25 files changed, 363 insertions, 45 deletions

diff --git a/src/common/uuid.h b/src/common/uuid.h
index 2e7a18405..0ffa37e7c 100644
--- a/src/common/uuid.h
+++ b/src/common/uuid.h
@@ -20,12 +20,11 @@ struct UUID {
     constexpr explicit UUID(const u64 lo, const u64 hi) : uuid{{lo, hi}} {}
 
     [[nodiscard]] constexpr explicit operator bool() const {
-        return uuid[0] != INVALID_UUID[0] && uuid[1] != INVALID_UUID[1];
+        return uuid != INVALID_UUID;
     }
 
     [[nodiscard]] constexpr bool operator==(const UUID& rhs) const {
-        // TODO(DarkLordZach): Replace with uuid == rhs.uuid with C++20
-        return uuid[0] == rhs.uuid[0] && uuid[1] == rhs.uuid[1];
+        return uuid == rhs.uuid;
     }
 
     [[nodiscard]] constexpr bool operator!=(const UUID& rhs) const {
diff --git a/src/core/hle/service/am/applets/applet_controller.cpp b/src/core/hle/service/am/applets/applet_controller.cpp
index 12682effe..2721679c1 100644
--- a/src/core/hle/service/am/applets/applet_controller.cpp
+++ b/src/core/hle/service/am/applets/applet_controller.cpp
@@ -87,6 +87,10 @@ void Controller::Initialize() {
         case sizeof(ControllerUpdateFirmwareArg):
             controller_private_arg.mode = ControllerSupportMode::ShowControllerFirmwareUpdate;
             break;
+        case sizeof(ControllerKeyRemappingArg):
+            controller_private_arg.mode =
+                ControllerSupportMode::ShowControllerKeyRemappingForSystem;
+            break;
         default:
             UNIMPLEMENTED_MSG("Unknown ControllerPrivateArg mode={} with arg_size={}",
                               controller_private_arg.mode, controller_private_arg.arg_size);
@@ -99,7 +103,9 @@ void Controller::Initialize() {
     // This is always 0 (Application) except with ShowControllerFirmwareUpdateForSystem.
     if (controller_private_arg.caller >= ControllerSupportCaller::MaxControllerSupportCaller) {
         if (controller_private_arg.flag_1 &&
-            controller_private_arg.mode == ControllerSupportMode::ShowControllerFirmwareUpdate) {
+            (controller_private_arg.mode == ControllerSupportMode::ShowControllerFirmwareUpdate ||
+             controller_private_arg.mode ==
+                 ControllerSupportMode::ShowControllerKeyRemappingForSystem)) {
             controller_private_arg.caller = ControllerSupportCaller::System;
         } else {
             controller_private_arg.caller = ControllerSupportCaller::Application;
@@ -121,6 +127,7 @@ void Controller::Initialize() {
             std::memcpy(&controller_user_arg_old, user_arg.data(), user_arg.size());
             break;
         case ControllerAppletVersion::Version7:
+        case ControllerAppletVersion::Version8:
             ASSERT(user_arg.size() == sizeof(ControllerSupportArgNew));
             std::memcpy(&controller_user_arg_new, user_arg.data(), user_arg.size());
             break;
@@ -143,6 +150,16 @@ void Controller::Initialize() {
         std::memcpy(&controller_update_arg, update_arg.data(), update_arg.size());
         break;
     }
+    case ControllerSupportMode::ShowControllerKeyRemappingForSystem: {
+        const auto remapping_arg_storage = broker.PopNormalDataToApplet();
+        ASSERT(remapping_arg_storage != nullptr);
+
+        const auto& remapping_arg = remapping_arg_storage->GetData();
+        ASSERT(remapping_arg.size() == sizeof(ControllerKeyRemappingArg));
+
+        std::memcpy(&controller_key_remapping_arg, remapping_arg.data(), remapping_arg.size());
+        break;
+    }
     default: {
         UNIMPLEMENTED_MSG("Unimplemented ControllerSupportMode={}", controller_private_arg.mode);
         break;
@@ -179,6 +196,7 @@ void Controller::Execute() {
                     std::vector<ExplainText>(controller_user_arg_old.explain_text.begin(),
                                              controller_user_arg_old.explain_text.end()));
             case ControllerAppletVersion::Version7:
+            case ControllerAppletVersion::Version8:
             default:
                 return ConvertToFrontendParameters(
                     controller_private_arg, controller_user_arg_new.header,
@@ -210,6 +228,7 @@ void Controller::Execute() {
     }
     case ControllerSupportMode::ShowControllerStrapGuide:
     case ControllerSupportMode::ShowControllerFirmwareUpdate:
+    case ControllerSupportMode::ShowControllerKeyRemappingForSystem:
         UNIMPLEMENTED_MSG("ControllerSupportMode={} is not implemented",
                           controller_private_arg.mode);
         ConfigurationComplete();
diff --git a/src/core/hle/service/am/applets/applet_controller.h b/src/core/hle/service/am/applets/applet_controller.h
index 20617e91f..0a34c4fc0 100644
--- a/src/core/hle/service/am/applets/applet_controller.h
+++ b/src/core/hle/service/am/applets/applet_controller.h
@@ -25,13 +25,15 @@ enum class ControllerAppletVersion : u32_le {
     Version3 = 0x3, // 1.0.0 - 2.3.0
     Version4 = 0x4, // 3.0.0 - 5.1.0
     Version5 = 0x5, // 6.0.0 - 7.0.1
-    Version7 = 0x7, // 8.0.0+
+    Version7 = 0x7, // 8.0.0 - 10.2.0
+    Version8 = 0x8, // 11.0.0+
 };
 
 enum class ControllerSupportMode : u8 {
     ShowControllerSupport,
     ShowControllerStrapGuide,
     ShowControllerFirmwareUpdate,
+    ShowControllerKeyRemappingForSystem,
 
     MaxControllerSupportMode,
 };
@@ -78,7 +80,7 @@ struct ControllerSupportArgOld {
 static_assert(sizeof(ControllerSupportArgOld) == 0x21C,
               "ControllerSupportArgOld has incorrect size.");
 
-// LibraryAppletVersion 0x7
+// LibraryAppletVersion 0x7, 0x8
 struct ControllerSupportArgNew {
     ControllerSupportArgHeader header{};
     std::array<IdentificationColor, 8> identification_colors{};
@@ -95,6 +97,14 @@ struct ControllerUpdateFirmwareArg {
 static_assert(sizeof(ControllerUpdateFirmwareArg) == 0x4,
               "ControllerUpdateFirmwareArg has incorrect size.");
 
+struct ControllerKeyRemappingArg {
+    u64 unknown{};
+    u32 unknown_2{};
+    INSERT_PADDING_WORDS(1);
+};
+static_assert(sizeof(ControllerKeyRemappingArg) == 0x10,
+              "ControllerKeyRemappingArg has incorrect size.");
+
 struct ControllerSupportResultInfo {
     s8 player_count{};
     INSERT_PADDING_BYTES(3);
@@ -128,6 +138,7 @@ private:
     ControllerSupportArgOld controller_user_arg_old;
     ControllerSupportArgNew controller_user_arg_new;
     ControllerUpdateFirmwareArg controller_update_arg;
+    ControllerKeyRemappingArg controller_key_remapping_arg;
     bool complete{false};
     ResultCode status{ResultSuccess};
     bool is_single_mode{false};
diff --git a/src/input_common/sdl/sdl_impl.cpp b/src/input_common/sdl/sdl_impl.cpp
index 1656b85fb..70a0ba09c 100644
--- a/src/input_common/sdl/sdl_impl.cpp
+++ b/src/input_common/sdl/sdl_impl.cpp
@@ -115,6 +115,41 @@ public:
         return state.buttons.at(button);
     }
 
+    bool ToggleButton(int button) {
+        std::lock_guard lock{mutex};
+
+        if (!state.toggle_buttons.contains(button) || !state.lock_buttons.contains(button)) {
+            state.toggle_buttons.insert_or_assign(button, false);
+            state.lock_buttons.insert_or_assign(button, false);
+        }
+
+        const bool button_state = state.toggle_buttons.at(button);
+        const bool button_lock = state.lock_buttons.at(button);
+
+        if (button_lock) {
+            return button_state;
+        }
+
+        state.lock_buttons.insert_or_assign(button, true);
+
+        if (button_state) {
+            state.toggle_buttons.insert_or_assign(button, false);
+        } else {
+            state.toggle_buttons.insert_or_assign(button, true);
+        }
+
+        return !button_state;
+    }
+
+    bool UnlockButton(int button) {
+        std::lock_guard lock{mutex};
+        if (!state.toggle_buttons.contains(button)) {
+            return false;
+        }
+        state.lock_buttons.insert_or_assign(button, false);
+        return state.toggle_buttons.at(button);
+    }
+
     void SetAxis(int axis, Sint16 value) {
         std::lock_guard lock{mutex};
         state.axes.insert_or_assign(axis, value);
@@ -241,6 +276,8 @@ public:
 private:
     struct State {
         std::unordered_map<int, bool> buttons;
+        std::unordered_map<int, bool> toggle_buttons{};
+        std::unordered_map<int, bool> lock_buttons{};
         std::unordered_map<int, Sint16> axes;
         std::unordered_map<int, Uint8> hats;
     } state;
@@ -402,16 +439,25 @@ void SDLState::CloseJoysticks() {
 
 class SDLButton final : public Input::ButtonDevice {
 public:
-    explicit SDLButton(std::shared_ptr<SDLJoystick> joystick_, int button_)
-        : joystick(std::move(joystick_)), button(button_) {}
+    explicit SDLButton(std::shared_ptr<SDLJoystick> joystick_, int button_, bool toggle_)
+        : joystick(std::move(joystick_)), button(button_), toggle(toggle_) {}
 
     bool GetStatus() const override {
-        return joystick->GetButton(button);
+        const bool button_state = joystick->GetButton(button);
+        if (!toggle) {
+            return button_state;
+        }
+
+        if (button_state) {
+            return joystick->ToggleButton(button);
+        }
+        return joystick->UnlockButton(button);
     }
 
 private:
     std::shared_ptr<SDLJoystick> joystick;
     int button;
+    bool toggle;
 };
 
 class SDLDirectionButton final : public Input::ButtonDevice {
@@ -635,6 +681,7 @@ public:
     std::unique_ptr<Input::ButtonDevice> Create(const Common::ParamPackage& params) override {
         const std::string guid = params.Get("guid", "0");
         const int port = params.Get("port", 0);
+        const auto toggle = params.Get("toggle", false);
 
         auto joystick = state.GetSDLJoystickByGUID(guid, port);
 
@@ -660,7 +707,8 @@ public:
 
         if (params.Has("axis")) {
             const int axis = params.Get("axis", 0);
-            const float threshold = params.Get("threshold", 0.5f);
+            // Convert range from (0.0, 1.0) to (-1.0, 1.0)
+            const float threshold = (params.Get("threshold", 0.5f) - 0.5f) * 2.0f;
             const std::string direction_name = params.Get("direction", "");
             bool trigger_if_greater;
             if (direction_name == "+") {
@@ -679,7 +727,7 @@ public:
         const int button = params.Get("button", 0);
         // This is necessary so accessing GetButton with button won't crash
         joystick->SetButton(button, false);
-        return std::make_unique<SDLButton>(joystick, button);
+        return std::make_unique<SDLButton>(joystick, button, toggle);
     }
 
 private:
@@ -933,12 +981,11 @@ Common::ParamPackage BuildAnalogParamPackageForButton(int port, std::string guid
     params.Set("port", port);
     params.Set("guid", std::move(guid));
     params.Set("axis", axis);
+    params.Set("threshold", "0.5");
     if (value > 0) {
         params.Set("direction", "+");
-        params.Set("threshold", "0.5");
     } else {
         params.Set("direction", "-");
-        params.Set("threshold", "-0.5");
     }
     return params;
 }
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 2871682f6..7373cb62d 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -164,11 +164,16 @@ public:
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
 
-    [[nodiscard]] bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount);
+    bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount);
+
+    bool DMAClear(GPUVAddr src_address, u64 amount, u32 value);
 
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
+    /// Return true when a region is registered on the cache
+    [[nodiscard]] bool IsRegionRegistered(VAddr addr, size_t size);
+
     /// Return true when a CPU region is modified from the CPU
     [[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size);
 
@@ -324,6 +329,8 @@ private:
 
     [[nodiscard]] bool HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept;
 
+    void ClearDownload(IntervalType subtract_interval);
+
     VideoCore::RasterizerInterface& rasterizer;
     Tegra::Engines::Maxwell3D& maxwell3d;
     Tegra::Engines::KeplerCompute& kepler_compute;
@@ -463,23 +470,28 @@ void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) {
 }
 
 template <class P>
+void BufferCache<P>::ClearDownload(IntervalType subtract_interval) {
+    uncommitted_ranges.subtract(subtract_interval);
+    for (auto& interval_set : committed_ranges) {
+        interval_set.subtract(subtract_interval);
+    }
+}
+
+template <class P>
 bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) {
     const std::optional<VAddr> cpu_src_address = gpu_memory.GpuToCpuAddress(src_address);
     const std::optional<VAddr> cpu_dest_address = gpu_memory.GpuToCpuAddress(dest_address);
     if (!cpu_src_address || !cpu_dest_address) {
         return false;
     }
-    const bool source_dirty = IsRegionGpuModified(*cpu_src_address, amount);
-    const bool dest_dirty = IsRegionGpuModified(*cpu_dest_address, amount);
+    const bool source_dirty = IsRegionRegistered(*cpu_src_address, amount);
+    const bool dest_dirty = IsRegionRegistered(*cpu_dest_address, amount);
     if (!source_dirty && !dest_dirty) {
         return false;
     }
 
     const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount};
-    uncommitted_ranges.subtract(subtract_interval);
-    for (auto& interval_set : committed_ranges) {
-        interval_set.subtract(subtract_interval);
-    }
+    ClearDownload(subtract_interval);
 
     BufferId buffer_a;
     BufferId buffer_b;
@@ -510,12 +522,13 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     ForEachWrittenRange(*cpu_src_address, amount, mirror);
     // This subtraction in this order is important for overlapping copies.
     common_ranges.subtract(subtract_interval);
+    bool atleast_1_download = tmp_intervals.size() != 0;
     for (const IntervalType add_interval : tmp_intervals) {
         common_ranges.add(add_interval);
     }
 
     runtime.CopyBuffer(dest_buffer, src_buffer, copies);
-    if (source_dirty) {
+    if (atleast_1_download) {
         dest_buffer.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
     std::vector<u8> tmp_buffer(amount);
@@ -525,6 +538,33 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
 }
 
 template <class P>
+bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
+    const std::optional<VAddr> cpu_dst_address = gpu_memory.GpuToCpuAddress(dst_address);
+    if (!cpu_dst_address) {
+        return false;
+    }
+    const bool dest_dirty = IsRegionRegistered(*cpu_dst_address, amount);
+    if (!dest_dirty) {
+        return false;
+    }
+
+    const size_t size = amount * sizeof(u32);
+    const IntervalType subtract_interval{*cpu_dst_address, *cpu_dst_address + size};
+    ClearDownload(subtract_interval);
+    common_ranges.subtract(subtract_interval);
+
+    BufferId buffer;
+    do {
+        has_deleted_buffers = false;
+        buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size));
+    } while (has_deleted_buffers);
+    auto& dest_buffer = slot_buffers[buffer];
+    const u32 offset = static_cast<u32>(*cpu_dst_address - dest_buffer.CpuAddr());
+    runtime.ClearBuffer(dest_buffer, offset, size, value);
+    return true;
+}
+
+template <class P>
 void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
                                                u32 size) {
     const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
@@ -782,6 +822,27 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
 }
 
 template <class P>
+bool BufferCache<P>::IsRegionRegistered(VAddr addr, size_t size) {
+    const VAddr end_addr = addr + size;
+    const u64 page_end = Common::DivCeil(end_addr, PAGE_SIZE);
+    for (u64 page = addr >> PAGE_BITS; page < page_end;) {
+        const BufferId buffer_id = page_table[page];
+        if (!buffer_id) {
+            ++page;
+            continue;
+        }
+        Buffer& buffer = slot_buffers[buffer_id];
+        const VAddr buf_start_addr = buffer.CpuAddr();
+        const VAddr buf_end_addr = buf_start_addr + buffer.SizeBytes();
+        if (buf_start_addr < end_addr && addr < buf_end_addr) {
+            return true;
+        }
+        page = Common::DivCeil(end_addr, PAGE_SIZE);
+    }
+    return false;
+}
+
+template <class P>
 bool BufferCache<P>::IsRegionCpuModified(VAddr addr, size_t size) {
     const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE);
     for (u64 page = addr >> PAGE_BITS; page < page_end;) {
@@ -1425,6 +1486,7 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
         const VAddr end_address = start_address + range_size;
         ForEachWrittenRange(start_address, range_size, add_download);
         const IntervalType subtract_interval{start_address, end_address};
+        ClearDownload(subtract_interval);
         common_ranges.subtract(subtract_interval);
     });
     if (total_size_bytes == 0) {
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 24481952b..c51776466 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -4,6 +4,7 @@
 
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -12,6 +13,9 @@
 #include "video_core/renderer_base.h"
 #include "video_core/textures/decoders.h"
 
+MICROPROFILE_DECLARE(GPU_DMAEngine);
+MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128));
+
 namespace Tegra::Engines {
 
 using namespace Texture;
@@ -43,6 +47,7 @@ void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
 }
 
 void MaxwellDMA::Launch() {
+    MICROPROFILE_SCOPE(GPU_DMAEngine);
     LOG_TRACE(Render_OpenGL, "DMA copy 0x{:x} -> 0x{:x}", static_cast<GPUVAddr>(regs.offset_in),
               static_cast<GPUVAddr>(regs.offset_out));
 
@@ -87,9 +92,11 @@ void MaxwellDMA::CopyPitchToPitch() {
         // TODO: allow multisized components.
         if (is_buffer_clear) {
             ASSERT(regs.remap_const.component_size_minus_one == 3);
+            accelerate.BufferClear(regs.offset_out, regs.line_length_in, regs.remap_consta_value);
             std::vector<u32> tmp_buffer(regs.line_length_in, regs.remap_consta_value);
-            memory_manager.WriteBlock(regs.offset_out, reinterpret_cast<u8*>(tmp_buffer.data()),
-                                      regs.line_length_in * sizeof(u32));
+            memory_manager.WriteBlockUnsafe(regs.offset_out,
+                                            reinterpret_cast<u8*>(tmp_buffer.data()),
+                                            regs.line_length_in * sizeof(u32));
             return;
         }
         UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
@@ -179,8 +186,13 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
         write_buffer.resize(dst_size);
     }
 
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
-    memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    if (Settings::IsGPULevelExtreme()) {
+        memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    } else {
+        memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
+    }
 
     // If the input is linear and the output is tiled, swizzle the input and copy it over.
     if (regs.dst_params.block_size.depth > 0) {
diff --git a/src/video_core/engines/maxwell_dma.h b/src/video_core/engines/maxwell_dma.h
index 4ed0d0996..d3329b0f8 100644
--- a/src/video_core/engines/maxwell_dma.h
+++ b/src/video_core/engines/maxwell_dma.h
@@ -31,6 +31,8 @@ class AccelerateDMAInterface {
 public:
     /// Write the value to the register identified by method.
     virtual bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) = 0;
+
+    virtual bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) = 0;
 };
 
 /**
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.cpp b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
index c225d1fc9..c4189fb60 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
@@ -98,6 +98,12 @@ void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
     }
 }
 
+void BufferCacheRuntime::ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value) {
+    glClearNamedBufferSubData(dest_buffer.Handle(), GL_R32UI, static_cast<GLintptr>(offset),
+                              static_cast<GLsizeiptr>(size / sizeof(u32)), GL_RGBA, GL_UNSIGNED_INT,
+                              &value);
+}
+
 void BufferCacheRuntime::BindIndexBuffer(Buffer& buffer, u32 offset, u32 size) {
     if (has_unified_vertex_buffers) {
         buffer.MakeResident(GL_READ_ONLY);
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.h b/src/video_core/renderer_opengl/gl_buffer_cache.h
index d8b20a9af..fe91aa452 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.h
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.h
@@ -57,6 +57,8 @@ public:
     void CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
                     std::span<const VideoCommon::BufferCopy> copies);
 
+    void ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value);
+
     void BindIndexBuffer(Buffer& buffer, u32 offset, u32 size);
 
     void BindVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size, u32 stride);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 82c84127a..ceb3abcb2 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -1407,4 +1407,9 @@ bool AccelerateDMA::BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64
     return buffer_cache.DMACopy(src_address, dest_address, amount);
 }
 
+bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) {
+    std::scoped_lock lock{buffer_cache.mutex};
+    return buffer_cache.DMAClear(src_address, amount, value);
+}
+
 } // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index ccee9ba33..d30ad698f 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -65,6 +65,8 @@ public:
 
     bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) override;
 
+    bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
+
 private:
     BufferCache& buffer_cache;
 };
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.h b/src/video_core/renderer_opengl/gl_texture_cache.h
index 25fe61566..cf3b789e3 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.h
+++ b/src/video_core/renderer_opengl/gl_texture_cache.h
@@ -122,7 +122,7 @@ private:
     bool has_broken_texture_view_formats = false;
 
     StagingBuffers upload_buffers{GL_MAP_WRITE_BIT, GL_MAP_WRITE_BIT | GL_MAP_FLUSH_EXPLICIT_BIT};
-    StagingBuffers download_buffers{GL_MAP_READ_BIT, GL_MAP_READ_BIT};
+    StagingBuffers download_buffers{GL_MAP_READ_BIT | GL_CLIENT_STORAGE_BIT, GL_MAP_READ_BIT};
 
     OGLTexture null_image_1d_array;
     OGLTexture null_image_cube_array;
diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
index 7d4e6ea7b..0def1e769 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
@@ -136,6 +136,30 @@ void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
     });
 }
 
+void BufferCacheRuntime::ClearBuffer(VkBuffer dest_buffer, u32 offset, size_t size, u32 value) {
+    static constexpr VkMemoryBarrier READ_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
+    };
+    static constexpr VkMemoryBarrier WRITE_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+    };
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([dest_buffer, offset, size, value](vk::CommandBuffer cmdbuf) {
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, READ_BARRIER);
+        cmdbuf.FillBuffer(dest_buffer, offset, size, value);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               0, WRITE_BARRIER);
+    });
+}
+
 void BufferCacheRuntime::BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format,
                                          u32 base_vertex, u32 num_indices, VkBuffer buffer,
                                          u32 offset, [[maybe_unused]] u32 size) {
diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.h b/src/video_core/renderer_vulkan/vk_buffer_cache.h
index 6ea8448d7..3bb81d5b3 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.h
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.h
@@ -60,6 +60,8 @@ public:
     void CopyBuffer(VkBuffer src_buffer, VkBuffer dst_buffer,
                     std::span<const VideoCommon::BufferCopy> copies);
 
+    void ClearBuffer(VkBuffer dest_buffer, u32 offset, size_t size, u32 value);
+
     void BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format, u32 num_indices,
                          u32 base_vertex, VkBuffer buffer, u32 offset, u32 size);
 
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index a8ffbe6ba..f57c15b37 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -706,6 +706,11 @@ void RasterizerVulkan::FlushWork() {
 
 AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_) : buffer_cache{buffer_cache_} {}
 
+bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) {
+    std::scoped_lock lock{buffer_cache.mutex};
+    return buffer_cache.DMAClear(src_address, amount, value);
+}
+
 bool AccelerateDMA::BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) {
     std::scoped_lock lock{buffer_cache.mutex};
     return buffer_cache.DMACopy(src_address, dest_address, amount);
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 3a78de258..2065209be 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -56,6 +56,8 @@ public:
 
     bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
 
+    bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
+
 private:
     BufferCache& buffer_cache;
 };
diff --git a/src/yuzu/bootmanager.cpp b/src/yuzu/bootmanager.cpp
index bfae73b60..d72ca5acc 100644
--- a/src/yuzu/bootmanager.cpp
+++ b/src/yuzu/bootmanager.cpp
@@ -567,6 +567,12 @@ std::unique_ptr<Core::Frontend::GraphicsContext> GRenderWindow::CreateSharedCont
 bool GRenderWindow::InitRenderTarget() {
     ReleaseRenderTarget();
 
+    {
+        // Create a dummy render widget so that Qt
+        // places the render window at the correct position.
+        const RenderWidget dummy_widget{this};
+    }
+
     first_frame = false;
 
     switch (Settings::values.renderer_backend.GetValue()) {
diff --git a/src/yuzu/configuration/configure_input_player.cpp b/src/yuzu/configuration/configure_input_player.cpp
index d5d624b96..6b9bd05f1 100644
--- a/src/yuzu/configuration/configure_input_player.cpp
+++ b/src/yuzu/configuration/configure_input_player.cpp
@@ -149,8 +149,9 @@ QString ButtonToText(const Common::ParamPackage& param) {
 
         if (param.Has("button")) {
             const QString button_str = QString::fromStdString(param.Get("button", ""));
+            const QString toggle = QString::fromStdString(param.Get("toggle", false) ? "~" : "");
 
-            return QObject::tr("Button %1").arg(button_str);
+            return QObject::tr("%1Button %2").arg(toggle, button_str);
         }
 
         if (param.Has("motion")) {
@@ -313,6 +314,24 @@ ConfigureInputPlayer::ConfigureInputPlayer(QWidget* parent, std::size_t player_i
                         buttons_param[button_id].Set("toggle", toggle_value);
                         button_map[button_id]->setText(ButtonToText(buttons_param[button_id]));
                     });
+                    if (buttons_param[button_id].Has("threshold")) {
+                        context_menu.addAction(tr("Set threshold"), [&] {
+                            const int button_threshold = static_cast<int>(
+                                buttons_param[button_id].Get("threshold", 0.5f) * 100.0f);
+                            const int new_threshold = QInputDialog::getInt(
+                                this, tr("Set threshold"), tr("Choose a value between 0% and 100%"),
+                                button_threshold, 0, 100);
+                            buttons_param[button_id].Set("threshold", new_threshold / 100.0f);
+
+                            if (button_id == Settings::NativeButton::ZL) {
+                                ui->sliderZLThreshold->setValue(new_threshold);
+                            }
+                            if (button_id == Settings::NativeButton::ZR) {
+                                ui->sliderZRThreshold->setValue(new_threshold);
+                            }
+                        });
+                    }
+
                     context_menu.exec(button_map[button_id]->mapToGlobal(menu_location));
                     ui->controllerFrame->SetPlayerInput(player_index, buttons_param, analogs_param);
                 });
@@ -341,6 +360,20 @@ ConfigureInputPlayer::ConfigureInputPlayer(QWidget* parent, std::size_t player_i
                 });
     }
 
+    connect(ui->sliderZLThreshold, &QSlider::valueChanged, [=, this] {
+        if (buttons_param[Settings::NativeButton::ZL].Has("threshold")) {
+            const auto slider_value = ui->sliderZLThreshold->value();
+            buttons_param[Settings::NativeButton::ZL].Set("threshold", slider_value / 100.0f);
+        }
+    });
+
+    connect(ui->sliderZRThreshold, &QSlider::valueChanged, [=, this] {
+        if (buttons_param[Settings::NativeButton::ZR].Has("threshold")) {
+            const auto slider_value = ui->sliderZRThreshold->value();
+            buttons_param[Settings::NativeButton::ZR].Set("threshold", slider_value / 100.0f);
+        }
+    });
+
     for (int analog_id = 0; analog_id < Settings::NativeAnalog::NumAnalogs; ++analog_id) {
         for (int sub_button_id = 0; sub_button_id < ANALOG_SUB_BUTTONS_NUM; ++sub_button_id) {
             auto* const analog_button = analog_map_buttons[analog_id][sub_button_id];
@@ -849,6 +882,18 @@ void ConfigureInputPlayer::UpdateUI() {
         button_map[button]->setText(ButtonToText(buttons_param[button]));
     }
 
+    if (buttons_param[Settings::NativeButton::ZL].Has("threshold")) {
+        const int button_threshold = static_cast<int>(
+            buttons_param[Settings::NativeButton::ZL].Get("threshold", 0.5f) * 100.0f);
+        ui->sliderZLThreshold->setValue(button_threshold);
+    }
+
+    if (buttons_param[Settings::NativeButton::ZR].Has("threshold")) {
+        const int button_threshold = static_cast<int>(
+            buttons_param[Settings::NativeButton::ZR].Get("threshold", 0.5f) * 100.0f);
+        ui->sliderZRThreshold->setValue(button_threshold);
+    }
+
     for (int motion_id = 0; motion_id < Settings::NativeMotion::NumMotions; ++motion_id) {
         motion_map[motion_id]->setText(ButtonToText(motions_param[motion_id]));
     }
diff --git a/src/yuzu/configuration/configure_input_player.ui b/src/yuzu/configuration/configure_input_player.ui
index e76aa484f..e7433912b 100644
--- a/src/yuzu/configuration/configure_input_player.ui
+++ b/src/yuzu/configuration/configure_input_player.ui
@@ -1334,6 +1334,12 @@
                 </item>
                 <item>
                  <widget class="QGroupBox" name="buttonShoulderButtonsButtonZLGroup">
+                   <property name="sizePolicy">
+                     <sizepolicy hsizetype="Maximum" vsizetype="Preferred">
+                       <horstretch>0</horstretch>
+                       <verstretch>0</verstretch>
+                     </sizepolicy>
+                   </property>
                   <property name="title">
                    <string>ZL</string>
                   </property>
@@ -1378,6 +1384,22 @@
                      </property>
                     </widget>
                    </item>
+                    <item>
+                      <widget class="QSlider" name="sliderZLThreshold">
+                        <property name="maximumSize">
+                          <size>
+                            <width>70</width>
+                            <height>15</height>
+                          </size>
+                        </property>
+                        <property name="maximum">
+                          <number>100</number>
+                        </property>
+                        <property name="orientation">
+                          <enum>Qt::Horizontal</enum>
+                        </property>
+                      </widget>
+                    </item>
                   </layout>
                  </widget>
                 </item>
@@ -1759,6 +1781,12 @@
                 </item>
                 <item>
                  <widget class="QGroupBox" name="buttonShoulderButtonsZRGroup">
+                   <property name="sizePolicy">
+                     <sizepolicy hsizetype="Maximum" vsizetype="Preferred">
+                       <horstretch>0</horstretch>
+                       <verstretch>0</verstretch>
+                     </sizepolicy>
+                   </property>
                   <property name="title">
                    <string>ZR</string>
                   </property>
@@ -1803,6 +1831,22 @@
                      </property>
                     </widget>
                    </item>
+                    <item>
+                      <widget class="QSlider" name="sliderZRThreshold">
+                        <property name="maximumSize">
+                          <size>
+                            <width>70</width>
+                            <height>15</height>
+                          </size>
+                        </property>
+                        <property name="maximum">
+                          <number>100</number>
+                        </property>
+                        <property name="orientation">
+                          <enum>Qt::Horizontal</enum>
+                        </property>
+                      </widget>
+                    </item>
                   </layout>
                  </widget>
                 </item>
diff --git a/src/yuzu_cmd/config.cpp b/src/yuzu_cmd/config.cpp
index b18056baf..3e22fee37 100644
--- a/src/yuzu_cmd/config.cpp
+++ b/src/yuzu_cmd/config.cpp
@@ -446,6 +446,7 @@ void Config::ReadValues() {
     ReadSetting("Renderer", Settings::values.renderer_debug);
     ReadSetting("Renderer", Settings::values.vulkan_device);
 
+    ReadSetting("Renderer", Settings::values.fullscreen_mode);
     ReadSetting("Renderer", Settings::values.aspect_ratio);
     ReadSetting("Renderer", Settings::values.max_anisotropy);
     ReadSetting("Renderer", Settings::values.use_frame_limit);
diff --git a/src/yuzu_cmd/default_ini.h b/src/yuzu_cmd/default_ini.h
index b362f10b4..88d33ecab 100644
--- a/src/yuzu_cmd/default_ini.h
+++ b/src/yuzu_cmd/default_ini.h
@@ -224,6 +224,10 @@ debug =
 # Which Vulkan physical device to use (defaults to 0)
 vulkan_device =
 
+# Whether to use fullscreen or borderless window mode
+# 0 (Windows default): Borderless window, 1 (All other default): Exclusive fullscreen
+fullscreen_mode =
+
 # Aspect ratio
 # 0: Default (16:9), 1: Force 4:3, 2: Force 21:9, 3: Stretch to Window
 aspect_ratio =
diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
index 896181f0b..353e51ea7 100644
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
@@ -6,6 +6,7 @@
 
 #include "common/logging/log.h"
 #include "common/scm_rev.h"
+#include "common/settings.h"
 #include "core/core.h"
 #include "core/perf_stats.h"
 #include "input_common/keyboard.h"
@@ -122,24 +123,37 @@ void EmuWindow_SDL2::OnResize() {
 }
 
 void EmuWindow_SDL2::Fullscreen() {
-    if (SDL_SetWindowFullscreen(render_window, SDL_WINDOW_FULLSCREEN) == 0) {
-        return;
-    }
-
-    LOG_ERROR(Frontend, "Fullscreening failed: {}", SDL_GetError());
+    switch (Settings::values.fullscreen_mode.GetValue()) {
+    case 1: // Exclusive fullscreen
+        // Set window size to render size before entering fullscreen -- SDL does not resize to
+        // display dimensions in this mode.
+        // TODO: Multiply the window size by resolution_factor (for both docked modes)
+        if (Settings::values.use_docked_mode) {
+            SDL_SetWindowSize(render_window, Layout::ScreenDocked::Width,
+                              Layout::ScreenDocked::Height);
+        }
 
-    // Try a different fullscreening method
-    LOG_INFO(Frontend, "Attempting to use borderless fullscreen...");
-    if (SDL_SetWindowFullscreen(render_window, SDL_WINDOW_FULLSCREEN_DESKTOP) == 0) {
-        return;
-    }
+        if (SDL_SetWindowFullscreen(render_window, SDL_WINDOW_FULLSCREEN) == 0) {
+            return;
+        }
 
-    LOG_ERROR(Frontend, "Borderless fullscreening failed: {}", SDL_GetError());
+        LOG_ERROR(Frontend, "Fullscreening failed: {}", SDL_GetError());
+        LOG_INFO(Frontend, "Attempting to use borderless fullscreen...");
+        [[fallthrough]];
+    case 0: // Borderless window
+        if (SDL_SetWindowFullscreen(render_window, SDL_WINDOW_FULLSCREEN_DESKTOP) == 0) {
+            return;
+        }
 
-    // Fallback algorithm: Maximise window.
-    // Works on all systems (unless something is seriously wrong), so no fallback for this one.
-    LOG_INFO(Frontend, "Falling back on a maximised window...");
-    SDL_MaximizeWindow(render_window);
+        LOG_ERROR(Frontend, "Borderless fullscreening failed: {}", SDL_GetError());
+        [[fallthrough]];
+    default:
+        // Fallback algorithm: Maximise window.
+        // Works on all systems (unless something is seriously wrong), so no fallback for this one.
+        LOG_INFO(Frontend, "Falling back on a maximised window...");
+        SDL_MaximizeWindow(render_window);
+        break;
+    }
 }
 
 void EmuWindow_SDL2::WaitEvent() {
diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp
index 152e56db8..d1473dbab 100644
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp
@@ -24,7 +24,7 @@
 #include <SDL.h>
 #include <SDL_syswm.h>
 
-EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem)
+EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem, bool fullscreen)
     : EmuWindow_SDL2{input_subsystem} {
     const std::string window_title = fmt::format("yuzu {} | {}-{} (Vulkan)", Common::g_build_name,
                                                  Common::g_scm_branch, Common::g_scm_desc);
@@ -42,6 +42,10 @@ EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsyste
 
     SetWindowIcon();
 
+    if (fullscreen) {
+        Fullscreen();
+    }
+
     switch (wm.subsystem) {
 #ifdef SDL_VIDEO_DRIVER_WINDOWS
     case SDL_SYSWM_TYPE::SDL_SYSWM_WINDOWS:
diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h
index bdfdc3c6f..de53844f0 100644
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h
@@ -19,7 +19,7 @@ class InputSubsystem;
 
 class EmuWindow_SDL2_VK final : public EmuWindow_SDL2 {
 public:
-    explicit EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem);
+    explicit EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem, bool fullscreen);
     ~EmuWindow_SDL2_VK() override;
 
     std::unique_ptr<Core::Frontend::GraphicsContext> CreateSharedContext() const override;
diff --git a/src/yuzu_cmd/yuzu.cpp b/src/yuzu_cmd/yuzu.cpp
index 9607cdcb1..ac4ea88d3 100644
--- a/src/yuzu_cmd/yuzu.cpp
+++ b/src/yuzu_cmd/yuzu.cpp
@@ -175,7 +175,7 @@ int main(int argc, char** argv) {
         emu_window = std::make_unique<EmuWindow_SDL2_GL>(&input_subsystem, fullscreen);
         break;
     case Settings::RendererBackend::Vulkan:
-        emu_window = std::make_unique<EmuWindow_SDL2_VK>(&input_subsystem);
+        emu_window = std::make_unique<EmuWindow_SDL2_VK>(&input_subsystem, fullscreen);
         break;
     }