATION("use_asynchronous_gpu_emulation"),
     RENDERER_ASYNC_PRESENTATION("async_presentation"),
     RENDERER_ASYNCHRONOUS_SHADERS("use_asynchronous_shaders"),
+    RENDERER_UNIFIED_MEMORY("use_unified_memory"),
     RENDERER_REACTIVE_FLUSHING("use_reactive_flushing"),
     ENABLE_BUFFER_HISTORY("enable_buffer_history"),
     USE_OPTIMIZED_VERTEX_BUFFERS("use_optimized_vertex_buffers"),
@@ -36,6 +37,8 @@ enum class BooleanSetting(override val key: String) : AbstractBooleanSetting {
     RENDERER_DEBUG("debug"),
     RENDERER_PATCH_OLD_QCOM_DRIVERS("patch_old_qcom_drivers"),
     RENDERER_VERTEX_INPUT_DYNAMIC_STATE("vertex_input_dynamic_state"),
+    RENDERER_DYNAMIC_RENDERING("dynamic_rendering"),
+    RENDERER_WORKGROUP_MEMORY_EXPLICIT_LAYOUT("workgroup_memory_explicit_layout"),
     RENDERER_SAMPLE_SHADING("sample_shading"),
     GPU_UNSWIZZLE_ENABLED("gpu_unswizzle_enabled"),
     PICTURE_IN_PICTURE("picture_in_picture"),

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

@@ -155,6 +155,20 @@ abstract class SettingsItem(
                     descriptionId = R.string.vertex_input_dynamic_state_description
                 )
             )
+            put(
+                SwitchSetting(
+                    BooleanSetting.RENDERER_DYNAMIC_RENDERING,
+                    titleId = R.string.dynamic_rendering,
+                    descriptionId = R.string.dynamic_rendering_description
+                )
+            )
+            put(
+                SwitchSetting(
+                    BooleanSetting.RENDERER_WORKGROUP_MEMORY_EXPLICIT_LAYOUT,
+                    titleId = R.string.workgroup_memory_explicit_layout,
+                    descriptionId = R.string.workgroup_memory_explicit_layout_description
+                )
+            )
             put(
                 SliderSetting(
                     IntSetting.RENDERER_SAMPLE_SHADING,
@@ -594,7 +608,7 @@ abstract class SettingsItem(
                     IntSetting.ANDROID_PIPELINE_WORKERS,
                     titleId = R.string.pipeline_worker_cores,
                     descriptionId = R.string.pipeline_worker_cores_description,
-                    min = 4,
+                    min = 2,
                     max = 8,
                     units = "cores"
                 )
@@ -685,6 +699,13 @@ abstract class SettingsItem(
                     descriptionId = R.string.renderer_asynchronous_shaders_description
                 )
             )
+            put(
+                SwitchSetting(
+                    BooleanSetting.RENDERER_UNIFIED_MEMORY,
+                    titleId = R.string.renderer_unified_memory,
+                    descriptionId = R.string.renderer_unified_memory_description
+                )
+            )
             put(
                 SingleChoiceSetting(
                     IntSetting.FAST_GPU_TIME,

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

@@ -304,6 +304,7 @@ class SettingsFragmentPresenter(
             add(BooleanSetting.EMULATE_BGR565.key)
             add(BooleanSetting.RESCALE_HACK.key)
             add(BooleanSetting.RENDERER_ASYNCHRONOUS_SHADERS.key)
+            add(BooleanSetting.RENDERER_UNIFIED_MEMORY.key)
             add(IntSetting.ANDROID_PIPELINE_WORKERS.key)
             add(BooleanSetting.RENDERER_ASYNCHRONOUS_GPU_EMULATION.key)
             add(BooleanSetting.RENDERER_ASYNC_PRESENTATION.key)
@@ -313,6 +314,8 @@ class SettingsFragmentPresenter(
 
             add(IntSetting.RENDERER_DYNA_STATE.key)
             add(BooleanSetting.RENDERER_VERTEX_INPUT_DYNAMIC_STATE.key)
+            add(BooleanSetting.RENDERER_DYNAMIC_RENDERING.key)
+            add(BooleanSetting.RENDERER_WORKGROUP_MEMORY_EXPLICIT_LAYOUT.key)
             add(IntSetting.RENDERER_SAMPLE_SHADING.key)
 
             add(HeaderSetting(R.string.display))

src/android/app/src/main/jni/android_settings.h

@@ -147,13 +147,6 @@ namespace AndroidSettings {
                                                       &show_performance_overlay};
 
 
-        Settings::Setting<s32> pipeline_worker_count{linkage, 4, "pipeline_worker_count",
-                                 Settings::Category::Android,
-                                 Settings::Specialization::Default,
-                                 true,
-                                 true};
-
-
         Settings::Setting<bool> show_input_overlay{linkage, true, "show_input_overlay",
                                                    Settings::Category::Overlay};
         Settings::Setting<bool> overlay_snap_to_grid{linkage, false, "overlay_snap_to_grid",

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

@@ -524,6 +524,8 @@
     <string name="rescale_hack_description">Enables a legacy handling for the rescale configuration pass for games by using a quick rescale path</string>
     <string name="renderer_asynchronous_shaders">Use asynchronous shaders</string>
     <string name="renderer_asynchronous_shaders_description">Compiles shaders asynchronously. This may reduce stutters but may also introduce glitches.</string>
+    <string name="renderer_unified_memory">Unified memory access (UMA)</string>
+    <string name="renderer_unified_memory_description">Allows GPU write buffer readbacks directly into guest memory, skipping the CPU staging copy.</string>
     <string name="gpu_unswizzle_settings">GPU Unswizzle Settings</string>
     <string name="gpu_unswizzle_settings_description">Configure GPU-based texture unswizzling parameters or disable it entirely. Adjust these settings to balance performance and texture loading quality.</string>
     <string name="gpu_unswizzle_enable">Enable GPU Unswizzle</string>
@@ -544,6 +546,10 @@
     <string name="disabled">Disabled</string>
     <string name="vertex_input_dynamic_state">Vertex Input Dynamic State</string>
     <string name="vertex_input_dynamic_state_description">Enabling this feature allows for more flexible vertex input handling, potentially reducing pipeline compilation time in vertex/buffer.</string>
+    <string name="dynamic_rendering">Dynamic Rendering</string>
+    <string name="dynamic_rendering_description">Render without render pass and framebuffer objects. Results vary by driver: some gain performance, others lose it.</string>
+    <string name="workgroup_memory_explicit_layout">Workgroup Memory Explicit Layout</string>
+    <string name="workgroup_memory_explicit_layout_description">Let shaders declare explicit layouts for workgroup memory. Disabled by default: some Qualcomm drivers are unstable with it.</string>
     <string name="sample_shading_fraction">Sample Shading</string>
     <string name="sample_shading_fraction_description">Allows the fragment shader to execute per sample in a multi-sampled fragment instead once per fragment. Improves graphics quality at the cost of some performance.</string>
 

src/common/host_memory.cpp

@@ -51,14 +51,60 @@
 
 #endif // ^^^ POSIX ^^^
 
+#include <atomic>
 #include <mutex>
 #include <random>
+#include <vector>
 
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/free_region_manager.h"
 #include "common/host_memory.h"
 #include "common/logging.h"
+#include "common/memory_detect.h"
+#include "common/settings.h"
+
+#ifdef __ANDROID__
+#include <dlfcn.h>
+#include <android/hardware_buffer.h>
+
+namespace {
+
+struct NativeHandle {
+    int version;
+    int numFds;
+    int numInts;
+    int data[1];
+};
+
+using PFN_AHardwareBuffer_getNativeHandle = const NativeHandle* (*)(const AHardwareBuffer*);
+using PFN_AHardwareBuffer_isSupported = int (*)(const AHardwareBuffer_Desc*);
+
+void* NativeWindowLibrary() {
+    static void* const lib = dlopen("libnativewindow.so", RTLD_NOW);
+    return lib;
+}
+
+PFN_AHardwareBuffer_getNativeHandle ResolveGetNativeHandle() {
+    void* const lib = NativeWindowLibrary();
+    if (lib == nullptr) {
+        return nullptr;
+    }
+    return reinterpret_cast<PFN_AHardwareBuffer_getNativeHandle>(
+        dlsym(lib, "AHardwareBuffer_getNativeHandle"));
+}
+
+PFN_AHardwareBuffer_isSupported ResolveIsSupported() {
+    void* const lib = NativeWindowLibrary();
+    if (lib == nullptr) {
+        return nullptr;
+    }
+    return reinterpret_cast<PFN_AHardwareBuffer_isSupported>(
+        dlsym(lib, "AHardwareBuffer_isSupported"));
+}
+
+} // namespace
+#endif
 
 #if defined(__ANDROID__) && __ANDROID_API__ < 30
 #include <sys/syscall.h>
@@ -75,6 +121,12 @@ namespace Common {
 [[maybe_unused]] constexpr size_t PageAlignment = 0x1000;
 [[maybe_unused]] constexpr size_t HugePageSize = 0x200000;
 
+static std::atomic<u64> committed_backing_size{};
+
+u64 GetCommittedBackingSize() noexcept {
+    return committed_backing_size.load(std::memory_order_relaxed);
+}
+
 #ifdef _WIN32
 
 // Manually imported for MinGW compatibility
@@ -509,6 +561,11 @@ public:
     bool Init() {
         long page_size = sysconf(_SC_PAGESIZE);
         ASSERT_MSG(page_size == 0x1000, "page size {:#x} is incompatible with 4K paging", page_size);
+#ifdef __ANDROID__
+        if (InitAhbBacking()) {
+            return InitVirtual();
+        }
+#endif
         // Backing memory initialization
 #if defined(__sun__) || defined(__HAIKU__) || defined(__NetBSD__) || defined(__DragonFly__)
         fd = shm_open_anon(O_RDWR | O_CREAT | O_EXCL | O_NOFOLLOW, 0600);
@@ -554,7 +611,10 @@ public:
             return false;
         }
 
-        // Virtual memory initialization
+        return InitVirtual();
+    }
+
+    bool InitVirtual() {
         virtual_base = virtual_map_base = static_cast<u8*>(ChooseVirtualBase(virtual_size));
         if (virtual_base == MAP_FAILED) {
             LOG_CRITICAL(HW_Memory, "mmap failed: {}", strerror(errno));
@@ -567,6 +627,180 @@ public:
         return true;
     }
 
+#ifdef __ANDROID__
+    static AHardwareBuffer_Desc MakeBlobDesc(size_t len) {
+        return AHardwareBuffer_Desc{
+            .width = static_cast<u32>(len),
+            .height = 1,
+            .layers = 1,
+            .format = AHARDWAREBUFFER_FORMAT_BLOB,
+            .usage = AHARDWAREBUFFER_USAGE_CPU_READ_OFTEN |
+                     AHARDWAREBUFFER_USAGE_CPU_WRITE_OFTEN |
+                     AHARDWAREBUFFER_USAGE_GPU_DATA_BUFFER,
+            .stride = 0,
+            .rfu0 = 0,
+            .rfu1 = 0,
+        };
+    }
+
+    static bool ProbeAhbBacking(PFN_AHardwareBuffer_getNativeHandle get_native_handle) {
+        const AHardwareBuffer_Desc desc = MakeBlobDesc(PageAlignment * 2);
+        AHardwareBuffer* buffer{};
+        if (AHardwareBuffer_allocate(&desc, &buffer) != 0 || buffer == nullptr) {
+            LOG_WARNING(HW_Memory, "Hardware buffer probe allocation failed");
+            return false;
+        }
+        const NativeHandle* const handle = get_native_handle(buffer);
+        if (handle == nullptr || handle->numFds < 1) {
+            LOG_WARNING(HW_Memory, "Hardware buffer has no mappable file descriptor");
+            AHardwareBuffer_release(buffer);
+            return false;
+        }
+        const int probe_fd = handle->data[0];
+        bool ok = true;
+        const auto try_map = [&](int prot, off_t offset, const char* what) {
+            if (!ok) {
+                return;
+            }
+            void* const ptr = mmap(nullptr, PageAlignment, prot, MAP_SHARED, probe_fd, offset);
+            if (ptr == MAP_FAILED) {
+                LOG_WARNING(HW_Memory, "Hardware buffer backing rejects {}: {}", what,
+                            strerror(errno));
+                ok = false;
+                return;
+            }
+            munmap(ptr, PageAlignment);
+        };
+        try_map(PROT_READ | PROT_WRITE, 0, "shared mappings");
+        try_map(PROT_READ | PROT_WRITE, static_cast<off_t>(PageAlignment), "mappings at an offset");
+#ifdef ARCHITECTURE_arm64
+        try_map(PROT_READ | PROT_EXEC, 0, "executable mappings");
+#endif
+        AHardwareBuffer_release(buffer);
+        return ok;
+    }
+
+    bool InitAhbBacking() {
+        if (!Settings::values.use_unified_memory.GetValue()) {
+            return false;
+        }
+        static const PFN_AHardwareBuffer_getNativeHandle get_native_handle =
+            ResolveGetNativeHandle();
+        if (get_native_handle == nullptr) {
+            LOG_WARNING(HW_Memory, "AHardwareBuffer_getNativeHandle is not available");
+            return false;
+        }
+        static const PFN_AHardwareBuffer_isSupported is_supported = ResolveIsSupported();
+        if (is_supported == nullptr) {
+            LOG_WARNING(HW_Memory, "AHardwareBuffer_isSupported is not available");
+            return false;
+        }
+        const u64 total_physical = Common::GetMemInfo().TotalPhysicalMemory;
+        if (total_physical != 0 && backing_size > total_physical / 2) {
+            LOG_WARNING(HW_Memory,
+                        "Hardware buffer backing would commit {} MiB on a {} MiB system, keeping "
+                        "lazily committed memory",
+                        backing_size >> 20, total_physical >> 20);
+            return false;
+        }
+        if (!ProbeAhbBacking(get_native_handle)) {
+            return false;
+        }
+        const auto try_window_size = [&](size_t window_size) -> bool {
+            const size_t num_windows = (backing_size + window_size - 1) / window_size;
+            std::vector<AHardwareBuffer*> buffers;
+            std::vector<int> buffer_fds;
+            const auto cleanup = [&] {
+                for (AHardwareBuffer* buffer : buffers) {
+                    AHardwareBuffer_release(buffer);
+                }
+                buffers.clear();
+                buffer_fds.clear();
+            };
+            for (size_t i = 0; i < num_windows; ++i) {
+                const size_t len = (std::min)(window_size, backing_size - i * window_size);
+                const AHardwareBuffer_Desc desc = MakeBlobDesc(len);
+                AHardwareBuffer* buffer{};
+                if (AHardwareBuffer_allocate(&desc, &buffer) != 0 || buffer == nullptr) {
+                    LOG_WARNING(HW_Memory, "Hardware buffer allocation failed for window {}", i);
+                    cleanup();
+                    return false;
+                }
+                buffers.push_back(buffer);
+                const NativeHandle* const handle = get_native_handle(buffer);
+                if (handle == nullptr || handle->numFds < 1) {
+                    LOG_WARNING(HW_Memory, "Hardware buffer has no mappable file descriptor");
+                    cleanup();
+                    return false;
+                }
+                const int buffer_fd = handle->data[0];
+                const off_t buffer_len = lseek(buffer_fd, 0, SEEK_END);
+                if (buffer_len < static_cast<off_t>(len)) {
+                    LOG_WARNING(HW_Memory, "Hardware buffer descriptor smaller than requested");
+                    cleanup();
+                    return false;
+                }
+                buffer_fds.push_back(buffer_fd);
+            }
+            u8* const base =
+                static_cast<u8*>(mmap(nullptr, backing_size, PROT_NONE,
+                                      MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0));
+            if (base == MAP_FAILED) {
+                cleanup();
+                return false;
+            }
+            for (size_t i = 0; i < num_windows; ++i) {
+                const size_t len = (std::min)(window_size, backing_size - i * window_size);
+                if (mmap(base + i * window_size, len, PROT_READ | PROT_WRITE,
+                         MAP_SHARED | MAP_FIXED, buffer_fds[i], 0) == MAP_FAILED) {
+                    LOG_WARNING(HW_Memory, "Hardware buffer mmap failed: {}", strerror(errno));
+                    munmap(base, backing_size);
+                    cleanup();
+                    return false;
+                }
+            }
+            backing_base = base;
+            ahb_windows = std::move(buffers);
+            ahb_fds = std::move(buffer_fds);
+            ahb_window_size = window_size;
+            ahb_backing = true;
+            committed_backing_size.store(backing_size, std::memory_order_relaxed);
+            LOG_INFO(HW_Memory,
+                     "Guest memory backed by {} hardware buffer windows of {} MiB, {} MiB committed",
+                     ahb_windows.size(), window_size >> 20, backing_size >> 20);
+            return true;
+        };
+        static constexpr size_t candidate_window_sizes[] = {
+            1024ULL << 20,
+            512ULL << 20,
+            256ULL << 20,
+            128ULL << 20,
+        };
+        for (const size_t candidate : candidate_window_sizes) {
+            const AHardwareBuffer_Desc window_desc = MakeBlobDesc(candidate);
+            if (is_supported(&window_desc) == 0) {
+                LOG_DEBUG(HW_Memory, "Allocator rejects {} MiB hardware buffer windows",
+                          candidate >> 20);
+                continue;
+            }
+            if (try_window_size(candidate)) {
+                return true;
+            }
+            LOG_WARNING(HW_Memory, "Could not back guest memory with {} MiB windows",
+                        candidate >> 20);
+        }
+        return false;
+    }
+
+    std::span<AHardwareBuffer* const> AhbWindows() const noexcept {
+        return ahb_windows;
+    }
+
+    size_t AhbWindowSize() const noexcept {
+        return ahb_backing ? ahb_window_size : 0;
+    }
+#endif
+
     ~Impl() {
         Release();
     }
@@ -587,6 +821,26 @@ public:
 #ifdef ARCHITECTURE_arm64
         if (True(perms & MemoryPermission::Execute))
             prot_flags |= PROT_EXEC;
+#endif
+#ifdef __ANDROID__
+        if (ahb_backing) {
+            size_t voff = virtual_offset;
+            size_t hoff = host_offset;
+            size_t remaining = length;
+            while (remaining > 0) {
+                const size_t window = hoff / ahb_window_size;
+                const size_t local = hoff % ahb_window_size;
+                const size_t chunk = (std::min)(remaining, ahb_window_size - local);
+                void* const ret =
+                    mmap(virtual_base + voff, chunk, prot_flags, MAP_SHARED | MAP_FIXED,
+                         ahb_fds[window], static_cast<off_t>(local));
+                ASSERT_MSG(ret != MAP_FAILED, "mmap: {}", strerror(errno));
+                voff += chunk;
+                hoff += chunk;
+                remaining -= chunk;
+            }
+            return;
+        }
 #endif
         int flags = (fd >= 0 ? MAP_SHARED : MAP_PRIVATE) | MAP_FIXED;
         void* ret = mmap(virtual_base + virtual_offset, length, prot_flags, flags, fd, host_offset);
@@ -656,6 +910,18 @@ private:
             int ret = close(fd);
             ASSERT_MSG(ret == 0, "close failed: {}", strerror(errno));
         }
+
+#ifdef __ANDROID__
+        for (AHardwareBuffer* buffer : ahb_windows) {
+            AHardwareBuffer_release(buffer);
+        }
+        ahb_windows.clear();
+        ahb_fds.clear();
+        if (ahb_backing) {
+            committed_backing_size.store(0, std::memory_order_relaxed);
+            ahb_backing = false;
+        }
+#endif
     }
 
     void AdjustMap(size_t* virtual_offset, size_t* length) {
@@ -681,6 +947,13 @@ private:
 
     int fd{-1}; // memfd file descriptor, -1 is the error value of memfd_create
     FreeRegionManager free_manager{};
+
+#ifdef __ANDROID__
+    bool ahb_backing{};
+    std::vector<AHardwareBuffer*> ahb_windows;
+    std::vector<int> ahb_fds;
+    size_t ahb_window_size{};
+#endif
 };
 
 #endif // ^^^ POSIX ^^^
@@ -767,6 +1040,22 @@ void HostMemory::ClearBackingRegion(size_t physical_offset, size_t length, u32 f
     std::memset(backing_base + physical_offset, fill_value, length);
 }
 
+std::span<AHardwareBuffer* const> HostMemory::BackingHardwareBuffers() const noexcept {
+#ifdef __ANDROID__
+    return impl ? impl->AhbWindows() : std::span<AHardwareBuffer* const>{};
+#else
+    return {};
+#endif
+}
+
+size_t HostMemory::BackingHardwareBufferWindowSize() const noexcept {
+#ifdef __ANDROID__
+    return impl ? impl->AhbWindowSize() : 0;
+#else
+    return 0;
+#endif
+}
+
 void HostMemory::EnableDirectMappedAddress() {
 #if !(defined(__OPENORBIS__) || defined(__managarm__))
     if (impl) {

src/common/host_memory.h

@@ -8,12 +8,17 @@
 
 #include <memory>
 #include <optional>
+#include <span>
 #include "common/common_funcs.h"
 #include "common/common_types.h"
 #include "common/virtual_buffer.h"
 
+struct AHardwareBuffer;
+
 namespace Common {
 
+[[nodiscard]] u64 GetCommittedBackingSize() noexcept;
+
 enum class MemoryPermission : u32 {
     Read = 1 << 0,
     Write = 1 << 1,
@@ -62,6 +67,14 @@ public:
         return backing_base;
     }
 
+    [[nodiscard]] size_t BackingSize() const noexcept {
+        return backing_size;
+    }
+
+    [[nodiscard]] std::span<AHardwareBuffer* const> BackingHardwareBuffers() const noexcept;
+
+    [[nodiscard]] size_t BackingHardwareBufferWindowSize() const noexcept;
+
     [[nodiscard]] u8* VirtualBasePointer() noexcept {
         return virtual_base;
     }

src/common/settings.h

@@ -587,6 +587,9 @@ struct Values {
     SwitchableSetting<bool> use_asynchronous_shaders{linkage, false, "use_asynchronous_shaders",
                                                      Category::RendererHacks};
 
+    SwitchableSetting<bool> use_unified_memory{linkage, false, "use_unified_memory",
+                                               Category::RendererHacks};
+
     SwitchableSetting<GpuUnswizzleSize> gpu_unswizzle_texture_size{linkage,
                                                   GpuUnswizzleSize::Large,
                                                   "gpu_unswizzle_texture_size",
@@ -635,6 +638,16 @@ struct Values {
 #endif
                                                        "vertex_input_dynamic_state", Category::RendererExtensions};
 
+    SwitchableSetting<bool> dynamic_rendering{linkage, true, "dynamic_rendering",
+                                              Category::RendererExtensions};
+
+    SwitchableSetting<bool> workgroup_memory_explicit_layout{
+        linkage, false, "workgroup_memory_explicit_layout", Category::RendererExtensions};
+
+    SwitchableSetting<s32, true> pipeline_worker_count{
+        linkage, 2, 2, 8, "pipeline_worker_count", Category::RendererAdvanced,
+        Specialization::Scalar};
+
     Setting<bool> renderer_debug{linkage, false, "debug", Category::RendererDebug};
     Setting<bool> renderer_shader_feedback{linkage, false, "shader_feedback",
                                            Category::RendererDebug};

src/common/thread.cpp

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

src/common/thread.h

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

src/common/thread_worker.h

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

src/core/arm/nce/arm_nce.cpp

@@ -157,6 +157,8 @@ bool ArmNce::HandleGuestAlignmentFault(GuestContext* guest_ctx, void* raw_info,
     return HandleFailedGuestFault(guest_ctx, raw_info, raw_context);
 }
 
+constexpr size_t NCE_WRITE_FAULT_CLUSTER_PAGES = 4;
+
 bool ArmNce::HandleGuestAccessFault(GuestContext* guest_ctx, void* raw_info, void* raw_context) {
     auto* info = static_cast<siginfo_t*>(raw_info);
 
@@ -165,7 +167,7 @@ bool ArmNce::HandleGuestAccessFault(GuestContext* guest_ctx, void* raw_info, voi
     const Common::ProcessAddress addr =
         (reinterpret_cast<u64>(info->si_addr) & ~Memory::YUZU_PAGEMASK);
     auto& memory = guest_ctx->parent->m_running_thread->GetOwnerProcess()->GetMemory();
-    if (memory.InvalidateNCE(addr, Memory::YUZU_PAGESIZE)) {
+    if (memory.InvalidateNCE(addr, Memory::YUZU_PAGESIZE * NCE_WRITE_FAULT_CLUSTER_PAGES)) {
         // We handled the access successfully and are returning to guest code.
         return true;
     }

src/core/core.cpp

@@ -118,6 +118,7 @@ struct System::Impl {
 
         is_multicore = Settings::values.use_multi_core.GetValue();
         extended_memory_layout = Settings::values.memory_layout_mode.GetValue() != Settings::MemoryLayout::Memory_4Gb;
+        unified_memory = Settings::values.use_unified_memory.GetValue();
 
         core_timing.SetMulticore(is_multicore);
         core_timing.Initialize([&system]() { system.RegisterHostThread(); });
@@ -145,7 +146,8 @@ struct System::Impl {
             !device_memory.has_value() ||
             is_multicore != Settings::values.use_multi_core.GetValue() ||
             extended_memory_layout != (Settings::values.memory_layout_mode.GetValue() !=
-                                       Settings::MemoryLayout::Memory_4Gb);
+                                       Settings::MemoryLayout::Memory_4Gb) ||
+            unified_memory != Settings::values.use_unified_memory.GetValue();
 
         if (!must_reinitialize) {
             return;
@@ -156,6 +158,7 @@ struct System::Impl {
         is_multicore = Settings::values.use_multi_core.GetValue();
         extended_memory_layout =
             Settings::values.memory_layout_mode.GetValue() != Settings::MemoryLayout::Memory_4Gb;
+        unified_memory = Settings::values.use_unified_memory.GetValue();
 
         Initialize(system);
     }
@@ -503,6 +506,7 @@ struct System::Impl {
     std::atomic_bool is_powered_on{};
     bool is_multicore : 1 = false;
     bool extended_memory_layout : 1 = false;
+    bool unified_memory : 1 = false;
     bool exit_locked : 1 = false;
     bool exit_requested : 1 = false;
     bool nvdec_active : 1 = false;

src/core/cpu_manager.cpp

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

src/core/device_memory_manager.h

@@ -20,6 +20,8 @@
 #include "common/scratch_buffer.h"
 #include "common/virtual_buffer.h"
 
+struct AHardwareBuffer;
+
 namespace Core {
 
 constexpr size_t DEVICE_PAGEBITS = 12ULL;
@@ -95,6 +97,26 @@ public:
         ApplyOpOnPAddr(address, buffer, operation);
     }
 
+    u8* GetPhysicalBase() noexcept {
+        return reinterpret_cast<u8*>(physical_base);
+    }
+
+    const u8* GetPhysicalBase() const noexcept {
+        return reinterpret_cast<const u8*>(physical_base);
+    }
+
+    size_t GetPhysicalSize() const noexcept {
+        return physical_size;
+    }
+
+    std::span<AHardwareBuffer* const> GetBackingHardwareBuffers() const noexcept {
+        return ahb_windows;
+    }
+
+    size_t GetBackingHardwareBufferWindowSize() const noexcept {
+        return ahb_window_size;
+    }
+
     PAddr GetPhysicalRawAddressFromDAddr(DAddr address) const {
         PAddr subbits = PAddr(address & page_mask);
         auto paddr = tracked_entries[(address >> page_bits)].compressed_physical_ptr;
@@ -126,6 +148,10 @@ public:
     // New batch API to update multiple ranges with a single lock acquisition.
     void UpdatePagesCachedBatch(std::span<const std::pair<DAddr, size_t>> ranges, s32 delta);
 
+    void UpdateTexturePagesCount(DAddr addr, size_t size, s32 delta);
+
+    [[nodiscard]] bool IsRegionTextureCached(DAddr addr, size_t size) const noexcept;
+
 private:
     struct TranslationEntry {
         DAddr guest_page{};
@@ -171,6 +197,9 @@ private:
     std::unique_ptr<DeviceMemoryManagerAllocator<Traits>> impl;
 
     const uintptr_t physical_base;
+    const size_t physical_size;
+    const std::span<AHardwareBuffer* const> ahb_windows;
+    const size_t ahb_window_size;
     DeviceInterface* device_inter;
 
     struct TrackedEntry {
@@ -234,6 +263,7 @@ private:
         (1ULL << (device_virtual_bits - page_bits)) / subentries;
     using CachedPages = std::array<CounterEntry, num_counter_entries>;
     std::unique_ptr<CachedPages> cached_pages;
+    std::unique_ptr<CachedPages> texture_cached_pages;
     Common::RangeMutex counter_guard;
     std::mutex mapping_guard;
 

src/core/device_memory_manager.inc

@@ -171,12 +171,16 @@ struct DeviceMemoryManagerAllocator {
 template <typename Traits>
 DeviceMemoryManager<Traits>::DeviceMemoryManager(const DeviceMemory& device_memory_)
     : physical_base{uintptr_t(device_memory_.buffer.BackingBasePointer())}
+    , physical_size{device_memory_.buffer.BackingSize()}
+    , ahb_windows{device_memory_.buffer.BackingHardwareBuffers()}
+    , ahb_window_size{device_memory_.buffer.BackingHardwareBufferWindowSize()}
     , device_inter{nullptr}
     , compressed_device_addr(1ULL << ((Settings::values.memory_layout_mode.GetValue() == Settings::MemoryLayout::Memory_4Gb ? physical_min_bits : physical_max_bits) - Memory::YUZU_PAGEBITS))
     , tracked_entries(device_as_size >> Memory::YUZU_PAGEBITS)
 {
     impl = std::make_unique<DeviceMemoryManagerAllocator<Traits>>();
     cached_pages = std::make_unique<CachedPages>();
+    texture_cached_pages = std::make_unique<CachedPages>();
 
     const size_t total_virtual = device_as_size >> Memory::YUZU_PAGEBITS;
     for (size_t i = 0; i < total_virtual; i++) {
@@ -625,6 +629,28 @@ void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size
     UpdatePagesCachedCountNoLock(addr, size, delta);
 }
 
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UpdateTexturePagesCount(DAddr addr, size_t size, s32 delta) {
+    Common::ScopedRangeLock lk(counter_guard, addr, size);
+    const size_t page_end = Common::DivCeil(addr + size, Memory::YUZU_PAGESIZE);
+    for (size_t page = addr >> Memory::YUZU_PAGEBITS; page != page_end; ++page) {
+        CounterAtomicType& count = texture_cached_pages->at(page >> subentries_shift).Count(page);
+        count.fetch_add(static_cast<CounterType>(delta), std::memory_order_release);
+    }
+}
+
+template <typename Traits>
+bool DeviceMemoryManager<Traits>::IsRegionTextureCached(DAddr addr, size_t size) const noexcept {
+    const size_t page_end = Common::DivCeil(addr + size, Memory::YUZU_PAGESIZE);
+    for (size_t page = addr >> Memory::YUZU_PAGEBITS; page != page_end; ++page) {
+        if (texture_cached_pages->at(page >> subentries_shift).Count(page).load(
+                std::memory_order_acquire) != 0) {
+            return true;
+        }
+    }
+    return false;
+}
+
 template <typename Traits>
 void DeviceMemoryManager<Traits>::UpdatePagesCachedBatch(std::span<const std::pair<DAddr, size_t>> ranges, s32 delta) {
     if (ranges.empty()) {

src/qt_common/config/shared_translation.cpp

@@ -230,6 +230,11 @@ std::unique_ptr<TranslationMap> InitializeTranslations(QObject* parent) {
            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, use_unified_memory, tr("Enable unified memory access (UMA)"),
+           tr("Lets the GPU write buffer readbacks directly into guest memory."));
+    INSERT(Settings, pipeline_worker_count, tr("Pipeline Worker Threads"),
+           tr("Number of threads used to build Vulkan pipelines.\n"
+              "Higher values speed up compilation at the cost of heat and power."));
     INSERT(Settings, fast_gpu_time, tr("Fast GPU Time"),
            tr("Overclocks the emulated GPU to increase dynamic resolution and render "
               "distance.\nUse 256 for maximal performance and 512 for maximal graphics fidelity."));
@@ -287,6 +292,12 @@ std::unique_ptr<TranslationMap> InitializeTranslations(QObject* parent) {
     INSERT(Settings, vertex_input_dynamic_state, tr("Vertex Input Dynamic State"),
            tr("Enables vertex input dynamic state feature for better quality and performance."));
 
+    INSERT(Settings, dynamic_rendering, tr("Dynamic Rendering"),
+           tr("Renders without render pass and framebuffer objects.\n"
+              "Results vary by driver: some gain performance, others lose it."));
+
+    INSERT(Settings, workgroup_memory_explicit_layout, QString(), QString());
+
     INSERT(
         Settings, sample_shading, tr("Sample Shading"),
         tr("Allows the fragment shader to execute per sample in a multi-sampled fragment "

src/shader_recompiler/backend/glasm/emit_glasm_instructions.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
@@ -665,6 +665,8 @@ void EmitShuffleDown(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU3
                      const IR::Value& clamp, const IR::Value& segmentation_mask);
 void EmitShuffleButterfly(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 index,
                           const IR::Value& clamp, const IR::Value& segmentation_mask);
+void EmitQuadBroadcast(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 lane);
+void EmitQuadSwap(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 direction);
 void EmitFSwizzleAdd(EmitContext& ctx, IR::Inst& inst, ScalarF32 op_a, ScalarF32 op_b,
                      ScalarU32 swizzle);
 void EmitDPdxFine(EmitContext& ctx, IR::Inst& inst, ScalarF32 op_a);

src/shader_recompiler/backend/glasm/emit_glasm_warp.cpp

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -97,6 +100,24 @@ void EmitShuffleButterfly(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, Sca
     Shuffle(ctx, inst, value, index, clamp, segmentation_mask, "XOR");
 }
 
+void EmitQuadBroadcast(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 lane) {
+    const Register ret{ctx.reg_alloc.Define(inst)};
+    ctx.Add("AND.U RC.x,{}.threadid,~3;"
+            "AND.U RC.y,{},3;"
+            "OR.U RC.x,RC.x,RC.y;"
+            "SHFIDX.U {},{},RC.x,0x1C03;"
+            "MOV.U {}.x,{}.y;",
+            ctx.stage_name, lane, ret, value, ret, ret);
+}
+
+void EmitQuadSwap(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 direction) {
+    const Register ret{ctx.reg_alloc.Define(inst)};
+    ctx.Add("ADD.U RC.x,{},1;"
+            "SHFXOR.U {},{},RC.x,0x1C03;"
+            "MOV.U {}.x,{}.y;",
+            direction, ret, value, ret, ret);
+}
+
 void EmitFSwizzleAdd(EmitContext& ctx, IR::Inst& inst, ScalarF32 op_a, ScalarF32 op_b,
                      ScalarU32 swizzle) {
     const auto ret{ctx.reg_alloc.Define(inst)};