CMakeLists: disable -Winvalid-offsetof

This Clang warning complains when offsetof is used on a non-standard-layout type (i.e. any class using various C++ features), even though it works fine (and is not undefined behavior as of C++17).
Merge pull request #4975 from comex/invalid-syncpoint-id
2020-11-27 17:42:02 -05:00 · 2020-11-26 01:27:24 -08:00 · 2020-11-25 21:48:57 -08:00 · 2020-11-25 21:09:13 -05:00 · 2020-11-25 23:57:11 +00:00 · 2020-11-25 20:44:53 -03:00
96 changed files with 1221 additions and 652 deletions
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -66,6 +66,7 @@ else()
        -Wextra
        -Wmissing-declarations
        -Wno-attributes
+        -Wno-invalid-offsetof
        -Wno-unused-parameter
    )

--- a/src/audio_core/audio_out.cpp
+++ b/src/audio_core/audio_out.cpp
@@ -43,6 +43,10 @@ std::vector<Buffer::Tag> AudioOut::GetTagsAndReleaseBuffers(StreamPtr stream,
    return stream->GetTagsAndReleaseBuffers(max_count);
 }

+std::vector<Buffer::Tag> AudioOut::GetTagsAndReleaseBuffers(StreamPtr stream) {
+    return stream->GetTagsAndReleaseBuffers();
+}
+
 void AudioOut::StartStream(StreamPtr stream) {
    stream->Play();
 }
--- a/src/audio_core/audio_out.h
+++ b/src/audio_core/audio_out.h
@@ -31,6 +31,9 @@ public:
    /// Returns a vector of recently released buffers specified by tag for the specified stream
    std::vector<Buffer::Tag> GetTagsAndReleaseBuffers(StreamPtr stream, std::size_t max_count);

+    /// Returns a vector of all recently released buffers specified by tag for the specified stream
+    std::vector<Buffer::Tag> GetTagsAndReleaseBuffers(StreamPtr stream);
+
    /// Starts an audio stream for playback
    void StartStream(StreamPtr stream);

--- a/src/audio_core/audio_renderer.cpp
+++ b/src/audio_core/audio_renderer.cpp
@@ -2,6 +2,7 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.

+#include <limits>
 #include <vector>

 #include "audio_core/audio_out.h"
@@ -14,6 +15,59 @@
 #include "core/memory.h"
 #include "core/settings.h"

+namespace {
+[[nodiscard]] static constexpr s16 ClampToS16(s32 value) {
+    return static_cast<s16>(std::clamp(value, s32{std::numeric_limits<s16>::min()},
+                                       s32{std::numeric_limits<s16>::max()}));
+}
+
+[[nodiscard]] static constexpr s16 Mix2To1(s16 l_channel, s16 r_channel) {
+    // Mix 50% from left and 50% from right channel
+    constexpr float l_mix_amount = 50.0f / 100.0f;
+    constexpr float r_mix_amount = 50.0f / 100.0f;
+    return ClampToS16(static_cast<s32>((static_cast<float>(l_channel) * l_mix_amount) +
+                                       (static_cast<float>(r_channel) * r_mix_amount)));
+}
+
+[[nodiscard]] static constexpr std::tuple<s16, s16> Mix6To2(s16 fl_channel, s16 fr_channel,
+                                                            s16 fc_channel,
+                                                            [[maybe_unused]] s16 lf_channel,
+                                                            s16 bl_channel, s16 br_channel) {
+    // Front channels are mixed 36.94%, Center channels are mixed to be 26.12% & the back channels
+    // are mixed to be 36.94%
+
+    constexpr float front_mix_amount = 36.94f / 100.0f;
+    constexpr float center_mix_amount = 26.12f / 100.0f;
+    constexpr float back_mix_amount = 36.94f / 100.0f;
+
+    // Mix 50% from left and 50% from right channel
+    const auto left = front_mix_amount * static_cast<float>(fl_channel) +
+                      center_mix_amount * static_cast<float>(fc_channel) +
+                      back_mix_amount * static_cast<float>(bl_channel);
+
+    const auto right = front_mix_amount * static_cast<float>(fr_channel) +
+                       center_mix_amount * static_cast<float>(fc_channel) +
+                       back_mix_amount * static_cast<float>(br_channel);
+
+    return {ClampToS16(static_cast<s32>(left)), ClampToS16(static_cast<s32>(right))};
+}
+
+[[nodiscard]] static constexpr std::tuple<s16, s16> Mix6To2WithCoefficients(
+    s16 fl_channel, s16 fr_channel, s16 fc_channel, s16 lf_channel, s16 bl_channel, s16 br_channel,
+    const std::array<float_le, 4>& coeff) {
+    const auto left =
+        static_cast<float>(fl_channel) * coeff[0] + static_cast<float>(fc_channel) * coeff[1] +
+        static_cast<float>(lf_channel) * coeff[2] + static_cast<float>(bl_channel) * coeff[0];
+
+    const auto right =
+        static_cast<float>(fr_channel) * coeff[0] + static_cast<float>(fc_channel) * coeff[1] +
+        static_cast<float>(lf_channel) * coeff[2] + static_cast<float>(br_channel) * coeff[0];
+
+    return {ClampToS16(static_cast<s32>(left)), ClampToS16(static_cast<s32>(right))};
+}
+
+} // namespace
+
 namespace AudioCore {
 AudioRenderer::AudioRenderer(Core::Timing::CoreTiming& core_timing, Core::Memory::Memory& memory_,
                             AudioCommon::AudioRendererParameter params,
@@ -62,10 +116,6 @@ Stream::State AudioRenderer::GetStreamState() const {
    return stream->GetState();
 }

-static constexpr s16 ClampToS16(s32 value) {
-    return static_cast<s16>(std::clamp(value, -32768, 32767));
-}
-
 ResultCode AudioRenderer::UpdateAudioRenderer(const std::vector<u8>& input_params,
                                              std::vector<u8>& output_params) {

@@ -104,8 +154,8 @@ ResultCode AudioRenderer::UpdateAudioRenderer(const std::vector<u8>& input_param
        }
    }

-    auto mix_result = info_updater.UpdateMixes(mix_context, worker_params.mix_buffer_count,
-                                               splitter_context, effect_context);
+    const auto mix_result = info_updater.UpdateMixes(mix_context, worker_params.mix_buffer_count,
+                                                     splitter_context, effect_context);

    if (mix_result.IsError()) {
        LOG_ERROR(Audio, "Failed to update mix parameters");
@@ -194,20 +244,22 @@ void AudioRenderer::QueueMixedBuffer(Buffer::Tag tag) {
        for (std::size_t i = 0; i < BUFFER_SIZE; i++) {
            if (channel_count == 1) {
                const auto sample = ClampToS16(mix_buffers[0][i]);
-                buffer[i * stream_channel_count + 0] = sample;
-                if (stream_channel_count > 1) {
-                    buffer[i * stream_channel_count + 1] = sample;
+
+                // Place sample in all channels
+                for (u32 channel = 0; channel < stream_channel_count; channel++) {
+                    buffer[i * stream_channel_count + channel] = sample;
                }
+
                if (stream_channel_count == 6) {
-                    buffer[i * stream_channel_count + 2] = sample;
-                    buffer[i * stream_channel_count + 4] = sample;
-                    buffer[i * stream_channel_count + 5] = sample;
+                    // Output stream has a LF channel, mute it!
+                    buffer[i * stream_channel_count + 3] = 0;
                }
+
            } else if (channel_count == 2) {
                const auto l_sample = ClampToS16(mix_buffers[0][i]);
                const auto r_sample = ClampToS16(mix_buffers[1][i]);
                if (stream_channel_count == 1) {
-                    buffer[i * stream_channel_count + 0] = l_sample;
+                    buffer[i * stream_channel_count + 0] = Mix2To1(l_sample, r_sample);
                } else if (stream_channel_count == 2) {
                    buffer[i * stream_channel_count + 0] = l_sample;
                    buffer[i * stream_channel_count + 1] = r_sample;
@@ -215,8 +267,8 @@ void AudioRenderer::QueueMixedBuffer(Buffer::Tag tag) {
                    buffer[i * stream_channel_count + 0] = l_sample;
                    buffer[i * stream_channel_count + 1] = r_sample;

-                    buffer[i * stream_channel_count + 2] =
-                        ClampToS16((static_cast<s32>(l_sample) + static_cast<s32>(r_sample)) / 2);
+                    // Combine both left and right channels to the center channel
+                    buffer[i * stream_channel_count + 2] = Mix2To1(l_sample, r_sample);

                    buffer[i * stream_channel_count + 4] = l_sample;
                    buffer[i * stream_channel_count + 5] = r_sample;
@@ -231,17 +283,25 @@ void AudioRenderer::QueueMixedBuffer(Buffer::Tag tag) {
                const auto br_sample = ClampToS16(mix_buffers[5][i]);

                if (stream_channel_count == 1) {
-                    buffer[i * stream_channel_count + 0] = fc_sample;
+                    // Games seem to ignore the center channel half the time, we use the front left
+                    // and right channel for mixing as that's where majority of the audio goes
+                    buffer[i * stream_channel_count + 0] = Mix2To1(fl_sample, fr_sample);
                } else if (stream_channel_count == 2) {
-                    buffer[i * stream_channel_count + 0] =
-                        static_cast<s16>(0.3694f * static_cast<float>(fl_sample) +
-                                         0.2612f * static_cast<float>(fc_sample) +
-                                         0.3694f * static_cast<float>(bl_sample));
-                    buffer[i * stream_channel_count + 1] =
-                        static_cast<s16>(0.3694f * static_cast<float>(fr_sample) +
-                                         0.2612f * static_cast<float>(fc_sample) +
-                                         0.3694f * static_cast<float>(br_sample));
+                    // Mix all channels into 2 channels
+                    if (sink_context.HasDownMixingCoefficients()) {
+                        const auto [left, right] = Mix6To2WithCoefficients(
+                            fl_sample, fr_sample, fc_sample, lf_sample, bl_sample, br_sample,
+                            sink_context.GetDownmixCoefficients());
+                        buffer[i * stream_channel_count + 0] = left;
+                        buffer[i * stream_channel_count + 1] = right;
+                    } else {
+                        const auto [left, right] = Mix6To2(fl_sample, fr_sample, fc_sample,
+                                                           lf_sample, bl_sample, br_sample);
+                        buffer[i * stream_channel_count + 0] = left;
+                        buffer[i * stream_channel_count + 1] = right;
+                    }
                } else if (stream_channel_count == 6) {
+                    // Pass through
                    buffer[i * stream_channel_count + 0] = fl_sample;
                    buffer[i * stream_channel_count + 1] = fr_sample;
                    buffer[i * stream_channel_count + 2] = fc_sample;
@@ -259,7 +319,7 @@ void AudioRenderer::QueueMixedBuffer(Buffer::Tag tag) {
 }

 void AudioRenderer::ReleaseAndQueueBuffers() {
-    const auto released_buffers{audio_out->GetTagsAndReleaseBuffers(stream, 2)};
+    const auto released_buffers{audio_out->GetTagsAndReleaseBuffers(stream)};
    for (const auto& tag : released_buffers) {
        QueueMixedBuffer(tag);
    }
--- a/src/audio_core/audio_renderer.h
+++ b/src/audio_core/audio_renderer.h
@@ -36,16 +36,10 @@ class Memory;
 }

 namespace AudioCore {
-using DSPStateHolder = std::array<VoiceState*, 6>;
+using DSPStateHolder = std::array<VoiceState*, AudioCommon::MAX_CHANNEL_COUNT>;

 class AudioOut;

-struct RendererInfo {
-    u64_le elasped_frame_count{};
-    INSERT_PADDING_WORDS(2);
-};
-static_assert(sizeof(RendererInfo) == 0x10, "RendererInfo is an invalid size");
-
 class AudioRenderer {
 public:
    AudioRenderer(Core::Timing::CoreTiming& core_timing, Core::Memory::Memory& memory_,
@@ -53,14 +47,14 @@ public:
                  std::shared_ptr<Kernel::WritableEvent> buffer_event, std::size_t instance_number);
    ~AudioRenderer();

-    ResultCode UpdateAudioRenderer(const std::vector<u8>& input_params,
-                                   std::vector<u8>& output_params);
+    [[nodiscard]] ResultCode UpdateAudioRenderer(const std::vector<u8>& input_params,
+                                                 std::vector<u8>& output_params);
    void QueueMixedBuffer(Buffer::Tag tag);
    void ReleaseAndQueueBuffers();
-    u32 GetSampleRate() const;
-    u32 GetSampleCount() const;
-    u32 GetMixBufferCount() const;
-    Stream::State GetStreamState() const;
+    [[nodiscard]] u32 GetSampleRate() const;
+    [[nodiscard]] u32 GetSampleCount() const;
+    [[nodiscard]] u32 GetMixBufferCount() const;
+    [[nodiscard]] Stream::State GetStreamState() const;

 private:
    BehaviorInfo behavior_info{};
--- a/src/audio_core/behavior_info.h
+++ b/src/audio_core/behavior_info.h
@@ -43,22 +43,22 @@ public:
    void ClearError();
    void UpdateFlags(u64_le dest_flags);
    void SetUserRevision(u32_le revision);
-    u32_le GetUserRevision() const;
-    u32_le GetProcessRevision() const;
+    [[nodiscard]] u32_le GetUserRevision() const;
+    [[nodiscard]] u32_le GetProcessRevision() const;

-    bool IsAdpcmLoopContextBugFixed() const;
-    bool IsSplitterSupported() const;
-    bool IsLongSizePreDelaySupported() const;
-    bool IsAudioRendererProcessingTimeLimit80PercentSupported() const;
-    bool IsAudioRendererProcessingTimeLimit75PercentSupported() const;
-    bool IsAudioRendererProcessingTimeLimit70PercentSupported() const;
-    bool IsElapsedFrameCountSupported() const;
-    bool IsMemoryPoolForceMappingEnabled() const;
-    bool IsFlushVoiceWaveBuffersSupported() const;
-    bool IsVoicePlayedSampleCountResetAtLoopPointSupported() const;
-    bool IsVoicePitchAndSrcSkippedSupported() const;
-    bool IsMixInParameterDirtyOnlyUpdateSupported() const;
-    bool IsSplitterBugFixed() const;
+    [[nodiscard]] bool IsAdpcmLoopContextBugFixed() const;
+    [[nodiscard]] bool IsSplitterSupported() const;
+    [[nodiscard]] bool IsLongSizePreDelaySupported() const;
+    [[nodiscard]] bool IsAudioRendererProcessingTimeLimit80PercentSupported() const;
+    [[nodiscard]] bool IsAudioRendererProcessingTimeLimit75PercentSupported() const;
+    [[nodiscard]] bool IsAudioRendererProcessingTimeLimit70PercentSupported() const;
+    [[nodiscard]] bool IsElapsedFrameCountSupported() const;
+    [[nodiscard]] bool IsMemoryPoolForceMappingEnabled() const;
+    [[nodiscard]] bool IsFlushVoiceWaveBuffersSupported() const;
+    [[nodiscard]] bool IsVoicePlayedSampleCountResetAtLoopPointSupported() const;
+    [[nodiscard]] bool IsVoicePitchAndSrcSkippedSupported() const;
+    [[nodiscard]] bool IsMixInParameterDirtyOnlyUpdateSupported() const;
+    [[nodiscard]] bool IsSplitterBugFixed() const;
    void CopyErrorInfo(OutParams& dst);

 private:
--- a/src/audio_core/command_generator.h
+++ b/src/audio_core/command_generator.h
@@ -39,13 +39,13 @@ public:
    void PreCommand();
    void PostCommand();

-    s32* GetChannelMixBuffer(s32 channel);
-    const s32* GetChannelMixBuffer(s32 channel) const;
-    s32* GetMixBuffer(std::size_t index);
-    const s32* GetMixBuffer(std::size_t index) const;
-    std::size_t GetMixChannelBufferOffset(s32 channel) const;
+    [[nodiscard]] s32* GetChannelMixBuffer(s32 channel);
+    [[nodiscard]] const s32* GetChannelMixBuffer(s32 channel) const;
+    [[nodiscard]] s32* GetMixBuffer(std::size_t index);
+    [[nodiscard]] const s32* GetMixBuffer(std::size_t index) const;
+    [[nodiscard]] std::size_t GetMixChannelBufferOffset(s32 channel) const;

-    std::size_t GetTotalMixBufferCount() const;
+    [[nodiscard]] std::size_t GetTotalMixBufferCount() const;

 private:
    void GenerateDataSourceCommand(ServerVoiceInfo& voice_info, VoiceState& dsp_state, s32 channel);
@@ -73,7 +73,7 @@ private:
    void GenerateI3dl2ReverbEffectCommand(s32 mix_buffer_offset, EffectBase* info, bool enabled);
    void GenerateBiquadFilterEffectCommand(s32 mix_buffer_offset, EffectBase* info, bool enabled);
    void GenerateAuxCommand(s32 mix_buffer_offset, EffectBase* info, bool enabled);
-    ServerSplitterDestinationData* GetDestinationData(s32 splitter_id, s32 index);
+    [[nodiscard]] ServerSplitterDestinationData* GetDestinationData(s32 splitter_id, s32 index);

    s32 WriteAuxBuffer(AuxInfoDSP& dsp_info, VAddr send_buffer, u32 max_samples, const s32* data,
                       u32 sample_count, u32 write_offset, u32 write_count);
--- a/src/audio_core/common.h
+++ b/src/audio_core/common.h
@@ -22,7 +22,7 @@ constexpr std::size_t MAX_CHANNEL_COUNT = 6;
 constexpr std::size_t MAX_WAVE_BUFFERS = 4;
 constexpr std::size_t MAX_SAMPLE_HISTORY = 4;
 constexpr u32 STREAM_SAMPLE_RATE = 48000;
-constexpr u32 STREAM_NUM_CHANNELS = 6;
+constexpr u32 STREAM_NUM_CHANNELS = 2;
 constexpr s32 NO_SPLITTER = -1;
 constexpr s32 NO_MIX = 0x7fffffff;
 constexpr s32 NO_FINAL_MIX = std::numeric_limits<s32>::min();
--- a/src/audio_core/effect_context.h
+++ b/src/audio_core/effect_context.h
@@ -189,11 +189,11 @@ public:

    virtual void Update(EffectInfo::InParams& in_params) = 0;
    virtual void UpdateForCommandGeneration() = 0;
-    UsageState GetUsage() const;
-    EffectType GetType() const;
-    bool IsEnabled() const;
-    s32 GetMixID() const;
-    s32 GetProcessingOrder() const;
+    [[nodiscard]] UsageState GetUsage() const;
+    [[nodiscard]] EffectType GetType() const;
+    [[nodiscard]] bool IsEnabled() const;
+    [[nodiscard]] s32 GetMixID() const;
+    [[nodiscard]] s32 GetProcessingOrder() const;

 protected:
    UsageState usage{UsageState::Invalid};
@@ -257,10 +257,10 @@ public:

    void Update(EffectInfo::InParams& in_params) override;
    void UpdateForCommandGeneration() override;
-    VAddr GetSendInfo() const;
-    VAddr GetSendBuffer() const;
-    VAddr GetRecvInfo() const;
-    VAddr GetRecvBuffer() const;
+    [[nodiscard]] VAddr GetSendInfo() const;
+    [[nodiscard]] VAddr GetSendBuffer() const;
+    [[nodiscard]] VAddr GetRecvInfo() const;
+    [[nodiscard]] VAddr GetRecvBuffer() const;

 private:
    VAddr send_info{};
@@ -309,10 +309,10 @@ public:
    explicit EffectContext(std::size_t effect_count);
    ~EffectContext();

-    std::size_t GetCount() const;
-    EffectBase* GetInfo(std::size_t i);
-    EffectBase* RetargetEffect(std::size_t i, EffectType effect);
-    const EffectBase* GetInfo(std::size_t i) const;
+    [[nodiscard]] std::size_t GetCount() const;
+    [[nodiscard]] EffectBase* GetInfo(std::size_t i);
+    [[nodiscard]] EffectBase* RetargetEffect(std::size_t i, EffectType effect);
+    [[nodiscard]] const EffectBase* GetInfo(std::size_t i) const;

 private:
    std::size_t effect_count{};
--- a/src/audio_core/mix_context.h
+++ b/src/audio_core/mix_context.h
@@ -62,17 +62,17 @@ public:
    ServerMixInfo();
    ~ServerMixInfo();

-    const ServerMixInfo::InParams& GetInParams() const;
-    ServerMixInfo::InParams& GetInParams();
+    [[nodiscard]] const ServerMixInfo::InParams& GetInParams() const;
+    [[nodiscard]] ServerMixInfo::InParams& GetInParams();

    bool Update(EdgeMatrix& edge_matrix, const MixInfo::InParams& mix_in,
                BehaviorInfo& behavior_info, SplitterContext& splitter_context,
                EffectContext& effect_context);
-    bool HasAnyConnection() const;
+    [[nodiscard]] bool HasAnyConnection() const;
    void Cleanup();
    void SetEffectCount(std::size_t count);
    void ResetEffectProcessingOrder();
-    s32 GetEffectOrder(std::size_t i) const;
+    [[nodiscard]] s32 GetEffectOrder(std::size_t i) const;

 private:
    std::vector<s32> effect_processing_order;
@@ -91,15 +91,15 @@ public:
    void SortInfo();
    bool TsortInfo(SplitterContext& splitter_context);

-    std::size_t GetCount() const;
-    ServerMixInfo& GetInfo(std::size_t i);
-    const ServerMixInfo& GetInfo(std::size_t i) const;
-    ServerMixInfo& GetSortedInfo(std::size_t i);
-    const ServerMixInfo& GetSortedInfo(std::size_t i) const;
-    ServerMixInfo& GetFinalMixInfo();
-    const ServerMixInfo& GetFinalMixInfo() const;
-    EdgeMatrix& GetEdgeMatrix();
-    const EdgeMatrix& GetEdgeMatrix() const;
+    [[nodiscard]] std::size_t GetCount() const;
+    [[nodiscard]] ServerMixInfo& GetInfo(std::size_t i);
+    [[nodiscard]] const ServerMixInfo& GetInfo(std::size_t i) const;
+    [[nodiscard]] ServerMixInfo& GetSortedInfo(std::size_t i);
+    [[nodiscard]] const ServerMixInfo& GetSortedInfo(std::size_t i) const;
+    [[nodiscard]] ServerMixInfo& GetFinalMixInfo();
+    [[nodiscard]] const ServerMixInfo& GetFinalMixInfo() const;
+    [[nodiscard]] EdgeMatrix& GetEdgeMatrix();
+    [[nodiscard]] const EdgeMatrix& GetEdgeMatrix() const;

 private:
    void CalcMixBufferOffset();
--- a/src/audio_core/sink_context.cpp
+++ b/src/audio_core/sink_context.cpp
@@ -12,10 +12,16 @@ std::size_t SinkContext::GetCount() const {
    return sink_count;
 }

-void SinkContext::UpdateMainSink(SinkInfo::InParams& in) {
+void SinkContext::UpdateMainSink(const SinkInfo::InParams& in) {
+    ASSERT(in.type == SinkTypes::Device);
+
+    has_downmix_coefs = in.device.down_matrix_enabled;
+    if (has_downmix_coefs) {
+        downmix_coefficients = in.device.down_matrix_coef;
+    }
    in_use = in.in_use;
    use_count = in.device.input_count;
-    std::memcpy(buffers.data(), in.device.input.data(), AudioCommon::MAX_CHANNEL_COUNT);
+    buffers = in.device.input;
 }

 bool SinkContext::InUse() const {
@@ -28,4 +34,12 @@ std::vector<u8> SinkContext::OutputBuffers() const {
    return buffer_ret;
 }

+bool SinkContext::HasDownMixingCoefficients() const {
+    return has_downmix_coefs;
+}
+
+const DownmixCoefficients& SinkContext::GetDownmixCoefficients() const {
+    return downmix_coefficients;
+}
+
 } // namespace AudioCore
--- a/src/audio_core/sink_context.h
+++ b/src/audio_core/sink_context.h
@@ -11,6 +11,8 @@

 namespace AudioCore {

+using DownmixCoefficients = std::array<float_le, 4>;
+
 enum class SinkTypes : u8 {
    Invalid = 0,
    Device = 1,
@@ -50,7 +52,7 @@ public:
        std::array<u8, AudioCommon::MAX_CHANNEL_COUNT> input;
        INSERT_UNION_PADDING_BYTES(1);
        bool down_matrix_enabled;
-        std::array<float_le, 4> down_matrix_coef;
+        DownmixCoefficients down_matrix_coef;
    };
    static_assert(sizeof(SinkInfo::DeviceIn) == 0x11c, "SinkInfo::DeviceIn is an invalid size");

@@ -74,16 +76,21 @@ public:
    explicit SinkContext(std::size_t sink_count);
    ~SinkContext();

-    std::size_t GetCount() const;
+    [[nodiscard]] std::size_t GetCount() const;

-    void UpdateMainSink(SinkInfo::InParams& in);
-    bool InUse() const;
-    std::vector<u8> OutputBuffers() const;
+    void UpdateMainSink(const SinkInfo::InParams& in);
+    [[nodiscard]] bool InUse() const;
+    [[nodiscard]] std::vector<u8> OutputBuffers() const;
+
+    [[nodiscard]] bool HasDownMixingCoefficients() const;
+    [[nodiscard]] const DownmixCoefficients& GetDownmixCoefficients() const;

 private:
    bool in_use{false};
    s32 use_count{};
    std::array<u8, AudioCommon::MAX_CHANNEL_COUNT> buffers{};
    std::size_t sink_count{};
+    bool has_downmix_coefs{false};
+    DownmixCoefficients downmix_coefficients{};
 };
 } // namespace AudioCore
--- a/src/audio_core/stream.cpp
+++ b/src/audio_core/stream.cpp
@@ -136,4 +136,14 @@ std::vector<Buffer::Tag> Stream::GetTagsAndReleaseBuffers(std::size_t max_count)
    return tags;
 }

+std::vector<Buffer::Tag> Stream::GetTagsAndReleaseBuffers() {
+    std::vector<Buffer::Tag> tags;
+    tags.reserve(released_buffers.size());
+    while (!released_buffers.empty()) {
+        tags.push_back(released_buffers.front()->GetTag());
+        released_buffers.pop();
+    }
+    return tags;
+}
+
 } // namespace AudioCore
--- a/src/audio_core/stream.h
+++ b/src/audio_core/stream.h
@@ -57,37 +57,40 @@ public:
    bool QueueBuffer(BufferPtr&& buffer);

    /// Returns true if the audio stream contains a buffer with the specified tag
-    bool ContainsBuffer(Buffer::Tag tag) const;
+    [[nodiscard]] bool ContainsBuffer(Buffer::Tag tag) const;

    /// Returns a vector of recently released buffers specified by tag
-    std::vector<Buffer::Tag> GetTagsAndReleaseBuffers(std::size_t max_count);
+    [[nodiscard]] std::vector<Buffer::Tag> GetTagsAndReleaseBuffers(std::size_t max_count);
+
+    /// Returns a vector of all recently released buffers specified by tag
+    [[nodiscard]] std::vector<Buffer::Tag> GetTagsAndReleaseBuffers();

    void SetVolume(float volume);

-    float GetVolume() const {
+    [[nodiscard]] float GetVolume() const {
        return game_volume;
    }

    /// Returns true if the stream is currently playing
-    bool IsPlaying() const {
+    [[nodiscard]] bool IsPlaying() const {
        return state == State::Playing;
    }

    /// Returns the number of queued buffers
-    std::size_t GetQueueSize() const {
+    [[nodiscard]] std::size_t GetQueueSize() const {
        return queued_buffers.size();
    }

    /// Gets the sample rate
-    u32 GetSampleRate() const {
+    [[nodiscard]] u32 GetSampleRate() const {
        return sample_rate;
    }

    /// Gets the number of channels
-    u32 GetNumChannels() const;
+    [[nodiscard]] u32 GetNumChannels() const;

    /// Get the state
-    State GetState() const;
+    [[nodiscard]] State GetState() const;

 private:
    /// Plays the next queued buffer in the audio stream, starting playback if necessary
@@ -97,7 +100,7 @@ private:
    void ReleaseActiveBuffer(std::chrono::nanoseconds ns_late = {});

    /// Gets the number of core cycles when the specified buffer will be released
-    std::chrono::nanoseconds GetBufferReleaseNS(const Buffer& buffer) const;
+    [[nodiscard]] std::chrono::nanoseconds GetBufferReleaseNS(const Buffer& buffer) const;

    u32 sample_rate;                  ///< Sample rate of the stream
    Format format;                    ///< Format of the stream
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -145,7 +145,7 @@ struct System::Impl {
    }

    ResultStatus Init(System& system, Frontend::EmuWindow& emu_window) {
-        LOG_DEBUG(HW_Memory, "initialized OK");
+        LOG_DEBUG(Core, "initialized OK");

        device_memory = std::make_unique<Core::DeviceMemory>();

@@ -187,7 +187,7 @@ struct System::Impl {

        service_manager = std::make_shared<Service::SM::ServiceManager>(kernel);

-        Service::Init(service_manager, system);
+        services = std::make_unique<Service::Services>(service_manager, system);
        GDBStub::DeferStart();

        interrupt_manager = std::make_unique<Core::Hardware::InterruptManager>(system);
@@ -208,9 +208,11 @@ struct System::Impl {
        return ResultStatus::Success;
    }

-    ResultStatus Load(System& system, Frontend::EmuWindow& emu_window,
-                      const std::string& filepath) {
-        app_loader = Loader::GetLoader(system, GetGameFileFromPath(virtual_filesystem, filepath));
+    ResultStatus Load(System& system, Frontend::EmuWindow& emu_window, const std::string& filepath,
+                      std::size_t program_index) {
+        app_loader = Loader::GetLoader(system, GetGameFileFromPath(virtual_filesystem, filepath),
+                                       program_index);
+
        if (!app_loader) {
            LOG_CRITICAL(Core, "Failed to obtain loader for {}!", filepath);
            return ResultStatus::ErrorGetLoader;
@@ -296,7 +298,7 @@ struct System::Impl {

        // Shutdown emulation session
        GDBStub::Shutdown();
-        Service::Shutdown();
+        services.reset();
        service_manager.reset();
        cheat_engine.reset();
        telemetry_session.reset();
@@ -306,8 +308,8 @@ struct System::Impl {
        cpu_manager.Shutdown();

        // Shutdown kernel and core timing
-        kernel.Shutdown();
        core_timing.Shutdown();
+        kernel.Shutdown();

        // Close app loader
        app_loader.reset();
@@ -398,6 +400,9 @@ struct System::Impl {
    /// Service manager
    std::shared_ptr<Service::SM::ServiceManager> service_manager;

+    /// Services
+    std::unique_ptr<Service::Services> services;
+
    /// Telemetry session for this emulation session
    std::unique_ptr<Core::TelemetrySession> telemetry_session;

@@ -413,6 +418,8 @@ struct System::Impl {
    bool is_multicore{};
    bool is_async_gpu{};

+    ExecuteProgramCallback execute_program_callback;
+
    std::array<u64, Core::Hardware::NUM_CPU_CORES> dynarmic_ticks{};
    std::array<MicroProfileToken, Core::Hardware::NUM_CPU_CORES> microprofile_dynarmic{};
 };
@@ -444,8 +451,13 @@ void System::InvalidateCpuInstructionCaches() {
    impl->kernel.InvalidateAllInstructionCaches();
 }

-System::ResultStatus System::Load(Frontend::EmuWindow& emu_window, const std::string& filepath) {
-    return impl->Load(*this, emu_window, filepath);
+void System::Shutdown() {
+    impl->Shutdown();
+}
+
+System::ResultStatus System::Load(Frontend::EmuWindow& emu_window, const std::string& filepath,
+                                  std::size_t program_index) {
+    return impl->Load(*this, emu_window, filepath, program_index);
 }

 bool System::IsPoweredOn() const {
@@ -632,7 +644,11 @@ const std::string& System::GetStatusDetails() const {
    return impl->status_details;
 }

-Loader::AppLoader& System::GetAppLoader() const {
+Loader::AppLoader& System::GetAppLoader() {
+    return *impl->app_loader;
+}
+
+const Loader::AppLoader& System::GetAppLoader() const {
    return *impl->app_loader;
 }

@@ -748,14 +764,6 @@ const System::CurrentBuildProcessID& System::GetCurrentProcessBuildID() const {
    return impl->build_id;
 }

-System::ResultStatus System::Init(Frontend::EmuWindow& emu_window) {
-    return impl->Init(*this, emu_window);
-}
-
-void System::Shutdown() {
-    impl->Shutdown();
-}
-
 Service::SM::ServiceManager& System::ServiceManager() {
    return *impl->service_manager;
 }
@@ -786,4 +794,16 @@ bool System::IsMulticore() const {
    return impl->is_multicore;
 }

+void System::RegisterExecuteProgramCallback(ExecuteProgramCallback&& callback) {
+    impl->execute_program_callback = std::move(callback);
+}
+
+void System::ExecuteProgram(std::size_t program_index) {
+    if (impl->execute_program_callback) {
+        impl->execute_program_callback(program_index);
+    } else {
+        LOG_CRITICAL(Core, "execute_program_callback must be initialized by the frontend");
+    }
+}
+
 } // namespace Core
--- a/src/core/core.h
+++ b/src/core/core.h
@@ -5,6 +5,7 @@
 #pragma once

 #include <cstddef>
+#include <functional>
 #include <memory>
 #include <string>
 #include <vector>
@@ -144,19 +145,19 @@ public:
     * Run the OS and Application
     * This function will start emulation and run the relevant devices
     */
-    ResultStatus Run();
+    [[nodiscard]] ResultStatus Run();

    /**
     * Pause the OS and Application
     * This function will pause emulation and stop the relevant devices
     */
-    ResultStatus Pause();
+    [[nodiscard]] ResultStatus Pause();

    /**
     * Step the CPU one instruction
     * @return Result status, indicating whether or not the operation succeeded.
     */
-    ResultStatus SingleStep();
+    [[nodiscard]] ResultStatus SingleStep();

    /**
     * Invalidate the CPU instruction caches
@@ -173,22 +174,24 @@ public:
     * @param emu_window Reference to the host-system window used for video output and keyboard
     *                   input.
     * @param filepath String path to the executable application to load on the host file system.
+     * @param program_index Specifies the index within the container of the program to launch.
     * @returns ResultStatus code, indicating if the operation succeeded.
     */
-    ResultStatus Load(Frontend::EmuWindow& emu_window, const std::string& filepath);
+    [[nodiscard]] ResultStatus Load(Frontend::EmuWindow& emu_window, const std::string& filepath,
+                                    std::size_t program_index = 0);

    /**
     * Indicates if the emulated system is powered on (all subsystems initialized and able to run an
     * application).
     * @returns True if the emulated system is powered on, otherwise false.
     */
-    bool IsPoweredOn() const;
+    [[nodiscard]] bool IsPoweredOn() const;

    /// Gets a reference to the telemetry session for this emulation session.
-    Core::TelemetrySession& TelemetrySession();
+    [[nodiscard]] Core::TelemetrySession& TelemetrySession();

    /// Gets a reference to the telemetry session for this emulation session.
-    const Core::TelemetrySession& TelemetrySession() const;
+    [[nodiscard]] const Core::TelemetrySession& TelemetrySession() const;

    /// Prepare the core emulation for a reschedule
    void PrepareReschedule();
@@ -197,185 +200,178 @@ public:
    void PrepareReschedule(u32 core_index);

    /// Gets and resets core performance statistics
-    PerfStatsResults GetAndResetPerfStats();
+    [[nodiscard]] PerfStatsResults GetAndResetPerfStats();

    /// Gets an ARM interface to the CPU core that is currently running
-    ARM_Interface& CurrentArmInterface();
+    [[nodiscard]] ARM_Interface& CurrentArmInterface();

    /// Gets an ARM interface to the CPU core that is currently running
-    const ARM_Interface& CurrentArmInterface() const;
+    [[nodiscard]] const ARM_Interface& CurrentArmInterface() const;

    /// Gets the index of the currently running CPU core
-    std::size_t CurrentCoreIndex() const;
+    [[nodiscard]] std::size_t CurrentCoreIndex() const;

    /// Gets the scheduler for the CPU core that is currently running
-    Kernel::Scheduler& CurrentScheduler();
+    [[nodiscard]] Kernel::Scheduler& CurrentScheduler();

    /// Gets the scheduler for the CPU core that is currently running
-    const Kernel::Scheduler& CurrentScheduler() const;
+    [[nodiscard]] const Kernel::Scheduler& CurrentScheduler() const;

    /// Gets the physical core for the CPU core that is currently running
-    Kernel::PhysicalCore& CurrentPhysicalCore();
+    [[nodiscard]] Kernel::PhysicalCore& CurrentPhysicalCore();

    /// Gets the physical core for the CPU core that is currently running
-    const Kernel::PhysicalCore& CurrentPhysicalCore() const;
+    [[nodiscard]] const Kernel::PhysicalCore& CurrentPhysicalCore() const;

    /// Gets a reference to an ARM interface for the CPU core with the specified index
-    ARM_Interface& ArmInterface(std::size_t core_index);
+    [[nodiscard]] ARM_Interface& ArmInterface(std::size_t core_index);

    /// Gets a const reference to an ARM interface from the CPU core with the specified index
-    const ARM_Interface& ArmInterface(std::size_t core_index) const;
+    [[nodiscard]] const ARM_Interface& ArmInterface(std::size_t core_index) const;

-    CpuManager& GetCpuManager();
+    /// Gets a reference to the underlying CPU manager.
+    [[nodiscard]] CpuManager& GetCpuManager();

-    const CpuManager& GetCpuManager() const;
+    /// Gets a const reference to the underlying CPU manager
+    [[nodiscard]] const CpuManager& GetCpuManager() const;

    /// Gets a reference to the exclusive monitor
-    ExclusiveMonitor& Monitor();
+    [[nodiscard]] ExclusiveMonitor& Monitor();

    /// Gets a constant reference to the exclusive monitor
-    const ExclusiveMonitor& Monitor() const;
+    [[nodiscard]] const ExclusiveMonitor& Monitor() const;

    /// Gets a mutable reference to the system memory instance.
-    Core::Memory::Memory& Memory();
+    [[nodiscard]] Core::Memory::Memory& Memory();

    /// Gets a constant reference to the system memory instance.
-    const Core::Memory::Memory& Memory() const;
+    [[nodiscard]] const Core::Memory::Memory& Memory() const;

    /// Gets a mutable reference to the GPU interface
-    Tegra::GPU& GPU();
+    [[nodiscard]] Tegra::GPU& GPU();

    /// Gets an immutable reference to the GPU interface.
-    const Tegra::GPU& GPU() const;
+    [[nodiscard]] const Tegra::GPU& GPU() const;

    /// Gets a mutable reference to the renderer.
-    VideoCore::RendererBase& Renderer();
+    [[nodiscard]] VideoCore::RendererBase& Renderer();

    /// Gets an immutable reference to the renderer.
-    const VideoCore::RendererBase& Renderer() const;
+    [[nodiscard]] const VideoCore::RendererBase& Renderer() const;

    /// Gets the scheduler for the CPU core with the specified index
-    Kernel::Scheduler& Scheduler(std::size_t core_index);
+    [[nodiscard]] Kernel::Scheduler& Scheduler(std::size_t core_index);

    /// Gets the scheduler for the CPU core with the specified index
-    const Kernel::Scheduler& Scheduler(std::size_t core_index) const;
+    [[nodiscard]] const Kernel::Scheduler& Scheduler(std::size_t core_index) const;

    /// Gets the global scheduler
-    Kernel::GlobalScheduler& GlobalScheduler();
+    [[nodiscard]] Kernel::GlobalScheduler& GlobalScheduler();

    /// Gets the global scheduler
-    const Kernel::GlobalScheduler& GlobalScheduler() const;
+    [[nodiscard]] const Kernel::GlobalScheduler& GlobalScheduler() const;

    /// Gets the manager for the guest device memory
-    Core::DeviceMemory& DeviceMemory();
+    [[nodiscard]] Core::DeviceMemory& DeviceMemory();

    /// Gets the manager for the guest device memory
-    const Core::DeviceMemory& DeviceMemory() const;
+    [[nodiscard]] const Core::DeviceMemory& DeviceMemory() const;

    /// Provides a pointer to the current process
-    Kernel::Process* CurrentProcess();
+    [[nodiscard]] Kernel::Process* CurrentProcess();

    /// Provides a constant pointer to the current process.
-    const Kernel::Process* CurrentProcess() const;
+    [[nodiscard]] const Kernel::Process* CurrentProcess() const;

    /// Provides a reference to the core timing instance.
-    Timing::CoreTiming& CoreTiming();
+    [[nodiscard]] Timing::CoreTiming& CoreTiming();

    /// Provides a constant reference to the core timing instance.
-    const Timing::CoreTiming& CoreTiming() const;
+    [[nodiscard]] const Timing::CoreTiming& CoreTiming() const;

    /// Provides a reference to the interrupt manager instance.
-    Core::Hardware::InterruptManager& InterruptManager();
+    [[nodiscard]] Core::Hardware::InterruptManager& InterruptManager();

    /// Provides a constant reference to the interrupt manager instance.
-    const Core::Hardware::InterruptManager& InterruptManager() const;
+    [[nodiscard]] const Core::Hardware::InterruptManager& InterruptManager() const;

    /// Provides a reference to the kernel instance.
-    Kernel::KernelCore& Kernel();
+    [[nodiscard]] Kernel::KernelCore& Kernel();

    /// Provides a constant reference to the kernel instance.
-    const Kernel::KernelCore& Kernel() const;
+    [[nodiscard]] const Kernel::KernelCore& Kernel() const;

    /// Provides a reference to the internal PerfStats instance.
-    Core::PerfStats& GetPerfStats();
+    [[nodiscard]] Core::PerfStats& GetPerfStats();

    /// Provides a constant reference to the internal PerfStats instance.
-    const Core::PerfStats& GetPerfStats() const;
+    [[nodiscard]] const Core::PerfStats& GetPerfStats() const;

    /// Provides a reference to the frame limiter;
-    Core::FrameLimiter& FrameLimiter();
+    [[nodiscard]] Core::FrameLimiter& FrameLimiter();

    /// Provides a constant referent to the frame limiter
-    const Core::FrameLimiter& FrameLimiter() const;
+    [[nodiscard]] const Core::FrameLimiter& FrameLimiter() const;

    /// Gets the name of the current game
-    Loader::ResultStatus GetGameName(std::string& out) const;
+    [[nodiscard]] Loader::ResultStatus GetGameName(std::string& out) const;

    void SetStatus(ResultStatus new_status, const char* details);

-    const std::string& GetStatusDetails() const;
+    [[nodiscard]] const std::string& GetStatusDetails() const;

-    Loader::AppLoader& GetAppLoader() const;
+    [[nodiscard]] Loader::AppLoader& GetAppLoader();
+    [[nodiscard]] const Loader::AppLoader& GetAppLoader() const;

-    Service::SM::ServiceManager& ServiceManager();
-    const Service::SM::ServiceManager& ServiceManager() const;
+    [[nodiscard]] Service::SM::ServiceManager& ServiceManager();
+    [[nodiscard]] const Service::SM::ServiceManager& ServiceManager() const;

    void SetFilesystem(FileSys::VirtualFilesystem vfs);

-    FileSys::VirtualFilesystem GetFilesystem() const;
+    [[nodiscard]] FileSys::VirtualFilesystem GetFilesystem() const;

    void RegisterCheatList(const std::vector<Memory::CheatEntry>& list,
                           const std::array<u8, 0x20>& build_id, VAddr main_region_begin,
                           u64 main_region_size);

    void SetAppletFrontendSet(Service::AM::Applets::AppletFrontendSet&& set);
-
    void SetDefaultAppletFrontendSet();

-    Service::AM::Applets::AppletManager& GetAppletManager();
-
-    const Service::AM::Applets::AppletManager& GetAppletManager() const;
+    [[nodiscard]] Service::AM::Applets::AppletManager& GetAppletManager();
+    [[nodiscard]] const Service::AM::Applets::AppletManager& GetAppletManager() const;

    void SetContentProvider(std::unique_ptr<FileSys::ContentProviderUnion> provider);

-    FileSys::ContentProvider& GetContentProvider();
+    [[nodiscard]] FileSys::ContentProvider& GetContentProvider();
+    [[nodiscard]] const FileSys::ContentProvider& GetContentProvider() const;

-    const FileSys::ContentProvider& GetContentProvider() const;
-
-    Service::FileSystem::FileSystemController& GetFileSystemController();
-
-    const Service::FileSystem::FileSystemController& GetFileSystemController() const;
+    [[nodiscard]] Service::FileSystem::FileSystemController& GetFileSystemController();
+    [[nodiscard]] const Service::FileSystem::FileSystemController& GetFileSystemController() const;

    void RegisterContentProvider(FileSys::ContentProviderUnionSlot slot,
                                 FileSys::ContentProvider* provider);

    void ClearContentProvider(FileSys::ContentProviderUnionSlot slot);

-    const Reporter& GetReporter() const;
+    [[nodiscard]] const Reporter& GetReporter() const;

-    Service::Glue::ARPManager& GetARPManager();
+    [[nodiscard]] Service::Glue::ARPManager& GetARPManager();
+    [[nodiscard]] const Service::Glue::ARPManager& GetARPManager() const;

-    const Service::Glue::ARPManager& GetARPManager() const;
+    [[nodiscard]] Service::APM::Controller& GetAPMController();
+    [[nodiscard]] const Service::APM::Controller& GetAPMController() const;

-    Service::APM::Controller& GetAPMController();
+    [[nodiscard]] Service::LM::Manager& GetLogManager();
+    [[nodiscard]] const Service::LM::Manager& GetLogManager() const;

-    const Service::APM::Controller& GetAPMController() const;
-
-    Service::LM::Manager& GetLogManager();
-
-    const Service::LM::Manager& GetLogManager() const;
-
-    Service::Time::TimeManager& GetTimeManager();
-
-    const Service::Time::TimeManager& GetTimeManager() const;
+    [[nodiscard]] Service::Time::TimeManager& GetTimeManager();
+    [[nodiscard]] const Service::Time::TimeManager& GetTimeManager() const;

    void SetExitLock(bool locked);
-
-    bool GetExitLock() const;
+    [[nodiscard]] bool GetExitLock() const;

    void SetCurrentProcessBuildID(const CurrentBuildProcessID& id);
-
-    const CurrentBuildProcessID& GetCurrentProcessBuildID() const;
+    [[nodiscard]] const CurrentBuildProcessID& GetCurrentProcessBuildID() const;

    /// Register a host thread as an emulated CPU Core.
    void RegisterCoreThread(std::size_t id);
@@ -390,19 +386,28 @@ public:
    void ExitDynarmicProfile();

    /// Tells if system is running on multicore.
-    bool IsMulticore() const;
+    [[nodiscard]] bool IsMulticore() const;
+
+    /// Type used for the frontend to designate a callback for System to re-launch the application
+    /// using a specified program index.
+    using ExecuteProgramCallback = std::function<void(std::size_t)>;
+
+    /**
+     * Registers a callback from the frontend for System to re-launch the application using a
+     * specified program index.
+     * @param callback Callback from the frontend to relaunch the application.
+     */
+    void RegisterExecuteProgramCallback(ExecuteProgramCallback&& callback);
+
+    /**
+     * Instructs the frontend to re-launch the application using the specified program_index.
+     * @param program_index Specifies the index within the application of the program to launch.
+     */
+    void ExecuteProgram(std::size_t program_index);

 private:
    System();

-    /**
-     * Initialize the emulated system.
-     * @param emu_window Reference to the host-system window used for video output and keyboard
-     *                   input.
-     * @return ResultStatus code, indicating if the operation succeeded.
-     */
-    ResultStatus Init(Frontend::EmuWindow& emu_window);
-
    struct Impl;
    std::unique_ptr<Impl> impl;

--- a/src/core/cpu_manager.cpp
+++ b/src/core/cpu_manager.cpp
@@ -4,6 +4,7 @@

 #include "common/fiber.h"
 #include "common/microprofile.h"
+#include "common/scope_exit.h"
 #include "common/thread.h"
 #include "core/arm/exclusive_monitor.h"
 #include "core/core.h"
@@ -343,6 +344,16 @@ void CpuManager::RunThread(std::size_t core) {
    data.initialized = true;
    const bool sc_sync = !is_async_gpu && !is_multicore;
    bool sc_sync_first_use = sc_sync;
+
+    // Cleanup
+    SCOPE_EXIT({
+        data.host_context->Exit();
+        data.enter_barrier.reset();
+        data.exit_barrier.reset();
+        data.initialized = false;
+        MicroProfileOnThreadExit();
+    });
+
    /// Running
    while (running_mode) {
        data.is_running = false;
@@ -351,6 +362,12 @@ void CpuManager::RunThread(std::size_t core) {
            system.GPU().ObtainContext();
            sc_sync_first_use = false;
        }
+
+        // Abort if emulation was killed before the session really starts
+        if (!system.IsPoweredOn()) {
+            return;
+        }
+
        auto& scheduler = system.Kernel().CurrentScheduler();
        Kernel::Thread* current_thread = scheduler.GetCurrentThread();
        data.is_running = true;
@@ -360,13 +377,6 @@ void CpuManager::RunThread(std::size_t core) {
        data.exit_barrier->Wait();
        data.is_paused = false;
    }
-    /// Time to cleanup
-    data.host_context->Exit();
-    data.enter_barrier.reset();
-    data.exit_barrier.reset();
-    data.initialized = false;
-
-    MicroProfileOnThreadExit();
 }

 } // namespace Core
--- a/src/core/file_sys/card_image.cpp
+++ b/src/core/file_sys/card_image.cpp
@@ -29,7 +29,7 @@ constexpr std::array partition_names{
    "logo",
 };

-XCI::XCI(VirtualFile file_)
+XCI::XCI(VirtualFile file_, std::size_t program_index)
    : file(std::move(file_)), program_nca_status{Loader::ResultStatus::ErrorXCIMissingProgramNCA},
      partitions(partition_names.size()),
      partitions_raw(partition_names.size()), keys{Core::Crypto::KeyManager::Instance()} {
@@ -62,7 +62,8 @@ XCI::XCI(VirtualFile file_)
    }

    secure_partition = std::make_shared<NSP>(
-        main_hfs.GetFile(partition_names[static_cast<std::size_t>(XCIPartition::Secure)]));
+        main_hfs.GetFile(partition_names[static_cast<std::size_t>(XCIPartition::Secure)]),
+        program_index);

    ncas = secure_partition->GetNCAsCollapsed();
    program =
--- a/src/core/file_sys/card_image.h
+++ b/src/core/file_sys/card_image.h
@@ -78,7 +78,7 @@ enum class XCIPartition : u8 { Update, Normal, Secure, Logo };

 class XCI : public ReadOnlyVfsDirectory {
 public:
-    explicit XCI(VirtualFile file);
+    explicit XCI(VirtualFile file, std::size_t program_index = 0);
    ~XCI() override;

    Loader::ResultStatus GetStatus() const;
--- a/src/core/file_sys/submission_package.cpp
+++ b/src/core/file_sys/submission_package.cpp
@@ -20,8 +20,8 @@

 namespace FileSys {

-NSP::NSP(VirtualFile file_)
-    : file(std::move(file_)), status{Loader::ResultStatus::Success},
+NSP::NSP(VirtualFile file_, std::size_t program_index)
+    : file(std::move(file_)), program_index(program_index), status{Loader::ResultStatus::Success},
      pfs(std::make_shared<PartitionFilesystem>(file)), keys{Core::Crypto::KeyManager::Instance()} {
    if (pfs->GetStatus() != Loader::ResultStatus::Success) {
        status = pfs->GetStatus();
@@ -146,7 +146,7 @@ std::shared_ptr<NCA> NSP::GetNCA(u64 title_id, ContentRecordType type, TitleType
    if (extracted)
        LOG_WARNING(Service_FS, "called on an NSP that is of type extracted.");

-    const auto title_id_iter = ncas.find(title_id);
+    const auto title_id_iter = ncas.find(title_id + program_index);
    if (title_id_iter == ncas.end())
        return nullptr;

--- a/src/core/file_sys/submission_package.h
+++ b/src/core/file_sys/submission_package.h
@@ -27,7 +27,7 @@ enum class ContentRecordType : u8;

 class NSP : public ReadOnlyVfsDirectory {
 public:
-    explicit NSP(VirtualFile file);
+    explicit NSP(VirtualFile file, std::size_t program_index = 0);
    ~NSP() override;

    Loader::ResultStatus GetStatus() const;
@@ -69,6 +69,8 @@ private:

    VirtualFile file;

+    const std::size_t program_index;
+
    bool extracted = false;
    Loader::ResultStatus status;
    std::map<u64, Loader::ResultStatus> program_status;
--- a/src/core/frontend/emu_window.h
+++ b/src/core/frontend/emu_window.h
@@ -102,8 +102,8 @@ public:
        float render_surface_scale = 1.0f;
    };

-    /// Polls window events
-    virtual void PollEvents() = 0;
+    /// Called from GPU thread when a frame is displayed.
+    virtual void OnFrameDisplayed() {}

    /**
     * Returns a GraphicsContext that the frontend provides to be used for rendering.
--- a/src/core/hle/kernel/svc.cpp
+++ b/src/core/hle/kernel/svc.cpp
@@ -681,7 +681,7 @@ static void Break32(Core::System& system, u32 reason, u32 info1, u32 info2) {
 }

 /// Used to output a message on a debug hardware unit - does nothing on a retail unit
-static void OutputDebugString([[maybe_unused]] Core::System& system, VAddr address, u64 len) {
+static void OutputDebugString(Core::System& system, VAddr address, u64 len) {
    if (len == 0) {
        return;
    }
--- a/src/core/hle/service/am/am.cpp
+++ b/src/core/hle/service/am/am.cpp
@@ -246,9 +246,8 @@ IDebugFunctions::IDebugFunctions() : ServiceFramework{"IDebugFunctions"} {

 IDebugFunctions::~IDebugFunctions() = default;

-ISelfController::ISelfController(Core::System& system,
-                                 std::shared_ptr<NVFlinger::NVFlinger> nvflinger)
-    : ServiceFramework("ISelfController"), system(system), nvflinger(std::move(nvflinger)) {
+ISelfController::ISelfController(Core::System& system, NVFlinger::NVFlinger& nvflinger)
+    : ServiceFramework("ISelfController"), system(system), nvflinger(nvflinger) {
    // clang-format off
    static const FunctionInfo functions[] = {
        {0, &ISelfController::Exit, "Exit"},
@@ -458,8 +457,8 @@ void ISelfController::CreateManagedDisplayLayer(Kernel::HLERequestContext& ctx)

    // TODO(Subv): Find out how AM determines the display to use, for now just
    // create the layer in the Default display.
-    const auto display_id = nvflinger->OpenDisplay("Default");
-    const auto layer_id = nvflinger->CreateLayer(*display_id);
+    const auto display_id = nvflinger.OpenDisplay("Default");
+    const auto layer_id = nvflinger.CreateLayer(*display_id);

    IPC::ResponseBuilder rb{ctx, 4};
    rb.Push(RESULT_SUCCESS);
@@ -476,8 +475,8 @@ void ISelfController::CreateManagedDisplaySeparableLayer(Kernel::HLERequestConte
    // Outputting 1 layer id instead of the expected 2 has not been observed to cause any adverse
    // side effects.
    // TODO: Support multiple layers
-    const auto display_id = nvflinger->OpenDisplay("Default");
-    const auto layer_id = nvflinger->CreateLayer(*display_id);
+    const auto display_id = nvflinger.OpenDisplay("Default");
+    const auto layer_id = nvflinger.CreateLayer(*display_id);

    IPC::ResponseBuilder rb{ctx, 4};
    rb.Push(RESULT_SUCCESS);
@@ -1189,9 +1188,9 @@ IApplicationFunctions::IApplicationFunctions(Core::System& system_)
        {102, &IApplicationFunctions::SetApplicationCopyrightVisibility, "SetApplicationCopyrightVisibility"},
        {110, &IApplicationFunctions::QueryApplicationPlayStatistics, "QueryApplicationPlayStatistics"},
        {111, &IApplicationFunctions::QueryApplicationPlayStatisticsByUid, "QueryApplicationPlayStatisticsByUid"},
-        {120, nullptr, "ExecuteProgram"},
-        {121, nullptr, "ClearUserChannel"},
-        {122, nullptr, "UnpopToUserChannel"},
+        {120, &IApplicationFunctions::ExecuteProgram, "ExecuteProgram"},
+        {121, &IApplicationFunctions::ClearUserChannel, "ClearUserChannel"},
+        {122, &IApplicationFunctions::UnpopToUserChannel, "UnpopToUserChannel"},
        {123, &IApplicationFunctions::GetPreviousProgramIndex, "GetPreviousProgramIndex"},
        {124, nullptr, "EnableApplicationAllThreadDumpOnCrash"},
        {130, &IApplicationFunctions::GetGpuErrorDetectedSystemEvent, "GetGpuErrorDetectedSystemEvent"},
@@ -1562,6 +1561,34 @@ void IApplicationFunctions::QueryApplicationPlayStatisticsByUid(Kernel::HLEReque
    rb.Push<u32>(0);
 }

+void IApplicationFunctions::ExecuteProgram(Kernel::HLERequestContext& ctx) {
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+
+    IPC::RequestParser rp{ctx};
+    [[maybe_unused]] const auto unk_1 = rp.Pop<u32>();
+    [[maybe_unused]] const auto unk_2 = rp.Pop<u32>();
+    const auto program_index = rp.Pop<u64>();
+
+    IPC::ResponseBuilder rb{ctx, 2};
+    rb.Push(RESULT_SUCCESS);
+
+    system.ExecuteProgram(program_index);
+}
+
+void IApplicationFunctions::ClearUserChannel(Kernel::HLERequestContext& ctx) {
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+
+    IPC::ResponseBuilder rb{ctx, 2};
+    rb.Push(RESULT_SUCCESS);
+}
+
+void IApplicationFunctions::UnpopToUserChannel(Kernel::HLERequestContext& ctx) {
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+
+    IPC::ResponseBuilder rb{ctx, 2};
+    rb.Push(RESULT_SUCCESS);
+}
+
 void IApplicationFunctions::GetPreviousProgramIndex(Kernel::HLERequestContext& ctx) {
    LOG_WARNING(Service_AM, "(STUBBED) called");

@@ -1586,8 +1613,8 @@ void IApplicationFunctions::GetFriendInvitationStorageChannelEvent(Kernel::HLERe
    rb.PushCopyObjects(friend_invitation_storage_channel_event.readable);
 }

-void InstallInterfaces(SM::ServiceManager& service_manager,
-                       std::shared_ptr<NVFlinger::NVFlinger> nvflinger, Core::System& system) {
+void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nvflinger,
+                       Core::System& system) {
    auto message_queue = std::make_shared<AppletMessageQueue>(system.Kernel());
    // Needed on game boot
    message_queue->PushMessage(AppletMessageQueue::AppletMessage::FocusStateChanged);
--- a/src/core/hle/service/am/am.h
+++ b/src/core/hle/service/am/am.h
@@ -121,8 +121,7 @@ public:

 class ISelfController final : public ServiceFramework<ISelfController> {
 public:
-    explicit ISelfController(Core::System& system_,
-                             std::shared_ptr<NVFlinger::NVFlinger> nvflinger_);
+    explicit ISelfController(Core::System& system_, NVFlinger::NVFlinger& nvflinger_);
    ~ISelfController() override;

 private:
@@ -156,7 +155,7 @@ private:
    };

    Core::System& system;
-    std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
+    NVFlinger::NVFlinger& nvflinger;
    Kernel::EventPair launchable_event;
    Kernel::EventPair accumulated_suspended_tick_changed_event;

@@ -288,6 +287,9 @@ private:
    void SetApplicationCopyrightVisibility(Kernel::HLERequestContext& ctx);
    void QueryApplicationPlayStatistics(Kernel::HLERequestContext& ctx);
    void QueryApplicationPlayStatisticsByUid(Kernel::HLERequestContext& ctx);
+    void ExecuteProgram(Kernel::HLERequestContext& ctx);
+    void ClearUserChannel(Kernel::HLERequestContext& ctx);
+    void UnpopToUserChannel(Kernel::HLERequestContext& ctx);
    void GetPreviousProgramIndex(Kernel::HLERequestContext& ctx);
    void GetGpuErrorDetectedSystemEvent(Kernel::HLERequestContext& ctx);
    void GetFriendInvitationStorageChannelEvent(Kernel::HLERequestContext& ctx);
@@ -332,7 +334,7 @@ public:
 };

 /// Registers all AM services with the specified service manager.
-void InstallInterfaces(SM::ServiceManager& service_manager,
-                       std::shared_ptr<NVFlinger::NVFlinger> nvflinger, Core::System& system);
+void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nvflinger,
+                       Core::System& system);

 } // namespace Service::AM
--- a/src/core/hle/service/am/applet_ae.cpp
+++ b/src/core/hle/service/am/applet_ae.cpp
@@ -13,10 +13,10 @@ namespace Service::AM {

 class ILibraryAppletProxy final : public ServiceFramework<ILibraryAppletProxy> {
 public:
-    explicit ILibraryAppletProxy(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
+    explicit ILibraryAppletProxy(NVFlinger::NVFlinger& nvflinger,
                                 std::shared_ptr<AppletMessageQueue> msg_queue,
                                 Core::System& system)
-        : ServiceFramework("ILibraryAppletProxy"), nvflinger(std::move(nvflinger)),
+        : ServiceFramework("ILibraryAppletProxy"), nvflinger(nvflinger),
          msg_queue(std::move(msg_queue)), system(system) {
        // clang-format off
        static const FunctionInfo functions[] = {
@@ -109,16 +109,16 @@ private:
        rb.PushIpcInterface<IApplicationFunctions>(system);
    }

-    std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
+    NVFlinger::NVFlinger& nvflinger;
    std::shared_ptr<AppletMessageQueue> msg_queue;
    Core::System& system;
 };

 class ISystemAppletProxy final : public ServiceFramework<ISystemAppletProxy> {
 public:
-    explicit ISystemAppletProxy(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
+    explicit ISystemAppletProxy(NVFlinger::NVFlinger& nvflinger,
                                std::shared_ptr<AppletMessageQueue> msg_queue, Core::System& system)
-        : ServiceFramework("ISystemAppletProxy"), nvflinger(std::move(nvflinger)),
+        : ServiceFramework("ISystemAppletProxy"), nvflinger(nvflinger),
          msg_queue(std::move(msg_queue)), system(system) {
        // clang-format off
        static const FunctionInfo functions[] = {
@@ -220,7 +220,8 @@ private:
        rb.Push(RESULT_SUCCESS);
        rb.PushIpcInterface<IApplicationCreator>();
    }
-    std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
+
+    NVFlinger::NVFlinger& nvflinger;
    std::shared_ptr<AppletMessageQueue> msg_queue;
    Core::System& system;
 };
@@ -249,10 +250,10 @@ void AppletAE::OpenLibraryAppletProxyOld(Kernel::HLERequestContext& ctx) {
    rb.PushIpcInterface<ILibraryAppletProxy>(nvflinger, msg_queue, system);
 }

-AppletAE::AppletAE(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
-                   std::shared_ptr<AppletMessageQueue> msg_queue, Core::System& system)
-    : ServiceFramework("appletAE"), nvflinger(std::move(nvflinger)),
-      msg_queue(std::move(msg_queue)), system(system) {
+AppletAE::AppletAE(NVFlinger::NVFlinger& nvflinger, std::shared_ptr<AppletMessageQueue> msg_queue,
+                   Core::System& system)
+    : ServiceFramework("appletAE"), nvflinger(nvflinger), msg_queue(std::move(msg_queue)),
+      system(system) {
    // clang-format off
    static const FunctionInfo functions[] = {
        {100, &AppletAE::OpenSystemAppletProxy, "OpenSystemAppletProxy"},
--- a/src/core/hle/service/am/applet_ae.h
+++ b/src/core/hle/service/am/applet_ae.h
@@ -23,7 +23,7 @@ class AppletMessageQueue;

 class AppletAE final : public ServiceFramework<AppletAE> {
 public:
-    explicit AppletAE(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
+    explicit AppletAE(NVFlinger::NVFlinger& nvflinger,
                      std::shared_ptr<AppletMessageQueue> msg_queue, Core::System& system);
    ~AppletAE() override;

@@ -34,7 +34,7 @@ private:
    void OpenLibraryAppletProxy(Kernel::HLERequestContext& ctx);
    void OpenLibraryAppletProxyOld(Kernel::HLERequestContext& ctx);

-    std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
+    NVFlinger::NVFlinger& nvflinger;
    std::shared_ptr<AppletMessageQueue> msg_queue;
    Core::System& system;
 };
--- a/src/core/hle/service/am/applet_oe.cpp
+++ b/src/core/hle/service/am/applet_oe.cpp
@@ -12,9 +12,9 @@ namespace Service::AM {

 class IApplicationProxy final : public ServiceFramework<IApplicationProxy> {
 public:
-    explicit IApplicationProxy(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
+    explicit IApplicationProxy(NVFlinger::NVFlinger& nvflinger,
                               std::shared_ptr<AppletMessageQueue> msg_queue, Core::System& system)
-        : ServiceFramework("IApplicationProxy"), nvflinger(std::move(nvflinger)),
+        : ServiceFramework("IApplicationProxy"), nvflinger(nvflinger),
          msg_queue(std::move(msg_queue)), system(system) {
        // clang-format off
        static const FunctionInfo functions[] = {
@@ -98,7 +98,7 @@ private:
        rb.PushIpcInterface<IApplicationFunctions>(system);
    }

-    std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
+    NVFlinger::NVFlinger& nvflinger;
    std::shared_ptr<AppletMessageQueue> msg_queue;
    Core::System& system;
 };
@@ -111,10 +111,10 @@ void AppletOE::OpenApplicationProxy(Kernel::HLERequestContext& ctx) {
    rb.PushIpcInterface<IApplicationProxy>(nvflinger, msg_queue, system);
 }

-AppletOE::AppletOE(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
-                   std::shared_ptr<AppletMessageQueue> msg_queue, Core::System& system)
-    : ServiceFramework("appletOE"), nvflinger(std::move(nvflinger)),
-      msg_queue(std::move(msg_queue)), system(system) {
+AppletOE::AppletOE(NVFlinger::NVFlinger& nvflinger, std::shared_ptr<AppletMessageQueue> msg_queue,
+                   Core::System& system)
+    : ServiceFramework("appletOE"), nvflinger(nvflinger), msg_queue(std::move(msg_queue)),
+      system(system) {
    static const FunctionInfo functions[] = {
        {0, &AppletOE::OpenApplicationProxy, "OpenApplicationProxy"},
    };
--- a/src/core/hle/service/am/applet_oe.h
+++ b/src/core/hle/service/am/applet_oe.h
@@ -23,7 +23,7 @@ class AppletMessageQueue;

 class AppletOE final : public ServiceFramework<AppletOE> {
 public:
-    explicit AppletOE(std::shared_ptr<NVFlinger::NVFlinger> nvflinger,
+    explicit AppletOE(NVFlinger::NVFlinger& nvflinger,
                      std::shared_ptr<AppletMessageQueue> msg_queue, Core::System& system);
    ~AppletOE() override;

@@ -32,7 +32,7 @@ public:
 private:
    void OpenApplicationProxy(Kernel::HLERequestContext& ctx);

-    std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
+    NVFlinger::NVFlinger& nvflinger;
    std::shared_ptr<AppletMessageQueue> msg_queue;
    Core::System& system;
 };
--- a/src/core/hle/service/nvdrv/devices/nvdevice.h
+++ b/src/core/hle/service/nvdrv/devices/nvdevice.h
@@ -31,8 +31,8 @@ public:
     * @param output A buffer where the output data will be written to.
     * @returns The result code of the ioctl.
     */
-    virtual NvResult Ioctl1(Ioctl command, const std::vector<u8>& input,
-                            std::vector<u8>& output) = 0;
+    virtual NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) = 0;

    /**
     * Handles an ioctl2 request.
@@ -43,7 +43,8 @@ public:
     * @returns The result code of the ioctl.
     */
    virtual NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                            const std::vector<u8>& inline_input, std::vector<u8>& output) = 0;
+                            const std::vector<u8>& inline_input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) = 0;

    /**
     * Handles an ioctl3 request.
@@ -54,7 +55,7 @@ public:
     * @returns The result code of the ioctl.
     */
    virtual NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                            std::vector<u8>& inline_output) = 0;
+                            std::vector<u8>& inline_output, IoctlCtrl& ctrl) = 0;

 protected:
    Core::System& system;
--- a/src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp
@@ -18,20 +18,21 @@ nvdisp_disp0::nvdisp_disp0(Core::System& system, std::shared_ptr<nvmap> nvmap_de
    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
 nvdisp_disp0 ::~nvdisp_disp0() = default;

-NvResult nvdisp_disp0::Ioctl1(Ioctl command, const std::vector<u8>& input,
-                              std::vector<u8>& output) {
+NvResult nvdisp_disp0::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                              IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvdisp_disp0::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                              const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                              const std::vector<u8>& inline_input, std::vector<u8>& output,
+                              IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvdisp_disp0::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                              std::vector<u8>& inline_output) {
+                              std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
--- a/src/core/hle/service/nvdrv/devices/nvdisp_disp0.h
+++ b/src/core/hle/service/nvdrv/devices/nvdisp_disp0.h
@@ -20,11 +20,13 @@ public:
    explicit nvdisp_disp0(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
    ~nvdisp_disp0() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

    /// Performs a screen flip, drawing the buffer pointed to by the handle.
    void flip(u32 buffer_handle, u32 offset, u32 format, u32 width, u32 height, u32 stride,
--- a/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
@@ -21,8 +21,8 @@ nvhost_as_gpu::nvhost_as_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_
    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
 nvhost_as_gpu::~nvhost_as_gpu() = default;

-NvResult nvhost_as_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input,
-                               std::vector<u8>& output) {
+NvResult nvhost_as_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                               IoctlCtrl& ctrl) {
    switch (command.group) {
    case 'A':
        switch (command.cmd) {
@@ -55,13 +55,14 @@ NvResult nvhost_as_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input,
 }

 NvResult nvhost_as_gpu::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                               const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                               const std::vector<u8>& inline_input, std::vector<u8>& output,
+                               IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvhost_as_gpu::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                               std::vector<u8>& inline_output) {
+                               std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    switch (command.group) {
    case 'A':
        switch (command.cmd) {
--- a/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.h
@@ -30,11 +30,13 @@ public:
    explicit nvhost_as_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
    ~nvhost_as_gpu() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

 private:
    class BufferMap final {
--- a/src/core/hle/service/nvdrv/devices/nvhost_ctrl.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_ctrl.cpp
@@ -20,7 +20,8 @@ nvhost_ctrl::nvhost_ctrl(Core::System& system, EventInterface& events_interface,
    : nvdevice(system), events_interface{events_interface}, syncpoint_manager{syncpoint_manager} {}
 nvhost_ctrl::~nvhost_ctrl() = default;

-NvResult nvhost_ctrl::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+NvResult nvhost_ctrl::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                             IoctlCtrl& ctrl) {
    switch (command.group) {
    case 0x0:
        switch (command.cmd) {
@@ -29,9 +30,9 @@ NvResult nvhost_ctrl::Ioctl1(Ioctl command, const std::vector<u8>& input, std::v
        case 0x1c:
            return IocCtrlClearEventWait(input, output);
        case 0x1d:
-            return IocCtrlEventWait(input, output, false);
+            return IocCtrlEventWait(input, output, false, ctrl);
        case 0x1e:
-            return IocCtrlEventWait(input, output, true);
+            return IocCtrlEventWait(input, output, true, ctrl);
        case 0x1f:
            return IocCtrlEventRegister(input, output);
        case 0x20:
@@ -47,13 +48,14 @@ NvResult nvhost_ctrl::Ioctl1(Ioctl command, const std::vector<u8>& input, std::v
 }

 NvResult nvhost_ctrl::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                             const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                             const std::vector<u8>& inline_input, std::vector<u8>& output,
+                             IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvhost_ctrl::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                             std::vector<u8>& inline_output) {
+                             std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
@@ -67,7 +69,7 @@ NvResult nvhost_ctrl::NvOsGetConfigU32(const std::vector<u8>& input, std::vector
 }

 NvResult nvhost_ctrl::IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output,
-                                       bool is_async) {
+                                       bool is_async, IoctlCtrl& ctrl) {
    IocCtrlEventWaitParams params{};
    std::memcpy(&params, input.data(), sizeof(params));
    LOG_DEBUG(Service_NVDRV, "syncpt_id={}, threshold={}, timeout={}, is_async={}",
@@ -139,7 +141,10 @@ NvResult nvhost_ctrl::IocCtrlEventWait(const std::vector<u8>& input, std::vector
        params.value |= event_id;
        event.event.writable->Clear();
        gpu.RegisterSyncptInterrupt(params.syncpt_id, target_value);
-        if (!is_async) {
+        if (!is_async && ctrl.fresh_call) {
+            ctrl.must_delay = true;
+            ctrl.timeout = params.timeout;
+            ctrl.event_id = event_id;
            return NvResult::Timeout;
        }
        std::memcpy(output.data(), &params, sizeof(params));
--- a/src/core/hle/service/nvdrv/devices/nvhost_ctrl.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_ctrl.h
@@ -18,11 +18,13 @@ public:
                         SyncpointManager& syncpoint_manager);
    ~nvhost_ctrl() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

 private:
    struct IocSyncptReadParams {
@@ -121,7 +123,8 @@ private:
    static_assert(sizeof(IocCtrlEventKill) == 8, "IocCtrlEventKill is incorrect size");

    NvResult NvOsGetConfigU32(const std::vector<u8>& input, std::vector<u8>& output);
-    NvResult IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output, bool is_async);
+    NvResult IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output, bool is_async,
+                              IoctlCtrl& ctrl);
    NvResult IocCtrlEventRegister(const std::vector<u8>& input, std::vector<u8>& output);
    NvResult IocCtrlEventUnregister(const std::vector<u8>& input, std::vector<u8>& output);
    NvResult IocCtrlClearEventWait(const std::vector<u8>& input, std::vector<u8>& output);
--- a/src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.cpp
@@ -16,7 +16,7 @@ nvhost_ctrl_gpu::nvhost_ctrl_gpu(Core::System& system) : nvdevice(system) {}
 nvhost_ctrl_gpu::~nvhost_ctrl_gpu() = default;

 NvResult nvhost_ctrl_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input,
-                                 std::vector<u8>& output) {
+                                 std::vector<u8>& output, IoctlCtrl& ctrl) {
    switch (command.group) {
    case 'G':
        switch (command.cmd) {
@@ -48,13 +48,15 @@ NvResult nvhost_ctrl_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input,
 }

 NvResult nvhost_ctrl_gpu::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                                 const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                                 const std::vector<u8>& inline_input, std::vector<u8>& output,
+                                 IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvhost_ctrl_gpu::Ioctl3(Ioctl command, const std::vector<u8>& input,
-                                 std::vector<u8>& output, std::vector<u8>& inline_output) {
+                                 std::vector<u8>& output, std::vector<u8>& inline_output,
+                                 IoctlCtrl& ctrl) {
    switch (command.group) {
    case 'G':
        switch (command.cmd) {
@@ -162,7 +164,7 @@ NvResult nvhost_ctrl_gpu::GetCharacteristics(const std::vector<u8>& input, std::
    params.gpu_characteristics_buf_size = 0xA0;
    params.gpu_characteristics_buf_addr = 0xdeadbeef; // Cannot be 0 (UNUSED)

-    std::memcpy(output.data(), input.data(), output.size());
+    std::memcpy(output.data(), &params, output.size());
    std::memcpy(inline_output.data(), &params.gc, inline_output.size());
    return NvResult::Success;
 }
--- a/src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.h
@@ -16,11 +16,13 @@ public:
    explicit nvhost_ctrl_gpu(Core::System& system);
    ~nvhost_ctrl_gpu() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

 private:
    struct IoctlGpuCharacteristics {
--- a/src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp
@@ -23,7 +23,8 @@ nvhost_gpu::nvhost_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,

 nvhost_gpu::~nvhost_gpu() = default;

-NvResult nvhost_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+NvResult nvhost_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) {
    switch (command.group) {
    case 0x0:
        switch (command.cmd) {
@@ -75,7 +76,8 @@ NvResult nvhost_gpu::Ioctl1(Ioctl command, const std::vector<u8>& input, std::ve
 };

 NvResult nvhost_gpu::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                            const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                            const std::vector<u8>& inline_input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) {
    switch (command.group) {
    case 'H':
        switch (command.cmd) {
@@ -89,7 +91,7 @@ NvResult nvhost_gpu::Ioctl2(Ioctl command, const std::vector<u8>& input,
 }

 NvResult nvhost_gpu::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                            std::vector<u8>& inline_output) {
+                            std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
--- a/src/core/hle/service/nvdrv/devices/nvhost_gpu.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_gpu.h
@@ -26,11 +26,13 @@ public:
                        SyncpointManager& syncpoint_manager);
    ~nvhost_gpu() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

 private:
    enum class CtxObjects : u32_le {
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp
@@ -15,8 +15,8 @@ nvhost_nvdec::nvhost_nvdec(Core::System& system, std::shared_ptr<nvmap> nvmap_de
    : nvhost_nvdec_common(system, std::move(nvmap_dev)) {}
 nvhost_nvdec::~nvhost_nvdec() = default;

-NvResult nvhost_nvdec::Ioctl1(Ioctl command, const std::vector<u8>& input,
-                              std::vector<u8>& output) {
+NvResult nvhost_nvdec::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                              IoctlCtrl& ctrl) {
    switch (command.group) {
    case 0x0:
        switch (command.cmd) {
@@ -58,13 +58,14 @@ NvResult nvhost_nvdec::Ioctl1(Ioctl command, const std::vector<u8>& input,
 }

 NvResult nvhost_nvdec::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                              const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                              const std::vector<u8>& inline_input, std::vector<u8>& output,
+                              IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvhost_nvdec::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                              std::vector<u8>& inline_output) {
+                              std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec.h
@@ -14,11 +14,13 @@ public:
    explicit nvhost_nvdec(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
    ~nvhost_nvdec() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;
 };

 } // namespace Service::Nvidia::Devices
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h
@@ -25,8 +25,8 @@ public:
     * @param output A buffer where the output data will be written to.
     * @returns The result code of the ioctl.
     */
-    virtual NvResult Ioctl1(Ioctl command, const std::vector<u8>& input,
-                            std::vector<u8>& output) = 0;
+    virtual NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) = 0;

    /**
     * Handles an ioctl2 request.
@@ -37,7 +37,8 @@ public:
     * @returns The result code of the ioctl.
     */
    virtual NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                            const std::vector<u8>& inline_input, std::vector<u8>& output) = 0;
+                            const std::vector<u8>& inline_input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) = 0;

    /**
     * Handles an ioctl3 request.
@@ -48,7 +49,7 @@ public:
     * @returns The result code of the ioctl.
     */
    virtual NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                            std::vector<u8>& inline_output) = 0;
+                            std::vector<u8>& inline_output, IoctlCtrl& ctrl) = 0;

 protected:
    class BufferMap final {
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvjpg.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_nvjpg.cpp
@@ -13,8 +13,8 @@ namespace Service::Nvidia::Devices {
 nvhost_nvjpg::nvhost_nvjpg(Core::System& system) : nvdevice(system) {}
 nvhost_nvjpg::~nvhost_nvjpg() = default;

-NvResult nvhost_nvjpg::Ioctl1(Ioctl command, const std::vector<u8>& input,
-                              std::vector<u8>& output) {
+NvResult nvhost_nvjpg::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                              IoctlCtrl& ctrl) {
    switch (command.group) {
    case 'H':
        switch (command.cmd) {
@@ -33,13 +33,14 @@ NvResult nvhost_nvjpg::Ioctl1(Ioctl command, const std::vector<u8>& input,
 }

 NvResult nvhost_nvjpg::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                              const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                              const std::vector<u8>& inline_input, std::vector<u8>& output,
+                              IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvhost_nvjpg::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                              std::vector<u8>& inline_output) {
+                              std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvjpg.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_nvjpg.h
@@ -16,11 +16,13 @@ public:
    explicit nvhost_nvjpg(Core::System& system);
    ~nvhost_nvjpg() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

 private:
    struct IoctlSetNvmapFD {
--- a/src/core/hle/service/nvdrv/devices/nvhost_vic.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_vic.cpp
@@ -15,7 +15,8 @@ nvhost_vic::nvhost_vic(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)

 nvhost_vic::~nvhost_vic() = default;

-NvResult nvhost_vic::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+NvResult nvhost_vic::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) {
    switch (command.group) {
    case 0x0:
        switch (command.cmd) {
@@ -50,13 +51,14 @@ NvResult nvhost_vic::Ioctl1(Ioctl command, const std::vector<u8>& input, std::ve
 }

 NvResult nvhost_vic::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                            const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                            const std::vector<u8>& inline_input, std::vector<u8>& output,
+                            IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvhost_vic::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                            std::vector<u8>& inline_output) {
+                            std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
--- a/src/core/hle/service/nvdrv/devices/nvhost_vic.h
+++ b/src/core/hle/service/nvdrv/devices/nvhost_vic.h
@@ -14,10 +14,12 @@ public:
    explicit nvhost_vic(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
    ~nvhost_vic();

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;
 };
 } // namespace Service::Nvidia::Devices
--- a/src/core/hle/service/nvdrv/devices/nvmap.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvmap.cpp
@@ -19,7 +19,8 @@ nvmap::nvmap(Core::System& system) : nvdevice(system) {

 nvmap::~nvmap() = default;

-NvResult nvmap::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+NvResult nvmap::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                       IoctlCtrl& ctrl) {
    switch (command.group) {
    case 0x1:
        switch (command.cmd) {
@@ -48,13 +49,14 @@ NvResult nvmap::Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<
 }

 NvResult nvmap::Ioctl2(Ioctl command, const std::vector<u8>& input,
-                       const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                       const std::vector<u8>& inline_input, std::vector<u8>& output,
+                       IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }

 NvResult nvmap::Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                       std::vector<u8>& inline_output) {
+                       std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
    return NvResult::NotImplemented;
 }
--- a/src/core/hle/service/nvdrv/devices/nvmap.h
+++ b/src/core/hle/service/nvdrv/devices/nvmap.h
@@ -19,11 +19,13 @@ public:
    explicit nvmap(Core::System& system);
    ~nvmap() override;

-    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl2(Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output) override;
+                    const std::vector<u8>& inline_input, std::vector<u8>& output,
+                    IoctlCtrl& ctrl) override;
    NvResult Ioctl3(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::vector<u8>& inline_output, IoctlCtrl& ctrl) override;

    /// Returns the allocated address of an nvmap object given its handle.
    VAddr GetObjectAddress(u32 handle) const;
--- a/src/core/hle/service/nvdrv/interface.cpp
+++ b/src/core/hle/service/nvdrv/interface.cpp
@@ -61,10 +61,32 @@ void NVDRV::Ioctl1(Kernel::HLERequestContext& ctx) {
    std::vector<u8> output_buffer(ctx.GetWriteBufferSize(0));
    const auto input_buffer = ctx.ReadBuffer(0);

-    const auto nv_result = nvdrv->Ioctl1(fd, command, input_buffer, output_buffer);
+    IoctlCtrl ctrl{};

-    if (command.is_out != 0) {
-        ctx.WriteBuffer(output_buffer);
+    const auto nv_result = nvdrv->Ioctl1(fd, command, input_buffer, output_buffer, ctrl);
+    if (ctrl.must_delay) {
+        ctrl.fresh_call = false;
+        ctx.SleepClientThread(
+            "NVServices::DelayedResponse", ctrl.timeout,
+            [=, this](std::shared_ptr<Kernel::Thread> thread, Kernel::HLERequestContext& ctx_,
+                      Kernel::ThreadWakeupReason reason) {
+                IoctlCtrl ctrl2{ctrl};
+                std::vector<u8> tmp_output = output_buffer;
+                const auto nv_result2 = nvdrv->Ioctl1(fd, command, input_buffer, tmp_output, ctrl2);
+
+                if (command.is_out != 0) {
+                    ctx.WriteBuffer(tmp_output);
+                }
+
+                IPC::ResponseBuilder rb{ctx_, 3};
+                rb.Push(RESULT_SUCCESS);
+                rb.PushEnum(nv_result2);
+            },
+            nvdrv->GetEventWriteable(ctrl.event_id));
+    } else {
+        if (command.is_out != 0) {
+            ctx.WriteBuffer(output_buffer);
+        }
    }

    IPC::ResponseBuilder rb{ctx, 3};
@@ -88,8 +110,35 @@ void NVDRV::Ioctl2(Kernel::HLERequestContext& ctx) {
    const auto input_inlined_buffer = ctx.ReadBuffer(1);
    std::vector<u8> output_buffer(ctx.GetWriteBufferSize(0));

+    IoctlCtrl ctrl{};
+
    const auto nv_result =
-        nvdrv->Ioctl2(fd, command, input_buffer, input_inlined_buffer, output_buffer);
+        nvdrv->Ioctl2(fd, command, input_buffer, input_inlined_buffer, output_buffer, ctrl);
+    if (ctrl.must_delay) {
+        ctrl.fresh_call = false;
+        ctx.SleepClientThread(
+            "NVServices::DelayedResponse", ctrl.timeout,
+            [=, this](std::shared_ptr<Kernel::Thread> thread, Kernel::HLERequestContext& ctx_,
+                      Kernel::ThreadWakeupReason reason) {
+                IoctlCtrl ctrl2{ctrl};
+                std::vector<u8> tmp_output = output_buffer;
+                const auto nv_result2 = nvdrv->Ioctl2(fd, command, input_buffer,
+                                                      input_inlined_buffer, tmp_output, ctrl2);
+
+                if (command.is_out != 0) {
+                    ctx.WriteBuffer(tmp_output);
+                }
+
+                IPC::ResponseBuilder rb{ctx_, 3};
+                rb.Push(RESULT_SUCCESS);
+                rb.PushEnum(nv_result2);
+            },
+            nvdrv->GetEventWriteable(ctrl.event_id));
+    } else {
+        if (command.is_out != 0) {
+            ctx.WriteBuffer(output_buffer);
+        }
+    }

    if (command.is_out != 0) {
        ctx.WriteBuffer(output_buffer);
@@ -116,12 +165,36 @@ void NVDRV::Ioctl3(Kernel::HLERequestContext& ctx) {
    std::vector<u8> output_buffer(ctx.GetWriteBufferSize(0));
    std::vector<u8> output_buffer_inline(ctx.GetWriteBufferSize(1));

+    IoctlCtrl ctrl{};
    const auto nv_result =
-        nvdrv->Ioctl3(fd, command, input_buffer, output_buffer, output_buffer_inline);
+        nvdrv->Ioctl3(fd, command, input_buffer, output_buffer, output_buffer_inline, ctrl);
+    if (ctrl.must_delay) {
+        ctrl.fresh_call = false;
+        ctx.SleepClientThread(
+            "NVServices::DelayedResponse", ctrl.timeout,
+            [=, this](std::shared_ptr<Kernel::Thread> thread, Kernel::HLERequestContext& ctx_,
+                      Kernel::ThreadWakeupReason reason) {
+                IoctlCtrl ctrl2{ctrl};
+                std::vector<u8> tmp_output = output_buffer;
+                std::vector<u8> tmp_output2 = output_buffer;
+                const auto nv_result2 =
+                    nvdrv->Ioctl3(fd, command, input_buffer, tmp_output, tmp_output2, ctrl2);

-    if (command.is_out != 0) {
-        ctx.WriteBuffer(output_buffer, 0);
-        ctx.WriteBuffer(output_buffer_inline, 1);
+                if (command.is_out != 0) {
+                    ctx.WriteBuffer(tmp_output, 0);
+                    ctx.WriteBuffer(tmp_output2, 1);
+                }
+
+                IPC::ResponseBuilder rb{ctx_, 3};
+                rb.Push(RESULT_SUCCESS);
+                rb.PushEnum(nv_result2);
+            },
+            nvdrv->GetEventWriteable(ctrl.event_id));
+    } else {
+        if (command.is_out != 0) {
+            ctx.WriteBuffer(output_buffer, 0);
+            ctx.WriteBuffer(output_buffer_inline, 1);
+        }
    }

    IPC::ResponseBuilder rb{ctx, 3};
--- a/src/core/hle/service/nvdrv/nvdata.h
+++ b/src/core/hle/service/nvdrv/nvdata.h
@@ -97,4 +97,15 @@ union Ioctl {
    BitField<31, 1, u32> is_out;
 };

+struct IoctlCtrl {
+    // First call done to the servioce for services that call itself again after a call.
+    bool fresh_call{true};
+    // Tells the Ioctl Wrapper that it must delay the IPC response and send the thread to sleep
+    bool must_delay{};
+    // Timeout for the delay
+    s64 timeout{};
+    // NV Event Id
+    s32 event_id{-1};
+};
+
 } // namespace Service::Nvidia
--- a/src/core/hle/service/nvdrv/nvdrv.cpp
+++ b/src/core/hle/service/nvdrv/nvdrv.cpp
@@ -91,7 +91,7 @@ DeviceFD Module::Open(const std::string& device_name) {
 }

 NvResult Module::Ioctl1(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
-                        std::vector<u8>& output) {
+                        std::vector<u8>& output, IoctlCtrl& ctrl) {
    if (fd < 0) {
        LOG_ERROR(Service_NVDRV, "Invalid DeviceFD={}!", fd);
        return NvResult::InvalidState;
@@ -104,11 +104,12 @@ NvResult Module::Ioctl1(DeviceFD fd, Ioctl command, const std::vector<u8>& input
        return NvResult::NotImplemented;
    }

-    return itr->second->Ioctl1(command, input, output);
+    return itr->second->Ioctl1(command, input, output, ctrl);
 }

 NvResult Module::Ioctl2(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
-                        const std::vector<u8>& inline_input, std::vector<u8>& output) {
+                        const std::vector<u8>& inline_input, std::vector<u8>& output,
+                        IoctlCtrl& ctrl) {
    if (fd < 0) {
        LOG_ERROR(Service_NVDRV, "Invalid DeviceFD={}!", fd);
        return NvResult::InvalidState;
@@ -121,11 +122,11 @@ NvResult Module::Ioctl2(DeviceFD fd, Ioctl command, const std::vector<u8>& input
        return NvResult::NotImplemented;
    }

-    return itr->second->Ioctl2(command, input, inline_input, output);
+    return itr->second->Ioctl2(command, input, inline_input, output, ctrl);
 }

 NvResult Module::Ioctl3(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
-                        std::vector<u8>& output, std::vector<u8>& inline_output) {
+                        std::vector<u8>& output, std::vector<u8>& inline_output, IoctlCtrl& ctrl) {
    if (fd < 0) {
        LOG_ERROR(Service_NVDRV, "Invalid DeviceFD={}!", fd);
        return NvResult::InvalidState;
@@ -138,7 +139,7 @@ NvResult Module::Ioctl3(DeviceFD fd, Ioctl command, const std::vector<u8>& input
        return NvResult::NotImplemented;
    }

-    return itr->second->Ioctl3(command, input, output, inline_output);
+    return itr->second->Ioctl3(command, input, output, inline_output, ctrl);
 }

 NvResult Module::Close(DeviceFD fd) {
--- a/src/core/hle/service/nvdrv/nvdrv.h
+++ b/src/core/hle/service/nvdrv/nvdrv.h
@@ -119,13 +119,13 @@ public:

    /// Sends an ioctl command to the specified file descriptor.
    NvResult Ioctl1(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
-                    std::vector<u8>& output);
+                    std::vector<u8>& output, IoctlCtrl& ctrl);

    NvResult Ioctl2(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
-                    const std::vector<u8>& inline_input, std::vector<u8>& output);
+                    const std::vector<u8>& inline_input, std::vector<u8>& output, IoctlCtrl& ctrl);

    NvResult Ioctl3(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
-                    std::vector<u8>& output, std::vector<u8>& inline_output);
+                    std::vector<u8>& output, std::vector<u8>& inline_output, IoctlCtrl& ctrl);

    /// Closes a device file descriptor and returns operation success.
    NvResult Close(DeviceFD fd);
--- a/src/core/hle/service/nvdrv/syncpoint_manager.h
+++ b/src/core/hle/service/nvdrv/syncpoint_manager.h
@@ -37,7 +37,7 @@ public:
     * @returns The lower bound for the specified syncpoint.
     */
    u32 GetSyncpointMin(u32 syncpoint_id) const {
-        return syncpoints[syncpoint_id].min.load(std::memory_order_relaxed);
+        return syncpoints.at(syncpoint_id).min.load(std::memory_order_relaxed);
    }

    /**
@@ -46,7 +46,7 @@ public:
     * @returns The upper bound for the specified syncpoint.
     */
    u32 GetSyncpointMax(u32 syncpoint_id) const {
-        return syncpoints[syncpoint_id].max.load(std::memory_order_relaxed);
+        return syncpoints.at(syncpoint_id).max.load(std::memory_order_relaxed);
    }

    /**
--- a/src/core/hle/service/service.cpp
+++ b/src/core/hle/service/service.cpp
@@ -188,17 +188,19 @@ ResultCode ServiceFrameworkBase::HandleSyncRequest(Kernel::HLERequestContext& co
    return RESULT_SUCCESS;
 }

-/// Initialize ServiceManager
-void Init(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system) {
+/// Initialize Services
+Services::Services(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system)
+    : nv_flinger{std::make_unique<NVFlinger::NVFlinger>(system)} {
+
    // NVFlinger needs to be accessed by several services like Vi and AppletOE so we instantiate it
    // here and pass it into the respective InstallInterfaces functions.
-    auto nv_flinger = std::make_shared<NVFlinger::NVFlinger>(system);
+
    system.GetFileSystemController().CreateFactories(*system.GetFilesystem(), false);

    SM::ServiceManager::InstallInterfaces(sm, system.Kernel());

    Account::InstallInterfaces(system);
-    AM::InstallInterfaces(*sm, nv_flinger, system);
+    AM::InstallInterfaces(*sm, *nv_flinger, system);
    AOC::InstallInterfaces(*sm, system);
    APM::InstallInterfaces(system);
    Audio::InstallInterfaces(*sm, system);
@@ -246,14 +248,10 @@ void Init(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system) {
    SSL::InstallInterfaces(*sm);
    Time::InstallInterfaces(system);
    USB::InstallInterfaces(*sm);
-    VI::InstallInterfaces(*sm, nv_flinger);
+    VI::InstallInterfaces(*sm, *nv_flinger);
    WLAN::InstallInterfaces(*sm);
-
-    LOG_DEBUG(Service, "initialized OK");
 }

-/// Shutdown ServiceManager
-void Shutdown() {
-    LOG_DEBUG(Service, "shutdown OK");
-}
+Services::~Services() = default;
+
 } // namespace Service
--- a/src/core/hle/service/service.h
+++ b/src/core/hle/service/service.h
@@ -29,7 +29,11 @@ namespace Service {

 namespace FileSystem {
 class FileSystemController;
-} // namespace FileSystem
+}
+
+namespace NVFlinger {
+class NVFlinger;
+}

 namespace SM {
 class ServiceManager;
@@ -181,10 +185,17 @@ private:
    }
 };

-/// Initialize ServiceManager
-void Init(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system);
+/**
+ * The purpose of this class is to own any objects that need to be shared across the other service
+ * implementations. Will be torn down when the global system instance is shutdown.
+ */
+class Services final {
+public:
+    explicit Services(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system);
+    ~Services();

-/// Shutdown ServiceManager
-void Shutdown();
+private:
+    std::unique_ptr<NVFlinger::NVFlinger> nv_flinger;
+};

 } // namespace Service
--- a/src/core/hle/service/vi/vi.cpp
+++ b/src/core/hle/service/vi/vi.cpp
@@ -492,8 +492,8 @@ private:

 class IHOSBinderDriver final : public ServiceFramework<IHOSBinderDriver> {
 public:
-    explicit IHOSBinderDriver(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
-        : ServiceFramework("IHOSBinderDriver"), nv_flinger(std::move(nv_flinger)) {
+    explicit IHOSBinderDriver(NVFlinger::NVFlinger& nv_flinger)
+        : ServiceFramework("IHOSBinderDriver"), nv_flinger(nv_flinger) {
        static const FunctionInfo functions[] = {
            {0, &IHOSBinderDriver::TransactParcel, "TransactParcel"},
            {1, &IHOSBinderDriver::AdjustRefcount, "AdjustRefcount"},
@@ -530,8 +530,8 @@ private:
        LOG_DEBUG(Service_VI, "called. id=0x{:08X} transaction={:X}, flags=0x{:08X}", id,
                  static_cast<u32>(transaction), flags);

-        const auto guard = nv_flinger->Lock();
-        auto& buffer_queue = nv_flinger->FindBufferQueue(id);
+        const auto guard = nv_flinger.Lock();
+        auto& buffer_queue = nv_flinger.FindBufferQueue(id);

        switch (transaction) {
        case TransactionId::Connect: {
@@ -570,8 +570,8 @@ private:
                    [=, this](std::shared_ptr<Kernel::Thread> thread,
                              Kernel::HLERequestContext& ctx, Kernel::ThreadWakeupReason reason) {
                        // Repeat TransactParcel DequeueBuffer when a buffer is available
-                        const auto guard = nv_flinger->Lock();
-                        auto& buffer_queue = nv_flinger->FindBufferQueue(id);
+                        const auto guard = nv_flinger.Lock();
+                        auto& buffer_queue = nv_flinger.FindBufferQueue(id);
                        auto result = buffer_queue.DequeueBuffer(width, height);
                        ASSERT_MSG(result != std::nullopt, "Could not dequeue buffer.");

@@ -676,7 +676,7 @@ private:

        LOG_WARNING(Service_VI, "(STUBBED) called id={}, unknown={:08X}", id, unknown);

-        const auto& buffer_queue = nv_flinger->FindBufferQueue(id);
+        const auto& buffer_queue = nv_flinger.FindBufferQueue(id);

        // TODO(Subv): Find out what this actually is.
        IPC::ResponseBuilder rb{ctx, 2, 1};
@@ -684,8 +684,8 @@ private:
        rb.PushCopyObjects(buffer_queue.GetBufferWaitEvent());
    }

-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger;
-}; // namespace VI
+    NVFlinger::NVFlinger& nv_flinger;
+};

 class ISystemDisplayService final : public ServiceFramework<ISystemDisplayService> {
 public:
@@ -790,8 +790,8 @@ private:

 class IManagerDisplayService final : public ServiceFramework<IManagerDisplayService> {
 public:
-    explicit IManagerDisplayService(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
-        : ServiceFramework("IManagerDisplayService"), nv_flinger(std::move(nv_flinger)) {
+    explicit IManagerDisplayService(NVFlinger::NVFlinger& nv_flinger)
+        : ServiceFramework("IManagerDisplayService"), nv_flinger(nv_flinger) {
        // clang-format off
        static const FunctionInfo functions[] = {
            {200, nullptr, "AllocateProcessHeapBlock"},
@@ -893,7 +893,7 @@ private:
                    "(STUBBED) called. unknown=0x{:08X}, display=0x{:016X}, aruid=0x{:016X}",
                    unknown, display, aruid);

-        const auto layer_id = nv_flinger->CreateLayer(display);
+        const auto layer_id = nv_flinger.CreateLayer(display);
        if (!layer_id) {
            LOG_ERROR(Service_VI, "Layer not found! display=0x{:016X}", display);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -930,12 +930,12 @@ private:
        rb.Push(RESULT_SUCCESS);
    }

-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger;
+    NVFlinger::NVFlinger& nv_flinger;
 };

 class IApplicationDisplayService final : public ServiceFramework<IApplicationDisplayService> {
 public:
-    explicit IApplicationDisplayService(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger);
+    explicit IApplicationDisplayService(NVFlinger::NVFlinger& nv_flinger);

 private:
    enum class ConvertedScaleMode : u64 {
@@ -1010,7 +1010,7 @@ private:

        ASSERT_MSG(name == "Default", "Non-default displays aren't supported yet");

-        const auto display_id = nv_flinger->OpenDisplay(name);
+        const auto display_id = nv_flinger.OpenDisplay(name);
        if (!display_id) {
            LOG_ERROR(Service_VI, "Display not found! display_name={}", name);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -1110,7 +1110,7 @@ private:

        LOG_DEBUG(Service_VI, "called. layer_id=0x{:016X}, aruid=0x{:016X}", layer_id, aruid);

-        const auto display_id = nv_flinger->OpenDisplay(display_name);
+        const auto display_id = nv_flinger.OpenDisplay(display_name);
        if (!display_id) {
            LOG_ERROR(Service_VI, "Layer not found! layer_id={}", layer_id);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -1118,7 +1118,7 @@ private:
            return;
        }

-        const auto buffer_queue_id = nv_flinger->FindBufferQueueId(*display_id, layer_id);
+        const auto buffer_queue_id = nv_flinger.FindBufferQueueId(*display_id, layer_id);
        if (!buffer_queue_id) {
            LOG_ERROR(Service_VI, "Buffer queue id not found! display_id={}", *display_id);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -1138,7 +1138,7 @@ private:

        LOG_DEBUG(Service_VI, "called. layer_id=0x{:016X}", layer_id);

-        nv_flinger->CloseLayer(layer_id);
+        nv_flinger.CloseLayer(layer_id);

        IPC::ResponseBuilder rb{ctx, 2};
        rb.Push(RESULT_SUCCESS);
@@ -1154,7 +1154,7 @@ private:

        // TODO(Subv): What's the difference between a Stray and a Managed layer?

-        const auto layer_id = nv_flinger->CreateLayer(display_id);
+        const auto layer_id = nv_flinger.CreateLayer(display_id);
        if (!layer_id) {
            LOG_ERROR(Service_VI, "Layer not found! layer_id={}", *layer_id);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -1162,7 +1162,7 @@ private:
            return;
        }

-        const auto buffer_queue_id = nv_flinger->FindBufferQueueId(display_id, *layer_id);
+        const auto buffer_queue_id = nv_flinger.FindBufferQueueId(display_id, *layer_id);
        if (!buffer_queue_id) {
            LOG_ERROR(Service_VI, "Buffer queue id not found! display_id={}", display_id);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -1193,7 +1193,7 @@ private:

        LOG_WARNING(Service_VI, "(STUBBED) called. display_id=0x{:016X}", display_id);

-        const auto vsync_event = nv_flinger->FindVsyncEvent(display_id);
+        const auto vsync_event = nv_flinger.FindVsyncEvent(display_id);
        if (!vsync_event) {
            LOG_ERROR(Service_VI, "Vsync event was not found for display_id={}", display_id);
            IPC::ResponseBuilder rb{ctx, 2};
@@ -1258,12 +1258,11 @@ private:
        }
    }

-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger;
+    NVFlinger::NVFlinger& nv_flinger;
 };

-IApplicationDisplayService::IApplicationDisplayService(
-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
-    : ServiceFramework("IApplicationDisplayService"), nv_flinger(std::move(nv_flinger)) {
+IApplicationDisplayService::IApplicationDisplayService(NVFlinger::NVFlinger& nv_flinger)
+    : ServiceFramework("IApplicationDisplayService"), nv_flinger(nv_flinger) {
    static const FunctionInfo functions[] = {
        {100, &IApplicationDisplayService::GetRelayService, "GetRelayService"},
        {101, &IApplicationDisplayService::GetSystemDisplayService, "GetSystemDisplayService"},
@@ -1304,8 +1303,7 @@ static bool IsValidServiceAccess(Permission permission, Policy policy) {
    return false;
 }

-void detail::GetDisplayServiceImpl(Kernel::HLERequestContext& ctx,
-                                   std::shared_ptr<NVFlinger::NVFlinger> nv_flinger,
+void detail::GetDisplayServiceImpl(Kernel::HLERequestContext& ctx, NVFlinger::NVFlinger& nv_flinger,
                                   Permission permission) {
    IPC::RequestParser rp{ctx};
    const auto policy = rp.PopEnum<Policy>();
@@ -1319,11 +1317,10 @@ void detail::GetDisplayServiceImpl(Kernel::HLERequestContext& ctx,

    IPC::ResponseBuilder rb{ctx, 2, 0, 1};
    rb.Push(RESULT_SUCCESS);
-    rb.PushIpcInterface<IApplicationDisplayService>(std::move(nv_flinger));
+    rb.PushIpcInterface<IApplicationDisplayService>(nv_flinger);
 }

-void InstallInterfaces(SM::ServiceManager& service_manager,
-                       std::shared_ptr<NVFlinger::NVFlinger> nv_flinger) {
+void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nv_flinger) {
    std::make_shared<VI_M>(nv_flinger)->InstallAsService(service_manager);
    std::make_shared<VI_S>(nv_flinger)->InstallAsService(service_manager);
    std::make_shared<VI_U>(nv_flinger)->InstallAsService(service_manager);
--- a/src/core/hle/service/vi/vi.h
+++ b/src/core/hle/service/vi/vi.h
@@ -43,12 +43,11 @@ enum class Policy {
 };

 namespace detail {
-void GetDisplayServiceImpl(Kernel::HLERequestContext& ctx,
-                           std::shared_ptr<NVFlinger::NVFlinger> nv_flinger, Permission permission);
+void GetDisplayServiceImpl(Kernel::HLERequestContext& ctx, NVFlinger::NVFlinger& nv_flinger,
+                           Permission permission);
 } // namespace detail

 /// Registers all VI services with the specified service manager.
-void InstallInterfaces(SM::ServiceManager& service_manager,
-                       std::shared_ptr<NVFlinger::NVFlinger> nv_flinger);
+void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nv_flinger);

 } // namespace Service::VI
--- a/src/core/hle/service/vi/vi_m.cpp
+++ b/src/core/hle/service/vi/vi_m.cpp
@@ -8,8 +8,7 @@

 namespace Service::VI {

-VI_M::VI_M(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
-    : ServiceFramework{"vi:m"}, nv_flinger{std::move(nv_flinger)} {
+VI_M::VI_M(NVFlinger::NVFlinger& nv_flinger) : ServiceFramework{"vi:m"}, nv_flinger{nv_flinger} {
    static const FunctionInfo functions[] = {
        {2, &VI_M::GetDisplayService, "GetDisplayService"},
        {3, nullptr, "GetDisplayServiceWithProxyNameExchange"},
--- a/src/core/hle/service/vi/vi_m.h
+++ b/src/core/hle/service/vi/vi_m.h
@@ -18,13 +18,13 @@ namespace Service::VI {

 class VI_M final : public ServiceFramework<VI_M> {
 public:
-    explicit VI_M(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger);
+    explicit VI_M(NVFlinger::NVFlinger& nv_flinger);
    ~VI_M() override;

 private:
    void GetDisplayService(Kernel::HLERequestContext& ctx);

-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger;
+    NVFlinger::NVFlinger& nv_flinger;
 };

 } // namespace Service::VI
--- a/src/core/hle/service/vi/vi_s.cpp
+++ b/src/core/hle/service/vi/vi_s.cpp
@@ -8,8 +8,7 @@

 namespace Service::VI {

-VI_S::VI_S(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
-    : ServiceFramework{"vi:s"}, nv_flinger{std::move(nv_flinger)} {
+VI_S::VI_S(NVFlinger::NVFlinger& nv_flinger) : ServiceFramework{"vi:s"}, nv_flinger{nv_flinger} {
    static const FunctionInfo functions[] = {
        {1, &VI_S::GetDisplayService, "GetDisplayService"},
        {3, nullptr, "GetDisplayServiceWithProxyNameExchange"},
--- a/src/core/hle/service/vi/vi_s.h
+++ b/src/core/hle/service/vi/vi_s.h
@@ -18,13 +18,13 @@ namespace Service::VI {

 class VI_S final : public ServiceFramework<VI_S> {
 public:
-    explicit VI_S(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger);
+    explicit VI_S(NVFlinger::NVFlinger& nv_flinger);
    ~VI_S() override;

 private:
    void GetDisplayService(Kernel::HLERequestContext& ctx);

-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger;
+    NVFlinger::NVFlinger& nv_flinger;
 };

 } // namespace Service::VI
--- a/src/core/hle/service/vi/vi_u.cpp
+++ b/src/core/hle/service/vi/vi_u.cpp
@@ -8,8 +8,7 @@

 namespace Service::VI {

-VI_U::VI_U(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
-    : ServiceFramework{"vi:u"}, nv_flinger{std::move(nv_flinger)} {
+VI_U::VI_U(NVFlinger::NVFlinger& nv_flinger) : ServiceFramework{"vi:u"}, nv_flinger{nv_flinger} {
    static const FunctionInfo functions[] = {
        {0, &VI_U::GetDisplayService, "GetDisplayService"},
        {1, nullptr, "GetDisplayServiceWithProxyNameExchange"},
--- a/src/core/hle/service/vi/vi_u.h
+++ b/src/core/hle/service/vi/vi_u.h
@@ -18,13 +18,13 @@ namespace Service::VI {

 class VI_U final : public ServiceFramework<VI_U> {
 public:
-    explicit VI_U(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger);
+    explicit VI_U(NVFlinger::NVFlinger& nv_flinger);
    ~VI_U() override;

 private:
    void GetDisplayService(Kernel::HLERequestContext& ctx);

-    std::shared_ptr<NVFlinger::NVFlinger> nv_flinger;
+    NVFlinger::NVFlinger& nv_flinger;
 };

 } // namespace Service::VI
--- a/src/core/loader/loader.cpp
+++ b/src/core/loader/loader.cpp
@@ -198,10 +198,11 @@ AppLoader::~AppLoader() = default;
 * @param system The system context to use.
 * @param file   The file to retrieve the loader for
 * @param type   The file type
+ * @param program_index Specifies the index within the container of the program to launch.
 * @return std::unique_ptr<AppLoader> a pointer to a loader object;  nullptr for unsupported type
 */
 static std::unique_ptr<AppLoader> GetFileLoader(Core::System& system, FileSys::VirtualFile file,
-                                                FileType type) {
+                                                FileType type, std::size_t program_index) {
    switch (type) {
    // Standard ELF file format.
    case FileType::ELF:
@@ -222,7 +223,7 @@ static std::unique_ptr<AppLoader> GetFileLoader(Core::System& system, FileSys::V
    // NX XCI (nX Card Image) file format.
    case FileType::XCI:
        return std::make_unique<AppLoader_XCI>(std::move(file), system.GetFileSystemController(),
-                                               system.GetContentProvider());
+                                               system.GetContentProvider(), program_index);

    // NX NAX (NintendoAesXts) file format.
    case FileType::NAX:
@@ -231,7 +232,7 @@ static std::unique_ptr<AppLoader> GetFileLoader(Core::System& system, FileSys::V
    // NX NSP (Nintendo Submission Package) file format
    case FileType::NSP:
        return std::make_unique<AppLoader_NSP>(std::move(file), system.GetFileSystemController(),
-                                               system.GetContentProvider());
+                                               system.GetContentProvider(), program_index);

    // NX KIP (Kernel Internal Process) file format
    case FileType::KIP:
@@ -246,7 +247,8 @@ static std::unique_ptr<AppLoader> GetFileLoader(Core::System& system, FileSys::V
    }
 }

-std::unique_ptr<AppLoader> GetLoader(Core::System& system, FileSys::VirtualFile file) {
+std::unique_ptr<AppLoader> GetLoader(Core::System& system, FileSys::VirtualFile file,
+                                     std::size_t program_index) {
    FileType type = IdentifyFile(file);
    const FileType filename_type = GuessFromFilename(file->GetName());

@@ -260,7 +262,7 @@ std::unique_ptr<AppLoader> GetLoader(Core::System& system, FileSys::VirtualFile

    LOG_DEBUG(Loader, "Loading file {} as {}...", file->GetName(), GetFileTypeString(type));

-    return GetFileLoader(system, std::move(file), type);
+    return GetFileLoader(system, std::move(file), type, program_index);
 }

 } // namespace Loader
--- a/src/core/loader/loader.h
+++ b/src/core/loader/loader.h
@@ -293,9 +293,11 @@ protected:
 *
 * @param system The system context.
 * @param file   The bootable file.
+ * @param program_index Specifies the index within the container of the program to launch.
 *
 * @return the best loader for this file.
 */
-std::unique_ptr<AppLoader> GetLoader(Core::System& system, FileSys::VirtualFile file);
+std::unique_ptr<AppLoader> GetLoader(Core::System& system, FileSys::VirtualFile file,
+                                     std::size_t program_index = 0);

 } // namespace Loader
--- a/src/core/loader/nsp.cpp
+++ b/src/core/loader/nsp.cpp
@@ -23,8 +23,9 @@ namespace Loader {

 AppLoader_NSP::AppLoader_NSP(FileSys::VirtualFile file,
                             const Service::FileSystem::FileSystemController& fsc,
-                             const FileSys::ContentProvider& content_provider)
-    : AppLoader(file), nsp(std::make_unique<FileSys::NSP>(file)),
+                             const FileSys::ContentProvider& content_provider,
+                             std::size_t program_index)
+    : AppLoader(file), nsp(std::make_unique<FileSys::NSP>(file, program_index)),
      title_id(nsp->GetProgramTitleID()) {

    if (nsp->GetStatus() != ResultStatus::Success) {
--- a/src/core/loader/nsp.h
+++ b/src/core/loader/nsp.h
@@ -28,7 +28,8 @@ class AppLoader_NSP final : public AppLoader {
 public:
    explicit AppLoader_NSP(FileSys::VirtualFile file,
                           const Service::FileSystem::FileSystemController& fsc,
-                           const FileSys::ContentProvider& content_provider);
+                           const FileSys::ContentProvider& content_provider,
+                           std::size_t program_index);
    ~AppLoader_NSP() override;

    /**
--- a/src/core/loader/xci.cpp
+++ b/src/core/loader/xci.cpp
@@ -22,8 +22,9 @@ namespace Loader {

 AppLoader_XCI::AppLoader_XCI(FileSys::VirtualFile file,
                             const Service::FileSystem::FileSystemController& fsc,
-                             const FileSys::ContentProvider& content_provider)
-    : AppLoader(file), xci(std::make_unique<FileSys::XCI>(file)),
+                             const FileSys::ContentProvider& content_provider,
+                             std::size_t program_index)
+    : AppLoader(file), xci(std::make_unique<FileSys::XCI>(file, program_index)),
      nca_loader(std::make_unique<AppLoader_NCA>(xci->GetProgramNCAFile())) {
    if (xci->GetStatus() != ResultStatus::Success) {
        return;
--- a/src/core/loader/xci.h
+++ b/src/core/loader/xci.h
@@ -28,7 +28,8 @@ class AppLoader_XCI final : public AppLoader {
 public:
    explicit AppLoader_XCI(FileSys::VirtualFile file,
                           const Service::FileSystem::FileSystemController& fsc,
-                           const FileSys::ContentProvider& content_provider);
+                           const FileSys::ContentProvider& content_provider,
+                           std::size_t program_index);
    ~AppLoader_XCI() override;

    /**
--- a/src/input_common/CMakeLists.txt
+++ b/src/input_common/CMakeLists.txt
@@ -56,8 +56,8 @@ else()
        -Werror=reorder
        -Werror=shadow
        -Werror=sign-compare
-        -Werror=unused-but-set-parameter
-        -Werror=unused-but-set-variable
+        $<$<CXX_COMPILER_ID:GNU>:-Werror=unused-but-set-parameter>
+        $<$<CXX_COMPILER_ID:GNU>:-Werror=unused-but-set-variable>
        -Werror=unused-variable
    )
 endif()
--- a/src/input_common/analog_from_button.cpp
+++ b/src/input_common/analog_from_button.cpp
@@ -2,6 +2,10 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.

+#include <chrono>
+#include <cmath>
+#include <thread>
+#include "common/math_util.h"
 #include "input_common/analog_from_button.h"

 namespace InputCommon {
@@ -11,31 +15,104 @@ public:
    using Button = std::unique_ptr<Input::ButtonDevice>;

    Analog(Button up_, Button down_, Button left_, Button right_, Button modifier_,
-           float modifier_scale_)
+           float modifier_scale_, float modifier_angle_)
        : up(std::move(up_)), down(std::move(down_)), left(std::move(left_)),
-          right(std::move(right_)), modifier(std::move(modifier_)),
-          modifier_scale(modifier_scale_) {}
+          right(std::move(right_)), modifier(std::move(modifier_)), modifier_scale(modifier_scale_),
+          modifier_angle(modifier_angle_) {
+        update_thread = std::thread(&Analog::UpdateStatus, this);
+    }
+
+    ~Analog() override {
+        update_thread_running = false;
+        if (update_thread.joinable()) {
+            update_thread.join();
+        }
+    }
+
+    void MoveToDirection(bool enable, float to_angle) {
+        if (!enable) {
+            return;
+        }
+        constexpr float TAU = Common::PI * 2.0f;
+        // Use wider angle to ease the transition.
+        constexpr float aperture = TAU * 0.15f;
+        const float top_limit = to_angle + aperture;
+        const float bottom_limit = to_angle - aperture;
+
+        if ((angle > to_angle && angle <= top_limit) ||
+            (angle + TAU > to_angle && angle + TAU <= top_limit)) {
+            angle -= modifier_angle;
+            if (angle < 0) {
+                angle += TAU;
+            }
+        } else if ((angle >= bottom_limit && angle < to_angle) ||
+                   (angle - TAU >= bottom_limit && angle - TAU < to_angle)) {
+            angle += modifier_angle;
+            if (angle >= TAU) {
+                angle -= TAU;
+            }
+        } else {
+            angle = to_angle;
+        }
+    }
+
+    void UpdateStatus() {
+        while (update_thread_running) {
+            const float coef = modifier->GetStatus() ? modifier_scale : 1.0f;
+
+            bool r = right->GetStatus();
+            bool l = left->GetStatus();
+            bool u = up->GetStatus();
+            bool d = down->GetStatus();
+
+            // Eliminate contradictory movements
+            if (r && l) {
+                r = false;
+                l = false;
+            }
+            if (u && d) {
+                u = false;
+                d = false;
+            }
+
+            // Move to the right
+            MoveToDirection(r && !u && !d, 0.0f);
+
+            // Move to the upper right
+            MoveToDirection(r && u && !d, Common::PI * 0.25f);
+
+            // Move up
+            MoveToDirection(u && !l && !r, Common::PI * 0.5f);
+
+            // Move to the upper left
+            MoveToDirection(l && u && !d, Common::PI * 0.75f);
+
+            // Move to the left
+            MoveToDirection(l && !u && !d, Common::PI);
+
+            // Move to the bottom left
+            MoveToDirection(l && !u && d, Common::PI * 1.25f);
+
+            // Move down
+            MoveToDirection(d && !l && !r, Common::PI * 1.5f);
+
+            // Move to the bottom right
+            MoveToDirection(r && !u && d, Common::PI * 1.75f);
+
+            // Move if a key is pressed
+            if (r || l || u || d) {
+                amplitude = coef;
+            } else {
+                amplitude = 0;
+            }
+
+            // Delay the update rate to 100hz
+            std::this_thread::sleep_for(std::chrono::milliseconds(10));
+        }
+    }

    std::tuple<float, float> GetStatus() const override {
-        constexpr float SQRT_HALF = 0.707106781f;
-        int x = 0, y = 0;
-
-        if (right->GetStatus()) {
-            ++x;
-        }
-        if (left->GetStatus()) {
-            --x;
-        }
-        if (up->GetStatus()) {
-            ++y;
-        }
-        if (down->GetStatus()) {
-            --y;
-        }
-
-        const float coef = modifier->GetStatus() ? modifier_scale : 1.0f;
-        return std::make_tuple(static_cast<float>(x) * coef * (y == 0 ? 1.0f : SQRT_HALF),
-                               static_cast<float>(y) * coef * (x == 0 ? 1.0f : SQRT_HALF));
+        return std::make_tuple(std::cos(angle) * amplitude, std::sin(angle) * amplitude);
    }

    bool GetAnalogDirectionStatus(Input::AnalogDirection direction) const override {
@@ -59,6 +136,11 @@ private:
    Button right;
    Button modifier;
    float modifier_scale;
+    float modifier_angle;
+    float angle{};
+    float amplitude{};
+    std::thread update_thread;
+    bool update_thread_running{true};
 };

 std::unique_ptr<Input::AnalogDevice> AnalogFromButton::Create(const Common::ParamPackage& params) {
@@ -69,8 +151,10 @@ std::unique_ptr<Input::AnalogDevice> AnalogFromButton::Create(const Common::Para
    auto right = Input::CreateDevice<Input::ButtonDevice>(params.Get("right", null_engine));
    auto modifier = Input::CreateDevice<Input::ButtonDevice>(params.Get("modifier", null_engine));
    auto modifier_scale = params.Get("modifier_scale", 0.5f);
+    auto modifier_angle = params.Get("modifier_angle", 0.035f);
    return std::make_unique<Analog>(std::move(up), std::move(down), std::move(left),
-                                    std::move(right), std::move(modifier), modifier_scale);
+                                    std::move(right), std::move(modifier), modifier_scale,
+                                    modifier_angle);
 }

 } // namespace InputCommon
--- a/src/input_common/gcadapter/gc_poller.cpp
+++ b/src/input_common/gcadapter/gc_poller.cpp
@@ -299,8 +299,8 @@ public:
        return gcadapter->RumblePlay(port, 0);
    }

-    bool SetRumblePlay(f32 amp_low, [[maybe_unused]] f32 freq_low, [[maybe_unused]] f32 amp_high,
-                       f32 freq_high) const override {
+    bool SetRumblePlay(f32 amp_low, [[maybe_unused]] f32 freq_low, f32 amp_high,
+                       [[maybe_unused]] f32 freq_high) const override {
        const auto mean_amplitude = (amp_low + amp_high) * 0.5f;
        const auto processed_amplitude =
            static_cast<u8>((mean_amplitude + std::pow(mean_amplitude, 0.3f)) * 0.5f * 0x8);
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -95,22 +95,29 @@ void GPU::WaitFence(u32 syncpoint_id, u32 value) {
    if (!is_async) {
        return;
    }
+    if (syncpoint_id == UINT32_MAX) {
+        // TODO: Research what this does.
+        LOG_ERROR(HW_GPU, "Waiting for syncpoint -1 not implemented");
+        return;
+    }
    MICROPROFILE_SCOPE(GPU_wait);
    std::unique_lock lock{sync_mutex};
-    sync_cv.wait(lock, [=, this] { return syncpoints[syncpoint_id].load() >= value; });
+    sync_cv.wait(lock, [=, this] { return syncpoints.at(syncpoint_id).load() >= value; });
 }

 void GPU::IncrementSyncPoint(const u32 syncpoint_id) {
-    syncpoints[syncpoint_id]++;
+    auto& syncpoint = syncpoints.at(syncpoint_id);
+    syncpoint++;
    std::lock_guard lock{sync_mutex};
    sync_cv.notify_all();
-    if (!syncpt_interrupts[syncpoint_id].empty()) {
-        u32 value = syncpoints[syncpoint_id].load();
-        auto it = syncpt_interrupts[syncpoint_id].begin();
-        while (it != syncpt_interrupts[syncpoint_id].end()) {
+    auto& interrupt = syncpt_interrupts.at(syncpoint_id);
+    if (!interrupt.empty()) {
+        u32 value = syncpoint.load();
+        auto it = interrupt.begin();
+        while (it != interrupt.end()) {
            if (value >= *it) {
                TriggerCpuInterrupt(syncpoint_id, *it);
-                it = syncpt_interrupts[syncpoint_id].erase(it);
+                it = interrupt.erase(it);
                continue;
            }
            it++;
@@ -119,22 +126,22 @@ void GPU::IncrementSyncPoint(const u32 syncpoint_id) {
 }

 u32 GPU::GetSyncpointValue(const u32 syncpoint_id) const {
-    return syncpoints[syncpoint_id].load();
+    return syncpoints.at(syncpoint_id).load();
 }

 void GPU::RegisterSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
-    auto& interrupt = syncpt_interrupts[syncpoint_id];
+    auto& interrupt = syncpt_interrupts.at(syncpoint_id);
    bool contains = std::any_of(interrupt.begin(), interrupt.end(),
                                [value](u32 in_value) { return in_value == value; });
    if (contains) {
        return;
    }
-    syncpt_interrupts[syncpoint_id].emplace_back(value);
+    interrupt.emplace_back(value);
 }

 bool GPU::CancelSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
    std::lock_guard lock{sync_mutex};
-    auto& interrupt = syncpt_interrupts[syncpoint_id];
+    auto& interrupt = syncpt_interrupts.at(syncpoint_id);
    const auto iter =
        std::find_if(interrupt.begin(), interrupt.end(),
                     [value](u32 interrupt_value) { return value == interrupt_value; });
--- a/src/video_core/renderer_opengl/renderer_opengl.cpp
+++ b/src/video_core/renderer_opengl/renderer_opengl.cpp
@@ -151,8 +151,8 @@ void RendererOpenGL::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {

    rasterizer->TickFrame();

-    render_window.PollEvents();
    context->SwapBuffers();
+    render_window.OnFrameDisplayed();
 }

 void RendererOpenGL::PrepareRendertarget(const Tegra::FramebufferConfig* framebuffer) {
--- a/src/video_core/renderer_vulkan/fixed_pipeline_state.cpp
+++ b/src/video_core/renderer_vulkan/fixed_pipeline_state.cpp
@@ -8,6 +8,7 @@

 #include <boost/functional/hash.hpp>

+#include "common/bit_cast.h"
 #include "common/cityhash.h"
 #include "common/common_types.h"
 #include "video_core/renderer_vulkan/fixed_pipeline_state.h"
@@ -60,7 +61,13 @@ void FixedPipelineState::Fill(const Maxwell& regs, bool has_extended_dynamic_sta
    rasterize_enable.Assign(regs.rasterize_enable != 0 ? 1 : 0);
    topology.Assign(regs.draw.topology);

-    std::memcpy(&point_size, &regs.point_size, sizeof(point_size)); // TODO: C++20 std::bit_cast
+    alpha_raw = 0;
+    const auto test_func =
+        regs.alpha_test_enabled == 1 ? regs.alpha_test_func : Maxwell::ComparisonOp::Always;
+    alpha_test_func.Assign(PackComparisonOp(test_func));
+    alpha_test_ref = Common::BitCast<u32>(regs.alpha_test_ref);
+
+    point_size = Common::BitCast<u32>(regs.point_size);

    for (std::size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
        binding_divisors[index] =
--- a/src/video_core/renderer_vulkan/fixed_pipeline_state.h
+++ b/src/video_core/renderer_vulkan/fixed_pipeline_state.h
@@ -187,6 +187,13 @@ struct FixedPipelineState {
        BitField<23, 1, u32> rasterize_enable;
        BitField<24, 4, Maxwell::PrimitiveTopology> topology;
    };
+
+    u32 alpha_test_ref; ///< Alpha test reference value
+    union {
+        u32 alpha_raw;
+        BitField<0, 3, u32> alpha_test_func;
+    };
+
    u32 point_size;
    std::array<u32, Maxwell::NumVertexArrays> binding_divisors;
    std::array<VertexAttribute, Maxwell::NumVertexAttributes> attributes;
--- a/src/video_core/renderer_vulkan/renderer_vulkan.cpp
+++ b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
@@ -252,8 +252,6 @@ RendererVulkan::~RendererVulkan() {
 }

 void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
-    render_window.PollEvents();
-
    if (!framebuffer) {
        return;
    }
@@ -283,7 +281,7 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
        rasterizer->TickFrame();
    }

-    render_window.PollEvents();
+    render_window.OnFrameDisplayed();
 }

 bool RendererVulkan::Init() {
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -7,6 +7,7 @@
 #include <memory>
 #include <vector>

+#include "common/bit_cast.h"
 #include "common/microprofile.h"
 #include "core/core.h"
 #include "core/memory.h"
@@ -344,6 +345,11 @@ VKPipelineCache::DecompileShaders(const FixedPipelineState& fixed_state) {
    }
    specialization.ndc_minus_one_to_one = fixed_state.ndc_minus_one_to_one;

+    // Alpha test
+    specialization.alpha_test_func =
+        FixedPipelineState::UnpackComparisonOp(fixed_state.alpha_test_func.Value());
+    specialization.alpha_test_ref = Common::BitCast<float>(fixed_state.alpha_test_ref);
+
    SPIRVProgram program;
    std::vector<VkDescriptorSetLayoutBinding> bindings;

--- a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
+++ b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
@@ -2075,6 +2075,45 @@ private:
        return {};
    }

+    Id MaxwellToSpirvComparison(Maxwell::ComparisonOp compare_op, Id operand_1, Id operand_2) {
+        using Compare = Maxwell::ComparisonOp;
+        switch (compare_op) {
+        case Compare::NeverOld:
+            return v_false; // Never let the test pass
+        case Compare::LessOld:
+            return OpFOrdLessThan(t_bool, operand_1, operand_2);
+        case Compare::EqualOld:
+            return OpFOrdEqual(t_bool, operand_1, operand_2);
+        case Compare::LessEqualOld:
+            return OpFOrdLessThanEqual(t_bool, operand_1, operand_2);
+        case Compare::GreaterOld:
+            return OpFOrdGreaterThan(t_bool, operand_1, operand_2);
+        case Compare::NotEqualOld:
+            return OpFOrdNotEqual(t_bool, operand_1, operand_2);
+        case Compare::GreaterEqualOld:
+            return OpFOrdGreaterThanEqual(t_bool, operand_1, operand_2);
+        default:
+            UNREACHABLE();
+        }
+    }
+
+    void AlphaTest(Id pointer) {
+        if (specialization.alpha_test_func == Maxwell::ComparisonOp::AlwaysOld) {
+            return;
+        }
+        const Id true_label = OpLabel();
+        const Id discard_label = OpLabel();
+        const Id alpha_reference = Constant(t_float, specialization.alpha_test_ref);
+        const Id alpha_value = OpLoad(t_float, pointer);
+        const Id condition =
+            MaxwellToSpirvComparison(specialization.alpha_test_func, alpha_value, alpha_reference);
+
+        OpBranchConditional(condition, true_label, discard_label);
+        AddLabel(discard_label);
+        OpKill();
+        AddLabel(true_label);
+    }
+
    void PreExit() {
        if (stage == ShaderType::Vertex && specialization.ndc_minus_one_to_one) {
            const u32 position_index = out_indices.position.value();
@@ -2097,8 +2136,6 @@ private:
            UNIMPLEMENTED_IF_MSG(header.ps.omap.sample_mask != 0,
                                 "Sample mask write is unimplemented");

-            // TODO(Rodrigo): Alpha testing
-
            // Write the color outputs using the data in the shader registers, disabled
            // rendertargets/components are skipped in the register assignment.
            u32 current_reg = 0;
@@ -2110,6 +2147,9 @@ private:
                    }
                    const Id pointer = AccessElement(t_out_float, frag_colors[rt], component);
                    OpStore(pointer, SafeGetRegister(current_reg));
+                    if (rt == 0 && component == 3) {
+                        AlphaTest(pointer);
+                    }
                    ++current_reg;
                }
            }
--- a/src/video_core/renderer_vulkan/vk_shader_decompiler.h
+++ b/src/video_core/renderer_vulkan/vk_shader_decompiler.h
@@ -95,6 +95,8 @@ struct Specialization final {
    std::bitset<Maxwell::NumVertexAttributes> enabled_attributes;
    std::array<Maxwell::VertexAttribute::Type, Maxwell::NumVertexAttributes> attribute_types{};
    bool ndc_minus_one_to_one{};
+    float alpha_test_ref{};
+    Maxwell::ComparisonOp alpha_test_func{};
 };
 // Old gcc versions don't consider this trivially copyable.
 // static_assert(std::is_trivially_copyable_v<Specialization>);
--- a/src/video_core/shader/decode/image.cpp
+++ b/src/video_core/shader/decode/image.cpp
@@ -212,10 +212,10 @@ u32 GetComponentSize(TextureFormat format, std::size_t component) {
        return 0;
    case TextureFormat::R8G24:
        if (component == 0) {
-            return 8;
+            return 24;
        }
        if (component == 1) {
-            return 24;
+            return 8;
        }
        return 0;
    case TextureFormat::R8G8:
--- a/src/yuzu/applets/controller.cpp
+++ b/src/yuzu/applets/controller.cpp
@@ -72,40 +72,6 @@ bool IsControllerCompatible(Settings::ControllerType controller_type,
    }
 }

-/// Maps the controller type combobox index to Controller Type enum
-constexpr Settings::ControllerType GetControllerTypeFromIndex(int index) {
-    switch (index) {
-    case 0:
-    default:
-        return Settings::ControllerType::ProController;
-    case 1:
-        return Settings::ControllerType::DualJoyconDetached;
-    case 2:
-        return Settings::ControllerType::LeftJoycon;
-    case 3:
-        return Settings::ControllerType::RightJoycon;
-    case 4:
-        return Settings::ControllerType::Handheld;
-    }
-}
-
-/// Maps the Controller Type enum to controller type combobox index
-constexpr int GetIndexFromControllerType(Settings::ControllerType type) {
-    switch (type) {
-    case Settings::ControllerType::ProController:
-    default:
-        return 0;
-    case Settings::ControllerType::DualJoyconDetached:
-        return 1;
-    case Settings::ControllerType::LeftJoycon:
-        return 2;
-    case Settings::ControllerType::RightJoycon:
-        return 3;
-    case Settings::ControllerType::Handheld:
-        return 4;
-    }
-}
-
 } // namespace

 QtControllerSelectorDialog::QtControllerSelectorDialog(
@@ -184,6 +150,11 @@ QtControllerSelectorDialog::QtControllerSelectorDialog(
    // This avoids unintentionally changing the states of elements while loading them in.
    SetSupportedControllers();
    DisableUnsupportedPlayers();
+
+    for (std::size_t player_index = 0; player_index < NUM_PLAYERS; ++player_index) {
+        SetEmulatedControllers(player_index);
+    }
+
    LoadConfiguration();

    for (std::size_t i = 0; i < NUM_PLAYERS; ++i) {
@@ -223,8 +194,8 @@ QtControllerSelectorDialog::QtControllerSelectorDialog(

        if (i == 0) {
            connect(emulated_controllers[i], qOverload<int>(&QComboBox::currentIndexChanged),
-                    [this](int index) {
-                        UpdateDockedState(GetControllerTypeFromIndex(index) ==
+                    [this, i](int index) {
+                        UpdateDockedState(GetControllerTypeFromIndex(index, i) ==
                                          Settings::ControllerType::Handheld);
                    });
        }
@@ -281,8 +252,8 @@ void QtControllerSelectorDialog::LoadConfiguration() {
            (index == 0 && Settings::values.players.GetValue()[HANDHELD_INDEX].connected);
        player_groupboxes[index]->setChecked(connected);
        connected_controller_checkboxes[index]->setChecked(connected);
-        emulated_controllers[index]->setCurrentIndex(
-            GetIndexFromControllerType(Settings::values.players.GetValue()[index].controller_type));
+        emulated_controllers[index]->setCurrentIndex(GetIndexFromControllerType(
+            Settings::values.players.GetValue()[index].controller_type, index));
    }

    UpdateDockedState(Settings::values.players.GetValue()[HANDHELD_INDEX].connected);
@@ -338,7 +309,7 @@ bool QtControllerSelectorDialog::CheckIfParametersMet() {
            }

            const auto compatible = IsControllerCompatible(
-                GetControllerTypeFromIndex(emulated_controllers[index]->currentIndex()),
+                GetControllerTypeFromIndex(emulated_controllers[index]->currentIndex(), index),
                parameters);

            // If any controller is found to be incompatible, return false early.
@@ -422,6 +393,63 @@ void QtControllerSelectorDialog::SetSupportedControllers() {
    }
 }

+void QtControllerSelectorDialog::SetEmulatedControllers(std::size_t player_index) {
+    auto& pairs = index_controller_type_pairs[player_index];
+
+    pairs.clear();
+    emulated_controllers[player_index]->clear();
+
+    pairs.emplace_back(emulated_controllers[player_index]->count(),
+                       Settings::ControllerType::ProController);
+    emulated_controllers[player_index]->addItem(tr("Pro Controller"));
+
+    pairs.emplace_back(emulated_controllers[player_index]->count(),
+                       Settings::ControllerType::DualJoyconDetached);
+    emulated_controllers[player_index]->addItem(tr("Dual Joycons"));
+
+    pairs.emplace_back(emulated_controllers[player_index]->count(),
+                       Settings::ControllerType::LeftJoycon);
+    emulated_controllers[player_index]->addItem(tr("Left Joycon"));
+
+    pairs.emplace_back(emulated_controllers[player_index]->count(),
+                       Settings::ControllerType::RightJoycon);
+    emulated_controllers[player_index]->addItem(tr("Right Joycon"));
+
+    if (player_index == 0) {
+        pairs.emplace_back(emulated_controllers[player_index]->count(),
+                           Settings::ControllerType::Handheld);
+        emulated_controllers[player_index]->addItem(tr("Handheld"));
+    }
+}
+
+Settings::ControllerType QtControllerSelectorDialog::GetControllerTypeFromIndex(
+    int index, std::size_t player_index) const {
+    const auto& pairs = index_controller_type_pairs[player_index];
+
+    const auto it = std::find_if(pairs.begin(), pairs.end(),
+                                 [index](const auto& pair) { return pair.first == index; });
+
+    if (it == pairs.end()) {
+        return Settings::ControllerType::ProController;
+    }
+
+    return it->second;
+}
+
+int QtControllerSelectorDialog::GetIndexFromControllerType(Settings::ControllerType type,
+                                                           std::size_t player_index) const {
+    const auto& pairs = index_controller_type_pairs[player_index];
+
+    const auto it = std::find_if(pairs.begin(), pairs.end(),
+                                 [type](const auto& pair) { return pair.second == type; });
+
+    if (it == pairs.end()) {
+        return 0;
+    }
+
+    return it->first;
+}
+
 void QtControllerSelectorDialog::UpdateControllerIcon(std::size_t player_index) {
    if (!player_groupboxes[player_index]->isChecked()) {
        connected_controller_icons[player_index]->setStyleSheet(QString{});
@@ -430,7 +458,8 @@ void QtControllerSelectorDialog::UpdateControllerIcon(std::size_t player_index)
    }

    const QString stylesheet = [this, player_index] {
-        switch (GetControllerTypeFromIndex(emulated_controllers[player_index]->currentIndex())) {
+        switch (GetControllerTypeFromIndex(emulated_controllers[player_index]->currentIndex(),
+                                           player_index)) {
        case Settings::ControllerType::ProController:
            return QStringLiteral("image: url(:/controller/applet_pro_controller%0); ");
        case Settings::ControllerType::DualJoyconDetached:
@@ -446,6 +475,12 @@ void QtControllerSelectorDialog::UpdateControllerIcon(std::size_t player_index)
        }
    }();

+    if (stylesheet.isEmpty()) {
+        connected_controller_icons[player_index]->setStyleSheet(QString{});
+        player_labels[player_index]->show();
+        return;
+    }
+
    const QString theme = [] {
        if (QIcon::themeName().contains(QStringLiteral("dark"))) {
            return QStringLiteral("_dark");
@@ -463,8 +498,8 @@ void QtControllerSelectorDialog::UpdateControllerIcon(std::size_t player_index)
 void QtControllerSelectorDialog::UpdateControllerState(std::size_t player_index) {
    auto& player = Settings::values.players.GetValue()[player_index];

-    const auto controller_type =
-        GetControllerTypeFromIndex(emulated_controllers[player_index]->currentIndex());
+    const auto controller_type = GetControllerTypeFromIndex(
+        emulated_controllers[player_index]->currentIndex(), player_index);
    const auto player_connected = player_groupboxes[player_index]->isChecked() &&
                                  controller_type != Settings::ControllerType::Handheld;

@@ -507,8 +542,8 @@ void QtControllerSelectorDialog::UpdateControllerState(std::size_t player_index)

 void QtControllerSelectorDialog::UpdateLEDPattern(std::size_t player_index) {
    if (!player_groupboxes[player_index]->isChecked() ||
-        GetControllerTypeFromIndex(emulated_controllers[player_index]->currentIndex()) ==
-            Settings::ControllerType::Handheld) {
+        GetControllerTypeFromIndex(emulated_controllers[player_index]->currentIndex(),
+                                   player_index) == Settings::ControllerType::Handheld) {
        led_patterns_boxes[player_index][0]->setChecked(false);
        led_patterns_boxes[player_index][1]->setChecked(false);
        led_patterns_boxes[player_index][2]->setChecked(false);
--- a/src/yuzu/applets/controller.h
+++ b/src/yuzu/applets/controller.h
@@ -22,6 +22,10 @@ namespace InputCommon {
 class InputSubsystem;
 }

+namespace Settings {
+enum class ControllerType;
+}
+
 namespace Ui {
 class QtControllerSelectorDialog;
 }
@@ -57,6 +61,15 @@ private:
    // Sets the controller icons for "Supported Controller Types".
    void SetSupportedControllers();

+    // Sets the emulated controllers per player.
+    void SetEmulatedControllers(std::size_t player_index);
+
+    // Gets the Controller Type for a given controller combobox index per player.
+    Settings::ControllerType GetControllerTypeFromIndex(int index, std::size_t player_index) const;
+
+    // Gets the controller combobox index for a given Controller Type per player.
+    int GetIndexFromControllerType(Settings::ControllerType type, std::size_t player_index) const;
+
    // Updates the controller icons per player.
    void UpdateControllerIcon(std::size_t player_index);

@@ -114,6 +127,10 @@ private:
    // Comboboxes with a list of emulated controllers per player.
    std::array<QComboBox*, NUM_PLAYERS> emulated_controllers;

+    /// Pairs of emulated controller index and Controller Type enum per player.
+    std::array<std::vector<std::pair<int, Settings::ControllerType>>, NUM_PLAYERS>
+        index_controller_type_pairs;
+
    // Labels representing the number of connected controllers
    // above the "Connected Controllers" checkboxes.
    std::array<QLabel*, NUM_PLAYERS> connected_controller_labels;
--- a/src/yuzu/bootmanager.cpp
+++ b/src/yuzu/bootmanager.cpp
@@ -302,13 +302,19 @@ GRenderWindow::GRenderWindow(GMainWindow* parent, EmuThread* emu_thread_,
    this->setMouseTracking(true);

    connect(this, &GRenderWindow::FirstFrameDisplayed, parent, &GMainWindow::OnLoadComplete);
+    connect(this, &GRenderWindow::ExecuteProgramSignal, parent, &GMainWindow::OnExecuteProgram,
+            Qt::QueuedConnection);
+}
+
+void GRenderWindow::ExecuteProgram(std::size_t program_index) {
+    emit ExecuteProgramSignal(program_index);
 }

 GRenderWindow::~GRenderWindow() {
    input_subsystem->Shutdown();
 }

-void GRenderWindow::PollEvents() {
+void GRenderWindow::OnFrameDisplayed() {
    if (!first_frame) {
        first_frame = true;
        emit FirstFrameDisplayed();
--- a/src/yuzu/bootmanager.h
+++ b/src/yuzu/bootmanager.h
@@ -131,7 +131,7 @@ public:
    ~GRenderWindow() override;

    // EmuWindow implementation.
-    void PollEvents() override;
+    void OnFrameDisplayed() override;
    bool IsShown() const override;
    std::unique_ptr<Core::Frontend::GraphicsContext> CreateSharedContext() const override;

@@ -166,6 +166,12 @@ public:

    std::pair<u32, u32> ScaleTouch(const QPointF& pos) const;

+    /**
+     * Instructs the window to re-launch the application using the specified program_index.
+     * @param program_index Specifies the index within the application of the program to launch.
+     */
+    void ExecuteProgram(std::size_t program_index);
+
 public slots:
    void OnEmulationStarting(EmuThread* emu_thread);
    void OnEmulationStopping();
@@ -175,6 +181,7 @@ signals:
    /// Emitted when the window is closed
    void Closed();
    void FirstFrameDisplayed();
+    void ExecuteProgramSignal(std::size_t program_index);

 private:
    void TouchBeginEvent(const QTouchEvent* event);
--- a/src/yuzu/configuration/configure_input_player.cpp
+++ b/src/yuzu/configuration/configure_input_player.cpp
@@ -27,6 +27,8 @@
 #include "yuzu/configuration/input_profiles.h"
 #include "yuzu/util/limitable_input_dialog.h"

+using namespace Service::HID;
+
 const std::array<std::string, ConfigureInputPlayer::ANALOG_SUB_BUTTONS_NUM>
    ConfigureInputPlayer::analog_sub_buttons{{
        "up",
@@ -47,48 +49,12 @@ void UpdateController(Settings::ControllerType controller_type, std::size_t npad
    }
    Service::SM::ServiceManager& sm = system.ServiceManager();

-    auto& npad =
-        sm.GetService<Service::HID::Hid>("hid")
-            ->GetAppletResource()
-            ->GetController<Service::HID::Controller_NPad>(Service::HID::HidController::NPad);
+    auto& npad = sm.GetService<Hid>("hid")->GetAppletResource()->GetController<Controller_NPad>(
+        HidController::NPad);

    npad.UpdateControllerAt(npad.MapSettingsTypeToNPad(controller_type), npad_index, connected);
 }

-/// Maps the controller type combobox index to Controller Type enum
-constexpr Settings::ControllerType GetControllerTypeFromIndex(int index) {
-    switch (index) {
-    case 0:
-    default:
-        return Settings::ControllerType::ProController;
-    case 1:
-        return Settings::ControllerType::DualJoyconDetached;
-    case 2:
-        return Settings::ControllerType::LeftJoycon;
-    case 3:
-        return Settings::ControllerType::RightJoycon;
-    case 4:
-        return Settings::ControllerType::Handheld;
-    }
-}
-
-/// Maps the Controller Type enum to controller type combobox index
-constexpr int GetIndexFromControllerType(Settings::ControllerType type) {
-    switch (type) {
-    case Settings::ControllerType::ProController:
-    default:
-        return 0;
-    case Settings::ControllerType::DualJoyconDetached:
-        return 1;
-    case Settings::ControllerType::LeftJoycon:
-        return 2;
-    case Settings::ControllerType::RightJoycon:
-        return 3;
-    case Settings::ControllerType::Handheld:
-        return 4;
-    }
-}
-
 QString GetKeyName(int key_code) {
    switch (key_code) {
    case Qt::LeftButton:
@@ -453,18 +419,7 @@ ConfigureInputPlayer::ConfigureInputPlayer(QWidget* parent, std::size_t player_i
    connect(ui->groupConnectedController, &QGroupBox::toggled,
            [this](bool checked) { emit Connected(checked); });

-    // Set up controller type. Only Player 1 can choose Handheld.
-    ui->comboControllerType->clear();
-
-    QStringList controller_types = {
-        tr("Pro Controller"),
-        tr("Dual Joycons"),
-        tr("Left Joycon"),
-        tr("Right Joycon"),
-    };
-
    if (player_index == 0) {
-        controller_types.append(tr("Handheld"));
        connect(ui->comboControllerType, qOverload<int>(&QComboBox::currentIndexChanged),
                [this](int index) {
                    emit HandheldStateChanged(GetControllerTypeFromIndex(index) ==
@@ -480,12 +435,9 @@ ConfigureInputPlayer::ConfigureInputPlayer(QWidget* parent, std::size_t player_i
    if (debug) {
        ui->buttonScreenshot->setEnabled(false);
        ui->buttonHome->setEnabled(false);
-        QStringList debug_controller_types = {
-            tr("Pro Controller"),
-        };
-        ui->comboControllerType->addItems(debug_controller_types);
+        ui->comboControllerType->addItem(tr("Pro Controller"));
    } else {
-        ui->comboControllerType->addItems(controller_types);
+        SetConnectableControllers();
    }

    UpdateControllerIcon();
@@ -667,7 +619,7 @@ void ConfigureInputPlayer::LoadConfiguration() {
        return;
    }

-    ui->comboControllerType->setCurrentIndex(static_cast<int>(player.controller_type));
+    ui->comboControllerType->setCurrentIndex(GetIndexFromControllerType(player.controller_type));
    ui->groupConnectedController->setChecked(
        player.connected ||
        (player_index == 0 && Settings::values.players.GetValue()[HANDHELD_INDEX].connected));
@@ -841,6 +793,82 @@ void ConfigureInputPlayer::UpdateUI() {
    }
 }

+void ConfigureInputPlayer::SetConnectableControllers() {
+    const auto add_controllers = [this](bool enable_all,
+                                        Controller_NPad::NpadStyleSet npad_style_set = {}) {
+        index_controller_type_pairs.clear();
+        ui->comboControllerType->clear();
+
+        if (enable_all || npad_style_set.pro_controller == 1) {
+            index_controller_type_pairs.emplace_back(ui->comboControllerType->count(),
+                                                     Settings::ControllerType::ProController);
+            ui->comboControllerType->addItem(tr("Pro Controller"));
+        }
+
+        if (enable_all || npad_style_set.joycon_dual == 1) {
+            index_controller_type_pairs.emplace_back(ui->comboControllerType->count(),
+                                                     Settings::ControllerType::DualJoyconDetached);
+            ui->comboControllerType->addItem(tr("Dual Joycons"));
+        }
+
+        if (enable_all || npad_style_set.joycon_left == 1) {
+            index_controller_type_pairs.emplace_back(ui->comboControllerType->count(),
+                                                     Settings::ControllerType::LeftJoycon);
+            ui->comboControllerType->addItem(tr("Left Joycon"));
+        }
+
+        if (enable_all || npad_style_set.joycon_right == 1) {
+            index_controller_type_pairs.emplace_back(ui->comboControllerType->count(),
+                                                     Settings::ControllerType::RightJoycon);
+            ui->comboControllerType->addItem(tr("Right Joycon"));
+        }
+
+        if (player_index == 0 && (enable_all || npad_style_set.handheld == 1)) {
+            index_controller_type_pairs.emplace_back(ui->comboControllerType->count(),
+                                                     Settings::ControllerType::Handheld);
+            ui->comboControllerType->addItem(tr("Handheld"));
+        }
+    };
+
+    Core::System& system{Core::System::GetInstance()};
+
+    if (!system.IsPoweredOn()) {
+        add_controllers(true);
+        return;
+    }
+
+    Service::SM::ServiceManager& sm = system.ServiceManager();
+
+    auto& npad = sm.GetService<Hid>("hid")->GetAppletResource()->GetController<Controller_NPad>(
+        HidController::NPad);
+
+    add_controllers(false, npad.GetSupportedStyleSet());
+}
+
+Settings::ControllerType ConfigureInputPlayer::GetControllerTypeFromIndex(int index) const {
+    const auto it =
+        std::find_if(index_controller_type_pairs.begin(), index_controller_type_pairs.end(),
+                     [index](const auto& pair) { return pair.first == index; });
+
+    if (it == index_controller_type_pairs.end()) {
+        return Settings::ControllerType::ProController;
+    }
+
+    return it->second;
+}
+
+int ConfigureInputPlayer::GetIndexFromControllerType(Settings::ControllerType type) const {
+    const auto it =
+        std::find_if(index_controller_type_pairs.begin(), index_controller_type_pairs.end(),
+                     [type](const auto& pair) { return pair.second == type; });
+
+    if (it == index_controller_type_pairs.end()) {
+        return -1;
+    }
+
+    return it->first;
+}
+
 void ConfigureInputPlayer::UpdateInputDevices() {
    input_devices = input_subsystem->GetInputDevices();
    ui->comboDevices->clear();
--- a/src/yuzu/configuration/configure_input_player.h
+++ b/src/yuzu/configuration/configure_input_player.h
@@ -9,6 +9,7 @@
 #include <memory>
 #include <optional>
 #include <string>
+#include <vector>

 #include <QWidget>

@@ -112,6 +113,15 @@ private:
    /// Update UI to reflect current configuration.
    void UpdateUI();

+    /// Sets the available controllers.
+    void SetConnectableControllers();
+
+    /// Gets the Controller Type for a given controller combobox index.
+    Settings::ControllerType GetControllerTypeFromIndex(int index) const;
+
+    /// Gets the controller combobox index for a given Controller Type.
+    int GetIndexFromControllerType(Settings::ControllerType type) const;
+
    /// Update the available input devices.
    void UpdateInputDevices();

@@ -151,6 +161,9 @@ private:
    std::unique_ptr<QTimer> timeout_timer;
    std::unique_ptr<QTimer> poll_timer;

+    /// Stores a pair of "Connected Controllers" combobox index and Controller Type enum.
+    std::vector<std::pair<int, Settings::ControllerType>> index_controller_type_pairs;
+
    static constexpr int PLAYER_COUNT = 8;
    std::array<QCheckBox*, PLAYER_COUNT> player_connected_checkbox;

--- a/src/yuzu/main.cpp
+++ b/src/yuzu/main.cpp
@@ -978,7 +978,7 @@ void GMainWindow::AllowOSSleep() {
 #endif
 }

-bool GMainWindow::LoadROM(const QString& filename) {
+bool GMainWindow::LoadROM(const QString& filename, std::size_t program_index) {
    // Shutdown previous session if the emu thread is still active...
    if (emu_thread != nullptr)
        ShutdownGame();
@@ -1003,7 +1003,8 @@ bool GMainWindow::LoadROM(const QString& filename) {

    system.RegisterHostThread();

-    const Core::System::ResultStatus result{system.Load(*render_window, filename.toStdString())};
+    const Core::System::ResultStatus result{
+        system.Load(*render_window, filename.toStdString(), program_index)};

    const auto drd_callout =
        (UISettings::values.callout_flags & static_cast<u32>(CalloutFlag::DRDDeprecation)) == 0;
@@ -1085,14 +1086,18 @@ void GMainWindow::SelectAndSetCurrentUser() {
    Settings::values.current_user = dialog.GetIndex();
 }

-void GMainWindow::BootGame(const QString& filename) {
+void GMainWindow::BootGame(const QString& filename, std::size_t program_index) {
    LOG_INFO(Frontend, "yuzu starting...");
    StoreRecentFile(filename); // Put the filename on top of the list

    u64 title_id{0};
+
+    last_filename_booted = filename;
+
    auto& system = Core::System::GetInstance();
    const auto v_file = Core::GetGameFileFromPath(vfs, filename.toUtf8().constData());
-    const auto loader = Loader::GetLoader(system, v_file);
+    const auto loader = Loader::GetLoader(system, v_file, program_index);
+
    if (!(loader == nullptr || loader->ReadProgramId(title_id) != Loader::ResultStatus::Success)) {
        // Load per game settings
        Config per_game_config(fmt::format("{:016X}", title_id), Config::ConfigType::PerGameConfig);
@@ -1106,7 +1111,7 @@ void GMainWindow::BootGame(const QString& filename) {
        SelectAndSetCurrentUser();
    }

-    if (!LoadROM(filename))
+    if (!LoadROM(filename, program_index))
        return;

    // Create and start the emulation thread
@@ -1114,6 +1119,10 @@ void GMainWindow::BootGame(const QString& filename) {
    emit EmulationStarting(emu_thread.get());
    emu_thread->start();

+    // Register an ExecuteProgram callback such that Core can execute a sub-program
+    system.RegisterExecuteProgramCallback(
+        [this](std::size_t program_index) { render_window->ExecuteProgram(program_index); });
+
    connect(render_window, &GRenderWindow::Closed, this, &GMainWindow::OnStopGame);
    // BlockingQueuedConnection is important here, it makes sure we've finished refreshing our views
    // before the CPU continues
@@ -2136,6 +2145,11 @@ void GMainWindow::OnLoadComplete() {
    loading_screen->OnLoadComplete();
 }

+void GMainWindow::OnExecuteProgram(std::size_t program_index) {
+    ShutdownGame();
+    BootGame(last_filename_booted, program_index);
+}
+
 void GMainWindow::ErrorDisplayDisplayError(QString body) {
    QMessageBox::critical(this, tr("Error Display"), body);
    emit ErrorDisplayFinished();
--- a/src/yuzu/main.h
+++ b/src/yuzu/main.h
@@ -131,6 +131,7 @@ signals:

 public slots:
    void OnLoadComplete();
+    void OnExecuteProgram(std::size_t program_index);
    void ControllerSelectorReconfigureControllers(
        const Core::Frontend::ControllerParameters& parameters);
    void ErrorDisplayDisplayError(QString body);
@@ -154,8 +155,8 @@ private:
    void PreventOSSleep();
    void AllowOSSleep();

-    bool LoadROM(const QString& filename);
-    void BootGame(const QString& filename);
+    bool LoadROM(const QString& filename, std::size_t program_index);
+    void BootGame(const QString& filename, std::size_t program_index = 0);
    void ShutdownGame();

    void ShowTelemetryCallout();
@@ -317,6 +318,9 @@ private:
    // Install progress dialog
    QProgressDialog* install_progress;

+    // Last game booted, used for multi-process apps
+    QString last_filename_booted;
+
 protected:
    void dropEvent(QDropEvent* event) override;
    void dragEnterEvent(QDragEnterEvent* event) override;
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp
@@ -121,62 +121,64 @@ void EmuWindow_SDL2::Fullscreen() {
    SDL_MaximizeWindow(render_window);
 }

-void EmuWindow_SDL2::PollEvents() {
+void EmuWindow_SDL2::WaitEvent() {
+    // Called on main thread
    SDL_Event event;

-    // SDL_PollEvent returns 0 when there are no more events in the event queue
-    while (SDL_PollEvent(&event)) {
-        switch (event.type) {
-        case SDL_WINDOWEVENT:
-            switch (event.window.event) {
-            case SDL_WINDOWEVENT_SIZE_CHANGED:
-            case SDL_WINDOWEVENT_RESIZED:
-            case SDL_WINDOWEVENT_MAXIMIZED:
-            case SDL_WINDOWEVENT_RESTORED:
-                OnResize();
-                break;
-            case SDL_WINDOWEVENT_MINIMIZED:
-            case SDL_WINDOWEVENT_EXPOSED:
-                is_shown = event.window.event == SDL_WINDOWEVENT_EXPOSED;
-                OnResize();
-                break;
-            case SDL_WINDOWEVENT_CLOSE:
-                is_open = false;
-                break;
-            }
+    if (!SDL_WaitEvent(&event)) {
+        LOG_CRITICAL(Frontend, "SDL_WaitEvent failed: {}", SDL_GetError());
+        exit(1);
+    }
+
+    switch (event.type) {
+    case SDL_WINDOWEVENT:
+        switch (event.window.event) {
+        case SDL_WINDOWEVENT_SIZE_CHANGED:
+        case SDL_WINDOWEVENT_RESIZED:
+        case SDL_WINDOWEVENT_MAXIMIZED:
+        case SDL_WINDOWEVENT_RESTORED:
+            OnResize();
            break;
-        case SDL_KEYDOWN:
-        case SDL_KEYUP:
-            OnKeyEvent(static_cast<int>(event.key.keysym.scancode), event.key.state);
+        case SDL_WINDOWEVENT_MINIMIZED:
+        case SDL_WINDOWEVENT_EXPOSED:
+            is_shown = event.window.event == SDL_WINDOWEVENT_EXPOSED;
+            OnResize();
            break;
-        case SDL_MOUSEMOTION:
-            // ignore if it came from touch
-            if (event.button.which != SDL_TOUCH_MOUSEID)
-                OnMouseMotion(event.motion.x, event.motion.y);
-            break;
-        case SDL_MOUSEBUTTONDOWN:
-        case SDL_MOUSEBUTTONUP:
-            // ignore if it came from touch
-            if (event.button.which != SDL_TOUCH_MOUSEID) {
-                OnMouseButton(event.button.button, event.button.state, event.button.x,
-                              event.button.y);
-            }
-            break;
-        case SDL_FINGERDOWN:
-            OnFingerDown(event.tfinger.x, event.tfinger.y);
-            break;
-        case SDL_FINGERMOTION:
-            OnFingerMotion(event.tfinger.x, event.tfinger.y);
-            break;
-        case SDL_FINGERUP:
-            OnFingerUp();
-            break;
-        case SDL_QUIT:
+        case SDL_WINDOWEVENT_CLOSE:
            is_open = false;
            break;
-        default:
-            break;
        }
+        break;
+    case SDL_KEYDOWN:
+    case SDL_KEYUP:
+        OnKeyEvent(static_cast<int>(event.key.keysym.scancode), event.key.state);
+        break;
+    case SDL_MOUSEMOTION:
+        // ignore if it came from touch
+        if (event.button.which != SDL_TOUCH_MOUSEID)
+            OnMouseMotion(event.motion.x, event.motion.y);
+        break;
+    case SDL_MOUSEBUTTONDOWN:
+    case SDL_MOUSEBUTTONUP:
+        // ignore if it came from touch
+        if (event.button.which != SDL_TOUCH_MOUSEID) {
+            OnMouseButton(event.button.button, event.button.state, event.button.x, event.button.y);
+        }
+        break;
+    case SDL_FINGERDOWN:
+        OnFingerDown(event.tfinger.x, event.tfinger.y);
+        break;
+    case SDL_FINGERMOTION:
+        OnFingerMotion(event.tfinger.x, event.tfinger.y);
+        break;
+    case SDL_FINGERUP:
+        OnFingerUp();
+        break;
+    case SDL_QUIT:
+        is_open = false;
+        break;
+    default:
+        break;
    }

    const u32 current_time = SDL_GetTicks();
--- a/src/yuzu_cmd/emu_window/emu_window_sdl2.h
+++ b/src/yuzu_cmd/emu_window/emu_window_sdl2.h
@@ -23,38 +23,38 @@ public:
    explicit EmuWindow_SDL2(InputCommon::InputSubsystem* input_subsystem);
    ~EmuWindow_SDL2();

-    /// Polls window events
-    void PollEvents() override;
-
    /// Whether the window is still open, and a close request hasn't yet been sent
    bool IsOpen() const;

    /// Returns if window is shown (not minimized)
    bool IsShown() const override;

+    /// Wait for the next event on the main thread.
+    void WaitEvent();
+
 protected:
-    /// Called by PollEvents when a key is pressed or released.
+    /// Called by WaitEvent when a key is pressed or released.
    void OnKeyEvent(int key, u8 state);

-    /// Called by PollEvents when the mouse moves.
+    /// Called by WaitEvent when the mouse moves.
    void OnMouseMotion(s32 x, s32 y);

-    /// Called by PollEvents when a mouse button is pressed or released
+    /// Called by WaitEvent when a mouse button is pressed or released
    void OnMouseButton(u32 button, u8 state, s32 x, s32 y);

    /// Translates pixel position (0..1) to pixel positions
    std::pair<unsigned, unsigned> TouchToPixelPos(float touch_x, float touch_y) const;

-    /// Called by PollEvents when a finger starts touching the touchscreen
+    /// Called by WaitEvent when a finger starts touching the touchscreen
    void OnFingerDown(float x, float y);

-    /// Called by PollEvents when a finger moves while touching the touchscreen
+    /// Called by WaitEvent when a finger moves while touching the touchscreen
    void OnFingerMotion(float x, float y);

-    /// Called by PollEvents when a finger stops touching the touchscreen
+    /// Called by WaitEvent when a finger stops touching the touchscreen
    void OnFingerUp();

-    /// Called by PollEvents when any event that may cause the window to be resized occurs
+    /// Called by WaitEvent when any event that may cause the window to be resized occurs
    void OnResize();

    /// Called when user passes the fullscreen parameter flag
--- a/src/yuzu_cmd/yuzu.cpp
+++ b/src/yuzu_cmd/yuzu.cpp
@@ -240,11 +240,11 @@ int main(int argc, char** argv) {
        system.CurrentProcess()->GetTitleID(), false,
        [](VideoCore::LoadCallbackStage, size_t value, size_t total) {});

-    system.Run();
+    void(system.Run());
    while (emu_window->IsOpen()) {
-        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+        emu_window->WaitEvent();
    }
-    system.Pause();
+    void(system.Pause());
    system.Shutdown();

    detached_tasks.WaitForAllTasks();
--- a/src/yuzu_tester/emu_window/emu_window_sdl2_hide.cpp
+++ b/src/yuzu_tester/emu_window/emu_window_sdl2_hide.cpp
@@ -109,8 +109,6 @@ EmuWindow_SDL2_Hide::~EmuWindow_SDL2_Hide() {
    SDL_Quit();
 }

-void EmuWindow_SDL2_Hide::PollEvents() {}
-
 bool EmuWindow_SDL2_Hide::IsShown() const {
    return false;
 }
--- a/src/yuzu_tester/emu_window/emu_window_sdl2_hide.h
+++ b/src/yuzu_tester/emu_window/emu_window_sdl2_hide.h
@@ -17,9 +17,6 @@ public:
    explicit EmuWindow_SDL2_Hide();
    ~EmuWindow_SDL2_Hide();

-    /// Polls window events
-    void PollEvents() override;
-
    /// Whether the screen is being shown or not.
    bool IsShown() const override;

--- a/src/yuzu_tester/yuzu.cpp
+++ b/src/yuzu_tester/yuzu.cpp
@@ -256,11 +256,11 @@ int main(int argc, char** argv) {

    system.GPU().Start();

-    system.Run();
+    void(system.Run());
    while (!finished) {
        std::this_thread::sleep_for(std::chrono::milliseconds(1));
    }
-    system.Pause();
+    void(system.Pause());

    detached_tasks.WaitForAllTasks();
    return return_value;
Author	SHA1	Message	Date
comex	32f3b6b865	CMakeLists: disable -Winvalid-offsetof This Clang warning complains when offsetof is used on a non-standard-layout type (i.e. any class using various C++ features), even though it works fine (and is not undefined behavior as of C++17).	2020-11-27 17:42:02 -05:00
bunnei	322349e8cc	Merge pull request #4975 from comex/invalid-syncpoint-id nvdrv, video_core: Don't index out of bounds when given invalid syncpoint ID	2020-11-26 01:27:24 -08:00
bunnei	ebcee03b0c	Merge pull request #4981 from ogniK5377/ioctl-ctrl nvservices: Reintroducee IoctlCtrl	2020-11-25 21:48:57 -08:00
LC	b70751ccb9	Merge pull request #5003 from jbeich/clang input_common: unbreak -Werror with Clang	2020-11-25 21:09:13 -05:00
Jan Beich	e48e9a406c	input_common: ignore some Clang warnings after `5c4774e8ce` error: unknown warning option '-Werror=unused-but-set-parameter'; did you mean '-Werror=unused-parameter'? [-Werror,-Wunknown-warning-option] error: unknown warning option '-Werror=unused-but-set-variable'; did you mean '-Werror=unused-const-variable'? [-Werror,-Wunknown-warning-option]	2020-11-25 23:57:11 +00:00
Rodrigo Locatti	0e15c68f54	Merge pull request #4976 from comex/poll-events Overhaul EmuWindow::PollEvents to fix yuzu-cmd calling SDL_PollEvents off main thread	2020-11-25 20:44:53 -03:00
Rodrigo Locatti	b834c21894	Merge pull request #4946 from ameerj/alpha-test renderer_vulkan: Implement alpha testing	2020-11-25 18:48:34 -03:00
ameerj	d52ee6d0a7	cleanup unneeded comments and newlines	2020-11-25 14:46:08 -05:00
bunnei	dcfa1992ea	Merge pull request #4959 from Morph1984/emulated-controller-styleset configure_input_player: Use the NpadStyleSet to limit the available controllers shown	2020-11-25 11:20:46 -08:00
bunnei	b7f1095980	Merge pull request #4932 from ogniK5377/misc-audio audren: Make use of nodiscard, rework downmixing, release all buffers	2020-11-25 10:50:52 -08:00
bunnei	9aeada734d	Merge pull request #4978 from bunnei/shutdown-crash core: cpu_manager: Fix shutdown crash when closing before emulation starts.	2020-11-25 01:20:01 -08:00
ameerj	e87670ee48	Refactor MaxwellToSpirvComparison. Use Common::BitCast Co-Authored-By: Rodrigo Locatti <reinuseslisp@airmail.cc>	2020-11-25 00:33:20 -05:00
ameerj	1dbf71ceb3	Address PR feedback from Rein	2020-11-24 22:46:45 -05:00
ameerj	9014861858	vulkan_renderer: Alpha Test Culling Implementation Used by various textures in many titles, e.g. SSBU menu.	2020-11-24 22:46:45 -05:00
bunnei	d1da7eb119	Merge pull request #4905 from german77/AnalogFromButton Allow to dial any angle with digital joystick	2020-11-24 16:04:55 -08:00
bunnei	0832da3e40	Merge pull request #4799 from bunnei/execute-program core: Refactor loader and implement ExecuteProgram	2020-11-24 15:27:22 -08:00
bunnei	3359e5ab70	core: cpu_manager: Fix shutdown crash when closing before emulation starts.	2020-11-24 15:25:39 -08:00
bunnei	4fbe4da911	frontend: yuzu (qt): Register a callback for ExecuteProgram.	2020-11-24 15:18:29 -08:00
bunnei	4fb5ca80c0	service: am: Implement ExecuteProgram and required stubs. - This is used by Super Mario 3D All-Stars.	2020-11-24 15:17:43 -08:00
bunnei	5f75d97125	core: loader: Implement support for loading indexed programs.	2020-11-24 15:16:24 -08:00
bunnei	7791cc8c2e	hle: services: Fix a crash with improper NVFlinger lifetime management. (#4977 ) * hle: services: Fix a crash with improper NVFlinger lifetime management. - This crash would happen when attempting to shutdown yuzu early on in boot.	2020-11-24 14:31:58 -08:00
comex	e8b2fd21d8	nvdrv, video_core: Don't index out of bounds when given invalid syncpoint ID - Use .at() instead of raw indexing when dealing with untrusted indices. - For the special case of WaitFence with syncpoint id UINT32_MAX, instead of crashing, log an error and ignore. This is what I get when running Super Mario Maker 2.	2020-11-24 12:59:41 -05:00
Rodrigo Locatti	fbda5e9ec9	Merge pull request #3681 from lioncash/component decoder/image: Fix incorrect G24R8 component sizes in GetComponentSize()	2020-11-24 04:38:03 -03:00
Rodrigo Locatti	410ed82922	Merge pull request #4942 from lioncash/system core: Make use of [[nodiscard]] with the System class	2020-11-24 04:26:59 -03:00
Rodrigo Locatti	7afb7a9494	Merge pull request #4972 from lioncash/unused4 svc: Remove unnecessary [[maybe_unused]] tag	2020-11-24 04:25:59 -03:00
bunnei	6694e11303	input_common: Fix typo in gc_poller.cpp with [[maybe_unused]].	2020-11-23 21:42:06 -08:00
Chloe Marcec	ab25d1fe9a	nvservices: Reintroducee IoctlCtrl Fixes regression caused by #4907 which caused games like Breath of the Wild 1.0.0 not to boot.	2020-11-24 16:40:23 +11:00
bunnei	5ec6a265bf	Merge pull request #4980 from bunnei/error-fixup input_common: Add more missing [[maybe_unused]] from #4927.	2020-11-23 21:00:59 -08:00
comex	994f497781	Overhaul EmuWindow::PollEvents to fix yuzu-cmd calling SDL_PollEvents off main thread EmuWindow::PollEvents was called from the GPU thread (or the CPU thread in sync-GPU mode) when swapping buffers. It had three implementations: - In GRenderWindow, it didn't actually poll events, just set a flag and emit a signal to indicate that a frame was displayed. - In EmuWindow_SDL2_Hide, it did nothing. - In EmuWindow_SDL2, it did call SDL_PollEvents, but this is wrong because SDL_PollEvents is supposed to be called on the thread that set up video - in this case, the main thread, which was sleeping in a busyloop (regardless of whether sync-GPU was enabled). On macOS this causes a crash. To fix this: - Rename EmuWindow::PollEvents to OnFrameDisplayed, and give it a default implementation that does nothing. - In EmuWindow_SDL2, do not override OnFrameDisplayed, but instead have the main thread call SDL_WaitEvent in a loop.	2020-11-23 17:58:49 -05:00
Lioncash	874be0e3e1	svc: Remove unnecessary [[maybe_unused]] tag The parameter is used in this function, so this suppression isn't necessary.	2020-11-23 10:17:20 -05:00
Morph	8758378dc4	applets/controller: Use a pair of emulated controller index to controller type	2020-11-20 22:22:22 -05:00
Morph	102630f2b2	configure_input_player: Use the npad style set to show the available controllers This will reduce the likelihood of an invalid controller type to be set within a game	2020-11-20 22:22:22 -05:00
Lioncash	aaf262bfed	core: Remove unused private Init function for the System class This isn't used, so it can be removed.	2020-11-18 02:09:08 -05:00
Lioncash	bcaadac22c	core: Make use of [[nodiscard]] with the System class Given this is a central class, we should flag cases where the return value of some functions not being used is likely a bug.	2020-11-18 02:06:44 -05:00
Chloe Marcec	908d3c5679	Addressed changes	2020-11-17 15:40:19 +11:00
Chloe Marcec	9a4beac95a	audren: Make use of nodiscard, rework downmixing, release all buffers Preliminary work for upmixing & general cleanup. Fixes basic issues in games such as Shovel Knight and slightly improves the LEGO games. Upmixing stitll needs to be implemented. Audio levels in a few games will be fixed as we now use the downmix coefficients when possible instead of supplying our own	2020-11-17 14:14:29 +11:00
german	f5110340e6	fix minor clang error	2020-11-10 10:38:15 -06:00
german	70df449d0a	Allow to dial any angle with digital joystick	2020-11-08 09:15:33 -06:00
Lioncash	24620bc4ea	decode/image: Fix typo in assert in GetComponentSize()	2020-04-15 22:29:51 -04:00
Lioncash	b178c9a349	decoder/image: Fix incorrect G24R8 component sizes in GetComponentSize() The components' sizes were mismatched. This corrects that.	2020-04-15 22:10:44 -04:00