23 files changed, 308 insertions, 86 deletions

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 6068c7a1f..a9f68a8f2 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -105,6 +105,8 @@ if (MSVC)
     set(CMAKE_EXE_LINKER_FLAGS_RELEASE "/DEBUG /MANIFEST:NO /INCREMENTAL:NO /OPT:REF,ICF" CACHE STRING "" FORCE)
 else()
     add_compile_options(
+        -fwrapv
+
         -Werror=all
         -Werror=extra
         -Werror=missing-declarations
@@ -129,7 +131,6 @@ else()
 
     if (ARCHITECTURE_x86_64)
         add_compile_options("-mcx16")
-        add_compile_options("-fwrapv")
     endif()
 
     if (APPLE AND CMAKE_CXX_COMPILER_ID STREQUAL Clang)
diff --git a/src/core/core.cpp b/src/core/core.cpp
index 2f67e60a9..e95ae80da 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -273,7 +273,8 @@ struct System::Impl {
         time_manager.Initialize();
 
         is_powered_on = true;
-        exit_lock = false;
+        exit_locked = false;
+        exit_requested = false;
 
         microprofile_cpu[0] = MICROPROFILE_TOKEN(ARM_CPU0);
         microprofile_cpu[1] = MICROPROFILE_TOKEN(ARM_CPU1);
@@ -398,7 +399,8 @@ struct System::Impl {
         }
 
         is_powered_on = false;
-        exit_lock = false;
+        exit_locked = false;
+        exit_requested = false;
 
         if (gpu_core != nullptr) {
             gpu_core->NotifyShutdown();
@@ -507,7 +509,8 @@ struct System::Impl {
 
     CpuManager cpu_manager;
     std::atomic_bool is_powered_on{};
-    bool exit_lock = false;
+    bool exit_locked = false;
+    bool exit_requested = false;
 
     bool nvdec_active{};
 
@@ -943,12 +946,20 @@ const Service::Time::TimeManager& System::GetTimeManager() const {
     return impl->time_manager;
 }
 
-void System::SetExitLock(bool locked) {
-    impl->exit_lock = locked;
+void System::SetExitLocked(bool locked) {
+    impl->exit_locked = locked;
 }
 
-bool System::GetExitLock() const {
-    return impl->exit_lock;
+bool System::GetExitLocked() const {
+    return impl->exit_locked;
+}
+
+void System::SetExitRequested(bool requested) {
+    impl->exit_requested = requested;
+}
+
+bool System::GetExitRequested() const {
+    return impl->exit_requested;
 }
 
 void System::SetApplicationProcessBuildID(const CurrentBuildProcessID& id) {
diff --git a/src/core/core.h b/src/core/core.h
index c70ea1965..a9ff9315e 100644
--- a/src/core/core.h
+++ b/src/core/core.h
@@ -412,8 +412,11 @@ public:
     /// Gets an immutable reference to the Room Network.
     [[nodiscard]] const Network::RoomNetwork& GetRoomNetwork() const;
 
-    void SetExitLock(bool locked);
-    [[nodiscard]] bool GetExitLock() const;
+    void SetExitLocked(bool locked);
+    bool GetExitLocked() const;
+
+    void SetExitRequested(bool requested);
+    bool GetExitRequested() const;
 
     void SetApplicationProcessBuildID(const CurrentBuildProcessID& id);
     [[nodiscard]] const CurrentBuildProcessID& GetApplicationProcessBuildID() const;
diff --git a/src/core/file_sys/content_archive.cpp b/src/core/file_sys/content_archive.cpp
index 44e6852fe..7d2f0abb8 100644
--- a/src/core/file_sys/content_archive.cpp
+++ b/src/core/file_sys/content_archive.cpp
@@ -22,6 +22,10 @@
 
 namespace FileSys {
 
+static u8 MasterKeyIdForKeyGeneration(u8 key_generation) {
+    return std::max<u8>(key_generation, 1) - 1;
+}
+
 NCA::NCA(VirtualFile file_, const NCA* base_nca)
     : file(std::move(file_)), keys{Core::Crypto::KeyManager::Instance()} {
     if (file == nullptr) {
@@ -41,12 +45,17 @@ NCA::NCA(VirtualFile file_, const NCA* base_nca)
         return;
     }
 
+    // Ensure we have the proper key area keys to continue.
+    const u8 master_key_id = MasterKeyIdForKeyGeneration(reader->GetKeyGeneration());
+    if (!keys.HasKey(Core::Crypto::S128KeyType::KeyArea, master_key_id, reader->GetKeyIndex())) {
+        status = Loader::ResultStatus::ErrorMissingKeyAreaKey;
+        return;
+    }
+
     RightsId rights_id{};
     reader->GetRightsId(rights_id.data(), rights_id.size());
     if (rights_id != RightsId{}) {
         // External decryption key required; provide it here.
-        const auto key_generation = std::max<s32>(reader->GetKeyGeneration(), 1) - 1;
-
         u128 rights_id_u128;
         std::memcpy(rights_id_u128.data(), rights_id.data(), sizeof(rights_id));
 
@@ -57,12 +66,12 @@ NCA::NCA(VirtualFile file_, const NCA* base_nca)
             return;
         }
 
-        if (!keys.HasKey(Core::Crypto::S128KeyType::Titlekek, key_generation)) {
+        if (!keys.HasKey(Core::Crypto::S128KeyType::Titlekek, master_key_id)) {
             status = Loader::ResultStatus::ErrorMissingTitlekek;
             return;
         }
 
-        auto titlekek = keys.GetKey(Core::Crypto::S128KeyType::Titlekek, key_generation);
+        auto titlekek = keys.GetKey(Core::Crypto::S128KeyType::Titlekek, master_key_id);
         Core::Crypto::AESCipher<Core::Crypto::Key128> cipher(titlekek, Core::Crypto::Mode::ECB);
         cipher.Transcode(titlekey.data(), titlekey.size(), titlekey.data(),
                          Core::Crypto::Op::Decrypt);
diff --git a/src/core/hle/service/am/am.cpp b/src/core/hle/service/am/am.cpp
index da33f0e44..e92f400de 100644
--- a/src/core/hle/service/am/am.cpp
+++ b/src/core/hle/service/am/am.cpp
@@ -341,7 +341,7 @@ void ISelfController::Exit(HLERequestContext& ctx) {
 void ISelfController::LockExit(HLERequestContext& ctx) {
     LOG_DEBUG(Service_AM, "called");
 
-    system.SetExitLock(true);
+    system.SetExitLocked(true);
 
     IPC::ResponseBuilder rb{ctx, 2};
     rb.Push(ResultSuccess);
@@ -350,10 +350,14 @@ void ISelfController::LockExit(HLERequestContext& ctx) {
 void ISelfController::UnlockExit(HLERequestContext& ctx) {
     LOG_DEBUG(Service_AM, "called");
 
-    system.SetExitLock(false);
+    system.SetExitLocked(false);
 
     IPC::ResponseBuilder rb{ctx, 2};
     rb.Push(ResultSuccess);
+
+    if (system.GetExitRequested()) {
+        system.Exit();
+    }
 }
 
 void ISelfController::EnterFatalSection(HLERequestContext& ctx) {
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
index 3ad668a47..d9872ecc2 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
@@ -558,7 +558,7 @@ void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
     if (multi_component) {
         if (info.num_derivates >= 3) {
             const auto offset_vec{ctx.var_alloc.Consume(offset)};
-            ctx.Add("{}=textureGrad({},{},vec3({}.xz, {}.x),vec3({}.yz, {}.y));", texel, texture,
+            ctx.Add("{}=textureGrad({},{},vec3({}.xz, {}.x),vec3({}.yw, {}.y));", texel, texture,
                     coords, derivatives_vec, offset_vec, derivatives_vec, offset_vec);
             return;
         }
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
index 7d901c04b..34240b36f 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
@@ -91,6 +91,34 @@ public:
         }
     }
 
+    explicit ImageOperands(EmitContext& ctx, bool has_lod_clamp, Id derivates_1, Id derivates_2,
+                           Id offset, Id lod_clamp) {
+        if (!Sirit::ValidId(derivates_1) || !Sirit::ValidId(derivates_2)) {
+            throw LogicError("Derivates must be present");
+        }
+        boost::container::static_vector<Id, 3> deriv_1_accum{
+            ctx.OpCompositeExtract(ctx.F32[1], derivates_1, 0),
+            ctx.OpCompositeExtract(ctx.F32[1], derivates_1, 2),
+            ctx.OpCompositeExtract(ctx.F32[1], derivates_2, 0),
+        };
+        boost::container::static_vector<Id, 3> deriv_2_accum{
+            ctx.OpCompositeExtract(ctx.F32[1], derivates_1, 1),
+            ctx.OpCompositeExtract(ctx.F32[1], derivates_1, 3),
+            ctx.OpCompositeExtract(ctx.F32[1], derivates_2, 1),
+        };
+        const Id derivates_id1{ctx.OpCompositeConstruct(
+            ctx.F32[3], std::span{deriv_1_accum.data(), deriv_1_accum.size()})};
+        const Id derivates_id2{ctx.OpCompositeConstruct(
+            ctx.F32[3], std::span{deriv_2_accum.data(), deriv_2_accum.size()})};
+        Add(spv::ImageOperandsMask::Grad, derivates_id1, derivates_id2);
+        if (Sirit::ValidId(offset)) {
+            Add(spv::ImageOperandsMask::Offset, offset);
+        }
+        if (has_lod_clamp) {
+            Add(spv::ImageOperandsMask::MinLod, lod_clamp);
+        }
+    }
+
     std::span<const Id> Span() const noexcept {
         return std::span{operands.data(), operands.size()};
     }
@@ -524,8 +552,11 @@ Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, I
 Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
                      Id derivates, Id offset, Id lod_clamp) {
     const auto info{inst->Flags<IR::TextureInstInfo>()};
-    const ImageOperands operands(ctx, info.has_lod_clamp != 0, derivates, info.num_derivates,
-                                 offset, lod_clamp);
+    const auto operands =
+        info.num_derivates == 3
+            ? ImageOperands(ctx, info.has_lod_clamp != 0, derivates, offset, {}, lod_clamp)
+            : ImageOperands(ctx, info.has_lod_clamp != 0, derivates, info.num_derivates, offset,
+                            lod_clamp);
     return Emit(&EmitContext::OpImageSparseSampleExplicitLod,
                 &EmitContext::OpImageSampleExplicitLod, ctx, inst, ctx.F32[4],
                 Texture(ctx, info, index), coords, operands.Mask(), operands.Span());
diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp
index 753c62098..e593132e6 100644
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp
@@ -161,7 +161,8 @@ enum class SpecialRegister : u64 {
         LOG_WARNING(Shader, "(STUBBED) SR_AFFINITY");
         return ir.Imm32(0); // This is the default value hardware returns.
     default:
-        throw NotImplementedException("S2R special register {}", special_register);
+        LOG_CRITICAL(Shader, "(STUBBED) Special register {}", special_register);
+        return ir.Imm32(0); // This is the default value hardware returns.
     }
 }
 } // Anonymous namespace
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 9f1b340a9..58ce0d8c2 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -14,6 +14,7 @@
 namespace Tegra {
 
 constexpr u32 MacroRegistersStart = 0xE00;
+constexpr u32 ComputeInline = 0x6D;
 
 DmaPusher::DmaPusher(Core::System& system_, GPU& gpu_, MemoryManager& memory_manager_,
                      Control::ChannelState& channel_state_)
@@ -83,12 +84,35 @@ bool DmaPusher::Step() {
                     dma_state.dma_get, command_list_header.size * sizeof(u32));
             }
         }
-        Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
-                                     Core::Memory::GuestMemoryFlags::UnsafeRead>
-            headers(memory_manager, dma_state.dma_get, command_list_header.size, &command_headers);
-        ProcessCommands(headers);
+        const auto safe_process = [&] {
+            Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
+                                         Core::Memory::GuestMemoryFlags::SafeRead>
+                headers(memory_manager, dma_state.dma_get, command_list_header.size,
+                        &command_headers);
+            ProcessCommands(headers);
+        };
+        const auto unsafe_process = [&] {
+            Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
+                                         Core::Memory::GuestMemoryFlags::UnsafeRead>
+                headers(memory_manager, dma_state.dma_get, command_list_header.size,
+                        &command_headers);
+            ProcessCommands(headers);
+        };
+        if (Settings::IsGPULevelHigh()) {
+            if (dma_state.method >= MacroRegistersStart) {
+                unsafe_process();
+                return true;
+            }
+            if (subchannel_type[dma_state.subchannel] == Engines::EngineTypes::KeplerCompute &&
+                dma_state.method == ComputeInline) {
+                unsafe_process();
+                return true;
+            }
+            safe_process();
+            return true;
+        }
+        unsafe_process();
     }
-
     return true;
 }
 
diff --git a/src/video_core/dma_pusher.h b/src/video_core/dma_pusher.h
index 8a2784cdc..c9fab2d90 100644
--- a/src/video_core/dma_pusher.h
+++ b/src/video_core/dma_pusher.h
@@ -130,8 +130,10 @@ public:
 
     void DispatchCalls();
 
-    void BindSubchannel(Engines::EngineInterface* engine, u32 subchannel_id) {
+    void BindSubchannel(Engines::EngineInterface* engine, u32 subchannel_id,
+                        Engines::EngineTypes engine_type) {
         subchannels[subchannel_id] = engine;
+        subchannel_type[subchannel_id] = engine_type;
     }
 
     void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
@@ -170,6 +172,7 @@ private:
     const bool ib_enable{true}; ///< IB mode enabled
 
     std::array<Engines::EngineInterface*, max_subchannels> subchannels{};
+    std::array<Engines::EngineTypes, max_subchannels> subchannel_type;
 
     GPU& gpu;
     Core::System& system;
diff --git a/src/video_core/engines/engine_interface.h b/src/video_core/engines/engine_interface.h
index 392322358..54631ee6c 100644
--- a/src/video_core/engines/engine_interface.h
+++ b/src/video_core/engines/engine_interface.h
@@ -11,6 +11,14 @@
 
 namespace Tegra::Engines {
 
+enum class EngineTypes : u32 {
+    KeplerCompute,
+    Maxwell3D,
+    Fermi2D,
+    MaxwellDMA,
+    KeplerMemory,
+};
+
 class EngineInterface {
 public:
     virtual ~EngineInterface() = default;
diff --git a/src/video_core/engines/engine_upload.h b/src/video_core/engines/engine_upload.h
index 7242d2529..21bf8aeb4 100644
--- a/src/video_core/engines/engine_upload.h
+++ b/src/video_core/engines/engine_upload.h
@@ -69,6 +69,14 @@ public:
     /// Binds a rasterizer to this engine.
     void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
 
+    GPUVAddr ExecTargetAddress() const {
+        return regs.dest.Address();
+    }
+
+    u32 GetUploadSize() const {
+        return copy_size;
+    }
+
 private:
     void ProcessData(std::span<const u8> read_buffer);
 
diff --git a/src/video_core/engines/kepler_compute.cpp b/src/video_core/engines/kepler_compute.cpp
index a38d9528a..cd61ab222 100644
--- a/src/video_core/engines/kepler_compute.cpp
+++ b/src/video_core/engines/kepler_compute.cpp
@@ -43,16 +43,33 @@ void KeplerCompute::CallMethod(u32 method, u32 method_argument, bool is_last_cal
 
     switch (method) {
     case KEPLER_COMPUTE_REG_INDEX(exec_upload): {
+        UploadInfo info{.upload_address = upload_address,
+                        .exec_address = upload_state.ExecTargetAddress(),
+                        .copy_size = upload_state.GetUploadSize()};
+        uploads.push_back(info);
         upload_state.ProcessExec(regs.exec_upload.linear != 0);
         break;
     }
     case KEPLER_COMPUTE_REG_INDEX(data_upload): {
+        upload_address = current_dma_segment;
         upload_state.ProcessData(method_argument, is_last_call);
         break;
     }
-    case KEPLER_COMPUTE_REG_INDEX(launch):
+    case KEPLER_COMPUTE_REG_INDEX(launch): {
+        const GPUVAddr launch_desc_loc = regs.launch_desc_loc.Address();
+
+        for (auto& data : uploads) {
+            const GPUVAddr offset = data.exec_address - launch_desc_loc;
+            if (offset / sizeof(u32) == LAUNCH_REG_INDEX(grid_dim_x) &&
+                memory_manager.IsMemoryDirty(data.upload_address, data.copy_size)) {
+                indirect_compute = {data.upload_address};
+            }
+        }
+        uploads.clear();
         ProcessLaunch();
+        indirect_compute = std::nullopt;
         break;
+    }
     default:
         break;
     }
@@ -62,6 +79,7 @@ void KeplerCompute::CallMultiMethod(u32 method, const u32* base_start, u32 amoun
                                     u32 methods_pending) {
     switch (method) {
     case KEPLER_COMPUTE_REG_INDEX(data_upload):
+        upload_address = current_dma_segment;
         upload_state.ProcessData(base_start, amount);
         return;
     default:
diff --git a/src/video_core/engines/kepler_compute.h b/src/video_core/engines/kepler_compute.h
index 2092e685f..735e05fb4 100644
--- a/src/video_core/engines/kepler_compute.h
+++ b/src/video_core/engines/kepler_compute.h
@@ -5,6 +5,7 @@
 
 #include <array>
 #include <cstddef>
+#include <optional>
 #include <vector>
 #include "common/bit_field.h"
 #include "common/common_funcs.h"
@@ -36,6 +37,9 @@ namespace Tegra::Engines {
 #define KEPLER_COMPUTE_REG_INDEX(field_name)                                                       \
     (offsetof(Tegra::Engines::KeplerCompute::Regs, field_name) / sizeof(u32))
 
+#define LAUNCH_REG_INDEX(field_name)                                                               \
+    (offsetof(Tegra::Engines::KeplerCompute::LaunchParams, field_name) / sizeof(u32))
+
 class KeplerCompute final : public EngineInterface {
 public:
     explicit KeplerCompute(Core::System& system, MemoryManager& memory_manager);
@@ -201,6 +205,10 @@ public:
     void CallMultiMethod(u32 method, const u32* base_start, u32 amount,
                          u32 methods_pending) override;
 
+    std::optional<GPUVAddr> GetIndirectComputeAddress() const {
+        return indirect_compute;
+    }
+
 private:
     void ProcessLaunch();
 
@@ -216,6 +224,15 @@ private:
     MemoryManager& memory_manager;
     VideoCore::RasterizerInterface* rasterizer = nullptr;
     Upload::State upload_state;
+    GPUVAddr upload_address;
+
+    struct UploadInfo {
+        GPUVAddr upload_address;
+        GPUVAddr exec_address;
+        u32 copy_size;
+    };
+    std::vector<UploadInfo> uploads;
+    std::optional<GPUVAddr> indirect_compute{};
 };
 
 #define ASSERT_REG_POSITION(field_name, position)                                                  \
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index c3696096d..06e349e43 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -257,6 +257,7 @@ u32 Maxwell3D::GetMaxCurrentVertices() {
         const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin);
         num_vertices = std::max(
             num_vertices, address_size / std::max(attribute.SizeInBytes(), array.stride.Value()));
+        break;
     }
     return num_vertices;
 }
@@ -269,10 +270,13 @@ size_t Maxwell3D::EstimateIndexBufferSize() {
                                                         std::numeric_limits<u32>::max()};
     const size_t byte_size = regs.index_buffer.FormatSizeInBytes();
     const size_t log2_byte_size = Common::Log2Ceil64(byte_size);
+    const size_t cap{GetMaxCurrentVertices() * 3 * byte_size};
+    const size_t lower_cap =
+        std::min<size_t>(static_cast<size_t>(end_address - start_address), cap);
     return std::min<size_t>(
         memory_manager.GetMemoryLayoutSize(start_address, byte_size * max_sizes[log2_byte_size]) /
             byte_size,
-        static_cast<size_t>(end_address - start_address));
+        lower_cap);
 }
 
 u32 Maxwell3D::ProcessShadowRam(u32 method, u32 argument) {
diff --git a/src/video_core/engines/puller.cpp b/src/video_core/engines/puller.cpp
index 7718a09b3..6de2543b7 100644
--- a/src/video_core/engines/puller.cpp
+++ b/src/video_core/engines/puller.cpp
@@ -34,19 +34,24 @@ void Puller::ProcessBindMethod(const MethodCall& method_call) {
     bound_engines[method_call.subchannel] = engine_id;
     switch (engine_id) {
     case EngineID::FERMI_TWOD_A:
-        dma_pusher.BindSubchannel(channel_state.fermi_2d.get(), method_call.subchannel);
+        dma_pusher.BindSubchannel(channel_state.fermi_2d.get(), method_call.subchannel,
+                                  EngineTypes::Fermi2D);
         break;
     case EngineID::MAXWELL_B:
-        dma_pusher.BindSubchannel(channel_state.maxwell_3d.get(), method_call.subchannel);
+        dma_pusher.BindSubchannel(channel_state.maxwell_3d.get(), method_call.subchannel,
+                                  EngineTypes::Maxwell3D);
         break;
     case EngineID::KEPLER_COMPUTE_B:
-        dma_pusher.BindSubchannel(channel_state.kepler_compute.get(), method_call.subchannel);
+        dma_pusher.BindSubchannel(channel_state.kepler_compute.get(), method_call.subchannel,
+                                  EngineTypes::KeplerCompute);
         break;
     case EngineID::MAXWELL_DMA_COPY_A:
-        dma_pusher.BindSubchannel(channel_state.maxwell_dma.get(), method_call.subchannel);
+        dma_pusher.BindSubchannel(channel_state.maxwell_dma.get(), method_call.subchannel,
+                                  EngineTypes::MaxwellDMA);
         break;
     case EngineID::KEPLER_INLINE_TO_MEMORY_B:
-        dma_pusher.BindSubchannel(channel_state.kepler_memory.get(), method_call.subchannel);
+        dma_pusher.BindSubchannel(channel_state.kepler_memory.get(), method_call.subchannel,
+                                  EngineTypes::KeplerMemory);
         break;
     default:
         UNIMPLEMENTED_MSG("Unimplemented engine {:04X}", engine_id);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 1ba31be88..dd03efecd 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -380,6 +380,17 @@ void RasterizerOpenGL::DispatchCompute() {
     pipeline->SetEngine(kepler_compute, gpu_memory);
     pipeline->Configure();
     const auto& qmd{kepler_compute->launch_description};
+    auto indirect_address = kepler_compute->GetIndirectComputeAddress();
+    if (indirect_address) {
+        // DispatchIndirect
+        static constexpr auto sync_info = VideoCommon::ObtainBufferSynchronize::FullSynchronize;
+        const auto post_op = VideoCommon::ObtainBufferOperation::DiscardWrite;
+        const auto [buffer, offset] =
+            buffer_cache.ObtainBuffer(*indirect_address, 12, sync_info, post_op);
+        glBindBuffer(GL_DISPATCH_INDIRECT_BUFFER, buffer->Handle());
+        glDispatchComputeIndirect(static_cast<GLintptr>(offset));
+        return;
+    }
     glDispatchCompute(qmd.grid_dim_x, qmd.grid_dim_y, qmd.grid_dim_z);
     ++num_queued_commands;
     has_written_global_memory |= pipeline->WritesGlobalMemory();
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index fe432dfe1..4f83a88e1 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -665,6 +665,19 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline(
         std::move(modules), infos);
 
 } catch (const Shader::Exception& exception) {
+    auto hash = key.Hash();
+    size_t env_index{0};
+    for (size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
+        if (key.unique_hashes[index] == 0) {
+            continue;
+        }
+        Shader::Environment& env{*envs[env_index]};
+        ++env_index;
+
+        const u32 cfg_offset{static_cast<u32>(env.StartAddress() + sizeof(Shader::ProgramHeader))};
+        Shader::Maxwell::Flow::CFG cfg(env, pools.flow_block, cfg_offset, index == 0);
+        env.Dump(hash, key.unique_hashes[index]);
+    }
     LOG_ERROR(Render_Vulkan, "{}", exception.what());
     return nullptr;
 }
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 032f694bc..01e76a82c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -463,6 +463,20 @@ void RasterizerVulkan::DispatchCompute() {
     pipeline->Configure(*kepler_compute, *gpu_memory, scheduler, buffer_cache, texture_cache);
 
     const auto& qmd{kepler_compute->launch_description};
+    auto indirect_address = kepler_compute->GetIndirectComputeAddress();
+    if (indirect_address) {
+        // DispatchIndirect
+        static constexpr auto sync_info = VideoCommon::ObtainBufferSynchronize::FullSynchronize;
+        const auto post_op = VideoCommon::ObtainBufferOperation::DiscardWrite;
+        const auto [buffer, offset] =
+            buffer_cache.ObtainBuffer(*indirect_address, 12, sync_info, post_op);
+        scheduler.RequestOutsideRenderPassOperationContext();
+        scheduler.Record([indirect_buffer = buffer->Handle(),
+                          indirect_offset = offset](vk::CommandBuffer cmdbuf) {
+            cmdbuf.DispatchIndirect(indirect_buffer, indirect_offset);
+        });
+        return;
+    }
     const std::array<u32, 3> dim{qmd.grid_dim_x, qmd.grid_dim_y, qmd.grid_dim_z};
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([dim](vk::CommandBuffer cmdbuf) { cmdbuf.Dispatch(dim[0], dim[1], dim[2]); });
diff --git a/src/video_core/vulkan_common/vulkan_wrapper.cpp b/src/video_core/vulkan_common/vulkan_wrapper.cpp
index 78e5a248f..c3f388d89 100644
--- a/src/video_core/vulkan_common/vulkan_wrapper.cpp
+++ b/src/video_core/vulkan_common/vulkan_wrapper.cpp
@@ -92,6 +92,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdCopyImage);
     X(vkCmdCopyImageToBuffer);
     X(vkCmdDispatch);
+    X(vkCmdDispatchIndirect);
     X(vkCmdDraw);
     X(vkCmdDrawIndexed);
     X(vkCmdDrawIndirect);
diff --git a/src/video_core/vulkan_common/vulkan_wrapper.h b/src/video_core/vulkan_common/vulkan_wrapper.h
index c226a2a29..049fa8038 100644
--- a/src/video_core/vulkan_common/vulkan_wrapper.h
+++ b/src/video_core/vulkan_common/vulkan_wrapper.h
@@ -203,6 +203,7 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkCmdCopyImage vkCmdCopyImage{};
     PFN_vkCmdCopyImageToBuffer vkCmdCopyImageToBuffer{};
     PFN_vkCmdDispatch vkCmdDispatch{};
+    PFN_vkCmdDispatchIndirect vkCmdDispatchIndirect{};
     PFN_vkCmdDraw vkCmdDraw{};
     PFN_vkCmdDrawIndexed vkCmdDrawIndexed{};
     PFN_vkCmdDrawIndirect vkCmdDrawIndirect{};
@@ -1209,6 +1210,10 @@ public:
         dld->vkCmdDispatch(handle, x, y, z);
     }
 
+    void DispatchIndirect(VkBuffer indirect_buffer, VkDeviceSize offset) const noexcept {
+        dld->vkCmdDispatchIndirect(handle, indirect_buffer, offset);
+    }
+
     void PipelineBarrier(VkPipelineStageFlags src_stage_mask, VkPipelineStageFlags dst_stage_mask,
                          VkDependencyFlags dependency_flags, Span<VkMemoryBarrier> memory_barriers,
                          Span<VkBufferMemoryBarrier> buffer_barriers,
diff --git a/src/yuzu/main.cpp b/src/yuzu/main.cpp
index 33c9fd0af..f2e6c03f0 100644
--- a/src/yuzu/main.cpp
+++ b/src/yuzu/main.cpp
@@ -2010,8 +2010,16 @@ bool GMainWindow::OnShutdownBegin() {
 
     emit EmulationStopping();
 
+    int shutdown_time = 1000;
+
+    if (system->DebuggerEnabled()) {
+        shutdown_time = 0;
+    } else if (system->GetExitLocked()) {
+        shutdown_time = 5000;
+    }
+
     shutdown_timer.setSingleShot(true);
-    shutdown_timer.start(system->DebuggerEnabled() ? 0 : 5000);
+    shutdown_timer.start(shutdown_time);
     connect(&shutdown_timer, &QTimer::timeout, this, &GMainWindow::OnEmulationStopTimeExpired);
     connect(emu_thread.get(), &QThread::finished, this, &GMainWindow::OnEmulationStopped);
 
@@ -2573,50 +2581,48 @@ void GMainWindow::OnGameListDumpRomFS(u64 program_id, const std::string& game_pa
         return;
     }
 
-    FileSys::VirtualFile base_romfs;
-    if (loader->ReadRomFS(base_romfs) != Loader::ResultStatus::Success) {
-        failed();
-        return;
-    }
+    FileSys::VirtualFile packed_update_raw{};
+    loader->ReadUpdateRaw(packed_update_raw);
 
     const auto& installed = system->GetContentProvider();
-    const auto romfs_title_id = SelectRomFSDumpTarget(installed, program_id);
 
-    if (!romfs_title_id) {
+    u64 title_id{};
+    u8 raw_type{};
+    if (!SelectRomFSDumpTarget(installed, program_id, &title_id, &raw_type)) {
         failed();
         return;
     }
 
-    const auto type = *romfs_title_id == program_id ? FileSys::ContentRecordType::Program
-                                                    : FileSys::ContentRecordType::Data;
-    const auto base_nca = installed.GetEntry(*romfs_title_id, type);
+    const auto type = static_cast<FileSys::ContentRecordType>(raw_type);
+    const auto base_nca = installed.GetEntry(title_id, type);
     if (!base_nca) {
         failed();
         return;
     }
 
+    const FileSys::NCA update_nca{packed_update_raw, nullptr};
+    if (type != FileSys::ContentRecordType::Program ||
+        update_nca.GetStatus() != Loader::ResultStatus::ErrorMissingBKTRBaseRomFS ||
+        update_nca.GetTitleId() != FileSys::GetUpdateTitleID(title_id)) {
+        packed_update_raw = {};
+    }
+
+    const auto base_romfs = base_nca->GetRomFS();
+    if (!base_romfs) {
+        failed();
+        return;
+    }
+
     const auto dump_dir =
         target == DumpRomFSTarget::Normal
             ? Common::FS::GetYuzuPath(Common::FS::YuzuPath::DumpDir)
             : Common::FS::GetYuzuPath(Common::FS::YuzuPath::SDMCDir) / "atmosphere" / "contents";
-    const auto romfs_dir = fmt::format("{:016X}/romfs", *romfs_title_id);
+    const auto romfs_dir = fmt::format("{:016X}/romfs", title_id);
 
     const auto path = Common::FS::PathToUTF8String(dump_dir / romfs_dir);
 
-    FileSys::VirtualFile romfs;
-
-    if (*romfs_title_id == program_id) {
-        const FileSys::PatchManager pm{program_id, system->GetFileSystemController(), installed};
-        romfs = pm.PatchRomFS(base_nca.get(), base_romfs, type, nullptr, false);
-    } else {
-        romfs = installed.GetEntry(*romfs_title_id, type)->GetRomFS();
-    }
-
-    const auto extracted = FileSys::ExtractRomFS(romfs, FileSys::RomFSExtractionType::Full);
-    if (extracted == nullptr) {
-        failed();
-        return;
-    }
+    const FileSys::PatchManager pm{title_id, system->GetFileSystemController(), installed};
+    auto romfs = pm.PatchRomFS(base_nca.get(), base_romfs, type, packed_update_raw, false);
 
     const auto out = VfsFilesystemCreateDirectoryWrapper(vfs, path, FileSys::Mode::ReadWrite);
 
@@ -2640,6 +2646,12 @@ void GMainWindow::OnGameListDumpRomFS(u64 program_id, const std::string& game_pa
         return;
     }
 
+    const auto extracted = FileSys::ExtractRomFS(romfs, FileSys::RomFSExtractionType::Full);
+    if (extracted == nullptr) {
+        failed();
+        return;
+    }
+
     const auto full = res == selections.constFirst();
     const auto entry_size = CalculateRomFSEntrySize(extracted, full);
 
@@ -3261,7 +3273,7 @@ void GMainWindow::OnPauseContinueGame() {
 }
 
 void GMainWindow::OnStopGame() {
-    if (system->GetExitLock() && !ConfirmForceLockedExit()) {
+    if (system->GetExitLocked() && !ConfirmForceLockedExit()) {
         return;
     }
 
@@ -4350,28 +4362,41 @@ bool GMainWindow::CheckSystemArchiveDecryption() {
     return mii_nca->GetRomFS().get() != nullptr;
 }
 
-std::optional<u64> GMainWindow::SelectRomFSDumpTarget(const FileSys::ContentProvider& installed,
-                                                      u64 program_id) {
-    const auto dlc_entries =
-        installed.ListEntriesFilter(FileSys::TitleType::AOC, FileSys::ContentRecordType::Data);
-    std::vector<FileSys::ContentProviderEntry> dlc_match;
-    dlc_match.reserve(dlc_entries.size());
-    std::copy_if(dlc_entries.begin(), dlc_entries.end(), std::back_inserter(dlc_match),
-                 [&program_id, &installed](const FileSys::ContentProviderEntry& entry) {
-                     return FileSys::GetBaseTitleID(entry.title_id) == program_id &&
-                            installed.GetEntry(entry)->GetStatus() == Loader::ResultStatus::Success;
-                 });
-
-    std::vector<u64> romfs_tids;
-    romfs_tids.push_back(program_id);
-    for (const auto& entry : dlc_match) {
-        romfs_tids.push_back(entry.title_id);
-    }
-
-    if (romfs_tids.size() > 1) {
-        QStringList list{QStringLiteral("Base")};
-        for (std::size_t i = 1; i < romfs_tids.size(); ++i) {
-            list.push_back(QStringLiteral("DLC %1").arg(romfs_tids[i] & 0x7FF));
+bool GMainWindow::SelectRomFSDumpTarget(const FileSys::ContentProvider& installed, u64 program_id,
+                                        u64* selected_title_id, u8* selected_content_record_type) {
+    using ContentInfo = std::pair<FileSys::TitleType, FileSys::ContentRecordType>;
+    boost::container::flat_map<u64, ContentInfo> available_title_ids;
+
+    const auto RetrieveEntries = [&](FileSys::TitleType title_type,
+                                     FileSys::ContentRecordType record_type) {
+        const auto entries = installed.ListEntriesFilter(title_type, record_type);
+        for (const auto& entry : entries) {
+            if (FileSys::GetBaseTitleID(entry.title_id) == program_id &&
+                installed.GetEntry(entry)->GetStatus() == Loader::ResultStatus::Success) {
+                available_title_ids[entry.title_id] = {title_type, record_type};
+            }
+        }
+    };
+
+    RetrieveEntries(FileSys::TitleType::Application, FileSys::ContentRecordType::Program);
+    RetrieveEntries(FileSys::TitleType::AOC, FileSys::ContentRecordType::Data);
+
+    if (available_title_ids.empty()) {
+        return false;
+    }
+
+    size_t title_index = 0;
+
+    if (available_title_ids.size() > 1) {
+        QStringList list;
+        for (auto& [title_id, content_info] : available_title_ids) {
+            const auto hex_title_id = QString::fromStdString(fmt::format("{:X}", title_id));
+            if (content_info.first == FileSys::TitleType::Application) {
+                list.push_back(QStringLiteral("Application [%1]").arg(hex_title_id));
+            } else {
+                list.push_back(
+                    QStringLiteral("DLC %1 [%2]").arg(title_id & 0x7FF).arg(hex_title_id));
+            }
         }
 
         bool ok;
@@ -4379,13 +4404,16 @@ std::optional<u64> GMainWindow::SelectRomFSDumpTarget(const FileSys::ContentProv
             this, tr("Select RomFS Dump Target"),
             tr("Please select which RomFS you would like to dump."), list, 0, false, &ok);
         if (!ok) {
-            return {};
+            return false;
         }
 
-        return romfs_tids[list.indexOf(res)];
+        title_index = list.indexOf(res);
     }
 
-    return program_id;
+    const auto selected_info = available_title_ids.nth(title_index);
+    *selected_title_id = selected_info->first;
+    *selected_content_record_type = static_cast<u8>(selected_info->second.second);
+    return true;
 }
 
 bool GMainWindow::ConfirmClose() {
@@ -4515,6 +4543,8 @@ void GMainWindow::RequestGameExit() {
     auto applet_ae = sm.GetService<Service::AM::AppletAE>("appletAE");
     bool has_signalled = false;
 
+    system->SetExitRequested(true);
+
     if (applet_oe != nullptr) {
         applet_oe->GetMessageQueue()->RequestExit();
         has_signalled = true;
diff --git a/src/yuzu/main.h b/src/yuzu/main.h
index 1b7055122..668dbc3b1 100644
--- a/src/yuzu/main.h
+++ b/src/yuzu/main.h
@@ -375,7 +375,8 @@ private:
     void RemoveAllTransferableShaderCaches(u64 program_id);
     void RemoveCustomConfiguration(u64 program_id, const std::string& game_path);
     void RemoveCacheStorage(u64 program_id);
-    std::optional<u64> SelectRomFSDumpTarget(const FileSys::ContentProvider&, u64 program_id);
+    bool SelectRomFSDumpTarget(const FileSys::ContentProvider&, u64 program_id,
+                               u64* selected_title_id, u8* selected_content_record_type);
     InstallResult InstallNSPXCI(const QString& filename);
     InstallResult InstallNCA(const QString& filename);
     void MigrateConfigFiles();