133 files changed, 1138 insertions, 679 deletions

diff --git a/src/audio_core/command_generator.cpp b/src/audio_core/command_generator.cpp
index b99d0fc91..45b2eef52 100644
--- a/src/audio_core/command_generator.cpp
+++ b/src/audio_core/command_generator.cpp
@@ -42,6 +42,15 @@ void ApplyMix(std::span<s32> output, std::span<const s32> input, s32 gain, s32 s
 
 s32 ApplyMixRamp(std::span<s32> output, std::span<const s32> input, float gain, float delta,
                  s32 sample_count) {
+    // XC2 passes in NaN mix volumes, causing further issues as we handle everything as s32 rather
+    // than float, so the NaN propogation is lost. As the samples get further modified for
+    // volume etc, they can get out of NaN range, so a later heuristic for catching this is
+    // more difficult. Handle it here by setting these samples to silence.
+    if (std::isnan(gain)) {
+        gain = 0.0f;
+        delta = 0.0f;
+    }
+
     s32 x = 0;
     for (s32 i = 0; i < sample_count; i++) {
         x = static_cast<s32>(static_cast<float>(input[i]) * gain);
diff --git a/src/common/settings.cpp b/src/common/settings.cpp
index 0061e29cc..bf5514386 100644
--- a/src/common/settings.cpp
+++ b/src/common/settings.cpp
@@ -93,7 +93,7 @@ bool IsGPULevelHigh() {
 }
 
 bool IsFastmemEnabled() {
-    if (values.cpu_accuracy.GetValue() == CPUAccuracy::DebugMode) {
+    if (values.cpu_debug_mode) {
         return static_cast<bool>(values.cpuopt_fastmem);
     }
     return true;
@@ -103,7 +103,7 @@ float Volume() {
     if (values.audio_muted) {
         return 0.0f;
     }
-    return values.volume.GetValue();
+    return values.volume.GetValue() / 100.0f;
 }
 
 void RestoreGlobalState(bool is_powered_on) {
diff --git a/src/common/settings.h b/src/common/settings.h
index bf83186f5..ce1bc647d 100644
--- a/src/common/settings.h
+++ b/src/common/settings.h
@@ -31,9 +31,9 @@ enum class GPUAccuracy : u32 {
 };
 
 enum class CPUAccuracy : u32 {
-    Accurate = 0,
-    Unsafe = 1,
-    DebugMode = 2,
+    Auto = 0,
+    Accurate = 1,
+    Unsafe = 2,
 };
 
 /** The BasicSetting class is a simple resource manager. It defines a label and default value
@@ -278,13 +278,16 @@ struct Values {
     BasicSetting<std::string> sink_id{"auto", "output_engine"};
     BasicSetting<bool> audio_muted{false, "audio_muted"};
     Setting<bool> enable_audio_stretching{true, "enable_audio_stretching"};
-    Setting<float> volume{1.0f, "volume"};
+    Setting<u8> volume{100, "volume"};
 
     // Core
     Setting<bool> use_multi_core{true, "use_multi_core"};
 
     // Cpu
-    Setting<CPUAccuracy> cpu_accuracy{CPUAccuracy::Accurate, "cpu_accuracy"};
+    Setting<CPUAccuracy> cpu_accuracy{CPUAccuracy::Auto, "cpu_accuracy"};
+    // TODO: remove cpu_accuracy_first_time, migration setting added 8 July 2021
+    BasicSetting<bool> cpu_accuracy_first_time{true, "cpu_accuracy_first_time"};
+    BasicSetting<bool> cpu_debug_mode{false, "cpu_debug_mode"};
 
     BasicSetting<bool> cpuopt_page_tables{true, "cpuopt_page_tables"};
     BasicSetting<bool> cpuopt_block_linking{true, "cpuopt_block_linking"};
@@ -327,15 +330,15 @@ struct Values {
     Setting<bool> use_nvdec_emulation{true, "use_nvdec_emulation"};
     Setting<bool> accelerate_astc{true, "accelerate_astc"};
     Setting<bool> use_vsync{true, "use_vsync"};
-    Setting<bool> disable_fps_limit{false, "disable_fps_limit"};
+    BasicSetting<bool> disable_fps_limit{false, "disable_fps_limit"};
     Setting<bool> use_assembly_shaders{false, "use_assembly_shaders"};
     Setting<bool> use_asynchronous_shaders{false, "use_asynchronous_shaders"};
     Setting<bool> use_fast_gpu_time{true, "use_fast_gpu_time"};
     Setting<bool> use_caches_gc{false, "use_caches_gc"};
 
-    Setting<float> bg_red{0.0f, "bg_red"};
-    Setting<float> bg_green{0.0f, "bg_green"};
-    Setting<float> bg_blue{0.0f, "bg_blue"};
+    Setting<u8> bg_red{0, "bg_red"};
+    Setting<u8> bg_green{0, "bg_green"};
+    Setting<u8> bg_blue{0, "bg_blue"};
 
     // System
     Setting<std::optional<u32>> rng_seed{std::optional<u32>(), "rng_seed"};
@@ -365,7 +368,7 @@ struct Values {
                                                 "udp_input_servers"};
 
     BasicSetting<bool> mouse_panning{false, "mouse_panning"};
-    BasicSetting<float> mouse_panning_sensitivity{1.0f, "mouse_panning_sensitivity"};
+    BasicSetting<u8> mouse_panning_sensitivity{10, "mouse_panning_sensitivity"};
     BasicSetting<bool> mouse_enabled{false, "mouse_enabled"};
     std::string mouse_device;
     MouseButtonsRaw mouse_buttons;
diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index b2b0dbe05..c7b899131 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -272,22 +272,22 @@ add_library(core STATIC
     hle/service/am/applet_ae.h
     hle/service/am/applet_oe.cpp
     hle/service/am/applet_oe.h
+    hle/service/am/applets/applet_controller.cpp
+    hle/service/am/applets/applet_controller.h
+    hle/service/am/applets/applet_error.cpp
+    hle/service/am/applets/applet_error.h
+    hle/service/am/applets/applet_general_backend.cpp
+    hle/service/am/applets/applet_general_backend.h
+    hle/service/am/applets/applet_profile_select.cpp
+    hle/service/am/applets/applet_profile_select.h
+    hle/service/am/applets/applet_software_keyboard.cpp
+    hle/service/am/applets/applet_software_keyboard.h
+    hle/service/am/applets/applet_software_keyboard_types.h
+    hle/service/am/applets/applet_web_browser.cpp
+    hle/service/am/applets/applet_web_browser.h
+    hle/service/am/applets/applet_web_browser_types.h
     hle/service/am/applets/applets.cpp
     hle/service/am/applets/applets.h
-    hle/service/am/applets/controller.cpp
-    hle/service/am/applets/controller.h
-    hle/service/am/applets/error.cpp
-    hle/service/am/applets/error.h
-    hle/service/am/applets/general_backend.cpp
-    hle/service/am/applets/general_backend.h
-    hle/service/am/applets/profile_select.cpp
-    hle/service/am/applets/profile_select.h
-    hle/service/am/applets/software_keyboard.cpp
-    hle/service/am/applets/software_keyboard.h
-    hle/service/am/applets/software_keyboard_types.h
-    hle/service/am/applets/web_browser.cpp
-    hle/service/am/applets/web_browser.h
-    hle/service/am/applets/web_types.h
     hle/service/am/idle.cpp
     hle/service/am/idle.h
     hle/service/am/omm.cpp
@@ -300,10 +300,10 @@ add_library(core STATIC
     hle/service/aoc/aoc_u.h
     hle/service/apm/apm.cpp
     hle/service/apm/apm.h
-    hle/service/apm/controller.cpp
-    hle/service/apm/controller.h
-    hle/service/apm/interface.cpp
-    hle/service/apm/interface.h
+    hle/service/apm/apm_controller.cpp
+    hle/service/apm/apm_controller.h
+    hle/service/apm/apm_interface.cpp
+    hle/service/apm/apm_interface.h
     hle/service/audio/audctl.cpp
     hle/service/audio/audctl.h
     hle/service/audio/auddbg.cpp
@@ -335,8 +335,8 @@ add_library(core STATIC
     hle/service/bcat/backend/backend.h
     hle/service/bcat/bcat.cpp
     hle/service/bcat/bcat.h
-    hle/service/bcat/module.cpp
-    hle/service/bcat/module.h
+    hle/service/bcat/bcat_module.cpp
+    hle/service/bcat/bcat_module.h
     hle/service/bpc/bpc.cpp
     hle/service/bpc/bpc.h
     hle/service/btdrv/btdrv.cpp
@@ -382,8 +382,8 @@ add_library(core STATIC
     hle/service/friend/errors.h
     hle/service/friend/friend.cpp
     hle/service/friend/friend.h
-    hle/service/friend/interface.cpp
-    hle/service/friend/interface.h
+    hle/service/friend/friend_interface.cpp
+    hle/service/friend/friend_interface.h
     hle/service/glue/arp.cpp
     hle/service/glue/arp.h
     hle/service/glue/bgtc.cpp
@@ -393,8 +393,8 @@ add_library(core STATIC
     hle/service/glue/errors.h
     hle/service/glue/glue.cpp
     hle/service/glue/glue.h
-    hle/service/glue/manager.cpp
-    hle/service/glue/manager.h
+    hle/service/glue/glue_manager.cpp
+    hle/service/glue/glue_manager.h
     hle/service/grc/grc.cpp
     hle/service/grc/grc.h
     hle/service/hid/hid.cpp
@@ -435,10 +435,10 @@ add_library(core STATIC
     hle/service/lm/lm.h
     hle/service/mig/mig.cpp
     hle/service/mig/mig.h
-    hle/service/mii/manager.cpp
-    hle/service/mii/manager.h
     hle/service/mii/mii.cpp
     hle/service/mii/mii.h
+    hle/service/mii/mii_manager.cpp
+    hle/service/mii/mii_manager.h
     hle/service/mii/raw_data.cpp
     hle/service/mii/raw_data.h
     hle/service/mii/types.h
@@ -486,11 +486,11 @@ add_library(core STATIC
     hle/service/nvdrv/devices/nvhost_vic.h
     hle/service/nvdrv/devices/nvmap.cpp
     hle/service/nvdrv/devices/nvmap.h
-    hle/service/nvdrv/interface.cpp
-    hle/service/nvdrv/interface.h
     hle/service/nvdrv/nvdata.h
     hle/service/nvdrv/nvdrv.cpp
     hle/service/nvdrv/nvdrv.h
+    hle/service/nvdrv/nvdrv_interface.cpp
+    hle/service/nvdrv/nvdrv_interface.h
     hle/service/nvdrv/nvmemp.cpp
     hle/service/nvdrv/nvmemp.h
     hle/service/nvdrv/syncpoint_manager.cpp
@@ -503,10 +503,10 @@ add_library(core STATIC
     hle/service/olsc/olsc.h
     hle/service/pcie/pcie.cpp
     hle/service/pcie/pcie.h
-    hle/service/pctl/module.cpp
-    hle/service/pctl/module.h
     hle/service/pctl/pctl.cpp
     hle/service/pctl/pctl.h
+    hle/service/pctl/pctl_module.cpp
+    hle/service/pctl/pctl_module.h
     hle/service/pcv/pcv.cpp
     hle/service/pcv/pcv.h
     hle/service/pm/pm.cpp
@@ -529,10 +529,10 @@ add_library(core STATIC
     hle/service/set/set_sys.h
     hle/service/set/settings.cpp
     hle/service/set/settings.h
-    hle/service/sm/controller.cpp
-    hle/service/sm/controller.h
     hle/service/sm/sm.cpp
     hle/service/sm/sm.h
+    hle/service/sm/sm_controller.cpp
+    hle/service/sm/sm_controller.h
     hle/service/sockets/bsd.cpp
     hle/service/sockets/bsd.h
     hle/service/sockets/ethc.cpp
@@ -547,10 +547,10 @@ add_library(core STATIC
     hle/service/sockets/sockets_translate.h
     hle/service/spl/csrng.cpp
     hle/service/spl/csrng.h
-    hle/service/spl/module.cpp
-    hle/service/spl/module.h
     hle/service/spl/spl.cpp
     hle/service/spl/spl.h
+    hle/service/spl/spl_module.cpp
+    hle/service/spl/spl_module.h
     hle/service/spl/spl_results.h
     hle/service/spl/spl_types.h
     hle/service/ssl/ssl.cpp
@@ -559,8 +559,6 @@ add_library(core STATIC
     hle/service/time/ephemeral_network_system_clock_context_writer.h
     hle/service/time/ephemeral_network_system_clock_core.h
     hle/service/time/errors.h
-    hle/service/time/interface.cpp
-    hle/service/time/interface.h
     hle/service/time/local_system_clock_context_writer.h
     hle/service/time/network_system_clock_context_writer.h
     hle/service/time/standard_local_system_clock_core.h
@@ -578,6 +576,8 @@ add_library(core STATIC
     hle/service/time/tick_based_steady_clock_core.h
     hle/service/time/time.cpp
     hle/service/time/time.h
+    hle/service/time/time_interface.cpp
+    hle/service/time/time_interface.h
     hle/service/time/time_manager.cpp
     hle/service/time/time_manager.h
     hle/service/time/time_sharedmemory.cpp
diff --git a/src/core/arm/dynarmic/arm_dynarmic_32.cpp b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
index 77a44f862..b0d89c539 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_32.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
@@ -150,7 +150,7 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
     config.far_code_offset = 400_MiB;
 
     // Safe optimizations
-    if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::DebugMode) {
+    if (Settings::values.cpu_debug_mode) {
         if (!Settings::values.cpuopt_page_tables) {
             config.page_table = nullptr;
         }
@@ -183,20 +183,28 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
     // Unsafe optimizations
     if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Unsafe) {
         config.unsafe_optimizations = true;
-        if (Settings::values.cpuopt_unsafe_unfuse_fma.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_unfuse_fma) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
         }
-        if (Settings::values.cpuopt_unsafe_reduce_fp_error.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_reduce_fp_error) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_ReducedErrorFP;
         }
-        if (Settings::values.cpuopt_unsafe_ignore_standard_fpcr.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_ignore_standard_fpcr) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_IgnoreStandardFPCRValue;
         }
-        if (Settings::values.cpuopt_unsafe_inaccurate_nan.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_inaccurate_nan) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
         }
     }
 
+    // Curated optimizations
+    if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Auto) {
+        config.unsafe_optimizations = true;
+        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
+        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_IgnoreStandardFPCRValue;
+        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
+    }
+
     return std::make_unique<Dynarmic::A32::Jit>(config);
 }
 
diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.cpp b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
index 75332e348..bf27ffe71 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
@@ -190,7 +190,7 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
     config.far_code_offset = 400_MiB;
 
     // Safe optimizations
-    if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::DebugMode) {
+    if (Settings::values.cpu_debug_mode) {
         if (!Settings::values.cpuopt_page_tables) {
             config.page_table = nullptr;
         }
@@ -223,20 +223,28 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
     // Unsafe optimizations
     if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Unsafe) {
         config.unsafe_optimizations = true;
-        if (Settings::values.cpuopt_unsafe_unfuse_fma.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_unfuse_fma) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
         }
-        if (Settings::values.cpuopt_unsafe_reduce_fp_error.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_reduce_fp_error) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_ReducedErrorFP;
         }
-        if (Settings::values.cpuopt_unsafe_inaccurate_nan.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_inaccurate_nan) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
         }
-        if (Settings::values.cpuopt_unsafe_fastmem_check.GetValue()) {
+        if (Settings::values.cpuopt_unsafe_fastmem_check) {
             config.fastmem_address_space_bits = 64;
         }
     }
 
+    // Curated optimizations
+    if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Auto) {
+        config.unsafe_optimizations = true;
+        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
+        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
+        config.fastmem_address_space_bits = 64;
+    }
+
     return std::make_shared<Dynarmic::A64::Jit>(config);
 }
 
diff --git a/src/core/core.cpp b/src/core/core.cpp
index 891f1cb49..406320ed6 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -35,9 +35,9 @@
 #include "core/hle/kernel/kernel.h"
 #include "core/hle/kernel/physical_core.h"
 #include "core/hle/service/am/applets/applets.h"
-#include "core/hle/service/apm/controller.h"
+#include "core/hle/service/apm/apm_controller.h"
 #include "core/hle/service/filesystem/filesystem.h"
-#include "core/hle/service/glue/manager.h"
+#include "core/hle/service/glue/glue_manager.h"
 #include "core/hle/service/hid/hid.h"
 #include "core/hle/service/service.h"
 #include "core/hle/service/sm/sm.h"
diff --git a/src/core/file_sys/content_archive.cpp b/src/core/file_sys/content_archive.cpp
index 24eff210f..7019a7a68 100644
--- a/src/core/file_sys/content_archive.cpp
+++ b/src/core/file_sys/content_archive.cpp
@@ -5,7 +5,6 @@
 #include <algorithm>
 #include <cstring>
 #include <optional>
-#include <ranges>
 #include <utility>
 
 #include "common/logging/log.h"
diff --git a/src/core/frontend/applets/software_keyboard.h b/src/core/frontend/applets/software_keyboard.h
index 506eb35bb..228a548d4 100644
--- a/src/core/frontend/applets/software_keyboard.h
+++ b/src/core/frontend/applets/software_keyboard.h
@@ -9,7 +9,7 @@
 
 #include "common/common_types.h"
 
-#include "core/hle/service/am/applets/software_keyboard_types.h"
+#include "core/hle/service/am/applets/applet_software_keyboard_types.h"
 
 namespace Core::Frontend {
 
diff --git a/src/core/frontend/applets/web_browser.h b/src/core/frontend/applets/web_browser.h
index d7bd44c27..915dde677 100644
--- a/src/core/frontend/applets/web_browser.h
+++ b/src/core/frontend/applets/web_browser.h
@@ -7,7 +7,7 @@
 #include <functional>
 #include <string_view>
 
-#include "core/hle/service/am/applets/web_types.h"
+#include "core/hle/service/am/applets/applet_web_browser_types.h"
 
 namespace Core::Frontend {
 
diff --git a/src/core/hardware_interrupt_manager.cpp b/src/core/hardware_interrupt_manager.cpp
index 645f26e91..290db505e 100644
--- a/src/core/hardware_interrupt_manager.cpp
+++ b/src/core/hardware_interrupt_manager.cpp
@@ -5,7 +5,7 @@
 #include "core/core.h"
 #include "core/core_timing.h"
 #include "core/hardware_interrupt_manager.h"
-#include "core/hle/service/nvdrv/interface.h"
+#include "core/hle/service/nvdrv/nvdrv_interface.h"
 #include "core/hle/service/sm/sm.h"
 
 namespace Core::Hardware {
diff --git a/src/core/hle/service/acc/acc.cpp b/src/core/hle/service/acc/acc.cpp
index d1c1fb71d..2e969f2a8 100644
--- a/src/core/hle/service/acc/acc.cpp
+++ b/src/core/hle/service/acc/acc.cpp
@@ -26,7 +26,7 @@
 #include "core/hle/service/acc/errors.h"
 #include "core/hle/service/acc/profile_manager.h"
 #include "core/hle/service/glue/arp.h"
-#include "core/hle/service/glue/manager.h"
+#include "core/hle/service/glue/glue_manager.h"
 #include "core/hle/service/sm/sm.h"
 #include "core/loader/loader.h"
 
diff --git a/src/core/hle/service/acc/acc.h b/src/core/hle/service/acc/acc.h
index 0e3ad8ec6..a83a480cd 100644
--- a/src/core/hle/service/acc/acc.h
+++ b/src/core/hle/service/acc/acc.h
@@ -5,7 +5,7 @@
 #pragma once
 
 #include "common/uuid.h"
-#include "core/hle/service/glue/manager.h"
+#include "core/hle/service/glue/glue_manager.h"
 #include "core/hle/service/service.h"
 
 namespace Service::Account {
diff --git a/src/core/hle/service/am/am.cpp b/src/core/hle/service/am/am.cpp
index 23ebc1138..a538f82e3 100644
--- a/src/core/hle/service/am/am.cpp
+++ b/src/core/hle/service/am/am.cpp
@@ -24,16 +24,16 @@
 #include "core/hle/service/am/am.h"
 #include "core/hle/service/am/applet_ae.h"
 #include "core/hle/service/am/applet_oe.h"
+#include "core/hle/service/am/applets/applet_profile_select.h"
+#include "core/hle/service/am/applets/applet_software_keyboard.h"
+#include "core/hle/service/am/applets/applet_web_browser.h"
 #include "core/hle/service/am/applets/applets.h"
-#include "core/hle/service/am/applets/profile_select.h"
-#include "core/hle/service/am/applets/software_keyboard.h"
-#include "core/hle/service/am/applets/web_browser.h"
 #include "core/hle/service/am/idle.h"
 #include "core/hle/service/am/omm.h"
 #include "core/hle/service/am/spsm.h"
 #include "core/hle/service/am/tcap.h"
-#include "core/hle/service/apm/controller.h"
-#include "core/hle/service/apm/interface.h"
+#include "core/hle/service/apm/apm_controller.h"
+#include "core/hle/service/apm/apm_interface.h"
 #include "core/hle/service/bcat/backend/backend.h"
 #include "core/hle/service/filesystem/filesystem.h"
 #include "core/hle/service/ns/ns.h"
diff --git a/src/core/hle/service/am/applets/controller.cpp b/src/core/hle/service/am/applets/applet_controller.cpp
index 218c8d1e4..12682effe 100644
--- a/src/core/hle/service/am/applets/controller.cpp
+++ b/src/core/hle/service/am/applets/applet_controller.cpp
@@ -12,7 +12,7 @@
 #include "core/frontend/applets/controller.h"
 #include "core/hle/result.h"
 #include "core/hle/service/am/am.h"
-#include "core/hle/service/am/applets/controller.h"
+#include "core/hle/service/am/applets/applet_controller.h"
 #include "core/hle/service/hid/controllers/npad.h"
 
 namespace Service::AM::Applets {
diff --git a/src/core/hle/service/am/applets/controller.h b/src/core/hle/service/am/applets/applet_controller.h
index 20617e91f..20617e91f 100644
--- a/src/core/hle/service/am/applets/controller.h
+++ b/src/core/hle/service/am/applets/applet_controller.h
diff --git a/src/core/hle/service/am/applets/error.cpp b/src/core/hle/service/am/applets/applet_error.cpp
index c724e5d5b..ef6854d62 100644
--- a/src/core/hle/service/am/applets/error.cpp
+++ b/src/core/hle/service/am/applets/applet_error.cpp
@@ -11,7 +11,7 @@
 #include "core/frontend/applets/error.h"
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/service/am/am.h"
-#include "core/hle/service/am/applets/error.h"
+#include "core/hle/service/am/applets/applet_error.h"
 #include "core/reporter.h"
 
 namespace Service::AM::Applets {
diff --git a/src/core/hle/service/am/applets/error.h b/src/core/hle/service/am/applets/applet_error.h
index 8aa9046a5..8aa9046a5 100644
--- a/src/core/hle/service/am/applets/error.h
+++ b/src/core/hle/service/am/applets/applet_error.h
diff --git a/src/core/hle/service/am/applets/general_backend.cpp b/src/core/hle/service/am/applets/applet_general_backend.cpp
index 9fcb9f95b..0f413f9a0 100644
--- a/src/core/hle/service/am/applets/general_backend.cpp
+++ b/src/core/hle/service/am/applets/applet_general_backend.cpp
@@ -12,7 +12,7 @@
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/result.h"
 #include "core/hle/service/am/am.h"
-#include "core/hle/service/am/applets/general_backend.h"
+#include "core/hle/service/am/applets/applet_general_backend.h"
 #include "core/reporter.h"
 
 namespace Service::AM::Applets {
diff --git a/src/core/hle/service/am/applets/general_backend.h b/src/core/hle/service/am/applets/applet_general_backend.h
index 7496ded88..7496ded88 100644
--- a/src/core/hle/service/am/applets/general_backend.h
+++ b/src/core/hle/service/am/applets/applet_general_backend.h
diff --git a/src/core/hle/service/am/applets/profile_select.cpp b/src/core/hle/service/am/applets/applet_profile_select.cpp
index 37048be26..bdc21778e 100644
--- a/src/core/hle/service/am/applets/profile_select.cpp
+++ b/src/core/hle/service/am/applets/applet_profile_select.cpp
@@ -9,7 +9,7 @@
 #include "core/core.h"
 #include "core/frontend/applets/profile_select.h"
 #include "core/hle/service/am/am.h"
-#include "core/hle/service/am/applets/profile_select.h"
+#include "core/hle/service/am/applets/applet_profile_select.h"
 
 namespace Service::AM::Applets {
 
diff --git a/src/core/hle/service/am/applets/profile_select.h b/src/core/hle/service/am/applets/applet_profile_select.h
index 8fb76e6c4..8fb76e6c4 100644
--- a/src/core/hle/service/am/applets/profile_select.h
+++ b/src/core/hle/service/am/applets/applet_profile_select.h
diff --git a/src/core/hle/service/am/applets/software_keyboard.cpp b/src/core/hle/service/am/applets/applet_software_keyboard.cpp
index 00dfe1675..7cae90609 100644
--- a/src/core/hle/service/am/applets/software_keyboard.cpp
+++ b/src/core/hle/service/am/applets/applet_software_keyboard.cpp
@@ -6,7 +6,7 @@
 #include "core/core.h"
 #include "core/frontend/applets/software_keyboard.h"
 #include "core/hle/service/am/am.h"
-#include "core/hle/service/am/applets/software_keyboard.h"
+#include "core/hle/service/am/applets/applet_software_keyboard.h"
 
 namespace Service::AM::Applets {
 
diff --git a/src/core/hle/service/am/applets/software_keyboard.h b/src/core/hle/service/am/applets/applet_software_keyboard.h
index e3fc733d0..9aef1bf11 100644
--- a/src/core/hle/service/am/applets/software_keyboard.h
+++ b/src/core/hle/service/am/applets/applet_software_keyboard.h
@@ -7,8 +7,8 @@
 #include "common/common_funcs.h"
 #include "common/common_types.h"
 #include "core/hle/result.h"
+#include "core/hle/service/am/applets/applet_software_keyboard_types.h"
 #include "core/hle/service/am/applets/applets.h"
-#include "core/hle/service/am/applets/software_keyboard_types.h"
 
 namespace Core {
 class System;
diff --git a/src/core/hle/service/am/applets/software_keyboard_types.h b/src/core/hle/service/am/applets/applet_software_keyboard_types.h
index 21aa8e800..21aa8e800 100644
--- a/src/core/hle/service/am/applets/software_keyboard_types.h
+++ b/src/core/hle/service/am/applets/applet_software_keyboard_types.h
diff --git a/src/core/hle/service/am/applets/web_browser.cpp b/src/core/hle/service/am/applets/applet_web_browser.cpp
index 3b28e829b..35f194961 100644
--- a/src/core/hle/service/am/applets/web_browser.cpp
+++ b/src/core/hle/service/am/applets/applet_web_browser.cpp
@@ -21,7 +21,7 @@
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/result.h"
 #include "core/hle/service/am/am.h"
-#include "core/hle/service/am/applets/web_browser.h"
+#include "core/hle/service/am/applets/applet_web_browser.h"
 #include "core/hle/service/filesystem/filesystem.h"
 #include "core/hle/service/ns/pl_u.h"
 
@@ -58,6 +58,16 @@ std::string GetMainURL(const std::string& url) {
     return url.substr(0, index);
 }
 
+std::string ResolveURL(const std::string& url) {
+    const auto index = url.find_first_of('%');
+
+    if (index == std::string::npos) {
+        return url;
+    }
+
+    return url.substr(0, index) + "lp1" + url.substr(index + 1);
+}
+
 WebArgInputTLVMap ReadWebArgs(const std::vector<u8>& web_arg, WebArgHeader& web_arg_header) {
     std::memcpy(&web_arg_header, web_arg.data(), sizeof(WebArgHeader));
 
@@ -407,6 +417,9 @@ void WebBrowser::InitializeShare() {}
 
 void WebBrowser::InitializeWeb() {
     external_url = ParseStringValue(GetInputTLVData(WebArgInputTLVType::InitialURL).value());
+
+    // Resolve Nintendo CDN URLs.
+    external_url = ResolveURL(external_url);
 }
 
 void WebBrowser::InitializeWifi() {}
diff --git a/src/core/hle/service/am/applets/web_browser.h b/src/core/hle/service/am/applets/applet_web_browser.h
index 9f81214b6..4f9e81b79 100644
--- a/src/core/hle/service/am/applets/web_browser.h
+++ b/src/core/hle/service/am/applets/applet_web_browser.h
@@ -11,8 +11,8 @@
 #include "common/common_types.h"
 #include "core/file_sys/vfs_types.h"
 #include "core/hle/result.h"
+#include "core/hle/service/am/applets/applet_web_browser_types.h"
 #include "core/hle/service/am/applets/applets.h"
-#include "core/hle/service/am/applets/web_types.h"
 
 namespace Core {
 class System;
diff --git a/src/core/hle/service/am/applets/web_types.h b/src/core/hle/service/am/applets/applet_web_browser_types.h
index 419c2bf79..419c2bf79 100644
--- a/src/core/hle/service/am/applets/web_types.h
+++ b/src/core/hle/service/am/applets/applet_web_browser_types.h
diff --git a/src/core/hle/service/am/applets/applets.cpp b/src/core/hle/service/am/applets/applets.cpp
index ae995df6b..2b7685d42 100644
--- a/src/core/hle/service/am/applets/applets.cpp
+++ b/src/core/hle/service/am/applets/applets.cpp
@@ -17,13 +17,13 @@
 #include "core/hle/service/am/am.h"
 #include "core/hle/service/am/applet_ae.h"
 #include "core/hle/service/am/applet_oe.h"
+#include "core/hle/service/am/applets/applet_controller.h"
+#include "core/hle/service/am/applets/applet_error.h"
+#include "core/hle/service/am/applets/applet_general_backend.h"
+#include "core/hle/service/am/applets/applet_profile_select.h"
+#include "core/hle/service/am/applets/applet_software_keyboard.h"
+#include "core/hle/service/am/applets/applet_web_browser.h"
 #include "core/hle/service/am/applets/applets.h"
-#include "core/hle/service/am/applets/controller.h"
-#include "core/hle/service/am/applets/error.h"
-#include "core/hle/service/am/applets/general_backend.h"
-#include "core/hle/service/am/applets/profile_select.h"
-#include "core/hle/service/am/applets/software_keyboard.h"
-#include "core/hle/service/am/applets/web_browser.h"
 #include "core/hle/service/sm/sm.h"
 
 namespace Service::AM::Applets {
diff --git a/src/core/hle/service/apm/apm.cpp b/src/core/hle/service/apm/apm.cpp
index 97d6619dd..f5ebfe8d6 100644
--- a/src/core/hle/service/apm/apm.cpp
+++ b/src/core/hle/service/apm/apm.cpp
@@ -5,7 +5,7 @@
 #include "core/core.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/service/apm/apm.h"
-#include "core/hle/service/apm/interface.h"
+#include "core/hle/service/apm/apm_interface.h"
 
 namespace Service::APM {
 
diff --git a/src/core/hle/service/apm/controller.cpp b/src/core/hle/service/apm/apm_controller.cpp
index 8bfa7c0e4..98839fe97 100644
--- a/src/core/hle/service/apm/controller.cpp
+++ b/src/core/hle/service/apm/apm_controller.cpp
@@ -9,7 +9,7 @@
 #include "common/logging/log.h"
 #include "common/settings.h"
 #include "core/core_timing.h"
-#include "core/hle/service/apm/controller.h"
+#include "core/hle/service/apm/apm_controller.h"
 
 namespace Service::APM {
 
diff --git a/src/core/hle/service/apm/controller.h b/src/core/hle/service/apm/apm_controller.h
index 8d48e0104..8d48e0104 100644
--- a/src/core/hle/service/apm/controller.h
+++ b/src/core/hle/service/apm/apm_controller.h
diff --git a/src/core/hle/service/apm/interface.cpp b/src/core/hle/service/apm/apm_interface.cpp
index d69ddd135..e58bad083 100644
--- a/src/core/hle/service/apm/interface.cpp
+++ b/src/core/hle/service/apm/apm_interface.cpp
@@ -5,8 +5,8 @@
 #include "common/logging/log.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/service/apm/apm.h"
-#include "core/hle/service/apm/controller.h"
-#include "core/hle/service/apm/interface.h"
+#include "core/hle/service/apm/apm_controller.h"
+#include "core/hle/service/apm/apm_interface.h"
 
 namespace Service::APM {
 
diff --git a/src/core/hle/service/apm/interface.h b/src/core/hle/service/apm/apm_interface.h
index 063ad5308..063ad5308 100644
--- a/src/core/hle/service/apm/interface.h
+++ b/src/core/hle/service/apm/apm_interface.h
diff --git a/src/core/hle/service/bcat/backend/boxcat.cpp b/src/core/hle/service/bcat/backend/boxcat.cpp
index dc15cf58b..7ca7f2aac 100644
--- a/src/core/hle/service/bcat/backend/boxcat.cpp
+++ b/src/core/hle/service/bcat/backend/boxcat.cpp
@@ -7,6 +7,9 @@
 #ifdef __GNUC__
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wshadow"
+#ifndef __clang__
+#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
+#endif
 #endif
 #include <httplib.h>
 #include <mbedtls/sha256.h>
diff --git a/src/core/hle/service/bcat/bcat.h b/src/core/hle/service/bcat/bcat.h
index d72798980..1eba477da 100644
--- a/src/core/hle/service/bcat/bcat.h
+++ b/src/core/hle/service/bcat/bcat.h
@@ -4,7 +4,7 @@
 
 #pragma once
 
-#include "core/hle/service/bcat/module.h"
+#include "core/hle/service/bcat/bcat_module.h"
 
 namespace Core {
 class System;
diff --git a/src/core/hle/service/bcat/module.cpp b/src/core/hle/service/bcat/bcat_module.cpp
index f85444da8..72294eb2e 100644
--- a/src/core/hle/service/bcat/module.cpp
+++ b/src/core/hle/service/bcat/bcat_module.cpp
@@ -17,7 +17,7 @@
 #include "core/hle/kernel/k_writable_event.h"
 #include "core/hle/service/bcat/backend/backend.h"
 #include "core/hle/service/bcat/bcat.h"
-#include "core/hle/service/bcat/module.h"
+#include "core/hle/service/bcat/bcat_module.h"
 #include "core/hle/service/filesystem/filesystem.h"
 
 namespace Service::BCAT {
diff --git a/src/core/hle/service/bcat/module.h b/src/core/hle/service/bcat/bcat_module.h
index 738731c06..738731c06 100644
--- a/src/core/hle/service/bcat/module.h
+++ b/src/core/hle/service/bcat/bcat_module.h
diff --git a/src/core/hle/service/friend/friend.cpp b/src/core/hle/service/friend/friend.cpp
index c5f88bce7..a3c939c0c 100644
--- a/src/core/hle/service/friend/friend.cpp
+++ b/src/core/hle/service/friend/friend.cpp
@@ -12,7 +12,7 @@
 #include "core/hle/kernel/k_writable_event.h"
 #include "core/hle/service/friend/errors.h"
 #include "core/hle/service/friend/friend.h"
-#include "core/hle/service/friend/interface.h"
+#include "core/hle/service/friend/friend_interface.h"
 
 namespace Service::Friend {
 
diff --git a/src/core/hle/service/friend/interface.cpp b/src/core/hle/service/friend/friend_interface.cpp
index 7368ccec2..9b18b2a32 100644
--- a/src/core/hle/service/friend/interface.cpp
+++ b/src/core/hle/service/friend/friend_interface.cpp
@@ -2,7 +2,7 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include "core/hle/service/friend/interface.h"
+#include "core/hle/service/friend/friend_interface.h"
 
 namespace Service::Friend {
 
diff --git a/src/core/hle/service/friend/interface.h b/src/core/hle/service/friend/friend_interface.h
index 43d914b32..43d914b32 100644
--- a/src/core/hle/service/friend/interface.h
+++ b/src/core/hle/service/friend/friend_interface.h
diff --git a/src/core/hle/service/glue/arp.cpp b/src/core/hle/service/glue/arp.cpp
index ca25df67e..5a3b54cc1 100644
--- a/src/core/hle/service/glue/arp.cpp
+++ b/src/core/hle/service/glue/arp.cpp
@@ -13,7 +13,7 @@
 #include "core/hle/kernel/kernel.h"
 #include "core/hle/service/glue/arp.h"
 #include "core/hle/service/glue/errors.h"
-#include "core/hle/service/glue/manager.h"
+#include "core/hle/service/glue/glue_manager.h"
 #include "core/hle/service/service.h"
 
 namespace Service::Glue {
diff --git a/src/core/hle/service/glue/manager.cpp b/src/core/hle/service/glue/glue_manager.cpp
index 9b1754cf8..aa9d48c0c 100644
--- a/src/core/hle/service/glue/manager.cpp
+++ b/src/core/hle/service/glue/glue_manager.cpp
@@ -3,7 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "core/hle/service/glue/errors.h"
-#include "core/hle/service/glue/manager.h"
+#include "core/hle/service/glue/glue_manager.h"
 
 namespace Service::Glue {
 
diff --git a/src/core/hle/service/glue/manager.h b/src/core/hle/service/glue/glue_manager.h
index a7f5ce3ee..a7f5ce3ee 100644
--- a/src/core/hle/service/glue/manager.h
+++ b/src/core/hle/service/glue/glue_manager.h
diff --git a/src/core/hle/service/hid/controllers/npad.cpp b/src/core/hle/service/hid/controllers/npad.cpp
index 1eb02aee2..6ce1360e3 100644
--- a/src/core/hle/service/hid/controllers/npad.cpp
+++ b/src/core/hle/service/hid/controllers/npad.cpp
@@ -941,6 +941,11 @@ void Controller_NPad::InitializeVibrationDevice(const DeviceHandle& vibration_de
 
 void Controller_NPad::InitializeVibrationDeviceAtIndex(std::size_t npad_index,
                                                        std::size_t device_index) {
+    if (!Settings::values.vibration_enabled.GetValue()) {
+        vibration_devices_mounted[npad_index][device_index] = false;
+        return;
+    }
+
     if (vibrations[npad_index][device_index]) {
         vibration_devices_mounted[npad_index][device_index] =
             vibrations[npad_index][device_index]->GetStatus() == 1;
diff --git a/src/core/hle/service/mii/mii.cpp b/src/core/hle/service/mii/mii.cpp
index bbd81a88a..9d863486a 100644
--- a/src/core/hle/service/mii/mii.cpp
+++ b/src/core/hle/service/mii/mii.cpp
@@ -7,8 +7,8 @@
 #include "common/logging/log.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/kernel/hle_ipc.h"
-#include "core/hle/service/mii/manager.h"
 #include "core/hle/service/mii/mii.h"
+#include "core/hle/service/mii/mii_manager.h"
 #include "core/hle/service/service.h"
 #include "core/hle/service/sm/sm.h"
 
diff --git a/src/core/hle/service/mii/manager.cpp b/src/core/hle/service/mii/mii_manager.cpp
index 869d2763f..4fef2aea4 100644
--- a/src/core/hle/service/mii/manager.cpp
+++ b/src/core/hle/service/mii/mii_manager.cpp
@@ -10,7 +10,7 @@
 #include "common/string_util.h"
 
 #include "core/hle/service/acc/profile_manager.h"
-#include "core/hle/service/mii/manager.h"
+#include "core/hle/service/mii/mii_manager.h"
 #include "core/hle/service/mii/raw_data.h"
 #include "core/hle/service/mii/types.h"
 
diff --git a/src/core/hle/service/mii/manager.h b/src/core/hle/service/mii/mii_manager.h
index 8e048fc56..8e048fc56 100644
--- a/src/core/hle/service/mii/manager.h
+++ b/src/core/hle/service/mii/mii_manager.h
diff --git a/src/core/hle/service/mii/raw_data.h b/src/core/hle/service/mii/raw_data.h
index 0e35d69d2..a0d2b9d3a 100644
--- a/src/core/hle/service/mii/raw_data.h
+++ b/src/core/hle/service/mii/raw_data.h
@@ -7,7 +7,7 @@
 #include <array>
 
 #include "common/common_types.h"
-#include "core/hle/service/mii/manager.h"
+#include "core/hle/service/mii/mii_manager.h"
 
 namespace Service::Mii::RawData {
 
diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp
index 98e6296f1..1403a39d0 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp
@@ -19,26 +19,29 @@
 namespace Service::Nvidia::Devices {
 
 namespace {
-// Splice vectors will copy count amount of type T from the input vector into the dst vector.
+// Copies count amount of type T from the input vector into the dst vector.
+// Returns the number of bytes written into dst.
 template <typename T>
-std::size_t SpliceVectors(const std::vector<u8>& input, std::vector<T>& dst, std::size_t count,
-                          std::size_t offset) {
-    if (!dst.empty()) {
-        std::memcpy(dst.data(), input.data() + offset, count * sizeof(T));
+std::size_t SliceVectors(const std::vector<u8>& input, std::vector<T>& dst, std::size_t count,
+                         std::size_t offset) {
+    if (dst.empty()) {
+        return 0;
     }
-    return 0;
+    const size_t bytes_copied = count * sizeof(T);
+    std::memcpy(dst.data(), input.data() + offset, bytes_copied);
+    return bytes_copied;
 }
 
-// Write vectors will write data to the output buffer
+// Writes the data in src to an offset into the dst vector. The offset is specified in bytes
+// Returns the number of bytes written into dst.
 template <typename T>
 std::size_t WriteVectors(std::vector<u8>& dst, const std::vector<T>& src, std::size_t offset) {
     if (src.empty()) {
         return 0;
-    } else {
-        std::memcpy(dst.data() + offset, src.data(), src.size() * sizeof(T));
-        offset += src.size() * sizeof(T);
-        return offset;
     }
+    const size_t bytes_copied = src.size() * sizeof(T);
+    std::memcpy(dst.data() + offset, src.data(), bytes_copied);
+    return bytes_copied;
 }
 } // Anonymous namespace
 
@@ -62,7 +65,6 @@ NvResult nvhost_nvdec_common::Submit(const std::vector<u8>& input, std::vector<u
     LOG_DEBUG(Service_NVDRV, "called NVDEC Submit, cmd_buffer_count={}", params.cmd_buffer_count);
 
     // Instantiate param buffers
-    std::size_t offset = sizeof(IoctlSubmit);
     std::vector<CommandBuffer> command_buffers(params.cmd_buffer_count);
     std::vector<Reloc> relocs(params.relocation_count);
     std::vector<u32> reloc_shifts(params.relocation_count);
@@ -70,13 +72,14 @@ NvResult nvhost_nvdec_common::Submit(const std::vector<u8>& input, std::vector<u
     std::vector<SyncptIncr> wait_checks(params.syncpoint_count);
     std::vector<Fence> fences(params.fence_count);
 
-    // Splice input into their respective buffers
-    offset = SpliceVectors(input, command_buffers, params.cmd_buffer_count, offset);
-    offset = SpliceVectors(input, relocs, params.relocation_count, offset);
-    offset = SpliceVectors(input, reloc_shifts, params.relocation_count, offset);
-    offset = SpliceVectors(input, syncpt_increments, params.syncpoint_count, offset);
-    offset = SpliceVectors(input, wait_checks, params.syncpoint_count, offset);
-    offset = SpliceVectors(input, fences, params.fence_count, offset);
+    // Slice input into their respective buffers
+    std::size_t offset = sizeof(IoctlSubmit);
+    offset += SliceVectors(input, command_buffers, params.cmd_buffer_count, offset);
+    offset += SliceVectors(input, relocs, params.relocation_count, offset);
+    offset += SliceVectors(input, reloc_shifts, params.relocation_count, offset);
+    offset += SliceVectors(input, syncpt_increments, params.syncpoint_count, offset);
+    offset += SliceVectors(input, wait_checks, params.syncpoint_count, offset);
+    offset += SliceVectors(input, fences, params.fence_count, offset);
 
     auto& gpu = system.GPU();
     if (gpu.UseNvdec()) {
@@ -88,35 +91,27 @@ NvResult nvhost_nvdec_common::Submit(const std::vector<u8>& input, std::vector<u
         }
     }
     for (const auto& cmd_buffer : command_buffers) {
-        auto object = nvmap_dev->GetObject(cmd_buffer.memory_id);
+        const auto object = nvmap_dev->GetObject(cmd_buffer.memory_id);
         ASSERT_OR_EXECUTE(object, return NvResult::InvalidState;);
-        const auto map = FindBufferMap(object->dma_map_addr);
-        if (!map) {
-            LOG_ERROR(Service_NVDRV, "Tried to submit an invalid offset 0x{:X} dma 0x{:X}",
-                      object->addr, object->dma_map_addr);
-            return NvResult::Success;
-        }
         Tegra::ChCommandHeaderList cmdlist(cmd_buffer.word_count);
-        gpu.MemoryManager().ReadBlock(map->StartAddr() + cmd_buffer.offset, cmdlist.data(),
-                                      cmdlist.size() * sizeof(u32));
+        system.Memory().ReadBlock(object->addr + cmd_buffer.offset, cmdlist.data(),
+                                  cmdlist.size() * sizeof(u32));
         gpu.PushCommandBuffer(cmdlist);
     }
     if (gpu.UseNvdec()) {
-
         fences[0].value = syncpoint_manager.IncreaseSyncpoint(fences[0].id, 1);
-
         Tegra::ChCommandHeaderList cmdlist{{(4 << 28) | fences[0].id}};
         gpu.PushCommandBuffer(cmdlist);
     }
     std::memcpy(output.data(), &params, sizeof(IoctlSubmit));
     // Some games expect command_buffers to be written back
     offset = sizeof(IoctlSubmit);
-    offset = WriteVectors(output, command_buffers, offset);
-    offset = WriteVectors(output, relocs, offset);
-    offset = WriteVectors(output, reloc_shifts, offset);
-    offset = WriteVectors(output, syncpt_increments, offset);
-    offset = WriteVectors(output, wait_checks, offset);
-    offset = WriteVectors(output, fences, offset);
+    offset += WriteVectors(output, command_buffers, offset);
+    offset += WriteVectors(output, relocs, offset);
+    offset += WriteVectors(output, reloc_shifts, offset);
+    offset += WriteVectors(output, syncpt_increments, offset);
+    offset += WriteVectors(output, wait_checks, offset);
+    offset += WriteVectors(output, fences, offset);
 
     return NvResult::Success;
 }
@@ -148,14 +143,14 @@ NvResult nvhost_nvdec_common::MapBuffer(const std::vector<u8>& input, std::vecto
     std::memcpy(&params, input.data(), sizeof(IoctlMapBuffer));
     std::vector<MapBufferEntry> cmd_buffer_handles(params.num_entries);
 
-    SpliceVectors(input, cmd_buffer_handles, params.num_entries, sizeof(IoctlMapBuffer));
+    SliceVectors(input, cmd_buffer_handles, params.num_entries, sizeof(IoctlMapBuffer));
 
     auto& gpu = system.GPU();
 
-    for (auto& cmf_buff : cmd_buffer_handles) {
-        auto object{nvmap_dev->GetObject(cmf_buff.map_handle)};
+    for (auto& cmd_buffer : cmd_buffer_handles) {
+        auto object{nvmap_dev->GetObject(cmd_buffer.map_handle)};
         if (!object) {
-            LOG_ERROR(Service_NVDRV, "invalid cmd_buffer nvmap_handle={:X}", cmf_buff.map_handle);
+            LOG_ERROR(Service_NVDRV, "invalid cmd_buffer nvmap_handle={:X}", cmd_buffer.map_handle);
             std::memcpy(output.data(), &params, output.size());
             return NvResult::InvalidState;
         }
@@ -170,7 +165,7 @@ NvResult nvhost_nvdec_common::MapBuffer(const std::vector<u8>& input, std::vecto
         if (!object->dma_map_addr) {
             LOG_ERROR(Service_NVDRV, "failed to map size={}", object->size);
         } else {
-            cmf_buff.map_address = object->dma_map_addr;
+            cmd_buffer.map_address = object->dma_map_addr;
             AddBufferMap(object->dma_map_addr, object->size, object->addr,
                          object->status == nvmap::Object::Status::Allocated);
         }
@@ -186,14 +181,14 @@ NvResult nvhost_nvdec_common::UnmapBuffer(const std::vector<u8>& input, std::vec
     IoctlMapBuffer params{};
     std::memcpy(&params, input.data(), sizeof(IoctlMapBuffer));
     std::vector<MapBufferEntry> cmd_buffer_handles(params.num_entries);
-    SpliceVectors(input, cmd_buffer_handles, params.num_entries, sizeof(IoctlMapBuffer));
+    SliceVectors(input, cmd_buffer_handles, params.num_entries, sizeof(IoctlMapBuffer));
 
     auto& gpu = system.GPU();
 
-    for (auto& cmf_buff : cmd_buffer_handles) {
-        const auto object{nvmap_dev->GetObject(cmf_buff.map_handle)};
+    for (auto& cmd_buffer : cmd_buffer_handles) {
+        const auto object{nvmap_dev->GetObject(cmd_buffer.map_handle)};
         if (!object) {
-            LOG_ERROR(Service_NVDRV, "invalid cmd_buffer nvmap_handle={:X}", cmf_buff.map_handle);
+            LOG_ERROR(Service_NVDRV, "invalid cmd_buffer nvmap_handle={:X}", cmd_buffer.map_handle);
             std::memcpy(output.data(), &params, output.size());
             return NvResult::InvalidState;
         }
diff --git a/src/core/hle/service/nvdrv/nvdrv.cpp b/src/core/hle/service/nvdrv/nvdrv.cpp
index 74796dce1..03992af5e 100644
--- a/src/core/hle/service/nvdrv/nvdrv.cpp
+++ b/src/core/hle/service/nvdrv/nvdrv.cpp
@@ -20,8 +20,8 @@
 #include "core/hle/service/nvdrv/devices/nvhost_nvjpg.h"
 #include "core/hle/service/nvdrv/devices/nvhost_vic.h"
 #include "core/hle/service/nvdrv/devices/nvmap.h"
-#include "core/hle/service/nvdrv/interface.h"
 #include "core/hle/service/nvdrv/nvdrv.h"
+#include "core/hle/service/nvdrv/nvdrv_interface.h"
 #include "core/hle/service/nvdrv/nvmemp.h"
 #include "core/hle/service/nvdrv/syncpoint_manager.h"
 #include "core/hle/service/nvflinger/nvflinger.h"
diff --git a/src/core/hle/service/nvdrv/interface.cpp b/src/core/hle/service/nvdrv/nvdrv_interface.cpp
index e4d495000..d61fb73dc 100644
--- a/src/core/hle/service/nvdrv/interface.cpp
+++ b/src/core/hle/service/nvdrv/nvdrv_interface.cpp
@@ -10,9 +10,9 @@
 #include "core/hle/kernel/k_thread.h"
 #include "core/hle/kernel/k_writable_event.h"
 #include "core/hle/kernel/kernel.h"
-#include "core/hle/service/nvdrv/interface.h"
 #include "core/hle/service/nvdrv/nvdata.h"
 #include "core/hle/service/nvdrv/nvdrv.h"
+#include "core/hle/service/nvdrv/nvdrv_interface.h"
 
 namespace Service::Nvidia {
 
diff --git a/src/core/hle/service/nvdrv/interface.h b/src/core/hle/service/nvdrv/nvdrv_interface.h
index 0e764c53f..0e764c53f 100644
--- a/src/core/hle/service/nvdrv/interface.h
+++ b/src/core/hle/service/nvdrv/nvdrv_interface.h
diff --git a/src/core/hle/service/pctl/pctl.h b/src/core/hle/service/pctl/pctl.h
index ea3b97823..1d28900b2 100644
--- a/src/core/hle/service/pctl/pctl.h
+++ b/src/core/hle/service/pctl/pctl.h
@@ -4,7 +4,7 @@
 
 #pragma once
 
-#include "core/hle/service/pctl/module.h"
+#include "core/hle/service/pctl/pctl_module.h"
 
 namespace Core {
 class System;
diff --git a/src/core/hle/service/pctl/module.cpp b/src/core/hle/service/pctl/pctl_module.cpp
index 1e31d05a6..6949fcf3b 100644
--- a/src/core/hle/service/pctl/module.cpp
+++ b/src/core/hle/service/pctl/pctl_module.cpp
@@ -8,8 +8,8 @@
 #include "core/file_sys/patch_manager.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/kernel/k_process.h"
-#include "core/hle/service/pctl/module.h"
 #include "core/hle/service/pctl/pctl.h"
+#include "core/hle/service/pctl/pctl_module.h"
 
 namespace Service::PCTL {
 
diff --git a/src/core/hle/service/pctl/module.h b/src/core/hle/service/pctl/pctl_module.h
index f25c5c557..f25c5c557 100644
--- a/src/core/hle/service/pctl/module.h
+++ b/src/core/hle/service/pctl/pctl_module.h
diff --git a/src/core/hle/service/service.cpp b/src/core/hle/service/service.cpp
index 663b83cd3..e6fba88b2 100644
--- a/src/core/hle/service/service.cpp
+++ b/src/core/hle/service/service.cpp
@@ -21,7 +21,7 @@
 #include "core/hle/service/aoc/aoc_u.h"
 #include "core/hle/service/apm/apm.h"
 #include "core/hle/service/audio/audio.h"
-#include "core/hle/service/bcat/module.h"
+#include "core/hle/service/bcat/bcat_module.h"
 #include "core/hle/service/bpc/bpc.h"
 #include "core/hle/service/btdrv/btdrv.h"
 #include "core/hle/service/btm/btm.h"
@@ -54,7 +54,7 @@
 #include "core/hle/service/nvflinger/nvflinger.h"
 #include "core/hle/service/olsc/olsc.h"
 #include "core/hle/service/pcie/pcie.h"
-#include "core/hle/service/pctl/module.h"
+#include "core/hle/service/pctl/pctl_module.h"
 #include "core/hle/service/pcv/pcv.h"
 #include "core/hle/service/pm/pm.h"
 #include "core/hle/service/prepo/prepo.h"
@@ -64,7 +64,7 @@
 #include "core/hle/service/set/settings.h"
 #include "core/hle/service/sm/sm.h"
 #include "core/hle/service/sockets/sockets.h"
-#include "core/hle/service/spl/module.h"
+#include "core/hle/service/spl/spl_module.h"
 #include "core/hle/service/ssl/ssl.h"
 #include "core/hle/service/time/time.h"
 #include "core/hle/service/usb/usb.h"
diff --git a/src/core/hle/service/sm/sm.cpp b/src/core/hle/service/sm/sm.cpp
index c7828c3bd..15034abed 100644
--- a/src/core/hle/service/sm/sm.cpp
+++ b/src/core/hle/service/sm/sm.cpp
@@ -14,8 +14,8 @@
 #include "core/hle/kernel/k_server_session.h"
 #include "core/hle/kernel/k_session.h"
 #include "core/hle/result.h"
-#include "core/hle/service/sm/controller.h"
 #include "core/hle/service/sm/sm.h"
+#include "core/hle/service/sm/sm_controller.h"
 
 namespace Service::SM {
 
diff --git a/src/core/hle/service/sm/controller.cpp b/src/core/hle/service/sm/sm_controller.cpp
index 8b9418e0f..b5fbc4569 100644
--- a/src/core/hle/service/sm/controller.cpp
+++ b/src/core/hle/service/sm/sm_controller.cpp
@@ -13,7 +13,7 @@
 #include "core/hle/kernel/k_server_port.h"
 #include "core/hle/kernel/k_server_session.h"
 #include "core/hle/kernel/k_session.h"
-#include "core/hle/service/sm/controller.h"
+#include "core/hle/service/sm/sm_controller.h"
 
 namespace Service::SM {
 
diff --git a/src/core/hle/service/sm/controller.h b/src/core/hle/service/sm/sm_controller.h
index 7494f898d..7494f898d 100644
--- a/src/core/hle/service/sm/controller.h
+++ b/src/core/hle/service/sm/sm_controller.h
diff --git a/src/core/hle/service/spl/csrng.h b/src/core/hle/service/spl/csrng.h
index 5c0bd2199..0d03cc6cb 100644
--- a/src/core/hle/service/spl/csrng.h
+++ b/src/core/hle/service/spl/csrng.h
@@ -4,7 +4,7 @@
 
 #pragma once
 
-#include "core/hle/service/spl/module.h"
+#include "core/hle/service/spl/spl_module.h"
 
 namespace Core {
 class System;
diff --git a/src/core/hle/service/spl/spl.h b/src/core/hle/service/spl/spl.h
index 9b35012ed..5599c0c01 100644
--- a/src/core/hle/service/spl/spl.h
+++ b/src/core/hle/service/spl/spl.h
@@ -4,7 +4,7 @@
 
 #pragma once
 
-#include "core/hle/service/spl/module.h"
+#include "core/hle/service/spl/spl_module.h"
 
 namespace Core {
 class System;
diff --git a/src/core/hle/service/spl/module.cpp b/src/core/hle/service/spl/spl_module.cpp
index ebb179aa8..918633af5 100644
--- a/src/core/hle/service/spl/module.cpp
+++ b/src/core/hle/service/spl/spl_module.cpp
@@ -13,8 +13,8 @@
 #include "core/hle/api_version.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/service/spl/csrng.h"
-#include "core/hle/service/spl/module.h"
 #include "core/hle/service/spl/spl.h"
+#include "core/hle/service/spl/spl_module.h"
 
 namespace Service::SPL {
 
diff --git a/src/core/hle/service/spl/module.h b/src/core/hle/service/spl/spl_module.h
index 61630df80..61630df80 100644
--- a/src/core/hle/service/spl/module.h
+++ b/src/core/hle/service/spl/spl_module.h
diff --git a/src/core/hle/service/time/time.cpp b/src/core/hle/service/time/time.cpp
index d6f710eba..8fdd5076f 100644
--- a/src/core/hle/service/time/time.cpp
+++ b/src/core/hle/service/time/time.cpp
@@ -11,8 +11,8 @@
 #include "core/hle/kernel/k_client_port.h"
 #include "core/hle/kernel/k_scheduler.h"
 #include "core/hle/kernel/kernel.h"
-#include "core/hle/service/time/interface.h"
 #include "core/hle/service/time/time.h"
+#include "core/hle/service/time/time_interface.h"
 #include "core/hle/service/time/time_sharedmemory.h"
 #include "core/hle/service/time/time_zone_service.h"
 
diff --git a/src/core/hle/service/time/interface.cpp b/src/core/hle/service/time/time_interface.cpp
index 53a204796..bb7b6b5c1 100644
--- a/src/core/hle/service/time/interface.cpp
+++ b/src/core/hle/service/time/time_interface.cpp
@@ -2,7 +2,7 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include "core/hle/service/time/interface.h"
+#include "core/hle/service/time/time_interface.h"
 
 namespace Service::Time {
 
diff --git a/src/core/hle/service/time/interface.h b/src/core/hle/service/time/time_interface.h
index c41766f1a..c41766f1a 100644
--- a/src/core/hle/service/time/interface.h
+++ b/src/core/hle/service/time/time_interface.h
diff --git a/src/input_common/main.cpp b/src/input_common/main.cpp
index 7399c3648..8de3d4520 100644
--- a/src/input_common/main.cpp
+++ b/src/input_common/main.cpp
@@ -294,8 +294,8 @@ void InputSubsystem::ReloadInputDevices() {
     impl->udp->ReloadSockets();
 }
 
-std::vector<std::unique_ptr<Polling::DevicePoller>> InputSubsystem::GetPollers(
-    Polling::DeviceType type) const {
+std::vector<std::unique_ptr<Polling::DevicePoller>> InputSubsystem::GetPollers([
+    [maybe_unused]] Polling::DeviceType type) const {
 #ifdef HAVE_SDL2
     return impl->sdl->GetPollers(type);
 #else
diff --git a/src/input_common/mouse/mouse_poller.cpp b/src/input_common/mouse/mouse_poller.cpp
index 45b3d7340..efcdd85d2 100644
--- a/src/input_common/mouse/mouse_poller.cpp
+++ b/src/input_common/mouse/mouse_poller.cpp
@@ -84,7 +84,7 @@ public:
         std::lock_guard lock{mutex};
         const auto axis_value =
             static_cast<float>(mouse_input->GetMouseState(button).axis.at(axis));
-        const float sensitivity = Settings::values.mouse_panning_sensitivity.GetValue();
+        const float sensitivity = Settings::values.mouse_panning_sensitivity.GetValue() * 0.10f;
         return axis_value * sensitivity / (100.0f * range);
     }
 
diff --git a/src/input_common/sdl/sdl_impl.h b/src/input_common/sdl/sdl_impl.h
index b77afcbd8..7a9ad6346 100644
--- a/src/input_common/sdl/sdl_impl.h
+++ b/src/input_common/sdl/sdl_impl.h
@@ -10,15 +10,7 @@
 #include <thread>
 #include <unordered_map>
 
-// Ignore -Wimplicit-fallthrough due to https://github.com/libsdl-org/SDL/issues/4307
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wimplicit-fallthrough"
-#endif
 #include <SDL.h>
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
 
 #include "common/common_types.h"
 #include "common/threadsafe_queue.h"
diff --git a/src/tests/video_core/buffer_base.cpp b/src/tests/video_core/buffer_base.cpp
index edced69bb..9f5a54de4 100644
--- a/src/tests/video_core/buffer_base.cpp
+++ b/src/tests/video_core/buffer_base.cpp
@@ -536,7 +536,7 @@ TEST_CASE("BufferBase: Cached write downloads") {
     REQUIRE(rasterizer.Count() == 63);
     buffer.MarkRegionAsGpuModified(c + PAGE, PAGE);
     int num = 0;
-    buffer.ForEachDownloadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
+    buffer.ForEachDownloadRangeAndClear(c, WORD, [&](u64 offset, u64 size) { ++num; });
     buffer.ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
     REQUIRE(num == 0);
     REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));
diff --git a/src/video_core/buffer_cache/buffer_base.h b/src/video_core/buffer_cache/buffer_base.h
index b121d36a3..c3318095c 100644
--- a/src/video_core/buffer_cache/buffer_base.h
+++ b/src/video_core/buffer_cache/buffer_base.h
@@ -226,19 +226,24 @@ public:
     /// Call 'func' for each CPU modified range and unmark those pages as CPU modified
     template <typename Func>
     void ForEachUploadRange(VAddr query_cpu_range, u64 size, Func&& func) {
-        ForEachModifiedRange<Type::CPU>(query_cpu_range, size, func);
+        ForEachModifiedRange<Type::CPU>(query_cpu_range, size, true, func);
     }
 
     /// Call 'func' for each GPU modified range and unmark those pages as GPU modified
     template <typename Func>
-    void ForEachDownloadRange(VAddr query_cpu_range, u64 size, Func&& func) {
-        ForEachModifiedRange<Type::GPU>(query_cpu_range, size, func);
+    void ForEachDownloadRange(VAddr query_cpu_range, u64 size, bool clear, Func&& func) {
+        ForEachModifiedRange<Type::GPU>(query_cpu_range, size, clear, func);
+    }
+
+    template <typename Func>
+    void ForEachDownloadRangeAndClear(VAddr query_cpu_range, u64 size, Func&& func) {
+        ForEachModifiedRange<Type::GPU>(query_cpu_range, size, true, func);
     }
 
     /// Call 'func' for each GPU modified range and unmark those pages as GPU modified
     template <typename Func>
     void ForEachDownloadRange(Func&& func) {
-        ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), func);
+        ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), true, func);
     }
 
     /// Mark buffer as picked
@@ -415,7 +420,7 @@ private:
      * @param func            Function to call for each turned off region
      */
     template <Type type, typename Func>
-    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) {
+    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, bool clear, Func&& func) {
         static_assert(type != Type::Untracked);
 
         const s64 difference = query_cpu_range - cpu_addr;
@@ -467,7 +472,9 @@ private:
             bits = (bits << left_offset) >> left_offset;
 
             const u64 current_word = state_words[word_index] & bits;
-            state_words[word_index] &= ~bits;
+            if (clear) {
+                state_words[word_index] &= ~bits;
+            }
 
             if constexpr (type == Type::CPU) {
                 const u64 current_bits = untracked_words[word_index] & bits;
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index cad7f902d..2871682f6 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -15,6 +15,7 @@
 #include <vector>
 
 #include <boost/container/small_vector.hpp>
+#include <boost/icl/interval_set.hpp>
 
 #include "common/common_types.h"
 #include "common/div_ceil.h"
@@ -77,6 +78,9 @@ class BufferCache {
     using Runtime = typename P::Runtime;
     using Buffer = typename P::Buffer;
 
+    using IntervalSet = boost::icl::interval_set<VAddr>;
+    using IntervalType = typename IntervalSet::interval_type;
+
     struct Empty {};
 
     struct OverlapResult {
@@ -148,18 +152,26 @@ public:
     /// Return true when there are uncommitted buffers to be downloaded
     [[nodiscard]] bool HasUncommittedFlushes() const noexcept;
 
+    void AccumulateFlushes();
+
     /// Return true when the caller should wait for async downloads
     [[nodiscard]] bool ShouldWaitAsyncFlushes() const noexcept;
 
     /// Commit asynchronous downloads
     void CommitAsyncFlushes();
+    void CommitAsyncFlushesHigh();
 
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
 
+    [[nodiscard]] bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount);
+
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
+    /// Return true when a CPU region is modified from the CPU
+    [[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size);
+
     std::mutex mutex;
 
 private:
@@ -190,6 +202,36 @@ private:
         }
     }
 
+    template <typename Func>
+    void ForEachWrittenRange(VAddr cpu_addr, u64 size, Func&& func) {
+        const VAddr start_address = cpu_addr;
+        const VAddr end_address = start_address + size;
+        const VAddr search_base =
+            static_cast<VAddr>(std::min<s64>(0LL, static_cast<s64>(start_address - size)));
+        const IntervalType search_interval{search_base, search_base + 1};
+        auto it = common_ranges.lower_bound(search_interval);
+        if (it == common_ranges.end()) {
+            it = common_ranges.begin();
+        }
+        for (; it != common_ranges.end(); it++) {
+            VAddr inter_addr_end = it->upper();
+            VAddr inter_addr = it->lower();
+            if (inter_addr >= end_address) {
+                break;
+            }
+            if (inter_addr_end <= start_address) {
+                continue;
+            }
+            if (inter_addr_end > end_address) {
+                inter_addr_end = end_address;
+            }
+            if (inter_addr < start_address) {
+                inter_addr = start_address;
+            }
+            func(inter_addr, inter_addr_end);
+        }
+    }
+
     static bool IsRangeGranular(VAddr cpu_addr, size_t size) {
         return (cpu_addr & ~Core::Memory::PAGE_MASK) ==
                ((cpu_addr + size) & ~Core::Memory::PAGE_MASK);
@@ -272,8 +314,6 @@ private:
 
     void DeleteBuffer(BufferId buffer_id);
 
-    void ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id);
-
     void NotifyBufferDeletion();
 
     [[nodiscard]] Binding StorageBufferBinding(GPUVAddr ssbo_addr) const;
@@ -327,9 +367,9 @@ private:
 
     std::vector<BufferId> cached_write_buffer_ids;
 
-    // TODO: This data structure is not optimal and it should be reworked
-    std::vector<BufferId> uncommitted_downloads;
-    std::deque<std::vector<BufferId>> committed_downloads;
+    IntervalSet uncommitted_ranges;
+    IntervalSet common_ranges;
+    std::deque<IntervalSet> committed_ranges;
 
     size_t immediate_buffer_capacity = 0;
     std::unique_ptr<u8[]> immediate_buffer_alloc;
@@ -352,6 +392,7 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
     // Ensure the first slot is used for the null buffer
     void(slot_buffers.insert(runtime, NullBufferParams{}));
     deletion_iterator = slot_buffers.end();
+    common_ranges.clear();
 }
 
 template <class P>
@@ -422,6 +463,68 @@ void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) {
 }
 
 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);
+    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);
+    }
+
+    BufferId buffer_a;
+    BufferId buffer_b;
+    do {
+        has_deleted_buffers = false;
+        buffer_a = FindBuffer(*cpu_src_address, static_cast<u32>(amount));
+        buffer_b = FindBuffer(*cpu_dest_address, static_cast<u32>(amount));
+    } while (has_deleted_buffers);
+    auto& src_buffer = slot_buffers[buffer_a];
+    auto& dest_buffer = slot_buffers[buffer_b];
+    SynchronizeBuffer(src_buffer, *cpu_src_address, static_cast<u32>(amount));
+    SynchronizeBuffer(dest_buffer, *cpu_dest_address, static_cast<u32>(amount));
+    std::array copies{BufferCopy{
+        .src_offset = src_buffer.Offset(*cpu_src_address),
+        .dst_offset = dest_buffer.Offset(*cpu_dest_address),
+        .size = amount,
+    }};
+
+    boost::container::small_vector<IntervalType, 4> tmp_intervals;
+    auto mirror = [&](VAddr base_address, VAddr base_address_end) {
+        const u64 size = base_address_end - base_address;
+        const VAddr diff = base_address - *cpu_src_address;
+        const VAddr new_base_address = *cpu_dest_address + diff;
+        const IntervalType add_interval{new_base_address, new_base_address + size};
+        uncommitted_ranges.add(add_interval);
+        tmp_intervals.push_back(add_interval);
+    };
+    ForEachWrittenRange(*cpu_src_address, amount, mirror);
+    // This subtraction in this order is important for overlapping copies.
+    common_ranges.subtract(subtract_interval);
+    for (const IntervalType add_interval : tmp_intervals) {
+        common_ranges.add(add_interval);
+    }
+
+    runtime.CopyBuffer(dest_buffer, src_buffer, copies);
+    if (source_dirty) {
+        dest_buffer.MarkRegionAsGpuModified(*cpu_dest_address, amount);
+    }
+    std::vector<u8> tmp_buffer(amount);
+    cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount);
+    cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount);
+    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);
@@ -547,29 +650,30 @@ void BufferCache<P>::FlushCachedWrites() {
 
 template <class P>
 bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
-    return !uncommitted_downloads.empty();
+    return !uncommitted_ranges.empty() || !committed_ranges.empty();
 }
 
 template <class P>
-bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
-    return !committed_downloads.empty() && !committed_downloads.front().empty();
+void BufferCache<P>::AccumulateFlushes() {
+    if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) {
+        uncommitted_ranges.clear();
+        return;
+    }
+    if (uncommitted_ranges.empty()) {
+        return;
+    }
+    committed_ranges.emplace_back(std::move(uncommitted_ranges));
 }
 
 template <class P>
-void BufferCache<P>::CommitAsyncFlushes() {
-    // This is intentionally passing the value by copy
-    committed_downloads.push_front(uncommitted_downloads);
-    uncommitted_downloads.clear();
+bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
+    return false;
 }
 
 template <class P>
-void BufferCache<P>::PopAsyncFlushes() {
-    if (committed_downloads.empty()) {
-        return;
-    }
-    auto scope_exit_pop_download = detail::ScopeExit([this] { committed_downloads.pop_back(); });
-    const std::span<const BufferId> download_ids = committed_downloads.back();
-    if (download_ids.empty()) {
+void BufferCache<P>::CommitAsyncFlushesHigh() {
+    AccumulateFlushes();
+    if (committed_ranges.empty()) {
         return;
     }
     MICROPROFILE_SCOPE(GPU_DownloadMemory);
@@ -577,20 +681,43 @@ void BufferCache<P>::PopAsyncFlushes() {
     boost::container::small_vector<std::pair<BufferCopy, BufferId>, 1> downloads;
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
-    for (const BufferId buffer_id : download_ids) {
-        slot_buffers[buffer_id].ForEachDownloadRange([&](u64 range_offset, u64 range_size) {
-            downloads.push_back({
-                BufferCopy{
-                    .src_offset = range_offset,
-                    .dst_offset = total_size_bytes,
-                    .size = range_size,
-                },
-                buffer_id,
+    for (const IntervalSet& intervals : committed_ranges) {
+        for (auto& interval : intervals) {
+            const std::size_t size = interval.upper() - interval.lower();
+            const VAddr cpu_addr = interval.lower();
+            ForEachBufferInRange(cpu_addr, size, [&](BufferId buffer_id, Buffer& buffer) {
+                boost::container::small_vector<BufferCopy, 1> copies;
+                buffer.ForEachDownloadRangeAndClear(
+                    cpu_addr, size, [&](u64 range_offset, u64 range_size) {
+                        const VAddr buffer_addr = buffer.CpuAddr();
+                        const auto add_download = [&](VAddr start, VAddr end) {
+                            const u64 new_offset = start - buffer_addr;
+                            const u64 new_size = end - start;
+                            downloads.push_back({
+                                BufferCopy{
+                                    .src_offset = new_offset,
+                                    .dst_offset = total_size_bytes,
+                                    .size = new_size,
+                                },
+                                buffer_id,
+                            });
+                            // Align up to avoid cache conflicts
+                            constexpr u64 align = 256ULL;
+                            constexpr u64 mask = ~(align - 1ULL);
+                            total_size_bytes += (new_size + align - 1) & mask;
+                            largest_copy = std::max(largest_copy, new_size);
+                        };
+
+                        const VAddr start_address = buffer_addr + range_offset;
+                        const VAddr end_address = start_address + range_size;
+                        ForEachWrittenRange(start_address, range_size, add_download);
+                        const IntervalType subtract_interval{start_address, end_address};
+                        common_ranges.subtract(subtract_interval);
+                    });
             });
-            total_size_bytes += range_size;
-            largest_copy = std::max(largest_copy, range_size);
-        });
+        }
     }
+    committed_ranges.clear();
     if (downloads.empty()) {
         return;
     }
@@ -623,6 +750,19 @@ void BufferCache<P>::PopAsyncFlushes() {
 }
 
 template <class P>
+void BufferCache<P>::CommitAsyncFlushes() {
+    if (Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High) {
+        CommitAsyncFlushesHigh();
+    } else {
+        uncommitted_ranges.clear();
+        committed_ranges.clear();
+    }
+}
+
+template <class P>
+void BufferCache<P>::PopAsyncFlushes() {}
+
+template <class P>
 bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
     const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE);
     for (u64 page = addr >> PAGE_BITS; page < page_end;) {
@@ -642,6 +782,25 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
 }
 
 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;) {
+        const BufferId image_id = page_table[page];
+        if (!image_id) {
+            ++page;
+            continue;
+        }
+        Buffer& buffer = slot_buffers[image_id];
+        if (buffer.IsRegionCpuModified(addr, size)) {
+            return true;
+        }
+        const VAddr end_addr = buffer.CpuAddr() + buffer.SizeBytes();
+        page = Common::DivCeil(end_addr, PAGE_SIZE);
+    }
+    return false;
+}
+
+template <class P>
 void BufferCache<P>::BindHostIndexBuffer() {
     Buffer& buffer = slot_buffers[index_buffer.buffer_id];
     TouchBuffer(buffer);
@@ -649,7 +808,9 @@ void BufferCache<P>::BindHostIndexBuffer() {
     const u32 size = index_buffer.size;
     SynchronizeBuffer(buffer, index_buffer.cpu_addr, size);
     if constexpr (HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) {
-        runtime.BindIndexBuffer(buffer, offset, size);
+        const u32 new_offset = offset + maxwell3d.regs.index_array.first *
+                                            maxwell3d.regs.index_array.FormatSizeInBytes();
+        runtime.BindIndexBuffer(buffer, new_offset, size);
     } else {
         runtime.BindIndexBuffer(maxwell3d.regs.draw.topology, maxwell3d.regs.index_array.format,
                                 maxwell3d.regs.index_array.first, maxwell3d.regs.index_array.count,
@@ -863,7 +1024,7 @@ void BufferCache<P>::UpdateIndexBuffer() {
     const GPUVAddr gpu_addr_end = index_array.EndAddress();
     const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr_begin);
     const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin);
-    const u32 draw_size = index_array.count * index_array.FormatSizeInBytes();
+    const u32 draw_size = (index_array.count + index_array.first) * index_array.FormatSizeInBytes();
     const u32 size = std::min(address_size, draw_size);
     if (size == 0 || !cpu_addr) {
         index_buffer = NULL_BINDING;
@@ -1010,16 +1171,16 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s
     Buffer& buffer = slot_buffers[buffer_id];
     buffer.MarkRegionAsGpuModified(cpu_addr, size);
 
-    const bool is_accuracy_high = Settings::IsGPULevelHigh();
+    const IntervalType base_interval{cpu_addr, cpu_addr + size};
+    common_ranges.add(base_interval);
+
+    const bool is_accuracy_high =
+        Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High;
     const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
-    if (!is_accuracy_high || !is_async) {
-        return;
-    }
-    if (std::ranges::find(uncommitted_downloads, buffer_id) != uncommitted_downloads.end()) {
-        // Already inserted
+    if (!is_async && !is_accuracy_high) {
         return;
     }
-    uncommitted_downloads.push_back(buffer_id);
+    uncommitted_ranges.add(base_interval);
 }
 
 template <class P>
@@ -1103,7 +1264,6 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
     if (!copies.empty()) {
         runtime.CopyBuffer(slot_buffers[new_buffer_id], overlap, copies);
     }
-    ReplaceBufferDownloads(overlap_id, new_buffer_id);
     DeleteBuffer(overlap_id);
 }
 
@@ -1244,14 +1404,28 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
     boost::container::small_vector<BufferCopy, 1> copies;
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
-    buffer.ForEachDownloadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
-        copies.push_back(BufferCopy{
-            .src_offset = range_offset,
-            .dst_offset = total_size_bytes,
-            .size = range_size,
-        });
-        total_size_bytes += range_size;
-        largest_copy = std::max(largest_copy, range_size);
+    buffer.ForEachDownloadRangeAndClear(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
+        const VAddr buffer_addr = buffer.CpuAddr();
+        const auto add_download = [&](VAddr start, VAddr end) {
+            const u64 new_offset = start - buffer_addr;
+            const u64 new_size = end - start;
+            copies.push_back(BufferCopy{
+                .src_offset = new_offset,
+                .dst_offset = total_size_bytes,
+                .size = new_size,
+            });
+            // Align up to avoid cache conflicts
+            constexpr u64 align = 256ULL;
+            constexpr u64 mask = ~(align - 1ULL);
+            total_size_bytes += (new_size + align - 1) & mask;
+            largest_copy = std::max(largest_copy, new_size);
+        };
+
+        const VAddr start_address = buffer_addr + range_offset;
+        const VAddr end_address = start_address + range_size;
+        ForEachWrittenRange(start_address, range_size, add_download);
+        const IntervalType subtract_interval{start_address, end_address};
+        common_ranges.subtract(subtract_interval);
     });
     if (total_size_bytes == 0) {
         return;
@@ -1316,18 +1490,6 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id) {
 }
 
 template <class P>
-void BufferCache<P>::ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id) {
-    const auto replace = [old_buffer_id, new_buffer_id](std::vector<BufferId>& buffers) {
-        std::ranges::replace(buffers, old_buffer_id, new_buffer_id);
-        if (auto it = std::ranges::find(buffers, new_buffer_id); it != buffers.end()) {
-            buffers.erase(std::remove(it + 1, buffers.end(), new_buffer_id), buffers.end());
-        }
-    };
-    replace(uncommitted_downloads);
-    std::ranges::for_each(committed_downloads, replace);
-}
-
-template <class P>
 void BufferCache<P>::NotifyBufferDeletion() {
     if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
         dirty_uniform_buffers.fill(~u32{0});
@@ -1349,15 +1511,9 @@ typename BufferCache<P>::Binding BufferCache<P>::StorageBufferBinding(GPUVAddr s
     if (!cpu_addr || size == 0) {
         return NULL_BINDING;
     }
-    // HACK(Rodrigo): This is the number of bytes bound in host beyond the guest API's range.
-    // It exists due to some games like Astral Chain operate out of bounds.
-    // Binding the whole map range would be technically correct, but games have large maps that make
-    // this approach unaffordable for now.
-    static constexpr u32 arbitrary_extra_bytes = 0xc000;
-    const u32 bytes_to_map_end = static_cast<u32>(gpu_memory.BytesToMapEnd(gpu_addr));
     const Binding binding{
         .cpu_addr = *cpu_addr,
-        .size = std::min(size + arbitrary_extra_bytes, bytes_to_map_end),
+        .size = size,
         .buffer_id = BufferId{},
     };
     return binding;
diff --git a/src/video_core/command_classes/vic.cpp b/src/video_core/command_classes/vic.cpp
index ff3db0aee..ffb7c82a1 100644
--- a/src/video_core/command_classes/vic.cpp
+++ b/src/video_core/command_classes/vic.cpp
@@ -129,28 +129,27 @@ void Vic::Execute() {
 
         const std::size_t surface_width = config.surface_width_minus1 + 1;
         const std::size_t surface_height = config.surface_height_minus1 + 1;
-        const std::size_t half_width = surface_width / 2;
-        const std::size_t half_height = config.surface_height_minus1 / 2;
+        const auto frame_width = std::min(surface_width, static_cast<size_t>(frame->width));
+        const auto frame_height = std::min(surface_height, static_cast<size_t>(frame->height));
+        const std::size_t half_width = frame_width / 2;
+        const std::size_t half_height = frame_height / 2;
         const std::size_t aligned_width = (surface_width + 0xff) & ~0xff;
 
         const auto* luma_ptr = frame->data[0];
         const auto* chroma_b_ptr = frame->data[1];
         const auto* chroma_r_ptr = frame->data[2];
-        const auto stride = frame->linesize[0];
-        const auto half_stride = frame->linesize[1];
+        const auto stride = static_cast<size_t>(frame->linesize[0]);
+        const auto half_stride = static_cast<size_t>(frame->linesize[1]);
 
         luma_buffer.resize(aligned_width * surface_height);
-        chroma_buffer.resize(aligned_width * half_height);
+        chroma_buffer.resize(aligned_width * surface_height / 2);
 
         // Populate luma buffer
-        for (std::size_t y = 0; y < surface_height - 1; ++y) {
+        for (std::size_t y = 0; y < frame_height; ++y) {
             const std::size_t src = y * stride;
             const std::size_t dst = y * aligned_width;
-
-            const std::size_t size = surface_width;
-
-            for (std::size_t offset = 0; offset < size; ++offset) {
-                luma_buffer[dst + offset] = luma_ptr[src + offset];
+            for (std::size_t x = 0; x < frame_width; ++x) {
+                luma_buffer[dst + x] = luma_ptr[src + x];
             }
         }
         gpu.MemoryManager().WriteBlock(output_surface_luma_address, luma_buffer.data(),
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 8b33c04ab..8d28bd884 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -4,6 +4,7 @@
 
 #include "common/cityhash.h"
 #include "common/microprofile.h"
+#include "common/settings.h"
 #include "core/core.h"
 #include "core/memory.h"
 #include "video_core/dma_pusher.h"
@@ -76,8 +77,13 @@ bool DmaPusher::Step() {
 
         // Push buffer non-empty, read a word
         command_headers.resize(command_list_header.size);
-        gpu.MemoryManager().ReadBlockUnsafe(dma_get, command_headers.data(),
-                                            command_list_header.size * sizeof(u32));
+        if (Settings::IsGPULevelHigh()) {
+            gpu.MemoryManager().ReadBlock(dma_get, command_headers.data(),
+                                          command_list_header.size * sizeof(u32));
+        } else {
+            gpu.MemoryManager().ReadBlockUnsafe(dma_get, command_headers.data(),
+                                                command_list_header.size * sizeof(u32));
+        }
     }
     for (std::size_t index = 0; index < command_headers.size();) {
         const CommandHeader& command_header = command_headers[index];
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 2ee980bab..24481952b 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -21,6 +21,10 @@ MaxwellDMA::MaxwellDMA(Core::System& system_, MemoryManager& memory_manager_)
 
 MaxwellDMA::~MaxwellDMA() = default;
 
+void MaxwellDMA::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
+    rasterizer = rasterizer_;
+}
+
 void MaxwellDMA::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
     ASSERT_MSG(method < NUM_REGS, "Invalid MaxwellDMA register");
 
@@ -44,7 +48,6 @@ void MaxwellDMA::Launch() {
 
     // TODO(Subv): Perform more research and implement all features of this engine.
     const LaunchDMA& launch = regs.launch_dma;
-    ASSERT(launch.remap_enable == 0);
     ASSERT(launch.semaphore_type == LaunchDMA::SemaphoreType::NONE);
     ASSERT(launch.interrupt_type == LaunchDMA::InterruptType::NONE);
     ASSERT(launch.data_transfer_type == LaunchDMA::DataTransferType::NON_PIPELINED);
@@ -77,11 +80,29 @@ void MaxwellDMA::CopyPitchToPitch() {
     // When `multi_line_enable` bit is disabled the copy is performed as if we were copying a 1D
     // buffer of length `line_length_in`.
     // Otherwise we copy a 2D image of dimensions (line_length_in, line_count).
+    auto& accelerate = rasterizer->AccessAccelerateDMA();
     if (!regs.launch_dma.multi_line_enable) {
-        memory_manager.CopyBlock(regs.offset_out, regs.offset_in, regs.line_length_in);
+        const bool is_buffer_clear = regs.launch_dma.remap_enable != 0 &&
+                                     regs.remap_const.dst_x == RemapConst::Swizzle::CONST_A;
+        // TODO: allow multisized components.
+        if (is_buffer_clear) {
+            ASSERT(regs.remap_const.component_size_minus_one == 3);
+            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));
+            return;
+        }
+        UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
+        if (!accelerate.BufferCopy(regs.offset_in, regs.offset_out, regs.line_length_in)) {
+            std::vector<u8> tmp_buffer(regs.line_length_in);
+            memory_manager.ReadBlockUnsafe(regs.offset_in, tmp_buffer.data(), regs.line_length_in);
+            memory_manager.WriteBlock(regs.offset_out, tmp_buffer.data(), regs.line_length_in);
+        }
         return;
     }
 
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
+
     // Perform a line-by-line copy.
     // We're going to take a subrect of size (line_length_in, line_count) from the source rectangle.
     // There is no need to manually flush/invalidate the regions because CopyBlock does that for us.
@@ -105,6 +126,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
     }
 
     // Deswizzle the input and copy it over.
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
     const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
     const Parameters& src_params = regs.src_params;
     const u32 width = src_params.width;
@@ -134,6 +156,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
 
 void MaxwellDMA::CopyPitchToBlockLinear() {
     UNIMPLEMENTED_IF_MSG(regs.dst_params.block_size.width != 0, "Block width is not one");
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
 
     const auto& dst_params = regs.dst_params;
     const u32 bytes_per_pixel = regs.pitch_in / regs.line_length_in;
@@ -156,13 +179,8 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
         write_buffer.resize(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);
-    }
+    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+    memory_manager.ReadBlock(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 c77f02a22..4ed0d0996 100644
--- a/src/video_core/engines/maxwell_dma.h
+++ b/src/video_core/engines/maxwell_dma.h
@@ -21,8 +21,18 @@ namespace Tegra {
 class MemoryManager;
 }
 
+namespace VideoCore {
+class RasterizerInterface;
+}
+
 namespace Tegra::Engines {
 
+class AccelerateDMAInterface {
+public:
+    /// Write the value to the register identified by method.
+    virtual bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) = 0;
+};
+
 /**
  * This engine is known as gk104_copy. Documentation can be found in:
  * https://github.com/NVIDIA/open-gpu-doc/blob/master/classes/dma-copy/clb0b5.h
@@ -187,6 +197,8 @@ public:
     };
     static_assert(sizeof(RemapConst) == 12);
 
+    void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
+
     explicit MaxwellDMA(Core::System& system_, MemoryManager& memory_manager_);
     ~MaxwellDMA() override;
 
@@ -213,6 +225,7 @@ private:
     Core::System& system;
 
     MemoryManager& memory_manager;
+    VideoCore::RasterizerInterface* rasterizer;
 
     std::vector<u8> read_buffer;
     std::vector<u8> write_buffer;
@@ -240,7 +253,9 @@ private:
                 u32 pitch_out;
                 u32 line_length_in;
                 u32 line_count;
-                u32 reserved06[0xb8];
+                u32 reserved06[0xb6];
+                u32 remap_consta_value;
+                u32 remap_constb_value;
                 RemapConst remap_const;
                 Parameters dst_params;
                 u32 reserved07[0x1];
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index f055b61e9..34dc6c596 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -8,6 +8,7 @@
 #include <queue>
 
 #include "common/common_types.h"
+#include "common/settings.h"
 #include "core/core.h"
 #include "video_core/delayed_destruction_ring.h"
 #include "video_core/gpu.h"
@@ -53,6 +54,12 @@ public:
         delayed_destruction_ring.Tick();
     }
 
+    // Unlike other fences, this one doesn't
+    void SignalOrdering() {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.AccumulateFlushes();
+    }
+
     void SignalSemaphore(GPUVAddr addr, u32 value) {
         TryReleasePendingFences();
         const bool should_flush = ShouldFlush();
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 35cc561be..ff024f530 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -50,6 +50,7 @@ void GPU::BindRenderer(std::unique_ptr<VideoCore::RendererBase> renderer_) {
     maxwell_3d->BindRasterizer(rasterizer);
     fermi_2d->BindRasterizer(rasterizer);
     kepler_compute->BindRasterizer(rasterizer);
+    maxwell_dma->BindRasterizer(rasterizer);
 }
 
 Engines::Maxwell3D& GPU::Maxwell3D() {
@@ -268,11 +269,13 @@ void GPU::CallPullerMethod(const MethodCall& method_call) {
     case BufferMethods::SemaphoreAddressHigh:
     case BufferMethods::SemaphoreAddressLow:
     case BufferMethods::SemaphoreSequence:
-    case BufferMethods::RefCnt:
     case BufferMethods::UnkCacheFlush:
     case BufferMethods::WrcacheFlush:
     case BufferMethods::FenceValue:
         break;
+    case BufferMethods::RefCnt:
+        rasterizer->SignalReference();
+        break;
     case BufferMethods::FenceAction:
         ProcessFenceActionMethod();
         break;
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 0cec4225b..58014c1c3 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -15,7 +15,10 @@
 
 namespace Tegra {
 class MemoryManager;
+namespace Engines {
+class AccelerateDMAInterface;
 }
+} // namespace Tegra
 
 namespace VideoCore {
 
@@ -63,6 +66,9 @@ public:
     /// Signal a GPU based syncpoint as a fence
     virtual void SignalSyncPoint(u32 value) = 0;
 
+    /// Signal a GPU based reference as point
+    virtual void SignalReference() = 0;
+
     /// Release all pending fences.
     virtual void ReleaseFences() = 0;
 
@@ -116,6 +122,8 @@ public:
         return false;
     }
 
+    [[nodiscard]] virtual Tegra::Engines::AccelerateDMAInterface& AccessAccelerateDMA() = 0;
+
     /// Attempt to use a faster method to display the framebuffer to screen
     [[nodiscard]] virtual bool AccelerateDisplay(const Tegra::FramebufferConfig& config,
                                                  VAddr framebuffer_addr, u32 pixel_stride) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 07ad0e205..82c84127a 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -171,7 +171,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache_runtime(device),
       buffer_cache(*this, maxwell3d, kepler_compute, gpu_memory, cpu_memory_, buffer_cache_runtime),
       shader_cache(*this, emu_window_, gpu, maxwell3d, kepler_compute, gpu_memory, device),
-      query_cache(*this, maxwell3d, gpu_memory),
+      query_cache(*this, maxwell3d, gpu_memory), accelerate_dma(buffer_cache),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache),
       async_shaders(emu_window_) {
     if (device.UseAsynchronousShaders()) {
@@ -634,6 +634,13 @@ void RasterizerOpenGL::SignalSyncPoint(u32 value) {
     fence_manager.SignalSyncPoint(value);
 }
 
+void RasterizerOpenGL::SignalReference() {
+    if (!gpu.IsAsync()) {
+        return;
+    }
+    fence_manager.SignalOrdering();
+}
+
 void RasterizerOpenGL::ReleaseFences() {
     if (!gpu.IsAsync()) {
         return;
@@ -650,6 +657,7 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
 
 void RasterizerOpenGL::WaitForIdle() {
     glMemoryBarrier(GL_ALL_BARRIER_BITS);
+    SignalReference();
 }
 
 void RasterizerOpenGL::FragmentBarrier() {
@@ -693,6 +701,10 @@ bool RasterizerOpenGL::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surf
     return true;
 }
 
+Tegra::Engines::AccelerateDMAInterface& RasterizerOpenGL::AccessAccelerateDMA() {
+    return accelerate_dma;
+}
+
 bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
                                          VAddr framebuffer_addr, u32 pixel_stride) {
     if (framebuffer_addr == 0) {
@@ -1388,4 +1400,11 @@ void RasterizerOpenGL::EndTransformFeedback() {
     glEndTransformFeedback();
 }
 
+AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_) : buffer_cache{buffer_cache_} {}
+
+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);
+}
+
 } // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 482efed7a..ccee9ba33 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -19,6 +19,7 @@
 #include "common/common_types.h"
 #include "video_core/engines/const_buffer_info.h"
 #include "video_core/engines/maxwell_3d.h"
+#include "video_core/engines/maxwell_dma.h"
 #include "video_core/rasterizer_accelerated.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_opengl/gl_buffer_cache.h"
@@ -58,6 +59,16 @@ struct BindlessSSBO {
 };
 static_assert(sizeof(BindlessSSBO) * CHAR_BIT == 128);
 
+class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface {
+public:
+    explicit AccelerateDMA(BufferCache& buffer_cache);
+
+    bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) override;
+
+private:
+    BufferCache& buffer_cache;
+};
+
 class RasterizerOpenGL : public VideoCore::RasterizerAccelerated {
 public:
     explicit RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_,
@@ -83,6 +94,7 @@ public:
     void ModifyGPUMemory(GPUVAddr addr, u64 size) override;
     void SignalSemaphore(GPUVAddr addr, u32 value) override;
     void SignalSyncPoint(u32 value) override;
+    void SignalReference() override;
     void ReleaseFences() override;
     void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
     void WaitForIdle() override;
@@ -93,6 +105,7 @@ public:
     bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
                                const Tegra::Engines::Fermi2D::Surface& dst,
                                const Tegra::Engines::Fermi2D::Config& copy_config) override;
+    Tegra::Engines::AccelerateDMAInterface& AccessAccelerateDMA() override;
     bool AccelerateDisplay(const Tegra::FramebufferConfig& config, VAddr framebuffer_addr,
                            u32 pixel_stride) override;
     void LoadDiskResources(u64 title_id, std::stop_token stop_loading,
@@ -233,6 +246,7 @@ private:
     BufferCache buffer_cache;
     ShaderCacheOpenGL shader_cache;
     QueryCache query_cache;
+    AccelerateDMA accelerate_dma;
     FenceManagerOpenGL fence_manager;
 
     VideoCommon::Shader::AsyncShaders async_shaders;
diff --git a/src/video_core/renderer_opengl/renderer_opengl.cpp b/src/video_core/renderer_opengl/renderer_opengl.cpp
index a718bff7a..c12929de6 100644
--- a/src/video_core/renderer_opengl/renderer_opengl.cpp
+++ b/src/video_core/renderer_opengl/renderer_opengl.cpp
@@ -229,9 +229,6 @@ void RendererOpenGL::LoadColorToActiveGLTexture(u8 color_r, u8 color_g, u8 color
 }
 
 void RendererOpenGL::InitOpenGLObjects() {
-    glClearColor(Settings::values.bg_red.GetValue(), Settings::values.bg_green.GetValue(),
-                 Settings::values.bg_blue.GetValue(), 0.0f);
-
     // Create shader programs
     OGLShader vertex_shader;
     vertex_shader.Create(HostShaders::OPENGL_PRESENT_VERT, GL_VERTEX_SHADER);
@@ -337,8 +334,9 @@ void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
 void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
     if (renderer_settings.set_background_color) {
         // Update background color before drawing
-        glClearColor(Settings::values.bg_red.GetValue(), Settings::values.bg_green.GetValue(),
-                     Settings::values.bg_blue.GetValue(), 0.0f);
+        glClearColor(Settings::values.bg_red.GetValue() / 255.0f,
+                     Settings::values.bg_green.GetValue() / 255.0f,
+                     Settings::values.bg_blue.GetValue() / 255.0f, 1.0f);
     }
 
     // Set projection matrix
diff --git a/src/video_core/renderer_vulkan/vk_blit_screen.cpp b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
index a1a32aabe..363134129 100644
--- a/src/video_core/renderer_vulkan/vk_blit_screen.cpp
+++ b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
@@ -225,8 +225,11 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
                       descriptor_set = descriptor_sets[image_index], buffer = *buffer,
                       size = swapchain.GetSize(), pipeline = *pipeline,
                       layout = *pipeline_layout](vk::CommandBuffer cmdbuf) {
+        const f32 bg_red = Settings::values.bg_red.GetValue() / 255.0f;
+        const f32 bg_green = Settings::values.bg_green.GetValue() / 255.0f;
+        const f32 bg_blue = Settings::values.bg_blue.GetValue() / 255.0f;
         const VkClearValue clear_color{
-            .color = {.float32 = {0.0f, 0.0f, 0.0f, 0.0f}},
+            .color = {.float32 = {bg_red, bg_green, bg_blue, 1.0f}},
         };
         const VkRenderPassBeginInfo renderpass_bi{
             .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.cpp b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
index 205cd3b05..4181d83ee 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
@@ -374,20 +374,20 @@ void ASTCDecoderPass::MakeDataBuffer() {
 
     scheduler.Record([src = staging_ref.buffer, offset = staging_ref.offset, dst = *data_buffer,
                       TOTAL_BUFFER_SIZE](vk::CommandBuffer cmdbuf) {
-        cmdbuf.CopyBuffer(src, dst,
-                          VkBufferCopy{
-                              .srcOffset = offset,
-                              .dstOffset = 0,
-                              .size = TOTAL_BUFFER_SIZE,
-                          });
-        cmdbuf.PipelineBarrier(
-            VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0,
-            VkMemoryBarrier{
-                .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = 0,
-                .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_SHADER_READ_BIT,
+        };
+        const VkBufferCopy copy{
+            .srcOffset = offset,
+            .dstOffset = 0,
+            .size = TOTAL_BUFFER_SIZE,
+        };
+        cmdbuf.CopyBuffer(src, dst, copy);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                               0, write_barrier);
     });
 }
 
@@ -411,7 +411,7 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
             const VkImageMemoryBarrier image_barrier{
                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
                 .pNext = nullptr,
-                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+                .srcAccessMask = is_initialized ? VK_ACCESS_SHADER_WRITE_BIT : VkAccessFlags{},
                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
                 .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
                 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
@@ -426,7 +426,8 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
                     .layerCount = VK_REMAINING_ARRAY_LAYERS,
                 },
             };
-            cmdbuf.PipelineBarrier(is_initialized ? VK_PIPELINE_STAGE_ALL_COMMANDS_BIT : 0,
+            cmdbuf.PipelineBarrier(is_initialized ? VK_PIPELINE_STAGE_ALL_COMMANDS_BIT
+                                                  : VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
                                    VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
             cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, vk_pipeline);
         });
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index bd4d649cc..a8ffbe6ba 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -251,7 +251,7 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache(*this, maxwell3d, kepler_compute, gpu_memory, cpu_memory_, buffer_cache_runtime),
       pipeline_cache(*this, gpu, maxwell3d, kepler_compute, gpu_memory, device, scheduler,
                      descriptor_pool, update_descriptor_queue),
-      query_cache{*this, maxwell3d, gpu_memory, device, scheduler},
+      query_cache{*this, maxwell3d, gpu_memory, device, scheduler}, accelerate_dma{buffer_cache},
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) {
     scheduler.SetQueryCache(query_cache);
@@ -357,11 +357,13 @@ void RasterizerVulkan::Clear() {
         .height = std::min(clear_rect.rect.extent.height, render_area.height),
     };
 
-    if (use_color) {
+    const u32 color_attachment = regs.clear_buffers.RT;
+    const auto attachment_aspect_mask = framebuffer->ImageRanges()[color_attachment].aspectMask;
+    const bool is_color_rt = (attachment_aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT) != 0;
+    if (use_color && is_color_rt) {
         VkClearValue clear_value;
         std::memcpy(clear_value.color.float32, regs.clear_color, sizeof(regs.clear_color));
 
-        const u32 color_attachment = regs.clear_buffers.RT;
         scheduler.Record([color_attachment, clear_value, clear_rect](vk::CommandBuffer cmdbuf) {
             const VkClearAttachment attachment{
                 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
@@ -580,6 +582,13 @@ void RasterizerVulkan::SignalSyncPoint(u32 value) {
     fence_manager.SignalSyncPoint(value);
 }
 
+void RasterizerVulkan::SignalReference() {
+    if (!gpu.IsAsync()) {
+        return;
+    }
+    fence_manager.SignalOrdering();
+}
+
 void RasterizerVulkan::ReleaseFences() {
     if (!gpu.IsAsync()) {
         return;
@@ -612,6 +621,7 @@ void RasterizerVulkan::WaitForIdle() {
         cmdbuf.SetEvent(event, flags);
         cmdbuf.WaitEvents(event, flags, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, {}, {}, {});
     });
+    SignalReference();
 }
 
 void RasterizerVulkan::FragmentBarrier() {
@@ -652,6 +662,10 @@ bool RasterizerVulkan::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surf
     return true;
 }
 
+Tegra::Engines::AccelerateDMAInterface& RasterizerVulkan::AccessAccelerateDMA() {
+    return accelerate_dma;
+}
+
 bool RasterizerVulkan::AccelerateDisplay(const Tegra::FramebufferConfig& config,
                                          VAddr framebuffer_addr, u32 pixel_stride) {
     if (!framebuffer_addr) {
@@ -690,6 +704,13 @@ void RasterizerVulkan::FlushWork() {
     draw_counter = 0;
 }
 
+AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_) : buffer_cache{buffer_cache_} {}
+
+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);
+}
+
 void RasterizerVulkan::SetupShaderDescriptors(
     const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders, bool is_indexed) {
     image_view_indices.clear();
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 41459c5c5..3a78de258 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -13,6 +13,7 @@
 #include <boost/container/static_vector.hpp>
 
 #include "common/common_types.h"
+#include "video_core/engines/maxwell_dma.h"
 #include "video_core/rasterizer_accelerated.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_vulkan/blit_image.h"
@@ -49,6 +50,16 @@ struct VKScreenInfo;
 
 class StateTracker;
 
+class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface {
+public:
+    explicit AccelerateDMA(BufferCache& buffer_cache);
+
+    bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
+
+private:
+    BufferCache& buffer_cache;
+};
+
 class RasterizerVulkan final : public VideoCore::RasterizerAccelerated {
 public:
     explicit RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_,
@@ -75,6 +86,7 @@ public:
     void ModifyGPUMemory(GPUVAddr addr, u64 size) override;
     void SignalSemaphore(GPUVAddr addr, u32 value) override;
     void SignalSyncPoint(u32 value) override;
+    void SignalReference() override;
     void ReleaseFences() override;
     void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
     void WaitForIdle() override;
@@ -85,6 +97,7 @@ public:
     bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
                                const Tegra::Engines::Fermi2D::Surface& dst,
                                const Tegra::Engines::Fermi2D::Config& copy_config) override;
+    Tegra::Engines::AccelerateDMAInterface& AccessAccelerateDMA() override;
     bool AccelerateDisplay(const Tegra::FramebufferConfig& config, VAddr framebuffer_addr,
                            u32 pixel_stride) override;
 
@@ -185,6 +198,7 @@ private:
     BufferCache buffer_cache;
     VKPipelineCache pipeline_cache;
     VKQueryCache query_cache;
+    AccelerateDMA accelerate_dma;
     VKFenceManager fence_manager;
 
     vk::Event wfi_event;
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index a2ab4d1ee..fd01c902c 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -608,7 +608,10 @@ void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst
     const VkImageAspectFlags aspect_mask = ImageAspectMask(src.format);
     const bool is_dst_msaa = dst.Samples() != VK_SAMPLE_COUNT_1_BIT;
     const bool is_src_msaa = src.Samples() != VK_SAMPLE_COUNT_1_BIT;
-    ASSERT(aspect_mask == ImageAspectMask(dst.format));
+    if (aspect_mask != ImageAspectMask(dst.format)) {
+        UNIMPLEMENTED_MSG("Incompatible blit from format {} to {}", src.format, dst.format);
+        return;
+    }
     if (aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT && !is_src_msaa && !is_dst_msaa) {
         blit_image_helper.BlitColor(dst_framebuffer, src, dst_region, src_region, filter,
                                     operation);
@@ -911,6 +914,7 @@ void Image::UploadMemory(const StagingBufferRef& map,
 
 void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferImageCopy> copies) {
     std::vector vk_copies = TransformBufferImageCopies(copies, map.offset, aspect_mask);
+    scheduler->RequestOutsideRenderPassOperationContext();
     scheduler->Record([buffer = map.buffer, image = *image, aspect_mask = aspect_mask,
                        vk_copies](vk::CommandBuffer cmdbuf) {
         const VkImageMemoryBarrier read_barrier{
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 01de2d498..85ce06d56 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -599,6 +599,12 @@ void TextureCache<P>::UpdateRenderTargets(bool is_clear) {
     using namespace VideoCommon::Dirty;
     auto& flags = maxwell3d.dirty.flags;
     if (!flags[Dirty::RenderTargets]) {
+        for (size_t index = 0; index < NUM_RT; ++index) {
+            ImageViewId& color_buffer_id = render_targets.color_buffer_ids[index];
+            PrepareImageView(color_buffer_id, true, is_clear && IsFullClear(color_buffer_id));
+        }
+        const ImageViewId depth_buffer_id = render_targets.depth_buffer_id;
+        PrepareImageView(depth_buffer_id, true, is_clear && IsFullClear(depth_buffer_id));
         return;
     }
     flags[Dirty::RenderTargets] = false;
diff --git a/src/video_core/texture_cache/types.h b/src/video_core/texture_cache/types.h
index 9fbdc1ac6..47a11cb2f 100644
--- a/src/video_core/texture_cache/types.h
+++ b/src/video_core/texture_cache/types.h
@@ -133,8 +133,8 @@ struct BufferImageCopy {
 };
 
 struct BufferCopy {
-    size_t src_offset;
-    size_t dst_offset;
+    u64 src_offset;
+    u64 dst_offset;
     size_t size;
 };
 
diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp
index c872517b8..59cf2f561 100644
--- a/src/video_core/texture_cache/util.cpp
+++ b/src/video_core/texture_cache/util.cpp
@@ -169,23 +169,6 @@ template <u32 GOB_EXTENT>
     return Common::DivCeil(AdjustMipSize(size, level), block_size);
 }
 
-[[nodiscard]] constexpr std::pair<int, int> Samples(int num_samples) {
-    switch (num_samples) {
-    case 1:
-        return {1, 1};
-    case 2:
-        return {2, 1};
-    case 4:
-        return {2, 2};
-    case 8:
-        return {4, 2};
-    case 16:
-        return {4, 4};
-    }
-    UNREACHABLE_MSG("Invalid number of samples={}", num_samples);
-    return {1, 1};
-}
-
 [[nodiscard]] constexpr Extent2D DefaultBlockSize(PixelFormat format) {
     return {DefaultBlockWidth(format), DefaultBlockHeight(format)};
 }
@@ -283,14 +266,13 @@ template <u32 GOB_EXTENT>
 }
 
 [[nodiscard]] constexpr LevelInfo MakeLevelInfo(PixelFormat format, Extent3D size, Extent3D block,
-                                                u32 num_samples, u32 tile_width_spacing) {
-    const auto [samples_x, samples_y] = Samples(num_samples);
+                                                u32 tile_width_spacing) {
     const u32 bytes_per_block = BytesPerBlock(format);
     return {
         .size =
             {
-                .width = size.width * samples_x,
-                .height = size.height * samples_y,
+                .width = size.width,
+                .height = size.height,
                 .depth = size.depth,
             },
         .block = block,
@@ -301,14 +283,12 @@ template <u32 GOB_EXTENT>
 }
 
 [[nodiscard]] constexpr LevelInfo MakeLevelInfo(const ImageInfo& info) {
-    return MakeLevelInfo(info.format, info.size, info.block, info.num_samples,
-                         info.tile_width_spacing);
+    return MakeLevelInfo(info.format, info.size, info.block, info.tile_width_spacing);
 }
 
 [[nodiscard]] constexpr u32 CalculateLevelOffset(PixelFormat format, Extent3D size, Extent3D block,
-                                                 u32 num_samples, u32 tile_width_spacing,
-                                                 u32 level) {
-    const LevelInfo info = MakeLevelInfo(format, size, block, num_samples, tile_width_spacing);
+                                                 u32 tile_width_spacing, u32 level) {
+    const LevelInfo info = MakeLevelInfo(format, size, block, tile_width_spacing);
     u32 offset = 0;
     for (u32 current_level = 0; current_level < level; ++current_level) {
         offset += CalculateLevelSize(info, current_level);
@@ -645,8 +625,8 @@ u32 CalculateLayerStride(const ImageInfo& info) noexcept {
 
 u32 CalculateLayerSize(const ImageInfo& info) noexcept {
     ASSERT(info.type != ImageType::Linear);
-    return CalculateLevelOffset(info.format, info.size, info.block, info.num_samples,
-                                info.tile_width_spacing, info.resources.levels);
+    return CalculateLevelOffset(info.format, info.size, info.block, info.tile_width_spacing,
+                                info.resources.levels);
 }
 
 LevelArray CalculateMipLevelOffsets(const ImageInfo& info) noexcept {
@@ -1195,37 +1175,37 @@ static_assert(CalculateLevelSize(LevelInfo{{1920, 1080, 1}, {0, 2, 0}, {1, 1}, 2
               0x7f8000);
 static_assert(CalculateLevelSize(LevelInfo{{32, 32, 1}, {0, 0, 4}, {1, 1}, 4, 0}, 0) == 0x4000);
 
-static_assert(CalculateLevelOffset(PixelFormat::R8_SINT, {1920, 1080, 1}, {0, 2, 0}, 1, 0, 7) ==
+static_assert(CalculateLevelOffset(PixelFormat::R8_SINT, {1920, 1080, 1}, {0, 2, 0}, 0, 7) ==
               0x2afc00);
-static_assert(CalculateLevelOffset(PixelFormat::ASTC_2D_12X12_UNORM, {8192, 4096, 1}, {0, 2, 0}, 1,
-                                   0, 12) == 0x50d200);
-
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   0) == 0);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   1) == 0x400000);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   2) == 0x500000);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   3) == 0x540000);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   4) == 0x550000);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   5) == 0x554000);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   6) == 0x555000);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   7) == 0x555400);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   8) == 0x555600);
-static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 1, 0,
-                                   9) == 0x555800);
+static_assert(CalculateLevelOffset(PixelFormat::ASTC_2D_12X12_UNORM, {8192, 4096, 1}, {0, 2, 0}, 0,
+                                   12) == 0x50d200);
+
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 0) ==
+              0);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 1) ==
+              0x400000);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 2) ==
+              0x500000);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 3) ==
+              0x540000);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 4) ==
+              0x550000);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 5) ==
+              0x554000);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 6) ==
+              0x555000);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 7) ==
+              0x555400);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 8) ==
+              0x555600);
+static_assert(CalculateLevelOffset(PixelFormat::A8B8G8R8_UNORM, {1024, 1024, 1}, {0, 4, 0}, 0, 9) ==
+              0x555800);
 
 constexpr u32 ValidateLayerSize(PixelFormat format, u32 width, u32 height, u32 block_height,
                                 u32 tile_width_spacing, u32 level) {
     const Extent3D size{width, height, 1};
     const Extent3D block{0, block_height, 0};
-    const u32 offset = CalculateLevelOffset(format, size, block, 1, tile_width_spacing, level);
+    const u32 offset = CalculateLevelOffset(format, size, block, tile_width_spacing, level);
     return AlignLayerSize(offset, size, block, DefaultBlockHeight(format), tile_width_spacing);
 }
 
diff --git a/src/web_service/web_backend.cpp b/src/web_service/web_backend.cpp
index e04f7dfc6..b1e02c57a 100644
--- a/src/web_service/web_backend.cpp
+++ b/src/web_service/web_backend.cpp
@@ -8,7 +8,17 @@
 #include <string>
 
 #include <fmt/format.h>
+
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#ifndef __clang__
+#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
+#endif
+#endif
 #include <httplib.h>
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
 
 #include "common/logging/log.h"
 #include "web_service/web_backend.h"
diff --git a/src/yuzu/CMakeLists.txt b/src/yuzu/CMakeLists.txt
index 634fe66a5..cb4bdcc7e 100644
--- a/src/yuzu/CMakeLists.txt
+++ b/src/yuzu/CMakeLists.txt
@@ -15,18 +15,19 @@ add_executable(yuzu
     about_dialog.cpp
     about_dialog.h
     aboutdialog.ui
-    applets/controller.cpp
-    applets/controller.h
-    applets/controller.ui
-    applets/error.cpp
-    applets/error.h
-    applets/profile_select.cpp
-    applets/profile_select.h
-    applets/software_keyboard.cpp
-    applets/software_keyboard.h
-    applets/software_keyboard.ui
-    applets/web_browser.cpp
-    applets/web_browser.h
+    applets/qt_controller.cpp
+    applets/qt_controller.h
+    applets/qt_controller.ui
+    applets/qt_error.cpp
+    applets/qt_error.h
+    applets/qt_profile_select.cpp
+    applets/qt_profile_select.h
+    applets/qt_software_keyboard.cpp
+    applets/qt_software_keyboard.h
+    applets/qt_software_keyboard.ui
+    applets/qt_web_browser.cpp
+    applets/qt_web_browser.h
+    applets/qt_web_browser_scripts.h
     bootmanager.cpp
     bootmanager.h
     compatdb.ui
@@ -52,6 +53,9 @@ add_executable(yuzu
     configuration/configure_debug_controller.cpp
     configuration/configure_debug_controller.h
     configuration/configure_debug_controller.ui
+    configuration/configure_debug_tab.cpp
+    configuration/configure_debug_tab.h
+    configuration/configure_debug_tab.ui
     configuration/configure_dialog.cpp
     configuration/configure_dialog.h
     configuration/configure_filesystem.cpp
diff --git a/src/yuzu/applets/controller.cpp b/src/yuzu/applets/qt_controller.cpp
index 836d90fda..97106d2cc 100644
--- a/src/yuzu/applets/controller.cpp
+++ b/src/yuzu/applets/qt_controller.cpp
@@ -12,8 +12,8 @@
 #include "core/hle/service/hid/controllers/npad.h"
 #include "core/hle/service/hid/hid.h"
 #include "core/hle/service/sm/sm.h"
-#include "ui_controller.h"
-#include "yuzu/applets/controller.h"
+#include "ui_qt_controller.h"
+#include "yuzu/applets/qt_controller.h"
 #include "yuzu/configuration/configure_input.h"
 #include "yuzu/configuration/configure_input_profile_dialog.h"
 #include "yuzu/configuration/configure_motion_touch.h"
diff --git a/src/yuzu/applets/controller.h b/src/yuzu/applets/qt_controller.h
index 9b57aea1a..9b57aea1a 100644
--- a/src/yuzu/applets/controller.h
+++ b/src/yuzu/applets/qt_controller.h
diff --git a/src/yuzu/applets/controller.ui b/src/yuzu/applets/qt_controller.ui
index c8cb6bcf3..c8cb6bcf3 100644
--- a/src/yuzu/applets/controller.ui
+++ b/src/yuzu/applets/qt_controller.ui
diff --git a/src/yuzu/applets/error.cpp b/src/yuzu/applets/qt_error.cpp
index 085688cd4..45cf64603 100644
--- a/src/yuzu/applets/error.cpp
+++ b/src/yuzu/applets/qt_error.cpp
@@ -4,7 +4,7 @@
 
 #include <QDateTime>
 #include "core/hle/lock.h"
-#include "yuzu/applets/error.h"
+#include "yuzu/applets/qt_error.h"
 #include "yuzu/main.h"
 
 QtErrorDisplay::QtErrorDisplay(GMainWindow& parent) {
diff --git a/src/yuzu/applets/error.h b/src/yuzu/applets/qt_error.h
index 8bd895a32..8bd895a32 100644
--- a/src/yuzu/applets/error.h
+++ b/src/yuzu/applets/qt_error.h
diff --git a/src/yuzu/applets/profile_select.cpp b/src/yuzu/applets/qt_profile_select.cpp
index 62fd1141c..a56638e21 100644
--- a/src/yuzu/applets/profile_select.cpp
+++ b/src/yuzu/applets/qt_profile_select.cpp
@@ -14,7 +14,7 @@
 #include "common/string_util.h"
 #include "core/constants.h"
 #include "core/hle/lock.h"
-#include "yuzu/applets/profile_select.h"
+#include "yuzu/applets/qt_profile_select.h"
 #include "yuzu/main.h"
 
 namespace {
diff --git a/src/yuzu/applets/profile_select.h b/src/yuzu/applets/qt_profile_select.h
index 4e9037488..4e9037488 100644
--- a/src/yuzu/applets/profile_select.h
+++ b/src/yuzu/applets/qt_profile_select.h
diff --git a/src/yuzu/applets/software_keyboard.cpp b/src/yuzu/applets/qt_software_keyboard.cpp
index aa453a79f..848801cec 100644
--- a/src/yuzu/applets/software_keyboard.cpp
+++ b/src/yuzu/applets/qt_software_keyboard.cpp
@@ -11,8 +11,8 @@
 #include "common/string_util.h"
 #include "core/core.h"
 #include "core/frontend/input_interpreter.h"
-#include "ui_software_keyboard.h"
-#include "yuzu/applets/software_keyboard.h"
+#include "ui_qt_software_keyboard.h"
+#include "yuzu/applets/qt_software_keyboard.h"
 #include "yuzu/main.h"
 #include "yuzu/util/overlay_dialog.h"
 
diff --git a/src/yuzu/applets/software_keyboard.h b/src/yuzu/applets/qt_software_keyboard.h
index 1a03c098c..1a03c098c 100644
--- a/src/yuzu/applets/software_keyboard.h
+++ b/src/yuzu/applets/qt_software_keyboard.h
diff --git a/src/yuzu/applets/software_keyboard.ui b/src/yuzu/applets/qt_software_keyboard.ui
index b0a1fcde9..b0a1fcde9 100644
--- a/src/yuzu/applets/software_keyboard.ui
+++ b/src/yuzu/applets/qt_software_keyboard.ui
diff --git a/src/yuzu/applets/web_browser.cpp b/src/yuzu/applets/qt_web_browser.cpp
index 34d3feb55..b112dd7b0 100644
--- a/src/yuzu/applets/web_browser.cpp
+++ b/src/yuzu/applets/qt_web_browser.cpp
@@ -17,8 +17,8 @@
 #include "core/frontend/input_interpreter.h"
 #include "input_common/keyboard.h"
 #include "input_common/main.h"
-#include "yuzu/applets/web_browser.h"
-#include "yuzu/applets/web_browser_scripts.h"
+#include "yuzu/applets/qt_web_browser.h"
+#include "yuzu/applets/qt_web_browser_scripts.h"
 #include "yuzu/main.h"
 #include "yuzu/util/url_request_interceptor.h"
 
diff --git a/src/yuzu/applets/web_browser.h b/src/yuzu/applets/qt_web_browser.h
index 7ad07409f..7ad07409f 100644
--- a/src/yuzu/applets/web_browser.h
+++ b/src/yuzu/applets/qt_web_browser.h
diff --git a/src/yuzu/applets/web_browser_scripts.h b/src/yuzu/applets/qt_web_browser_scripts.h
index 992837a85..992837a85 100644
--- a/src/yuzu/applets/web_browser_scripts.h
+++ b/src/yuzu/applets/qt_web_browser_scripts.h
diff --git a/src/yuzu/bootmanager.cpp b/src/yuzu/bootmanager.cpp
index 7524e3c40..bfae73b60 100644
--- a/src/yuzu/bootmanager.cpp
+++ b/src/yuzu/bootmanager.cpp
@@ -411,8 +411,9 @@ void GRenderWindow::mousePressEvent(QMouseEvent* event) {
     if (event->source() == Qt::MouseEventSynthesizedBySystem) {
         return;
     }
-
-    auto pos = event->pos();
+    // Qt sometimes returns the parent coordinates. To avoid this we read the global mouse
+    // coordinates and map them to the current render area
+    const auto pos = mapFromGlobal(QCursor::pos());
     const auto [x, y] = ScaleTouch(pos);
     const auto button = QtButtonToMouseButton(event->button());
     input_subsystem->GetMouse()->PressButton(x, y, button);
@@ -429,7 +430,9 @@ void GRenderWindow::mouseMoveEvent(QMouseEvent* event) {
     if (event->source() == Qt::MouseEventSynthesizedBySystem) {
         return;
     }
-    auto pos = event->pos();
+    // Qt sometimes returns the parent coordinates. To avoid this we read the global mouse
+    // coordinates and map them to the current render area
+    const auto pos = mapFromGlobal(QCursor::pos());
     const auto [x, y] = ScaleTouch(pos);
     const int center_x = width() / 2;
     const int center_y = height() / 2;
diff --git a/src/yuzu/configuration/config.cpp b/src/yuzu/configuration/config.cpp
index 1a0f75373..a5e032959 100644
--- a/src/yuzu/configuration/config.cpp
+++ b/src/yuzu/configuration/config.cpp
@@ -311,16 +311,6 @@ void Config::WriteBasicSetting(const Settings::BasicSetting<std::string>& settin
     qt_config->setValue(name, QString::fromStdString(value));
 }
 
-// Explicit float definition: use a double as Qt doesn't write legible floats to config files
-template <>
-void Config::WriteBasicSetting(const Settings::BasicSetting<float>& setting) {
-    const QString name = QString::fromStdString(setting.GetLabel());
-    const double value = setting.GetValue();
-    qt_config->setValue(name + QStringLiteral("/default"),
-                        setting.GetValue() == setting.GetDefault());
-    qt_config->setValue(name, value);
-}
-
 template <typename Type>
 void Config::WriteBasicSetting(const Settings::BasicSetting<Type>& setting) {
     const QString name = QString::fromStdString(setting.GetLabel());
@@ -329,21 +319,6 @@ void Config::WriteBasicSetting(const Settings::BasicSetting<Type>& setting) {
     qt_config->setValue(name, value);
 }
 
-// Explicit float definition: use a double as Qt doesn't write legible floats to config files
-template <>
-void Config::WriteGlobalSetting(const Settings::Setting<float>& setting) {
-    const QString name = QString::fromStdString(setting.GetLabel());
-    const double value = setting.GetValue(global);
-    if (!global) {
-        qt_config->setValue(name + QStringLiteral("/use_global"), setting.UsingGlobal());
-    }
-    if (global || !setting.UsingGlobal()) {
-        qt_config->setValue(name + QStringLiteral("/default"),
-                            setting.GetValue(global) == setting.GetDefault());
-        qt_config->setValue(name, value);
-    }
-}
-
 template <typename Type>
 void Config::WriteGlobalSetting(const Settings::Setting<Type>& setting) {
     const QString name = QString::fromStdString(setting.GetLabel());
@@ -793,7 +768,13 @@ void Config::ReadPathValues() {
 void Config::ReadCpuValues() {
     qt_config->beginGroup(QStringLiteral("Cpu"));
 
-    ReadGlobalSetting(Settings::values.cpu_accuracy);
+    ReadBasicSetting(Settings::values.cpu_accuracy_first_time);
+    if (Settings::values.cpu_accuracy_first_time) {
+        Settings::values.cpu_accuracy.SetValue(Settings::values.cpu_accuracy.GetDefault());
+        Settings::values.cpu_accuracy_first_time.SetValue(false);
+    } else {
+        ReadGlobalSetting(Settings::values.cpu_accuracy);
+    }
 
     ReadGlobalSetting(Settings::values.cpuopt_unsafe_unfuse_fma);
     ReadGlobalSetting(Settings::values.cpuopt_unsafe_reduce_fp_error);
@@ -802,6 +783,7 @@ void Config::ReadCpuValues() {
     ReadGlobalSetting(Settings::values.cpuopt_unsafe_fastmem_check);
 
     if (global) {
+        ReadBasicSetting(Settings::values.cpu_debug_mode);
         ReadBasicSetting(Settings::values.cpuopt_page_tables);
         ReadBasicSetting(Settings::values.cpuopt_block_linking);
         ReadBasicSetting(Settings::values.cpuopt_return_stack_buffer);
@@ -820,7 +802,6 @@ void Config::ReadRendererValues() {
     qt_config->beginGroup(QStringLiteral("Renderer"));
 
     ReadGlobalSetting(Settings::values.renderer_backend);
-    ReadBasicSetting(Settings::values.renderer_debug);
     ReadGlobalSetting(Settings::values.vulkan_device);
     ReadGlobalSetting(Settings::values.fullscreen_mode);
     ReadGlobalSetting(Settings::values.aspect_ratio);
@@ -833,7 +814,6 @@ void Config::ReadRendererValues() {
     ReadGlobalSetting(Settings::values.use_nvdec_emulation);
     ReadGlobalSetting(Settings::values.accelerate_astc);
     ReadGlobalSetting(Settings::values.use_vsync);
-    ReadGlobalSetting(Settings::values.disable_fps_limit);
     ReadGlobalSetting(Settings::values.use_assembly_shaders);
     ReadGlobalSetting(Settings::values.use_asynchronous_shaders);
     ReadGlobalSetting(Settings::values.use_fast_gpu_time);
@@ -842,6 +822,10 @@ void Config::ReadRendererValues() {
     ReadGlobalSetting(Settings::values.bg_green);
     ReadGlobalSetting(Settings::values.bg_blue);
 
+    if (global) {
+        ReadBasicSetting(Settings::values.renderer_debug);
+    }
+
     qt_config->endGroup();
 }
 
@@ -1309,6 +1293,7 @@ void Config::SavePathValues() {
 void Config::SaveCpuValues() {
     qt_config->beginGroup(QStringLiteral("Cpu"));
 
+    WriteBasicSetting(Settings::values.cpu_accuracy_first_time);
     WriteSetting(QStringLiteral("cpu_accuracy"),
                  static_cast<u32>(Settings::values.cpu_accuracy.GetValue(global)),
                  static_cast<u32>(Settings::values.cpu_accuracy.GetDefault()),
@@ -1321,6 +1306,7 @@ void Config::SaveCpuValues() {
     WriteGlobalSetting(Settings::values.cpuopt_unsafe_fastmem_check);
 
     if (global) {
+        WriteBasicSetting(Settings::values.cpu_debug_mode);
         WriteBasicSetting(Settings::values.cpuopt_page_tables);
         WriteBasicSetting(Settings::values.cpuopt_block_linking);
         WriteBasicSetting(Settings::values.cpuopt_return_stack_buffer);
@@ -1342,7 +1328,6 @@ void Config::SaveRendererValues() {
                  static_cast<u32>(Settings::values.renderer_backend.GetValue(global)),
                  static_cast<u32>(Settings::values.renderer_backend.GetDefault()),
                  Settings::values.renderer_backend.UsingGlobal());
-    WriteBasicSetting(Settings::values.renderer_debug);
     WriteGlobalSetting(Settings::values.vulkan_device);
     WriteGlobalSetting(Settings::values.fullscreen_mode);
     WriteGlobalSetting(Settings::values.aspect_ratio);
@@ -1358,7 +1343,6 @@ void Config::SaveRendererValues() {
     WriteGlobalSetting(Settings::values.use_nvdec_emulation);
     WriteGlobalSetting(Settings::values.accelerate_astc);
     WriteGlobalSetting(Settings::values.use_vsync);
-    WriteGlobalSetting(Settings::values.disable_fps_limit);
     WriteGlobalSetting(Settings::values.use_assembly_shaders);
     WriteGlobalSetting(Settings::values.use_asynchronous_shaders);
     WriteGlobalSetting(Settings::values.use_fast_gpu_time);
@@ -1367,6 +1351,10 @@ void Config::SaveRendererValues() {
     WriteGlobalSetting(Settings::values.bg_green);
     WriteGlobalSetting(Settings::values.bg_blue);
 
+    if (global) {
+        WriteBasicSetting(Settings::values.renderer_debug);
+    }
+
     qt_config->endGroup();
 }
 
diff --git a/src/yuzu/configuration/configure.ui b/src/yuzu/configuration/configure.ui
index f92c3aff3..fca9aed5f 100644
--- a/src/yuzu/configuration/configure.ui
+++ b/src/yuzu/configuration/configure.ui
@@ -41,7 +41,7 @@
      <item>
       <widget class="QTabWidget" name="tabWidget">
        <property name="currentIndex">
-        <number>0</number>
+        <number>11</number>
        </property>
        <widget class="ConfigureGeneral" name="generalTab">
         <property name="accessibleName">
@@ -107,14 +107,6 @@
          <string>CPU</string>
         </attribute>
        </widget>
-       <widget class="ConfigureCpuDebug" name="cpuDebugTab">
-        <property name="accessibleName">
-         <string>Debug</string>
-        </property>
-        <attribute name="title">
-         <string>Debug</string>
-        </attribute>
-       </widget>
        <widget class="ConfigureGraphics" name="graphicsTab">
         <property name="accessibleName">
          <string>Graphics</string>
@@ -139,7 +131,7 @@
          <string>Audio</string>
         </attribute>
        </widget>
-       <widget class="ConfigureDebug" name="debugTab">
+       <widget class="ConfigureDebugTab" name="debugTab">
         <property name="accessibleName">
          <string>Debug</string>
         </property>
@@ -208,24 +200,12 @@
    <container>1</container>
   </customwidget>
   <customwidget>
-   <class>ConfigureDebug</class>
-   <extends>QWidget</extends>
-   <header>configuration/configure_debug.h</header>
-   <container>1</container>
-  </customwidget>
-  <customwidget>
    <class>ConfigureCpu</class>
    <extends>QWidget</extends>
    <header>configuration/configure_cpu.h</header>
    <container>1</container>
   </customwidget>
   <customwidget>
-   <class>ConfigureCpuDebug</class>
-   <extends>QWidget</extends>
-   <header>configuration/configure_cpu_debug.h</header>
-   <container>1</container>
-  </customwidget>
-  <customwidget>
    <class>ConfigureGraphics</class>
    <extends>QWidget</extends>
    <header>configuration/configure_graphics.h</header>
@@ -267,6 +247,12 @@
    <header>configuration/configure_service.h</header>
    <container>1</container>
   </customwidget>
+  <customwidget>
+   <class>ConfigureDebugTab</class>
+   <extends>QWidget</extends>
+   <header>configuration/configure_debug_tab.h</header>
+   <container>1</container>
+  </customwidget>
  </customwidgets>
  <resources/>
  <connections>
@@ -275,12 +261,32 @@
    <signal>accepted()</signal>
    <receiver>ConfigureDialog</receiver>
    <slot>accept()</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>20</x>
+     <y>20</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>20</x>
+     <y>20</y>
+    </hint>
+   </hints>
   </connection>
   <connection>
    <sender>buttonBox</sender>
    <signal>rejected()</signal>
    <receiver>ConfigureDialog</receiver>
    <slot>reject()</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>20</x>
+     <y>20</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>20</x>
+     <y>20</y>
+    </hint>
+   </hints>
   </connection>
  </connections>
 </ui>
diff --git a/src/yuzu/configuration/configure_audio.cpp b/src/yuzu/configuration/configure_audio.cpp
index 5aba1a3b2..1d84bf4ed 100644
--- a/src/yuzu/configuration/configure_audio.cpp
+++ b/src/yuzu/configuration/configure_audio.cpp
@@ -47,7 +47,8 @@ void ConfigureAudio::SetConfiguration() {
 
     SetAudioDeviceFromDeviceID();
 
-    ui->volume_slider->setValue(Settings::values.volume.GetValue() * ui->volume_slider->maximum());
+    const auto volume_value = static_cast<int>(Settings::values.volume.GetValue());
+    ui->volume_slider->setValue(volume_value);
 
     ui->toggle_audio_stretching->setChecked(Settings::values.enable_audio_stretching.GetValue());
 
@@ -112,18 +113,16 @@ void ConfigureAudio::ApplyConfiguration() {
 
         // Guard if during game and set to game-specific value
         if (Settings::values.volume.UsingGlobal()) {
-            Settings::values.volume.SetValue(
-                static_cast<float>(ui->volume_slider->sliderPosition()) /
-                ui->volume_slider->maximum());
+            const auto volume = static_cast<u8>(ui->volume_slider->value());
+            Settings::values.volume.SetValue(volume);
         }
     } else {
         if (ui->volume_combo_box->currentIndex() == 0) {
             Settings::values.volume.SetGlobal(true);
         } else {
             Settings::values.volume.SetGlobal(false);
-            Settings::values.volume.SetValue(
-                static_cast<float>(ui->volume_slider->sliderPosition()) /
-                ui->volume_slider->maximum());
+            const auto volume = static_cast<u8>(ui->volume_slider->value());
+            Settings::values.volume.SetValue(volume);
         }
     }
 }
diff --git a/src/yuzu/configuration/configure_cpu.cpp b/src/yuzu/configuration/configure_cpu.cpp
index 13db2ba98..8d7171487 100644
--- a/src/yuzu/configuration/configure_cpu.cpp
+++ b/src/yuzu/configuration/configure_cpu.cpp
@@ -20,8 +20,6 @@ ConfigureCpu::ConfigureCpu(QWidget* parent) : QWidget(parent), ui(new Ui::Config
 
     SetConfiguration();
 
-    connect(ui->accuracy, qOverload<int>(&QComboBox::activated), this,
-            &ConfigureCpu::AccuracyUpdated);
     connect(ui->accuracy, qOverload<int>(&QComboBox::currentIndexChanged), this,
             &ConfigureCpu::UpdateGroup);
 }
@@ -58,20 +56,6 @@ void ConfigureCpu::SetConfiguration() {
     UpdateGroup(ui->accuracy->currentIndex());
 }
 
-void ConfigureCpu::AccuracyUpdated(int index) {
-    if (Settings::IsConfiguringGlobal() &&
-        static_cast<Settings::CPUAccuracy>(index) == Settings::CPUAccuracy::DebugMode) {
-        const auto result = QMessageBox::warning(this, tr("Setting CPU to Debug Mode"),
-                                                 tr("CPU Debug Mode is only intended for developer "
-                                                    "use. Are you sure you want to enable this?"),
-                                                 QMessageBox::Yes | QMessageBox::No);
-        if (result == QMessageBox::No) {
-            ui->accuracy->setCurrentIndex(static_cast<int>(Settings::CPUAccuracy::Accurate));
-            UpdateGroup(static_cast<int>(Settings::CPUAccuracy::Accurate));
-        }
-    }
-}
-
 void ConfigureCpu::UpdateGroup(int index) {
     if (!Settings::IsConfiguringGlobal()) {
         index -= ConfigurationShared::USE_GLOBAL_OFFSET;
@@ -134,8 +118,6 @@ void ConfigureCpu::SetupPerGameUI() {
     ConfigurationShared::SetColoredComboBox(
         ui->accuracy, ui->widget_accuracy,
         static_cast<u32>(Settings::values.cpu_accuracy.GetValue(true)));
-    ui->accuracy->removeItem(static_cast<u32>(Settings::CPUAccuracy::DebugMode) +
-                             ConfigurationShared::USE_GLOBAL_OFFSET);
 
     ConfigurationShared::SetColoredTristate(ui->cpuopt_unsafe_unfuse_fma,
                                             Settings::values.cpuopt_unsafe_unfuse_fma,
diff --git a/src/yuzu/configuration/configure_cpu.h b/src/yuzu/configuration/configure_cpu.h
index b2b5f1671..154931482 100644
--- a/src/yuzu/configuration/configure_cpu.h
+++ b/src/yuzu/configuration/configure_cpu.h
@@ -29,7 +29,6 @@ private:
     void changeEvent(QEvent* event) override;
     void RetranslateUI();
 
-    void AccuracyUpdated(int index);
     void UpdateGroup(int index);
 
     void SetConfiguration();
diff --git a/src/yuzu/configuration/configure_cpu.ui b/src/yuzu/configuration/configure_cpu.ui
index 0e296d4e5..5b9457faf 100644
--- a/src/yuzu/configuration/configure_cpu.ui
+++ b/src/yuzu/configuration/configure_cpu.ui
@@ -6,8 +6,8 @@
    <rect>
     <x>0</x>
     <y>0</y>
-    <width>400</width>
-    <height>321</height>
+    <width>448</width>
+    <height>433</height>
    </rect>
   </property>
   <property name="windowTitle">
@@ -17,7 +17,7 @@
    <item>
     <layout class="QVBoxLayout">
      <item>
-      <widget class="QGroupBox">
+      <widget class="QGroupBox" name="groupBox">
        <property name="title">
         <string>General</string>
        </property>
@@ -36,17 +36,17 @@
             <widget class="QComboBox" name="accuracy">
              <item>
               <property name="text">
-               <string>Accurate</string>
+               <string>Auto</string>
               </property>
              </item>
              <item>
               <property name="text">
-               <string>Unsafe</string>
+               <string>Accurate</string>
               </property>
              </item>
              <item>
               <property name="text">
-               <string>Enable Debug Mode</string>
+               <string>Unsafe</string>
               </property>
              </item>
             </widget>
@@ -57,7 +57,7 @@
         <item>
          <widget class="QLabel" name="label_recommended_accuracy">
           <property name="text">
-           <string>We recommend setting accuracy to &quot;Accurate&quot;.</string>
+           <string>We recommend setting accuracy to &quot;Auto&quot;.</string>
           </property>
           <property name="wordWrap">
            <bool>false</bool>
diff --git a/src/yuzu/configuration/configure_cpu_debug.h b/src/yuzu/configuration/configure_cpu_debug.h
index 10de55099..1b0d8050c 100644
--- a/src/yuzu/configuration/configure_cpu_debug.h
+++ b/src/yuzu/configuration/configure_cpu_debug.h
@@ -6,7 +6,6 @@
 
 #include <memory>
 #include <QWidget>
-#include "common/settings.h"
 
 namespace Ui {
 class ConfigureCpuDebug;
diff --git a/src/yuzu/configuration/configure_cpu_debug.ui b/src/yuzu/configuration/configure_cpu_debug.ui
index c43f89a5a..abf469b55 100644
--- a/src/yuzu/configuration/configure_cpu_debug.ui
+++ b/src/yuzu/configuration/configure_cpu_debug.ui
@@ -6,8 +6,8 @@
    <rect>
     <x>0</x>
     <y>0</y>
-    <width>400</width>
-    <height>321</height>
+    <width>592</width>
+    <height>503</height>
    </rect>
   </property>
   <property name="windowTitle">
@@ -17,140 +17,132 @@
    <item>
     <layout class="QVBoxLayout">
      <item>
-      <widget class="QGroupBox">
+      <widget class="QGroupBox" name="groupBox">
        <property name="title">
         <string>Toggle CPU Optimizations</string>
        </property>
        <layout class="QVBoxLayout">
         <item>
-         <widget class="QLabel">
-          <property name="wordWrap">
-            <bool>1</bool>
-          </property>
+         <widget class="QLabel" name="label">
           <property name="text">
-           <string>
-            &lt;div&gt;
-            &lt;b&gt;For debugging only.&lt;/b&gt;
-            &lt;br&gt;
-            If you're not sure what these do, keep all of these enabled.
-            &lt;br&gt;
-            These settings, when disabled, only take effect when CPU Accuracy is "Debug Mode".
-            &lt;/div&gt;
-           </string>
+           <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;&lt;span style=&quot; font-weight:600;&quot;&gt;For debugging only.&lt;/span&gt;&lt;br/&gt;If you're not sure what these do, keep all of these enabled. &lt;br/&gt;These settings, when disabled, only take effect when CPU Debugging is enabled. &lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+          </property>
+          <property name="wordWrap">
+           <bool>false</bool>
           </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_page_tables">
-          <property name="text">
-           <string>Enable inline page tables</string>
-          </property>
           <property name="toolTip">
            <string>
-            &lt;div style="white-space: nowrap"&gt;This optimization speeds up memory accesses by the guest program.&lt;/div&gt;
-            &lt;div style="white-space: nowrap"&gt;Enabling it inlines accesses to PageTable::pointers into emitted code.&lt;/div&gt;
-            &lt;div style="white-space: nowrap"&gt;Disabling this forces all memory accesses to go through the Memory::Read/Memory::Write functions.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;This optimization speeds up memory accesses by the guest program.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;Enabling it inlines accesses to PageTable::pointers into emitted code.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;Disabling this forces all memory accesses to go through the Memory::Read/Memory::Write functions.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable inline page tables</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_block_linking">
-          <property name="text">
-           <string>Enable block linking</string>
-          </property>
           <property name="toolTip">
            <string>
             &lt;div&gt;This optimization avoids dispatcher lookups by allowing emitted basic blocks to jump directly to other basic blocks if the destination PC is static.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable block linking</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_return_stack_buffer">
-          <property name="text">
-           <string>Enable return stack buffer</string>
-          </property>
           <property name="toolTip">
            <string>
             &lt;div&gt;This optimization avoids dispatcher lookups by keeping track potential return addresses of BL instructions. This approximates what happens with a return stack buffer on a real CPU.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable return stack buffer</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_fast_dispatcher">
-          <property name="text">
-           <string>Enable fast dispatcher</string>
-          </property>
           <property name="toolTip">
            <string>
             &lt;div&gt;Enable a two-tiered dispatch system. A faster dispatcher written in assembly has a small MRU cache of jump destinations is used first. If that fails, dispatch falls back to the slower C++ dispatcher.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable fast dispatcher</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_context_elimination">
-          <property name="text">
-           <string>Enable context elimination</string>
-          </property>
           <property name="toolTip">
            <string>
             &lt;div&gt;Enables an IR optimization that reduces unnecessary accesses to the CPU context structure.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable context elimination</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_const_prop">
-          <property name="text">
-           <string>Enable constant propagation</string>
-          </property>
           <property name="toolTip">
            <string>
             &lt;div&gt;Enables IR optimizations that involve constant propagation.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable constant propagation</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_misc_ir">
-          <property name="text">
-           <string>Enable miscellaneous optimizations</string>
-          </property>
           <property name="toolTip">
            <string>
             &lt;div&gt;Enables miscellaneous IR optimizations.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable miscellaneous optimizations</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_reduce_misalign_checks">
-          <property name="text">
-           <string>Enable misalignment check reduction</string>
-          </property>
           <property name="toolTip">
            <string>
-            &lt;div style="white-space: nowrap"&gt;When enabled, a misalignment is only triggered when an access crosses a page boundary.&lt;/div&gt;
-            &lt;div style="white-space: nowrap"&gt;When disabled, a misalignment is triggered on all misaligned accesses.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;When enabled, a misalignment is only triggered when an access crosses a page boundary.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;When disabled, a misalignment is triggered on all misaligned accesses.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable misalignment check reduction</string>
+          </property>
          </widget>
         </item>
         <item>
          <widget class="QCheckBox" name="cpuopt_fastmem">
-          <property name="text">
-           <string>Enable Host MMU Emulation</string>
-          </property>
           <property name="toolTip">
            <string>
-            &lt;div style="white-space: nowrap"&gt;This optimization speeds up memory accesses by the guest program.&lt;/div&gt;
-            &lt;div style="white-space: nowrap"&gt;Enabling it causes guest memory reads/writes to be done directly into memory and make use of Host's MMU.&lt;/div&gt;
-            &lt;div style="white-space: nowrap"&gt;Disabling this forces all memory accesses to use Software MMU Emulation.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;This optimization speeds up memory accesses by the guest program.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;Enabling it causes guest memory reads/writes to be done directly into memory and make use of Host's MMU.&lt;/div&gt;
+            &lt;div style=&quot;white-space: nowrap&quot;&gt;Disabling this forces all memory accesses to use Software MMU Emulation.&lt;/div&gt;
            </string>
           </property>
+          <property name="text">
+           <string>Enable Host MMU Emulation</string>
+          </property>
          </widget>
         </item>
        </layout>
diff --git a/src/yuzu/configuration/configure_debug.cpp b/src/yuzu/configuration/configure_debug.cpp
index cbe45a305..8fceb3878 100644
--- a/src/yuzu/configuration/configure_debug.cpp
+++ b/src/yuzu/configuration/configure_debug.cpp
@@ -43,6 +43,8 @@ void ConfigureDebug::SetConfiguration() {
     ui->use_auto_stub->setChecked(Settings::values.use_auto_stub.GetValue());
     ui->enable_graphics_debugging->setEnabled(runtime_lock);
     ui->enable_graphics_debugging->setChecked(Settings::values.renderer_debug.GetValue());
+    ui->enable_cpu_debugging->setEnabled(runtime_lock);
+    ui->enable_cpu_debugging->setChecked(Settings::values.cpu_debug_mode.GetValue());
     ui->disable_macro_jit->setEnabled(runtime_lock);
     ui->disable_macro_jit->setChecked(Settings::values.disable_macro_jit.GetValue());
     ui->extended_logging->setChecked(Settings::values.extended_logging.GetValue());
@@ -58,6 +60,7 @@ void ConfigureDebug::ApplyConfiguration() {
     Settings::values.use_debug_asserts = ui->use_debug_asserts->isChecked();
     Settings::values.use_auto_stub = ui->use_auto_stub->isChecked();
     Settings::values.renderer_debug = ui->enable_graphics_debugging->isChecked();
+    Settings::values.cpu_debug_mode = ui->enable_cpu_debugging->isChecked();
     Settings::values.disable_macro_jit = ui->disable_macro_jit->isChecked();
     Settings::values.extended_logging = ui->extended_logging->isChecked();
     Debugger::ToggleConsole();
diff --git a/src/yuzu/configuration/configure_debug.ui b/src/yuzu/configuration/configure_debug.ui
index c8087542f..1260ad6f0 100644
--- a/src/yuzu/configuration/configure_debug.ui
+++ b/src/yuzu/configuration/configure_debug.ui
@@ -7,7 +7,7 @@
     <x>0</x>
     <y>0</y>
     <width>400</width>
-    <height>486</height>
+    <height>777</height>
    </rect>
   </property>
   <property name="windowTitle">
@@ -192,34 +192,41 @@
         </property>
        </widget>
       </item>
-       <item>
-         <widget class="QCheckBox" name="use_debug_asserts">
-           <property name="text">
-             <string>Enable Debug Asserts</string>
-           </property>
-         </widget>
-       </item>
-       <item>
-         <widget class="QCheckBox" name="use_auto_stub">
-           <property name="text">
-             <string>Enable Auto-Stub</string>
-           </property>
-         </widget>
-       </item>
       <item>
-        <widget class="QLabel" name="label_5">
-          <property name="font">
-            <font>
-              <italic>true</italic>
-            </font>
-          </property>
-          <property name="text">
-            <string>This will be reset automatically when yuzu closes.</string>
-          </property>
-          <property name="indent">
-            <number>20</number>
-          </property>
-        </widget>
+       <widget class="QCheckBox" name="enable_cpu_debugging">
+        <property name="text">
+         <string>Enable CPU Debugging</string>
+        </property>
+       </widget>
+      </item>
+      <item>
+       <widget class="QCheckBox" name="use_debug_asserts">
+        <property name="text">
+         <string>Enable Debug Asserts</string>
+        </property>
+       </widget>
+      </item>
+      <item>
+       <widget class="QCheckBox" name="use_auto_stub">
+        <property name="text">
+         <string>Enable Auto-Stub</string>
+        </property>
+       </widget>
+      </item>
+      <item>
+       <widget class="QLabel" name="label_5">
+        <property name="font">
+         <font>
+          <italic>true</italic>
+         </font>
+        </property>
+        <property name="text">
+         <string>This will be reset automatically when yuzu closes.</string>
+        </property>
+        <property name="indent">
+         <number>20</number>
+        </property>
+       </widget>
       </item>
      </layout>
     </widget>
diff --git a/src/yuzu/configuration/configure_debug_tab.cpp b/src/yuzu/configuration/configure_debug_tab.cpp
new file mode 100644
index 000000000..67d369249
--- /dev/null
+++ b/src/yuzu/configuration/configure_debug_tab.cpp
@@ -0,0 +1,38 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "ui_configure_debug_tab.h"
+#include "yuzu/configuration/configure_debug_tab.h"
+
+ConfigureDebugTab::ConfigureDebugTab(QWidget* parent)
+    : QWidget(parent), ui(new Ui::ConfigureDebugTab) {
+    ui->setupUi(this);
+
+    SetConfiguration();
+}
+
+ConfigureDebugTab::~ConfigureDebugTab() = default;
+
+void ConfigureDebugTab::ApplyConfiguration() {
+    ui->debugTab->ApplyConfiguration();
+    ui->cpuDebugTab->ApplyConfiguration();
+}
+
+void ConfigureDebugTab::SetCurrentIndex(int index) {
+    ui->tabWidget->setCurrentIndex(index);
+}
+
+void ConfigureDebugTab::changeEvent(QEvent* event) {
+    if (event->type() == QEvent::LanguageChange) {
+        RetranslateUI();
+    }
+
+    QWidget::changeEvent(event);
+}
+
+void ConfigureDebugTab::RetranslateUI() {
+    ui->retranslateUi(this);
+}
+
+void ConfigureDebugTab::SetConfiguration() {}
diff --git a/src/yuzu/configuration/configure_debug_tab.h b/src/yuzu/configuration/configure_debug_tab.h
new file mode 100644
index 000000000..0a96d43d0
--- /dev/null
+++ b/src/yuzu/configuration/configure_debug_tab.h
@@ -0,0 +1,32 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <QWidget>
+
+namespace Ui {
+class ConfigureDebugTab;
+}
+
+class ConfigureDebugTab : public QWidget {
+    Q_OBJECT
+
+public:
+    explicit ConfigureDebugTab(QWidget* parent = nullptr);
+    ~ConfigureDebugTab() override;
+
+    void ApplyConfiguration();
+
+    void SetCurrentIndex(int index);
+
+private:
+    void changeEvent(QEvent* event) override;
+    void RetranslateUI();
+
+    void SetConfiguration();
+
+    std::unique_ptr<Ui::ConfigureDebugTab> ui;
+};
diff --git a/src/yuzu/configuration/configure_debug_tab.ui b/src/yuzu/configuration/configure_debug_tab.ui
new file mode 100644
index 000000000..7dc6dd704
--- /dev/null
+++ b/src/yuzu/configuration/configure_debug_tab.ui
@@ -0,0 +1,52 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<ui version="4.0">
+ <class>ConfigureDebugTab</class>
+ <widget class="QWidget" name="ConfigureDebugTab">
+  <property name="geometry">
+   <rect>
+    <x>0</x>
+    <y>0</y>
+    <width>320</width>
+    <height>240</height>
+   </rect>
+  </property>
+  <property name="windowTitle">
+   <string>Form</string>
+  </property>
+  <layout class="QVBoxLayout" name="verticalLayout">
+   <item>
+    <widget class="QTabWidget" name="tabWidget">
+     <property name="currentIndex">
+      <number>1</number>
+     </property>
+     <widget class="ConfigureDebug" name="debugTab">
+      <attribute name="title">
+       <string>General</string>
+      </attribute>
+     </widget>
+     <widget class="ConfigureCpuDebug" name="cpuDebugTab">
+      <attribute name="title">
+       <string>CPU</string>
+      </attribute>
+     </widget>
+    </widget>
+   </item>
+  </layout>
+ </widget>
+ <customwidgets>
+  <customwidget>
+   <class>ConfigureDebug</class>
+   <extends>QWidget</extends>
+   <header>configuration/configure_debug.h</header>
+   <container>1</container>
+  </customwidget>
+  <customwidget>
+   <class>ConfigureCpuDebug</class>
+   <extends>QWidget</extends>
+   <header>configuration/configure_cpu_debug.h</header>
+   <container>1</container>
+  </customwidget>
+ </customwidgets>
+ <resources/>
+ <connections/>
+</ui>
diff --git a/src/yuzu/configuration/configure_dialog.cpp b/src/yuzu/configuration/configure_dialog.cpp
index 371bc01b1..bc009b6b3 100644
--- a/src/yuzu/configuration/configure_dialog.cpp
+++ b/src/yuzu/configuration/configure_dialog.cpp
@@ -8,6 +8,7 @@
 #include <QListWidgetItem>
 #include <QPushButton>
 #include <QSignalBlocker>
+#include <QTabWidget>
 #include "common/settings.h"
 #include "core/core.h"
 #include "ui_configure.h"
@@ -32,6 +33,8 @@ ConfigureDialog::ConfigureDialog(QWidget* parent, HotkeyRegistry& registry,
     SetConfiguration();
     PopulateSelectionList();
 
+    connect(ui->tabWidget, &QTabWidget::currentChanged, this,
+            [this]() { ui->debugTab->SetCurrentIndex(0); });
     connect(ui->uiTab, &ConfigureUi::LanguageChanged, this, &ConfigureDialog::OnLanguageChanged);
     connect(ui->selectorList, &QListWidget::itemSelectionChanged, this,
             &ConfigureDialog::UpdateVisibleTabs);
@@ -59,7 +62,6 @@ void ConfigureDialog::ApplyConfiguration() {
     ui->inputTab->ApplyConfiguration();
     ui->hotkeysTab->ApplyConfiguration(registry);
     ui->cpuTab->ApplyConfiguration();
-    ui->cpuDebugTab->ApplyConfiguration();
     ui->graphicsTab->ApplyConfiguration();
     ui->graphicsAdvancedTab->ApplyConfiguration();
     ui->audioTab->ApplyConfiguration();
@@ -102,7 +104,7 @@ void ConfigureDialog::PopulateSelectionList() {
     const std::array<std::pair<QString, QList<QWidget*>>, 6> items{
         {{tr("General"), {ui->generalTab, ui->hotkeysTab, ui->uiTab, ui->webTab, ui->debugTab}},
          {tr("System"), {ui->systemTab, ui->profileManagerTab, ui->serviceTab, ui->filesystemTab}},
-         {tr("CPU"), {ui->cpuTab, ui->cpuDebugTab}},
+         {tr("CPU"), {ui->cpuTab}},
          {tr("Graphics"), {ui->graphicsTab, ui->graphicsAdvancedTab}},
          {tr("Audio"), {ui->audioTab}},
          {tr("Controls"), ui->inputTab->GetSubTabs()}},
diff --git a/src/yuzu/configuration/configure_graphics.cpp b/src/yuzu/configuration/configure_graphics.cpp
index 41a69d9b8..4d5b4c0e6 100644
--- a/src/yuzu/configuration/configure_graphics.cpp
+++ b/src/yuzu/configuration/configure_graphics.cpp
@@ -101,9 +101,9 @@ void ConfigureGraphics::SetConfiguration() {
         ConfigurationShared::SetHighlight(ui->bg_layout, !Settings::values.bg_red.UsingGlobal());
     }
 
-    UpdateBackgroundColorButton(QColor::fromRgbF(Settings::values.bg_red.GetValue(),
-                                                 Settings::values.bg_green.GetValue(),
-                                                 Settings::values.bg_blue.GetValue()));
+    UpdateBackgroundColorButton(QColor::fromRgb(Settings::values.bg_red.GetValue(),
+                                                Settings::values.bg_green.GetValue(),
+                                                Settings::values.bg_blue.GetValue()));
     UpdateDeviceComboBox();
 }
 
@@ -132,9 +132,9 @@ void ConfigureGraphics::ApplyConfiguration() {
             Settings::values.vulkan_device.SetValue(vulkan_device);
         }
         if (Settings::values.bg_red.UsingGlobal()) {
-            Settings::values.bg_red.SetValue(static_cast<float>(bg_color.redF()));
-            Settings::values.bg_green.SetValue(static_cast<float>(bg_color.greenF()));
-            Settings::values.bg_blue.SetValue(static_cast<float>(bg_color.blueF()));
+            Settings::values.bg_red.SetValue(static_cast<u8>(bg_color.red()));
+            Settings::values.bg_green.SetValue(static_cast<u8>(bg_color.green()));
+            Settings::values.bg_blue.SetValue(static_cast<u8>(bg_color.blue()));
         }
     } else {
         if (ui->api->currentIndex() == ConfigurationShared::USE_GLOBAL_INDEX) {
@@ -159,9 +159,9 @@ void ConfigureGraphics::ApplyConfiguration() {
             Settings::values.bg_red.SetGlobal(false);
             Settings::values.bg_green.SetGlobal(false);
             Settings::values.bg_blue.SetGlobal(false);
-            Settings::values.bg_red.SetValue(static_cast<float>(bg_color.redF()));
-            Settings::values.bg_green.SetValue(static_cast<float>(bg_color.greenF()));
-            Settings::values.bg_blue.SetValue(static_cast<float>(bg_color.blueF()));
+            Settings::values.bg_red.SetValue(static_cast<u8>(bg_color.red()));
+            Settings::values.bg_green.SetValue(static_cast<u8>(bg_color.green()));
+            Settings::values.bg_blue.SetValue(static_cast<u8>(bg_color.blue()));
         }
     }
 }
diff --git a/src/yuzu/configuration/configure_graphics_advanced.cpp b/src/yuzu/configuration/configure_graphics_advanced.cpp
index 8d13c9857..a9e611125 100644
--- a/src/yuzu/configuration/configure_graphics_advanced.cpp
+++ b/src/yuzu/configuration/configure_graphics_advanced.cpp
@@ -28,7 +28,6 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
     ui->anisotropic_filtering_combobox->setEnabled(runtime_lock);
 
     ui->use_vsync->setChecked(Settings::values.use_vsync.GetValue());
-    ui->disable_fps_limit->setChecked(Settings::values.disable_fps_limit.GetValue());
     ui->use_assembly_shaders->setChecked(Settings::values.use_assembly_shaders.GetValue());
     ui->use_asynchronous_shaders->setChecked(Settings::values.use_asynchronous_shaders.GetValue());
     ui->use_caches_gc->setChecked(Settings::values.use_caches_gc.GetValue());
@@ -59,8 +58,6 @@ void ConfigureGraphicsAdvanced::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.max_anisotropy,
                                              ui->anisotropic_filtering_combobox);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_vsync, ui->use_vsync, use_vsync);
-    ConfigurationShared::ApplyPerGameSetting(&Settings::values.disable_fps_limit,
-                                             ui->disable_fps_limit, disable_fps_limit);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_assembly_shaders,
                                              ui->use_assembly_shaders, use_assembly_shaders);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_asynchronous_shaders,
@@ -103,7 +100,6 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
     if (Settings::IsConfiguringGlobal()) {
         ui->gpu_accuracy->setEnabled(Settings::values.gpu_accuracy.UsingGlobal());
         ui->use_vsync->setEnabled(Settings::values.use_vsync.UsingGlobal());
-        ui->disable_fps_limit->setEnabled(Settings::values.disable_fps_limit.UsingGlobal());
         ui->use_assembly_shaders->setEnabled(Settings::values.use_assembly_shaders.UsingGlobal());
         ui->use_asynchronous_shaders->setEnabled(
             Settings::values.use_asynchronous_shaders.UsingGlobal());
@@ -116,8 +112,6 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
     }
 
     ConfigurationShared::SetColoredTristate(ui->use_vsync, Settings::values.use_vsync, use_vsync);
-    ConfigurationShared::SetColoredTristate(ui->disable_fps_limit,
-                                            Settings::values.disable_fps_limit, disable_fps_limit);
     ConfigurationShared::SetColoredTristate(
         ui->use_assembly_shaders, Settings::values.use_assembly_shaders, use_assembly_shaders);
     ConfigurationShared::SetColoredTristate(ui->use_asynchronous_shaders,
diff --git a/src/yuzu/configuration/configure_graphics_advanced.h b/src/yuzu/configuration/configure_graphics_advanced.h
index 6ac5f20ec..9148aacf2 100644
--- a/src/yuzu/configuration/configure_graphics_advanced.h
+++ b/src/yuzu/configuration/configure_graphics_advanced.h
@@ -35,7 +35,6 @@ private:
     std::unique_ptr<Ui::ConfigureGraphicsAdvanced> ui;
 
     ConfigurationShared::CheckState use_vsync;
-    ConfigurationShared::CheckState disable_fps_limit;
     ConfigurationShared::CheckState use_assembly_shaders;
     ConfigurationShared::CheckState use_asynchronous_shaders;
     ConfigurationShared::CheckState use_fast_gpu_time;
diff --git a/src/yuzu/configuration/configure_graphics_advanced.ui b/src/yuzu/configuration/configure_graphics_advanced.ui
index 18c43629e..ad0840355 100644
--- a/src/yuzu/configuration/configure_graphics_advanced.ui
+++ b/src/yuzu/configuration/configure_graphics_advanced.ui
@@ -77,24 +77,6 @@
          </widget>
         </item>
         <item>
-         <widget class="QCheckBox" name="disable_fps_limit">
-          <property name="enabled">
-           <bool>true</bool>
-          </property>
-          <property name="toolTip">
-           <string>
-            &lt;html&gt;&lt;head/&gt;&lt;body&gt;
-            &lt;p&gt;Presents guest frames as they become available, disabling the FPS limit in most titles.&lt;/p&gt;
-            &lt;p&gt;NOTE: Will cause instabilities.&lt;/p&gt;
-            &lt;/body&gt;&lt;/html&gt;
-           </string>
-          </property>
-          <property name="text">
-           <string>Disable framerate limit (experimental)</string>
-          </property>
-         </widget>
-        </item>
-        <item>
          <widget class="QCheckBox" name="use_assembly_shaders">
           <property name="toolTip">
            <string>Enabling this reduces shader stutter. Enables OpenGL assembly shaders on supported Nvidia devices (NV_gpu_program5 is required). This feature is experimental.</string>
diff --git a/src/yuzu/configuration/configure_input_advanced.ui b/src/yuzu/configuration/configure_input_advanced.ui
index 173130d8d..d3ef5bd06 100644
--- a/src/yuzu/configuration/configure_input_advanced.ui
+++ b/src/yuzu/configuration/configure_input_advanced.ui
@@ -2573,27 +2573,24 @@
                   </widget>
                 </item>
                 <item row="2" column="2">
-                    <widget class="QDoubleSpinBox" name="mouse_panning_sensitivity">
+                    <widget class="QSpinBox" name="mouse_panning_sensitivity">
                       <property name="toolTip">
                         <string>Mouse sensitivity</string>
                       </property>
                       <property name="alignment">
                         <set>Qt::AlignCenter</set>
                       </property>
-                      <property name="decimals">
-                        <number>2</number>
+                      <property name="suffix">
+                        <string>%</string>
                       </property>
                       <property name="minimum">
-                        <double>0.100000000000000</double>
+                        <number>1</number>
                       </property>
                       <property name="maximum">
-                        <double>16.000000000000000</double>
-                      </property>
-                      <property name="singleStep">
-                        <double>0.010000000000000</double>
+                        <number>100</number>
                       </property>
                       <property name="value">
-                        <double>1.000000000000000</double>
+                        <number>100</number>
                       </property>
                     </widget>
                 </item>
diff --git a/src/yuzu/main.cpp b/src/yuzu/main.cpp
index 5ed3b90b8..b7fd33ae7 100644
--- a/src/yuzu/main.cpp
+++ b/src/yuzu/main.cpp
@@ -11,11 +11,11 @@
 #endif
 
 // VFS includes must be before glad as they will conflict with Windows file api, which uses defines.
-#include "applets/controller.h"
-#include "applets/error.h"
-#include "applets/profile_select.h"
-#include "applets/software_keyboard.h"
-#include "applets/web_browser.h"
+#include "applets/qt_controller.h"
+#include "applets/qt_error.h"
+#include "applets/qt_profile_select.h"
+#include "applets/qt_software_keyboard.h"
+#include "applets/qt_web_browser.h"
 #include "common/nvidia_flags.h"
 #include "configuration/configure_input.h"
 #include "configuration/configure_per_game.h"
@@ -1355,6 +1355,9 @@ void GMainWindow::BootGame(const QString& filename, std::size_t program_index, S
 
     ConfigureVibration::SetAllVibrationDevices();
 
+    // Disable fps limit toggle when booting a new title
+    Settings::values.disable_fps_limit.SetValue(false);
+
     // Save configurations
     UpdateUISettings();
     game_list->SaveInterfaceLayout();
@@ -1428,8 +1431,10 @@ void GMainWindow::BootGame(const QString& filename, std::size_t program_index, S
             std::filesystem::path{filename.toStdU16String()}.filename());
     }
     const bool is_64bit = system.Kernel().CurrentProcess()->Is64BitProcess();
-    const auto instruction_set_suffix = is_64bit ? " (64-bit)" : " (32-bit)";
-    title_name += instruction_set_suffix;
+    const auto instruction_set_suffix = is_64bit ? tr("(64-bit)") : tr("(32-bit)");
+    title_name = tr("%1 %2", "%1 is the title name. %2 indicates if the title is 64-bit or 32-bit")
+                     .arg(QString::fromStdString(title_name), instruction_set_suffix)
+                     .toStdString();
     LOG_INFO(Frontend, "Booting game: {:016X} | {} | {}", title_id, title_name, title_version);
     const auto gpu_vendor = system.GPU().Renderer().GetDeviceVendor();
     UpdateWindowTitle(title_name, title_version, gpu_vendor);
@@ -2913,7 +2918,12 @@ void GMainWindow::UpdateStatusBar() {
     } else {
         emu_speed_label->setText(tr("Speed: %1%").arg(results.emulation_speed * 100.0, 0, 'f', 0));
     }
-    game_fps_label->setText(tr("Game: %1 FPS").arg(results.average_game_fps, 0, 'f', 0));
+    if (Settings::values.disable_fps_limit) {
+        game_fps_label->setText(
+            tr("Game: %1 FPS (Limit off)").arg(results.average_game_fps, 0, 'f', 0));
+    } else {
+        game_fps_label->setText(tr("Game: %1 FPS").arg(results.average_game_fps, 0, 'f', 0));
+    }
     emu_frametime_label->setText(tr("Frame: %1 ms").arg(results.frametime * 1000.0, 0, 'f', 2));
 
     emu_speed_label->setVisible(!Settings::values.use_multi_core.GetValue());
diff --git a/src/yuzu_cmd/CMakeLists.txt b/src/yuzu_cmd/CMakeLists.txt
index 4bf25727b..e55a19649 100644
--- a/src/yuzu_cmd/CMakeLists.txt
+++ b/src/yuzu_cmd/CMakeLists.txt
@@ -38,6 +38,11 @@ target_include_directories(yuzu-cmd PRIVATE ${RESOURCES_DIR})
 
 target_include_directories(yuzu-cmd PRIVATE ../../externals/Vulkan-Headers/include)
 
+if (YUZU_USE_EXTERNAL_SDL2)
+    target_compile_definitions(yuzu-cmd PRIVATE -DYUZU_USE_EXTERNAL_SDL2)
+    target_include_directories(yuzu-cmd PRIVATE ${PROJECT_BINARY_DIR}/externals/SDL/include)
+endif()
+
 if(UNIX AND NOT APPLE)
     install(TARGETS yuzu-cmd RUNTIME DESTINATION "${CMAKE_INSTALL_PREFIX}/bin")
 endif()
diff --git a/src/yuzu_cmd/config.cpp b/src/yuzu_cmd/config.cpp
index 325584a1a..b18056baf 100644
--- a/src/yuzu_cmd/config.cpp
+++ b/src/yuzu_cmd/config.cpp
@@ -242,17 +242,15 @@ static const std::array<int, 8> keyboard_mods{
 };
 
 template <>
-void Config::ReadSetting(const std::string& group, Settings::BasicSetting<float>& setting) {
-    setting = sdl2_config->GetReal(group, setting.GetLabel(), setting.GetDefault());
-}
-template <>
 void Config::ReadSetting(const std::string& group, Settings::BasicSetting<std::string>& setting) {
     setting = sdl2_config->Get(group, setting.GetLabel(), setting.GetDefault());
 }
+
 template <>
 void Config::ReadSetting(const std::string& group, Settings::BasicSetting<bool>& setting) {
     setting = sdl2_config->GetBoolean(group, setting.GetLabel(), setting.GetDefault());
 }
+
 template <typename Type>
 void Config::ReadSetting(const std::string& group, Settings::BasicSetting<Type>& setting) {
     setting = static_cast<Type>(sdl2_config->GetInteger(group, setting.GetLabel(),
@@ -294,6 +292,8 @@ void Config::ReadValues() {
 
     ReadSetting("ControlsGeneral", Settings::values.motion_device);
 
+    ReadSetting("ControlsGeneral", Settings::values.touch_device);
+
     ReadSetting("ControlsGeneral", Settings::values.keyboard_enabled);
 
     ReadSetting("ControlsGeneral", Settings::values.debug_pad_enabled);
@@ -416,11 +416,31 @@ void Config::ReadValues() {
     }
 
     ReadSetting("System", Settings::values.language_index);
+    ReadSetting("System", Settings::values.region_index);
     ReadSetting("System", Settings::values.time_zone_index);
+    ReadSetting("System", Settings::values.sound_index);
 
     // Core
     ReadSetting("Core", Settings::values.use_multi_core);
 
+    // Cpu
+    ReadSetting("Cpu", Settings::values.cpu_accuracy);
+    ReadSetting("Cpu", Settings::values.cpu_debug_mode);
+    ReadSetting("Cpu", Settings::values.cpuopt_page_tables);
+    ReadSetting("Cpu", Settings::values.cpuopt_block_linking);
+    ReadSetting("Cpu", Settings::values.cpuopt_return_stack_buffer);
+    ReadSetting("Cpu", Settings::values.cpuopt_fast_dispatcher);
+    ReadSetting("Cpu", Settings::values.cpuopt_context_elimination);
+    ReadSetting("Cpu", Settings::values.cpuopt_const_prop);
+    ReadSetting("Cpu", Settings::values.cpuopt_misc_ir);
+    ReadSetting("Cpu", Settings::values.cpuopt_reduce_misalign_checks);
+    ReadSetting("Cpu", Settings::values.cpuopt_fastmem);
+    ReadSetting("Cpu", Settings::values.cpuopt_unsafe_unfuse_fma);
+    ReadSetting("Cpu", Settings::values.cpuopt_unsafe_reduce_fp_error);
+    ReadSetting("Cpu", Settings::values.cpuopt_unsafe_ignore_standard_fpcr);
+    ReadSetting("Cpu", Settings::values.cpuopt_unsafe_inaccurate_nan);
+    ReadSetting("Cpu", Settings::values.cpuopt_unsafe_fastmem_check);
+
     // Renderer
     ReadSetting("Renderer", Settings::values.renderer_backend);
     ReadSetting("Renderer", Settings::values.renderer_debug);
@@ -440,6 +460,7 @@ void Config::ReadValues() {
     ReadSetting("Renderer", Settings::values.use_nvdec_emulation);
     ReadSetting("Renderer", Settings::values.accelerate_astc);
     ReadSetting("Renderer", Settings::values.use_fast_gpu_time);
+    ReadSetting("Renderer", Settings::values.use_caches_gc);
 
     ReadSetting("Renderer", Settings::values.bg_red);
     ReadSetting("Renderer", Settings::values.bg_green);
@@ -460,7 +481,6 @@ void Config::ReadValues() {
     // Debugging
     Settings::values.record_frame_times =
         sdl2_config->GetBoolean("Debugging", "record_frame_times", false);
-    ReadSetting("Debugging", Settings::values.program_args);
     ReadSetting("Debugging", Settings::values.dump_exefs);
     ReadSetting("Debugging", Settings::values.dump_nso);
     ReadSetting("Debugging", Settings::values.enable_fs_access_log);
diff --git a/src/yuzu_cmd/default_ini.h b/src/yuzu_cmd/default_ini.h
index cc9850aad..b362f10b4 100644
--- a/src/yuzu_cmd/default_ini.h
+++ b/src/yuzu_cmd/default_ini.h
@@ -65,6 +65,13 @@ button_screenshot=
 lstick=
 rstick=
 
+# To use the debug_pad, prepend `debug_pad_` before each button setting above.
+# i.e. debug_pad_button_a=
+
+# Enable debug pad inputs to the guest
+# 0 (default): Disabled, 1: Enabled
+debug_pad_enabled =
+
 # Whether to enable or disable vibration
 # 0: Disabled, 1 (default): Enabled
 vibration_enabled=
@@ -73,6 +80,10 @@ vibration_enabled=
 # 0 (default): Disabled, 1: Enabled
 enable_accurate_vibrations=
 
+# Enables controller motion inputs
+# 0: Disabled, 1 (default): Enabled
+motion_enabled =
+
 # for motion input, the following devices are available:
 #  - "motion_emu" (default) for emulating motion input from mouse input. Required parameters:
 #      - "update_period": update period in milliseconds (default to 100)
@@ -98,19 +109,30 @@ use_touch_from_button=
 #touch_from_button_maps_0_bind_1=bar
 # etc.
 
-# Most desktop operating systems do not expose a way to poll the motion state of the controllers
-# so as a way around it, cemuhook created a udp client/server protocol to broadcast the data directly
-# from a controller device to the client program. Citra has a client that can connect and read
-# from any cemuhook compatible motion program.
+# List of Cemuhook UDP servers, delimited by ','.
+# Default: 127.0.0.1:26760
+# Example: 127.0.0.1:26760,123.4.5.67:26761
+udp_input_servers =
 
-# IPv4 address of the udp input server (Default "127.0.0.1")
-udp_input_address=127.0.0.1
+# Enable controlling an axis via a mouse input.
+# 0 (default): Off, 1: On
+mouse_panning =
+
+# Set mouse sensitivity.
+# Default: 1.0
+mouse_panning_sensitivity =
+
+# Emulate an analog control stick from keyboard inputs.
+# 0 (default): Disabled, 1: Enabled
+emulate_analog_keyboard =
 
-# Port of the udp input server. (Default 26760)
-udp_input_port=
+# Enable mouse inputs to the guest
+# 0 (default): Disabled, 1: Enabled
+mouse_enabled =
 
-# The pad to request data on. Should be between 0 (Pad 1) and 3 (Pad 4). (Default 0)
-udp_pad_index=
+# Enable keyboard inputs to the guest
+# 0 (default): Disabled, 1: Enabled
+keyboard_enabled =
 
 [Core]
 # Whether to use multi-core for CPU emulation
@@ -118,6 +140,17 @@ udp_pad_index=
 use_multi_core=
 
 [Cpu]
+# Adjusts various optimizations.
+# Auto-select mode enables choice unsafe optimizations.
+# Accurate enables only safe optimizations.
+# Unsafe allows any unsafe optimizations.
+# 0 (default): Auto-select, 1: Accurate, 2: Enable unsafe optimizations
+cpu_accuracy =
+
+# Allow disabling safe optimizations.
+# 0 (default): Disabled, 1: Enabled
+cpu_debug_mode =
+
 # Enable inline page tables optimization (faster guest memory access)
 # 0: Disabled, 1 (default): Enabled
 cpuopt_page_tables =
@@ -154,6 +187,31 @@ cpuopt_reduce_misalign_checks =
 # 0: Disabled, 1 (default): Enabled
 cpuopt_fastmem =
 
+# Enable unfuse FMA (improve performance on CPUs without FMA)
+# Only enabled if cpu_accuracy is set to Unsafe. Automatically chosen with cpu_accuracy = Auto-select.
+# 0: Disabled, 1 (default): Enabled
+cpuopt_unsafe_unfuse_fma =
+
+# Enable faster FRSQRTE and FRECPE
+# Only enabled if cpu_accuracy is set to Unsafe.
+# 0: Disabled, 1 (default): Enabled
+cpuopt_unsafe_reduce_fp_error =
+
+# Enable faster ASIMD instructions (32 bits only)
+# Only enabled if cpu_accuracy is set to Unsafe. Automatically chosen with cpu_accuracy = Auto-select.
+# 0: Disabled, 1 (default): Enabled
+cpuopt_unsafe_ignore_standard_fpcr =
+
+# Enable inaccurate NaN handling
+# Only enabled if cpu_accuracy is set to Unsafe. Automatically chosen with cpu_accuracy = Auto-select.
+# 0: Disabled, 1 (default): Enabled
+cpuopt_unsafe_inaccurate_nan =
+
+# Disable address space checks (64 bits only)
+# Only enabled if cpu_accuracy is set to Unsafe. Automatically chosen with cpu_accuracy = Auto-select.
+# 0: Disabled, 1 (default): Enabled
+cpuopt_unsafe_fastmem_check =
+
 [Renderer]
 # Which backend API to use.
 # 0 (default): OpenGL, 1: Vulkan
@@ -166,14 +224,6 @@ debug =
 # Which Vulkan physical device to use (defaults to 0)
 vulkan_device =
 
-# Whether to use software or hardware rendering.
-# 0: Software, 1 (default): Hardware
-use_hw_renderer =
-
-# Whether to use the Just-In-Time (JIT) compiler for shader emulation
-# 0: Interpreter (slow), 1 (default): JIT (fast)
-use_shader_jit =
-
 # Aspect ratio
 # 0: Default (16:9), 1: Force 4:3, 2: Force 21:9, 3: Stretch to Window
 aspect_ratio =
@@ -211,57 +261,31 @@ use_frame_limit =
 frame_limit =
 
 # Whether to use disk based shader cache
-# 0 (default): Off, 1 : On
+# 0: Off, 1 (default): On
 use_disk_shader_cache =
 
 # Which gpu accuracy level to use
-# 0 (Normal), 1 (High), 2 (Extreme)
+# 0: Normal, 1 (default): High, 2: Extreme (Very slow)
 gpu_accuracy =
 
 # Whether to use asynchronous GPU emulation
 # 0 : Off (slow), 1 (default): On (fast)
 use_asynchronous_gpu_emulation =
 
-# Forces VSync on the display thread. Usually doesn't impact performance, but on some drivers it can
-# so only turn this off if you notice a speed difference.
+# Inform the guest that GPU operations completed more quickly than they did.
 # 0: Off, 1 (default): On
-use_vsync =
+use_fast_gpu_time =
 
 # Whether to use garbage collection or not for GPU caches.
 # 0 (default): Off, 1: On
 use_caches_gc =
 
 # The clear color for the renderer. What shows up on the sides of the bottom screen.
-# Must be in range of 0.0-1.0. Defaults to 1.0 for all.
+# Must be in range of 0-255. Defaults to 0 for all.
 bg_red =
 bg_blue =
 bg_green =
 
-[Layout]
-# Layout for the screen inside the render window.
-# 0 (default): Default Top Bottom Screen, 1: Single Screen Only, 2: Large Screen Small Screen
-layout_option =
-
-# Toggle custom layout (using the settings below) on or off.
-# 0 (default): Off, 1: On
-custom_layout =
-
-# Screen placement when using Custom layout option
-# 0x, 0y is the top left corner of the render window.
-custom_top_left =
-custom_top_top =
-custom_top_right =
-custom_top_bottom =
-custom_bottom_left =
-custom_bottom_top =
-custom_bottom_right =
-custom_bottom_bottom =
-
-# Swaps the prominent screen with the other screen.
-# For example, if Single Screen is chosen, setting this to 1 will display the bottom screen instead of the top screen.
-# 0 (default): Top Screen is prominent, 1: Bottom Screen is prominent
-swap_screen =
-
 [Audio]
 # Which audio output engine to use.
 # auto (default): Auto-select
@@ -281,7 +305,7 @@ enable_audio_stretching =
 output_device =
 
 # Output volume.
-# 1.0 (default): 100%, 0.0; mute
+# 100 (default): 100%, 0; mute
 volume =
 
 [Data Storage]
@@ -308,10 +332,6 @@ gamecard_path =
 # 1 (default): Yes, 0: No
 use_docked_mode =
 
-# Allow the use of NFC in games
-# 1 (default): Yes, 0 : No
-enable_nfc =
-
 # Sets the seed for the RNG generator built into the switch
 # rng_seed will be ignored and randomly generated if rng_seed_enabled is false
 rng_seed_enabled =
@@ -323,10 +343,6 @@ rng_seed =
 custom_rtc_enabled =
 custom_rtc =
 
-# Sets the account username, max length is 32 characters
-# yuzu (default)
-username = yuzu
-
 # Sets the systems language index
 # 0: Japanese, 1: English (default), 2: French, 3: German, 4: Italian, 5: Spanish, 6: Chinese,
 # 7: Korean, 8: Dutch, 9: Portuguese, 10: Russian, 11: Taiwanese, 12: British English, 13: Canadian French,
@@ -335,17 +351,25 @@ language_index =
 
 # The system region that yuzu will use during emulation
 # -1: Auto-select (default), 0: Japan, 1: USA, 2: Europe, 3: Australia, 4: China, 5: Korea, 6: Taiwan
-region_value =
+region_index =
 
 # The system time zone that yuzu will use during emulation
 # 0: Auto-select (default), 1: Default (system archive value), Others: Index for specified time zone
 time_zone_index =
 
+# Sets the sound output mode.
+# 0: Mono, 1 (default): Stereo, 2: Surround
+sound_index =
+
 [Miscellaneous]
 # A filter which removes logs below a certain logging level.
 # Examples: *:Debug Kernel.SVC:Trace Service.*:Critical
 log_filter = *:Trace
 
+# Use developer keys
+# 0 (default): Disabled, 1: Enabled
+use_dev_keys =
+
 [Debugging]
 # Record frame time data, can be found in the log directory. Boolean value
 record_frame_times =
@@ -355,6 +379,8 @@ dump_exefs=false
 dump_nso=false
 # Determines whether or not yuzu will save the filesystem access log.
 enable_fs_access_log=false
+# Enables verbose reporting services
+reporting_services =
 # Determines whether or not yuzu will report to the game that the emulated console is in Kiosk Mode
 # false: Retail/Normal Mode (default), true: Kiosk Mode
 quest_flag =
@@ -393,4 +419,4 @@ title_ids =
 # For each title ID, have a key/value pair called `disabled_<title_id>` equal to the names of the add-ons to disable (sep. by '|')
 # e.x. disabled_0100000000010000 = Update|DLC <- disables Updates and DLC on Super Mario Odyssey
 )";
-}
+} // namespace DefaultINI
diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
index 06b20c975..896181f0b 100644
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
@@ -2,15 +2,7 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-// Ignore -Wimplicit-fallthrough due to https://github.com/libsdl-org/SDL/issues/4307
-#ifdef __clang__
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wimplicit-fallthrough"
-#endif
 #include <SDL.h>
-#ifdef __clang__
-#pragma clang diagnostic pop
-#endif
 
 #include "common/logging/log.h"
 #include "common/scm_rev.h"
diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp
index 837a44be7..eadb41790 100644
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp
@@ -7,15 +7,7 @@
 #include <string>
 
 #define SDL_MAIN_HANDLED
-// Ignore -Wimplicit-fallthrough due to https://github.com/libsdl-org/SDL/issues/4307
-#ifdef __clang__
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wimplicit-fallthrough"
-#endif
 #include <SDL.h>
-#ifdef __clang__
-#pragma clang diagnostic pop
-#endif
 
 #include <fmt/format.h>
 #include <glad/glad.h>
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 3401ad4b4..152e56db8 100644
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp
@@ -15,16 +15,13 @@
 #include "video_core/renderer_vulkan/renderer_vulkan.h"
 #include "yuzu_cmd/emu_window/emu_window_sdl2_vk.h"
 
-// Include these late to avoid polluting everything with Xlib macros
-// Ignore -Wimplicit-fallthrough due to https://github.com/libsdl-org/SDL/issues/4307
-#ifdef __clang__
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wimplicit-fallthrough"
+#ifdef YUZU_USE_EXTERNAL_SDL2
+// Include this before SDL.h to prevent the external from including a dummy
+#define USING_GENERATED_CONFIG_H
+#include <SDL_config.h>
 #endif
+
 #include <SDL.h>
-#ifdef __clang__
-#pragma clang diagnostic pop
-#endif
 #include <SDL_syswm.h>
 
 EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem)
@@ -51,6 +48,11 @@ EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsyste
         window_info.type = Core::Frontend::WindowSystemType::Windows;
         window_info.render_surface = reinterpret_cast<void*>(wm.info.win.window);
         break;
+#else
+    case SDL_SYSWM_TYPE::SDL_SYSWM_WINDOWS:
+        LOG_CRITICAL(Frontend, "Window manager subsystem Windows not compiled");
+        std::exit(EXIT_FAILURE);
+        break;
 #endif
 #ifdef SDL_VIDEO_DRIVER_X11
     case SDL_SYSWM_TYPE::SDL_SYSWM_X11:
@@ -58,6 +60,11 @@ EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsyste
         window_info.display_connection = wm.info.x11.display;
         window_info.render_surface = reinterpret_cast<void*>(wm.info.x11.window);
         break;
+#else
+    case SDL_SYSWM_TYPE::SDL_SYSWM_X11:
+        LOG_CRITICAL(Frontend, "Window manager subsystem X11 not compiled");
+        std::exit(EXIT_FAILURE);
+        break;
 #endif
 #ifdef SDL_VIDEO_DRIVER_WAYLAND
     case SDL_SYSWM_TYPE::SDL_SYSWM_WAYLAND:
@@ -65,6 +72,11 @@ EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsyste
         window_info.display_connection = wm.info.wl.display;
         window_info.render_surface = wm.info.wl.surface;
         break;
+#else
+    case SDL_SYSWM_TYPE::SDL_SYSWM_WAYLAND:
+        LOG_CRITICAL(Frontend, "Window manager subsystem Wayland not compiled");
+        std::exit(EXIT_FAILURE);
+        break;
 #endif
     default:
         LOG_CRITICAL(Frontend, "Window manager subsystem not implemented");