move sunshine to src

- this will allow for common cpp workflow files within org

src/platform/common.h

@@ -0,0 +1,315 @@
+//
+// Created by loki on 6/21/19.
+//
+
+#ifndef SUNSHINE_COMMON_H
+#define SUNSHINE_COMMON_H
+
+#include <bitset>
+#include <filesystem>
+#include <functional>
+#include <mutex>
+#include <string>
+
+#include "src/thread_safe.h"
+#include "src/utility.h"
+
+struct sockaddr;
+struct AVFrame;
+
+namespace platf {
+constexpr auto MAX_GAMEPADS = 32;
+
+constexpr std::uint16_t DPAD_UP      = 0x0001;
+constexpr std::uint16_t DPAD_DOWN    = 0x0002;
+constexpr std::uint16_t DPAD_LEFT    = 0x0004;
+constexpr std::uint16_t DPAD_RIGHT   = 0x0008;
+constexpr std::uint16_t START        = 0x0010;
+constexpr std::uint16_t BACK         = 0x0020;
+constexpr std::uint16_t LEFT_STICK   = 0x0040;
+constexpr std::uint16_t RIGHT_STICK  = 0x0080;
+constexpr std::uint16_t LEFT_BUTTON  = 0x0100;
+constexpr std::uint16_t RIGHT_BUTTON = 0x0200;
+constexpr std::uint16_t HOME         = 0x0400;
+constexpr std::uint16_t A            = 0x1000;
+constexpr std::uint16_t B            = 0x2000;
+constexpr std::uint16_t X            = 0x4000;
+constexpr std::uint16_t Y            = 0x8000;
+
+struct rumble_t {
+  KITTY_DEFAULT_CONSTR(rumble_t)
+
+  rumble_t(std::uint16_t id, std::uint16_t lowfreq, std::uint16_t highfreq)
+      : id { id }, lowfreq { lowfreq }, highfreq { highfreq } {}
+
+  std::uint16_t id;
+  std::uint16_t lowfreq;
+  std::uint16_t highfreq;
+};
+using rumble_queue_t = safe::mail_raw_t::queue_t<rumble_t>;
+
+namespace speaker {
+enum speaker_e {
+  FRONT_LEFT,
+  FRONT_RIGHT,
+  FRONT_CENTER,
+  LOW_FREQUENCY,
+  BACK_LEFT,
+  BACK_RIGHT,
+  SIDE_LEFT,
+  SIDE_RIGHT,
+  MAX_SPEAKERS,
+};
+
+constexpr std::uint8_t map_stereo[] {
+  FRONT_LEFT, FRONT_RIGHT
+};
+constexpr std::uint8_t map_surround51[] {
+  FRONT_LEFT,
+  FRONT_RIGHT,
+  FRONT_CENTER,
+  LOW_FREQUENCY,
+  BACK_LEFT,
+  BACK_RIGHT,
+};
+constexpr std::uint8_t map_surround71[] {
+  FRONT_LEFT,
+  FRONT_RIGHT,
+  FRONT_CENTER,
+  LOW_FREQUENCY,
+  LOW_FREQUENCY,
+  BACK_LEFT,
+  BACK_RIGHT,
+  SIDE_LEFT,
+  SIDE_RIGHT,
+};
+} // namespace speaker
+
+enum class mem_type_e {
+  system,
+  vaapi,
+  dxgi,
+  cuda,
+  unknown
+};
+
+enum class pix_fmt_e {
+  yuv420p,
+  yuv420p10,
+  nv12,
+  p010,
+  unknown
+};
+
+inline std::string_view from_pix_fmt(pix_fmt_e pix_fmt) {
+  using namespace std::literals;
+#define _CONVERT(x)  \
+  case pix_fmt_e::x: \
+    return #x##sv
+  switch(pix_fmt) {
+    _CONVERT(yuv420p);
+    _CONVERT(yuv420p10);
+    _CONVERT(nv12);
+    _CONVERT(p010);
+    _CONVERT(unknown);
+  }
+#undef _CONVERT
+
+  return "unknown"sv;
+}
+
+// Dimensions for touchscreen input
+struct touch_port_t {
+  int offset_x, offset_y;
+  int width, height;
+};
+
+struct gamepad_state_t {
+  std::uint16_t buttonFlags;
+  std::uint8_t lt;
+  std::uint8_t rt;
+  std::int16_t lsX;
+  std::int16_t lsY;
+  std::int16_t rsX;
+  std::int16_t rsY;
+};
+
+class deinit_t {
+public:
+  virtual ~deinit_t() = default;
+};
+
+struct img_t {
+public:
+  img_t() = default;
+
+  img_t(img_t &&)      = delete;
+  img_t(const img_t &) = delete;
+  img_t &operator=(img_t &&) = delete;
+  img_t &operator=(const img_t &) = delete;
+
+  std::uint8_t *data {};
+  std::int32_t width {};
+  std::int32_t height {};
+  std::int32_t pixel_pitch {};
+  std::int32_t row_pitch {};
+
+  virtual ~img_t() = default;
+};
+
+struct sink_t {
+  // Play on host PC
+  std::string host;
+
+  // On macOS and Windows, it is not possible to create a virtual sink
+  // Therefore, it is optional
+  struct null_t {
+    std::string stereo;
+    std::string surround51;
+    std::string surround71;
+  };
+  std::optional<null_t> null;
+};
+
+struct hwdevice_t {
+  void *data {};
+  AVFrame *frame {};
+
+  virtual int convert(platf::img_t &img) {
+    return -1;
+  }
+
+  /**
+   * implementations must take ownership of 'frame'
+   */
+  virtual int set_frame(AVFrame *frame) {
+    std::abort(); // ^ This function must never be called
+    return -1;
+  };
+
+  virtual void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) {};
+
+  virtual ~hwdevice_t() = default;
+};
+
+enum class capture_e : int {
+  ok,
+  reinit,
+  timeout,
+  error
+};
+
+class display_t {
+public:
+  /**
+   * When display has a new image ready, this callback will be called with the new image.
+   * 
+   * On Break Request -->
+   *    Returns nullptr
+   * 
+   * On Success -->
+   *    Returns the image object that should be filled next.
+   *    This may or may not be the image send with the callback
+   */
+  using snapshot_cb_t = std::function<std::shared_ptr<img_t>(std::shared_ptr<img_t> &img)>;
+
+  display_t() noexcept : offset_x { 0 }, offset_y { 0 } {}
+
+  /**
+   * snapshot_cb --> the callback
+   * std::shared_ptr<img_t> img --> The first image to use
+   * bool *cursor --> A pointer to the flag that indicates wether the cursor should be captured as well
+   * 
+   * Returns either:
+   *    capture_e::ok when stopping
+   *    capture_e::error on error
+   *    capture_e::reinit when need of reinitialization
+   */
+  virtual capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) = 0;
+
+  virtual std::shared_ptr<img_t> alloc_img() = 0;
+
+  virtual int dummy_img(img_t *img) = 0;
+
+  virtual std::shared_ptr<hwdevice_t> make_hwdevice(pix_fmt_e pix_fmt) {
+    return std::make_shared<hwdevice_t>();
+  }
+
+  virtual ~display_t() = default;
+
+  // Offsets for when streaming a specific monitor. By default, they are 0.
+  int offset_x, offset_y;
+  int env_width, env_height;
+
+  int width, height;
+};
+
+class mic_t {
+public:
+  virtual capture_e sample(std::vector<std::int16_t> &frame_buffer) = 0;
+
+  virtual ~mic_t() = default;
+};
+
+class audio_control_t {
+public:
+  virtual int set_sink(const std::string &sink) = 0;
+
+  virtual std::unique_ptr<mic_t> microphone(const std::uint8_t *mapping, int channels, std::uint32_t sample_rate, std::uint32_t frame_size) = 0;
+
+  virtual std::optional<sink_t> sink_info() = 0;
+
+  virtual ~audio_control_t() = default;
+};
+
+void freeInput(void *);
+
+using input_t = util::safe_ptr<void, freeInput>;
+
+std::filesystem::path appdata();
+
+std::string get_mac_address(const std::string_view &address);
+
+std::string from_sockaddr(const sockaddr *const);
+std::pair<std::uint16_t, std::string> from_sockaddr_ex(const sockaddr *const);
+
+std::unique_ptr<audio_control_t> audio_control();
+
+/**
+ * display_name --> The name of the monitor that SHOULD be displayed
+ *    If display_name is empty --> Use the first monitor that's compatible you can find
+ *    If you require to use this parameter in a seperate thread --> make a copy of it.
+ * 
+ * framerate --> The peak number of images per second
+ * 
+ * Returns display_t based on hwdevice_type
+ */
+std::shared_ptr<display_t> display(mem_type_e hwdevice_type, const std::string &display_name, int framerate);
+
+// A list of names of displays accepted as display_name with the mem_type_e
+std::vector<std::string> display_names(mem_type_e hwdevice_type);
+
+input_t input();
+void move_mouse(input_t &input, int deltaX, int deltaY);
+void abs_mouse(input_t &input, const touch_port_t &touch_port, float x, float y);
+void button_mouse(input_t &input, int button, bool release);
+void scroll(input_t &input, int distance);
+void keyboard(input_t &input, uint16_t modcode, bool release);
+void gamepad(input_t &input, int nr, const gamepad_state_t &gamepad_state);
+
+int alloc_gamepad(input_t &input, int nr, rumble_queue_t rumble_queue);
+void free_gamepad(input_t &input, int nr);
+
+#define SERVICE_NAME "Sunshine"
+#define SERVICE_TYPE "_nvstream._tcp"
+
+namespace publish {
+[[nodiscard]] std::unique_ptr<deinit_t> start();
+}
+
+[[nodiscard]] std::unique_ptr<deinit_t> init();
+
+std::vector<std::string_view> &supported_gamepads();
+} // namespace platf
+
+#endif //SUNSHINE_COMMON_H

src/platform/linux/audio.cpp

@@ -0,0 +1,506 @@
+//
+// Created by loki on 5/16/21.
+//
+#include <bitset>
+#include <sstream>
+
+#include <boost/regex.hpp>
+
+#include <pulse/error.h>
+#include <pulse/pulseaudio.h>
+#include <pulse/simple.h>
+
+#include "src/platform/common.h"
+
+#include "src/config.h"
+#include "src/main.h"
+#include "src/thread_safe.h"
+
+namespace platf {
+using namespace std::literals;
+
+constexpr pa_channel_position_t position_mapping[] {
+  PA_CHANNEL_POSITION_FRONT_LEFT,
+  PA_CHANNEL_POSITION_FRONT_RIGHT,
+  PA_CHANNEL_POSITION_FRONT_CENTER,
+  PA_CHANNEL_POSITION_LFE,
+  PA_CHANNEL_POSITION_REAR_LEFT,
+  PA_CHANNEL_POSITION_REAR_RIGHT,
+  PA_CHANNEL_POSITION_SIDE_LEFT,
+  PA_CHANNEL_POSITION_SIDE_RIGHT,
+};
+
+std::string to_string(const char *name, const std::uint8_t *mapping, int channels) {
+  std::stringstream ss;
+
+  ss << "rate=48000 sink_name="sv << name << " format=s16le channels="sv << channels << " channel_map="sv;
+  std::for_each_n(mapping, channels - 1, [&ss](std::uint8_t pos) {
+    ss << pa_channel_position_to_string(position_mapping[pos]) << ',';
+  });
+
+  ss << pa_channel_position_to_string(position_mapping[mapping[channels - 1]]);
+
+  ss << " sink_properties=device.description="sv << name;
+  auto result = ss.str();
+
+  BOOST_LOG(debug) << "null-sink args: "sv << result;
+  return result;
+}
+
+struct mic_attr_t : public mic_t {
+  util::safe_ptr<pa_simple, pa_simple_free> mic;
+
+  capture_e sample(std::vector<std::int16_t> &sample_buf) override {
+    auto sample_size = sample_buf.size();
+
+    auto buf = sample_buf.data();
+    int status;
+    if(pa_simple_read(mic.get(), buf, sample_size * 2, &status)) {
+      BOOST_LOG(error) << "pa_simple_read() failed: "sv << pa_strerror(status);
+
+      return capture_e::error;
+    }
+
+    return capture_e::ok;
+  }
+};
+
+std::unique_ptr<mic_t> microphone(const std::uint8_t *mapping, int channels, std::uint32_t sample_rate, std::uint32_t frame_size, std::string source_name) {
+  auto mic = std::make_unique<mic_attr_t>();
+
+  pa_sample_spec ss { PA_SAMPLE_S16LE, sample_rate, (std::uint8_t)channels };
+  pa_channel_map pa_map;
+
+  pa_map.channels = channels;
+  std::for_each_n(pa_map.map, pa_map.channels, [mapping](auto &channel) mutable {
+    channel = position_mapping[*mapping++];
+  });
+
+  pa_buffer_attr pa_attr = {};
+  pa_attr.maxlength      = frame_size * 8;
+
+  int status;
+
+  mic->mic.reset(
+    pa_simple_new(nullptr, "sunshine",
+      pa_stream_direction_t::PA_STREAM_RECORD, source_name.c_str(),
+      "sunshine-record", &ss, &pa_map, &pa_attr, &status));
+
+  if(!mic->mic) {
+    auto err_str = pa_strerror(status);
+    BOOST_LOG(error) << "pa_simple_new() failed: "sv << err_str;
+
+    log_flush();
+    std::abort();
+  }
+
+  return mic;
+}
+
+namespace pa {
+template<bool B, class T>
+struct add_const_helper;
+
+template<class T>
+struct add_const_helper<true, T> {
+  using type = const std::remove_pointer_t<T> *;
+};
+
+template<class T>
+struct add_const_helper<false, T> {
+  using type = const T *;
+};
+
+template<class T>
+using add_const_t = typename add_const_helper<std::is_pointer_v<T>, T>::type;
+
+template<class T>
+void pa_free(T *p) {
+  pa_xfree(p);
+}
+using ctx_t    = util::safe_ptr<pa_context, pa_context_unref>;
+using loop_t   = util::safe_ptr<pa_mainloop, pa_mainloop_free>;
+using op_t     = util::safe_ptr<pa_operation, pa_operation_unref>;
+using string_t = util::safe_ptr<char, pa_free<char>>;
+
+template<class T>
+using cb_simple_t = std::function<void(ctx_t::pointer, add_const_t<T> i)>;
+
+template<class T>
+void cb(ctx_t::pointer ctx, add_const_t<T> i, void *userdata) {
+  auto &f = *(cb_simple_t<T> *)userdata;
+
+  // Cannot similarly filter on eol here. Unless reported otherwise assume
+  // we have no need for special filtering like cb?
+  f(ctx, i);
+}
+
+template<class T>
+using cb_t = std::function<void(ctx_t::pointer, add_const_t<T> i, int eol)>;
+
+template<class T>
+void cb(ctx_t::pointer ctx, add_const_t<T> i, int eol, void *userdata) {
+  auto &f = *(cb_t<T> *)userdata;
+
+  // For some reason, pulseaudio calls this callback after disconnecting
+  if(i && eol) {
+    return;
+  }
+
+  f(ctx, i, eol);
+}
+
+void cb_i(ctx_t::pointer ctx, std::uint32_t i, void *userdata) {
+  auto alarm = (safe::alarm_raw_t<int> *)userdata;
+
+  alarm->ring(i);
+}
+
+void ctx_state_cb(ctx_t::pointer ctx, void *userdata) {
+  auto &f = *(std::function<void(ctx_t::pointer)> *)userdata;
+
+  f(ctx);
+}
+
+void success_cb(ctx_t::pointer ctx, int status, void *userdata) {
+  assert(userdata != nullptr);
+
+  auto alarm = (safe::alarm_raw_t<int> *)userdata;
+  alarm->ring(status ? 0 : 1);
+}
+
+class server_t : public audio_control_t {
+  enum ctx_event_e : int {
+    ready,
+    terminated,
+    failed
+  };
+
+public:
+  loop_t loop;
+  ctx_t ctx;
+  std::string requested_sink;
+
+  struct {
+    std::uint32_t stereo     = PA_INVALID_INDEX;
+    std::uint32_t surround51 = PA_INVALID_INDEX;
+    std::uint32_t surround71 = PA_INVALID_INDEX;
+  } index;
+
+  std::unique_ptr<safe::event_t<ctx_event_e>> events;
+  std::unique_ptr<std::function<void(ctx_t::pointer)>> events_cb;
+
+  std::thread worker;
+  int init() {
+    events = std::make_unique<safe::event_t<ctx_event_e>>();
+    loop.reset(pa_mainloop_new());
+    ctx.reset(pa_context_new(pa_mainloop_get_api(loop.get()), "sunshine"));
+
+    events_cb = std::make_unique<std::function<void(ctx_t::pointer)>>([this](ctx_t::pointer ctx) {
+      switch(pa_context_get_state(ctx)) {
+      case PA_CONTEXT_READY:
+        events->raise(ready);
+        break;
+      case PA_CONTEXT_TERMINATED:
+        BOOST_LOG(debug) << "Pulseadio context terminated"sv;
+        events->raise(terminated);
+        break;
+      case PA_CONTEXT_FAILED:
+        BOOST_LOG(debug) << "Pulseadio context failed"sv;
+        events->raise(failed);
+        break;
+      case PA_CONTEXT_CONNECTING:
+        BOOST_LOG(debug) << "Connecting to pulseaudio"sv;
+      case PA_CONTEXT_UNCONNECTED:
+      case PA_CONTEXT_AUTHORIZING:
+      case PA_CONTEXT_SETTING_NAME:
+        break;
+      }
+    });
+
+    pa_context_set_state_callback(ctx.get(), ctx_state_cb, events_cb.get());
+
+    auto status = pa_context_connect(ctx.get(), nullptr, PA_CONTEXT_NOFLAGS, nullptr);
+    if(status) {
+      BOOST_LOG(error) << "Couldn't connect to pulseaudio: "sv << pa_strerror(status);
+      return -1;
+    }
+
+    worker = std::thread {
+      [](loop_t::pointer loop) {
+        int retval;
+        auto status = pa_mainloop_run(loop, &retval);
+
+        if(status < 0) {
+          BOOST_LOG(fatal) << "Couldn't run pulseaudio main loop"sv;
+
+          log_flush();
+          std::abort();
+        }
+      },
+      loop.get()
+    };
+
+    auto event = events->pop();
+    if(event == failed) {
+      return -1;
+    }
+
+    return 0;
+  }
+
+  int load_null(const char *name, const std::uint8_t *channel_mapping, int channels) {
+    auto alarm = safe::make_alarm<int>();
+
+    op_t op {
+      pa_context_load_module(
+        ctx.get(),
+        "module-null-sink",
+        to_string(name, channel_mapping, channels).c_str(),
+        cb_i,
+        alarm.get()),
+    };
+
+    alarm->wait();
+    return *alarm->status();
+  }
+
+  int unload_null(std::uint32_t i) {
+    if(i == PA_INVALID_INDEX) {
+      return 0;
+    }
+
+    auto alarm = safe::make_alarm<int>();
+
+    op_t op {
+      pa_context_unload_module(ctx.get(), i, success_cb, alarm.get())
+    };
+
+    alarm->wait();
+
+    if(*alarm->status()) {
+      BOOST_LOG(error) << "Couldn't unload null-sink with index ["sv << i << "]: "sv << pa_strerror(pa_context_errno(ctx.get()));
+      return -1;
+    }
+
+    return 0;
+  }
+
+  std::optional<sink_t> sink_info() override {
+    constexpr auto stereo     = "sink-sunshine-stereo";
+    constexpr auto surround51 = "sink-sunshine-surround51";
+    constexpr auto surround71 = "sink-sunshine-surround71";
+
+    auto alarm = safe::make_alarm<int>();
+
+    sink_t sink;
+
+    // Count of all virtual sinks that are created by us
+    int nullcount = 0;
+
+    cb_t<pa_sink_info *> f = [&](ctx_t::pointer ctx, const pa_sink_info *sink_info, int eol) {
+      if(!sink_info) {
+        if(!eol) {
+          BOOST_LOG(error) << "Couldn't get pulseaudio sink info: "sv << pa_strerror(pa_context_errno(ctx));
+
+          alarm->ring(-1);
+        }
+
+        alarm->ring(0);
+        return;
+      }
+
+      // Ensure Sunshine won't create a sink that already exists.
+      if(!std::strcmp(sink_info->name, stereo)) {
+        index.stereo = sink_info->owner_module;
+
+        ++nullcount;
+      }
+      else if(!std::strcmp(sink_info->name, surround51)) {
+        index.surround51 = sink_info->owner_module;
+
+        ++nullcount;
+      }
+      else if(!std::strcmp(sink_info->name, surround71)) {
+        index.surround71 = sink_info->owner_module;
+
+        ++nullcount;
+      }
+    };
+
+    op_t op { pa_context_get_sink_info_list(ctx.get(), cb<pa_sink_info *>, &f) };
+
+    if(!op) {
+      BOOST_LOG(error) << "Couldn't create card info operation: "sv << pa_strerror(pa_context_errno(ctx.get()));
+
+      return std::nullopt;
+    }
+
+    alarm->wait();
+
+    if(*alarm->status()) {
+      return std::nullopt;
+    }
+
+    auto sink_name = get_default_sink_name();
+    if(sink_name.empty()) {
+      BOOST_LOG(warning) << "Couldn't find an active sink"sv;
+    }
+    else {
+      sink.host = sink_name;
+    }
+
+    if(index.stereo == PA_INVALID_INDEX) {
+      index.stereo = load_null(stereo, speaker::map_stereo, sizeof(speaker::map_stereo));
+      if(index.stereo == PA_INVALID_INDEX) {
+        BOOST_LOG(warning) << "Couldn't create virtual sink for stereo: "sv << pa_strerror(pa_context_errno(ctx.get()));
+      }
+      else {
+        ++nullcount;
+      }
+    }
+
+    if(index.surround51 == PA_INVALID_INDEX) {
+      index.surround51 = load_null(surround51, speaker::map_surround51, sizeof(speaker::map_surround51));
+      if(index.surround51 == PA_INVALID_INDEX) {
+        BOOST_LOG(warning) << "Couldn't create virtual sink for surround-51: "sv << pa_strerror(pa_context_errno(ctx.get()));
+      }
+      else {
+        ++nullcount;
+      }
+    }
+
+    if(index.surround71 == PA_INVALID_INDEX) {
+      index.surround71 = load_null(surround71, speaker::map_surround71, sizeof(speaker::map_surround71));
+      if(index.surround71 == PA_INVALID_INDEX) {
+        BOOST_LOG(warning) << "Couldn't create virtual sink for surround-71: "sv << pa_strerror(pa_context_errno(ctx.get()));
+      }
+      else {
+        ++nullcount;
+      }
+    }
+
+    if(nullcount == 3) {
+      sink.null = std::make_optional(sink_t::null_t { stereo, surround51, surround71 });
+    }
+
+    return std::make_optional(std::move(sink));
+  }
+
+  std::string get_default_sink_name() {
+    std::string sink_name = "@DEFAULT_SINK@"s;
+    auto alarm            = safe::make_alarm<int>();
+
+    cb_simple_t<pa_server_info *> server_f = [&](ctx_t::pointer ctx, const pa_server_info *server_info) {
+      if(!server_info) {
+        BOOST_LOG(error) << "Couldn't get pulseaudio server info: "sv << pa_strerror(pa_context_errno(ctx));
+        alarm->ring(-1);
+      }
+
+      sink_name = server_info->default_sink_name;
+      alarm->ring(0);
+    };
+
+    op_t server_op { pa_context_get_server_info(ctx.get(), cb<pa_server_info *>, &server_f) };
+    alarm->wait();
+    // No need to check status. If it failed just return default name.
+    return sink_name;
+  }
+
+  std::string get_monitor_name(const std::string &sink_name) {
+    std::string monitor_name = "@DEFAULT_MONITOR@"s;
+    auto alarm               = safe::make_alarm<int>();
+
+    cb_t<pa_sink_info *> sink_f = [&](ctx_t::pointer ctx, const pa_sink_info *sink_info, int eol) {
+      if(!sink_info) {
+        if(!eol) {
+          BOOST_LOG(error) << "Couldn't get pulseaudio sink info for ["sv << sink_name
+                           << "]: "sv << pa_strerror(pa_context_errno(ctx));
+          alarm->ring(-1);
+        }
+
+        alarm->ring(0);
+        return;
+      }
+
+      monitor_name = sink_info->monitor_source_name;
+    };
+
+    op_t sink_op { pa_context_get_sink_info_by_name(ctx.get(), sink_name.c_str(), cb<pa_sink_info *>, &sink_f) };
+
+    alarm->wait();
+    // No need to check status. If it failed just return default name.
+    BOOST_LOG(info) << "Found default monitor by name: "sv << monitor_name;
+    return monitor_name;
+  }
+
+  std::unique_ptr<mic_t> microphone(const std::uint8_t *mapping, int channels, std::uint32_t sample_rate, std::uint32_t frame_size) override {
+    // Sink choice priority:
+    // 1. Config sink
+    // 2. Last sink swapped to (Usually virtual in this case)
+    // 3. Default Sink
+    // An attempt was made to always use default to match the switching mechanic,
+    // but this happens right after the swap so the default returned by PA was not
+    // the new one just set!
+    auto sink_name = config::audio.sink;
+    if(sink_name.empty()) sink_name = requested_sink;
+    if(sink_name.empty()) sink_name = get_default_sink_name();
+
+    return ::platf::microphone(mapping, channels, sample_rate, frame_size, get_monitor_name(sink_name));
+  }
+
+  int set_sink(const std::string &sink) override {
+    auto alarm = safe::make_alarm<int>();
+
+    BOOST_LOG(info) << "Setting default sink to: ["sv << sink << "]"sv;
+    op_t op {
+      pa_context_set_default_sink(
+        ctx.get(), sink.c_str(), success_cb, alarm.get()),
+    };
+
+    if(!op) {
+      BOOST_LOG(error) << "Couldn't create set default-sink operation: "sv << pa_strerror(pa_context_errno(ctx.get()));
+      return -1;
+    }
+
+    alarm->wait();
+    if(*alarm->status()) {
+      BOOST_LOG(error) << "Couldn't set default-sink ["sv << sink << "]: "sv << pa_strerror(pa_context_errno(ctx.get()));
+
+      return -1;
+    }
+
+    requested_sink = sink;
+
+    return 0;
+  }
+
+  ~server_t() override {
+    unload_null(index.stereo);
+    unload_null(index.surround51);
+    unload_null(index.surround71);
+
+    if(worker.joinable()) {
+      pa_context_disconnect(ctx.get());
+
+      KITTY_WHILE_LOOP(auto event = events->pop(), event != terminated && event != failed, {
+        event = events->pop();
+      })
+
+      pa_mainloop_quit(loop.get(), 0);
+      worker.join();
+    }
+  }
+};
+} // namespace pa
+
+std::unique_ptr<audio_control_t> audio_control() {
+  auto audio = std::make_unique<pa::server_t>();
+
+  if(audio->init()) {
+    return nullptr;
+  }
+
+  return audio;
+}
+} // namespace platf

src/platform/linux/cuda.cpp

@@ -0,0 +1,717 @@
+#include <bitset>
+
+#include <NvFBC.h>
+#include <ffnvcodec/dynlink_loader.h>
+
+extern "C" {
+#include <libavcodec/avcodec.h>
+#include <libavutil/hwcontext_cuda.h>
+#include <libavutil/imgutils.h>
+}
+
+#include "cuda.h"
+#include "graphics.h"
+#include "src/main.h"
+#include "src/utility.h"
+#include "wayland.h"
+
+#define SUNSHINE_STRINGVIEW_HELPER(x) x##sv
+#define SUNSHINE_STRINGVIEW(x) SUNSHINE_STRINGVIEW_HELPER(x)
+
+#define CU_CHECK(x, y) \
+  if(check((x), SUNSHINE_STRINGVIEW(y ": "))) return -1
+
+#define CU_CHECK_IGNORE(x, y) \
+  check((x), SUNSHINE_STRINGVIEW(y ": "))
+
+using namespace std::literals;
+namespace cuda {
+constexpr auto cudaDevAttrMaxThreadsPerBlock          = (CUdevice_attribute)1;
+constexpr auto cudaDevAttrMaxThreadsPerMultiProcessor = (CUdevice_attribute)39;
+
+void pass_error(const std::string_view &sv, const char *name, const char *description) {
+  BOOST_LOG(error) << sv << name << ':' << description;
+}
+
+void cff(CudaFunctions *cf) {
+  cuda_free_functions(&cf);
+}
+
+using cdf_t = util::safe_ptr<CudaFunctions, cff>;
+
+static cdf_t cdf;
+
+inline static int check(CUresult result, const std::string_view &sv) {
+  if(result != CUDA_SUCCESS) {
+    const char *name;
+    const char *description;
+
+    cdf->cuGetErrorName(result, &name);
+    cdf->cuGetErrorString(result, &description);
+
+    BOOST_LOG(error) << sv << name << ':' << description;
+    return -1;
+  }
+
+  return 0;
+}
+
+void freeStream(CUstream stream) {
+  CU_CHECK_IGNORE(cdf->cuStreamDestroy(stream), "Couldn't destroy cuda stream");
+}
+
+class img_t : public platf::img_t {
+public:
+  tex_t tex;
+};
+
+int init() {
+  auto status = cuda_load_functions(&cdf, nullptr);
+  if(status) {
+    BOOST_LOG(error) << "Couldn't load cuda: "sv << status;
+
+    return -1;
+  }
+
+  CU_CHECK(cdf->cuInit(0), "Couldn't initialize cuda");
+
+  return 0;
+}
+
+class cuda_t : public platf::hwdevice_t {
+public:
+  int init(int in_width, int in_height) {
+    if(!cdf) {
+      BOOST_LOG(warning) << "cuda not initialized"sv;
+      return -1;
+    }
+
+    data = (void *)0x1;
+
+    width  = in_width;
+    height = in_height;
+
+    return 0;
+  }
+
+  int set_frame(AVFrame *frame) override {
+    this->hwframe.reset(frame);
+    this->frame = frame;
+
+    auto hwframe_ctx = (AVHWFramesContext *)frame->hw_frames_ctx->data;
+    if(hwframe_ctx->sw_format != AV_PIX_FMT_NV12) {
+      BOOST_LOG(error) << "cuda::cuda_t doesn't support any format other than AV_PIX_FMT_NV12"sv;
+      return -1;
+    }
+
+    if(av_hwframe_get_buffer(frame->hw_frames_ctx, frame, 0)) {
+      BOOST_LOG(error) << "Couldn't get hwframe for NVENC"sv;
+
+      return -1;
+    }
+
+    auto cuda_ctx = (AVCUDADeviceContext *)hwframe_ctx->device_ctx->hwctx;
+
+    stream = make_stream();
+    if(!stream) {
+      return -1;
+    }
+
+    cuda_ctx->stream = stream.get();
+
+    auto sws_opt = sws_t::make(width, height, frame->width, frame->height, width * 4);
+    if(!sws_opt) {
+      return -1;
+    }
+
+    sws = std::move(*sws_opt);
+
+    linear_interpolation = width != frame->width || height != frame->height;
+
+    return 0;
+  }
+
+  void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) override {
+    sws.set_colorspace(colorspace, color_range);
+
+    auto tex = tex_t::make(height, width * 4);
+    if(!tex) {
+      return;
+    }
+
+    // The default green color is ugly.
+    // Update the background color
+    platf::img_t img;
+    img.width       = width;
+    img.height      = height;
+    img.pixel_pitch = 4;
+    img.row_pitch   = img.width * img.pixel_pitch;
+
+    std::vector<std::uint8_t> image_data;
+    image_data.resize(img.row_pitch * img.height);
+
+    img.data = image_data.data();
+
+    if(sws.load_ram(img, tex->array)) {
+      return;
+    }
+
+    sws.convert(frame->data[0], frame->data[1], frame->linesize[0], frame->linesize[1], tex->texture.linear, stream.get(), { frame->width, frame->height, 0, 0 });
+  }
+
+  cudaTextureObject_t tex_obj(const tex_t &tex) const {
+    return linear_interpolation ? tex.texture.linear : tex.texture.point;
+  }
+
+  stream_t stream;
+  frame_t hwframe;
+
+  int width, height;
+
+  // When heigth and width don't change, it's not necessary to use linear interpolation
+  bool linear_interpolation;
+
+  sws_t sws;
+};
+
+class cuda_ram_t : public cuda_t {
+public:
+  int convert(platf::img_t &img) override {
+    return sws.load_ram(img, tex.array) || sws.convert(frame->data[0], frame->data[1], frame->linesize[0], frame->linesize[1], tex_obj(tex), stream.get());
+  }
+
+  int set_frame(AVFrame *frame) {
+    if(cuda_t::set_frame(frame)) {
+      return -1;
+    }
+
+    auto tex_opt = tex_t::make(height, width * 4);
+    if(!tex_opt) {
+      return -1;
+    }
+
+    tex = std::move(*tex_opt);
+
+    return 0;
+  }
+
+  tex_t tex;
+};
+
+class cuda_vram_t : public cuda_t {
+public:
+  int convert(platf::img_t &img) override {
+    return sws.convert(frame->data[0], frame->data[1], frame->linesize[0], frame->linesize[1], tex_obj(((img_t *)&img)->tex), stream.get());
+  }
+};
+
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, bool vram) {
+  if(init()) {
+    return nullptr;
+  }
+
+  std::shared_ptr<cuda_t> cuda;
+
+  if(vram) {
+    cuda = std::make_shared<cuda_vram_t>();
+  }
+  else {
+    cuda = std::make_shared<cuda_ram_t>();
+  }
+
+  if(cuda->init(width, height)) {
+    return nullptr;
+  }
+
+  return cuda;
+}
+
+namespace nvfbc {
+static PNVFBCCREATEINSTANCE createInstance {};
+static NVFBC_API_FUNCTION_LIST func { NVFBC_VERSION };
+
+static constexpr inline NVFBC_BOOL nv_bool(bool b) {
+  return b ? NVFBC_TRUE : NVFBC_FALSE;
+}
+
+static void *handle { nullptr };
+int init() {
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libnvidia-fbc.so.1", "libnvidia-fbc.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&createInstance, "NvFBCCreateInstance" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    dlclose(handle);
+    handle = nullptr;
+
+    return -1;
+  }
+
+  auto status = cuda::nvfbc::createInstance(&cuda::nvfbc::func);
+  if(status) {
+    BOOST_LOG(error) << "Unable to create NvFBC instance"sv;
+
+    dlclose(handle);
+    handle = nullptr;
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+class ctx_t {
+public:
+  ctx_t(NVFBC_SESSION_HANDLE handle) {
+    NVFBC_BIND_CONTEXT_PARAMS params { NVFBC_BIND_CONTEXT_PARAMS_VER };
+
+    if(func.nvFBCBindContext(handle, &params)) {
+      BOOST_LOG(error) << "Couldn't bind NvFBC context to current thread: " << func.nvFBCGetLastErrorStr(handle);
+    }
+
+    this->handle = handle;
+  }
+
+  ~ctx_t() {
+    NVFBC_RELEASE_CONTEXT_PARAMS params { NVFBC_RELEASE_CONTEXT_PARAMS_VER };
+    if(func.nvFBCReleaseContext(handle, &params)) {
+      BOOST_LOG(error) << "Couldn't release NvFBC context from current thread: " << func.nvFBCGetLastErrorStr(handle);
+    }
+  }
+
+  NVFBC_SESSION_HANDLE handle;
+};
+
+class handle_t {
+  enum flag_e {
+    SESSION_HANDLE,
+    SESSION_CAPTURE,
+    MAX_FLAGS,
+  };
+
+public:
+  handle_t() = default;
+  handle_t(handle_t &&other) : handle_flags { other.handle_flags }, handle { other.handle } {
+    other.handle_flags.reset();
+  }
+
+  handle_t &operator=(handle_t &&other) {
+    std::swap(handle_flags, other.handle_flags);
+    std::swap(handle, other.handle);
+
+    return *this;
+  }
+
+  static std::optional<handle_t> make() {
+    NVFBC_CREATE_HANDLE_PARAMS params { NVFBC_CREATE_HANDLE_PARAMS_VER };
+
+    handle_t handle;
+    auto status = func.nvFBCCreateHandle(&handle.handle, &params);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create session: "sv << handle.last_error();
+
+      return std::nullopt;
+    }
+
+    handle.handle_flags[SESSION_HANDLE] = true;
+
+    return std::move(handle);
+  }
+
+  const char *last_error() {
+    return func.nvFBCGetLastErrorStr(handle);
+  }
+
+  std::optional<NVFBC_GET_STATUS_PARAMS> status() {
+    NVFBC_GET_STATUS_PARAMS params { NVFBC_GET_STATUS_PARAMS_VER };
+
+    auto status = func.nvFBCGetStatus(handle, &params);
+    if(status) {
+      BOOST_LOG(error) << "Failed to get NvFBC status: "sv << last_error();
+
+      return std::nullopt;
+    }
+
+    return params;
+  }
+
+  int capture(NVFBC_CREATE_CAPTURE_SESSION_PARAMS &capture_params) {
+    if(func.nvFBCCreateCaptureSession(handle, &capture_params)) {
+      BOOST_LOG(error) << "Failed to start capture session: "sv << last_error();
+      return -1;
+    }
+
+    handle_flags[SESSION_CAPTURE] = true;
+
+    NVFBC_TOCUDA_SETUP_PARAMS setup_params {
+      NVFBC_TOCUDA_SETUP_PARAMS_VER,
+      NVFBC_BUFFER_FORMAT_BGRA,
+    };
+
+    if(func.nvFBCToCudaSetUp(handle, &setup_params)) {
+      BOOST_LOG(error) << "Failed to setup cuda interop with nvFBC: "sv << last_error();
+      return -1;
+    }
+    return 0;
+  }
+
+  int stop() {
+    if(!handle_flags[SESSION_CAPTURE]) {
+      return 0;
+    }
+
+    NVFBC_DESTROY_CAPTURE_SESSION_PARAMS params { NVFBC_DESTROY_CAPTURE_SESSION_PARAMS_VER };
+
+    if(func.nvFBCDestroyCaptureSession(handle, &params)) {
+      BOOST_LOG(error) << "Couldn't destroy capture session: "sv << last_error();
+
+      return -1;
+    }
+
+    handle_flags[SESSION_CAPTURE] = false;
+
+    return 0;
+  }
+
+  int reset() {
+    if(!handle_flags[SESSION_HANDLE]) {
+      return 0;
+    }
+
+    stop();
+
+    NVFBC_DESTROY_HANDLE_PARAMS params { NVFBC_DESTROY_HANDLE_PARAMS_VER };
+
+    if(func.nvFBCDestroyHandle(handle, &params)) {
+      BOOST_LOG(error) << "Couldn't destroy session handle: "sv << func.nvFBCGetLastErrorStr(handle);
+    }
+
+    handle_flags[SESSION_HANDLE] = false;
+
+    return 0;
+  }
+
+  ~handle_t() {
+    reset();
+  }
+
+  std::bitset<MAX_FLAGS> handle_flags;
+
+  NVFBC_SESSION_HANDLE handle;
+};
+
+class display_t : public platf::display_t {
+public:
+  int init(const std::string_view &display_name, int framerate) {
+    auto handle = handle_t::make();
+    if(!handle) {
+      return -1;
+    }
+
+    ctx_t ctx { handle->handle };
+
+    auto status_params = handle->status();
+    if(!status_params) {
+      return -1;
+    }
+
+    int streamedMonitor = -1;
+    if(!display_name.empty()) {
+      if(status_params->bXRandRAvailable) {
+        auto monitor_nr = util::from_view(display_name);
+
+        if(monitor_nr < 0 || monitor_nr >= status_params->dwOutputNum) {
+          BOOST_LOG(warning) << "Can't stream monitor ["sv << monitor_nr << "], it needs to be between [0] and ["sv << status_params->dwOutputNum - 1 << "], defaulting to virtual desktop"sv;
+        }
+        else {
+          streamedMonitor = monitor_nr;
+        }
+      }
+      else {
+        BOOST_LOG(warning) << "XrandR not available, streaming entire virtual desktop"sv;
+      }
+    }
+
+    delay = std::chrono::nanoseconds { 1s } / framerate;
+
+    capture_params = NVFBC_CREATE_CAPTURE_SESSION_PARAMS { NVFBC_CREATE_CAPTURE_SESSION_PARAMS_VER };
+
+    capture_params.eCaptureType                = NVFBC_CAPTURE_SHARED_CUDA;
+    capture_params.bDisableAutoModesetRecovery = nv_bool(true);
+
+    capture_params.dwSamplingRateMs = 1000 /* ms */ / framerate;
+
+    if(streamedMonitor != -1) {
+      auto &output = status_params->outputs[streamedMonitor];
+
+      width    = output.trackedBox.w;
+      height   = output.trackedBox.h;
+      offset_x = output.trackedBox.x;
+      offset_y = output.trackedBox.y;
+
+      capture_params.eTrackingType = NVFBC_TRACKING_OUTPUT;
+      capture_params.dwOutputId    = output.dwId;
+    }
+    else {
+      capture_params.eTrackingType = NVFBC_TRACKING_SCREEN;
+
+      width  = status_params->screenSize.w;
+      height = status_params->screenSize.h;
+    }
+
+    env_width  = status_params->screenSize.w;
+    env_height = status_params->screenSize.h;
+
+    this->handle = std::move(*handle);
+    return 0;
+  }
+
+  platf::capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<platf::img_t> img, bool *cursor) override {
+    auto next_frame = std::chrono::steady_clock::now();
+
+    // Force display_t::capture to initialize handle_t::capture
+    cursor_visible = !*cursor;
+
+    ctx_t ctx { handle.handle };
+    auto fg = util::fail_guard([&]() {
+      handle.reset();
+    });
+
+    while(img) {
+      auto now = std::chrono::steady_clock::now();
+      if(next_frame > now) {
+        std::this_thread::sleep_for((next_frame - now) / 3 * 2);
+      }
+      while(next_frame > now) {
+        std::this_thread::sleep_for(1ns);
+        now = std::chrono::steady_clock::now();
+      }
+      next_frame = now + delay;
+
+      auto status = snapshot(img.get(), 150ms, *cursor);
+      switch(status) {
+      case platf::capture_e::reinit:
+      case platf::capture_e::error:
+        return status;
+      case platf::capture_e::timeout:
+        std::this_thread::sleep_for(1ms);
+        continue;
+      case platf::capture_e::ok:
+        img = snapshot_cb(img);
+        break;
+      default:
+        BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+        return status;
+      }
+    }
+
+    return platf::capture_e::ok;
+  }
+
+  // Reinitialize the capture session.
+  platf::capture_e reinit(bool cursor) {
+    if(handle.stop()) {
+      return platf::capture_e::error;
+    }
+
+    cursor_visible = cursor;
+    if(cursor) {
+      capture_params.bPushModel          = nv_bool(false);
+      capture_params.bWithCursor         = nv_bool(true);
+      capture_params.bAllowDirectCapture = nv_bool(false);
+    }
+    else {
+      capture_params.bPushModel          = nv_bool(true);
+      capture_params.bWithCursor         = nv_bool(false);
+      capture_params.bAllowDirectCapture = nv_bool(true);
+    }
+
+    if(handle.capture(capture_params)) {
+      return platf::capture_e::error;
+    }
+
+    // If trying to capture directly, test if it actually does.
+    if(capture_params.bAllowDirectCapture) {
+      CUdeviceptr device_ptr;
+      NVFBC_FRAME_GRAB_INFO info;
+
+      NVFBC_TOCUDA_GRAB_FRAME_PARAMS grab {
+        NVFBC_TOCUDA_GRAB_FRAME_PARAMS_VER,
+        NVFBC_TOCUDA_GRAB_FLAGS_NOWAIT,
+        &device_ptr,
+        &info,
+        0,
+      };
+
+      // Direct Capture may fail the first few times, even if it's possible
+      for(int x = 0; x < 3; ++x) {
+        if(auto status = func.nvFBCToCudaGrabFrame(handle.handle, &grab)) {
+          if(status == NVFBC_ERR_MUST_RECREATE) {
+            return platf::capture_e::reinit;
+          }
+
+          BOOST_LOG(error) << "Couldn't capture nvFramebuffer: "sv << handle.last_error();
+
+          return platf::capture_e::error;
+        }
+
+        if(info.bDirectCapture) {
+          break;
+        }
+
+        BOOST_LOG(debug) << "Direct capture failed attempt ["sv << x << ']';
+      }
+
+      if(!info.bDirectCapture) {
+        BOOST_LOG(debug) << "Direct capture failed, trying the extra copy method"sv;
+        // Direct capture failed
+        capture_params.bPushModel          = nv_bool(false);
+        capture_params.bWithCursor         = nv_bool(false);
+        capture_params.bAllowDirectCapture = nv_bool(false);
+
+        if(handle.stop() || handle.capture(capture_params)) {
+          return platf::capture_e::error;
+        }
+      }
+    }
+
+    return platf::capture_e::ok;
+  }
+
+  platf::capture_e snapshot(platf::img_t *img, std::chrono::milliseconds timeout, bool cursor) {
+    if(cursor != cursor_visible) {
+      auto status = reinit(cursor);
+      if(status != platf::capture_e::ok) {
+        return status;
+      }
+    }
+
+    CUdeviceptr device_ptr;
+    NVFBC_FRAME_GRAB_INFO info;
+
+    NVFBC_TOCUDA_GRAB_FRAME_PARAMS grab {
+      NVFBC_TOCUDA_GRAB_FRAME_PARAMS_VER,
+      NVFBC_TOCUDA_GRAB_FLAGS_NOWAIT,
+      &device_ptr,
+      &info,
+      (std::uint32_t)timeout.count(),
+    };
+
+    if(auto status = func.nvFBCToCudaGrabFrame(handle.handle, &grab)) {
+      if(status == NVFBC_ERR_MUST_RECREATE) {
+        return platf::capture_e::reinit;
+      }
+
+      BOOST_LOG(error) << "Couldn't capture nvFramebuffer: "sv << handle.last_error();
+      return platf::capture_e::error;
+    }
+
+    if(((img_t *)img)->tex.copy((std::uint8_t *)device_ptr, img->height, img->row_pitch)) {
+      return platf::capture_e::error;
+    }
+
+    return platf::capture_e::ok;
+  }
+
+  std::shared_ptr<platf::hwdevice_t> make_hwdevice(platf::pix_fmt_e pix_fmt) override {
+    return ::cuda::make_hwdevice(width, height, true);
+  }
+
+  std::shared_ptr<platf::img_t> alloc_img() override {
+    auto img = std::make_shared<cuda::img_t>();
+
+    img->data        = nullptr;
+    img->width       = width;
+    img->height      = height;
+    img->pixel_pitch = 4;
+    img->row_pitch   = img->width * img->pixel_pitch;
+
+    auto tex_opt = tex_t::make(height, width * img->pixel_pitch);
+    if(!tex_opt) {
+      return nullptr;
+    }
+
+    img->tex = std::move(*tex_opt);
+
+    return img;
+  };
+
+  int dummy_img(platf::img_t *) override {
+    return 0;
+  }
+
+  std::chrono::nanoseconds delay;
+
+  bool cursor_visible;
+  handle_t handle;
+
+  NVFBC_CREATE_CAPTURE_SESSION_PARAMS capture_params;
+};
+} // namespace nvfbc
+} // namespace cuda
+
+namespace platf {
+std::shared_ptr<display_t> nvfbc_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+  if(hwdevice_type != mem_type_e::cuda) {
+    BOOST_LOG(error) << "Could not initialize nvfbc display with the given hw device type"sv;
+    return nullptr;
+  }
+
+  auto display = std::make_shared<cuda::nvfbc::display_t>();
+
+  if(display->init(display_name, framerate)) {
+    return nullptr;
+  }
+
+  return display;
+}
+
+std::vector<std::string> nvfbc_display_names() {
+  if(cuda::init() || cuda::nvfbc::init()) {
+    return {};
+  }
+
+  std::vector<std::string> display_names;
+
+  auto handle = cuda::nvfbc::handle_t::make();
+  if(!handle) {
+    return {};
+  }
+
+  auto status_params = handle->status();
+  if(!status_params) {
+    return {};
+  }
+
+  if(!status_params->bIsCapturePossible) {
+    BOOST_LOG(error) << "NVidia driver doesn't support NvFBC screencasting"sv;
+  }
+
+  BOOST_LOG(info) << "Found ["sv << status_params->dwOutputNum << "] outputs"sv;
+  BOOST_LOG(info) << "Virtual Desktop: "sv << status_params->screenSize.w << 'x' << status_params->screenSize.h;
+  BOOST_LOG(info) << "XrandR: "sv << (status_params->bXRandRAvailable ? "available"sv : "unavailable"sv);
+
+  for(auto x = 0; x < status_params->dwOutputNum; ++x) {
+    auto &output = status_params->outputs[x];
+    BOOST_LOG(info) << "-- Output --"sv;
+    BOOST_LOG(debug) << "  ID: "sv << output.dwId;
+    BOOST_LOG(debug) << "  Name: "sv << output.name;
+    BOOST_LOG(info) << "  Resolution: "sv << output.trackedBox.w << 'x' << output.trackedBox.h;
+    BOOST_LOG(info) << "  Offset: "sv << output.trackedBox.x << 'x' << output.trackedBox.y;
+    display_names.emplace_back(std::to_string(x));
+  }
+
+  return display_names;
+}
+} // namespace platf

src/platform/linux/cuda.cu

@@ -0,0 +1,331 @@
+// #include <algorithm>
+#include <helper_math.h>
+#include <limits>
+#include <memory>
+#include <optional>
+#include <string_view>
+
+#include "cuda.h"
+
+using namespace std::literals;
+
+#define SUNSHINE_STRINGVIEW_HELPER(x) x##sv
+#define SUNSHINE_STRINGVIEW(x) SUNSHINE_STRINGVIEW_HELPER(x)
+
+#define CU_CHECK(x, y) \
+  if(check((x), SUNSHINE_STRINGVIEW(y ": "))) return -1
+
+#define CU_CHECK_VOID(x, y) \
+  if(check((x), SUNSHINE_STRINGVIEW(y ": "))) return;
+
+#define CU_CHECK_PTR(x, y) \
+  if(check((x), SUNSHINE_STRINGVIEW(y ": "))) return nullptr;
+
+#define CU_CHECK_OPT(x, y) \
+  if(check((x), SUNSHINE_STRINGVIEW(y ": "))) return std::nullopt;
+
+#define CU_CHECK_IGNORE(x, y) \
+  check((x), SUNSHINE_STRINGVIEW(y ": "))
+
+using namespace std::literals;
+
+//////////////////// Special desclarations
+/**
+ * NVCC segfaults when including <chrono>
+ * Therefore, some declarations need to be added explicitely
+ */
+namespace platf {
+struct img_t {
+public:
+  std::uint8_t *data {};
+  std::int32_t width {};
+  std::int32_t height {};
+  std::int32_t pixel_pitch {};
+  std::int32_t row_pitch {};
+
+  virtual ~img_t() = default;
+};
+} // namespace platf
+
+namespace video {
+using __float4 = float[4];
+using __float3 = float[3];
+using __float2 = float[2];
+
+struct __attribute__((__aligned__(16))) color_t {
+  float4 color_vec_y;
+  float4 color_vec_u;
+  float4 color_vec_v;
+  float2 range_y;
+  float2 range_uv;
+};
+
+struct __attribute__((__aligned__(16))) color_extern_t {
+  __float4 color_vec_y;
+  __float4 color_vec_u;
+  __float4 color_vec_v;
+  __float2 range_y;
+  __float2 range_uv;
+};
+
+static_assert(sizeof(video::color_t) == sizeof(video::color_extern_t), "color matrix struct mismatch");
+
+extern color_t colors[4];
+} // namespace video
+
+//////////////////// End special declarations
+
+namespace cuda {
+auto constexpr INVALID_TEXTURE = std::numeric_limits<cudaTextureObject_t>::max();
+
+template<class T>
+inline T div_align(T l, T r) {
+  return (l + r - 1) / r;
+}
+
+void pass_error(const std::string_view &sv, const char *name, const char *description);
+inline static int check(cudaError_t result, const std::string_view &sv) {
+  if(result) {
+    auto name        = cudaGetErrorName(result);
+    auto description = cudaGetErrorString(result);
+
+    pass_error(sv, name, description);
+    return -1;
+  }
+
+  return 0;
+}
+
+template<class T>
+ptr_t make_ptr() {
+  void *p;
+  CU_CHECK_PTR(cudaMalloc(&p, sizeof(T)), "Couldn't allocate color matrix");
+
+  ptr_t ptr { p };
+
+  return ptr;
+}
+
+void freeCudaPtr_t::operator()(void *ptr) {
+  CU_CHECK_IGNORE(cudaFree(ptr), "Couldn't free cuda device pointer");
+}
+
+void freeCudaStream_t::operator()(cudaStream_t ptr) {
+  CU_CHECK_IGNORE(cudaStreamDestroy(ptr), "Couldn't free cuda stream");
+}
+
+stream_t make_stream(int flags) {
+  cudaStream_t stream;
+
+  if(!flags) {
+    CU_CHECK_PTR(cudaStreamCreate(&stream), "Couldn't create cuda stream");
+  }
+  else {
+    CU_CHECK_PTR(cudaStreamCreateWithFlags(&stream, flags), "Couldn't create cuda stream with flags");
+  }
+
+  return stream_t { stream };
+}
+
+inline __device__ float3 bgra_to_rgb(uchar4 vec) {
+  return make_float3((float)vec.z, (float)vec.y, (float)vec.x);
+}
+
+inline __device__ float3 bgra_to_rgb(float4 vec) {
+  return make_float3(vec.z, vec.y, vec.x);
+}
+
+inline __device__ float2 calcUV(float3 pixel, const video::color_t *const color_matrix) {
+  float4 vec_u = color_matrix->color_vec_u;
+  float4 vec_v = color_matrix->color_vec_v;
+
+  float u = dot(pixel, make_float3(vec_u)) + vec_u.w;
+  float v = dot(pixel, make_float3(vec_v)) + vec_v.w;
+
+  u = u * color_matrix->range_uv.x + color_matrix->range_uv.y;
+  v = (v * color_matrix->range_uv.x + color_matrix->range_uv.y) * 224.0f / 256.0f + 0.0625f;
+
+  return make_float2(u, v);
+}
+
+inline __device__ float calcY(float3 pixel, const video::color_t *const color_matrix) {
+  float4 vec_y = color_matrix->color_vec_y;
+
+  return (dot(pixel, make_float3(vec_y)) + vec_y.w) * color_matrix->range_y.x + color_matrix->range_y.y;
+}
+
+__global__ void RGBA_to_NV12(
+  cudaTextureObject_t srcImage, std::uint8_t *dstY, std::uint8_t *dstUV,
+  std::uint32_t dstPitchY, std::uint32_t dstPitchUV,
+  float scale, const viewport_t viewport, const video::color_t *const color_matrix) {
+
+  int idX = (threadIdx.x + blockDim.x * blockIdx.x) * 2;
+  int idY = (threadIdx.y + blockDim.y * blockIdx.y);
+
+  if(idX >= viewport.width) return;
+  if(idY >= viewport.height) return;
+
+  float x = idX * scale;
+  float y = idY * scale;
+
+  idX += viewport.offsetX;
+  idY += viewport.offsetY;
+
+  dstY  = dstY + idX + idY * dstPitchY;
+  dstUV = dstUV + idX + (idY / 2 * dstPitchUV);
+
+  float3 rgb_l = bgra_to_rgb(tex2D<float4>(srcImage, x, y));
+  float3 rgb_r = bgra_to_rgb(tex2D<float4>(srcImage, x + scale, y));
+
+  float2 uv = calcUV((rgb_l + rgb_r) * 0.5f, color_matrix) * 256.0f;
+
+  dstUV[0] = uv.x;
+  dstUV[1] = uv.y;
+  dstY[0]  = calcY(rgb_l, color_matrix) * 245.0f; // 245.0f is a magic number to ensure slight changes in luminosity are more visisble
+  dstY[1]  = calcY(rgb_r, color_matrix) * 245.0f; // 245.0f is a magic number to ensure slight changes in luminosity are more visisble
+}
+
+int tex_t::copy(std::uint8_t *src, int height, int pitch) {
+  CU_CHECK(cudaMemcpy2DToArray(array, 0, 0, src, pitch, pitch, height, cudaMemcpyDeviceToDevice), "Couldn't copy to cuda array from deviceptr");
+
+  return 0;
+}
+
+std::optional<tex_t> tex_t::make(int height, int pitch) {
+  tex_t tex;
+
+  auto format = cudaCreateChannelDesc<uchar4>();
+  CU_CHECK_OPT(cudaMallocArray(&tex.array, &format, pitch, height, cudaArrayDefault), "Couldn't allocate cuda array");
+
+  cudaResourceDesc res {};
+  res.resType         = cudaResourceTypeArray;
+  res.res.array.array = tex.array;
+
+  cudaTextureDesc desc {};
+
+  desc.readMode         = cudaReadModeNormalizedFloat;
+  desc.filterMode       = cudaFilterModePoint;
+  desc.normalizedCoords = false;
+
+  std::fill_n(std::begin(desc.addressMode), 2, cudaAddressModeClamp);
+
+  CU_CHECK_OPT(cudaCreateTextureObject(&tex.texture.point, &res, &desc, nullptr), "Couldn't create cuda texture that uses point interpolation");
+
+  desc.filterMode = cudaFilterModeLinear;
+
+  CU_CHECK_OPT(cudaCreateTextureObject(&tex.texture.linear, &res, &desc, nullptr), "Couldn't create cuda texture that uses linear interpolation");
+
+  return std::move(tex);
+}
+
+tex_t::tex_t() : array {}, texture { INVALID_TEXTURE } {}
+tex_t::tex_t(tex_t &&other) : array { other.array }, texture { other.texture } {
+  other.array          = 0;
+  other.texture.point  = INVALID_TEXTURE;
+  other.texture.linear = INVALID_TEXTURE;
+}
+
+tex_t &tex_t::operator=(tex_t &&other) {
+  std::swap(array, other.array);
+  std::swap(texture, other.texture);
+
+  return *this;
+}
+
+tex_t::~tex_t() {
+  if(texture.point != INVALID_TEXTURE) {
+    CU_CHECK_IGNORE(cudaDestroyTextureObject(texture.point), "Couldn't deallocate cuda texture that uses point interpolation");
+
+    texture.point = INVALID_TEXTURE;
+  }
+
+  if(texture.linear != INVALID_TEXTURE) {
+    CU_CHECK_IGNORE(cudaDestroyTextureObject(texture.linear), "Couldn't deallocate cuda texture that uses linear interpolation");
+
+    texture.linear = INVALID_TEXTURE;
+  }
+
+  if(array) {
+    CU_CHECK_IGNORE(cudaFreeArray(array), "Couldn't deallocate cuda array");
+
+    array = cudaArray_t {};
+  }
+}
+
+sws_t::sws_t(int in_width, int in_height, int out_width, int out_height, int pitch, int threadsPerBlock, ptr_t &&color_matrix)
+    : threadsPerBlock { threadsPerBlock }, color_matrix { std::move(color_matrix) } {
+  // Ensure aspect ratio is maintained
+  auto scalar       = std::fminf(out_width / (float)in_width, out_height / (float)in_height);
+  auto out_width_f  = in_width * scalar;
+  auto out_height_f = in_height * scalar;
+
+  // result is always positive
+  auto offsetX_f = (out_width - out_width_f) / 2;
+  auto offsetY_f = (out_height - out_height_f) / 2;
+
+  viewport.width  = out_width_f;
+  viewport.height = out_height_f;
+
+  viewport.offsetX = offsetX_f;
+  viewport.offsetY = offsetY_f;
+
+  scale = 1.0f / scalar;
+}
+
+std::optional<sws_t> sws_t::make(int in_width, int in_height, int out_width, int out_height, int pitch) {
+  cudaDeviceProp props;
+  int device;
+  CU_CHECK_OPT(cudaGetDevice(&device), "Couldn't get cuda device");
+  CU_CHECK_OPT(cudaGetDeviceProperties(&props, device), "Couldn't get cuda device properties");
+
+  auto ptr = make_ptr<video::color_t>();
+  if(!ptr) {
+    return std::nullopt;
+  }
+
+  return std::make_optional<sws_t>(in_width, in_height, out_width, out_height, pitch, props.maxThreadsPerMultiProcessor / props.maxBlocksPerMultiProcessor, std::move(ptr));
+}
+
+int sws_t::convert(std::uint8_t *Y, std::uint8_t *UV, std::uint32_t pitchY, std::uint32_t pitchUV, cudaTextureObject_t texture, stream_t::pointer stream) {
+  return convert(Y, UV, pitchY, pitchUV, texture, stream, viewport);
+}
+
+int sws_t::convert(std::uint8_t *Y, std::uint8_t *UV, std::uint32_t pitchY, std::uint32_t pitchUV, cudaTextureObject_t texture, stream_t::pointer stream, const viewport_t &viewport) {
+  int threadsX = viewport.width / 2;
+  int threadsY = viewport.height;
+
+  dim3 block(threadsPerBlock);
+  dim3 grid(div_align(threadsX, threadsPerBlock), threadsY);
+
+  RGBA_to_NV12<<<grid, block, 0, stream>>>(texture, Y, UV, pitchY, pitchUV, scale, viewport, (video::color_t *)color_matrix.get());
+
+  return CU_CHECK_IGNORE(cudaGetLastError(), "RGBA_to_NV12 failed");
+}
+
+void sws_t::set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) {
+  video::color_t *color_p;
+  switch(colorspace) {
+  case 5: // SWS_CS_SMPTE170M
+    color_p = &video::colors[0];
+    break;
+  case 1: // SWS_CS_ITU709
+    color_p = &video::colors[2];
+    break;
+  case 9: // SWS_CS_BT2020
+  default:
+    color_p = &video::colors[0];
+  };
+
+  if(color_range > 1) {
+    // Full range
+    ++color_p;
+  }
+
+  CU_CHECK_IGNORE(cudaMemcpy(color_matrix.get(), color_p, sizeof(video::color_t), cudaMemcpyHostToDevice), "Couldn't copy color matrix to cuda");
+}
+
+int sws_t::load_ram(platf::img_t &img, cudaArray_t array) {
+  return CU_CHECK_IGNORE(cudaMemcpy2DToArray(array, 0, 0, img.data, img.row_pitch, img.width * img.pixel_pitch, img.height, cudaMemcpyHostToDevice), "Couldn't copy to cuda array");
+}
+
+} // namespace cuda

src/platform/linux/cuda.h

@@ -0,0 +1,107 @@
+#if !defined(SUNSHINE_PLATFORM_CUDA_H) && defined(SUNSHINE_BUILD_CUDA)
+#define SUNSHINE_PLATFORM_CUDA_H
+
+#include <memory>
+#include <optional>
+#include <string>
+#include <vector>
+
+namespace platf {
+class hwdevice_t;
+class img_t;
+} // namespace platf
+
+namespace cuda {
+
+namespace nvfbc {
+std::vector<std::string> display_names();
+}
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, bool vram);
+int init();
+} // namespace cuda
+
+typedef struct cudaArray *cudaArray_t;
+
+#if !defined(__CUDACC__)
+typedef struct CUstream_st *cudaStream_t;
+typedef unsigned long long cudaTextureObject_t;
+#else  /* defined(__CUDACC__) */
+typedef __location__(device_builtin) struct CUstream_st *cudaStream_t;
+typedef __location__(device_builtin) unsigned long long cudaTextureObject_t;
+#endif /* !defined(__CUDACC__) */
+
+namespace cuda {
+
+class freeCudaPtr_t {
+public:
+  void operator()(void *ptr);
+};
+
+class freeCudaStream_t {
+public:
+  void operator()(cudaStream_t ptr);
+};
+
+using ptr_t    = std::unique_ptr<void, freeCudaPtr_t>;
+using stream_t = std::unique_ptr<CUstream_st, freeCudaStream_t>;
+
+stream_t make_stream(int flags = 0);
+
+struct viewport_t {
+  int width, height;
+  int offsetX, offsetY;
+};
+
+class tex_t {
+public:
+  static std::optional<tex_t> make(int height, int pitch);
+
+  tex_t();
+  tex_t(tex_t &&);
+
+  tex_t &operator=(tex_t &&other);
+
+  ~tex_t();
+
+  int copy(std::uint8_t *src, int height, int pitch);
+
+  cudaArray_t array;
+
+  struct texture {
+    cudaTextureObject_t point;
+    cudaTextureObject_t linear;
+  } texture;
+};
+
+class sws_t {
+public:
+  sws_t() = default;
+  sws_t(int in_width, int in_height, int out_width, int out_height, int pitch, int threadsPerBlock, ptr_t &&color_matrix);
+
+  /**
+   * in_width, in_height -- The width and height of the captured image in pixels
+   * out_width, out_height -- the width and height of the NV12 image in pixels
+   * 
+   * pitch -- The size of a single row of pixels in bytes
+   */
+  static std::optional<sws_t> make(int in_width, int in_height, int out_width, int out_height, int pitch);
+
+  // Converts loaded image into a CUDevicePtr
+  int convert(std::uint8_t *Y, std::uint8_t *UV, std::uint32_t pitchY, std::uint32_t pitchUV, cudaTextureObject_t texture, stream_t::pointer stream);
+  int convert(std::uint8_t *Y, std::uint8_t *UV, std::uint32_t pitchY, std::uint32_t pitchUV, cudaTextureObject_t texture, stream_t::pointer stream, const viewport_t &viewport);
+
+  void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range);
+
+  int load_ram(platf::img_t &img, cudaArray_t array);
+
+  ptr_t color_matrix;
+
+  int threadsPerBlock;
+
+  viewport_t viewport;
+
+  float scale;
+};
+} // namespace cuda
+
+#endif

src/platform/linux/graphics.cpp

@@ -0,0 +1,866 @@
+#include "graphics.h"
+#include "src/video.h"
+
+#include <fcntl.h>
+
+// I want to have as little build dependencies as possible
+// There aren't that many DRM_FORMAT I need to use, so define them here
+//
+// They aren't likely to change any time soon.
+#define fourcc_code(a, b, c, d) ((std::uint32_t)(a) | ((std::uint32_t)(b) << 8) | \
+                                 ((std::uint32_t)(c) << 16) | ((std::uint32_t)(d) << 24))
+#define fourcc_mod_code(vendor, val) ((((uint64_t)vendor) << 56) | ((val)&0x00ffffffffffffffULL))
+#define DRM_FORMAT_R8 fourcc_code('R', '8', ' ', ' ')       /* [7:0] R */
+#define DRM_FORMAT_GR88 fourcc_code('G', 'R', '8', '8')     /* [15:0] G:R 8:8 little endian */
+#define DRM_FORMAT_ARGB8888 fourcc_code('A', 'R', '2', '4') /* [31:0] A:R:G:B 8:8:8:8 little endian */
+#define DRM_FORMAT_XRGB8888 fourcc_code('X', 'R', '2', '4') /* [31:0] x:R:G:B 8:8:8:8 little endian */
+#define DRM_FORMAT_XBGR8888 fourcc_code('X', 'B', '2', '4') /* [31:0] x:B:G:R 8:8:8:8 little endian */
+#define DRM_FORMAT_MOD_INVALID fourcc_mod_code(0, ((1ULL << 56) - 1))
+
+#define SUNSHINE_SHADERS_DIR SUNSHINE_ASSETS_DIR "/shaders/opengl"
+
+using namespace std::literals;
+namespace gl {
+GladGLContext ctx;
+
+void drain_errors(const std::string_view &prefix) {
+  GLenum err;
+  while((err = ctx.GetError()) != GL_NO_ERROR) {
+    BOOST_LOG(error) << "GL: "sv << prefix << ": ["sv << util::hex(err).to_string_view() << ']';
+  }
+}
+
+tex_t::~tex_t() {
+  if(!size() == 0) {
+    ctx.DeleteTextures(size(), begin());
+  }
+}
+
+tex_t tex_t::make(std::size_t count) {
+  tex_t textures { count };
+
+  ctx.GenTextures(textures.size(), textures.begin());
+
+  float color[] = { 0.0f, 0.0f, 0.0f, 1.0f };
+
+  for(auto tex : textures) {
+    gl::ctx.BindTexture(GL_TEXTURE_2D, tex);
+    gl::ctx.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); // x
+    gl::ctx.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); // y
+    gl::ctx.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+    gl::ctx.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+    gl::ctx.TexParameterfv(GL_TEXTURE_2D, GL_TEXTURE_BORDER_COLOR, color);
+  }
+
+  return textures;
+}
+
+frame_buf_t::~frame_buf_t() {
+  if(begin()) {
+    ctx.DeleteFramebuffers(size(), begin());
+  }
+}
+
+frame_buf_t frame_buf_t::make(std::size_t count) {
+  frame_buf_t frame_buf { count };
+
+  ctx.GenFramebuffers(frame_buf.size(), frame_buf.begin());
+
+  return frame_buf;
+}
+
+void frame_buf_t::copy(int id, int texture, int offset_x, int offset_y, int width, int height) {
+  gl::ctx.BindFramebuffer(GL_FRAMEBUFFER, (*this)[id]);
+  gl::ctx.ReadBuffer(GL_COLOR_ATTACHMENT0 + id);
+  gl::ctx.BindTexture(GL_TEXTURE_2D, texture);
+  gl::ctx.CopyTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, offset_x, offset_y, width, height);
+}
+
+std::string shader_t::err_str() {
+  int length;
+  ctx.GetShaderiv(handle(), GL_INFO_LOG_LENGTH, &length);
+
+  std::string string;
+  string.resize(length);
+
+  ctx.GetShaderInfoLog(handle(), length, &length, string.data());
+
+  string.resize(length - 1);
+
+  return string;
+}
+
+util::Either<shader_t, std::string> shader_t::compile(const std::string_view &source, GLenum type) {
+  shader_t shader;
+
+  auto data    = source.data();
+  GLint length = source.length();
+
+  shader._shader.el = ctx.CreateShader(type);
+  ctx.ShaderSource(shader.handle(), 1, &data, &length);
+  ctx.CompileShader(shader.handle());
+
+  int status = 0;
+  ctx.GetShaderiv(shader.handle(), GL_COMPILE_STATUS, &status);
+
+  if(!status) {
+    return shader.err_str();
+  }
+
+  return shader;
+}
+
+GLuint shader_t::handle() const {
+  return _shader.el;
+}
+
+buffer_t buffer_t::make(util::buffer_t<GLint> &&offsets, const char *block, const std::string_view &data) {
+  buffer_t buffer;
+  buffer._block   = block;
+  buffer._size    = data.size();
+  buffer._offsets = std::move(offsets);
+
+  ctx.GenBuffers(1, &buffer._buffer.el);
+  ctx.BindBuffer(GL_UNIFORM_BUFFER, buffer.handle());
+  ctx.BufferData(GL_UNIFORM_BUFFER, data.size(), (const std::uint8_t *)data.data(), GL_DYNAMIC_DRAW);
+
+  return buffer;
+}
+
+GLuint buffer_t::handle() const {
+  return _buffer.el;
+}
+
+const char *buffer_t::block() const {
+  return _block;
+}
+
+void buffer_t::update(const std::string_view &view, std::size_t offset) {
+  ctx.BindBuffer(GL_UNIFORM_BUFFER, handle());
+  ctx.BufferSubData(GL_UNIFORM_BUFFER, offset, view.size(), (const void *)view.data());
+}
+
+void buffer_t::update(std::string_view *members, std::size_t count, std::size_t offset) {
+  util::buffer_t<std::uint8_t> buffer { _size };
+
+  for(int x = 0; x < count; ++x) {
+    auto val = members[x];
+
+    std::copy_n((const std::uint8_t *)val.data(), val.size(), &buffer[_offsets[x]]);
+  }
+
+  update(util::view(buffer.begin(), buffer.end()), offset);
+}
+
+std::string program_t::err_str() {
+  int length;
+  ctx.GetProgramiv(handle(), GL_INFO_LOG_LENGTH, &length);
+
+  std::string string;
+  string.resize(length);
+
+  ctx.GetShaderInfoLog(handle(), length, &length, string.data());
+
+  string.resize(length - 1);
+
+  return string;
+}
+
+util::Either<program_t, std::string> program_t::link(const shader_t &vert, const shader_t &frag) {
+  program_t program;
+
+  program._program.el = ctx.CreateProgram();
+
+  ctx.AttachShader(program.handle(), vert.handle());
+  ctx.AttachShader(program.handle(), frag.handle());
+
+  // p_handle stores a copy of the program handle, since program will be moved before
+  // the fail guard funcion is called.
+  auto fg = util::fail_guard([p_handle = program.handle(), &vert, &frag]() {
+    ctx.DetachShader(p_handle, vert.handle());
+    ctx.DetachShader(p_handle, frag.handle());
+  });
+
+  ctx.LinkProgram(program.handle());
+
+  int status = 0;
+  ctx.GetProgramiv(program.handle(), GL_LINK_STATUS, &status);
+
+  if(!status) {
+    return program.err_str();
+  }
+
+  return program;
+}
+
+void program_t::bind(const buffer_t &buffer) {
+  ctx.UseProgram(handle());
+  auto i = ctx.GetUniformBlockIndex(handle(), buffer.block());
+
+  ctx.BindBufferBase(GL_UNIFORM_BUFFER, i, buffer.handle());
+}
+
+std::optional<buffer_t> program_t::uniform(const char *block, std::pair<const char *, std::string_view> *members, std::size_t count) {
+  auto i = ctx.GetUniformBlockIndex(handle(), block);
+  if(i == GL_INVALID_INDEX) {
+    BOOST_LOG(error) << "Couldn't find index of ["sv << block << ']';
+    return std::nullopt;
+  }
+
+  int size;
+  ctx.GetActiveUniformBlockiv(handle(), i, GL_UNIFORM_BLOCK_DATA_SIZE, &size);
+
+  bool error_flag = false;
+
+  util::buffer_t<GLint> offsets { count };
+  auto indices = (std::uint32_t *)alloca(count * sizeof(std::uint32_t));
+  auto names   = (const char **)alloca(count * sizeof(const char *));
+  auto names_p = names;
+
+  std::for_each_n(members, count, [names_p](auto &member) mutable {
+    *names_p++ = std::get<0>(member);
+  });
+
+  std::fill_n(indices, count, GL_INVALID_INDEX);
+  ctx.GetUniformIndices(handle(), count, names, indices);
+
+  for(int x = 0; x < count; ++x) {
+    if(indices[x] == GL_INVALID_INDEX) {
+      error_flag = true;
+
+      BOOST_LOG(error) << "Couldn't find ["sv << block << '.' << members[x].first << ']';
+    }
+  }
+
+  if(error_flag) {
+    return std::nullopt;
+  }
+
+  ctx.GetActiveUniformsiv(handle(), count, indices, GL_UNIFORM_OFFSET, offsets.begin());
+  util::buffer_t<std::uint8_t> buffer { (std::size_t)size };
+
+  for(int x = 0; x < count; ++x) {
+    auto val = std::get<1>(members[x]);
+
+    std::copy_n((const std::uint8_t *)val.data(), val.size(), &buffer[offsets[x]]);
+  }
+
+  return buffer_t::make(std::move(offsets), block, std::string_view { (char *)buffer.begin(), buffer.size() });
+}
+
+GLuint program_t::handle() const {
+  return _program.el;
+}
+
+} // namespace gl
+
+namespace gbm {
+device_destroy_fn device_destroy;
+create_device_fn create_device;
+
+int init() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libgbm.so.1", "libgbm.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<GLADapiproc *, const char *>> funcs {
+    { (GLADapiproc *)&device_destroy, "gbm_device_destroy" },
+    { (GLADapiproc *)&create_device, "gbm_create_device" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+} // namespace gbm
+
+namespace egl {
+constexpr auto EGL_LINUX_DMA_BUF_EXT              = 0x3270;
+constexpr auto EGL_LINUX_DRM_FOURCC_EXT           = 0x3271;
+constexpr auto EGL_DMA_BUF_PLANE0_FD_EXT          = 0x3272;
+constexpr auto EGL_DMA_BUF_PLANE0_OFFSET_EXT      = 0x3273;
+constexpr auto EGL_DMA_BUF_PLANE0_PITCH_EXT       = 0x3274;
+constexpr auto EGL_DMA_BUF_PLANE1_FD_EXT          = 0x3275;
+constexpr auto EGL_DMA_BUF_PLANE1_OFFSET_EXT      = 0x3276;
+constexpr auto EGL_DMA_BUF_PLANE1_PITCH_EXT       = 0x3277;
+constexpr auto EGL_DMA_BUF_PLANE2_FD_EXT          = 0x3278;
+constexpr auto EGL_DMA_BUF_PLANE2_OFFSET_EXT      = 0x3279;
+constexpr auto EGL_DMA_BUF_PLANE2_PITCH_EXT       = 0x327A;
+constexpr auto EGL_DMA_BUF_PLANE3_FD_EXT          = 0x3440;
+constexpr auto EGL_DMA_BUF_PLANE3_OFFSET_EXT      = 0x3441;
+constexpr auto EGL_DMA_BUF_PLANE3_PITCH_EXT       = 0x3442;
+constexpr auto EGL_DMA_BUF_PLANE0_MODIFIER_LO_EXT = 0x3443;
+constexpr auto EGL_DMA_BUF_PLANE0_MODIFIER_HI_EXT = 0x3444;
+constexpr auto EGL_DMA_BUF_PLANE1_MODIFIER_LO_EXT = 0x3445;
+constexpr auto EGL_DMA_BUF_PLANE1_MODIFIER_HI_EXT = 0x3446;
+constexpr auto EGL_DMA_BUF_PLANE2_MODIFIER_LO_EXT = 0x3447;
+constexpr auto EGL_DMA_BUF_PLANE2_MODIFIER_HI_EXT = 0x3448;
+constexpr auto EGL_DMA_BUF_PLANE3_MODIFIER_LO_EXT = 0x3449;
+constexpr auto EGL_DMA_BUF_PLANE3_MODIFIER_HI_EXT = 0x344A;
+
+bool fail() {
+  return eglGetError() != EGL_SUCCESS;
+}
+
+display_t make_display(std::variant<gbm::gbm_t::pointer, wl_display *, _XDisplay *> native_display) {
+  constexpr auto EGL_PLATFORM_GBM_MESA    = 0x31D7;
+  constexpr auto EGL_PLATFORM_WAYLAND_KHR = 0x31D8;
+  constexpr auto EGL_PLATFORM_X11_KHR     = 0x31D5;
+
+  int egl_platform;
+  void *native_display_p;
+
+  switch(native_display.index()) {
+  case 0:
+    egl_platform     = EGL_PLATFORM_GBM_MESA;
+    native_display_p = std::get<0>(native_display);
+    break;
+  case 1:
+    egl_platform     = EGL_PLATFORM_WAYLAND_KHR;
+    native_display_p = std::get<1>(native_display);
+    break;
+  case 2:
+    egl_platform     = EGL_PLATFORM_X11_KHR;
+    native_display_p = std::get<2>(native_display);
+    break;
+  default:
+    BOOST_LOG(error) << "egl::make_display(): Index ["sv << native_display.index() << "] not implemented"sv;
+    return nullptr;
+  }
+
+  // native_display.left() equals native_display.right()
+  display_t display = eglGetPlatformDisplay(egl_platform, native_display_p, nullptr);
+
+  if(fail()) {
+    BOOST_LOG(error) << "Couldn't open EGL display: ["sv << util::hex(eglGetError()).to_string_view() << ']';
+    return nullptr;
+  }
+
+  int major, minor;
+  if(!eglInitialize(display.get(), &major, &minor)) {
+    BOOST_LOG(error) << "Couldn't initialize EGL display: ["sv << util::hex(eglGetError()).to_string_view() << ']';
+    return nullptr;
+  }
+
+  const char *extension_st = eglQueryString(display.get(), EGL_EXTENSIONS);
+  const char *version      = eglQueryString(display.get(), EGL_VERSION);
+  const char *vendor       = eglQueryString(display.get(), EGL_VENDOR);
+  const char *apis         = eglQueryString(display.get(), EGL_CLIENT_APIS);
+
+  BOOST_LOG(debug) << "EGL: ["sv << vendor << "]: version ["sv << version << ']';
+  BOOST_LOG(debug) << "API's supported: ["sv << apis << ']';
+
+  const char *extensions[] {
+    "EGL_KHR_create_context",
+    "EGL_KHR_surfaceless_context",
+    "EGL_EXT_image_dma_buf_import",
+  };
+
+  for(auto ext : extensions) {
+    if(!std::strstr(extension_st, ext)) {
+      BOOST_LOG(error) << "Missing extension: ["sv << ext << ']';
+      return nullptr;
+    }
+  }
+
+  return display;
+}
+
+std::optional<ctx_t> make_ctx(display_t::pointer display) {
+  constexpr int conf_attr[] {
+    EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT, EGL_NONE
+  };
+
+  int count;
+  EGLConfig conf;
+  if(!eglChooseConfig(display, conf_attr, &conf, 1, &count)) {
+    BOOST_LOG(error) << "Couldn't set config attributes: ["sv << util::hex(eglGetError()).to_string_view() << ']';
+    return std::nullopt;
+  }
+
+  if(!eglBindAPI(EGL_OPENGL_API)) {
+    BOOST_LOG(error) << "Couldn't bind API: ["sv << util::hex(eglGetError()).to_string_view() << ']';
+    return std::nullopt;
+  }
+
+  constexpr int attr[] {
+    EGL_CONTEXT_CLIENT_VERSION, 3, EGL_NONE
+  };
+
+  ctx_t ctx { display, eglCreateContext(display, conf, EGL_NO_CONTEXT, attr) };
+  if(fail()) {
+    BOOST_LOG(error) << "Couldn't create EGL context: ["sv << util::hex(eglGetError()).to_string_view() << ']';
+    return std::nullopt;
+  }
+
+  TUPLE_EL_REF(ctx_p, 1, ctx.el);
+  if(!eglMakeCurrent(display, EGL_NO_SURFACE, EGL_NO_SURFACE, ctx_p)) {
+    BOOST_LOG(error) << "Couldn't make current display"sv;
+    return std::nullopt;
+  }
+
+  if(!gladLoadGLContext(&gl::ctx, eglGetProcAddress)) {
+    BOOST_LOG(error) << "Couldn't load OpenGL library"sv;
+    return std::nullopt;
+  }
+
+  BOOST_LOG(debug) << "GL: vendor: "sv << gl::ctx.GetString(GL_VENDOR);
+  BOOST_LOG(debug) << "GL: renderer: "sv << gl::ctx.GetString(GL_RENDERER);
+  BOOST_LOG(debug) << "GL: version: "sv << gl::ctx.GetString(GL_VERSION);
+  BOOST_LOG(debug) << "GL: shader: "sv << gl::ctx.GetString(GL_SHADING_LANGUAGE_VERSION);
+
+  gl::ctx.PixelStorei(GL_UNPACK_ALIGNMENT, 1);
+
+  return ctx;
+}
+
+struct plane_attr_t {
+  EGLAttrib fd;
+  EGLAttrib offset;
+  EGLAttrib pitch;
+  EGLAttrib lo;
+  EGLAttrib hi;
+};
+
+inline plane_attr_t get_plane(std::uint32_t plane_indice) {
+  switch(plane_indice) {
+  case 0:
+    return {
+      EGL_DMA_BUF_PLANE0_FD_EXT,
+      EGL_DMA_BUF_PLANE0_OFFSET_EXT,
+      EGL_DMA_BUF_PLANE0_PITCH_EXT,
+      EGL_DMA_BUF_PLANE0_MODIFIER_LO_EXT,
+      EGL_DMA_BUF_PLANE0_MODIFIER_HI_EXT,
+    };
+  case 1:
+    return {
+      EGL_DMA_BUF_PLANE1_FD_EXT,
+      EGL_DMA_BUF_PLANE1_OFFSET_EXT,
+      EGL_DMA_BUF_PLANE1_PITCH_EXT,
+      EGL_DMA_BUF_PLANE1_MODIFIER_LO_EXT,
+      EGL_DMA_BUF_PLANE1_MODIFIER_HI_EXT,
+    };
+  case 2:
+    return {
+      EGL_DMA_BUF_PLANE2_FD_EXT,
+      EGL_DMA_BUF_PLANE2_OFFSET_EXT,
+      EGL_DMA_BUF_PLANE2_PITCH_EXT,
+      EGL_DMA_BUF_PLANE2_MODIFIER_LO_EXT,
+      EGL_DMA_BUF_PLANE2_MODIFIER_HI_EXT,
+    };
+  case 3:
+    return {
+      EGL_DMA_BUF_PLANE3_FD_EXT,
+      EGL_DMA_BUF_PLANE3_OFFSET_EXT,
+      EGL_DMA_BUF_PLANE3_PITCH_EXT,
+      EGL_DMA_BUF_PLANE3_MODIFIER_LO_EXT,
+      EGL_DMA_BUF_PLANE3_MODIFIER_HI_EXT,
+    };
+  }
+
+  // Avoid warning
+  return {};
+}
+
+std::optional<rgb_t> import_source(display_t::pointer egl_display, const surface_descriptor_t &xrgb) {
+  EGLAttrib attribs[47];
+  int atti        = 0;
+  attribs[atti++] = EGL_WIDTH;
+  attribs[atti++] = xrgb.width;
+  attribs[atti++] = EGL_HEIGHT;
+  attribs[atti++] = xrgb.height;
+  attribs[atti++] = EGL_LINUX_DRM_FOURCC_EXT;
+  attribs[atti++] = xrgb.fourcc;
+
+  for(auto x = 0; x < 4; ++x) {
+    auto fd = xrgb.fds[x];
+
+    if(fd < 0) {
+      continue;
+    }
+
+    auto plane_attr = get_plane(x);
+
+    attribs[atti++] = plane_attr.fd;
+    attribs[atti++] = fd;
+    attribs[atti++] = plane_attr.offset;
+    attribs[atti++] = xrgb.offsets[x];
+    attribs[atti++] = plane_attr.pitch;
+    attribs[atti++] = xrgb.pitches[x];
+
+    if(xrgb.modifier != DRM_FORMAT_MOD_INVALID) {
+      attribs[atti++] = plane_attr.lo;
+      attribs[atti++] = xrgb.modifier & 0xFFFFFFFF;
+      attribs[atti++] = plane_attr.hi;
+      attribs[atti++] = xrgb.modifier >> 32;
+    }
+  }
+  attribs[atti++] = EGL_NONE;
+
+  rgb_t rgb {
+    egl_display,
+    eglCreateImage(egl_display, EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, nullptr, attribs),
+    gl::tex_t::make(1)
+  };
+
+  if(!rgb->xrgb8) {
+    BOOST_LOG(error) << "Couldn't import RGB Image: "sv << util::hex(eglGetError()).to_string_view();
+
+    return std::nullopt;
+  }
+
+  gl::ctx.BindTexture(GL_TEXTURE_2D, rgb->tex[0]);
+  gl::ctx.EGLImageTargetTexture2DOES(GL_TEXTURE_2D, rgb->xrgb8);
+
+  gl::ctx.BindTexture(GL_TEXTURE_2D, 0);
+
+  gl_drain_errors;
+
+  return rgb;
+}
+
+std::optional<nv12_t> import_target(display_t::pointer egl_display, std::array<file_t, nv12_img_t::num_fds> &&fds, const surface_descriptor_t &r8, const surface_descriptor_t &gr88) {
+  EGLAttrib img_attr_planes[2][13] {
+    { EGL_LINUX_DRM_FOURCC_EXT, DRM_FORMAT_R8,
+      EGL_WIDTH, r8.width,
+      EGL_HEIGHT, r8.height,
+      EGL_DMA_BUF_PLANE0_FD_EXT, r8.fds[0],
+      EGL_DMA_BUF_PLANE0_OFFSET_EXT, r8.offsets[0],
+      EGL_DMA_BUF_PLANE0_PITCH_EXT, r8.pitches[0],
+      EGL_NONE },
+
+    { EGL_LINUX_DRM_FOURCC_EXT, DRM_FORMAT_GR88,
+      EGL_WIDTH, gr88.width,
+      EGL_HEIGHT, gr88.height,
+      EGL_DMA_BUF_PLANE0_FD_EXT, r8.fds[0],
+      EGL_DMA_BUF_PLANE0_OFFSET_EXT, gr88.offsets[0],
+      EGL_DMA_BUF_PLANE0_PITCH_EXT, gr88.pitches[0],
+      EGL_NONE },
+  };
+
+  nv12_t nv12 {
+    egl_display,
+    eglCreateImage(egl_display, EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, nullptr, img_attr_planes[0]),
+    eglCreateImage(egl_display, EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, nullptr, img_attr_planes[1]),
+    gl::tex_t::make(2),
+    gl::frame_buf_t::make(2),
+    std::move(fds)
+  };
+
+  if(!nv12->r8 || !nv12->bg88) {
+    BOOST_LOG(error) << "Couldn't create KHR Image"sv;
+
+    return std::nullopt;
+  }
+
+  gl::ctx.BindTexture(GL_TEXTURE_2D, nv12->tex[0]);
+  gl::ctx.EGLImageTargetTexture2DOES(GL_TEXTURE_2D, nv12->r8);
+
+  gl::ctx.BindTexture(GL_TEXTURE_2D, nv12->tex[1]);
+  gl::ctx.EGLImageTargetTexture2DOES(GL_TEXTURE_2D, nv12->bg88);
+
+  nv12->buf.bind(std::begin(nv12->tex), std::end(nv12->tex));
+
+  gl_drain_errors;
+
+  return nv12;
+}
+
+void sws_t::set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) {
+  video::color_t *color_p;
+  switch(colorspace) {
+  case 5: // SWS_CS_SMPTE170M
+    color_p = &video::colors[0];
+    break;
+  case 1: // SWS_CS_ITU709
+    color_p = &video::colors[2];
+    break;
+  case 9: // SWS_CS_BT2020
+  default:
+    BOOST_LOG(warning) << "Colorspace: ["sv << colorspace << "] not yet supported: switching to default"sv;
+    color_p = &video::colors[0];
+  };
+
+  if(color_range > 1) {
+    // Full range
+    ++color_p;
+  }
+
+  std::string_view members[] {
+    util::view(color_p->color_vec_y),
+    util::view(color_p->color_vec_u),
+    util::view(color_p->color_vec_v),
+    util::view(color_p->range_y),
+    util::view(color_p->range_uv),
+  };
+
+  color_matrix.update(members, sizeof(members) / sizeof(decltype(members[0])));
+
+  program[0].bind(color_matrix);
+  program[1].bind(color_matrix);
+}
+
+std::optional<sws_t> sws_t::make(int in_width, int in_height, int out_width, int out_heigth, gl::tex_t &&tex) {
+  sws_t sws;
+
+  sws.serial = std::numeric_limits<std::uint64_t>::max();
+
+  // Ensure aspect ratio is maintained
+  auto scalar       = std::fminf(out_width / (float)in_width, out_heigth / (float)in_height);
+  auto out_width_f  = in_width * scalar;
+  auto out_height_f = in_height * scalar;
+
+  // result is always positive
+  auto offsetX_f = (out_width - out_width_f) / 2;
+  auto offsetY_f = (out_heigth - out_height_f) / 2;
+
+  sws.out_width  = out_width_f;
+  sws.out_height = out_height_f;
+
+  sws.in_width  = in_width;
+  sws.in_height = in_height;
+
+  sws.offsetX = offsetX_f;
+  sws.offsetY = offsetY_f;
+
+  auto width_i = 1.0f / sws.out_width;
+
+  {
+    const char *sources[] {
+      SUNSHINE_SHADERS_DIR "/ConvertUV.frag",
+      SUNSHINE_SHADERS_DIR "/ConvertUV.vert",
+      SUNSHINE_SHADERS_DIR "/ConvertY.frag",
+      SUNSHINE_SHADERS_DIR "/Scene.vert",
+      SUNSHINE_SHADERS_DIR "/Scene.frag",
+    };
+
+    GLenum shader_type[2] {
+      GL_FRAGMENT_SHADER,
+      GL_VERTEX_SHADER,
+    };
+
+    constexpr auto count = sizeof(sources) / sizeof(const char *);
+
+    util::Either<gl::shader_t, std::string> compiled_sources[count];
+
+    bool error_flag = false;
+    for(int x = 0; x < count; ++x) {
+      auto &compiled_source = compiled_sources[x];
+
+      compiled_source = gl::shader_t::compile(read_file(sources[x]), shader_type[x % 2]);
+      gl_drain_errors;
+
+      if(compiled_source.has_right()) {
+        BOOST_LOG(error) << sources[x] << ": "sv << compiled_source.right();
+        error_flag = true;
+      }
+    }
+
+    if(error_flag) {
+      return std::nullopt;
+    }
+
+    auto program = gl::program_t::link(compiled_sources[3].left(), compiled_sources[4].left());
+    if(program.has_right()) {
+      BOOST_LOG(error) << "GL linker: "sv << program.right();
+      return std::nullopt;
+    }
+
+    // Cursor - shader
+    sws.program[2] = std::move(program.left());
+
+    program = gl::program_t::link(compiled_sources[1].left(), compiled_sources[0].left());
+    if(program.has_right()) {
+      BOOST_LOG(error) << "GL linker: "sv << program.right();
+      return std::nullopt;
+    }
+
+    // UV - shader
+    sws.program[1] = std::move(program.left());
+
+    program = gl::program_t::link(compiled_sources[3].left(), compiled_sources[2].left());
+    if(program.has_right()) {
+      BOOST_LOG(error) << "GL linker: "sv << program.right();
+      return std::nullopt;
+    }
+
+    // Y - shader
+    sws.program[0] = std::move(program.left());
+  }
+
+  auto loc_width_i = gl::ctx.GetUniformLocation(sws.program[1].handle(), "width_i");
+  if(loc_width_i < 0) {
+    BOOST_LOG(error) << "Couldn't find uniform [width_i]"sv;
+    return std::nullopt;
+  }
+
+  gl::ctx.UseProgram(sws.program[1].handle());
+  gl::ctx.Uniform1fv(loc_width_i, 1, &width_i);
+
+  auto color_p = &video::colors[0];
+  std::pair<const char *, std::string_view> members[] {
+    std::make_pair("color_vec_y", util::view(color_p->color_vec_y)),
+    std::make_pair("color_vec_u", util::view(color_p->color_vec_u)),
+    std::make_pair("color_vec_v", util::view(color_p->color_vec_v)),
+    std::make_pair("range_y", util::view(color_p->range_y)),
+    std::make_pair("range_uv", util::view(color_p->range_uv)),
+  };
+
+  auto color_matrix = sws.program[0].uniform("ColorMatrix", members, sizeof(members) / sizeof(decltype(members[0])));
+  if(!color_matrix) {
+    return std::nullopt;
+  }
+
+  sws.color_matrix = std::move(*color_matrix);
+
+  sws.tex = std::move(tex);
+
+  sws.cursor_framebuffer = gl::frame_buf_t::make(1);
+  sws.cursor_framebuffer.bind(&sws.tex[0], &sws.tex[1]);
+
+  sws.program[0].bind(sws.color_matrix);
+  sws.program[1].bind(sws.color_matrix);
+
+  gl::ctx.BlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
+
+  gl_drain_errors;
+
+  return std::move(sws);
+}
+
+int sws_t::blank(gl::frame_buf_t &fb, int offsetX, int offsetY, int width, int height) {
+  auto f = [&]() {
+    std::swap(offsetX, this->offsetX);
+    std::swap(offsetY, this->offsetY);
+    std::swap(width, this->out_width);
+    std::swap(height, this->out_height);
+  };
+
+  f();
+  auto fg = util::fail_guard(f);
+
+  return convert(fb);
+}
+
+std::optional<sws_t> sws_t::make(int in_width, int in_height, int out_width, int out_heigth) {
+  auto tex = gl::tex_t::make(2);
+  gl::ctx.BindTexture(GL_TEXTURE_2D, tex[0]);
+  gl::ctx.TexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, in_width, in_height);
+
+  return make(in_width, in_height, out_width, out_heigth, std::move(tex));
+}
+
+void sws_t::load_ram(platf::img_t &img) {
+  loaded_texture = tex[0];
+
+  gl::ctx.BindTexture(GL_TEXTURE_2D, loaded_texture);
+  gl::ctx.TexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, img.width, img.height, GL_BGRA, GL_UNSIGNED_BYTE, img.data);
+}
+
+void sws_t::load_vram(img_descriptor_t &img, int offset_x, int offset_y, int texture) {
+  // When only a sub-part of the image must be encoded...
+  const bool copy = offset_x || offset_y || img.sd.width != in_width || img.sd.height != in_height;
+  if(copy) {
+    auto framebuf = gl::frame_buf_t::make(1);
+    framebuf.bind(&texture, &texture + 1);
+
+    loaded_texture = tex[0];
+    framebuf.copy(0, loaded_texture, offset_x, offset_y, in_width, in_height);
+  }
+  else {
+    loaded_texture = texture;
+  }
+
+  if(img.data) {
+    GLenum attachment = GL_COLOR_ATTACHMENT0;
+
+    gl::ctx.BindFramebuffer(GL_FRAMEBUFFER, cursor_framebuffer[0]);
+    gl::ctx.UseProgram(program[2].handle());
+
+    // When a copy has already been made...
+    if(!copy) {
+      gl::ctx.BindTexture(GL_TEXTURE_2D, texture);
+      gl::ctx.DrawBuffers(1, &attachment);
+
+      gl::ctx.Viewport(0, 0, in_width, in_height);
+      gl::ctx.DrawArrays(GL_TRIANGLES, 0, 3);
+
+      loaded_texture = tex[0];
+    }
+
+    gl::ctx.BindTexture(GL_TEXTURE_2D, tex[1]);
+    if(serial != img.serial) {
+      serial = img.serial;
+
+      gl::ctx.TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, img.width, img.height, 0, GL_BGRA, GL_UNSIGNED_BYTE, img.data);
+    }
+
+    gl::ctx.Enable(GL_BLEND);
+
+    gl::ctx.DrawBuffers(1, &attachment);
+
+#ifndef NDEBUG
+    auto status = gl::ctx.CheckFramebufferStatus(GL_FRAMEBUFFER);
+    if(status != GL_FRAMEBUFFER_COMPLETE) {
+      BOOST_LOG(error) << "Pass Cursor: CheckFramebufferStatus() --> [0x"sv << util::hex(status).to_string_view() << ']';
+      return;
+    }
+#endif
+
+    gl::ctx.Viewport(img.x, img.y, img.width, img.height);
+    gl::ctx.DrawArrays(GL_TRIANGLES, 0, 3);
+
+    gl::ctx.Disable(GL_BLEND);
+
+    gl::ctx.BindTexture(GL_TEXTURE_2D, 0);
+    gl::ctx.BindFramebuffer(GL_FRAMEBUFFER, 0);
+  }
+}
+
+int sws_t::convert(gl::frame_buf_t &fb) {
+  gl::ctx.BindTexture(GL_TEXTURE_2D, loaded_texture);
+
+  GLenum attachments[] {
+    GL_COLOR_ATTACHMENT0,
+    GL_COLOR_ATTACHMENT1
+  };
+
+  for(int x = 0; x < sizeof(attachments) / sizeof(decltype(attachments[0])); ++x) {
+    gl::ctx.BindFramebuffer(GL_FRAMEBUFFER, fb[x]);
+    gl::ctx.DrawBuffers(1, &attachments[x]);
+
+#ifndef NDEBUG
+    auto status = gl::ctx.CheckFramebufferStatus(GL_FRAMEBUFFER);
+    if(status != GL_FRAMEBUFFER_COMPLETE) {
+      BOOST_LOG(error) << "Pass "sv << x << ": CheckFramebufferStatus() --> [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+#endif
+
+    gl::ctx.UseProgram(program[x].handle());
+    gl::ctx.Viewport(offsetX / (x + 1), offsetY / (x + 1), out_width / (x + 1), out_height / (x + 1));
+    gl::ctx.DrawArrays(GL_TRIANGLES, 0, 3);
+  }
+
+  gl::ctx.BindTexture(GL_TEXTURE_2D, 0);
+
+  gl::ctx.Flush();
+
+  return 0;
+}
+} // namespace egl
+
+void free_frame(AVFrame *frame) {
+  av_frame_free(&frame);
+}

src/platform/linux/graphics.h

@@ -0,0 +1,319 @@
+#ifndef SUNSHINE_PLATFORM_LINUX_OPENGL_H
+#define SUNSHINE_PLATFORM_LINUX_OPENGL_H
+
+#include <optional>
+#include <string_view>
+
+#include <glad/egl.h>
+#include <glad/gl.h>
+
+#include "misc.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+#define SUNSHINE_STRINGIFY_HELPER(x) #x
+#define SUNSHINE_STRINGIFY(x) SUNSHINE_STRINGIFY_HELPER(x)
+#define gl_drain_errors_helper(x) gl::drain_errors(x)
+#define gl_drain_errors gl_drain_errors_helper(__FILE__ ":" SUNSHINE_STRINGIFY(__LINE__))
+
+extern "C" int close(int __fd);
+
+// X11 Display
+extern "C" struct _XDisplay;
+
+struct AVFrame;
+void free_frame(AVFrame *frame);
+
+using frame_t = util::safe_ptr<AVFrame, free_frame>;
+
+namespace gl {
+extern GladGLContext ctx;
+void drain_errors(const std::string_view &prefix);
+
+class tex_t : public util::buffer_t<GLuint> {
+  using util::buffer_t<GLuint>::buffer_t;
+
+public:
+  tex_t(tex_t &&) = default;
+  tex_t &operator=(tex_t &&) = default;
+
+  ~tex_t();
+
+  static tex_t make(std::size_t count);
+};
+
+class frame_buf_t : public util::buffer_t<GLuint> {
+  using util::buffer_t<GLuint>::buffer_t;
+
+public:
+  frame_buf_t(frame_buf_t &&) = default;
+  frame_buf_t &operator=(frame_buf_t &&) = default;
+
+  ~frame_buf_t();
+
+  static frame_buf_t make(std::size_t count);
+
+  inline void bind(std::nullptr_t, std::nullptr_t) {
+    int x = 0;
+    for(auto fb : (*this)) {
+      ctx.BindFramebuffer(GL_FRAMEBUFFER, fb);
+      ctx.FramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + x, 0, 0);
+
+      ++x;
+    }
+    return;
+  }
+
+  template<class It>
+  void bind(It it_begin, It it_end) {
+    using namespace std::literals;
+    if(std::distance(it_begin, it_end) > size()) {
+      BOOST_LOG(warning) << "To many elements to bind"sv;
+      return;
+    }
+
+    int x = 0;
+    std::for_each(it_begin, it_end, [&](auto tex) {
+      ctx.BindFramebuffer(GL_FRAMEBUFFER, (*this)[x]);
+      ctx.BindTexture(GL_TEXTURE_2D, tex);
+
+      ctx.FramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + x, tex, 0);
+
+      ++x;
+    });
+  }
+
+  /**
+   * Copies a part of the framebuffer to texture
+   */
+  void copy(int id, int texture, int offset_x, int offset_y, int width, int height);
+};
+
+class shader_t {
+  KITTY_USING_MOVE_T(shader_internal_t, GLuint, std::numeric_limits<GLuint>::max(), {
+    if(el != std::numeric_limits<GLuint>::max()) {
+      ctx.DeleteShader(el);
+    }
+  });
+
+public:
+  std::string err_str();
+
+  static util::Either<shader_t, std::string> compile(const std::string_view &source, GLenum type);
+
+  GLuint handle() const;
+
+private:
+  shader_internal_t _shader;
+};
+
+class buffer_t {
+  KITTY_USING_MOVE_T(buffer_internal_t, GLuint, std::numeric_limits<GLuint>::max(), {
+    if(el != std::numeric_limits<GLuint>::max()) {
+      ctx.DeleteBuffers(1, &el);
+    }
+  });
+
+public:
+  static buffer_t make(util::buffer_t<GLint> &&offsets, const char *block, const std::string_view &data);
+
+  GLuint handle() const;
+
+  const char *block() const;
+
+  void update(const std::string_view &view, std::size_t offset = 0);
+  void update(std::string_view *members, std::size_t count, std::size_t offset = 0);
+
+private:
+  const char *_block;
+
+  std::size_t _size;
+
+  util::buffer_t<GLint> _offsets;
+
+  buffer_internal_t _buffer;
+};
+
+class program_t {
+  KITTY_USING_MOVE_T(program_internal_t, GLuint, std::numeric_limits<GLuint>::max(), {
+    if(el != std::numeric_limits<GLuint>::max()) {
+      ctx.DeleteProgram(el);
+    }
+  });
+
+public:
+  std::string err_str();
+
+  static util::Either<program_t, std::string> link(const shader_t &vert, const shader_t &frag);
+
+  void bind(const buffer_t &buffer);
+
+  std::optional<buffer_t> uniform(const char *block, std::pair<const char *, std::string_view> *members, std::size_t count);
+
+  GLuint handle() const;
+
+private:
+  program_internal_t _program;
+};
+} // namespace gl
+
+namespace gbm {
+struct device;
+typedef void (*device_destroy_fn)(device *gbm);
+typedef device *(*create_device_fn)(int fd);
+
+extern device_destroy_fn device_destroy;
+extern create_device_fn create_device;
+
+using gbm_t = util::dyn_safe_ptr<device, &device_destroy>;
+
+int init();
+
+} // namespace gbm
+
+namespace egl {
+using display_t = util::dyn_safe_ptr_v2<void, EGLBoolean, &eglTerminate>;
+
+struct rgb_img_t {
+  display_t::pointer display;
+  EGLImage xrgb8;
+
+  gl::tex_t tex;
+};
+
+struct nv12_img_t {
+  display_t::pointer display;
+  EGLImage r8;
+  EGLImage bg88;
+
+  gl::tex_t tex;
+  gl::frame_buf_t buf;
+
+  // sizeof(va::DRMPRIMESurfaceDescriptor::objects) / sizeof(va::DRMPRIMESurfaceDescriptor::objects[0]);
+  static constexpr std::size_t num_fds = 4;
+
+  std::array<file_t, num_fds> fds;
+};
+
+KITTY_USING_MOVE_T(rgb_t, rgb_img_t, , {
+  if(el.xrgb8) {
+    eglDestroyImage(el.display, el.xrgb8);
+  }
+});
+
+KITTY_USING_MOVE_T(nv12_t, nv12_img_t, , {
+  if(el.r8) {
+    eglDestroyImage(el.display, el.r8);
+  }
+
+  if(el.bg88) {
+    eglDestroyImage(el.display, el.bg88);
+  }
+});
+
+KITTY_USING_MOVE_T(ctx_t, (std::tuple<display_t::pointer, EGLContext>), , {
+  TUPLE_2D_REF(disp, ctx, el);
+  if(ctx) {
+    eglMakeCurrent(disp, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
+    eglDestroyContext(disp, ctx);
+  }
+});
+
+struct surface_descriptor_t {
+  int width;
+  int height;
+  int fds[4];
+  std::uint32_t fourcc;
+  std::uint64_t modifier;
+  std::uint32_t pitches[4];
+  std::uint32_t offsets[4];
+};
+
+display_t make_display(std::variant<gbm::gbm_t::pointer, wl_display *, _XDisplay *> native_display);
+std::optional<ctx_t> make_ctx(display_t::pointer display);
+
+std::optional<rgb_t> import_source(
+  display_t::pointer egl_display,
+  const surface_descriptor_t &xrgb);
+
+std::optional<nv12_t> import_target(
+  display_t::pointer egl_display,
+  std::array<file_t, nv12_img_t::num_fds> &&fds,
+  const surface_descriptor_t &r8, const surface_descriptor_t &gr88);
+
+class cursor_t : public platf::img_t {
+public:
+  int x, y;
+
+  unsigned long serial;
+
+  std::vector<std::uint8_t> buffer;
+};
+
+// Allow cursor and the underlying image to be kept together
+class img_descriptor_t : public cursor_t {
+public:
+  ~img_descriptor_t() {
+    reset();
+  }
+
+  void reset() {
+    for(auto x = 0; x < 4; ++x) {
+      if(sd.fds[x] >= 0) {
+        close(sd.fds[x]);
+
+        sd.fds[x] = -1;
+      }
+    }
+  }
+
+  surface_descriptor_t sd;
+
+  // Increment sequence when new rgb_t needs to be created
+  std::uint64_t sequence;
+};
+
+class sws_t {
+public:
+  static std::optional<sws_t> make(int in_width, int in_height, int out_width, int out_heigth, gl::tex_t &&tex);
+  static std::optional<sws_t> make(int in_width, int in_height, int out_width, int out_heigth);
+
+  // Convert the loaded image into the first two framebuffers
+  int convert(gl::frame_buf_t &fb);
+
+  // Make an area of the image black
+  int blank(gl::frame_buf_t &fb, int offsetX, int offsetY, int width, int height);
+
+  void load_ram(platf::img_t &img);
+  void load_vram(img_descriptor_t &img, int offset_x, int offset_y, int texture);
+
+  void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range);
+
+  // The first texture is the monitor image.
+  // The second texture is the cursor image
+  gl::tex_t tex;
+
+  // The cursor image will be blended into this framebuffer
+  gl::frame_buf_t cursor_framebuffer;
+  gl::frame_buf_t copy_framebuffer;
+
+  // Y - shader, UV - shader, Cursor - shader
+  gl::program_t program[3];
+  gl::buffer_t color_matrix;
+
+  int out_width, out_height;
+  int in_width, in_height;
+  int offsetX, offsetY;
+
+  // Pointer to the texture to be converted to nv12
+  int loaded_texture;
+
+  // Store latest cursor for load_vram
+  std::uint64_t serial;
+};
+
+bool fail();
+} // namespace egl
+
+#endif

src/platform/linux/misc.cpp

@@ -0,0 +1,302 @@
+#include <arpa/inet.h>
+#include <dlfcn.h>
+#include <fcntl.h>
+#include <ifaddrs.h>
+#include <pwd.h>
+#include <unistd.h>
+
+#include <fstream>
+
+#include "graphics.h"
+#include "misc.h"
+#include "vaapi.h"
+
+#include "src/main.h"
+#include "src/platform/common.h"
+
+#ifdef __GNUC__
+#define SUNSHINE_GNUC_EXTENSION __extension__
+#else
+#define SUNSHINE_GNUC_EXTENSION
+#endif
+
+using namespace std::literals;
+namespace fs = std::filesystem;
+
+window_system_e window_system;
+
+namespace dyn {
+void *handle(const std::vector<const char *> &libs) {
+  void *handle;
+
+  for(auto lib : libs) {
+    handle = dlopen(lib, RTLD_LAZY | RTLD_LOCAL);
+    if(handle) {
+      return handle;
+    }
+  }
+
+  std::stringstream ss;
+  ss << "Couldn't find any of the following libraries: ["sv << libs.front();
+  std::for_each(std::begin(libs) + 1, std::end(libs), [&](auto lib) {
+    ss << ", "sv << lib;
+  });
+
+  ss << ']';
+
+  BOOST_LOG(error) << ss.str();
+
+  return nullptr;
+}
+
+int load(void *handle, const std::vector<std::tuple<apiproc *, const char *>> &funcs, bool strict) {
+  int err = 0;
+  for(auto &func : funcs) {
+    TUPLE_2D_REF(fn, name, func);
+
+    *fn = SUNSHINE_GNUC_EXTENSION(apiproc) dlsym(handle, name);
+
+    if(!*fn && strict) {
+      BOOST_LOG(error) << "Couldn't find function: "sv << name;
+
+      err = -1;
+    }
+  }
+
+  return err;
+}
+} // namespace dyn
+namespace platf {
+using ifaddr_t = util::safe_ptr<ifaddrs, freeifaddrs>;
+
+ifaddr_t get_ifaddrs() {
+  ifaddrs *p { nullptr };
+
+  getifaddrs(&p);
+
+  return ifaddr_t { p };
+}
+
+fs::path appdata() {
+  const char *homedir;
+  if((homedir = getenv("HOME")) == nullptr) {
+    homedir = getpwuid(geteuid())->pw_dir;
+  }
+
+  return fs::path { homedir } / ".config/sunshine"sv;
+}
+
+std::string from_sockaddr(const sockaddr *const ip_addr) {
+  char data[INET6_ADDRSTRLEN];
+
+  auto family = ip_addr->sa_family;
+  if(family == AF_INET6) {
+    inet_ntop(AF_INET6, &((sockaddr_in6 *)ip_addr)->sin6_addr, data,
+      INET6_ADDRSTRLEN);
+  }
+
+  if(family == AF_INET) {
+    inet_ntop(AF_INET, &((sockaddr_in *)ip_addr)->sin_addr, data,
+      INET_ADDRSTRLEN);
+  }
+
+  return std::string { data };
+}
+
+std::pair<std::uint16_t, std::string> from_sockaddr_ex(const sockaddr *const ip_addr) {
+  char data[INET6_ADDRSTRLEN];
+
+  auto family = ip_addr->sa_family;
+  std::uint16_t port;
+  if(family == AF_INET6) {
+    inet_ntop(AF_INET6, &((sockaddr_in6 *)ip_addr)->sin6_addr, data,
+      INET6_ADDRSTRLEN);
+    port = ((sockaddr_in6 *)ip_addr)->sin6_port;
+  }
+
+  if(family == AF_INET) {
+    inet_ntop(AF_INET, &((sockaddr_in *)ip_addr)->sin_addr, data,
+      INET_ADDRSTRLEN);
+    port = ((sockaddr_in *)ip_addr)->sin_port;
+  }
+
+  return { port, std::string { data } };
+}
+
+std::string get_mac_address(const std::string_view &address) {
+  auto ifaddrs = get_ifaddrs();
+  for(auto pos = ifaddrs.get(); pos != nullptr; pos = pos->ifa_next) {
+    if(pos->ifa_addr && address == from_sockaddr(pos->ifa_addr)) {
+      std::ifstream mac_file("/sys/class/net/"s + pos->ifa_name + "/address");
+      if(mac_file.good()) {
+        std::string mac_address;
+        std::getline(mac_file, mac_address);
+        return mac_address;
+      }
+    }
+  }
+
+  BOOST_LOG(warning) << "Unable to find MAC address for "sv << address;
+  return "00:00:00:00:00:00"s;
+}
+
+namespace source {
+enum source_e : std::size_t {
+#ifdef SUNSHINE_BUILD_CUDA
+  NVFBC,
+#endif
+#ifdef SUNSHINE_BUILD_WAYLAND
+  WAYLAND,
+#endif
+#ifdef SUNSHINE_BUILD_DRM
+  KMS,
+#endif
+#ifdef SUNSHINE_BUILD_X11
+  X11,
+#endif
+  MAX_FLAGS
+};
+} // namespace source
+
+static std::bitset<source::MAX_FLAGS> sources;
+
+#ifdef SUNSHINE_BUILD_CUDA
+std::vector<std::string> nvfbc_display_names();
+std::shared_ptr<display_t> nvfbc_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate);
+
+bool verify_nvfbc() {
+  return !nvfbc_display_names().empty();
+}
+#endif
+
+#ifdef SUNSHINE_BUILD_WAYLAND
+std::vector<std::string> wl_display_names();
+std::shared_ptr<display_t> wl_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate);
+
+bool verify_wl() {
+  return window_system == window_system_e::WAYLAND && !wl_display_names().empty();
+}
+#endif
+
+#ifdef SUNSHINE_BUILD_DRM
+std::vector<std::string> kms_display_names();
+std::shared_ptr<display_t> kms_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate);
+
+bool verify_kms() {
+  return !kms_display_names().empty();
+}
+#endif
+
+#ifdef SUNSHINE_BUILD_X11
+std::vector<std::string> x11_display_names();
+std::shared_ptr<display_t> x11_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate);
+
+bool verify_x11() {
+  return window_system == window_system_e::X11 && !x11_display_names().empty();
+}
+#endif
+
+std::vector<std::string> display_names(mem_type_e hwdevice_type) {
+#ifdef SUNSHINE_BUILD_CUDA
+  // display using NvFBC only supports mem_type_e::cuda
+  if(sources[source::NVFBC] && hwdevice_type == mem_type_e::cuda) return nvfbc_display_names();
+#endif
+#ifdef SUNSHINE_BUILD_WAYLAND
+  if(sources[source::WAYLAND]) return wl_display_names();
+#endif
+#ifdef SUNSHINE_BUILD_DRM
+  if(sources[source::KMS]) return kms_display_names();
+#endif
+#ifdef SUNSHINE_BUILD_X11
+  if(sources[source::X11]) return x11_display_names();
+#endif
+  return {};
+}
+
+std::shared_ptr<display_t> display(mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+#ifdef SUNSHINE_BUILD_CUDA
+  if(sources[source::NVFBC] && hwdevice_type == mem_type_e::cuda) {
+    BOOST_LOG(info) << "Screencasting with NvFBC"sv;
+    return nvfbc_display(hwdevice_type, display_name, framerate);
+  }
+#endif
+#ifdef SUNSHINE_BUILD_WAYLAND
+  if(sources[source::WAYLAND]) {
+    BOOST_LOG(info) << "Screencasting with Wayland's protocol"sv;
+    return wl_display(hwdevice_type, display_name, framerate);
+  }
+#endif
+#ifdef SUNSHINE_BUILD_DRM
+  if(sources[source::KMS]) {
+    BOOST_LOG(info) << "Screencasting with KMS"sv;
+    return kms_display(hwdevice_type, display_name, framerate);
+  }
+#endif
+#ifdef SUNSHINE_BUILD_X11
+  if(sources[source::X11]) {
+    BOOST_LOG(info) << "Screencasting with X11"sv;
+    return x11_display(hwdevice_type, display_name, framerate);
+  }
+#endif
+
+  return nullptr;
+}
+
+std::unique_ptr<deinit_t> init() {
+  // These are allowed to fail.
+  gbm::init();
+  va::init();
+
+  window_system = window_system_e::NONE;
+#ifdef SUNSHINE_BUILD_WAYLAND
+  if(std::getenv("WAYLAND_DISPLAY")) {
+    window_system = window_system_e::WAYLAND;
+  }
+#endif
+#if defined(SUNSHINE_BUILD_X11) || defined(SUNSHINE_BUILD_CUDA)
+  if(std::getenv("DISPLAY") && window_system != window_system_e::WAYLAND) {
+    if(std::getenv("WAYLAND_DISPLAY")) {
+      BOOST_LOG(warning) << "Wayland detected, yet sunshine will use X11 for screencasting, screencasting will only work on XWayland applications"sv;
+    }
+
+    window_system = window_system_e::X11;
+  }
+#endif
+#ifdef SUNSHINE_BUILD_CUDA
+  if(verify_nvfbc()) {
+    sources[source::NVFBC] = true;
+  }
+#endif
+#ifdef SUNSHINE_BUILD_WAYLAND
+  if(verify_wl()) {
+    sources[source::WAYLAND] = true;
+  }
+#endif
+#ifdef SUNSHINE_BUILD_DRM
+  if(verify_kms()) {
+    if(window_system == window_system_e::WAYLAND) {
+      // On Wayland, using KMS, the cursor is unreliable.
+      // Hide it by default
+      display_cursor = false;
+    }
+
+    sources[source::KMS] = true;
+  }
+#endif
+#ifdef SUNSHINE_BUILD_X11
+  if(verify_x11()) {
+    sources[source::X11] = true;
+  }
+#endif
+
+  if(sources.none()) {
+    return nullptr;
+  }
+
+  if(!gladLoaderLoadEGL(EGL_NO_DISPLAY) || !eglGetPlatformDisplay) {
+    BOOST_LOG(warning) << "Couldn't load EGL library"sv;
+  }
+
+  return std::make_unique<deinit_t>();
+}
+} // namespace platf

src/platform/linux/misc.h

@@ -0,0 +1,31 @@
+#ifndef SUNSHINE_PLATFORM_MISC_H
+#define SUNSHINE_PLATFORM_MISC_H
+
+#include <unistd.h>
+#include <vector>
+
+#include "src/utility.h"
+
+KITTY_USING_MOVE_T(file_t, int, -1, {
+  if(el >= 0) {
+    close(el);
+  }
+});
+
+enum class window_system_e {
+  NONE,
+  X11,
+  WAYLAND,
+};
+
+extern window_system_e window_system;
+
+namespace dyn {
+typedef void (*apiproc)(void);
+
+int load(void *handle, const std::vector<std::tuple<apiproc *, const char *>> &funcs, bool strict = true);
+void *handle(const std::vector<const char *> &libs);
+
+} // namespace dyn
+
+#endif

src/platform/linux/publish.cpp

@@ -0,0 +1,429 @@
+
+// adapted from https://www.avahi.org/doxygen/html/client-publish-service_8c-example.html
+#include <thread>
+
+#include "misc.h"
+#include "src/main.h"
+#include "src/nvhttp.h"
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+using namespace std::literals;
+
+namespace avahi {
+
+/** Error codes used by avahi */
+enum err_e {
+  OK                      = 0,  /**< OK */
+  ERR_FAILURE             = -1, /**< Generic error code */
+  ERR_BAD_STATE           = -2, /**< Object was in a bad state */
+  ERR_INVALID_HOST_NAME   = -3, /**< Invalid host name */
+  ERR_INVALID_DOMAIN_NAME = -4, /**< Invalid domain name */
+  ERR_NO_NETWORK          = -5, /**< No suitable network protocol available */
+  ERR_INVALID_TTL         = -6, /**< Invalid DNS TTL */
+  ERR_IS_PATTERN          = -7, /**< RR key is pattern */
+  ERR_COLLISION           = -8, /**< Name collision */
+  ERR_INVALID_RECORD      = -9, /**< Invalid RR */
+
+  ERR_INVALID_SERVICE_NAME = -10, /**< Invalid service name */
+  ERR_INVALID_SERVICE_TYPE = -11, /**< Invalid service type */
+  ERR_INVALID_PORT         = -12, /**< Invalid port number */
+  ERR_INVALID_KEY          = -13, /**< Invalid key */
+  ERR_INVALID_ADDRESS      = -14, /**< Invalid address */
+  ERR_TIMEOUT              = -15, /**< Timeout reached */
+  ERR_TOO_MANY_CLIENTS     = -16, /**< Too many clients */
+  ERR_TOO_MANY_OBJECTS     = -17, /**< Too many objects */
+  ERR_TOO_MANY_ENTRIES     = -18, /**< Too many entries */
+  ERR_OS                   = -19, /**< OS error */
+
+  ERR_ACCESS_DENIED     = -20, /**< Access denied */
+  ERR_INVALID_OPERATION = -21, /**< Invalid operation */
+  ERR_DBUS_ERROR        = -22, /**< An unexpected D-Bus error occurred */
+  ERR_DISCONNECTED      = -23, /**< Daemon connection failed */
+  ERR_NO_MEMORY         = -24, /**< Memory exhausted */
+  ERR_INVALID_OBJECT    = -25, /**< The object passed to this function was invalid */
+  ERR_NO_DAEMON         = -26, /**< Daemon not running */
+  ERR_INVALID_INTERFACE = -27, /**< Invalid interface */
+  ERR_INVALID_PROTOCOL  = -28, /**< Invalid protocol */
+  ERR_INVALID_FLAGS     = -29, /**< Invalid flags */
+
+  ERR_NOT_FOUND               = -30, /**< Not found */
+  ERR_INVALID_CONFIG          = -31, /**< Configuration error */
+  ERR_VERSION_MISMATCH        = -32, /**< Verson mismatch */
+  ERR_INVALID_SERVICE_SUBTYPE = -33, /**< Invalid service subtype */
+  ERR_INVALID_PACKET          = -34, /**< Invalid packet */
+  ERR_INVALID_DNS_ERROR       = -35, /**< Invlaid DNS return code */
+  ERR_DNS_FORMERR             = -36, /**< DNS Error: Form error */
+  ERR_DNS_SERVFAIL            = -37, /**< DNS Error: Server Failure */
+  ERR_DNS_NXDOMAIN            = -38, /**< DNS Error: No such domain */
+  ERR_DNS_NOTIMP              = -39, /**< DNS Error: Not implemented */
+
+  ERR_DNS_REFUSED       = -40, /**< DNS Error: Operation refused */
+  ERR_DNS_YXDOMAIN      = -41,
+  ERR_DNS_YXRRSET       = -42,
+  ERR_DNS_NXRRSET       = -43,
+  ERR_DNS_NOTAUTH       = -44, /**< DNS Error: Not authorized */
+  ERR_DNS_NOTZONE       = -45,
+  ERR_INVALID_RDATA     = -46, /**< Invalid RDATA */
+  ERR_INVALID_DNS_CLASS = -47, /**< Invalid DNS class */
+  ERR_INVALID_DNS_TYPE  = -48, /**< Invalid DNS type */
+  ERR_NOT_SUPPORTED     = -49, /**< Not supported */
+
+  ERR_NOT_PERMITTED    = -50, /**< Operation not permitted */
+  ERR_INVALID_ARGUMENT = -51, /**< Invalid argument */
+  ERR_IS_EMPTY         = -52, /**< Is empty */
+  ERR_NO_CHANGE        = -53, /**< The requested operation is invalid because it is redundant */
+
+  ERR_MAX = -54
+};
+
+constexpr auto IF_UNSPEC = -1;
+enum proto {
+  PROTO_INET   = 0, /**< IPv4 */
+  PROTO_INET6  = 1, /**< IPv6 */
+  PROTO_UNSPEC = -1 /**< Unspecified/all protocol(s) */
+};
+
+enum ServerState {
+  SERVER_INVALID,     /**< Invalid state (initial) */
+  SERVER_REGISTERING, /**< Host RRs are being registered */
+  SERVER_RUNNING,     /**< All host RRs have been established */
+  SERVER_COLLISION,   /**< There is a collision with a host RR. All host RRs have been withdrawn, the user should set a new host name via avahi_server_set_host_name() */
+  SERVER_FAILURE      /**< Some fatal failure happened, the server is unable to proceed */
+};
+
+enum ClientState {
+  CLIENT_S_REGISTERING = SERVER_REGISTERING, /**< Server state: REGISTERING */
+  CLIENT_S_RUNNING     = SERVER_RUNNING,     /**< Server state: RUNNING */
+  CLIENT_S_COLLISION   = SERVER_COLLISION,   /**< Server state: COLLISION */
+  CLIENT_FAILURE       = 100,                /**< Some kind of error happened on the client side */
+  CLIENT_CONNECTING    = 101                 /**< We're still connecting. This state is only entered when AVAHI_CLIENT_NO_FAIL has been passed to avahi_client_new() and the daemon is not yet available. */
+};
+
+enum EntryGroupState {
+  ENTRY_GROUP_UNCOMMITED,  /**< The group has not yet been commited, the user must still call avahi_entry_group_commit() */
+  ENTRY_GROUP_REGISTERING, /**< The entries of the group are currently being registered */
+  ENTRY_GROUP_ESTABLISHED, /**< The entries have successfully been established */
+  ENTRY_GROUP_COLLISION,   /**< A name collision for one of the entries in the group has been detected, the entries have been withdrawn */
+  ENTRY_GROUP_FAILURE      /**< Some kind of failure happened, the entries have been withdrawn */
+};
+
+enum ClientFlags {
+  CLIENT_IGNORE_USER_CONFIG = 1, /**< Don't read user configuration */
+  CLIENT_NO_FAIL            = 2  /**< Don't fail if the daemon is not available when avahi_client_new() is called, instead enter CLIENT_CONNECTING state and wait for the daemon to appear */
+};
+
+/** Some flags for publishing functions */
+enum PublishFlags {
+  PUBLISH_UNIQUE         = 1,  /**< For raw records: The RRset is intended to be unique */
+  PUBLISH_NO_PROBE       = 2,  /**< For raw records: Though the RRset is intended to be unique no probes shall be sent */
+  PUBLISH_NO_ANNOUNCE    = 4,  /**< For raw records: Do not announce this RR to other hosts */
+  PUBLISH_ALLOW_MULTIPLE = 8,  /**< For raw records: Allow multiple local records of this type, even if they are intended to be unique */
+                               /** \cond fulldocs */
+  PUBLISH_NO_REVERSE = 16,     /**< For address records: don't create a reverse (PTR) entry */
+  PUBLISH_NO_COOKIE  = 32,     /**< For service records: do not implicitly add the local service cookie to TXT data */
+                               /** \endcond */
+  PUBLISH_UPDATE = 64,         /**< Update existing records instead of adding new ones */
+                               /** \cond fulldocs */
+  PUBLISH_USE_WIDE_AREA = 128, /**< Register the record using wide area DNS (i.e. unicast DNS update) */
+  PUBLISH_USE_MULTICAST = 256  /**< Register the record using multicast DNS */
+  /** \endcond */
+};
+
+using IfIndex  = int;
+using Protocol = int;
+
+struct EntryGroup;
+struct Poll;
+struct SimplePoll;
+struct Client;
+
+typedef void (*ClientCallback)(Client *, ClientState, void *userdata);
+typedef void (*EntryGroupCallback)(EntryGroup *g, EntryGroupState state, void *userdata);
+
+typedef void (*free_fn)(void *);
+
+typedef Client *(*client_new_fn)(const Poll *poll_api, ClientFlags flags, ClientCallback callback, void *userdata, int *error);
+typedef void (*client_free_fn)(Client *);
+typedef char *(*alternative_service_name_fn)(char *);
+
+typedef Client *(*entry_group_get_client_fn)(EntryGroup *);
+
+typedef EntryGroup *(*entry_group_new_fn)(Client *, EntryGroupCallback, void *userdata);
+typedef int (*entry_group_add_service_fn)(
+  EntryGroup *group,
+  IfIndex interface,
+  Protocol protocol,
+  PublishFlags flags,
+  const char *name,
+  const char *type,
+  const char *domain,
+  const char *host,
+  uint16_t port,
+  ...);
+
+typedef int (*entry_group_is_empty_fn)(EntryGroup *);
+typedef int (*entry_group_reset_fn)(EntryGroup *);
+typedef int (*entry_group_commit_fn)(EntryGroup *);
+
+typedef char *(*strdup_fn)(const char *);
+typedef char *(*strerror_fn)(int);
+typedef int (*client_errno_fn)(Client *);
+
+typedef Poll *(*simple_poll_get_fn)(SimplePoll *);
+typedef int (*simple_poll_loop_fn)(SimplePoll *);
+typedef void (*simple_poll_quit_fn)(SimplePoll *);
+typedef SimplePoll *(*simple_poll_new_fn)();
+typedef void (*simple_poll_free_fn)(SimplePoll *);
+
+free_fn free;
+client_new_fn client_new;
+client_free_fn client_free;
+alternative_service_name_fn alternative_service_name;
+entry_group_get_client_fn entry_group_get_client;
+entry_group_new_fn entry_group_new;
+entry_group_add_service_fn entry_group_add_service;
+entry_group_is_empty_fn entry_group_is_empty;
+entry_group_reset_fn entry_group_reset;
+entry_group_commit_fn entry_group_commit;
+strdup_fn strdup;
+strerror_fn strerror;
+client_errno_fn client_errno;
+simple_poll_get_fn simple_poll_get;
+simple_poll_loop_fn simple_poll_loop;
+simple_poll_quit_fn simple_poll_quit;
+simple_poll_new_fn simple_poll_new;
+simple_poll_free_fn simple_poll_free;
+
+
+int init_common() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libavahi-common.so.3", "libavahi-common.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&alternative_service_name, "avahi_alternative_service_name" },
+    { (dyn::apiproc *)&free, "avahi_free" },
+    { (dyn::apiproc *)&strdup, "avahi_strdup" },
+    { (dyn::apiproc *)&strerror, "avahi_strerror" },
+    { (dyn::apiproc *)&simple_poll_get, "avahi_simple_poll_get" },
+    { (dyn::apiproc *)&simple_poll_loop, "avahi_simple_poll_loop" },
+    { (dyn::apiproc *)&simple_poll_quit, "avahi_simple_poll_quit" },
+    { (dyn::apiproc *)&simple_poll_new, "avahi_simple_poll_new" },
+    { (dyn::apiproc *)&simple_poll_free, "avahi_simple_poll_free" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+int init_client() {
+  if(init_common()) {
+    return -1;
+  }
+
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libavahi-client.so.3", "libavahi-client.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&client_new, "avahi_client_new" },
+    { (dyn::apiproc *)&client_free, "avahi_client_free" },
+    { (dyn::apiproc *)&entry_group_get_client, "avahi_entry_group_get_client" },
+    { (dyn::apiproc *)&entry_group_new, "avahi_entry_group_new" },
+    { (dyn::apiproc *)&entry_group_add_service, "avahi_entry_group_add_service" },
+    { (dyn::apiproc *)&entry_group_is_empty, "avahi_entry_group_is_empty" },
+    { (dyn::apiproc *)&entry_group_reset, "avahi_entry_group_reset" },
+    { (dyn::apiproc *)&entry_group_commit, "avahi_entry_group_commit" },
+    { (dyn::apiproc *)&client_errno, "avahi_client_errno" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+} // namespace avahi
+
+namespace platf::publish {
+
+template<class T>
+void free(T *p) {
+  avahi::free(p);
+}
+
+template<class T>
+using ptr_t    = util::safe_ptr<T, free<T>>;
+using client_t = util::dyn_safe_ptr<avahi::Client, &avahi::client_free>;
+using poll_t   = util::dyn_safe_ptr<avahi::SimplePoll, &avahi::simple_poll_free>;
+
+avahi::EntryGroup *group = nullptr;
+
+poll_t poll;
+client_t client;
+
+ptr_t<char> name;
+
+void create_services(avahi::Client *c);
+
+void entry_group_callback(avahi::EntryGroup *g, avahi::EntryGroupState state, void *) {
+  group = g;
+
+  switch(state) {
+  case avahi::ENTRY_GROUP_ESTABLISHED:
+    BOOST_LOG(info) << "Avahi service " << name.get() << " successfully established.";
+    break;
+  case avahi::ENTRY_GROUP_COLLISION:
+    name.reset(avahi::alternative_service_name(name.get()));
+
+    BOOST_LOG(info) << "Avahi service name collision, renaming service to " << name.get();
+
+    create_services(avahi::entry_group_get_client(g));
+    break;
+  case avahi::ENTRY_GROUP_FAILURE:
+    BOOST_LOG(error) << "Avahi entry group failure: " << avahi::strerror(avahi::client_errno(avahi::entry_group_get_client(g)));
+    avahi::simple_poll_quit(poll.get());
+    break;
+  case avahi::ENTRY_GROUP_UNCOMMITED:
+  case avahi::ENTRY_GROUP_REGISTERING:;
+  }
+}
+
+void create_services(avahi::Client *c) {
+  int ret;
+
+  auto fg = util::fail_guard([]() {
+    avahi::simple_poll_quit(poll.get());
+  });
+
+  if(!group) {
+    if(!(group = avahi::entry_group_new(c, entry_group_callback, nullptr))) {
+      BOOST_LOG(error) << "avahi::entry_group_new() failed: "sv << avahi::strerror(avahi::client_errno(c));
+      return;
+    }
+  }
+
+  if(avahi::entry_group_is_empty(group)) {
+    BOOST_LOG(info) << "Adding avahi service "sv << name.get();
+
+    ret = avahi::entry_group_add_service(
+      group,
+      avahi::IF_UNSPEC, avahi::PROTO_UNSPEC,
+      avahi::PublishFlags(0),
+      name.get(),
+      SERVICE_TYPE,
+      nullptr, nullptr,
+      map_port(nvhttp::PORT_HTTP),
+      nullptr);
+
+    if(ret < 0) {
+      if(ret == avahi::ERR_COLLISION) {
+        // A service name collision with a local service happened. Let's pick a new name
+        name.reset(avahi::alternative_service_name(name.get()));
+        BOOST_LOG(info) << "Service name collision, renaming service to "sv << name.get();
+
+        avahi::entry_group_reset(group);
+
+        create_services(c);
+
+        fg.disable();
+        return;
+      }
+
+      BOOST_LOG(error) << "Failed to add "sv << SERVICE_TYPE << " service: "sv << avahi::strerror(ret);
+      return;
+    }
+
+    ret = avahi::entry_group_commit(group);
+    if(ret < 0) {
+      BOOST_LOG(error) << "Failed to commit entry group: "sv << avahi::strerror(ret);
+      return;
+    }
+  }
+
+  fg.disable();
+}
+
+void client_callback(avahi::Client *c, avahi::ClientState state, void *) {
+  switch(state) {
+  case avahi::CLIENT_S_RUNNING:
+    create_services(c);
+    break;
+  case avahi::CLIENT_FAILURE:
+    BOOST_LOG(error) << "Client failure: "sv << avahi::strerror(avahi::client_errno(c));
+    avahi::simple_poll_quit(poll.get());
+    break;
+  case avahi::CLIENT_S_COLLISION:
+  case avahi::CLIENT_S_REGISTERING:
+    if(group)
+      avahi::entry_group_reset(group);
+    break;
+  case avahi::CLIENT_CONNECTING:;
+  }
+}
+
+class deinit_t : public ::platf::deinit_t {
+public:
+  std::thread poll_thread;
+
+  deinit_t(std::thread poll_thread) : poll_thread { std::move(poll_thread) } {}
+
+  ~deinit_t() override {
+    if(avahi::simple_poll_quit && poll) {
+      avahi::simple_poll_quit(poll.get());
+    }
+
+    if(poll_thread.joinable()) {
+      poll_thread.join();
+    }
+  }
+};
+
+[[nodiscard]] std::unique_ptr<::platf::deinit_t> start() {
+  if(avahi::init_client()) {
+    return nullptr;
+  }
+
+  int avhi_error;
+
+  poll.reset(avahi::simple_poll_new());
+  if(!poll) {
+    BOOST_LOG(error) << "Failed to create simple poll object."sv;
+    return nullptr;
+  }
+
+  name.reset(avahi::strdup(SERVICE_NAME));
+
+  client.reset(
+    avahi::client_new(avahi::simple_poll_get(poll.get()), avahi::ClientFlags(0), client_callback, nullptr, &avhi_error));
+
+  if(!client) {
+    BOOST_LOG(error) << "Failed to create client: "sv << avahi::strerror(avhi_error);
+    return nullptr;
+  }
+
+  return std::make_unique<deinit_t>(std::thread { avahi::simple_poll_loop, poll.get() });
+}
+} // namespace platf::publish

src/platform/linux/vaapi.cpp

@@ -0,0 +1,651 @@
+#include <sstream>
+#include <string>
+
+#include <fcntl.h>
+
+extern "C" {
+#include <libavcodec/avcodec.h>
+}
+
+#include "graphics.h"
+#include "misc.h"
+#include "src/config.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+using namespace std::literals;
+
+extern "C" struct AVBufferRef;
+
+namespace va {
+constexpr auto SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2 = 0x40000000;
+constexpr auto EXPORT_SURFACE_WRITE_ONLY           = 0x0002;
+constexpr auto EXPORT_SURFACE_COMPOSED_LAYERS      = 0x0008;
+
+using VADisplay   = void *;
+using VAStatus    = int;
+using VAGenericID = unsigned int;
+using VASurfaceID = VAGenericID;
+
+struct DRMPRIMESurfaceDescriptor {
+  // VA Pixel format fourcc of the whole surface (VA_FOURCC_*).
+  uint32_t fourcc;
+
+  uint32_t width;
+  uint32_t height;
+
+  // Number of distinct DRM objects making up the surface.
+  uint32_t num_objects;
+
+  struct {
+    // DRM PRIME file descriptor for this object.
+    // Needs to be closed manually
+    int fd;
+
+    /*
+     *  Total size of this object (may include regions which are
+     *  not part of the surface).
+     */
+    uint32_t size;
+    // Format modifier applied to this object, not sure what that means
+    uint64_t drm_format_modifier;
+  } objects[4];
+
+  // Number of layers making up the surface.
+  uint32_t num_layers;
+  struct {
+    // DRM format fourcc of this layer (DRM_FOURCC_*).
+    uint32_t drm_format;
+
+    // Number of planes in this layer.
+    uint32_t num_planes;
+
+    // references objects --> DRMPRIMESurfaceDescriptor.objects[object_index[0]]
+    uint32_t object_index[4];
+
+    // Offset within the object of each plane.
+    uint32_t offset[4];
+
+    // Pitch of each plane.
+    uint32_t pitch[4];
+  } layers[4];
+};
+
+/** Currently defined profiles */
+enum class profile_e {
+  // Profile ID used for video processing.
+  ProfileNone             = -1,
+  MPEG2Simple             = 0,
+  MPEG2Main               = 1,
+  MPEG4Simple             = 2,
+  MPEG4AdvancedSimple     = 3,
+  MPEG4Main               = 4,
+  H264Baseline            = 5,
+  H264Main                = 6,
+  H264High                = 7,
+  VC1Simple               = 8,
+  VC1Main                 = 9,
+  VC1Advanced             = 10,
+  H263Baseline            = 11,
+  JPEGBaseline            = 12,
+  H264ConstrainedBaseline = 13,
+  VP8Version0_3           = 14,
+  H264MultiviewHigh       = 15,
+  H264StereoHigh          = 16,
+  HEVCMain                = 17,
+  HEVCMain10              = 18,
+  VP9Profile0             = 19,
+  VP9Profile1             = 20,
+  VP9Profile2             = 21,
+  VP9Profile3             = 22,
+  HEVCMain12              = 23,
+  HEVCMain422_10          = 24,
+  HEVCMain422_12          = 25,
+  HEVCMain444             = 26,
+  HEVCMain444_10          = 27,
+  HEVCMain444_12          = 28,
+  HEVCSccMain             = 29,
+  HEVCSccMain10           = 30,
+  HEVCSccMain444          = 31,
+  AV1Profile0             = 32,
+  AV1Profile1             = 33,
+  HEVCSccMain444_10       = 34,
+
+  // Profile ID used for protected video playback.
+  Protected = 35
+};
+
+enum class entry_e {
+  VLD        = 1,
+  IZZ        = 2,
+  IDCT       = 3,
+  MoComp     = 4,
+  Deblocking = 5,
+  EncSlice   = 6, /* slice level encode */
+  EncPicture = 7, /* pictuer encode, JPEG, etc */
+  /*
+     * For an implementation that supports a low power/high performance variant
+     * for slice level encode, it can choose to expose the
+     * VAEntrypointEncSliceLP entrypoint. Certain encoding tools may not be
+     * available with this entrypoint (e.g. interlace, MBAFF) and the
+     * application can query the encoding configuration attributes to find
+     * out more details if this entrypoint is supported.
+     */
+  EncSliceLP = 8,
+  VideoProc  = 10, /**< Video pre/post-processing. */
+  /**
+     * \brief FEI
+     *
+     * The purpose of FEI (Flexible Encoding Infrastructure) is to allow applications to
+     * have more controls and trade off quality for speed with their own IPs.
+     * The application can optionally provide input to ENC for extra encode control
+     * and get the output from ENC. Application can chose to modify the ENC
+     * output/PAK input during encoding, but the performance impact is significant.
+     *
+     * On top of the existing buffers for normal encode, there will be
+     * one extra input buffer (VAEncMiscParameterFEIFrameControl) and
+     * three extra output buffers (VAEncFEIMVBufferType, VAEncFEIMBModeBufferType
+     * and VAEncFEIDistortionBufferType) for FEI entry function.
+     * If separate PAK is set, two extra input buffers
+     * (VAEncFEIMVBufferType, VAEncFEIMBModeBufferType) are needed for PAK input.
+     **/
+  FEI = 11,
+  /**
+     * \brief Stats
+     *
+     * A pre-processing function for getting some statistics and motion vectors is added,
+     * and some extra controls for Encode pipeline are provided. The application can
+     * optionally call the statistics function to get motion vectors and statistics like
+     * variances, distortions before calling Encode function via this entry point.
+     *
+     * Checking whether Statistics is supported can be performed with vaQueryConfigEntrypoints().
+     * If Statistics entry point is supported, then the list of returned entry-points will
+     * include #Stats. Supported pixel format, maximum resolution and statistics
+     * specific attributes can be obtained via normal attribute query. One input buffer
+     * (VAStatsStatisticsParameterBufferType) and one or two output buffers
+     * (VAStatsStatisticsBufferType, VAStatsStatisticsBottomFieldBufferType (for interlace only)
+     * and VAStatsMVBufferType) are needed for this entry point.
+     **/
+  Stats = 12,
+  /**
+     * \brief ProtectedTEEComm
+     *
+     * A function for communicating with TEE (Trusted Execution Environment).
+     **/
+  ProtectedTEEComm = 13,
+  /**
+     * \brief ProtectedContent
+     *
+     * A function for protected content to decrypt encrypted content.
+     **/
+  ProtectedContent = 14,
+};
+
+
+typedef VAStatus (*queryConfigEntrypoints_fn)(VADisplay dpy, profile_e profile, entry_e *entrypoint_list, int *num_entrypoints);
+typedef int (*maxNumEntrypoints_fn)(VADisplay dpy);
+typedef VADisplay (*getDisplayDRM_fn)(int fd);
+typedef VAStatus (*terminate_fn)(VADisplay dpy);
+typedef VAStatus (*initialize_fn)(VADisplay dpy, int *major_version, int *minor_version);
+typedef const char *(*errorStr_fn)(VAStatus error_status);
+typedef void (*VAMessageCallback)(void *user_context, const char *message);
+typedef VAMessageCallback (*setErrorCallback_fn)(VADisplay dpy, VAMessageCallback callback, void *user_context);
+typedef VAMessageCallback (*setInfoCallback_fn)(VADisplay dpy, VAMessageCallback callback, void *user_context);
+typedef const char *(*queryVendorString_fn)(VADisplay dpy);
+typedef VAStatus (*exportSurfaceHandle_fn)(
+  VADisplay dpy, VASurfaceID surface_id,
+  uint32_t mem_type, uint32_t flags,
+  void *descriptor);
+
+static maxNumEntrypoints_fn maxNumEntrypoints;
+static queryConfigEntrypoints_fn queryConfigEntrypoints;
+static getDisplayDRM_fn getDisplayDRM;
+static terminate_fn terminate;
+static initialize_fn initialize;
+static errorStr_fn errorStr;
+static setErrorCallback_fn setErrorCallback;
+static setInfoCallback_fn setInfoCallback;
+static queryVendorString_fn queryVendorString;
+static exportSurfaceHandle_fn exportSurfaceHandle;
+
+using display_t = util::dyn_safe_ptr_v2<void, VAStatus, &terminate>;
+
+int init_main_va() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libva.so.2", "libva.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&maxNumEntrypoints, "vaMaxNumEntrypoints" },
+    { (dyn::apiproc *)&queryConfigEntrypoints, "vaQueryConfigEntrypoints" },
+    { (dyn::apiproc *)&terminate, "vaTerminate" },
+    { (dyn::apiproc *)&initialize, "vaInitialize" },
+    { (dyn::apiproc *)&errorStr, "vaErrorStr" },
+    { (dyn::apiproc *)&setErrorCallback, "vaSetErrorCallback" },
+    { (dyn::apiproc *)&setInfoCallback, "vaSetInfoCallback" },
+    { (dyn::apiproc *)&queryVendorString, "vaQueryVendorString" },
+    { (dyn::apiproc *)&exportSurfaceHandle, "vaExportSurfaceHandle" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+int init() {
+  if(init_main_va()) {
+    return -1;
+  }
+
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libva-drm.so.2", "libva-drm.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&getDisplayDRM, "vaGetDisplayDRM" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+int vaapi_make_hwdevice_ctx(platf::hwdevice_t *base, AVBufferRef **hw_device_buf);
+
+class va_t : public platf::hwdevice_t {
+public:
+  int init(int in_width, int in_height, file_t &&render_device) {
+    file = std::move(render_device);
+
+    if(!va::initialize || !gbm::create_device) {
+      if(!va::initialize) BOOST_LOG(warning) << "libva not initialized"sv;
+      if(!gbm::create_device) BOOST_LOG(warning) << "libgbm not initialized"sv;
+      return -1;
+    }
+
+    this->data = (void *)vaapi_make_hwdevice_ctx;
+
+    gbm.reset(gbm::create_device(file.el));
+    if(!gbm) {
+      char string[1024];
+      BOOST_LOG(error) << "Couldn't create GBM device: ["sv << strerror_r(errno, string, sizeof(string)) << ']';
+      return -1;
+    }
+
+    display = egl::make_display(gbm.get());
+    if(!display) {
+      return -1;
+    }
+
+    auto ctx_opt = egl::make_ctx(display.get());
+    if(!ctx_opt) {
+      return -1;
+    }
+
+    ctx = std::move(*ctx_opt);
+
+    width  = in_width;
+    height = in_height;
+
+    return 0;
+  }
+
+  int set_frame(AVFrame *frame) override {
+    this->hwframe.reset(frame);
+    this->frame = frame;
+
+    if(av_hwframe_get_buffer(frame->hw_frames_ctx, frame, 0)) {
+      BOOST_LOG(error) << "Couldn't get hwframe for VAAPI"sv;
+
+      return -1;
+    }
+
+    va::DRMPRIMESurfaceDescriptor prime;
+    va::VASurfaceID surface = (std::uintptr_t)frame->data[3];
+
+    auto status = va::exportSurfaceHandle(
+      this->va_display,
+      surface,
+      va::SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2,
+      va::EXPORT_SURFACE_WRITE_ONLY | va::EXPORT_SURFACE_COMPOSED_LAYERS,
+      &prime);
+    if(status) {
+
+      BOOST_LOG(error) << "Couldn't export va surface handle: ["sv << (int)surface << "]: "sv << va::errorStr(status);
+
+      return -1;
+    }
+
+    // Keep track of file descriptors
+    std::array<file_t, egl::nv12_img_t::num_fds> fds;
+    for(int x = 0; x < prime.num_objects; ++x) {
+      fds[x] = prime.objects[x].fd;
+    }
+
+    auto nv12_opt = egl::import_target(
+      display.get(),
+      std::move(fds),
+      { (int)prime.width,
+        (int)prime.height,
+        { prime.objects[prime.layers[0].object_index[0]].fd, -1, -1, -1 },
+        0,
+        0,
+        { prime.layers[0].pitch[0] },
+        { prime.layers[0].offset[0] } },
+      { (int)prime.width / 2,
+        (int)prime.height / 2,
+        { prime.objects[prime.layers[0].object_index[1]].fd, -1, -1, -1 },
+        0,
+        0,
+        { prime.layers[0].pitch[1] },
+        { prime.layers[0].offset[1] } });
+
+    if(!nv12_opt) {
+      return -1;
+    }
+
+    auto sws_opt = egl::sws_t::make(width, height, frame->width, frame->height);
+    if(!sws_opt) {
+      return -1;
+    }
+
+    this->sws  = std::move(*sws_opt);
+    this->nv12 = std::move(*nv12_opt);
+
+    return 0;
+  }
+
+  void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) override {
+    sws.set_colorspace(colorspace, color_range);
+  }
+
+  va::display_t::pointer va_display;
+  file_t file;
+
+  frame_t hwframe;
+
+  gbm::gbm_t gbm;
+  egl::display_t display;
+  egl::ctx_t ctx;
+
+  egl::sws_t sws;
+  egl::nv12_t nv12;
+
+  int width, height;
+};
+
+class va_ram_t : public va_t {
+public:
+  int convert(platf::img_t &img) override {
+    sws.load_ram(img);
+
+    sws.convert(nv12->buf);
+    return 0;
+  }
+};
+
+class va_vram_t : public va_t {
+public:
+  int convert(platf::img_t &img) override {
+    auto &descriptor = (egl::img_descriptor_t &)img;
+
+    if(descriptor.sequence > sequence) {
+      sequence = descriptor.sequence;
+
+      rgb = egl::rgb_t {};
+
+      auto rgb_opt = egl::import_source(display.get(), descriptor.sd);
+
+      if(!rgb_opt) {
+        return -1;
+      }
+
+      rgb = std::move(*rgb_opt);
+    }
+
+    sws.load_vram(descriptor, offset_x, offset_y, rgb->tex[0]);
+
+    sws.convert(nv12->buf);
+    return 0;
+  }
+
+  int init(int in_width, int in_height, file_t &&render_device, int offset_x, int offset_y) {
+    if(va_t::init(in_width, in_height, std::move(render_device))) {
+      return -1;
+    }
+
+    sequence = 0;
+
+    this->offset_x = offset_x;
+    this->offset_y = offset_y;
+
+    return 0;
+  }
+
+  std::uint64_t sequence;
+  egl::rgb_t rgb;
+
+  int offset_x, offset_y;
+};
+
+/**
+ * This is a private structure of FFmpeg, I need this to manually create
+ * a VAAPI hardware context
+ * 
+ * xdisplay will not be used internally by FFmpeg
+ */
+typedef struct VAAPIDevicePriv {
+  union {
+    void *xdisplay;
+    int fd;
+  } drm;
+  int drm_fd;
+} VAAPIDevicePriv;
+
+/**
+ * VAAPI connection details.
+ *
+ * Allocated as AVHWDeviceContext.hwctx
+ */
+typedef struct AVVAAPIDeviceContext {
+  /**
+   * The VADisplay handle, to be filled by the user.
+   */
+  va::VADisplay display;
+  /**
+   * Driver quirks to apply - this is filled by av_hwdevice_ctx_init(),
+   * with reference to a table of known drivers, unless the
+   * AV_VAAPI_DRIVER_QUIRK_USER_SET bit is already present.  The user
+   * may need to refer to this field when performing any later
+   * operations using VAAPI with the same VADisplay.
+   */
+  unsigned int driver_quirks;
+} AVVAAPIDeviceContext;
+
+static void __log(void *level, const char *msg) {
+  BOOST_LOG(*(boost::log::sources::severity_logger<int> *)level) << msg;
+}
+
+int vaapi_make_hwdevice_ctx(platf::hwdevice_t *base, AVBufferRef **hw_device_buf) {
+  if(!va::initialize) {
+    BOOST_LOG(warning) << "libva not loaded"sv;
+    return -1;
+  }
+
+  if(!va::getDisplayDRM) {
+    BOOST_LOG(warning) << "libva-drm not loaded"sv;
+    return -1;
+  }
+
+  auto va = (va::va_t *)base;
+  auto fd = dup(va->file.el);
+
+  auto *priv   = (VAAPIDevicePriv *)av_mallocz(sizeof(VAAPIDevicePriv));
+  priv->drm_fd = fd;
+  priv->drm.fd = fd;
+
+  auto fg = util::fail_guard([fd, priv]() {
+    close(fd);
+    av_free(priv);
+  });
+
+  va::display_t display { va::getDisplayDRM(fd) };
+  if(!display) {
+    auto render_device = config::video.adapter_name.empty() ? "/dev/dri/renderD128" : config::video.adapter_name.c_str();
+
+    BOOST_LOG(error) << "Couldn't open a va display from DRM with device: "sv << render_device;
+    return -1;
+  }
+
+  va->va_display = display.get();
+
+  va::setErrorCallback(display.get(), __log, &error);
+  va::setErrorCallback(display.get(), __log, &info);
+
+  int major, minor;
+  auto status = va::initialize(display.get(), &major, &minor);
+  if(status) {
+    BOOST_LOG(error) << "Couldn't initialize va display: "sv << va::errorStr(status);
+    return -1;
+  }
+
+  BOOST_LOG(debug) << "vaapi vendor: "sv << va::queryVendorString(display.get());
+
+  *hw_device_buf = av_hwdevice_ctx_alloc(AV_HWDEVICE_TYPE_VAAPI);
+  auto ctx       = (AVVAAPIDeviceContext *)((AVHWDeviceContext *)(*hw_device_buf)->data)->hwctx;
+  ctx->display   = display.release();
+
+  fg.disable();
+
+  auto err = av_hwdevice_ctx_init(*hw_device_buf);
+  if(err) {
+    char err_str[AV_ERROR_MAX_STRING_SIZE] { 0 };
+    BOOST_LOG(error) << "Failed to create FFMpeg hardware device context: "sv << av_make_error_string(err_str, AV_ERROR_MAX_STRING_SIZE, err);
+
+    return err;
+  }
+
+  return 0;
+}
+
+static bool query(display_t::pointer display, profile_e profile) {
+  std::vector<entry_e> entrypoints;
+  entrypoints.resize(maxNumEntrypoints(display));
+
+  int count;
+  auto status = queryConfigEntrypoints(display, profile, entrypoints.data(), &count);
+  if(status) {
+    BOOST_LOG(error) << "Couldn't query entrypoints: "sv << va::errorStr(status);
+    return false;
+  }
+  entrypoints.resize(count);
+
+  for(auto entrypoint : entrypoints) {
+    if(entrypoint == entry_e::EncSlice || entrypoint == entry_e::EncSliceLP) {
+      return true;
+    }
+  }
+
+  return false;
+}
+
+bool validate(int fd) {
+  if(init()) {
+    return false;
+  }
+
+  va::display_t display { va::getDisplayDRM(fd) };
+  if(!display) {
+    char string[1024];
+
+    auto bytes = readlink(("/proc/self/fd/" + std::to_string(fd)).c_str(), string, sizeof(string));
+
+    std::string_view render_device { string, (std::size_t)bytes };
+
+    BOOST_LOG(error) << "Couldn't open a va display from DRM with device: "sv << render_device;
+    return false;
+  }
+
+  int major, minor;
+  auto status = initialize(display.get(), &major, &minor);
+  if(status) {
+    BOOST_LOG(error) << "Couldn't initialize va display: "sv << va::errorStr(status);
+    return false;
+  }
+
+  if(!query(display.get(), profile_e::H264Main)) {
+    return false;
+  }
+
+  if(config::video.hevc_mode > 1 && !query(display.get(), profile_e::HEVCMain)) {
+    return false;
+  }
+
+  if(config::video.hevc_mode > 2 && !query(display.get(), profile_e::HEVCMain10)) {
+    return false;
+  }
+
+  return true;
+}
+
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, file_t &&card, int offset_x, int offset_y, bool vram) {
+  if(vram) {
+    auto egl = std::make_shared<va::va_vram_t>();
+    if(egl->init(width, height, std::move(card), offset_x, offset_y)) {
+      return nullptr;
+    }
+
+    return egl;
+  }
+
+  else {
+    auto egl = std::make_shared<va::va_ram_t>();
+    if(egl->init(width, height, std::move(card))) {
+      return nullptr;
+    }
+
+    return egl;
+  }
+}
+
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, int offset_x, int offset_y, bool vram) {
+  auto render_device = config::video.adapter_name.empty() ? "/dev/dri/renderD128" : config::video.adapter_name.c_str();
+
+  file_t file = open(render_device, O_RDWR);
+  if(file.el < 0) {
+    char string[1024];
+    BOOST_LOG(error) << "Couldn't open "sv << render_device << ": " << strerror_r(errno, string, sizeof(string));
+
+    return nullptr;
+  }
+
+  return make_hwdevice(width, height, std::move(file), offset_x, offset_y, vram);
+}
+
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, bool vram) {
+  return make_hwdevice(width, height, 0, 0, vram);
+}
+} // namespace va

src/platform/linux/vaapi.h

@@ -0,0 +1,27 @@
+#ifndef SUNSHINE_VAAPI_H
+#define SUNSHINE_VAAPI_H
+
+#include "misc.h"
+#include "src/platform/common.h"
+
+namespace egl {
+struct surface_descriptor_t;
+}
+namespace va {
+/**
+ * Width --> Width of the image
+ * Height --> Height of the image
+ * offset_x --> Horizontal offset of the image in the texture
+ * offset_y --> Vertical offset of the image in the texture
+ * file_t card --> The file descriptor of the render device used for encoding
+ */
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, bool vram);
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, int offset_x, int offset_y, bool vram);
+std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, file_t &&card, int offset_x, int offset_y, bool vram);
+
+// Ensure the render device pointed to by fd is capable of encoding h264 with the hevc_mode configured
+bool validate(int fd);
+
+int init();
+} // namespace va
+#endif

src/platform/linux/wayland.cpp

@@ -0,0 +1,268 @@
+#include <wayland-client.h>
+#include <wayland-util.h>
+
+#include <cstdlib>
+
+#include "graphics.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+#include "src/round_robin.h"
+#include "src/utility.h"
+#include "wayland.h"
+
+extern const wl_interface wl_output_interface;
+
+using namespace std::literals;
+
+// Disable warning for converting incompatible functions
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wpedantic"
+#pragma GCC diagnostic ignored "-Wpmf-conversions"
+
+namespace wl {
+int display_t::init(const char *display_name) {
+  if(!display_name) {
+    display_name = std::getenv("WAYLAND_DISPLAY");
+  }
+
+  if(!display_name) {
+    BOOST_LOG(error) << "Environment variable WAYLAND_DISPLAY has not been defined"sv;
+    return -1;
+  }
+
+  display_internal.reset(wl_display_connect(display_name));
+  if(!display_internal) {
+    BOOST_LOG(error) << "Couldn't connect to Wayland display: "sv << display_name;
+    return -1;
+  }
+
+  BOOST_LOG(info) << "Found display ["sv << display_name << ']';
+
+  return 0;
+}
+
+void display_t::roundtrip() {
+  wl_display_roundtrip(display_internal.get());
+}
+
+wl_registry *display_t::registry() {
+  return wl_display_get_registry(display_internal.get());
+}
+
+inline monitor_t::monitor_t(wl_output *output) : output { output } {}
+
+inline void monitor_t::xdg_name(zxdg_output_v1 *, const char *name) {
+  this->name = name;
+
+  BOOST_LOG(info) << "Name: "sv << this->name;
+}
+
+void monitor_t::xdg_description(zxdg_output_v1 *, const char *description) {
+  this->description = description;
+
+  BOOST_LOG(info) << "Found monitor: "sv << this->description;
+}
+
+void monitor_t::xdg_position(zxdg_output_v1 *, std::int32_t x, std::int32_t y) {
+  viewport.offset_x = x;
+  viewport.offset_y = y;
+
+  BOOST_LOG(info) << "Offset: "sv << x << 'x' << y;
+}
+
+void monitor_t::xdg_size(zxdg_output_v1 *, std::int32_t width, std::int32_t height) {
+  viewport.width  = width;
+  viewport.height = height;
+
+  BOOST_LOG(info) << "Resolution: "sv << width << 'x' << height;
+}
+
+void monitor_t::xdg_done(zxdg_output_v1 *) {
+  BOOST_LOG(info) << "All info about monitor ["sv << name << "] has been send"sv;
+}
+
+void monitor_t::listen(zxdg_output_manager_v1 *output_manager) {
+  auto xdg_output = zxdg_output_manager_v1_get_xdg_output(output_manager, output);
+
+#define CLASS_CALL(x, y) x = (decltype(x))&y
+
+  CLASS_CALL(listener.name, monitor_t::xdg_name);
+  CLASS_CALL(listener.logical_size, monitor_t::xdg_size);
+  CLASS_CALL(listener.logical_position, monitor_t::xdg_position);
+  CLASS_CALL(listener.done, monitor_t::xdg_done);
+  CLASS_CALL(listener.description, monitor_t::xdg_description);
+
+#undef CLASS_CALL
+  zxdg_output_v1_add_listener(xdg_output, &listener, this);
+}
+
+interface_t::interface_t() noexcept
+    : output_manager { nullptr }, listener {
+        (decltype(wl_registry_listener::global))&interface_t::add_interface,
+        (decltype(wl_registry_listener::global_remove))&interface_t::del_interface,
+      } {}
+
+void interface_t::listen(wl_registry *registry) {
+  wl_registry_add_listener(registry, &listener, this);
+}
+
+void interface_t::add_interface(wl_registry *registry, std::uint32_t id, const char *interface, std::uint32_t version) {
+  BOOST_LOG(debug) << "Available interface: "sv << interface << '(' << id << ") version "sv << version;
+
+  if(!std::strcmp(interface, wl_output_interface.name)) {
+    BOOST_LOG(info) << "Found interface: "sv << interface << '(' << id << ") version "sv << version;
+    monitors.emplace_back(
+      std::make_unique<monitor_t>(
+        (wl_output *)wl_registry_bind(registry, id, &wl_output_interface, version)));
+  }
+  else if(!std::strcmp(interface, zxdg_output_manager_v1_interface.name)) {
+    BOOST_LOG(info) << "Found interface: "sv << interface << '(' << id << ") version "sv << version;
+    output_manager = (zxdg_output_manager_v1 *)wl_registry_bind(registry, id, &zxdg_output_manager_v1_interface, version);
+
+    this->interface[XDG_OUTPUT] = true;
+  }
+  else if(!std::strcmp(interface, zwlr_export_dmabuf_manager_v1_interface.name)) {
+    BOOST_LOG(info) << "Found interface: "sv << interface << '(' << id << ") version "sv << version;
+    dmabuf_manager = (zwlr_export_dmabuf_manager_v1 *)wl_registry_bind(registry, id, &zwlr_export_dmabuf_manager_v1_interface, version);
+
+    this->interface[WLR_EXPORT_DMABUF] = true;
+  }
+}
+
+void interface_t::del_interface(wl_registry *registry, uint32_t id) {
+  BOOST_LOG(info) << "Delete: "sv << id;
+}
+
+dmabuf_t::dmabuf_t()
+    : status { READY }, frames {}, current_frame { &frames[0] }, listener {
+        (decltype(zwlr_export_dmabuf_frame_v1_listener::frame))&dmabuf_t::frame,
+        (decltype(zwlr_export_dmabuf_frame_v1_listener::object))&dmabuf_t::object,
+        (decltype(zwlr_export_dmabuf_frame_v1_listener::ready))&dmabuf_t::ready,
+        (decltype(zwlr_export_dmabuf_frame_v1_listener::cancel))&dmabuf_t::cancel,
+      } {
+}
+
+void dmabuf_t::listen(zwlr_export_dmabuf_manager_v1 *dmabuf_manager, wl_output *output, bool blend_cursor) {
+  auto frame = zwlr_export_dmabuf_manager_v1_capture_output(dmabuf_manager, blend_cursor, output);
+  zwlr_export_dmabuf_frame_v1_add_listener(frame, &listener, this);
+
+  status = WAITING;
+}
+
+dmabuf_t::~dmabuf_t() {
+  for(auto &frame : frames) {
+    frame.destroy();
+  }
+}
+
+void dmabuf_t::frame(
+  zwlr_export_dmabuf_frame_v1 *frame,
+  std::uint32_t width, std::uint32_t height,
+  std::uint32_t x, std::uint32_t y,
+  std::uint32_t buffer_flags, std::uint32_t flags,
+  std::uint32_t format,
+  std::uint32_t high, std::uint32_t low,
+  std::uint32_t obj_count) {
+  auto next_frame = get_next_frame();
+
+  next_frame->sd.fourcc   = format;
+  next_frame->sd.width    = width;
+  next_frame->sd.height   = height;
+  next_frame->sd.modifier = (((std::uint64_t)high) << 32) | low;
+}
+
+void dmabuf_t::object(
+  zwlr_export_dmabuf_frame_v1 *frame,
+  std::uint32_t index,
+  std::int32_t fd,
+  std::uint32_t size,
+  std::uint32_t offset,
+  std::uint32_t stride,
+  std::uint32_t plane_index) {
+  auto next_frame = get_next_frame();
+
+  next_frame->sd.fds[plane_index]     = fd;
+  next_frame->sd.pitches[plane_index] = stride;
+  next_frame->sd.offsets[plane_index] = offset;
+}
+
+void dmabuf_t::ready(
+  zwlr_export_dmabuf_frame_v1 *frame,
+  std::uint32_t tv_sec_hi, std::uint32_t tv_sec_lo, std::uint32_t tv_nsec) {
+
+  zwlr_export_dmabuf_frame_v1_destroy(frame);
+
+  current_frame->destroy();
+  current_frame = get_next_frame();
+
+  status = READY;
+}
+
+void dmabuf_t::cancel(
+  zwlr_export_dmabuf_frame_v1 *frame,
+  zwlr_export_dmabuf_frame_v1_cancel_reason reason) {
+
+  zwlr_export_dmabuf_frame_v1_destroy(frame);
+
+  auto next_frame = get_next_frame();
+  next_frame->destroy();
+
+  status = REINIT;
+}
+
+void frame_t::destroy() {
+  for(auto x = 0; x < 4; ++x) {
+    if(sd.fds[x] >= 0) {
+      close(sd.fds[x]);
+
+      sd.fds[x] = -1;
+    }
+  }
+}
+
+frame_t::frame_t() {
+  // File descriptors aren't open
+  std::fill_n(sd.fds, 4, -1);
+};
+
+std::vector<std::unique_ptr<monitor_t>> monitors(const char *display_name) {
+  display_t display;
+
+  if(display.init(display_name)) {
+    return {};
+  }
+
+  interface_t interface;
+  interface.listen(display.registry());
+
+  display.roundtrip();
+
+  if(!interface[interface_t::XDG_OUTPUT]) {
+    BOOST_LOG(error) << "Missing Wayland wire XDG_OUTPUT"sv;
+    return {};
+  }
+
+  for(auto &monitor : interface.monitors) {
+    monitor->listen(interface.output_manager);
+  }
+
+  display.roundtrip();
+
+  return std::move(interface.monitors);
+}
+
+static bool validate() {
+  display_t display;
+
+  return display.init() == 0;
+}
+
+int init() {
+  static bool validated = validate();
+
+  return !validated;
+}
+
+} // namespace wl
+
+#pragma GCC diagnostic pop

src/platform/linux/wayland.h

@@ -0,0 +1,216 @@
+#ifndef SUNSHINE_WAYLAND_H
+#define SUNSHINE_WAYLAND_H
+
+#include <bitset>
+
+#ifdef SUNSHINE_BUILD_WAYLAND
+#include <wlr-export-dmabuf-unstable-v1.h>
+#include <xdg-output-unstable-v1.h>
+#endif
+
+#include "graphics.h"
+
+/**
+ * The classes defined in this macro block should only be used by
+ * cpp files whose compilation depends on SUNSHINE_BUILD_WAYLAND
+ */
+#ifdef SUNSHINE_BUILD_WAYLAND
+
+namespace wl {
+using display_internal_t = util::safe_ptr<wl_display, wl_display_disconnect>;
+
+class frame_t {
+public:
+  frame_t();
+  egl::surface_descriptor_t sd;
+
+  void destroy();
+};
+
+class dmabuf_t {
+public:
+  enum status_e {
+    WAITING,
+    READY,
+    REINIT,
+  };
+
+  dmabuf_t(dmabuf_t &&)      = delete;
+  dmabuf_t(const dmabuf_t &) = delete;
+
+  dmabuf_t &operator=(const dmabuf_t &) = delete;
+  dmabuf_t &operator=(dmabuf_t &&) = delete;
+
+  dmabuf_t();
+
+  void listen(zwlr_export_dmabuf_manager_v1 *dmabuf_manager, wl_output *output, bool blend_cursor = false);
+
+  ~dmabuf_t();
+
+  void frame(
+    zwlr_export_dmabuf_frame_v1 *frame,
+    std::uint32_t width, std::uint32_t height,
+    std::uint32_t x, std::uint32_t y,
+    std::uint32_t buffer_flags, std::uint32_t flags,
+    std::uint32_t format,
+    std::uint32_t high, std::uint32_t low,
+    std::uint32_t obj_count);
+
+  void object(
+    zwlr_export_dmabuf_frame_v1 *frame,
+    std::uint32_t index,
+    std::int32_t fd,
+    std::uint32_t size,
+    std::uint32_t offset,
+    std::uint32_t stride,
+    std::uint32_t plane_index);
+
+  void ready(
+    zwlr_export_dmabuf_frame_v1 *frame,
+    std::uint32_t tv_sec_hi, std::uint32_t tv_sec_lo, std::uint32_t tv_nsec);
+
+  void cancel(
+    zwlr_export_dmabuf_frame_v1 *frame,
+    zwlr_export_dmabuf_frame_v1_cancel_reason reason);
+
+  inline frame_t *get_next_frame() {
+    return current_frame == &frames[0] ? &frames[1] : &frames[0];
+  }
+
+  status_e status;
+
+  std::array<frame_t, 2> frames;
+  frame_t *current_frame;
+
+  zwlr_export_dmabuf_frame_v1_listener listener;
+};
+
+class monitor_t {
+public:
+  monitor_t(monitor_t &&)      = delete;
+  monitor_t(const monitor_t &) = delete;
+
+  monitor_t &operator=(const monitor_t &) = delete;
+  monitor_t &operator=(monitor_t &&) = delete;
+
+  monitor_t(wl_output *output);
+
+  void xdg_name(zxdg_output_v1 *, const char *name);
+  void xdg_description(zxdg_output_v1 *, const char *description);
+  void xdg_position(zxdg_output_v1 *, std::int32_t x, std::int32_t y);
+  void xdg_size(zxdg_output_v1 *, std::int32_t width, std::int32_t height);
+  void xdg_done(zxdg_output_v1 *);
+
+  void listen(zxdg_output_manager_v1 *output_manager);
+
+  wl_output *output;
+
+  std::string name;
+  std::string description;
+
+  platf::touch_port_t viewport;
+
+  zxdg_output_v1_listener listener;
+};
+
+class interface_t {
+  struct bind_t {
+    std::uint32_t id;
+    std::uint32_t version;
+  };
+
+public:
+  enum interface_e {
+    XDG_OUTPUT,
+    WLR_EXPORT_DMABUF,
+    MAX_INTERFACES,
+  };
+
+  interface_t(interface_t &&)      = delete;
+  interface_t(const interface_t &) = delete;
+
+  interface_t &operator=(const interface_t &) = delete;
+  interface_t &operator=(interface_t &&) = delete;
+
+  interface_t() noexcept;
+
+  void listen(wl_registry *registry);
+
+  std::vector<std::unique_ptr<monitor_t>> monitors;
+
+  zwlr_export_dmabuf_manager_v1 *dmabuf_manager;
+  zxdg_output_manager_v1 *output_manager;
+
+  bool operator[](interface_e bit) const {
+    return interface[bit];
+  }
+
+private:
+  void add_interface(wl_registry *registry, std::uint32_t id, const char *interface, std::uint32_t version);
+  void del_interface(wl_registry *registry, uint32_t id);
+
+  std::bitset<MAX_INTERFACES> interface;
+
+  wl_registry_listener listener;
+};
+
+class display_t {
+public:
+  /**
+   * Initialize display with display_name
+   * If display_name == nullptr -> display_name = std::getenv("WAYLAND_DISPLAY")
+   */
+  int init(const char *display_name = nullptr);
+
+  // Roundtrip with Wayland connection
+  void roundtrip();
+
+  // Get the registry associated with the display
+  // No need to manually free the registry
+  wl_registry *registry();
+
+  inline display_internal_t::pointer get() {
+    return display_internal.get();
+  }
+
+private:
+  display_internal_t display_internal;
+};
+
+std::vector<std::unique_ptr<monitor_t>> monitors(const char *display_name = nullptr);
+
+int init();
+} // namespace wl
+#else
+
+struct wl_output;
+struct zxdg_output_manager_v1;
+
+namespace wl {
+class monitor_t {
+public:
+  monitor_t(monitor_t &&)      = delete;
+  monitor_t(const monitor_t &) = delete;
+
+  monitor_t &operator=(const monitor_t &) = delete;
+  monitor_t &operator=(monitor_t &&) = delete;
+
+  monitor_t(wl_output *output);
+
+  void listen(zxdg_output_manager_v1 *output_manager);
+
+  wl_output *output;
+
+  std::string name;
+  std::string description;
+
+  platf::touch_port_t viewport;
+};
+
+inline std::vector<std::unique_ptr<monitor_t>> monitors(const char *display_name = nullptr) { return {}; }
+
+inline int init() { return -1; }
+} // namespace wl
+#endif
+
+#endif

src/platform/linux/wlgrab.cpp

@@ -0,0 +1,369 @@
+#include "src/platform/common.h"
+
+#include "src/main.h"
+#include "vaapi.h"
+#include "wayland.h"
+
+using namespace std::literals;
+namespace wl {
+static int env_width;
+static int env_height;
+
+struct img_t : public platf::img_t {
+  ~img_t() override {
+    delete[] data;
+    data = nullptr;
+  }
+};
+
+class wlr_t : public platf::display_t {
+public:
+  int init(platf::mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+    delay    = std::chrono::nanoseconds { 1s } / framerate;
+    mem_type = hwdevice_type;
+
+    if(display.init()) {
+      return -1;
+    }
+
+    interface.listen(display.registry());
+
+    display.roundtrip();
+
+    if(!interface[wl::interface_t::XDG_OUTPUT]) {
+      BOOST_LOG(error) << "Missing Wayland wire for xdg_output"sv;
+      return -1;
+    }
+
+    if(!interface[wl::interface_t::WLR_EXPORT_DMABUF]) {
+      BOOST_LOG(error) << "Missing Wayland wire for wlr-export-dmabuf"sv;
+      return -1;
+    }
+
+    auto monitor = interface.monitors[0].get();
+
+    if(!display_name.empty()) {
+      auto streamedMonitor = util::from_view(display_name);
+
+      if(streamedMonitor >= 0 && streamedMonitor < interface.monitors.size()) {
+        monitor = interface.monitors[streamedMonitor].get();
+      }
+    }
+
+    monitor->listen(interface.output_manager);
+
+    display.roundtrip();
+
+    output = monitor->output;
+
+    offset_x = monitor->viewport.offset_x;
+    offset_y = monitor->viewport.offset_y;
+    width    = monitor->viewport.width;
+    height   = monitor->viewport.height;
+
+    this->env_width  = ::wl::env_width;
+    this->env_height = ::wl::env_height;
+
+    BOOST_LOG(info) << "Selected monitor ["sv << monitor->description << "] for streaming"sv;
+    BOOST_LOG(debug) << "Offset: "sv << offset_x << 'x' << offset_y;
+    BOOST_LOG(debug) << "Resolution: "sv << width << 'x' << height;
+    BOOST_LOG(debug) << "Desktop Resolution: "sv << env_width << 'x' << env_height;
+
+    return 0;
+  }
+
+  int dummy_img(platf::img_t *img) override {
+    return 0;
+  }
+
+  inline platf::capture_e snapshot(platf::img_t *img_out_base, std::chrono::milliseconds timeout, bool cursor) {
+    auto to = std::chrono::steady_clock::now() + timeout;
+
+    dmabuf.listen(interface.dmabuf_manager, output, cursor);
+    do {
+      display.roundtrip();
+
+      if(to < std::chrono::steady_clock::now()) {
+        return platf::capture_e::timeout;
+      }
+    } while(dmabuf.status == dmabuf_t::WAITING);
+
+    auto current_frame = dmabuf.current_frame;
+
+    if(
+      dmabuf.status == dmabuf_t::REINIT ||
+      current_frame->sd.width != width ||
+      current_frame->sd.height != height) {
+
+      return platf::capture_e::reinit;
+    }
+
+    return platf::capture_e::ok;
+  }
+
+  platf::mem_type_e mem_type;
+
+  std::chrono::nanoseconds delay;
+
+  wl::display_t display;
+  interface_t interface;
+  dmabuf_t dmabuf;
+
+  wl_output *output;
+};
+
+class wlr_ram_t : public wlr_t {
+public:
+  platf::capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<platf::img_t> img, bool *cursor) override {
+    auto next_frame = std::chrono::steady_clock::now();
+
+    while(img) {
+      auto now = std::chrono::steady_clock::now();
+
+      if(next_frame > now) {
+        std::this_thread::sleep_for((next_frame - now) / 3 * 2);
+      }
+      while(next_frame > now) {
+        now = std::chrono::steady_clock::now();
+      }
+      next_frame = now + delay;
+
+      auto status = snapshot(img.get(), 1000ms, *cursor);
+      switch(status) {
+      case platf::capture_e::reinit:
+      case platf::capture_e::error:
+        return status;
+      case platf::capture_e::timeout:
+        continue;
+      case platf::capture_e::ok:
+        img = snapshot_cb(img);
+        break;
+      default:
+        BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+        return status;
+      }
+    }
+
+    return platf::capture_e::ok;
+  }
+
+  platf::capture_e snapshot(platf::img_t *img_out_base, std::chrono::milliseconds timeout, bool cursor) {
+    auto status = wlr_t::snapshot(img_out_base, timeout, cursor);
+    if(status != platf::capture_e::ok) {
+      return status;
+    }
+
+    auto current_frame = dmabuf.current_frame;
+
+    auto rgb_opt = egl::import_source(egl_display.get(), current_frame->sd);
+
+    if(!rgb_opt) {
+      return platf::capture_e::reinit;
+    }
+
+    gl::ctx.BindTexture(GL_TEXTURE_2D, (*rgb_opt)->tex[0]);
+    gl::ctx.GetTextureSubImage((*rgb_opt)->tex[0], 0, 0, 0, 0, width, height, 1, GL_BGRA, GL_UNSIGNED_BYTE, img_out_base->height * img_out_base->row_pitch, img_out_base->data);
+    gl::ctx.BindTexture(GL_TEXTURE_2D, 0);
+
+    return platf::capture_e::ok;
+  }
+
+  int init(platf::mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+    if(wlr_t::init(hwdevice_type, display_name, framerate)) {
+      return -1;
+    }
+
+    egl_display = egl::make_display(display.get());
+    if(!egl_display) {
+      return -1;
+    }
+
+    auto ctx_opt = egl::make_ctx(egl_display.get());
+    if(!ctx_opt) {
+      return -1;
+    }
+
+    ctx = std::move(*ctx_opt);
+
+    return 0;
+  }
+
+  std::shared_ptr<platf::hwdevice_t> make_hwdevice(platf::pix_fmt_e pix_fmt) override {
+    if(mem_type == platf::mem_type_e::vaapi) {
+      return va::make_hwdevice(width, height, false);
+    }
+
+    return std::make_shared<platf::hwdevice_t>();
+  }
+
+  std::shared_ptr<platf::img_t> alloc_img() override {
+    auto img         = std::make_shared<img_t>();
+    img->width       = width;
+    img->height      = height;
+    img->pixel_pitch = 4;
+    img->row_pitch   = img->pixel_pitch * width;
+    img->data        = new std::uint8_t[height * img->row_pitch];
+
+    return img;
+  }
+
+  egl::display_t egl_display;
+  egl::ctx_t ctx;
+};
+
+class wlr_vram_t : public wlr_t {
+public:
+  platf::capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<platf::img_t> img, bool *cursor) override {
+    auto next_frame = std::chrono::steady_clock::now();
+
+    while(img) {
+      auto now = std::chrono::steady_clock::now();
+
+      if(next_frame > now) {
+        std::this_thread::sleep_for((next_frame - now) / 3 * 2);
+      }
+      while(next_frame > now) {
+        now = std::chrono::steady_clock::now();
+      }
+      next_frame = now + delay;
+
+      auto status = snapshot(img.get(), 1000ms, *cursor);
+      switch(status) {
+      case platf::capture_e::reinit:
+      case platf::capture_e::error:
+        return status;
+      case platf::capture_e::timeout:
+        continue;
+      case platf::capture_e::ok:
+        img = snapshot_cb(img);
+        break;
+      default:
+        BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+        return status;
+      }
+    }
+
+    return platf::capture_e::ok;
+  }
+
+  platf::capture_e snapshot(platf::img_t *img_out_base, std::chrono::milliseconds timeout, bool cursor) {
+    auto status = wlr_t::snapshot(img_out_base, timeout, cursor);
+    if(status != platf::capture_e::ok) {
+      return status;
+    }
+
+    auto img = (egl::img_descriptor_t *)img_out_base;
+    img->reset();
+
+    auto current_frame = dmabuf.current_frame;
+
+    ++sequence;
+    img->sequence = sequence;
+
+    img->sd = current_frame->sd;
+
+    // Prevent dmabuf from closing the file descriptors.
+    std::fill_n(current_frame->sd.fds, 4, -1);
+
+    return platf::capture_e::ok;
+  }
+
+  std::shared_ptr<platf::img_t> alloc_img() override {
+    auto img = std::make_shared<egl::img_descriptor_t>();
+
+    img->sequence = 0;
+    img->serial   = std::numeric_limits<decltype(img->serial)>::max();
+    img->data     = nullptr;
+
+    // File descriptors aren't open
+    std::fill_n(img->sd.fds, 4, -1);
+
+    return img;
+  }
+
+  std::shared_ptr<platf::hwdevice_t> make_hwdevice(platf::pix_fmt_e pix_fmt) override {
+    if(mem_type == platf::mem_type_e::vaapi) {
+      return va::make_hwdevice(width, height, 0, 0, true);
+    }
+
+    return std::make_shared<platf::hwdevice_t>();
+  }
+
+  int dummy_img(platf::img_t *img) override {
+    return snapshot(img, 1000ms, false) != platf::capture_e::ok;
+  }
+
+  std::uint64_t sequence {};
+};
+
+} // namespace wl
+
+namespace platf {
+std::shared_ptr<display_t> wl_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+  if(hwdevice_type != platf::mem_type_e::system && hwdevice_type != platf::mem_type_e::vaapi && hwdevice_type != platf::mem_type_e::cuda) {
+    BOOST_LOG(error) << "Could not initialize display with the given hw device type."sv;
+    return nullptr;
+  }
+
+  if(hwdevice_type == platf::mem_type_e::vaapi) {
+    auto wlr = std::make_shared<wl::wlr_vram_t>();
+    if(wlr->init(hwdevice_type, display_name, framerate)) {
+      return nullptr;
+    }
+
+    return wlr;
+  }
+
+  auto wlr = std::make_shared<wl::wlr_ram_t>();
+  if(wlr->init(hwdevice_type, display_name, framerate)) {
+    return nullptr;
+  }
+
+  return wlr;
+}
+
+std::vector<std::string> wl_display_names() {
+  std::vector<std::string> display_names;
+
+  wl::display_t display;
+  if(display.init()) {
+    return {};
+  }
+
+  wl::interface_t interface;
+  interface.listen(display.registry());
+
+  display.roundtrip();
+
+  if(!interface[wl::interface_t::XDG_OUTPUT]) {
+    BOOST_LOG(warning) << "Missing Wayland wire for xdg_output"sv;
+    return {};
+  }
+
+  if(!interface[wl::interface_t::WLR_EXPORT_DMABUF]) {
+    BOOST_LOG(warning) << "Missing Wayland wire for wlr-export-dmabuf"sv;
+    return {};
+  }
+
+  wl::env_width  = 0;
+  wl::env_height = 0;
+
+  for(auto &monitor : interface.monitors) {
+    monitor->listen(interface.output_manager);
+  }
+
+  display.roundtrip();
+
+  for(int x = 0; x < interface.monitors.size(); ++x) {
+    auto monitor = interface.monitors[x].get();
+
+    wl::env_width  = std::max(wl::env_width, (int)(monitor->viewport.offset_x + monitor->viewport.width));
+    wl::env_height = std::max(wl::env_height, (int)(monitor->viewport.offset_y + monitor->viewport.height));
+
+    display_names.emplace_back(std::to_string(x));
+  }
+
+  return display_names;
+}
+
+} // namespace platf

src/platform/linux/x11grab.cpp

@@ -0,0 +1,837 @@
+//
+// Created by loki on 6/21/19.
+//
+
+#include "src/platform/common.h"
+
+#include <fstream>
+
+#include <X11/X.h>
+#include <X11/Xlib.h>
+#include <X11/Xutil.h>
+#include <X11/extensions/Xfixes.h>
+#include <X11/extensions/Xrandr.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+#include <xcb/shm.h>
+#include <xcb/xfixes.h>
+
+#include "src/config.h"
+#include "src/main.h"
+#include "src/task_pool.h"
+
+#include "cuda.h"
+#include "graphics.h"
+#include "misc.h"
+#include "vaapi.h"
+#include "x11grab.h"
+
+using namespace std::literals;
+
+namespace platf {
+int load_xcb();
+int load_x11();
+
+namespace x11 {
+#define _FN(x, ret, args)    \
+  typedef ret(*x##_fn) args; \
+  static x##_fn x
+
+_FN(GetImage, XImage *,
+  (
+    Display * display,
+    Drawable d,
+    int x, int y,
+    unsigned int width, unsigned int height,
+    unsigned long plane_mask,
+    int format));
+
+_FN(OpenDisplay, Display *, (_Xconst char *display_name));
+_FN(GetWindowAttributes, Status,
+  (
+    Display * display,
+    Window w,
+    XWindowAttributes *window_attributes_return));
+
+_FN(CloseDisplay, int, (Display * display));
+_FN(Free, int, (void *data));
+_FN(InitThreads, Status, (void));
+
+namespace rr {
+_FN(GetScreenResources, XRRScreenResources *, (Display * dpy, Window window));
+_FN(GetOutputInfo, XRROutputInfo *, (Display * dpy, XRRScreenResources *resources, RROutput output));
+_FN(GetCrtcInfo, XRRCrtcInfo *, (Display * dpy, XRRScreenResources *resources, RRCrtc crtc));
+_FN(FreeScreenResources, void, (XRRScreenResources * resources));
+_FN(FreeOutputInfo, void, (XRROutputInfo * outputInfo));
+_FN(FreeCrtcInfo, void, (XRRCrtcInfo * crtcInfo));
+
+int init() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libXrandr.so.2", "libXrandr.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&GetScreenResources, "XRRGetScreenResources" },
+    { (dyn::apiproc *)&GetOutputInfo, "XRRGetOutputInfo" },
+    { (dyn::apiproc *)&GetCrtcInfo, "XRRGetCrtcInfo" },
+    { (dyn::apiproc *)&FreeScreenResources, "XRRFreeScreenResources" },
+    { (dyn::apiproc *)&FreeOutputInfo, "XRRFreeOutputInfo" },
+    { (dyn::apiproc *)&FreeCrtcInfo, "XRRFreeCrtcInfo" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+} // namespace rr
+namespace fix {
+_FN(GetCursorImage, XFixesCursorImage *, (Display * dpy));
+
+int init() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libXfixes.so.3", "libXfixes.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&GetCursorImage, "XFixesGetCursorImage" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+} // namespace fix
+
+int init() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libX11.so.6", "libX11.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&GetImage, "XGetImage" },
+    { (dyn::apiproc *)&OpenDisplay, "XOpenDisplay" },
+    { (dyn::apiproc *)&GetWindowAttributes, "XGetWindowAttributes" },
+    { (dyn::apiproc *)&Free, "XFree" },
+    { (dyn::apiproc *)&CloseDisplay, "XCloseDisplay" },
+    { (dyn::apiproc *)&InitThreads, "XInitThreads" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+} // namespace x11
+
+namespace xcb {
+static xcb_extension_t *shm_id;
+
+_FN(shm_get_image_reply, xcb_shm_get_image_reply_t *,
+  (
+    xcb_connection_t * c,
+    xcb_shm_get_image_cookie_t cookie,
+    xcb_generic_error_t **e));
+
+_FN(shm_get_image_unchecked, xcb_shm_get_image_cookie_t,
+  (
+    xcb_connection_t * c,
+    xcb_drawable_t drawable,
+    int16_t x, int16_t y,
+    uint16_t width, uint16_t height,
+    uint32_t plane_mask,
+    uint8_t format,
+    xcb_shm_seg_t shmseg,
+    uint32_t offset));
+
+_FN(shm_attach, xcb_void_cookie_t,
+  (xcb_connection_t * c,
+    xcb_shm_seg_t shmseg,
+    uint32_t shmid,
+    uint8_t read_only));
+
+_FN(get_extension_data, xcb_query_extension_reply_t *,
+  (xcb_connection_t * c, xcb_extension_t *ext));
+
+_FN(get_setup, xcb_setup_t *, (xcb_connection_t * c));
+_FN(disconnect, void, (xcb_connection_t * c));
+_FN(connection_has_error, int, (xcb_connection_t * c));
+_FN(connect, xcb_connection_t *, (const char *displayname, int *screenp));
+_FN(setup_roots_iterator, xcb_screen_iterator_t, (const xcb_setup_t *R));
+_FN(generate_id, std::uint32_t, (xcb_connection_t * c));
+
+int init_shm() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libxcb-shm.so.0", "libxcb-shm.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&shm_id, "xcb_shm_id" },
+    { (dyn::apiproc *)&shm_get_image_reply, "xcb_shm_get_image_reply" },
+    { (dyn::apiproc *)&shm_get_image_unchecked, "xcb_shm_get_image_unchecked" },
+    { (dyn::apiproc *)&shm_attach, "xcb_shm_attach" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+int init() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libxcb.so.1", "libxcb.so" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&get_extension_data, "xcb_get_extension_data" },
+    { (dyn::apiproc *)&get_setup, "xcb_get_setup" },
+    { (dyn::apiproc *)&disconnect, "xcb_disconnect" },
+    { (dyn::apiproc *)&connection_has_error, "xcb_connection_has_error" },
+    { (dyn::apiproc *)&connect, "xcb_connect" },
+    { (dyn::apiproc *)&setup_roots_iterator, "xcb_setup_roots_iterator" },
+    { (dyn::apiproc *)&generate_id, "xcb_generate_id" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+#undef _FN
+} // namespace xcb
+
+void freeImage(XImage *);
+void freeX(XFixesCursorImage *);
+
+using xcb_connect_t = util::dyn_safe_ptr<xcb_connection_t, &xcb::disconnect>;
+using xcb_img_t     = util::c_ptr<xcb_shm_get_image_reply_t>;
+
+using ximg_t    = util::safe_ptr<XImage, freeImage>;
+using xcursor_t = util::safe_ptr<XFixesCursorImage, freeX>;
+
+using crtc_info_t   = util::dyn_safe_ptr<_XRRCrtcInfo, &x11::rr::FreeCrtcInfo>;
+using output_info_t = util::dyn_safe_ptr<_XRROutputInfo, &x11::rr::FreeOutputInfo>;
+using screen_res_t  = util::dyn_safe_ptr<_XRRScreenResources, &x11::rr::FreeScreenResources>;
+
+class shm_id_t {
+public:
+  shm_id_t() : id { -1 } {}
+  shm_id_t(int id) : id { id } {}
+  shm_id_t(shm_id_t &&other) noexcept : id(other.id) {
+    other.id = -1;
+  }
+
+  ~shm_id_t() {
+    if(id != -1) {
+      shmctl(id, IPC_RMID, nullptr);
+      id = -1;
+    }
+  }
+  int id;
+};
+
+class shm_data_t {
+public:
+  shm_data_t() : data { (void *)-1 } {}
+  shm_data_t(void *data) : data { data } {}
+
+  shm_data_t(shm_data_t &&other) noexcept : data(other.data) {
+    other.data = (void *)-1;
+  }
+
+  ~shm_data_t() {
+    if((std::uintptr_t)data != -1) {
+      shmdt(data);
+    }
+  }
+
+  void *data;
+};
+
+struct x11_img_t : public img_t {
+  ximg_t img;
+};
+
+struct shm_img_t : public img_t {
+  ~shm_img_t() override {
+    delete[] data;
+    data = nullptr;
+  }
+};
+
+static void blend_cursor(Display *display, img_t &img, int offsetX, int offsetY) {
+  xcursor_t overlay { x11::fix::GetCursorImage(display) };
+
+  if(!overlay) {
+    BOOST_LOG(error) << "Couldn't get cursor from XFixesGetCursorImage"sv;
+    return;
+  }
+
+  overlay->x -= overlay->xhot;
+  overlay->y -= overlay->yhot;
+
+  overlay->x -= offsetX;
+  overlay->y -= offsetY;
+
+  overlay->x = std::max((short)0, overlay->x);
+  overlay->y = std::max((short)0, overlay->y);
+
+  auto pixels = (int *)img.data;
+
+  auto screen_height = img.height;
+  auto screen_width  = img.width;
+
+  auto delta_height = std::min<uint16_t>(overlay->height, std::max(0, screen_height - overlay->y));
+  auto delta_width  = std::min<uint16_t>(overlay->width, std::max(0, screen_width - overlay->x));
+  for(auto y = 0; y < delta_height; ++y) {
+    auto overlay_begin = &overlay->pixels[y * overlay->width];
+    auto overlay_end   = &overlay->pixels[y * overlay->width + delta_width];
+
+    auto pixels_begin = &pixels[(y + overlay->y) * (img.row_pitch / img.pixel_pitch) + overlay->x];
+
+    std::for_each(overlay_begin, overlay_end, [&](long pixel) {
+      int *pixel_p = (int *)&pixel;
+
+      auto colors_in = (uint8_t *)pixels_begin;
+
+      auto alpha = (*(uint *)pixel_p) >> 24u;
+      if(alpha == 255) {
+        *pixels_begin = *pixel_p;
+      }
+      else {
+        auto colors_out = (uint8_t *)pixel_p;
+        colors_in[0]    = colors_out[0] + (colors_in[0] * (255 - alpha) + 255 / 2) / 255;
+        colors_in[1]    = colors_out[1] + (colors_in[1] * (255 - alpha) + 255 / 2) / 255;
+        colors_in[2]    = colors_out[2] + (colors_in[2] * (255 - alpha) + 255 / 2) / 255;
+      }
+      ++pixels_begin;
+    });
+  }
+}
+
+struct x11_attr_t : public display_t {
+  std::chrono::nanoseconds delay;
+
+  x11::xdisplay_t xdisplay;
+  Window xwindow;
+  XWindowAttributes xattr;
+
+  mem_type_e mem_type;
+
+  /*
+   * Last X (NOT the streamed monitor!) size.
+   * This way we can trigger reinitialization if the dimensions changed while streaming
+   */
+  // int env_width, env_height;
+
+  x11_attr_t(mem_type_e mem_type) : xdisplay { x11::OpenDisplay(nullptr) }, xwindow {}, xattr {}, mem_type { mem_type } {
+    x11::InitThreads();
+  }
+
+  int init(const std::string &display_name, int framerate) {
+    if(!xdisplay) {
+      BOOST_LOG(error) << "Could not open X11 display"sv;
+      return -1;
+    }
+
+    delay = std::chrono::nanoseconds { 1s } / framerate;
+
+    xwindow = DefaultRootWindow(xdisplay.get());
+
+    refresh();
+
+    int streamedMonitor = -1;
+    if(!display_name.empty()) {
+      streamedMonitor = (int)util::from_view(display_name);
+    }
+
+    if(streamedMonitor != -1) {
+      BOOST_LOG(info) << "Configuring selected monitor ("sv << streamedMonitor << ") to stream"sv;
+      screen_res_t screenr { x11::rr::GetScreenResources(xdisplay.get(), xwindow) };
+      int output = screenr->noutput;
+
+      output_info_t result;
+      int monitor = 0;
+      for(int x = 0; x < output; ++x) {
+        output_info_t out_info { x11::rr::GetOutputInfo(xdisplay.get(), screenr.get(), screenr->outputs[x]) };
+        if(out_info && out_info->connection == RR_Connected) {
+          if(monitor++ == streamedMonitor) {
+            result = std::move(out_info);
+            break;
+          }
+        }
+      }
+
+      if(!result) {
+        BOOST_LOG(error) << "Could not stream display number ["sv << streamedMonitor << "], there are only ["sv << monitor << "] displays."sv;
+        return -1;
+      }
+
+      if(result->crtc) {
+        crtc_info_t crt_info { x11::rr::GetCrtcInfo(xdisplay.get(), screenr.get(), result->crtc) };
+        BOOST_LOG(info)
+          << "Streaming display: "sv << result->name << " with res "sv << crt_info->width << 'x' << crt_info->height << " offset by "sv << crt_info->x << 'x' << crt_info->y;
+
+        width    = crt_info->width;
+        height   = crt_info->height;
+        offset_x = crt_info->x;
+        offset_y = crt_info->y;
+      }
+      else {
+        BOOST_LOG(warning) << "Couldn't get requested display info, defaulting to recording entire virtual desktop"sv;
+        width  = xattr.width;
+        height = xattr.height;
+      }
+    }
+    else {
+      width  = xattr.width;
+      height = xattr.height;
+    }
+
+    env_width  = xattr.width;
+    env_height = xattr.height;
+
+    return 0;
+  }
+
+  /**
+   * Called when the display attributes should change.
+   */
+  void refresh() {
+    x11::GetWindowAttributes(xdisplay.get(), xwindow, &xattr); //Update xattr's
+  }
+
+  capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) override {
+    auto next_frame = std::chrono::steady_clock::now();
+
+    while(img) {
+      auto now = std::chrono::steady_clock::now();
+
+      if(next_frame > now) {
+        std::this_thread::sleep_for((next_frame - now) / 3 * 2);
+      }
+      while(next_frame > now) {
+        std::this_thread::sleep_for(1ns);
+        now = std::chrono::steady_clock::now();
+      }
+      next_frame = now + delay;
+
+      auto status = snapshot(img.get(), 1000ms, *cursor);
+      switch(status) {
+      case platf::capture_e::reinit:
+      case platf::capture_e::error:
+        return status;
+      case platf::capture_e::timeout:
+        std::this_thread::sleep_for(1ms);
+        continue;
+      case platf::capture_e::ok:
+        img = snapshot_cb(img);
+        break;
+      default:
+        BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+        return status;
+      }
+    }
+
+    return capture_e::ok;
+  }
+
+  capture_e snapshot(img_t *img_out_base, std::chrono::milliseconds timeout, bool cursor) {
+    refresh();
+
+    //The whole X server changed, so we gotta reinit everything
+    if(xattr.width != env_width || xattr.height != env_height) {
+      BOOST_LOG(warning) << "X dimensions changed in non-SHM mode, request reinit"sv;
+      return capture_e::reinit;
+    }
+    XImage *img { x11::GetImage(xdisplay.get(), xwindow, offset_x, offset_y, width, height, AllPlanes, ZPixmap) };
+
+    auto img_out         = (x11_img_t *)img_out_base;
+    img_out->width       = img->width;
+    img_out->height      = img->height;
+    img_out->data        = (uint8_t *)img->data;
+    img_out->row_pitch   = img->bytes_per_line;
+    img_out->pixel_pitch = img->bits_per_pixel / 8;
+    img_out->img.reset(img);
+
+    if(cursor) {
+      blend_cursor(xdisplay.get(), *img_out_base, offset_x, offset_y);
+    }
+
+    return capture_e::ok;
+  }
+
+  std::shared_ptr<img_t> alloc_img() override {
+    return std::make_shared<x11_img_t>();
+  }
+
+  std::shared_ptr<hwdevice_t> make_hwdevice(pix_fmt_e pix_fmt) override {
+    if(mem_type == mem_type_e::vaapi) {
+      return va::make_hwdevice(width, height, false);
+    }
+
+#ifdef SUNSHINE_BUILD_CUDA
+    if(mem_type == mem_type_e::cuda) {
+      return cuda::make_hwdevice(width, height, false);
+    }
+#endif
+
+    return std::make_shared<hwdevice_t>();
+  }
+
+  int dummy_img(img_t *img) override {
+    snapshot(img, 0s, true);
+    return 0;
+  }
+};
+
+struct shm_attr_t : public x11_attr_t {
+  x11::xdisplay_t shm_xdisplay; // Prevent race condition with x11_attr_t::xdisplay
+  xcb_connect_t xcb;
+  xcb_screen_t *display;
+  std::uint32_t seg;
+
+  shm_id_t shm_id;
+
+  shm_data_t data;
+
+  util::TaskPool::task_id_t refresh_task_id;
+
+  void delayed_refresh() {
+    refresh();
+
+    refresh_task_id = task_pool.pushDelayed(&shm_attr_t::delayed_refresh, 2s, this).task_id;
+  }
+
+  shm_attr_t(mem_type_e mem_type) : x11_attr_t(mem_type), shm_xdisplay { x11::OpenDisplay(nullptr) } {
+    refresh_task_id = task_pool.pushDelayed(&shm_attr_t::delayed_refresh, 2s, this).task_id;
+  }
+
+  ~shm_attr_t() override {
+    while(!task_pool.cancel(refresh_task_id))
+      ;
+  }
+
+  capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) override {
+    auto next_frame = std::chrono::steady_clock::now();
+
+    while(img) {
+      auto now = std::chrono::steady_clock::now();
+
+      if(next_frame > now) {
+        std::this_thread::sleep_for((next_frame - now) / 3 * 2);
+      }
+      while(next_frame > now) {
+        std::this_thread::sleep_for(1ns);
+        now = std::chrono::steady_clock::now();
+      }
+      next_frame = now + delay;
+
+      auto status = snapshot(img.get(), 1000ms, *cursor);
+      switch(status) {
+      case platf::capture_e::reinit:
+      case platf::capture_e::error:
+        return status;
+      case platf::capture_e::timeout:
+        std::this_thread::sleep_for(1ms);
+        continue;
+      case platf::capture_e::ok:
+        img = snapshot_cb(img);
+        break;
+      default:
+        BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+        return status;
+      }
+    }
+
+    return capture_e::ok;
+  }
+
+  capture_e snapshot(img_t *img, std::chrono::milliseconds timeout, bool cursor) {
+    //The whole X server changed, so we gotta reinit everything
+    if(xattr.width != env_width || xattr.height != env_height) {
+      BOOST_LOG(warning) << "X dimensions changed in SHM mode, request reinit"sv;
+      return capture_e::reinit;
+    }
+    else {
+      auto img_cookie = xcb::shm_get_image_unchecked(xcb.get(), display->root, offset_x, offset_y, width, height, ~0, XCB_IMAGE_FORMAT_Z_PIXMAP, seg, 0);
+
+      xcb_img_t img_reply { xcb::shm_get_image_reply(xcb.get(), img_cookie, nullptr) };
+      if(!img_reply) {
+        BOOST_LOG(error) << "Could not get image reply"sv;
+        return capture_e::reinit;
+      }
+
+      std::copy_n((std::uint8_t *)data.data, frame_size(), img->data);
+
+      if(cursor) {
+        blend_cursor(shm_xdisplay.get(), *img, offset_x, offset_y);
+      }
+
+      return capture_e::ok;
+    }
+  }
+
+  std::shared_ptr<img_t> alloc_img() override {
+    auto img         = std::make_shared<shm_img_t>();
+    img->width       = width;
+    img->height      = height;
+    img->pixel_pitch = 4;
+    img->row_pitch   = img->pixel_pitch * width;
+    img->data        = new std::uint8_t[height * img->row_pitch];
+
+    return img;
+  }
+
+  int dummy_img(platf::img_t *img) override {
+    return 0;
+  }
+
+  int init(const std::string &display_name, int framerate) {
+    if(x11_attr_t::init(display_name, framerate)) {
+      return 1;
+    }
+
+    shm_xdisplay.reset(x11::OpenDisplay(nullptr));
+    xcb.reset(xcb::connect(nullptr, nullptr));
+    if(xcb::connection_has_error(xcb.get())) {
+      return -1;
+    }
+
+    if(!xcb::get_extension_data(xcb.get(), xcb::shm_id)->present) {
+      BOOST_LOG(error) << "Missing SHM extension"sv;
+
+      return -1;
+    }
+
+    auto iter = xcb::setup_roots_iterator(xcb::get_setup(xcb.get()));
+    display   = iter.data;
+    seg       = xcb::generate_id(xcb.get());
+
+    shm_id.id = shmget(IPC_PRIVATE, frame_size(), IPC_CREAT | 0777);
+    if(shm_id.id == -1) {
+      BOOST_LOG(error) << "shmget failed"sv;
+      return -1;
+    }
+
+    xcb::shm_attach(xcb.get(), seg, shm_id.id, false);
+    data.data = shmat(shm_id.id, nullptr, 0);
+
+    if((uintptr_t)data.data == -1) {
+      BOOST_LOG(error) << "shmat failed"sv;
+
+      return -1;
+    }
+
+    return 0;
+  }
+
+  std::uint32_t frame_size() {
+    return width * height * 4;
+  }
+};
+
+std::shared_ptr<display_t> x11_display(platf::mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+  if(hwdevice_type != platf::mem_type_e::system && hwdevice_type != platf::mem_type_e::vaapi && hwdevice_type != platf::mem_type_e::cuda) {
+    BOOST_LOG(error) << "Could not initialize x11 display with the given hw device type"sv;
+    return nullptr;
+  }
+
+  if(xcb::init_shm() || xcb::init() || x11::init() || x11::rr::init() || x11::fix::init()) {
+    BOOST_LOG(error) << "Couldn't init x11 libraries"sv;
+
+    return nullptr;
+  }
+
+  // Attempt to use shared memory X11 to avoid copying the frame
+  auto shm_disp = std::make_shared<shm_attr_t>(hwdevice_type);
+
+  auto status = shm_disp->init(display_name, framerate);
+  if(status > 0) {
+    // x11_attr_t::init() failed, don't bother trying again.
+    return nullptr;
+  }
+
+  if(status == 0) {
+    return shm_disp;
+  }
+
+  // Fallback
+  auto x11_disp = std::make_shared<x11_attr_t>(hwdevice_type);
+  if(x11_disp->init(display_name, framerate)) {
+    return nullptr;
+  }
+
+  return x11_disp;
+}
+
+std::vector<std::string> x11_display_names() {
+  if(load_x11() || load_xcb()) {
+    BOOST_LOG(error) << "Couldn't init x11 libraries"sv;
+
+    return {};
+  }
+
+  BOOST_LOG(info) << "Detecting connected monitors"sv;
+
+  x11::xdisplay_t xdisplay { x11::OpenDisplay(nullptr) };
+  if(!xdisplay) {
+    return {};
+  }
+
+  auto xwindow = DefaultRootWindow(xdisplay.get());
+  screen_res_t screenr { x11::rr::GetScreenResources(xdisplay.get(), xwindow) };
+  int output = screenr->noutput;
+
+  int monitor = 0;
+  for(int x = 0; x < output; ++x) {
+    output_info_t out_info { x11::rr::GetOutputInfo(xdisplay.get(), screenr.get(), screenr->outputs[x]) };
+    if(out_info && out_info->connection == RR_Connected) {
+      ++monitor;
+    }
+  }
+
+  std::vector<std::string> names;
+  names.reserve(monitor);
+
+  for(auto x = 0; x < monitor; ++x) {
+    names.emplace_back(std::to_string(x));
+  }
+
+  return names;
+}
+
+void freeImage(XImage *p) {
+  XDestroyImage(p);
+}
+void freeX(XFixesCursorImage *p) {
+  x11::Free(p);
+}
+
+int load_xcb() {
+  // This will be called once only
+  static int xcb_status = xcb::init_shm() || xcb::init();
+
+  return xcb_status;
+}
+
+int load_x11() {
+  // This will be called once only
+  static int x11_status =
+    window_system == window_system_e::NONE ||
+    x11::init() || x11::rr::init() || x11::fix::init();
+
+  return x11_status;
+}
+
+namespace x11 {
+std::optional<cursor_t> cursor_t::make() {
+  if(load_x11()) {
+    return std::nullopt;
+  }
+
+  cursor_t cursor;
+
+  cursor.ctx.reset((cursor_ctx_t::pointer)x11::OpenDisplay(nullptr));
+
+  return cursor;
+}
+
+void cursor_t::capture(egl::cursor_t &img) {
+  auto display = (xdisplay_t::pointer)ctx.get();
+
+  xcursor_t xcursor = fix::GetCursorImage(display);
+
+  if(img.serial != xcursor->cursor_serial) {
+    auto buf_size = xcursor->width * xcursor->height * sizeof(int);
+
+    if(img.buffer.size() < buf_size) {
+      img.buffer.resize(buf_size);
+    }
+
+    std::transform(xcursor->pixels, xcursor->pixels + buf_size / 4, (int *)img.buffer.data(), [](long pixel) -> int {
+      return pixel;
+    });
+  }
+
+  img.data        = img.buffer.data();
+  img.width       = xcursor->width;
+  img.height      = xcursor->height;
+  img.x           = xcursor->x - xcursor->xhot;
+  img.y           = xcursor->y - xcursor->yhot;
+  img.pixel_pitch = 4;
+  img.row_pitch   = img.pixel_pitch * img.width;
+  img.serial      = xcursor->cursor_serial;
+}
+
+void cursor_t::blend(img_t &img, int offsetX, int offsetY) {
+  blend_cursor((xdisplay_t::pointer)ctx.get(), img, offsetX, offsetY);
+}
+
+xdisplay_t make_display() {
+  return OpenDisplay(nullptr);
+}
+
+void freeDisplay(_XDisplay *xdisplay) {
+  CloseDisplay(xdisplay);
+}
+
+void freeCursorCtx(cursor_ctx_t::pointer ctx) {
+  CloseDisplay((xdisplay_t::pointer)ctx);
+}
+} // namespace x11
+} // namespace platf

src/platform/linux/x11grab.h

@@ -0,0 +1,60 @@
+#ifndef SUNSHINE_X11_GRAB
+#define SUNSHINE_X11_GRAB
+
+#include <optional>
+
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+// X11 Display
+extern "C" struct _XDisplay;
+
+namespace egl {
+class cursor_t;
+}
+
+namespace platf::x11 {
+
+#ifdef SUNSHINE_BUILD_X11
+struct cursor_ctx_raw_t;
+void freeCursorCtx(cursor_ctx_raw_t *ctx);
+void freeDisplay(_XDisplay *xdisplay);
+
+using cursor_ctx_t = util::safe_ptr<cursor_ctx_raw_t, freeCursorCtx>;
+using xdisplay_t   = util::safe_ptr<_XDisplay, freeDisplay>;
+
+class cursor_t {
+public:
+  static std::optional<cursor_t> make();
+
+  void capture(egl::cursor_t &img);
+
+  /**
+   * Capture and blend the cursor into the image
+   * 
+   * img <-- destination image
+   * offsetX, offsetY <--- Top left corner of the virtual screen
+   */
+  void blend(img_t &img, int offsetX, int offsetY);
+
+  cursor_ctx_t ctx;
+};
+
+xdisplay_t make_display();
+#else
+// It's never something different from nullptr
+util::safe_ptr<_XDisplay, std::default_delete<_XDisplay>>;
+
+class cursor_t {
+public:
+  static std::optional<cursor_t> make() { return std::nullopt; }
+
+  void capture(egl::cursor_t &) {}
+  void blend(img_t &, int, int) {}
+};
+
+xdisplay_t make_display() { return nullptr; }
+#endif
+} // namespace platf::x11
+
+#endif

src/platform/macos/av_audio.h

@@ -0,0 +1,26 @@
+#ifndef SUNSHINE_PLATFORM_AV_AUDIO_H
+#define SUNSHINE_PLATFORM_AV_AUDIO_H
+
+#import <AVFoundation/AVFoundation.h>
+
+#include "third-party/TPCircularBuffer/TPCircularBuffer.h"
+
+#define kBufferLength 2048
+
+@interface AVAudio : NSObject <AVCaptureAudioDataOutputSampleBufferDelegate> {
+@public
+  TPCircularBuffer audioSampleBuffer;
+}
+
+@property(nonatomic, assign) AVCaptureSession *audioCaptureSession;
+@property(nonatomic, assign) AVCaptureConnection *audioConnection;
+@property(nonatomic, assign) NSCondition *samplesArrivedSignal;
+
++ (NSArray *)microphoneNames;
++ (AVCaptureDevice *)findMicrophone:(NSString *)name;
+
+- (int)setupMicrophone:(AVCaptureDevice *)device sampleRate:(UInt32)sampleRate frameSize:(UInt32)frameSize channels:(UInt8)channels;
+
+@end
+
+#endif //SUNSHINE_PLATFORM_AV_AUDIO_H

src/platform/macos/av_audio.m

@@ -0,0 +1,120 @@
+#import "av_audio.h"
+
+@implementation AVAudio
+
++ (NSArray<AVCaptureDevice *> *)microphones {
+  AVCaptureDeviceDiscoverySession *discoverySession = [AVCaptureDeviceDiscoverySession discoverySessionWithDeviceTypes:@[AVCaptureDeviceTypeBuiltInMicrophone,
+    AVCaptureDeviceTypeExternalUnknown]
+                                                                                                             mediaType:AVMediaTypeAudio
+                                                                                                              position:AVCaptureDevicePositionUnspecified];
+  return discoverySession.devices;
+}
+
++ (NSArray<NSString *> *)microphoneNames {
+  NSMutableArray *result = [[NSMutableArray alloc] init];
+
+  for(AVCaptureDevice *device in [AVAudio microphones]) {
+    [result addObject:[device localizedName]];
+  }
+
+  return result;
+}
+
++ (AVCaptureDevice *)findMicrophone:(NSString *)name {
+  for(AVCaptureDevice *device in [AVAudio microphones]) {
+    if([[device localizedName] isEqualToString:name]) {
+      return device;
+    }
+  }
+
+  return nil;
+}
+
+- (void)dealloc {
+  // make sure we don't process any further samples
+  self.audioConnection = nil;
+  // make sure nothing gets stuck on this signal
+  [self.samplesArrivedSignal signal];
+  [self.samplesArrivedSignal release];
+  TPCircularBufferCleanup(&audioSampleBuffer);
+  [super dealloc];
+}
+
+- (int)setupMicrophone:(AVCaptureDevice *)device sampleRate:(UInt32)sampleRate frameSize:(UInt32)frameSize channels:(UInt8)channels {
+  self.audioCaptureSession = [[AVCaptureSession alloc] init];
+
+  NSError *error;
+  AVCaptureDeviceInput *audioInput = [AVCaptureDeviceInput deviceInputWithDevice:device error:&error];
+  if(audioInput == nil) {
+    return -1;
+  }
+
+  if([self.audioCaptureSession canAddInput:audioInput]) {
+    [self.audioCaptureSession addInput:audioInput];
+  }
+  else {
+    [audioInput dealloc];
+    return -1;
+  }
+
+  AVCaptureAudioDataOutput *audioOutput = [[AVCaptureAudioDataOutput alloc] init];
+
+  [audioOutput setAudioSettings:@{
+    (NSString *)AVFormatIDKey: [NSNumber numberWithUnsignedInt:kAudioFormatLinearPCM],
+    (NSString *)AVSampleRateKey: [NSNumber numberWithUnsignedInt:sampleRate],
+    (NSString *)AVNumberOfChannelsKey: [NSNumber numberWithUnsignedInt:channels],
+    (NSString *)AVLinearPCMBitDepthKey: [NSNumber numberWithUnsignedInt:16],
+    (NSString *)AVLinearPCMIsFloatKey: @NO,
+    (NSString *)AVLinearPCMIsNonInterleaved: @NO
+  }];
+
+  dispatch_queue_attr_t qos       = dispatch_queue_attr_make_with_qos_class(DISPATCH_QUEUE_CONCURRENT,
+          QOS_CLASS_USER_INITIATED,
+          DISPATCH_QUEUE_PRIORITY_HIGH);
+  dispatch_queue_t recordingQueue = dispatch_queue_create("audioSamplingQueue", qos);
+
+  [audioOutput setSampleBufferDelegate:self queue:recordingQueue];
+
+  if([self.audioCaptureSession canAddOutput:audioOutput]) {
+    [self.audioCaptureSession addOutput:audioOutput];
+  }
+  else {
+    [audioInput release];
+    [audioOutput release];
+    return -1;
+  }
+
+  self.audioConnection = [audioOutput connectionWithMediaType:AVMediaTypeAudio];
+
+  [self.audioCaptureSession startRunning];
+
+  [audioInput release];
+  [audioOutput release];
+
+  self.samplesArrivedSignal = [[NSCondition alloc] init];
+  TPCircularBufferInit(&self->audioSampleBuffer, kBufferLength * channels);
+
+  return 0;
+}
+
+- (void)captureOutput:(AVCaptureOutput *)output
+  didOutputSampleBuffer:(CMSampleBufferRef)sampleBuffer
+         fromConnection:(AVCaptureConnection *)connection {
+  if(connection == self.audioConnection) {
+    AudioBufferList audioBufferList;
+    CMBlockBufferRef blockBuffer;
+
+    CMSampleBufferGetAudioBufferListWithRetainedBlockBuffer(sampleBuffer, NULL, &audioBufferList, sizeof(audioBufferList), NULL, NULL, 0, &blockBuffer);
+
+    //NSAssert(audioBufferList.mNumberBuffers == 1, @"Expected interlveaved PCM format but buffer contained %u streams", audioBufferList.mNumberBuffers);
+
+    // this is safe, because an interleaved PCM stream has exactly one buffer
+    // and we don't want to do sanity checks in a performance critical exec path
+    AudioBuffer audioBuffer = audioBufferList.mBuffers[0];
+
+    TPCircularBufferProduceBytes(&self->audioSampleBuffer, audioBuffer.mData, audioBuffer.mDataByteSize);
+    [self.samplesArrivedSignal signal];
+  }
+}
+
+@end

src/platform/macos/av_img_t.h

@@ -0,0 +1,18 @@
+#ifndef av_img_t_h
+#define av_img_t_h
+
+#include "src/platform/common.h"
+
+#include <CoreMedia/CoreMedia.h>
+#include <CoreVideo/CoreVideo.h>
+
+namespace platf {
+struct av_img_t : public img_t {
+  CVPixelBufferRef pixel_buffer   = nullptr;
+  CMSampleBufferRef sample_buffer = nullptr;
+
+  ~av_img_t();
+};
+} // namespace platf
+
+#endif /* av_img_t_h */

src/platform/macos/av_video.h

@@ -0,0 +1,43 @@
+#ifndef SUNSHINE_PLATFORM_AV_VIDEO_H
+#define SUNSHINE_PLATFORM_AV_VIDEO_H
+
+#import <AVFoundation/AVFoundation.h>
+
+
+struct CaptureSession {
+  AVCaptureVideoDataOutput *output;
+  NSCondition *captureStopped;
+};
+
+@interface AVVideo : NSObject <AVCaptureVideoDataOutputSampleBufferDelegate>
+
+#define kMaxDisplays 32
+
+@property(nonatomic, assign) CGDirectDisplayID displayID;
+@property(nonatomic, assign) CMTime minFrameDuration;
+@property(nonatomic, assign) OSType pixelFormat;
+@property(nonatomic, assign) int frameWidth;
+@property(nonatomic, assign) int frameHeight;
+@property(nonatomic, assign) float scaling;
+@property(nonatomic, assign) int paddingLeft;
+@property(nonatomic, assign) int paddingRight;
+@property(nonatomic, assign) int paddingTop;
+@property(nonatomic, assign) int paddingBottom;
+
+typedef bool (^FrameCallbackBlock)(CMSampleBufferRef);
+
+@property(nonatomic, assign) AVCaptureSession *session;
+@property(nonatomic, assign) NSMapTable<AVCaptureConnection *, AVCaptureVideoDataOutput *> *videoOutputs;
+@property(nonatomic, assign) NSMapTable<AVCaptureConnection *, FrameCallbackBlock> *captureCallbacks;
+@property(nonatomic, assign) NSMapTable<AVCaptureConnection *, dispatch_semaphore_t> *captureSignals;
+
++ (NSArray<NSDictionary *> *)displayNames;
+
+- (id)initWithDisplay:(CGDirectDisplayID)displayID frameRate:(int)frameRate;
+
+- (void)setFrameWidth:(int)frameWidth frameHeight:(int)frameHeight;
+- (dispatch_semaphore_t)capture:(FrameCallbackBlock)frameCallback;
+
+@end
+
+#endif //SUNSHINE_PLATFORM_AV_VIDEO_H

src/platform/macos/av_video.m

@@ -0,0 +1,184 @@
+#import "av_video.h"
+
+@implementation AVVideo
+
+// XXX: Currently, this function only returns the screen IDs as names,
+// which is not very helpful to the user. The API to retrieve names
+// was deprecated with 10.9+.
+// However, there is a solution with little external code that can be used:
+// https://stackoverflow.com/questions/20025868/cgdisplayioserviceport-is-deprecated-in-os-x-10-9-how-to-replace
++ (NSArray<NSDictionary *> *)displayNames {
+  CGDirectDisplayID displays[kMaxDisplays];
+  uint32_t count;
+  if(CGGetActiveDisplayList(kMaxDisplays, displays, &count) != kCGErrorSuccess) {
+    return [NSArray array];
+  }
+
+  NSMutableArray *result = [NSMutableArray array];
+
+  for(uint32_t i = 0; i < count; i++) {
+    [result addObject:@{
+      @"id": [NSNumber numberWithUnsignedInt:displays[i]],
+      @"name": [NSString stringWithFormat:@"%d", displays[i]]
+    }];
+  }
+
+  return [NSArray arrayWithArray:result];
+}
+
+- (id)initWithDisplay:(CGDirectDisplayID)displayID frameRate:(int)frameRate {
+  self = [super init];
+
+  CGDisplayModeRef mode = CGDisplayCopyDisplayMode(displayID);
+
+  self.displayID        = displayID;
+  self.pixelFormat      = kCVPixelFormatType_32BGRA;
+  self.frameWidth       = CGDisplayModeGetPixelWidth(mode);
+  self.frameHeight      = CGDisplayModeGetPixelHeight(mode);
+  self.scaling          = CGDisplayPixelsWide(displayID) / CGDisplayModeGetPixelWidth(mode);
+  self.paddingLeft      = 0;
+  self.paddingRight     = 0;
+  self.paddingTop       = 0;
+  self.paddingBottom    = 0;
+  self.minFrameDuration = CMTimeMake(1, frameRate);
+  self.session          = [[AVCaptureSession alloc] init];
+  self.videoOutputs     = [[NSMapTable alloc] init];
+  self.captureCallbacks = [[NSMapTable alloc] init];
+  self.captureSignals   = [[NSMapTable alloc] init];
+
+  CFRelease(mode);
+
+  AVCaptureScreenInput *screenInput = [[AVCaptureScreenInput alloc] initWithDisplayID:self.displayID];
+  [screenInput setMinFrameDuration:self.minFrameDuration];
+
+  if([self.session canAddInput:screenInput]) {
+    [self.session addInput:screenInput];
+  }
+  else {
+    [screenInput release];
+    return nil;
+  }
+
+  [self.session startRunning];
+
+  return self;
+}
+
+- (void)dealloc {
+  [self.videoOutputs release];
+  [self.captureCallbacks release];
+  [self.captureSignals release];
+  [self.session stopRunning];
+  [super dealloc];
+}
+
+- (void)setFrameWidth:(int)frameWidth frameHeight:(int)frameHeight {
+  CGImageRef screenshot = CGDisplayCreateImage(self.displayID);
+
+  self.frameWidth  = frameWidth;
+  self.frameHeight = frameHeight;
+
+  double screenRatio = (double)CGImageGetWidth(screenshot) / (double)CGImageGetHeight(screenshot);
+  double streamRatio = (double)frameWidth / (double)frameHeight;
+
+  if(screenRatio < streamRatio) {
+    int padding        = frameWidth - (frameHeight * screenRatio);
+    self.paddingLeft   = padding / 2;
+    self.paddingRight  = padding - self.paddingLeft;
+    self.paddingTop    = 0;
+    self.paddingBottom = 0;
+  }
+  else {
+    int padding        = frameHeight - (frameWidth / screenRatio);
+    self.paddingLeft   = 0;
+    self.paddingRight  = 0;
+    self.paddingTop    = padding / 2;
+    self.paddingBottom = padding - self.paddingTop;
+  }
+
+  // XXX: if the streamed image is larger than the native resolution, we add a black box around
+  // the frame. Instead the frame should be resized entirely.
+  int delta_width = frameWidth - (CGImageGetWidth(screenshot) + self.paddingLeft + self.paddingRight);
+  if(delta_width > 0) {
+    int adjust_left  = delta_width / 2;
+    int adjust_right = delta_width - adjust_left;
+    self.paddingLeft += adjust_left;
+    self.paddingRight += adjust_right;
+  }
+
+  int delta_height = frameHeight - (CGImageGetHeight(screenshot) + self.paddingTop + self.paddingBottom);
+  if(delta_height > 0) {
+    int adjust_top    = delta_height / 2;
+    int adjust_bottom = delta_height - adjust_top;
+    self.paddingTop += adjust_top;
+    self.paddingBottom += adjust_bottom;
+  }
+
+  CFRelease(screenshot);
+}
+
+- (dispatch_semaphore_t)capture:(FrameCallbackBlock)frameCallback {
+  @synchronized(self) {
+    AVCaptureVideoDataOutput *videoOutput = [[AVCaptureVideoDataOutput alloc] init];
+
+    [videoOutput setVideoSettings:@{
+      (NSString *)kCVPixelBufferPixelFormatTypeKey: [NSNumber numberWithUnsignedInt:self.pixelFormat],
+      (NSString *)kCVPixelBufferWidthKey: [NSNumber numberWithInt:self.frameWidth],
+      (NSString *)kCVPixelBufferExtendedPixelsRightKey: [NSNumber numberWithInt:self.paddingRight],
+      (NSString *)kCVPixelBufferExtendedPixelsLeftKey: [NSNumber numberWithInt:self.paddingLeft],
+      (NSString *)kCVPixelBufferExtendedPixelsTopKey: [NSNumber numberWithInt:self.paddingTop],
+      (NSString *)kCVPixelBufferExtendedPixelsBottomKey: [NSNumber numberWithInt:self.paddingBottom],
+      (NSString *)kCVPixelBufferHeightKey: [NSNumber numberWithInt:self.frameHeight]
+    }];
+
+    dispatch_queue_attr_t qos       = dispatch_queue_attr_make_with_qos_class(DISPATCH_QUEUE_SERIAL,
+            QOS_CLASS_USER_INITIATED,
+            DISPATCH_QUEUE_PRIORITY_HIGH);
+    dispatch_queue_t recordingQueue = dispatch_queue_create("videoCaptureQueue", qos);
+    [videoOutput setSampleBufferDelegate:self queue:recordingQueue];
+
+    [self.session stopRunning];
+
+    if([self.session canAddOutput:videoOutput]) {
+      [self.session addOutput:videoOutput];
+    }
+    else {
+      [videoOutput release];
+      return nil;
+    }
+
+    AVCaptureConnection *videoConnection = [videoOutput connectionWithMediaType:AVMediaTypeVideo];
+    dispatch_semaphore_t signal          = dispatch_semaphore_create(0);
+
+    [self.videoOutputs setObject:videoOutput forKey:videoConnection];
+    [self.captureCallbacks setObject:frameCallback forKey:videoConnection];
+    [self.captureSignals setObject:signal forKey:videoConnection];
+
+    [self.session startRunning];
+
+    return signal;
+  }
+}
+
+- (void)captureOutput:(AVCaptureOutput *)captureOutput
+  didOutputSampleBuffer:(CMSampleBufferRef)sampleBuffer
+         fromConnection:(AVCaptureConnection *)connection {
+
+  FrameCallbackBlock callback = [self.captureCallbacks objectForKey:connection];
+
+  if(callback != nil) {
+    if(!callback(sampleBuffer)) {
+      @synchronized(self) {
+        [self.session stopRunning];
+        [self.captureCallbacks removeObjectForKey:connection];
+        [self.session removeOutput:[self.videoOutputs objectForKey:connection]];
+        [self.videoOutputs removeObjectForKey:connection];
+        dispatch_semaphore_signal([self.captureSignals objectForKey:connection]);
+        [self.captureSignals removeObjectForKey:connection];
+        [self.session startRunning];
+      }
+    }
+  }
+}
+
+@end

src/platform/macos/display.mm

@@ -0,0 +1,196 @@
+#include "src/platform/common.h"
+#include "src/platform/macos/av_img_t.h"
+#include "src/platform/macos/av_video.h"
+#include "src/platform/macos/nv12_zero_device.h"
+
+#include "src/config.h"
+#include "src/main.h"
+
+namespace fs = std::filesystem;
+
+namespace platf {
+using namespace std::literals;
+
+av_img_t::~av_img_t() {
+  if(pixel_buffer != NULL) {
+    CVPixelBufferUnlockBaseAddress(pixel_buffer, 0);
+  }
+
+  if(sample_buffer != nullptr) {
+    CFRelease(sample_buffer);
+  }
+
+  data = nullptr;
+}
+
+struct av_display_t : public display_t {
+  AVVideo *av_capture;
+  CGDirectDisplayID display_id;
+
+  ~av_display_t() {
+    [av_capture release];
+  }
+
+  capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) override {
+    __block auto img_next = std::move(img);
+
+    auto signal = [av_capture capture:^(CMSampleBufferRef sampleBuffer) {
+      auto av_img_next = std::static_pointer_cast<av_img_t>(img_next);
+
+      CFRetain(sampleBuffer);
+
+      CVPixelBufferRef pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer);
+      CVPixelBufferLockBaseAddress(pixelBuffer, kCVPixelBufferLock_ReadOnly);
+
+      if(av_img_next->pixel_buffer != nullptr)
+        CVPixelBufferUnlockBaseAddress(av_img_next->pixel_buffer, 0);
+
+      if(av_img_next->sample_buffer != nullptr)
+        CFRelease(av_img_next->sample_buffer);
+
+      av_img_next->sample_buffer = sampleBuffer;
+      av_img_next->pixel_buffer  = pixelBuffer;
+      img_next->data             = (uint8_t *)CVPixelBufferGetBaseAddress(pixelBuffer);
+
+      size_t extraPixels[4];
+      CVPixelBufferGetExtendedPixels(pixelBuffer, &extraPixels[0], &extraPixels[1], &extraPixels[2], &extraPixels[3]);
+
+      img_next->width       = CVPixelBufferGetWidth(pixelBuffer) + extraPixels[0] + extraPixels[1];
+      img_next->height      = CVPixelBufferGetHeight(pixelBuffer) + extraPixels[2] + extraPixels[3];
+      img_next->row_pitch   = CVPixelBufferGetBytesPerRow(pixelBuffer);
+      img_next->pixel_pitch = img_next->row_pitch / img_next->width;
+
+      img_next = snapshot_cb(img_next);
+
+      return img_next != nullptr;
+    }];
+
+    dispatch_semaphore_wait(signal, DISPATCH_TIME_FOREVER);
+
+    return capture_e::ok;
+  }
+
+  std::shared_ptr<img_t> alloc_img() override {
+    return std::make_shared<av_img_t>();
+  }
+
+  std::shared_ptr<hwdevice_t> make_hwdevice(pix_fmt_e pix_fmt) override {
+    if(pix_fmt == pix_fmt_e::yuv420p) {
+      av_capture.pixelFormat = kCVPixelFormatType_32BGRA;
+
+      return std::make_shared<hwdevice_t>();
+    }
+    else if(pix_fmt == pix_fmt_e::nv12) {
+      auto device = std::make_shared<nv12_zero_device>();
+
+      device->init(static_cast<void *>(av_capture), setResolution, setPixelFormat);
+
+      return device;
+    }
+    else {
+      BOOST_LOG(error) << "Unsupported Pixel Format."sv;
+      return nullptr;
+    }
+  }
+
+  int dummy_img(img_t *img) override {
+    auto signal = [av_capture capture:^(CMSampleBufferRef sampleBuffer) {
+      auto av_img = (av_img_t *)img;
+
+      CFRetain(sampleBuffer);
+
+      CVPixelBufferRef pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer);
+      CVPixelBufferLockBaseAddress(pixelBuffer, kCVPixelBufferLock_ReadOnly);
+
+      // XXX: next_img->img should be moved to a smart pointer with
+      // the CFRelease as custom deallocator
+      if(av_img->pixel_buffer != nullptr)
+        CVPixelBufferUnlockBaseAddress(((av_img_t *)img)->pixel_buffer, 0);
+
+      if(av_img->sample_buffer != nullptr)
+        CFRelease(av_img->sample_buffer);
+
+      av_img->sample_buffer = sampleBuffer;
+      av_img->pixel_buffer  = pixelBuffer;
+      img->data             = (uint8_t *)CVPixelBufferGetBaseAddress(pixelBuffer);
+
+      size_t extraPixels[4];
+      CVPixelBufferGetExtendedPixels(pixelBuffer, &extraPixels[0], &extraPixels[1], &extraPixels[2], &extraPixels[3]);
+
+      img->width       = CVPixelBufferGetWidth(pixelBuffer) + extraPixels[0] + extraPixels[1];
+      img->height      = CVPixelBufferGetHeight(pixelBuffer) + extraPixels[2] + extraPixels[3];
+      img->row_pitch   = CVPixelBufferGetBytesPerRow(pixelBuffer);
+      img->pixel_pitch = img->row_pitch / img->width;
+
+      return false;
+    }];
+
+    dispatch_semaphore_wait(signal, DISPATCH_TIME_FOREVER);
+
+    return 0;
+  }
+
+  /**
+   * A bridge from the pure C++ code of the hwdevice_t class to the pure Objective C code.
+   *
+   * display --> an opaque pointer to an object of this class
+   * width --> the intended capture width
+   * height --> the intended capture height
+   */
+  static void setResolution(void *display, int width, int height) {
+    [static_cast<AVVideo *>(display) setFrameWidth:width frameHeight:height];
+  }
+
+  static void setPixelFormat(void *display, OSType pixelFormat) {
+    static_cast<AVVideo *>(display).pixelFormat = pixelFormat;
+  }
+};
+
+std::shared_ptr<display_t> display(platf::mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+  if(hwdevice_type != platf::mem_type_e::system) {
+    BOOST_LOG(error) << "Could not initialize display with the given hw device type."sv;
+    return nullptr;
+  }
+
+  auto display = std::make_shared<av_display_t>();
+
+  display->display_id = CGMainDisplayID();
+  if(!display_name.empty()) {
+    auto display_array = [AVVideo displayNames];
+
+    for(NSDictionary *item in display_array) {
+      NSString *name = item[@"name"];
+      if(name.UTF8String == display_name) {
+        NSNumber *display_id = item[@"id"];
+        display->display_id  = [display_id unsignedIntValue];
+      }
+    }
+  }
+
+  display->av_capture = [[AVVideo alloc] initWithDisplay:display->display_id frameRate:framerate];
+
+  if(!display->av_capture) {
+    BOOST_LOG(error) << "Video setup failed."sv;
+    return nullptr;
+  }
+
+  display->width  = display->av_capture.frameWidth;
+  display->height = display->av_capture.frameHeight;
+
+  return display;
+}
+
+std::vector<std::string> display_names(mem_type_e hwdevice_type) {
+  __block std::vector<std::string> display_names;
+
+  auto display_array = [AVVideo displayNames];
+
+  display_names.reserve([display_array count]);
+  [display_array enumerateObjectsUsingBlock:^(NSDictionary *_Nonnull obj, NSUInteger idx, BOOL *_Nonnull stop) {
+    NSString *name = obj[@"name"];
+    display_names.push_back(name.UTF8String);
+  }];
+
+  return display_names;
+}
+}

src/platform/macos/input.cpp

@@ -0,0 +1,465 @@
+#import <Carbon/Carbon.h>
+#include <mach/mach.h>
+#include <mach/mach_time.h>
+
+#include "src/main.h"
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+// Delay for a double click
+// FIXME: we probably want to make this configurable
+#define MULTICLICK_DELAY_NS 500000000
+
+namespace platf {
+using namespace std::literals;
+
+struct macos_input_t {
+public:
+  CGDirectDisplayID display;
+  CGFloat displayScaling;
+  CGEventSourceRef source;
+
+  // keyboard related stuff
+  CGEventRef kb_event;
+  CGEventFlags kb_flags;
+
+  // mouse related stuff
+  CGEventRef mouse_event;          // mouse event source
+  bool mouse_down[3];              // mouse button status
+  uint64_t last_mouse_event[3][2]; // timestamp of last mouse events
+};
+
+// A struct to hold a Windows keycode to Mac virtual keycode mapping.
+struct KeyCodeMap {
+  int win_keycode;
+  int mac_keycode;
+};
+
+// Customized less operator for using std::lower_bound() on a KeyCodeMap array.
+bool operator<(const KeyCodeMap &a, const KeyCodeMap &b) {
+  return a.win_keycode < b.win_keycode;
+}
+
+// clang-format off
+const KeyCodeMap kKeyCodesMap[] = {
+  { 0x08 /* VKEY_BACK */,                      kVK_Delete              },
+  { 0x09 /* VKEY_TAB */,                       kVK_Tab                 },
+  { 0x0A /* VKEY_BACKTAB */,                   0x21E4                  },
+  { 0x0C /* VKEY_CLEAR */,                     kVK_ANSI_KeypadClear    },
+  { 0x0D /* VKEY_RETURN */,                    kVK_Return              },
+  { 0x10 /* VKEY_SHIFT */,                     kVK_Shift               },
+  { 0x11 /* VKEY_CONTROL */,                   kVK_Control             },
+  { 0x12 /* VKEY_MENU */,                      kVK_Option              },
+  { 0x13 /* VKEY_PAUSE */,                     -1                      },
+  { 0x14 /* VKEY_CAPITAL */,                   kVK_CapsLock            },
+  { 0x15 /* VKEY_KANA */,                      kVK_JIS_Kana            },
+  { 0x15 /* VKEY_HANGUL */,                    -1                      },
+  { 0x17 /* VKEY_JUNJA */,                     -1                      },
+  { 0x18 /* VKEY_FINAL */,                     -1                      },
+  { 0x19 /* VKEY_HANJA */,                     -1                      },
+  { 0x19 /* VKEY_KANJI */,                     -1                      },
+  { 0x1B /* VKEY_ESCAPE */,                    kVK_Escape              },
+  { 0x1C /* VKEY_CONVERT */,                   -1                      },
+  { 0x1D /* VKEY_NONCONVERT */,                -1                      },
+  { 0x1E /* VKEY_ACCEPT */,                    -1                      },
+  { 0x1F /* VKEY_MODECHANGE */,                -1                      },
+  { 0x20 /* VKEY_SPACE */,                     kVK_Space               },
+  { 0x21 /* VKEY_PRIOR */,                     kVK_PageUp              },
+  { 0x22 /* VKEY_NEXT */,                      kVK_PageDown            },
+  { 0x23 /* VKEY_END */,                       kVK_End                 },
+  { 0x24 /* VKEY_HOME */,                      kVK_Home                },
+  { 0x25 /* VKEY_LEFT */,                      kVK_LeftArrow           },
+  { 0x26 /* VKEY_UP */,                        kVK_UpArrow             },
+  { 0x27 /* VKEY_RIGHT */,                     kVK_RightArrow          },
+  { 0x28 /* VKEY_DOWN */,                      kVK_DownArrow           },
+  { 0x29 /* VKEY_SELECT */,                    -1                      },
+  { 0x2A /* VKEY_PRINT */,                     -1                      },
+  { 0x2B /* VKEY_EXECUTE */,                   -1                      },
+  { 0x2C /* VKEY_SNAPSHOT */,                  -1                      },
+  { 0x2D /* VKEY_INSERT */,                    kVK_Help                },
+  { 0x2E /* VKEY_DELETE */,                    kVK_ForwardDelete       },
+  { 0x2F /* VKEY_HELP */,                      kVK_Help                },
+  { 0x30 /* VKEY_0 */,                         kVK_ANSI_0              },
+  { 0x31 /* VKEY_1 */,                         kVK_ANSI_1              },
+  { 0x32 /* VKEY_2 */,                         kVK_ANSI_2              },
+  { 0x33 /* VKEY_3 */,                         kVK_ANSI_3              },
+  { 0x34 /* VKEY_4 */,                         kVK_ANSI_4              },
+  { 0x35 /* VKEY_5 */,                         kVK_ANSI_5              },
+  { 0x36 /* VKEY_6 */,                         kVK_ANSI_6              },
+  { 0x37 /* VKEY_7 */,                         kVK_ANSI_7              },
+  { 0x38 /* VKEY_8 */,                         kVK_ANSI_8              },
+  { 0x39 /* VKEY_9 */,                         kVK_ANSI_9              },
+  { 0x41 /* VKEY_A */,                         kVK_ANSI_A              },
+  { 0x42 /* VKEY_B */,                         kVK_ANSI_B              },
+  { 0x43 /* VKEY_C */,                         kVK_ANSI_C              },
+  { 0x44 /* VKEY_D */,                         kVK_ANSI_D              },
+  { 0x45 /* VKEY_E */,                         kVK_ANSI_E              },
+  { 0x46 /* VKEY_F */,                         kVK_ANSI_F              },
+  { 0x47 /* VKEY_G */,                         kVK_ANSI_G              },
+  { 0x48 /* VKEY_H */,                         kVK_ANSI_H              },
+  { 0x49 /* VKEY_I */,                         kVK_ANSI_I              },
+  { 0x4A /* VKEY_J */,                         kVK_ANSI_J              },
+  { 0x4B /* VKEY_K */,                         kVK_ANSI_K              },
+  { 0x4C /* VKEY_L */,                         kVK_ANSI_L              },
+  { 0x4D /* VKEY_M */,                         kVK_ANSI_M              },
+  { 0x4E /* VKEY_N */,                         kVK_ANSI_N              },
+  { 0x4F /* VKEY_O */,                         kVK_ANSI_O              },
+  { 0x50 /* VKEY_P */,                         kVK_ANSI_P              },
+  { 0x51 /* VKEY_Q */,                         kVK_ANSI_Q              },
+  { 0x52 /* VKEY_R */,                         kVK_ANSI_R              },
+  { 0x53 /* VKEY_S */,                         kVK_ANSI_S              },
+  { 0x54 /* VKEY_T */,                         kVK_ANSI_T              },
+  { 0x55 /* VKEY_U */,                         kVK_ANSI_U              },
+  { 0x56 /* VKEY_V */,                         kVK_ANSI_V              },
+  { 0x57 /* VKEY_W */,                         kVK_ANSI_W              },
+  { 0x58 /* VKEY_X */,                         kVK_ANSI_X              },
+  { 0x59 /* VKEY_Y */,                         kVK_ANSI_Y              },
+  { 0x5A /* VKEY_Z */,                         kVK_ANSI_Z              },
+  { 0x5B /* VKEY_LWIN */,                      kVK_Command             },
+  { 0x5C /* VKEY_RWIN */,                      kVK_RightCommand        },
+  { 0x5D /* VKEY_APPS */,                      kVK_RightCommand        },
+  { 0x5F /* VKEY_SLEEP */,                     -1                      },
+  { 0x60 /* VKEY_NUMPAD0 */,                   kVK_ANSI_Keypad0        },
+  { 0x61 /* VKEY_NUMPAD1 */,                   kVK_ANSI_Keypad1        },
+  { 0x62 /* VKEY_NUMPAD2 */,                   kVK_ANSI_Keypad2        },
+  { 0x63 /* VKEY_NUMPAD3 */,                   kVK_ANSI_Keypad3        },
+  { 0x64 /* VKEY_NUMPAD4 */,                   kVK_ANSI_Keypad4        },
+  { 0x65 /* VKEY_NUMPAD5 */,                   kVK_ANSI_Keypad5        },
+  { 0x66 /* VKEY_NUMPAD6 */,                   kVK_ANSI_Keypad6        },
+  { 0x67 /* VKEY_NUMPAD7 */,                   kVK_ANSI_Keypad7        },
+  { 0x68 /* VKEY_NUMPAD8 */,                   kVK_ANSI_Keypad8        },
+  { 0x69 /* VKEY_NUMPAD9 */,                   kVK_ANSI_Keypad9        },
+  { 0x6A /* VKEY_MULTIPLY */,                  kVK_ANSI_KeypadMultiply },
+  { 0x6B /* VKEY_ADD */,                       kVK_ANSI_KeypadPlus     },
+  { 0x6C /* VKEY_SEPARATOR */,                 -1                      },
+  { 0x6D /* VKEY_SUBTRACT */,                  kVK_ANSI_KeypadMinus    },
+  { 0x6E /* VKEY_DECIMAL */,                   kVK_ANSI_KeypadDecimal  },
+  { 0x6F /* VKEY_DIVIDE */,                    kVK_ANSI_KeypadDivide   },
+  { 0x70 /* VKEY_F1 */,                        kVK_F1                  },
+  { 0x71 /* VKEY_F2 */,                        kVK_F2                  },
+  { 0x72 /* VKEY_F3 */,                        kVK_F3                  },
+  { 0x73 /* VKEY_F4 */,                        kVK_F4                  },
+  { 0x74 /* VKEY_F5 */,                        kVK_F5                  },
+  { 0x75 /* VKEY_F6 */,                        kVK_F6                  },
+  { 0x76 /* VKEY_F7 */,                        kVK_F7                  },
+  { 0x77 /* VKEY_F8 */,                        kVK_F8                  },
+  { 0x78 /* VKEY_F9 */,                        kVK_F9                  },
+  { 0x79 /* VKEY_F10 */,                       kVK_F10                 },
+  { 0x7A /* VKEY_F11 */,                       kVK_F11                 },
+  { 0x7B /* VKEY_F12 */,                       kVK_F12                 },
+  { 0x7C /* VKEY_F13 */,                       kVK_F13                 },
+  { 0x7D /* VKEY_F14 */,                       kVK_F14                 },
+  { 0x7E /* VKEY_F15 */,                       kVK_F15                 },
+  { 0x7F /* VKEY_F16 */,                       kVK_F16                 },
+  { 0x80 /* VKEY_F17 */,                       kVK_F17                 },
+  { 0x81 /* VKEY_F18 */,                       kVK_F18                 },
+  { 0x82 /* VKEY_F19 */,                       kVK_F19                 },
+  { 0x83 /* VKEY_F20 */,                       kVK_F20                 },
+  { 0x84 /* VKEY_F21 */,                       -1                      },
+  { 0x85 /* VKEY_F22 */,                       -1                      },
+  { 0x86 /* VKEY_F23 */,                       -1                      },
+  { 0x87 /* VKEY_F24 */,                       -1                      },
+  { 0x90 /* VKEY_NUMLOCK */,                   -1                      },
+  { 0x91 /* VKEY_SCROLL */,                    -1                      },
+  { 0xA0 /* VKEY_LSHIFT */,                    kVK_Shift               },
+  { 0xA1 /* VKEY_RSHIFT */,                    kVK_RightShift          },
+  { 0xA2 /* VKEY_LCONTROL */,                  kVK_Control             },
+  { 0xA3 /* VKEY_RCONTROL */,                  kVK_RightControl        },
+  { 0xA4 /* VKEY_LMENU */,                     kVK_Option              },
+  { 0xA5 /* VKEY_RMENU */,                     kVK_RightOption         },
+  { 0xA6 /* VKEY_BROWSER_BACK */,              -1                      },
+  { 0xA7 /* VKEY_BROWSER_FORWARD */,           -1                      },
+  { 0xA8 /* VKEY_BROWSER_REFRESH */,           -1                      },
+  { 0xA9 /* VKEY_BROWSER_STOP */,              -1                      },
+  { 0xAA /* VKEY_BROWSER_SEARCH */,            -1                      },
+  { 0xAB /* VKEY_BROWSER_FAVORITES */,         -1                      },
+  { 0xAC /* VKEY_BROWSER_HOME */,              -1                      },
+  { 0xAD /* VKEY_VOLUME_MUTE */,               -1                      },
+  { 0xAE /* VKEY_VOLUME_DOWN */,               -1                      },
+  { 0xAF /* VKEY_VOLUME_UP */,                 -1                      },
+  { 0xB0 /* VKEY_MEDIA_NEXT_TRACK */,          -1                      },
+  { 0xB1 /* VKEY_MEDIA_PREV_TRACK */,          -1                      },
+  { 0xB2 /* VKEY_MEDIA_STOP */,                -1                      },
+  { 0xB3 /* VKEY_MEDIA_PLAY_PAUSE */,          -1                      },
+  { 0xB4 /* VKEY_MEDIA_LAUNCH_MAIL */,         -1                      },
+  { 0xB5 /* VKEY_MEDIA_LAUNCH_MEDIA_SELECT */, -1                      },
+  { 0xB6 /* VKEY_MEDIA_LAUNCH_APP1 */,         -1                      },
+  { 0xB7 /* VKEY_MEDIA_LAUNCH_APP2 */,         -1                      },
+  { 0xBA /* VKEY_OEM_1 */,                     kVK_ANSI_Semicolon      },
+  { 0xBB /* VKEY_OEM_PLUS */,                  kVK_ANSI_Equal          },
+  { 0xBC /* VKEY_OEM_COMMA */,                 kVK_ANSI_Comma          },
+  { 0xBD /* VKEY_OEM_MINUS */,                 kVK_ANSI_Minus          },
+  { 0xBE /* VKEY_OEM_PERIOD */,                kVK_ANSI_Period         },
+  { 0xBF /* VKEY_OEM_2 */,                     kVK_ANSI_Slash          },
+  { 0xC0 /* VKEY_OEM_3 */,                     kVK_ANSI_Grave          },
+  { 0xDB /* VKEY_OEM_4 */,                     kVK_ANSI_LeftBracket    },
+  { 0xDC /* VKEY_OEM_5 */,                     kVK_ANSI_Backslash      },
+  { 0xDD /* VKEY_OEM_6 */,                     kVK_ANSI_RightBracket   },
+  { 0xDE /* VKEY_OEM_7 */,                     kVK_ANSI_Quote          },
+  { 0xDF /* VKEY_OEM_8 */,                     -1                      },
+  { 0xE2 /* VKEY_OEM_102 */,                   -1                      },
+  { 0xE5 /* VKEY_PROCESSKEY */,                -1                      },
+  { 0xE7 /* VKEY_PACKET */,                    -1                      },
+  { 0xF6 /* VKEY_ATTN */,                      -1                      },
+  { 0xF7 /* VKEY_CRSEL */,                     -1                      },
+  { 0xF8 /* VKEY_EXSEL */,                     -1                      },
+  { 0xF9 /* VKEY_EREOF */,                     -1                      },
+  { 0xFA /* VKEY_PLAY */,                      -1                      },
+  { 0xFB /* VKEY_ZOOM */,                      -1                      },
+  { 0xFC /* VKEY_NONAME */,                    -1                      },
+  { 0xFD /* VKEY_PA1 */,                       -1                      },
+  { 0xFE /* VKEY_OEM_CLEAR */,                 kVK_ANSI_KeypadClear    }
+};
+// clang-format on
+
+int keysym(int keycode) {
+  KeyCodeMap key_map;
+
+  key_map.win_keycode        = keycode;
+  const KeyCodeMap *temp_map = std::lower_bound(
+    kKeyCodesMap, kKeyCodesMap + sizeof(kKeyCodesMap) / sizeof(kKeyCodesMap[0]), key_map);
+
+  if(temp_map >= kKeyCodesMap + sizeof(kKeyCodesMap) / sizeof(kKeyCodesMap[0]) ||
+     temp_map->win_keycode != keycode || temp_map->mac_keycode == -1) {
+    return -1;
+  }
+
+  return temp_map->mac_keycode;
+}
+
+void keyboard(input_t &input, uint16_t modcode, bool release) {
+  auto key = keysym(modcode);
+
+  BOOST_LOG(debug) << "got keycode: 0x"sv << std::hex << modcode << ", translated to: 0x" << std::hex << key << ", release:" << release;
+
+  if(key < 0) {
+    return;
+  }
+
+  auto macos_input = ((macos_input_t *)input.get());
+  auto event       = macos_input->kb_event;
+
+  if(key == kVK_Shift || key == kVK_RightShift ||
+     key == kVK_Command || key == kVK_RightCommand ||
+     key == kVK_Option || key == kVK_RightOption ||
+     key == kVK_Control || key == kVK_RightControl) {
+
+    CGEventFlags mask;
+
+    switch(key) {
+    case kVK_Shift:
+    case kVK_RightShift:
+      mask = kCGEventFlagMaskShift;
+      break;
+    case kVK_Command:
+    case kVK_RightCommand:
+      mask = kCGEventFlagMaskCommand;
+      break;
+    case kVK_Option:
+    case kVK_RightOption:
+      mask = kCGEventFlagMaskAlternate;
+      break;
+    case kVK_Control:
+    case kVK_RightControl:
+      mask = kCGEventFlagMaskControl;
+      break;
+    }
+
+    macos_input->kb_flags = release ? macos_input->kb_flags & ~mask : macos_input->kb_flags | mask;
+    CGEventSetType(event, kCGEventFlagsChanged);
+    CGEventSetFlags(event, macos_input->kb_flags);
+  }
+  else {
+    CGEventSetIntegerValueField(event, kCGKeyboardEventKeycode, key);
+    CGEventSetType(event, release ? kCGEventKeyUp : kCGEventKeyDown);
+  }
+
+  CGEventPost(kCGHIDEventTap, event);
+}
+
+int alloc_gamepad(input_t &input, int nr, rumble_queue_t rumble_queue) {
+  BOOST_LOG(info) << "alloc_gamepad: Gamepad not yet implemented for MacOS."sv;
+  return -1;
+}
+
+void free_gamepad(input_t &input, int nr) {
+  BOOST_LOG(info) << "free_gamepad: Gamepad not yet implemented for MacOS."sv;
+}
+
+void gamepad(input_t &input, int nr, const gamepad_state_t &gamepad_state) {
+  BOOST_LOG(info) << "gamepad: Gamepad not yet implemented for MacOS."sv;
+}
+
+// returns current mouse location:
+inline CGPoint get_mouse_loc(input_t &input) {
+  return CGEventGetLocation(((macos_input_t *)input.get())->mouse_event);
+}
+
+void post_mouse(input_t &input, CGMouseButton button, CGEventType type, CGPoint location, int click_count) {
+  BOOST_LOG(debug) << "mouse_event: "sv << button << ", type: "sv << type << ", location:"sv << location.x << ":"sv << location.y << " click_count: "sv << click_count;
+
+  auto macos_input = (macos_input_t *)input.get();
+  auto display     = macos_input->display;
+  auto event       = macos_input->mouse_event;
+
+  if(location.x < 0)
+    location.x = 0;
+  if(location.x >= CGDisplayPixelsWide(display))
+    location.x = CGDisplayPixelsWide(display) - 1;
+
+  if(location.y < 0)
+    location.y = 0;
+  if(location.y >= CGDisplayPixelsHigh(display))
+    location.y = CGDisplayPixelsHigh(display) - 1;
+
+  CGEventSetType(event, type);
+  CGEventSetLocation(event, location);
+  CGEventSetIntegerValueField(event, kCGMouseEventButtonNumber, button);
+  CGEventSetIntegerValueField(event, kCGMouseEventClickState, click_count);
+
+  CGEventPost(kCGHIDEventTap, event);
+
+  // For why this is here, see:
+  // https://stackoverflow.com/questions/15194409/simulated-mouseevent-not-working-properly-osx
+  CGWarpMouseCursorPosition(location);
+}
+
+inline CGEventType event_type_mouse(input_t &input) {
+  auto macos_input = ((macos_input_t *)input.get());
+
+  if(macos_input->mouse_down[0]) {
+    return kCGEventLeftMouseDragged;
+  }
+  else if(macos_input->mouse_down[1]) {
+    return kCGEventOtherMouseDragged;
+  }
+  else if(macos_input->mouse_down[2]) {
+    return kCGEventRightMouseDragged;
+  }
+  else {
+    return kCGEventMouseMoved;
+  }
+}
+
+void move_mouse(input_t &input, int deltaX, int deltaY) {
+  auto current = get_mouse_loc(input);
+
+  CGPoint location = CGPointMake(current.x + deltaX, current.y + deltaY);
+
+  post_mouse(input, kCGMouseButtonLeft, event_type_mouse(input), location, 0);
+}
+
+void abs_mouse(input_t &input, const touch_port_t &touch_port, float x, float y) {
+  auto scaling = ((macos_input_t *)input.get())->displayScaling;
+
+  CGPoint location = CGPointMake(x * scaling, y * scaling);
+
+  post_mouse(input, kCGMouseButtonLeft, event_type_mouse(input), location, 0);
+}
+
+uint64_t time_diff(uint64_t start) {
+  uint64_t elapsed;
+  Nanoseconds elapsedNano;
+
+  elapsed     = mach_absolute_time() - start;
+  elapsedNano = AbsoluteToNanoseconds(*(AbsoluteTime *)&elapsed);
+
+  return *(uint64_t *)&elapsedNano;
+}
+
+void button_mouse(input_t &input, int button, bool release) {
+  CGMouseButton mac_button;
+  CGEventType event;
+
+  auto mouse = ((macos_input_t *)input.get());
+
+  switch(button) {
+  case 1:
+    mac_button = kCGMouseButtonLeft;
+    event      = release ? kCGEventLeftMouseUp : kCGEventLeftMouseDown;
+    break;
+  case 2:
+    mac_button = kCGMouseButtonCenter;
+    event      = release ? kCGEventOtherMouseUp : kCGEventOtherMouseDown;
+    break;
+  case 3:
+    mac_button = kCGMouseButtonRight;
+    event      = release ? kCGEventRightMouseUp : kCGEventRightMouseDown;
+    break;
+  default:
+    BOOST_LOG(warning) << "Unsupported mouse button for MacOS: "sv << button;
+    return;
+  }
+
+  mouse->mouse_down[mac_button] = !release;
+
+  // if the last mouse down was less than MULTICLICK_DELAY_NS, we send a double click event
+  if(time_diff(mouse->last_mouse_event[mac_button][release]) < MULTICLICK_DELAY_NS) {
+    post_mouse(input, mac_button, event, get_mouse_loc(input), 2);
+  }
+  else {
+    post_mouse(input, mac_button, event, get_mouse_loc(input), 1);
+  }
+
+  mouse->last_mouse_event[mac_button][release] = mach_absolute_time();
+}
+
+void scroll(input_t &input, int high_res_distance) {
+  CGEventRef upEvent = CGEventCreateScrollWheelEvent(
+    NULL,
+    kCGScrollEventUnitLine,
+    2, high_res_distance > 0 ? 1 : -1, high_res_distance);
+  CGEventPost(kCGHIDEventTap, upEvent);
+  CFRelease(upEvent);
+}
+
+input_t input() {
+  input_t result { new macos_input_t() };
+
+  auto macos_input = (macos_input_t *)result.get();
+
+  // If we don't use the main display in the future, this has to be adapted
+  macos_input->display = CGMainDisplayID();
+
+  // Input coordinates are based on the virtual resolution not the physical, so we need the scaling factor
+  CGDisplayModeRef mode       = CGDisplayCopyDisplayMode(macos_input->display);
+  macos_input->displayScaling = ((CGFloat)CGDisplayPixelsWide(macos_input->display)) / ((CGFloat)CGDisplayModeGetPixelWidth(mode));
+  CFRelease(mode);
+
+  macos_input->source = CGEventSourceCreate(kCGEventSourceStateHIDSystemState);
+
+  macos_input->kb_event = CGEventCreate(macos_input->source);
+  macos_input->kb_flags = 0;
+
+  macos_input->mouse_event            = CGEventCreate(macos_input->source);
+  macos_input->mouse_down[0]          = false;
+  macos_input->mouse_down[1]          = false;
+  macos_input->mouse_down[2]          = false;
+  macos_input->last_mouse_event[0][0] = 0;
+  macos_input->last_mouse_event[0][1] = 0;
+  macos_input->last_mouse_event[1][0] = 0;
+  macos_input->last_mouse_event[1][1] = 0;
+  macos_input->last_mouse_event[2][0] = 0;
+  macos_input->last_mouse_event[2][1] = 0;
+
+  BOOST_LOG(debug) << "Display "sv << macos_input->display << ", pixel dimention: " << CGDisplayPixelsWide(macos_input->display) << "x"sv << CGDisplayPixelsHigh(macos_input->display);
+
+  return result;
+}
+
+void freeInput(void *p) {
+  auto *input = (macos_input_t *)p;
+
+  CFRelease(input->source);
+  CFRelease(input->kb_event);
+  CFRelease(input->mouse_event);
+
+  delete input;
+}
+
+std::vector<std::string_view> &supported_gamepads() {
+  static std::vector<std::string_view> gamepads { ""sv };
+
+  return gamepads;
+}
+} // namespace platf

src/platform/macos/microphone.mm

@@ -0,0 +1,87 @@
+#include "src/platform/common.h"
+#include "src/platform/macos/av_audio.h"
+
+#include "src/config.h"
+#include "src/main.h"
+
+namespace platf {
+using namespace std::literals;
+
+struct av_mic_t : public mic_t {
+  AVAudio *av_audio_capture;
+
+  ~av_mic_t() {
+    [av_audio_capture release];
+  }
+
+  capture_e sample(std::vector<std::int16_t> &sample_in) override {
+    auto sample_size = sample_in.size();
+
+    uint32_t length        = 0;
+    void *byteSampleBuffer = TPCircularBufferTail(&av_audio_capture->audioSampleBuffer, &length);
+
+    while(length < sample_size * sizeof(std::int16_t)) {
+      [av_audio_capture.samplesArrivedSignal wait];
+      byteSampleBuffer = TPCircularBufferTail(&av_audio_capture->audioSampleBuffer, &length);
+    }
+
+    const int16_t *sampleBuffer = (int16_t *)byteSampleBuffer;
+    std::vector<int16_t> vectorBuffer(sampleBuffer, sampleBuffer + sample_size);
+
+    std::copy_n(std::begin(vectorBuffer), sample_size, std::begin(sample_in));
+
+    TPCircularBufferConsume(&av_audio_capture->audioSampleBuffer, sample_size * sizeof(std::int16_t));
+
+    return capture_e::ok;
+  }
+};
+
+struct macos_audio_control_t : public audio_control_t {
+  AVCaptureDevice *audio_capture_device;
+
+public:
+  int set_sink(const std::string &sink) override {
+    BOOST_LOG(warning) << "audio_control_t::set_sink() unimplemented: "sv << sink;
+    return 0;
+  }
+
+  std::unique_ptr<mic_t> microphone(const std::uint8_t *mapping, int channels, std::uint32_t sample_rate, std::uint32_t frame_size) override {
+    auto mic               = std::make_unique<av_mic_t>();
+    const char *audio_sink = "";
+
+    if(!config::audio.sink.empty()) {
+      audio_sink = config::audio.sink.c_str();
+    }
+
+    if((audio_capture_device = [AVAudio findMicrophone:[NSString stringWithUTF8String:audio_sink]]) == nullptr) {
+      BOOST_LOG(error) << "opening microphone '"sv << audio_sink << "' failed. Please set a valid input source in the Sunshine config."sv;
+      BOOST_LOG(error) << "Available inputs:"sv;
+
+      for(NSString *name in [AVAudio microphoneNames]) {
+        BOOST_LOG(error) << "\t"sv << [name UTF8String];
+      }
+
+      return nullptr;
+    }
+
+    mic->av_audio_capture = [[AVAudio alloc] init];
+
+    if([mic->av_audio_capture setupMicrophone:audio_capture_device sampleRate:sample_rate frameSize:frame_size channels:channels]) {
+      BOOST_LOG(error) << "Failed to setup microphone."sv;
+      return nullptr;
+    }
+
+    return mic;
+  }
+
+  std::optional<sink_t> sink_info() override {
+    sink_t sink;
+
+    return sink;
+  }
+};
+
+std::unique_ptr<audio_control_t> audio_control() {
+  return std::make_unique<macos_audio_control_t>();
+}
+}

src/platform/macos/misc.cpp

@@ -0,0 +1,161 @@
+#include <arpa/inet.h>
+#include <dlfcn.h>
+#include <fcntl.h>
+#include <ifaddrs.h>
+#include <net/if_dl.h>
+#include <pwd.h>
+
+#include "misc.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+
+using namespace std::literals;
+namespace fs = std::filesystem;
+
+namespace platf {
+std::unique_ptr<deinit_t> init() {
+  if(!CGPreflightScreenCaptureAccess()) {
+    BOOST_LOG(error) << "No screen capture permission!"sv;
+    BOOST_LOG(error) << "Please activate it in 'System Preferences' -> 'Privacy' -> 'Screen Recording'"sv;
+    CGRequestScreenCaptureAccess();
+    return nullptr;
+  }
+  return std::make_unique<deinit_t>();
+}
+
+fs::path appdata() {
+  const char *homedir;
+  if((homedir = getenv("HOME")) == nullptr) {
+    homedir = getpwuid(geteuid())->pw_dir;
+  }
+
+  return fs::path { homedir } / ".config/sunshine"sv;
+}
+
+using ifaddr_t = util::safe_ptr<ifaddrs, freeifaddrs>;
+
+ifaddr_t get_ifaddrs() {
+  ifaddrs *p { nullptr };
+
+  getifaddrs(&p);
+
+  return ifaddr_t { p };
+}
+
+std::string from_sockaddr(const sockaddr *const ip_addr) {
+  char data[INET6_ADDRSTRLEN];
+
+  auto family = ip_addr->sa_family;
+  if(family == AF_INET6) {
+    inet_ntop(AF_INET6, &((sockaddr_in6 *)ip_addr)->sin6_addr, data,
+      INET6_ADDRSTRLEN);
+  }
+
+  if(family == AF_INET) {
+    inet_ntop(AF_INET, &((sockaddr_in *)ip_addr)->sin_addr, data,
+      INET_ADDRSTRLEN);
+  }
+
+  return std::string { data };
+}
+
+std::pair<std::uint16_t, std::string> from_sockaddr_ex(const sockaddr *const ip_addr) {
+  char data[INET6_ADDRSTRLEN];
+
+  auto family = ip_addr->sa_family;
+  std::uint16_t port;
+  if(family == AF_INET6) {
+    inet_ntop(AF_INET6, &((sockaddr_in6 *)ip_addr)->sin6_addr, data,
+      INET6_ADDRSTRLEN);
+    port = ((sockaddr_in6 *)ip_addr)->sin6_port;
+  }
+
+  if(family == AF_INET) {
+    inet_ntop(AF_INET, &((sockaddr_in *)ip_addr)->sin_addr, data,
+      INET_ADDRSTRLEN);
+    port = ((sockaddr_in *)ip_addr)->sin_port;
+  }
+
+  return { port, std::string { data } };
+}
+
+std::string get_mac_address(const std::string_view &address) {
+  auto ifaddrs = get_ifaddrs();
+
+  for(auto pos = ifaddrs.get(); pos != nullptr; pos = pos->ifa_next) {
+    if(pos->ifa_addr && address == from_sockaddr(pos->ifa_addr)) {
+      BOOST_LOG(verbose) << "Looking for MAC of "sv << pos->ifa_name;
+
+      struct ifaddrs *ifap, *ifaptr;
+      unsigned char *ptr;
+      std::string mac_address;
+
+      if(getifaddrs(&ifap) == 0) {
+        for(ifaptr = ifap; ifaptr != NULL; ifaptr = (ifaptr)->ifa_next) {
+          if(!strcmp((ifaptr)->ifa_name, pos->ifa_name) && (((ifaptr)->ifa_addr)->sa_family == AF_LINK)) {
+            ptr = (unsigned char *)LLADDR((struct sockaddr_dl *)(ifaptr)->ifa_addr);
+            char buff[100];
+
+            snprintf(buff, sizeof(buff), "%02x:%02x:%02x:%02x:%02x:%02x",
+              *ptr, *(ptr + 1), *(ptr + 2), *(ptr + 3), *(ptr + 4), *(ptr + 5));
+            mac_address = buff;
+            break;
+          }
+        }
+
+        freeifaddrs(ifap);
+
+        if(ifaptr != NULL) {
+          BOOST_LOG(verbose) << "Found MAC of "sv << pos->ifa_name << ": "sv << mac_address;
+          return mac_address;
+        }
+      }
+    }
+  }
+
+  BOOST_LOG(warning) << "Unable to find MAC address for "sv << address;
+  return "00:00:00:00:00:00"s;
+}
+} // namespace platf
+
+namespace dyn {
+void *handle(const std::vector<const char *> &libs) {
+  void *handle;
+
+  for(auto lib : libs) {
+    handle = dlopen(lib, RTLD_LAZY | RTLD_LOCAL);
+    if(handle) {
+      return handle;
+    }
+  }
+
+  std::stringstream ss;
+  ss << "Couldn't find any of the following libraries: ["sv << libs.front();
+  std::for_each(std::begin(libs) + 1, std::end(libs), [&](auto lib) {
+    ss << ", "sv << lib;
+  });
+
+  ss << ']';
+
+  BOOST_LOG(error) << ss.str();
+
+  return nullptr;
+}
+
+int load(void *handle, const std::vector<std::tuple<apiproc *, const char *>> &funcs, bool strict) {
+  int err = 0;
+  for(auto &func : funcs) {
+    TUPLE_2D_REF(fn, name, func);
+
+    *fn = (void (*)())dlsym(handle, name);
+
+    if(!*fn && strict) {
+      BOOST_LOG(error) << "Couldn't find function: "sv << name;
+
+      err = -1;
+    }
+  }
+
+  return err;
+}
+} // namespace dyn

src/platform/macos/misc.h

@@ -0,0 +1,16 @@
+#ifndef SUNSHINE_PLATFORM_MISC_H
+#define SUNSHINE_PLATFORM_MISC_H
+
+#include <vector>
+
+#include <CoreGraphics/CoreGraphics.h>
+
+namespace dyn {
+typedef void (*apiproc)(void);
+
+int load(void *handle, const std::vector<std::tuple<apiproc *, const char *>> &funcs, bool strict = true);
+void *handle(const std::vector<const char *> &libs);
+
+} // namespace dyn
+
+#endif

src/platform/macos/nv12_zero_device.cpp

@@ -0,0 +1,82 @@
+#include "src/platform/macos/nv12_zero_device.h"
+#include "src/platform/macos/av_img_t.h"
+
+#include "src/video.h"
+
+extern "C" {
+#include "libavutil/imgutils.h"
+}
+
+namespace platf {
+
+void free_frame(AVFrame *frame) {
+  av_frame_free(&frame);
+}
+
+util::safe_ptr<AVFrame, free_frame> av_frame;
+
+int nv12_zero_device::convert(platf::img_t &img) {
+  av_frame_make_writable(av_frame.get());
+
+  av_img_t *av_img = (av_img_t *)&img;
+
+  size_t left_pad, right_pad, top_pad, bottom_pad;
+  CVPixelBufferGetExtendedPixels(av_img->pixel_buffer, &left_pad, &right_pad, &top_pad, &bottom_pad);
+
+  const uint8_t *data = (const uint8_t *)CVPixelBufferGetBaseAddressOfPlane(av_img->pixel_buffer, 0) - left_pad - (top_pad * img.width);
+
+  int result = av_image_fill_arrays(av_frame->data, av_frame->linesize, data, (AVPixelFormat)av_frame->format, img.width, img.height, 32);
+
+  // We will create the black bars for the padding top/bottom or left/right here in very cheap way.
+  // The luminance is 0, therefore, we simply need to set the chroma values to 128 for each pixel
+  // for black bars (instead of green with chroma 0). However, this only works 100% correct, when
+  // the resolution is devisable by 32. This could be improved by calculating the chroma values for
+  // the outer content pixels, which should introduce only a minor performance hit.
+  //
+  // XXX: Improve the algorithm to take into account the outer pixels
+
+  size_t uv_plane_height = CVPixelBufferGetHeightOfPlane(av_img->pixel_buffer, 1);
+
+  if(left_pad || right_pad) {
+    for(int l = 0; l < uv_plane_height + (top_pad / 2); l++) {
+      int line = l * av_frame->linesize[1];
+      memset((void *)&av_frame->data[1][line], 128, (size_t)left_pad);
+      memset((void *)&av_frame->data[1][line + img.width - right_pad], 128, right_pad);
+    }
+  }
+
+  if(top_pad || bottom_pad) {
+    memset((void *)&av_frame->data[1][0], 128, (top_pad / 2) * av_frame->linesize[1]);
+    memset((void *)&av_frame->data[1][((top_pad / 2) + uv_plane_height) * av_frame->linesize[1]], 128, bottom_pad / 2 * av_frame->linesize[1]);
+  }
+
+  return result > 0 ? 0 : -1;
+}
+
+int nv12_zero_device::set_frame(AVFrame *frame) {
+  this->frame = frame;
+
+  av_frame.reset(frame);
+
+  resolution_fn(this->display, frame->width, frame->height);
+
+  return 0;
+}
+
+void nv12_zero_device::set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) {
+}
+
+int nv12_zero_device::init(void *display, resolution_fn_t resolution_fn, pixel_format_fn_t pixel_format_fn) {
+  pixel_format_fn(display, '420v');
+
+  this->display       = display;
+  this->resolution_fn = resolution_fn;
+
+  // we never use this pointer but it's existence is checked/used
+  // by the platform independed code
+  data = this;
+
+  return 0;
+}
+
+} // namespace platf

src/platform/macos/nv12_zero_device.h

@@ -0,0 +1,29 @@
+#ifndef vtdevice_h
+#define vtdevice_h
+
+#include "src/platform/common.h"
+
+namespace platf {
+
+class nv12_zero_device : public hwdevice_t {
+  // display holds a pointer to an av_video object. Since the namespaces of AVFoundation
+  // and FFMPEG collide, we need this opaque pointer and cannot use the definition
+  void *display;
+
+public:
+  // this function is used to set the resolution on an av_video object that we cannot
+  // call directly because of namespace collisions between AVFoundation and FFMPEG
+  using resolution_fn_t = std::function<void(void *display, int width, int height)>;
+  resolution_fn_t resolution_fn;
+  using pixel_format_fn_t = std::function<void(void *display, int pixelFormat)>;
+
+  int init(void *display, resolution_fn_t resolution_fn, pixel_format_fn_t pixel_format_fn);
+
+  int convert(img_t &img);
+  int set_frame(AVFrame *frame);
+  void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range);
+};
+
+} // namespace platf
+
+#endif /* vtdevice_h */

src/platform/macos/publish.cpp

@@ -0,0 +1,429 @@
+
+// adapted from https://www.avahi.org/doxygen/html/client-publish-service_8c-example.html
+#include <thread>
+
+#include "misc.h"
+#include "src/main.h"
+#include "src/nvhttp.h"
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+using namespace std::literals;
+
+namespace avahi {
+
+/** Error codes used by avahi */
+enum err_e {
+  OK                      = 0,  /**< OK */
+  ERR_FAILURE             = -1, /**< Generic error code */
+  ERR_BAD_STATE           = -2, /**< Object was in a bad state */
+  ERR_INVALID_HOST_NAME   = -3, /**< Invalid host name */
+  ERR_INVALID_DOMAIN_NAME = -4, /**< Invalid domain name */
+  ERR_NO_NETWORK          = -5, /**< No suitable network protocol available */
+  ERR_INVALID_TTL         = -6, /**< Invalid DNS TTL */
+  ERR_IS_PATTERN          = -7, /**< RR key is pattern */
+  ERR_COLLISION           = -8, /**< Name collision */
+  ERR_INVALID_RECORD      = -9, /**< Invalid RR */
+
+  ERR_INVALID_SERVICE_NAME = -10, /**< Invalid service name */
+  ERR_INVALID_SERVICE_TYPE = -11, /**< Invalid service type */
+  ERR_INVALID_PORT         = -12, /**< Invalid port number */
+  ERR_INVALID_KEY          = -13, /**< Invalid key */
+  ERR_INVALID_ADDRESS      = -14, /**< Invalid address */
+  ERR_TIMEOUT              = -15, /**< Timeout reached */
+  ERR_TOO_MANY_CLIENTS     = -16, /**< Too many clients */
+  ERR_TOO_MANY_OBJECTS     = -17, /**< Too many objects */
+  ERR_TOO_MANY_ENTRIES     = -18, /**< Too many entries */
+  ERR_OS                   = -19, /**< OS error */
+
+  ERR_ACCESS_DENIED     = -20, /**< Access denied */
+  ERR_INVALID_OPERATION = -21, /**< Invalid operation */
+  ERR_DBUS_ERROR        = -22, /**< An unexpected D-Bus error occurred */
+  ERR_DISCONNECTED      = -23, /**< Daemon connection failed */
+  ERR_NO_MEMORY         = -24, /**< Memory exhausted */
+  ERR_INVALID_OBJECT    = -25, /**< The object passed to this function was invalid */
+  ERR_NO_DAEMON         = -26, /**< Daemon not running */
+  ERR_INVALID_INTERFACE = -27, /**< Invalid interface */
+  ERR_INVALID_PROTOCOL  = -28, /**< Invalid protocol */
+  ERR_INVALID_FLAGS     = -29, /**< Invalid flags */
+
+  ERR_NOT_FOUND               = -30, /**< Not found */
+  ERR_INVALID_CONFIG          = -31, /**< Configuration error */
+  ERR_VERSION_MISMATCH        = -32, /**< Verson mismatch */
+  ERR_INVALID_SERVICE_SUBTYPE = -33, /**< Invalid service subtype */
+  ERR_INVALID_PACKET          = -34, /**< Invalid packet */
+  ERR_INVALID_DNS_ERROR       = -35, /**< Invlaid DNS return code */
+  ERR_DNS_FORMERR             = -36, /**< DNS Error: Form error */
+  ERR_DNS_SERVFAIL            = -37, /**< DNS Error: Server Failure */
+  ERR_DNS_NXDOMAIN            = -38, /**< DNS Error: No such domain */
+  ERR_DNS_NOTIMP              = -39, /**< DNS Error: Not implemented */
+
+  ERR_DNS_REFUSED       = -40, /**< DNS Error: Operation refused */
+  ERR_DNS_YXDOMAIN      = -41,
+  ERR_DNS_YXRRSET       = -42,
+  ERR_DNS_NXRRSET       = -43,
+  ERR_DNS_NOTAUTH       = -44, /**< DNS Error: Not authorized */
+  ERR_DNS_NOTZONE       = -45,
+  ERR_INVALID_RDATA     = -46, /**< Invalid RDATA */
+  ERR_INVALID_DNS_CLASS = -47, /**< Invalid DNS class */
+  ERR_INVALID_DNS_TYPE  = -48, /**< Invalid DNS type */
+  ERR_NOT_SUPPORTED     = -49, /**< Not supported */
+
+  ERR_NOT_PERMITTED    = -50, /**< Operation not permitted */
+  ERR_INVALID_ARGUMENT = -51, /**< Invalid argument */
+  ERR_IS_EMPTY         = -52, /**< Is empty */
+  ERR_NO_CHANGE        = -53, /**< The requested operation is invalid because it is redundant */
+
+  ERR_MAX = -54
+};
+
+constexpr auto IF_UNSPEC = -1;
+enum proto {
+  PROTO_INET   = 0, /**< IPv4 */
+  PROTO_INET6  = 1, /**< IPv6 */
+  PROTO_UNSPEC = -1 /**< Unspecified/all protocol(s) */
+};
+
+enum ServerState {
+  SERVER_INVALID,     /**< Invalid state (initial) */
+  SERVER_REGISTERING, /**< Host RRs are being registered */
+  SERVER_RUNNING,     /**< All host RRs have been established */
+  SERVER_COLLISION,   /**< There is a collision with a host RR. All host RRs have been withdrawn, the user should set a new host name via avahi_server_set_host_name() */
+  SERVER_FAILURE      /**< Some fatal failure happened, the server is unable to proceed */
+};
+
+enum ClientState {
+  CLIENT_S_REGISTERING = SERVER_REGISTERING, /**< Server state: REGISTERING */
+  CLIENT_S_RUNNING     = SERVER_RUNNING,     /**< Server state: RUNNING */
+  CLIENT_S_COLLISION   = SERVER_COLLISION,   /**< Server state: COLLISION */
+  CLIENT_FAILURE       = 100,                /**< Some kind of error happened on the client side */
+  CLIENT_CONNECTING    = 101                 /**< We're still connecting. This state is only entered when AVAHI_CLIENT_NO_FAIL has been passed to avahi_client_new() and the daemon is not yet available. */
+};
+
+enum EntryGroupState {
+  ENTRY_GROUP_UNCOMMITED,  /**< The group has not yet been commited, the user must still call avahi_entry_group_commit() */
+  ENTRY_GROUP_REGISTERING, /**< The entries of the group are currently being registered */
+  ENTRY_GROUP_ESTABLISHED, /**< The entries have successfully been established */
+  ENTRY_GROUP_COLLISION,   /**< A name collision for one of the entries in the group has been detected, the entries have been withdrawn */
+  ENTRY_GROUP_FAILURE      /**< Some kind of failure happened, the entries have been withdrawn */
+};
+
+enum ClientFlags {
+  CLIENT_IGNORE_USER_CONFIG = 1, /**< Don't read user configuration */
+  CLIENT_NO_FAIL            = 2  /**< Don't fail if the daemon is not available when avahi_client_new() is called, instead enter CLIENT_CONNECTING state and wait for the daemon to appear */
+};
+
+/** Some flags for publishing functions */
+enum PublishFlags {
+  PUBLISH_UNIQUE         = 1,  /**< For raw records: The RRset is intended to be unique */
+  PUBLISH_NO_PROBE       = 2,  /**< For raw records: Though the RRset is intended to be unique no probes shall be sent */
+  PUBLISH_NO_ANNOUNCE    = 4,  /**< For raw records: Do not announce this RR to other hosts */
+  PUBLISH_ALLOW_MULTIPLE = 8,  /**< For raw records: Allow multiple local records of this type, even if they are intended to be unique */
+                               /** \cond fulldocs */
+  PUBLISH_NO_REVERSE = 16,     /**< For address records: don't create a reverse (PTR) entry */
+  PUBLISH_NO_COOKIE  = 32,     /**< For service records: do not implicitly add the local service cookie to TXT data */
+                               /** \endcond */
+  PUBLISH_UPDATE = 64,         /**< Update existing records instead of adding new ones */
+                               /** \cond fulldocs */
+  PUBLISH_USE_WIDE_AREA = 128, /**< Register the record using wide area DNS (i.e. unicast DNS update) */
+  PUBLISH_USE_MULTICAST = 256  /**< Register the record using multicast DNS */
+  /** \endcond */
+};
+
+using IfIndex  = int;
+using Protocol = int;
+
+struct EntryGroup;
+struct Poll;
+struct SimplePoll;
+struct Client;
+
+typedef void (*ClientCallback)(Client *, ClientState, void *userdata);
+typedef void (*EntryGroupCallback)(EntryGroup *g, EntryGroupState state, void *userdata);
+
+typedef void (*free_fn)(void *);
+
+typedef Client *(*client_new_fn)(const Poll *poll_api, ClientFlags flags, ClientCallback callback, void *userdata, int *error);
+typedef void (*client_free_fn)(Client *);
+typedef char *(*alternative_service_name_fn)(char *);
+
+typedef Client *(*entry_group_get_client_fn)(EntryGroup *);
+
+typedef EntryGroup *(*entry_group_new_fn)(Client *, EntryGroupCallback, void *userdata);
+typedef int (*entry_group_add_service_fn)(
+  EntryGroup *group,
+  IfIndex interface,
+  Protocol protocol,
+  PublishFlags flags,
+  const char *name,
+  const char *type,
+  const char *domain,
+  const char *host,
+  uint16_t port,
+  ...);
+
+typedef int (*entry_group_is_empty_fn)(EntryGroup *);
+typedef int (*entry_group_reset_fn)(EntryGroup *);
+typedef int (*entry_group_commit_fn)(EntryGroup *);
+
+typedef char *(*strdup_fn)(const char *);
+typedef char *(*strerror_fn)(int);
+typedef int (*client_errno_fn)(Client *);
+
+typedef Poll *(*simple_poll_get_fn)(SimplePoll *);
+typedef int (*simple_poll_loop_fn)(SimplePoll *);
+typedef void (*simple_poll_quit_fn)(SimplePoll *);
+typedef SimplePoll *(*simple_poll_new_fn)();
+typedef void (*simple_poll_free_fn)(SimplePoll *);
+
+free_fn free;
+client_new_fn client_new;
+client_free_fn client_free;
+alternative_service_name_fn alternative_service_name;
+entry_group_get_client_fn entry_group_get_client;
+entry_group_new_fn entry_group_new;
+entry_group_add_service_fn entry_group_add_service;
+entry_group_is_empty_fn entry_group_is_empty;
+entry_group_reset_fn entry_group_reset;
+entry_group_commit_fn entry_group_commit;
+strdup_fn strdup;
+strerror_fn strerror;
+client_errno_fn client_errno;
+simple_poll_get_fn simple_poll_get;
+simple_poll_loop_fn simple_poll_loop;
+simple_poll_quit_fn simple_poll_quit;
+simple_poll_new_fn simple_poll_new;
+simple_poll_free_fn simple_poll_free;
+
+
+int init_common() {
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libavahi-common.3.dylib", "libavahi-common.dylib" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&alternative_service_name, "avahi_alternative_service_name" },
+    { (dyn::apiproc *)&free, "avahi_free" },
+    { (dyn::apiproc *)&strdup, "avahi_strdup" },
+    { (dyn::apiproc *)&strerror, "avahi_strerror" },
+    { (dyn::apiproc *)&simple_poll_get, "avahi_simple_poll_get" },
+    { (dyn::apiproc *)&simple_poll_loop, "avahi_simple_poll_loop" },
+    { (dyn::apiproc *)&simple_poll_quit, "avahi_simple_poll_quit" },
+    { (dyn::apiproc *)&simple_poll_new, "avahi_simple_poll_new" },
+    { (dyn::apiproc *)&simple_poll_free, "avahi_simple_poll_free" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+
+int init_client() {
+  if(init_common()) {
+    return -1;
+  }
+
+  static void *handle { nullptr };
+  static bool funcs_loaded = false;
+
+  if(funcs_loaded) return 0;
+
+  if(!handle) {
+    handle = dyn::handle({ "libavahi-client.3.dylib", "libavahi-client.dylib" });
+    if(!handle) {
+      return -1;
+    }
+  }
+
+  std::vector<std::tuple<dyn::apiproc *, const char *>> funcs {
+    { (dyn::apiproc *)&client_new, "avahi_client_new" },
+    { (dyn::apiproc *)&client_free, "avahi_client_free" },
+    { (dyn::apiproc *)&entry_group_get_client, "avahi_entry_group_get_client" },
+    { (dyn::apiproc *)&entry_group_new, "avahi_entry_group_new" },
+    { (dyn::apiproc *)&entry_group_add_service, "avahi_entry_group_add_service" },
+    { (dyn::apiproc *)&entry_group_is_empty, "avahi_entry_group_is_empty" },
+    { (dyn::apiproc *)&entry_group_reset, "avahi_entry_group_reset" },
+    { (dyn::apiproc *)&entry_group_commit, "avahi_entry_group_commit" },
+    { (dyn::apiproc *)&client_errno, "avahi_client_errno" },
+  };
+
+  if(dyn::load(handle, funcs)) {
+    return -1;
+  }
+
+  funcs_loaded = true;
+  return 0;
+}
+} // namespace avahi
+
+namespace platf::publish {
+
+template<class T>
+void free(T *p) {
+  avahi::free(p);
+}
+
+template<class T>
+using ptr_t    = util::safe_ptr<T, free<T>>;
+using client_t = util::dyn_safe_ptr<avahi::Client, &avahi::client_free>;
+using poll_t   = util::dyn_safe_ptr<avahi::SimplePoll, &avahi::simple_poll_free>;
+
+avahi::EntryGroup *group = nullptr;
+
+poll_t poll;
+client_t client;
+
+ptr_t<char> name;
+
+void create_services(avahi::Client *c);
+
+void entry_group_callback(avahi::EntryGroup *g, avahi::EntryGroupState state, void *) {
+  group = g;
+
+  switch(state) {
+  case avahi::ENTRY_GROUP_ESTABLISHED:
+    BOOST_LOG(info) << "Avahi service " << name.get() << " successfully established.";
+    break;
+  case avahi::ENTRY_GROUP_COLLISION:
+    name.reset(avahi::alternative_service_name(name.get()));
+
+    BOOST_LOG(info) << "Avahi service name collision, renaming service to " << name.get();
+
+    create_services(avahi::entry_group_get_client(g));
+    break;
+  case avahi::ENTRY_GROUP_FAILURE:
+    BOOST_LOG(error) << "Avahi entry group failure: " << avahi::strerror(avahi::client_errno(avahi::entry_group_get_client(g)));
+    avahi::simple_poll_quit(poll.get());
+    break;
+  case avahi::ENTRY_GROUP_UNCOMMITED:
+  case avahi::ENTRY_GROUP_REGISTERING:;
+  }
+}
+
+void create_services(avahi::Client *c) {
+  int ret;
+
+  auto fg = util::fail_guard([]() {
+    avahi::simple_poll_quit(poll.get());
+  });
+
+  if(!group) {
+    if(!(group = avahi::entry_group_new(c, entry_group_callback, nullptr))) {
+      BOOST_LOG(error) << "avahi::entry_group_new() failed: "sv << avahi::strerror(avahi::client_errno(c));
+      return;
+    }
+  }
+
+  if(avahi::entry_group_is_empty(group)) {
+    BOOST_LOG(info) << "Adding avahi service "sv << name.get();
+
+    ret = avahi::entry_group_add_service(
+      group,
+      avahi::IF_UNSPEC, avahi::PROTO_UNSPEC,
+      avahi::PublishFlags(0),
+      name.get(),
+      SERVICE_TYPE,
+      nullptr, nullptr,
+      map_port(nvhttp::PORT_HTTP),
+      nullptr);
+
+    if(ret < 0) {
+      if(ret == avahi::ERR_COLLISION) {
+        // A service name collision with a local service happened. Let's pick a new name
+        name.reset(avahi::alternative_service_name(name.get()));
+        BOOST_LOG(info) << "Service name collision, renaming service to "sv << name.get();
+
+        avahi::entry_group_reset(group);
+
+        create_services(c);
+
+        fg.disable();
+        return;
+      }
+
+      BOOST_LOG(error) << "Failed to add "sv << SERVICE_TYPE << " service: "sv << avahi::strerror(ret);
+      return;
+    }
+
+    ret = avahi::entry_group_commit(group);
+    if(ret < 0) {
+      BOOST_LOG(error) << "Failed to commit entry group: "sv << avahi::strerror(ret);
+      return;
+    }
+  }
+
+  fg.disable();
+}
+
+void client_callback(avahi::Client *c, avahi::ClientState state, void *) {
+  switch(state) {
+  case avahi::CLIENT_S_RUNNING:
+    create_services(c);
+    break;
+  case avahi::CLIENT_FAILURE:
+    BOOST_LOG(error) << "Client failure: "sv << avahi::strerror(avahi::client_errno(c));
+    avahi::simple_poll_quit(poll.get());
+    break;
+  case avahi::CLIENT_S_COLLISION:
+  case avahi::CLIENT_S_REGISTERING:
+    if(group)
+      avahi::entry_group_reset(group);
+    break;
+  case avahi::CLIENT_CONNECTING:;
+  }
+}
+
+class deinit_t : public ::platf::deinit_t {
+public:
+  std::thread poll_thread;
+
+  deinit_t(std::thread poll_thread) : poll_thread { std::move(poll_thread) } {}
+
+  ~deinit_t() override {
+    if(avahi::simple_poll_quit && poll) {
+      avahi::simple_poll_quit(poll.get());
+    }
+
+    if(poll_thread.joinable()) {
+      poll_thread.join();
+    }
+  }
+};
+
+[[nodiscard]] std::unique_ptr<::platf::deinit_t> start() {
+  if(avahi::init_client()) {
+    return nullptr;
+  }
+
+  int avhi_error;
+
+  poll.reset(avahi::simple_poll_new());
+  if(!poll) {
+    BOOST_LOG(error) << "Failed to create simple poll object."sv;
+    return nullptr;
+  }
+
+  name.reset(avahi::strdup(SERVICE_NAME));
+
+  client.reset(
+    avahi::client_new(avahi::simple_poll_get(poll.get()), avahi::ClientFlags(0), client_callback, nullptr, &avhi_error));
+
+  if(!client) {
+    BOOST_LOG(error) << "Failed to create client: "sv << avahi::strerror(avhi_error);
+    return nullptr;
+  }
+
+  return std::make_unique<deinit_t>(std::thread { avahi::simple_poll_loop, poll.get() });
+}
+}; // namespace platf::publish

src/platform/windows/PolicyConfig.h

@@ -0,0 +1,164 @@
+// ----------------------------------------------------------------------------
+// PolicyConfig.h
+// Undocumented COM-interface IPolicyConfig.
+// Use for set default audio render endpoint
+// @author EreTIk
+// http://eretik.omegahg.com/
+// ----------------------------------------------------------------------------
+
+
+#pragma once
+
+#ifdef __MINGW32__
+#undef DEFINE_GUID
+#ifdef __cplusplus
+#define DEFINE_GUID(name, l, w1, w2, b1, b2, b3, b4, b5, b6, b7, b8) EXTERN_C const GUID DECLSPEC_SELECTANY name = { l, w1, w2, { b1, b2, b3, b4, b5, b6, b7, b8 } }
+#else
+#define DEFINE_GUID(name, l, w1, w2, b1, b2, b3, b4, b5, b6, b7, b8) const GUID DECLSPEC_SELECTANY name = { l, w1, w2, { b1, b2, b3, b4, b5, b6, b7, b8 } }
+#endif
+
+DEFINE_GUID(IID_IPolicyConfig, 0xf8679f50, 0x850a, 0x41cf, 0x9c, 0x72, 0x43, 0x0f, 0x29, 0x02, 0x90, 0xc8);
+DEFINE_GUID(CLSID_CPolicyConfigClient, 0x870af99c, 0x171d, 0x4f9e, 0xaf, 0x0d, 0xe6, 0x3d, 0xf4, 0x0c, 0x2b, 0xc9);
+
+#endif
+
+interface DECLSPEC_UUID("f8679f50-850a-41cf-9c72-430f290290c8") IPolicyConfig;
+class DECLSPEC_UUID("870af99c-171d-4f9e-af0d-e63df40c2bc9") CPolicyConfigClient;
+// ----------------------------------------------------------------------------
+// class CPolicyConfigClient
+// {870af99c-171d-4f9e-af0d-e63df40c2bc9}
+//
+// interface IPolicyConfig
+// {f8679f50-850a-41cf-9c72-430f290290c8}
+//
+// Query interface:
+// CComPtr<IPolicyConfig> PolicyConfig;
+// PolicyConfig.CoCreateInstance(__uuidof(CPolicyConfigClient));
+//
+// @compatible: Windows 7 and Later
+// ----------------------------------------------------------------------------
+interface IPolicyConfig : public IUnknown {
+public:
+  virtual HRESULT GetMixFormat(
+    PCWSTR,
+    WAVEFORMATEX **);
+
+  virtual HRESULT STDMETHODCALLTYPE GetDeviceFormat(
+    PCWSTR,
+    INT,
+    WAVEFORMATEX **);
+
+  virtual HRESULT STDMETHODCALLTYPE ResetDeviceFormat(
+    PCWSTR);
+
+  virtual HRESULT STDMETHODCALLTYPE SetDeviceFormat(
+    PCWSTR,
+    WAVEFORMATEX *,
+    WAVEFORMATEX *);
+
+  virtual HRESULT STDMETHODCALLTYPE GetProcessingPeriod(
+    PCWSTR,
+    INT,
+    PINT64,
+    PINT64);
+
+  virtual HRESULT STDMETHODCALLTYPE SetProcessingPeriod(
+    PCWSTR,
+    PINT64);
+
+  virtual HRESULT STDMETHODCALLTYPE GetShareMode(
+    PCWSTR,
+    struct DeviceShareMode *);
+
+  virtual HRESULT STDMETHODCALLTYPE SetShareMode(
+    PCWSTR,
+    struct DeviceShareMode *);
+
+  virtual HRESULT STDMETHODCALLTYPE GetPropertyValue(
+    PCWSTR,
+    const PROPERTYKEY &,
+    PROPVARIANT *);
+
+  virtual HRESULT STDMETHODCALLTYPE SetPropertyValue(
+    PCWSTR,
+    const PROPERTYKEY &,
+    PROPVARIANT *);
+
+  virtual HRESULT STDMETHODCALLTYPE SetDefaultEndpoint(
+    PCWSTR wszDeviceId,
+    ERole eRole);
+
+  virtual HRESULT STDMETHODCALLTYPE SetEndpointVisibility(
+    PCWSTR,
+    INT);
+};
+
+interface DECLSPEC_UUID("568b9108-44bf-40b4-9006-86afe5b5a620") IPolicyConfigVista;
+class DECLSPEC_UUID("294935CE-F637-4E7C-A41B-AB255460B862") CPolicyConfigVistaClient;
+// ----------------------------------------------------------------------------
+// class CPolicyConfigVistaClient
+// {294935CE-F637-4E7C-A41B-AB255460B862}
+//
+// interface IPolicyConfigVista
+// {568b9108-44bf-40b4-9006-86afe5b5a620}
+//
+// Query interface:
+// CComPtr<IPolicyConfigVista> PolicyConfig;
+// PolicyConfig.CoCreateInstance(__uuidof(CPolicyConfigVistaClient));
+//
+// @compatible: Windows Vista and Later
+// ----------------------------------------------------------------------------
+interface IPolicyConfigVista : public IUnknown {
+public:
+  virtual HRESULT GetMixFormat(
+    PCWSTR,
+    WAVEFORMATEX **); // not available on Windows 7, use method from IPolicyConfig
+
+  virtual HRESULT STDMETHODCALLTYPE GetDeviceFormat(
+    PCWSTR,
+    INT,
+    WAVEFORMATEX **);
+
+  virtual HRESULT STDMETHODCALLTYPE SetDeviceFormat(
+    PCWSTR,
+    WAVEFORMATEX *,
+    WAVEFORMATEX *);
+
+  virtual HRESULT STDMETHODCALLTYPE GetProcessingPeriod(
+    PCWSTR,
+    INT,
+    PINT64,
+    PINT64); // not available on Windows 7, use method from IPolicyConfig
+
+  virtual HRESULT STDMETHODCALLTYPE SetProcessingPeriod(
+    PCWSTR,
+    PINT64); // not available on Windows 7, use method from IPolicyConfig
+
+  virtual HRESULT STDMETHODCALLTYPE GetShareMode(
+    PCWSTR,
+    struct DeviceShareMode *); // not available on Windows 7, use method from IPolicyConfig
+
+  virtual HRESULT STDMETHODCALLTYPE SetShareMode(
+    PCWSTR,
+    struct DeviceShareMode *); // not available on Windows 7, use method from IPolicyConfig
+
+  virtual HRESULT STDMETHODCALLTYPE GetPropertyValue(
+    PCWSTR,
+    const PROPERTYKEY &,
+    PROPVARIANT *);
+
+  virtual HRESULT STDMETHODCALLTYPE SetPropertyValue(
+    PCWSTR,
+    const PROPERTYKEY &,
+    PROPVARIANT *);
+
+  virtual HRESULT STDMETHODCALLTYPE SetDefaultEndpoint(
+    PCWSTR wszDeviceId,
+    ERole eRole);
+
+  virtual HRESULT STDMETHODCALLTYPE SetEndpointVisibility(
+    PCWSTR,
+    INT); // not available on Windows 7, use method from IPolicyConfig
+};
+
+// ----------------------------------------------------------------------------

src/platform/windows/audio.cpp

@@ -0,0 +1,988 @@
+//
+// Created by loki on 1/12/20.
+//
+
+#include <audioclient.h>
+#include <mmdeviceapi.h>
+#include <roapi.h>
+
+#include <codecvt>
+
+#include <synchapi.h>
+
+#define INITGUID
+#include <propkeydef.h>
+#undef INITGUID
+
+#include "src/config.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+
+// Must be the last included file
+// clang-format off
+#include "PolicyConfig.h"
+// clang-format on
+
+DEFINE_PROPERTYKEY(PKEY_Device_DeviceDesc, 0xa45c254e, 0xdf1c, 0x4efd, 0x80, 0x20, 0x67, 0xd1, 0x46, 0xa8, 0x50, 0xe0, 2);    // DEVPROP_TYPE_STRING
+DEFINE_PROPERTYKEY(PKEY_Device_FriendlyName, 0xa45c254e, 0xdf1c, 0x4efd, 0x80, 0x20, 0x67, 0xd1, 0x46, 0xa8, 0x50, 0xe0, 14); // DEVPROP_TYPE_STRING
+DEFINE_PROPERTYKEY(PKEY_DeviceInterface_FriendlyName, 0x026e516e, 0xb814, 0x414b, 0x83, 0xcd, 0x85, 0x6d, 0x6f, 0xef, 0x48, 0x22, 2);
+
+const CLSID CLSID_MMDeviceEnumerator = __uuidof(MMDeviceEnumerator);
+const IID IID_IMMDeviceEnumerator    = __uuidof(IMMDeviceEnumerator);
+const IID IID_IAudioClient           = __uuidof(IAudioClient);
+const IID IID_IAudioCaptureClient    = __uuidof(IAudioCaptureClient);
+
+using namespace std::literals;
+namespace platf::audio {
+constexpr auto SAMPLE_RATE = 48000;
+
+template<class T>
+void Release(T *p) {
+  p->Release();
+}
+
+template<class T>
+void co_task_free(T *p) {
+  CoTaskMemFree((LPVOID)p);
+}
+
+using device_enum_t   = util::safe_ptr<IMMDeviceEnumerator, Release<IMMDeviceEnumerator>>;
+using device_t        = util::safe_ptr<IMMDevice, Release<IMMDevice>>;
+using collection_t    = util::safe_ptr<IMMDeviceCollection, Release<IMMDeviceCollection>>;
+using audio_client_t  = util::safe_ptr<IAudioClient, Release<IAudioClient>>;
+using audio_capture_t = util::safe_ptr<IAudioCaptureClient, Release<IAudioCaptureClient>>;
+using wave_format_t   = util::safe_ptr<WAVEFORMATEX, co_task_free<WAVEFORMATEX>>;
+using wstring_t       = util::safe_ptr<WCHAR, co_task_free<WCHAR>>;
+using handle_t        = util::safe_ptr_v2<void, BOOL, CloseHandle>;
+using policy_t        = util::safe_ptr<IPolicyConfig, Release<IPolicyConfig>>;
+using prop_t          = util::safe_ptr<IPropertyStore, Release<IPropertyStore>>;
+
+class co_init_t : public deinit_t {
+public:
+  co_init_t() {
+    CoInitializeEx(nullptr, COINIT_MULTITHREADED | COINIT_SPEED_OVER_MEMORY);
+  }
+
+  ~co_init_t() override {
+    CoUninitialize();
+  }
+};
+
+class prop_var_t {
+public:
+  prop_var_t() {
+    PropVariantInit(&prop);
+  }
+
+  ~prop_var_t() {
+    PropVariantClear(&prop);
+  }
+
+  PROPVARIANT prop;
+};
+
+class audio_pipe_t {
+public:
+  static constexpr auto stereo     = 2;
+  static constexpr auto channels51 = 6;
+  static constexpr auto channels71 = 8;
+
+  using samples_t = std::vector<std::int16_t>;
+  using buf_t     = util::buffer_t<std::int16_t>;
+
+  virtual void to_stereo(samples_t &out, const buf_t &in) = 0;
+  virtual void to_51(samples_t &out, const buf_t &in)     = 0;
+  virtual void to_71(samples_t &out, const buf_t &in)     = 0;
+};
+
+class mono_t : public audio_pipe_t {
+public:
+  void to_stereo(samples_t &out, const buf_t &in) override {
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end;) {
+      *sample_out_p++ = *sample_in_pos * 7 / 10;
+      *sample_out_p++ = *sample_in_pos++ * 7 / 10;
+    }
+  }
+
+  void to_51(samples_t &out, const buf_t &in) override {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += channels51) {
+      int left = *sample_in_pos++;
+
+      auto fl = (left * 7 / 10);
+
+      sample_out_p[FRONT_LEFT]    = fl;
+      sample_out_p[FRONT_RIGHT]   = fl;
+      sample_out_p[FRONT_CENTER]  = fl * 6;
+      sample_out_p[LOW_FREQUENCY] = fl / 10;
+      sample_out_p[BACK_LEFT]     = left * 4 / 10;
+      sample_out_p[BACK_RIGHT]    = left * 4 / 10;
+    }
+  }
+
+  void to_71(samples_t &out, const buf_t &in) override {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += channels71) {
+      int left = *sample_in_pos++;
+
+      auto fl = (left * 7 / 10);
+
+      sample_out_p[FRONT_LEFT]    = fl;
+      sample_out_p[FRONT_RIGHT]   = fl;
+      sample_out_p[FRONT_CENTER]  = fl * 6;
+      sample_out_p[LOW_FREQUENCY] = fl / 10;
+      sample_out_p[BACK_LEFT]     = left * 4 / 10;
+      sample_out_p[BACK_RIGHT]    = left * 4 / 10;
+      sample_out_p[SIDE_LEFT]     = left * 5 / 10;
+      sample_out_p[SIDE_RIGHT]    = left * 5 / 10;
+    }
+  }
+};
+
+class stereo_t : public audio_pipe_t {
+public:
+  void to_stereo(samples_t &out, const buf_t &in) override {
+    std::copy_n(std::begin(in), out.size(), std::begin(out));
+  }
+
+  void to_51(samples_t &out, const buf_t &in) override {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += channels51) {
+      int left  = sample_in_pos[speaker::FRONT_LEFT];
+      int right = sample_in_pos[speaker::FRONT_RIGHT];
+
+      sample_in_pos += 2;
+
+      auto fl = (left * 7 / 10);
+      auto fr = (right * 7 / 10);
+
+      auto mix = (fl + fr) / 2;
+
+      sample_out_p[FRONT_LEFT]    = fl;
+      sample_out_p[FRONT_RIGHT]   = fr;
+      sample_out_p[FRONT_CENTER]  = mix;
+      sample_out_p[LOW_FREQUENCY] = mix / 2;
+      sample_out_p[BACK_LEFT]     = left * 4 / 10;
+      sample_out_p[BACK_RIGHT]    = right * 4 / 10;
+    }
+  }
+
+  void to_71(samples_t &out, const buf_t &in) override {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += channels71) {
+      int left  = sample_in_pos[speaker::FRONT_LEFT];
+      int right = sample_in_pos[speaker::FRONT_RIGHT];
+
+      sample_in_pos += 2;
+
+      auto fl = (left * 7 / 10);
+      auto fr = (right * 7 / 10);
+
+      auto mix = (fl + fr) / 2;
+
+      sample_out_p[FRONT_LEFT]    = fl;
+      sample_out_p[FRONT_RIGHT]   = fr;
+      sample_out_p[FRONT_CENTER]  = mix;
+      sample_out_p[LOW_FREQUENCY] = mix / 2;
+      sample_out_p[BACK_LEFT]     = left * 4 / 10;
+      sample_out_p[BACK_RIGHT]    = right * 4 / 10;
+      sample_out_p[SIDE_LEFT]     = left * 5 / 10;
+      sample_out_p[SIDE_RIGHT]    = right * 5 / 10;
+    }
+  }
+};
+
+class surr51_t : public audio_pipe_t {
+public:
+  void to_stereo(samples_t &out, const buf_t &in) {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += stereo) {
+      int left {}, right {};
+
+      left += sample_in_pos[FRONT_LEFT];
+      left += sample_in_pos[FRONT_CENTER] * 9 / 10;
+      left += sample_in_pos[LOW_FREQUENCY] * 3 / 10;
+      left += sample_in_pos[BACK_LEFT] * 7 / 10;
+      left += sample_in_pos[BACK_RIGHT] * 3 / 10;
+
+      right += sample_in_pos[FRONT_RIGHT];
+      right += sample_in_pos[FRONT_CENTER] * 9 / 10;
+      right += sample_in_pos[LOW_FREQUENCY] * 3 / 10;
+      right += sample_in_pos[BACK_LEFT] * 3 / 10;
+      right += sample_in_pos[BACK_RIGHT] * 7 / 10;
+
+      sample_out_p[0] = left;
+      sample_out_p[1] = right;
+
+      sample_in_pos += channels51;
+    }
+  }
+
+  void to_51(samples_t &out, const buf_t &in) override {
+    std::copy_n(std::begin(in), out.size(), std::begin(out));
+  }
+
+  void to_71(samples_t &out, const buf_t &in) override {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += channels71) {
+      int fl = sample_in_pos[FRONT_LEFT];
+      int fr = sample_in_pos[FRONT_RIGHT];
+      int bl = sample_in_pos[BACK_LEFT];
+      int br = sample_in_pos[BACK_RIGHT];
+
+      auto mix_l = (fl + bl) / 2;
+      auto mix_r = (bl + br) / 2;
+
+      sample_out_p[FRONT_LEFT]    = fl;
+      sample_out_p[FRONT_RIGHT]   = fr;
+      sample_out_p[FRONT_CENTER]  = sample_in_pos[FRONT_CENTER];
+      sample_out_p[LOW_FREQUENCY] = sample_in_pos[LOW_FREQUENCY];
+      sample_out_p[BACK_LEFT]     = bl;
+      sample_out_p[BACK_RIGHT]    = br;
+      sample_out_p[SIDE_LEFT]     = mix_l;
+      sample_out_p[SIDE_RIGHT]    = mix_r;
+
+      sample_in_pos += channels51;
+    }
+  }
+};
+
+class surr71_t : public audio_pipe_t {
+public:
+  void to_stereo(samples_t &out, const buf_t &in) {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += stereo) {
+      int left {}, right {};
+
+      left += sample_in_pos[FRONT_LEFT];
+      left += sample_in_pos[FRONT_CENTER] * 9 / 10;
+      left += sample_in_pos[LOW_FREQUENCY] * 3 / 10;
+      left += sample_in_pos[BACK_LEFT] * 7 / 10;
+      left += sample_in_pos[BACK_RIGHT] * 3 / 10;
+      left += sample_in_pos[SIDE_LEFT];
+
+      right += sample_in_pos[FRONT_RIGHT];
+      right += sample_in_pos[FRONT_CENTER] * 9 / 10;
+      right += sample_in_pos[LOW_FREQUENCY] * 3 / 10;
+      right += sample_in_pos[BACK_LEFT] * 3 / 10;
+      right += sample_in_pos[BACK_RIGHT] * 7 / 10;
+      right += sample_in_pos[SIDE_RIGHT];
+
+      sample_out_p[0] = left;
+      sample_out_p[1] = right;
+
+      sample_in_pos += channels71;
+    }
+  }
+
+  void to_51(samples_t &out, const buf_t &in) override {
+    using namespace speaker;
+
+    auto sample_in_pos = std::begin(in);
+    auto sample_end    = std::begin(out) + out.size();
+
+    for(auto sample_out_p = std::begin(out); sample_out_p != sample_end; sample_out_p += channels51) {
+      auto sl = (int)sample_out_p[SIDE_LEFT] * 3 / 10;
+      auto sr = (int)sample_out_p[SIDE_RIGHT] * 3 / 10;
+
+      sample_out_p[FRONT_LEFT]    = sample_in_pos[FRONT_LEFT] + sl;
+      sample_out_p[FRONT_RIGHT]   = sample_in_pos[FRONT_RIGHT] + sr;
+      sample_out_p[FRONT_CENTER]  = sample_in_pos[FRONT_CENTER];
+      sample_out_p[LOW_FREQUENCY] = sample_in_pos[LOW_FREQUENCY];
+      sample_out_p[BACK_LEFT]     = sample_in_pos[BACK_LEFT] + sl;
+      sample_out_p[BACK_RIGHT]    = sample_in_pos[BACK_RIGHT] + sr;
+
+      sample_in_pos += channels71;
+    }
+  }
+
+  void to_71(samples_t &out, const buf_t &in) override {
+    std::copy_n(std::begin(in), out.size(), std::begin(out));
+  }
+};
+
+static std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
+struct format_t {
+  enum type_e : int {
+    none,
+    mono,
+    stereo,
+    surr51,
+    surr71,
+  } type;
+
+  std::string_view name;
+  int channels;
+  int channel_mask;
+} formats[] {
+  {
+    format_t::mono,
+    "Mono"sv,
+    1,
+    SPEAKER_FRONT_CENTER,
+  },
+  {
+    format_t::stereo,
+    "Stereo"sv,
+    2,
+    SPEAKER_FRONT_LEFT | SPEAKER_FRONT_RIGHT,
+  },
+  {
+    format_t::surr51,
+    "Surround 5.1"sv,
+    6,
+    SPEAKER_FRONT_LEFT |
+      SPEAKER_FRONT_RIGHT |
+      SPEAKER_FRONT_CENTER |
+      SPEAKER_LOW_FREQUENCY |
+      SPEAKER_BACK_LEFT |
+      SPEAKER_BACK_RIGHT,
+  },
+  {
+    format_t::surr71,
+    "Surround 7.1"sv,
+    8,
+    SPEAKER_FRONT_LEFT |
+      SPEAKER_FRONT_RIGHT |
+      SPEAKER_FRONT_CENTER |
+      SPEAKER_LOW_FREQUENCY |
+      SPEAKER_BACK_LEFT |
+      SPEAKER_BACK_RIGHT |
+      SPEAKER_SIDE_LEFT |
+      SPEAKER_SIDE_RIGHT,
+  },
+};
+
+static format_t surround_51_side_speakers {
+  format_t::surr51,
+  "Surround 5.1"sv,
+  6,
+  SPEAKER_FRONT_LEFT |
+    SPEAKER_FRONT_RIGHT |
+    SPEAKER_FRONT_CENTER |
+    SPEAKER_LOW_FREQUENCY |
+    SPEAKER_SIDE_LEFT |
+    SPEAKER_SIDE_RIGHT,
+};
+
+void set_wave_format(audio::wave_format_t &wave_format, const format_t &format) {
+  wave_format->nChannels       = format.channels;
+  wave_format->nBlockAlign     = wave_format->nChannels * wave_format->wBitsPerSample / 8;
+  wave_format->nAvgBytesPerSec = wave_format->nSamplesPerSec * wave_format->nBlockAlign;
+
+  if(wave_format->wFormatTag == WAVE_FORMAT_EXTENSIBLE) {
+    ((PWAVEFORMATEXTENSIBLE)wave_format.get())->dwChannelMask = format.channel_mask;
+  }
+}
+
+int init_wave_format(audio::wave_format_t &wave_format, DWORD sample_rate) {
+  wave_format->wBitsPerSample = 16;
+  wave_format->nSamplesPerSec = sample_rate;
+  switch(wave_format->wFormatTag) {
+  case WAVE_FORMAT_PCM:
+    break;
+  case WAVE_FORMAT_IEEE_FLOAT:
+    break;
+  case WAVE_FORMAT_EXTENSIBLE: {
+    auto wave_ex = (PWAVEFORMATEXTENSIBLE)wave_format.get();
+    if(IsEqualGUID(KSDATAFORMAT_SUBTYPE_IEEE_FLOAT, wave_ex->SubFormat)) {
+      wave_ex->Samples.wValidBitsPerSample = 16;
+      wave_ex->SubFormat                   = KSDATAFORMAT_SUBTYPE_PCM;
+      break;
+    }
+
+    BOOST_LOG(error) << "Unsupported Sub Format for WAVE_FORMAT_EXTENSIBLE: [0x"sv << util::hex(wave_ex->SubFormat).to_string_view() << ']';
+  }
+  default:
+    BOOST_LOG(error) << "Unsupported Wave Format: [0x"sv << util::hex(wave_format->wFormatTag).to_string_view() << ']';
+    return -1;
+  };
+
+  return 0;
+}
+
+audio_client_t make_audio_client(device_t &device, const format_t &format, int sample_rate) {
+  audio_client_t audio_client;
+  auto status = device->Activate(
+    IID_IAudioClient,
+    CLSCTX_ALL,
+    nullptr,
+    (void **)&audio_client);
+
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Couldn't activate Device: [0x"sv << util::hex(status).to_string_view() << ']';
+
+    return nullptr;
+  }
+
+  wave_format_t wave_format;
+  status = audio_client->GetMixFormat(&wave_format);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Couldn't acquire Wave Format [0x"sv << util::hex(status).to_string_view() << ']';
+
+    return nullptr;
+  }
+
+  if(init_wave_format(wave_format, sample_rate)) {
+    return nullptr;
+  }
+  set_wave_format(wave_format, format);
+
+  status = audio_client->Initialize(
+    AUDCLNT_SHAREMODE_SHARED,
+    AUDCLNT_STREAMFLAGS_LOOPBACK | AUDCLNT_STREAMFLAGS_EVENTCALLBACK,
+    0, 0,
+    wave_format.get(),
+    nullptr);
+
+  if(status) {
+    BOOST_LOG(debug) << "Couldn't initialize audio client for ["sv << format.name << "]: [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return audio_client;
+}
+
+const wchar_t *no_null(const wchar_t *str) {
+  return str ? str : L"Unknown";
+}
+
+format_t::type_e validate_device(device_t &device, int sample_rate) {
+  for(const auto &format : formats) {
+    // Ensure WaveFromat is compatible
+    auto audio_client = make_audio_client(device, format, sample_rate);
+
+    BOOST_LOG(debug) << format.name << ": "sv << (!audio_client ? "unsupported"sv : "supported"sv);
+
+    if(audio_client) {
+      return format.type;
+    }
+  }
+
+  return format_t::none;
+}
+
+device_t default_device(device_enum_t &device_enum) {
+  device_t device;
+  HRESULT status;
+  status = device_enum->GetDefaultAudioEndpoint(
+    eRender,
+    eConsole,
+    &device);
+
+
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Couldn't create audio Device [0x"sv << util::hex(status).to_string_view() << ']';
+
+    return nullptr;
+  }
+
+  return device;
+}
+
+class mic_wasapi_t : public mic_t {
+public:
+  capture_e sample(std::vector<std::int16_t> &sample_out) override {
+    auto sample_size = sample_out.size() / channels_out * channels_in;
+    while(sample_buf_pos - std::begin(sample_buf) < sample_size) {
+      //FIXME: Use IAudioClient3 instead of IAudioClient, that would allows for adjusting the latency of the audio samples
+      auto capture_result = _fill_buffer();
+
+      if(capture_result != capture_e::ok) {
+        return capture_result;
+      }
+    }
+
+    switch(channels_out) {
+    case 2:
+      pipe->to_stereo(sample_out, sample_buf);
+      break;
+    case 6:
+      pipe->to_51(sample_out, sample_buf);
+      break;
+    case 8:
+      pipe->to_71(sample_out, sample_buf);
+      break;
+    default:
+      BOOST_LOG(error) << "converting to ["sv << channels_out << "] channels is not supported"sv;
+      return capture_e::error;
+    }
+
+    // The excess samples should be in front of the queue
+    std::move(&sample_buf[sample_size], sample_buf_pos, std::begin(sample_buf));
+    sample_buf_pos -= sample_size;
+
+    return capture_e::ok;
+  }
+
+
+  int init(std::uint32_t sample_rate, std::uint32_t frame_size, std::uint32_t channels_out) {
+    audio_event.reset(CreateEventA(nullptr, FALSE, FALSE, nullptr));
+    if(!audio_event) {
+      BOOST_LOG(error) << "Couldn't create Event handle"sv;
+
+      return -1;
+    }
+
+    HRESULT status;
+
+    status = CoCreateInstance(
+      CLSID_MMDeviceEnumerator,
+      nullptr,
+      CLSCTX_ALL,
+      IID_IMMDeviceEnumerator,
+      (void **)&device_enum);
+
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't create Device Enumerator [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    auto device = default_device(device_enum);
+    if(!device) {
+      return -1;
+    }
+
+    for(auto &format : formats) {
+      BOOST_LOG(debug) << "Trying audio format ["sv << format.name << ']';
+      audio_client = make_audio_client(device, format, sample_rate);
+
+      if(audio_client) {
+        BOOST_LOG(debug) << "Found audio format ["sv << format.name << ']';
+        channels_in        = format.channels;
+        this->channels_out = channels_out;
+
+        switch(channels_in) {
+        case 1:
+          pipe = std::make_unique<mono_t>();
+          break;
+        case 2:
+          pipe = std::make_unique<stereo_t>();
+          break;
+        case 6:
+          pipe = std::make_unique<surr51_t>();
+          break;
+        case 8:
+          pipe = std::make_unique<surr71_t>();
+          break;
+        default:
+          BOOST_LOG(error) << "converting from ["sv << channels_in << "] channels is not supported"sv;
+          return -1;
+        }
+        break;
+      }
+    }
+
+    if(!audio_client) {
+      BOOST_LOG(error) << "Couldn't find supported format for audio"sv;
+      return -1;
+    }
+
+    REFERENCE_TIME default_latency;
+    audio_client->GetDevicePeriod(&default_latency, nullptr);
+    default_latency_ms = default_latency / 1000;
+
+    std::uint32_t frames;
+    status = audio_client->GetBufferSize(&frames);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't acquire the number of audio frames [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    // *2 --> needs to fit double
+    sample_buf     = util::buffer_t<std::int16_t> { std::max(frames, frame_size) * 2 * channels_in };
+    sample_buf_pos = std::begin(sample_buf);
+
+    status = audio_client->GetService(IID_IAudioCaptureClient, (void **)&audio_capture);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't initialize audio capture client [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    status = audio_client->SetEventHandle(audio_event.get());
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't set event handle [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    status = audio_client->Start();
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't start recording [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    return 0;
+  }
+
+  ~mic_wasapi_t() override {
+    if(audio_client) {
+      audio_client->Stop();
+    }
+  }
+
+private:
+  capture_e _fill_buffer() {
+    HRESULT status;
+
+    // Total number of samples
+    struct sample_aligned_t {
+      std::uint32_t uninitialized;
+      std::int16_t *samples;
+    } sample_aligned;
+
+    // number of samples / number of channels
+    struct block_aligned_t {
+      std::uint32_t audio_sample_size;
+    } block_aligned;
+
+    status = WaitForSingleObjectEx(audio_event.get(), default_latency_ms, FALSE);
+    switch(status) {
+    case WAIT_OBJECT_0:
+      break;
+    case WAIT_TIMEOUT:
+      return capture_e::timeout;
+    default:
+      BOOST_LOG(error) << "Couldn't wait for audio event: [0x"sv << util::hex(status).to_string_view() << ']';
+      return capture_e::error;
+    }
+
+    std::uint32_t packet_size {};
+    for(
+      status = audio_capture->GetNextPacketSize(&packet_size);
+      SUCCEEDED(status) && packet_size > 0;
+      status = audio_capture->GetNextPacketSize(&packet_size)) {
+      DWORD buffer_flags;
+      status = audio_capture->GetBuffer(
+        (BYTE **)&sample_aligned.samples,
+        &block_aligned.audio_sample_size,
+        &buffer_flags,
+        nullptr, nullptr);
+
+      switch(status) {
+      case S_OK:
+        break;
+      case AUDCLNT_E_DEVICE_INVALIDATED:
+        return capture_e::reinit;
+      default:
+        BOOST_LOG(error) << "Couldn't capture audio [0x"sv << util::hex(status).to_string_view() << ']';
+        return capture_e::error;
+      }
+
+      sample_aligned.uninitialized = std::end(sample_buf) - sample_buf_pos;
+      auto n                       = std::min(sample_aligned.uninitialized, block_aligned.audio_sample_size * channels_in);
+
+      if(buffer_flags & AUDCLNT_BUFFERFLAGS_SILENT) {
+        std::fill_n(sample_buf_pos, n, 0);
+      }
+      else {
+        std::copy_n(sample_aligned.samples, n, sample_buf_pos);
+      }
+
+      sample_buf_pos += n;
+
+      audio_capture->ReleaseBuffer(block_aligned.audio_sample_size);
+    }
+
+    if(status == AUDCLNT_E_DEVICE_INVALIDATED) {
+      return capture_e::reinit;
+    }
+
+    if(FAILED(status)) {
+      return capture_e::error;
+    }
+
+    return capture_e::ok;
+  }
+
+public:
+  handle_t audio_event;
+
+  device_enum_t device_enum;
+  device_t device;
+  audio_client_t audio_client;
+  audio_capture_t audio_capture;
+
+  REFERENCE_TIME default_latency_ms;
+
+  util::buffer_t<std::int16_t> sample_buf;
+  std::int16_t *sample_buf_pos;
+
+  // out --> our audio output
+  int channels_out;
+  // in --> our wasapi input
+  int channels_in;
+
+  std::unique_ptr<audio_pipe_t> pipe;
+};
+
+class audio_control_t : public ::platf::audio_control_t {
+public:
+  std::optional<sink_t> sink_info() override {
+    auto virtual_adapter_name = L"Steam Streaming Speakers"sv;
+
+    sink_t sink;
+
+    audio::device_enum_t device_enum;
+    auto status = CoCreateInstance(
+      CLSID_MMDeviceEnumerator,
+      nullptr,
+      CLSCTX_ALL,
+      IID_IMMDeviceEnumerator,
+      (void **)&device_enum);
+
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't create Device Enumerator: [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return std::nullopt;
+    }
+
+    auto device = default_device(device_enum);
+    if(!device) {
+      return std::nullopt;
+    }
+
+    audio::wstring_t wstring;
+    device->GetId(&wstring);
+
+    sink.host = converter.to_bytes(wstring.get());
+
+    collection_t collection;
+    status = device_enum->EnumAudioEndpoints(eRender, DEVICE_STATE_ACTIVE, &collection);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't enumerate: [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return std::nullopt;
+    }
+
+    UINT count;
+    collection->GetCount(&count);
+
+    std::string virtual_device_id = config::audio.virtual_sink;
+    for(auto x = 0; x < count; ++x) {
+      audio::device_t device;
+      collection->Item(x, &device);
+
+      auto type = validate_device(device, SAMPLE_RATE);
+      if(type == format_t::none) {
+        continue;
+      }
+
+      audio::wstring_t wstring;
+      device->GetId(&wstring);
+
+      audio::prop_t prop;
+      device->OpenPropertyStore(STGM_READ, &prop);
+
+      prop_var_t adapter_friendly_name;
+      prop_var_t device_friendly_name;
+      prop_var_t device_desc;
+
+      prop->GetValue(PKEY_Device_FriendlyName, &device_friendly_name.prop);
+      prop->GetValue(PKEY_DeviceInterface_FriendlyName, &adapter_friendly_name.prop);
+      prop->GetValue(PKEY_Device_DeviceDesc, &device_desc.prop);
+
+      auto adapter_name = no_null((LPWSTR)adapter_friendly_name.prop.pszVal);
+      BOOST_LOG(verbose)
+        << L"===== Device ====="sv << std::endl
+        << L"Device ID          : "sv << wstring.get() << std::endl
+        << L"Device name        : "sv << no_null((LPWSTR)device_friendly_name.prop.pszVal) << std::endl
+        << L"Adapter name       : "sv << adapter_name << std::endl
+        << L"Device description : "sv << no_null((LPWSTR)device_desc.prop.pszVal) << std::endl
+        << std::endl;
+
+      if(virtual_device_id.empty() && adapter_name == virtual_adapter_name) {
+        virtual_device_id = converter.to_bytes(wstring.get());
+      }
+    }
+
+    if(!virtual_device_id.empty()) {
+      sink.null = std::make_optional(sink_t::null_t {
+        "virtual-"s.append(formats[format_t::stereo - 1].name) + virtual_device_id,
+        "virtual-"s.append(formats[format_t::surr51 - 1].name) + virtual_device_id,
+        "virtual-"s.append(formats[format_t::surr71 - 1].name) + virtual_device_id,
+      });
+    }
+
+    return sink;
+  }
+
+  std::unique_ptr<mic_t> microphone(const std::uint8_t *mapping, int channels, std::uint32_t sample_rate, std::uint32_t frame_size) override {
+    auto mic = std::make_unique<mic_wasapi_t>();
+
+    if(mic->init(sample_rate, frame_size, channels)) {
+      return nullptr;
+    }
+
+    return mic;
+  }
+
+  /**
+   * If the requested sink is a virtual sink, meaning no speakers attached to
+   * the host, then we can seamlessly set the format to stereo and surround sound.
+   * 
+   * Any virtual sink detected will be prefixed by:
+   *    virtual-(format name)
+   * If it doesn't contain that prefix, then the format will not be changed
+   */
+  std::optional<std::wstring> set_format(const std::string &sink) {
+    std::string_view sv { sink.c_str(), sink.size() };
+
+    format_t::type_e type = format_t::none;
+    // sink format:
+    // [virtual-(format name)]device_id
+    auto prefix = "virtual-"sv;
+    if(sv.find(prefix) == 0) {
+      sv = sv.substr(prefix.size(), sv.size() - prefix.size());
+
+      for(auto &format : formats) {
+        auto &name = format.name;
+        if(sv.find(name) == 0) {
+          type = format.type;
+          sv   = sv.substr(name.size(), sv.size() - name.size());
+
+          break;
+        }
+      }
+    }
+
+    auto wstring_device_id = converter.from_bytes(sv.data());
+
+    if(type == format_t::none) {
+      // wstring_device_id does not contain virtual-(format name)
+      // It's a simple deviceId, just pass it back
+      return std::make_optional(std::move(wstring_device_id));
+    }
+
+    wave_format_t wave_format;
+    auto status = policy->GetMixFormat(wstring_device_id.c_str(), &wave_format);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't acquire Wave Format [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return std::nullopt;
+    }
+
+    if(init_wave_format(wave_format, SAMPLE_RATE)) {
+      return std::nullopt;
+    }
+    set_wave_format(wave_format, formats[(int)type - 1]);
+
+    WAVEFORMATEXTENSIBLE p {};
+    status = policy->SetDeviceFormat(wstring_device_id.c_str(), wave_format.get(), (WAVEFORMATEX *)&p);
+
+    // Surround 5.1 might contain side-{left, right} instead of speaker in the back
+    // Try again with different speaker mask.
+    if(status == 0x88890008 && type == format_t::surr51) {
+      set_wave_format(wave_format, surround_51_side_speakers);
+      status = policy->SetDeviceFormat(wstring_device_id.c_str(), wave_format.get(), (WAVEFORMATEX *)&p);
+    }
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't set Wave Format [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return std::nullopt;
+    }
+
+    return std::make_optional(std::move(wstring_device_id));
+  }
+
+  int set_sink(const std::string &sink) override {
+    auto wstring_device_id = set_format(sink);
+    if(!wstring_device_id) {
+      return -1;
+    }
+
+    int failure {};
+    for(int x = 0; x < (int)ERole_enum_count; ++x) {
+      auto status = policy->SetDefaultEndpoint(wstring_device_id->c_str(), (ERole)x);
+      if(status) {
+        BOOST_LOG(warning) << "Couldn't set ["sv << sink << "] to role ["sv << x << ']';
+
+        ++failure;
+      }
+    }
+
+    return failure;
+  }
+
+  int init() {
+    auto status = CoCreateInstance(
+      CLSID_CPolicyConfigClient,
+      nullptr,
+      CLSCTX_ALL,
+      IID_IPolicyConfig,
+      (void **)&policy);
+
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't create audio policy config: [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    return 0;
+  }
+
+  ~audio_control_t() override {}
+
+  policy_t policy;
+};
+} // namespace platf::audio
+
+namespace platf {
+
+// It's not big enough to justify it's own source file :/
+namespace dxgi {
+int init();
+}
+
+std::unique_ptr<audio_control_t> audio_control() {
+  auto control = std::make_unique<audio::audio_control_t>();
+
+  if(control->init()) {
+    return nullptr;
+  }
+
+  return control;
+}
+
+std::unique_ptr<deinit_t> init() {
+  if(dxgi::init()) {
+    return nullptr;
+  }
+  return std::make_unique<platf::audio::co_init_t>();
+}
+} // namespace platf

src/platform/windows/display.h

@@ -0,0 +1,177 @@
+//
+// Created by loki on 4/23/20.
+//
+
+#ifndef SUNSHINE_DISPLAY_H
+#define SUNSHINE_DISPLAY_H
+
+#include <d3d11.h>
+#include <d3d11_4.h>
+#include <d3dcommon.h>
+#include <dwmapi.h>
+#include <dxgi.h>
+#include <dxgi1_2.h>
+
+#include "src/platform/common.h"
+#include "src/utility.h"
+
+namespace platf::dxgi {
+extern const char *format_str[];
+
+template<class T>
+void Release(T *dxgi) {
+  dxgi->Release();
+}
+
+using factory1_t            = util::safe_ptr<IDXGIFactory1, Release<IDXGIFactory1>>;
+using dxgi_t                = util::safe_ptr<IDXGIDevice, Release<IDXGIDevice>>;
+using dxgi1_t               = util::safe_ptr<IDXGIDevice1, Release<IDXGIDevice1>>;
+using device_t              = util::safe_ptr<ID3D11Device, Release<ID3D11Device>>;
+using device_ctx_t          = util::safe_ptr<ID3D11DeviceContext, Release<ID3D11DeviceContext>>;
+using adapter_t             = util::safe_ptr<IDXGIAdapter1, Release<IDXGIAdapter1>>;
+using output_t              = util::safe_ptr<IDXGIOutput, Release<IDXGIOutput>>;
+using output1_t             = util::safe_ptr<IDXGIOutput1, Release<IDXGIOutput1>>;
+using dup_t                 = util::safe_ptr<IDXGIOutputDuplication, Release<IDXGIOutputDuplication>>;
+using texture2d_t           = util::safe_ptr<ID3D11Texture2D, Release<ID3D11Texture2D>>;
+using texture1d_t           = util::safe_ptr<ID3D11Texture1D, Release<ID3D11Texture1D>>;
+using resource_t            = util::safe_ptr<IDXGIResource, Release<IDXGIResource>>;
+using multithread_t         = util::safe_ptr<ID3D11Multithread, Release<ID3D11Multithread>>;
+using vs_t                  = util::safe_ptr<ID3D11VertexShader, Release<ID3D11VertexShader>>;
+using ps_t                  = util::safe_ptr<ID3D11PixelShader, Release<ID3D11PixelShader>>;
+using blend_t               = util::safe_ptr<ID3D11BlendState, Release<ID3D11BlendState>>;
+using input_layout_t        = util::safe_ptr<ID3D11InputLayout, Release<ID3D11InputLayout>>;
+using render_target_t       = util::safe_ptr<ID3D11RenderTargetView, Release<ID3D11RenderTargetView>>;
+using shader_res_t          = util::safe_ptr<ID3D11ShaderResourceView, Release<ID3D11ShaderResourceView>>;
+using buf_t                 = util::safe_ptr<ID3D11Buffer, Release<ID3D11Buffer>>;
+using raster_state_t        = util::safe_ptr<ID3D11RasterizerState, Release<ID3D11RasterizerState>>;
+using sampler_state_t       = util::safe_ptr<ID3D11SamplerState, Release<ID3D11SamplerState>>;
+using blob_t                = util::safe_ptr<ID3DBlob, Release<ID3DBlob>>;
+using depth_stencil_state_t = util::safe_ptr<ID3D11DepthStencilState, Release<ID3D11DepthStencilState>>;
+using depth_stencil_view_t  = util::safe_ptr<ID3D11DepthStencilView, Release<ID3D11DepthStencilView>>;
+
+namespace video {
+using device_t         = util::safe_ptr<ID3D11VideoDevice, Release<ID3D11VideoDevice>>;
+using ctx_t            = util::safe_ptr<ID3D11VideoContext, Release<ID3D11VideoContext>>;
+using processor_t      = util::safe_ptr<ID3D11VideoProcessor, Release<ID3D11VideoProcessor>>;
+using processor_out_t  = util::safe_ptr<ID3D11VideoProcessorOutputView, Release<ID3D11VideoProcessorOutputView>>;
+using processor_in_t   = util::safe_ptr<ID3D11VideoProcessorInputView, Release<ID3D11VideoProcessorInputView>>;
+using processor_enum_t = util::safe_ptr<ID3D11VideoProcessorEnumerator, Release<ID3D11VideoProcessorEnumerator>>;
+} // namespace video
+
+class hwdevice_t;
+struct cursor_t {
+  std::vector<std::uint8_t> img_data;
+
+  DXGI_OUTDUPL_POINTER_SHAPE_INFO shape_info;
+  int x, y;
+  bool visible;
+};
+
+class gpu_cursor_t {
+public:
+  gpu_cursor_t() : cursor_view { 0, 0, 0, 0, 0.0f, 1.0f } {};
+  void set_pos(LONG rel_x, LONG rel_y, bool visible) {
+    cursor_view.TopLeftX = rel_x;
+    cursor_view.TopLeftY = rel_y;
+
+    this->visible = visible;
+  }
+
+  void set_texture(LONG width, LONG height, texture2d_t &&texture) {
+    cursor_view.Width  = width;
+    cursor_view.Height = height;
+
+    this->texture = std::move(texture);
+  }
+
+  texture2d_t texture;
+  shader_res_t input_res;
+
+  D3D11_VIEWPORT cursor_view;
+
+  bool visible;
+};
+
+class duplication_t {
+public:
+  dup_t dup;
+  bool has_frame {};
+  bool use_dwmflush {};
+
+  capture_e next_frame(DXGI_OUTDUPL_FRAME_INFO &frame_info, std::chrono::milliseconds timeout, resource_t::pointer *res_p);
+  capture_e reset(dup_t::pointer dup_p = dup_t::pointer());
+  capture_e release_frame();
+
+  ~duplication_t();
+};
+
+class display_base_t : public display_t {
+public:
+  int init(int framerate, const std::string &display_name);
+
+  std::chrono::nanoseconds delay;
+
+  factory1_t factory;
+  adapter_t adapter;
+  output_t output;
+  device_t device;
+  device_ctx_t device_ctx;
+  duplication_t dup;
+
+  DXGI_FORMAT format;
+  D3D_FEATURE_LEVEL feature_level;
+
+  typedef enum _D3DKMT_SCHEDULINGPRIORITYCLASS {
+    D3DKMT_SCHEDULINGPRIORITYCLASS_IDLE,
+    D3DKMT_SCHEDULINGPRIORITYCLASS_BELOW_NORMAL,
+    D3DKMT_SCHEDULINGPRIORITYCLASS_NORMAL,
+    D3DKMT_SCHEDULINGPRIORITYCLASS_ABOVE_NORMAL,
+    D3DKMT_SCHEDULINGPRIORITYCLASS_HIGH,
+    D3DKMT_SCHEDULINGPRIORITYCLASS_REALTIME
+  } D3DKMT_SCHEDULINGPRIORITYCLASS;
+
+  typedef NTSTATUS WINAPI (*PD3DKMTSetProcessSchedulingPriorityClass)(HANDLE, D3DKMT_SCHEDULINGPRIORITYCLASS);
+};
+
+class display_ram_t : public display_base_t {
+public:
+  capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) override;
+  capture_e snapshot(img_t *img, std::chrono::milliseconds timeout, bool cursor_visible);
+
+
+  std::shared_ptr<img_t> alloc_img() override;
+  int dummy_img(img_t *img) override;
+
+  int init(int framerate, const std::string &display_name);
+
+  cursor_t cursor;
+  D3D11_MAPPED_SUBRESOURCE img_info;
+  texture2d_t texture;
+};
+
+class display_vram_t : public display_base_t, public std::enable_shared_from_this<display_vram_t> {
+public:
+  capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) override;
+  capture_e snapshot(img_t *img, std::chrono::milliseconds timeout, bool cursor_visible);
+
+  std::shared_ptr<img_t> alloc_img() override;
+  int dummy_img(img_t *img_base) override;
+
+  int init(int framerate, const std::string &display_name);
+
+  std::shared_ptr<platf::hwdevice_t> make_hwdevice(pix_fmt_e pix_fmt) override;
+
+  sampler_state_t sampler_linear;
+
+  blend_t blend_enable;
+  blend_t blend_disable;
+
+  ps_t scene_ps;
+  vs_t scene_vs;
+
+  texture2d_t src;
+  gpu_cursor_t cursor;
+};
+} // namespace platf::dxgi
+
+#endif

src/platform/windows/display_base.cpp

@@ -0,0 +1,533 @@
+//
+// Created by loki on 1/12/20.
+//
+
+#include <cmath>
+#include <codecvt>
+
+#include "display.h"
+#include "misc.h"
+#include "src/config.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+
+namespace platf {
+using namespace std::literals;
+}
+namespace platf::dxgi {
+capture_e duplication_t::next_frame(DXGI_OUTDUPL_FRAME_INFO &frame_info, std::chrono::milliseconds timeout, resource_t::pointer *res_p) {
+  auto capture_status = release_frame();
+  if(capture_status != capture_e::ok) {
+    return capture_status;
+  }
+
+  if(use_dwmflush) {
+    DwmFlush();
+  }
+
+  auto status = dup->AcquireNextFrame(timeout.count(), &frame_info, res_p);
+
+  switch(status) {
+  case S_OK:
+    has_frame = true;
+    return capture_e::ok;
+  case DXGI_ERROR_WAIT_TIMEOUT:
+    return capture_e::timeout;
+  case WAIT_ABANDONED:
+  case DXGI_ERROR_ACCESS_LOST:
+  case DXGI_ERROR_ACCESS_DENIED:
+    return capture_e::reinit;
+  default:
+    BOOST_LOG(error) << "Couldn't acquire next frame [0x"sv << util::hex(status).to_string_view();
+    return capture_e::error;
+  }
+}
+
+capture_e duplication_t::reset(dup_t::pointer dup_p) {
+  auto capture_status = release_frame();
+
+  dup.reset(dup_p);
+
+  return capture_status;
+}
+
+capture_e duplication_t::release_frame() {
+  if(!has_frame) {
+    return capture_e::ok;
+  }
+
+  auto status = dup->ReleaseFrame();
+  switch(status) {
+  case S_OK:
+    has_frame = false;
+    return capture_e::ok;
+  case DXGI_ERROR_WAIT_TIMEOUT:
+    return capture_e::timeout;
+  case WAIT_ABANDONED:
+  case DXGI_ERROR_ACCESS_LOST:
+  case DXGI_ERROR_ACCESS_DENIED:
+    has_frame = false;
+    return capture_e::reinit;
+  default:
+    BOOST_LOG(error) << "Couldn't release frame [0x"sv << util::hex(status).to_string_view();
+    return capture_e::error;
+  }
+}
+
+duplication_t::~duplication_t() {
+  release_frame();
+}
+
+int display_base_t::init(int framerate, const std::string &display_name) {
+  /* Uncomment when use of IDXGIOutput5 is implemented
+  std::call_once(windows_cpp_once_flag, []() {
+    DECLARE_HANDLE(DPI_AWARENESS_CONTEXT);
+    const auto DPI_AWARENESS_CONTEXT_PER_MONITOR_AWARE_V2 = ((DPI_AWARENESS_CONTEXT)-4);
+
+    typedef BOOL (*User32_SetProcessDpiAwarenessContext)(DPI_AWARENESS_CONTEXT value);
+
+    auto user32 = LoadLibraryA("user32.dll");
+    auto f = (User32_SetProcessDpiAwarenessContext)GetProcAddress(user32, "SetProcessDpiAwarenessContext");
+    if(f) {
+      f(DPI_AWARENESS_CONTEXT_PER_MONITOR_AWARE_V2);
+    }
+
+    FreeLibrary(user32);
+  });
+*/
+
+  // Ensure we can duplicate the current display
+  syncThreadDesktop();
+
+  delay = std::chrono::nanoseconds { 1s } / framerate;
+
+  // Get rectangle of full desktop for absolute mouse coordinates
+  env_width  = GetSystemMetrics(SM_CXVIRTUALSCREEN);
+  env_height = GetSystemMetrics(SM_CYVIRTUALSCREEN);
+
+  HRESULT status;
+
+  status = CreateDXGIFactory1(IID_IDXGIFactory1, (void **)&factory);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create DXGIFactory1 [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
+
+  auto adapter_name = converter.from_bytes(config::video.adapter_name);
+  auto output_name  = converter.from_bytes(display_name);
+
+  adapter_t::pointer adapter_p;
+  for(int x = 0; factory->EnumAdapters1(x, &adapter_p) != DXGI_ERROR_NOT_FOUND; ++x) {
+    dxgi::adapter_t adapter_tmp { adapter_p };
+
+    DXGI_ADAPTER_DESC1 adapter_desc;
+    adapter_tmp->GetDesc1(&adapter_desc);
+
+    if(!adapter_name.empty() && adapter_desc.Description != adapter_name) {
+      continue;
+    }
+
+    dxgi::output_t::pointer output_p;
+    for(int y = 0; adapter_tmp->EnumOutputs(y, &output_p) != DXGI_ERROR_NOT_FOUND; ++y) {
+      dxgi::output_t output_tmp { output_p };
+
+      DXGI_OUTPUT_DESC desc;
+      output_tmp->GetDesc(&desc);
+
+      if(!output_name.empty() && desc.DeviceName != output_name) {
+        continue;
+      }
+
+      if(desc.AttachedToDesktop) {
+        output = std::move(output_tmp);
+
+        offset_x = desc.DesktopCoordinates.left;
+        offset_y = desc.DesktopCoordinates.top;
+        width    = desc.DesktopCoordinates.right - offset_x;
+        height   = desc.DesktopCoordinates.bottom - offset_y;
+
+        // left and bottom may be negative, yet absolute mouse coordinates start at 0x0
+        // Ensure offset starts at 0x0
+        offset_x -= GetSystemMetrics(SM_XVIRTUALSCREEN);
+        offset_y -= GetSystemMetrics(SM_YVIRTUALSCREEN);
+      }
+    }
+
+    if(output) {
+      adapter = std::move(adapter_tmp);
+      break;
+    }
+  }
+
+  if(!output) {
+    BOOST_LOG(error) << "Failed to locate an output device"sv;
+    return -1;
+  }
+
+  D3D_FEATURE_LEVEL featureLevels[] {
+    D3D_FEATURE_LEVEL_11_1,
+    D3D_FEATURE_LEVEL_11_0,
+    D3D_FEATURE_LEVEL_10_1,
+    D3D_FEATURE_LEVEL_10_0,
+    D3D_FEATURE_LEVEL_9_3,
+    D3D_FEATURE_LEVEL_9_2,
+    D3D_FEATURE_LEVEL_9_1
+  };
+
+  status = adapter->QueryInterface(IID_IDXGIAdapter, (void **)&adapter_p);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to query IDXGIAdapter interface"sv;
+    return -1;
+  }
+
+  status = D3D11CreateDevice(
+    adapter_p,
+    D3D_DRIVER_TYPE_UNKNOWN,
+    nullptr,
+    D3D11_CREATE_DEVICE_VIDEO_SUPPORT,
+    featureLevels, sizeof(featureLevels) / sizeof(D3D_FEATURE_LEVEL),
+    D3D11_SDK_VERSION,
+    &device,
+    &feature_level,
+    &device_ctx);
+
+  adapter_p->Release();
+
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create D3D11 device [0x"sv << util::hex(status).to_string_view() << ']';
+
+    return -1;
+  }
+
+  DXGI_ADAPTER_DESC adapter_desc;
+  adapter->GetDesc(&adapter_desc);
+
+  auto description = converter.to_bytes(adapter_desc.Description);
+  BOOST_LOG(info)
+    << std::endl
+    << "Device Description : " << description << std::endl
+    << "Device Vendor ID   : 0x"sv << util::hex(adapter_desc.VendorId).to_string_view() << std::endl
+    << "Device Device ID   : 0x"sv << util::hex(adapter_desc.DeviceId).to_string_view() << std::endl
+    << "Device Video Mem   : "sv << adapter_desc.DedicatedVideoMemory / 1048576 << " MiB"sv << std::endl
+    << "Device Sys Mem     : "sv << adapter_desc.DedicatedSystemMemory / 1048576 << " MiB"sv << std::endl
+    << "Share Sys Mem      : "sv << adapter_desc.SharedSystemMemory / 1048576 << " MiB"sv << std::endl
+    << "Feature Level      : 0x"sv << util::hex(feature_level).to_string_view() << std::endl
+    << "Capture size       : "sv << width << 'x' << height << std::endl
+    << "Offset             : "sv << offset_x << 'x' << offset_y << std::endl
+    << "Virtual Desktop    : "sv << env_width << 'x' << env_height;
+
+  // Enable DwmFlush() only if the current refresh rate can match the client framerate.
+  auto refresh_rate = framerate;
+  DWM_TIMING_INFO timing_info;
+  timing_info.cbSize = sizeof(timing_info);
+
+  status = DwmGetCompositionTimingInfo(NULL, &timing_info);
+  if(FAILED(status)) {
+    BOOST_LOG(warning) << "Failed to detect active refresh rate.";
+  }
+  else {
+    refresh_rate = std::round((double)timing_info.rateRefresh.uiNumerator / (double)timing_info.rateRefresh.uiDenominator);
+  }
+
+  dup.use_dwmflush = config::video.dwmflush && !(framerate > refresh_rate) ? true : false;
+
+  // Bump up thread priority
+  {
+    const DWORD flags = TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY;
+    TOKEN_PRIVILEGES tp;
+    HANDLE token;
+    LUID val;
+
+    if(OpenProcessToken(GetCurrentProcess(), flags, &token) &&
+       !!LookupPrivilegeValue(NULL, SE_INC_BASE_PRIORITY_NAME, &val)) {
+      tp.PrivilegeCount           = 1;
+      tp.Privileges[0].Luid       = val;
+      tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
+
+      if(!AdjustTokenPrivileges(token, false, &tp, sizeof(tp), NULL, NULL)) {
+        BOOST_LOG(warning) << "Could not set privilege to increase GPU priority";
+      }
+    }
+
+    CloseHandle(token);
+
+    HMODULE gdi32 = GetModuleHandleA("GDI32");
+    if(gdi32) {
+      PD3DKMTSetProcessSchedulingPriorityClass fn =
+        (PD3DKMTSetProcessSchedulingPriorityClass)GetProcAddress(gdi32, "D3DKMTSetProcessSchedulingPriorityClass");
+      if(fn) {
+        status = fn(GetCurrentProcess(), D3DKMT_SCHEDULINGPRIORITYCLASS_REALTIME);
+        if(FAILED(status)) {
+          BOOST_LOG(warning) << "Failed to set realtime GPU priority. Please run application as administrator for optimal performance.";
+        }
+      }
+    }
+
+    dxgi::dxgi_t dxgi;
+    status = device->QueryInterface(IID_IDXGIDevice, (void **)&dxgi);
+    if(FAILED(status)) {
+      BOOST_LOG(warning) << "Failed to query DXGI interface from device [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    dxgi->SetGPUThreadPriority(7);
+  }
+
+  // Try to reduce latency
+  {
+    dxgi::dxgi1_t dxgi {};
+    status = device->QueryInterface(IID_IDXGIDevice, (void **)&dxgi);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to query DXGI interface from device [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = dxgi->SetMaximumFrameLatency(1);
+    if(FAILED(status)) {
+      BOOST_LOG(warning) << "Failed to set maximum frame latency [0x"sv << util::hex(status).to_string_view() << ']';
+    }
+  }
+
+  //FIXME: Duplicate output on RX580 in combination with DOOM (2016) --> BSOD
+  //TODO: Use IDXGIOutput5 for improved performance
+  {
+    dxgi::output1_t output1 {};
+    status = output->QueryInterface(IID_IDXGIOutput1, (void **)&output1);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to query IDXGIOutput1 from the output"sv;
+      return -1;
+    }
+
+    // We try this twice, in case we still get an error on reinitialization
+    for(int x = 0; x < 2; ++x) {
+      status = output1->DuplicateOutput((IUnknown *)device.get(), &dup.dup);
+      if(SUCCEEDED(status)) {
+        break;
+      }
+      std::this_thread::sleep_for(200ms);
+    }
+
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "DuplicateOutput Failed [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+  }
+
+  DXGI_OUTDUPL_DESC dup_desc;
+  dup.dup->GetDesc(&dup_desc);
+
+  format = dup_desc.ModeDesc.Format;
+
+  BOOST_LOG(debug) << "Source format ["sv << format_str[dup_desc.ModeDesc.Format] << ']';
+
+  return 0;
+}
+
+const char *format_str[] = {
+  "DXGI_FORMAT_UNKNOWN",
+  "DXGI_FORMAT_R32G32B32A32_TYPELESS",
+  "DXGI_FORMAT_R32G32B32A32_FLOAT",
+  "DXGI_FORMAT_R32G32B32A32_UINT",
+  "DXGI_FORMAT_R32G32B32A32_SINT",
+  "DXGI_FORMAT_R32G32B32_TYPELESS",
+  "DXGI_FORMAT_R32G32B32_FLOAT",
+  "DXGI_FORMAT_R32G32B32_UINT",
+  "DXGI_FORMAT_R32G32B32_SINT",
+  "DXGI_FORMAT_R16G16B16A16_TYPELESS",
+  "DXGI_FORMAT_R16G16B16A16_FLOAT",
+  "DXGI_FORMAT_R16G16B16A16_UNORM",
+  "DXGI_FORMAT_R16G16B16A16_UINT",
+  "DXGI_FORMAT_R16G16B16A16_SNORM",
+  "DXGI_FORMAT_R16G16B16A16_SINT",
+  "DXGI_FORMAT_R32G32_TYPELESS",
+  "DXGI_FORMAT_R32G32_FLOAT",
+  "DXGI_FORMAT_R32G32_UINT",
+  "DXGI_FORMAT_R32G32_SINT",
+  "DXGI_FORMAT_R32G8X24_TYPELESS",
+  "DXGI_FORMAT_D32_FLOAT_S8X24_UINT",
+  "DXGI_FORMAT_R32_FLOAT_X8X24_TYPELESS",
+  "DXGI_FORMAT_X32_TYPELESS_G8X24_UINT",
+  "DXGI_FORMAT_R10G10B10A2_TYPELESS",
+  "DXGI_FORMAT_R10G10B10A2_UNORM",
+  "DXGI_FORMAT_R10G10B10A2_UINT",
+  "DXGI_FORMAT_R11G11B10_FLOAT",
+  "DXGI_FORMAT_R8G8B8A8_TYPELESS",
+  "DXGI_FORMAT_R8G8B8A8_UNORM",
+  "DXGI_FORMAT_R8G8B8A8_UNORM_SRGB",
+  "DXGI_FORMAT_R8G8B8A8_UINT",
+  "DXGI_FORMAT_R8G8B8A8_SNORM",
+  "DXGI_FORMAT_R8G8B8A8_SINT",
+  "DXGI_FORMAT_R16G16_TYPELESS",
+  "DXGI_FORMAT_R16G16_FLOAT",
+  "DXGI_FORMAT_R16G16_UNORM",
+  "DXGI_FORMAT_R16G16_UINT",
+  "DXGI_FORMAT_R16G16_SNORM",
+  "DXGI_FORMAT_R16G16_SINT",
+  "DXGI_FORMAT_R32_TYPELESS",
+  "DXGI_FORMAT_D32_FLOAT",
+  "DXGI_FORMAT_R32_FLOAT",
+  "DXGI_FORMAT_R32_UINT",
+  "DXGI_FORMAT_R32_SINT",
+  "DXGI_FORMAT_R24G8_TYPELESS",
+  "DXGI_FORMAT_D24_UNORM_S8_UINT",
+  "DXGI_FORMAT_R24_UNORM_X8_TYPELESS",
+  "DXGI_FORMAT_X24_TYPELESS_G8_UINT",
+  "DXGI_FORMAT_R8G8_TYPELESS",
+  "DXGI_FORMAT_R8G8_UNORM",
+  "DXGI_FORMAT_R8G8_UINT",
+  "DXGI_FORMAT_R8G8_SNORM",
+  "DXGI_FORMAT_R8G8_SINT",
+  "DXGI_FORMAT_R16_TYPELESS",
+  "DXGI_FORMAT_R16_FLOAT",
+  "DXGI_FORMAT_D16_UNORM",
+  "DXGI_FORMAT_R16_UNORM",
+  "DXGI_FORMAT_R16_UINT",
+  "DXGI_FORMAT_R16_SNORM",
+  "DXGI_FORMAT_R16_SINT",
+  "DXGI_FORMAT_R8_TYPELESS",
+  "DXGI_FORMAT_R8_UNORM",
+  "DXGI_FORMAT_R8_UINT",
+  "DXGI_FORMAT_R8_SNORM",
+  "DXGI_FORMAT_R8_SINT",
+  "DXGI_FORMAT_A8_UNORM",
+  "DXGI_FORMAT_R1_UNORM",
+  "DXGI_FORMAT_R9G9B9E5_SHAREDEXP",
+  "DXGI_FORMAT_R8G8_B8G8_UNORM",
+  "DXGI_FORMAT_G8R8_G8B8_UNORM",
+  "DXGI_FORMAT_BC1_TYPELESS",
+  "DXGI_FORMAT_BC1_UNORM",
+  "DXGI_FORMAT_BC1_UNORM_SRGB",
+  "DXGI_FORMAT_BC2_TYPELESS",
+  "DXGI_FORMAT_BC2_UNORM",
+  "DXGI_FORMAT_BC2_UNORM_SRGB",
+  "DXGI_FORMAT_BC3_TYPELESS",
+  "DXGI_FORMAT_BC3_UNORM",
+  "DXGI_FORMAT_BC3_UNORM_SRGB",
+  "DXGI_FORMAT_BC4_TYPELESS",
+  "DXGI_FORMAT_BC4_UNORM",
+  "DXGI_FORMAT_BC4_SNORM",
+  "DXGI_FORMAT_BC5_TYPELESS",
+  "DXGI_FORMAT_BC5_UNORM",
+  "DXGI_FORMAT_BC5_SNORM",
+  "DXGI_FORMAT_B5G6R5_UNORM",
+  "DXGI_FORMAT_B5G5R5A1_UNORM",
+  "DXGI_FORMAT_B8G8R8A8_UNORM",
+  "DXGI_FORMAT_B8G8R8X8_UNORM",
+  "DXGI_FORMAT_R10G10B10_XR_BIAS_A2_UNORM",
+  "DXGI_FORMAT_B8G8R8A8_TYPELESS",
+  "DXGI_FORMAT_B8G8R8A8_UNORM_SRGB",
+  "DXGI_FORMAT_B8G8R8X8_TYPELESS",
+  "DXGI_FORMAT_B8G8R8X8_UNORM_SRGB",
+  "DXGI_FORMAT_BC6H_TYPELESS",
+  "DXGI_FORMAT_BC6H_UF16",
+  "DXGI_FORMAT_BC6H_SF16",
+  "DXGI_FORMAT_BC7_TYPELESS",
+  "DXGI_FORMAT_BC7_UNORM",
+  "DXGI_FORMAT_BC7_UNORM_SRGB",
+  "DXGI_FORMAT_AYUV",
+  "DXGI_FORMAT_Y410",
+  "DXGI_FORMAT_Y416",
+  "DXGI_FORMAT_NV12",
+  "DXGI_FORMAT_P010",
+  "DXGI_FORMAT_P016",
+  "DXGI_FORMAT_420_OPAQUE",
+  "DXGI_FORMAT_YUY2",
+  "DXGI_FORMAT_Y210",
+  "DXGI_FORMAT_Y216",
+  "DXGI_FORMAT_NV11",
+  "DXGI_FORMAT_AI44",
+  "DXGI_FORMAT_IA44",
+  "DXGI_FORMAT_P8",
+  "DXGI_FORMAT_A8P8",
+  "DXGI_FORMAT_B4G4R4A4_UNORM",
+
+  NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
+
+  "DXGI_FORMAT_P208",
+  "DXGI_FORMAT_V208",
+  "DXGI_FORMAT_V408"
+};
+
+} // namespace platf::dxgi
+
+namespace platf {
+std::shared_ptr<display_t> display(mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
+  if(hwdevice_type == mem_type_e::dxgi) {
+    auto disp = std::make_shared<dxgi::display_vram_t>();
+
+    if(!disp->init(framerate, display_name)) {
+      return disp;
+    }
+  }
+  else if(hwdevice_type == mem_type_e::system) {
+    auto disp = std::make_shared<dxgi::display_ram_t>();
+
+    if(!disp->init(framerate, display_name)) {
+      return disp;
+    }
+  }
+
+  return nullptr;
+}
+
+std::vector<std::string> display_names(mem_type_e) {
+  std::vector<std::string> display_names;
+
+  HRESULT status;
+
+  BOOST_LOG(debug) << "Detecting monitors..."sv;
+
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
+
+  dxgi::factory1_t factory;
+  status = CreateDXGIFactory1(IID_IDXGIFactory1, (void **)&factory);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create DXGIFactory1 [0x"sv << util::hex(status).to_string_view() << ']' << std::endl;
+    return {};
+  }
+
+  dxgi::adapter_t adapter;
+  for(int x = 0; factory->EnumAdapters1(x, &adapter) != DXGI_ERROR_NOT_FOUND; ++x) {
+    DXGI_ADAPTER_DESC1 adapter_desc;
+    adapter->GetDesc1(&adapter_desc);
+
+    BOOST_LOG(debug)
+      << std::endl
+      << "====== ADAPTER ====="sv << std::endl
+      << "Device Name      : "sv << converter.to_bytes(adapter_desc.Description) << std::endl
+      << "Device Vendor ID : 0x"sv << util::hex(adapter_desc.VendorId).to_string_view() << std::endl
+      << "Device Device ID : 0x"sv << util::hex(adapter_desc.DeviceId).to_string_view() << std::endl
+      << "Device Video Mem : "sv << adapter_desc.DedicatedVideoMemory / 1048576 << " MiB"sv << std::endl
+      << "Device Sys Mem   : "sv << adapter_desc.DedicatedSystemMemory / 1048576 << " MiB"sv << std::endl
+      << "Share Sys Mem    : "sv << adapter_desc.SharedSystemMemory / 1048576 << " MiB"sv << std::endl
+      << std::endl
+      << "    ====== OUTPUT ======"sv << std::endl;
+
+    dxgi::output_t::pointer output_p {};
+    for(int y = 0; adapter->EnumOutputs(y, &output_p) != DXGI_ERROR_NOT_FOUND; ++y) {
+      dxgi::output_t output { output_p };
+
+      DXGI_OUTPUT_DESC desc;
+      output->GetDesc(&desc);
+
+      auto device_name = converter.to_bytes(desc.DeviceName);
+
+      auto width  = desc.DesktopCoordinates.right - desc.DesktopCoordinates.left;
+      auto height = desc.DesktopCoordinates.bottom - desc.DesktopCoordinates.top;
+
+      BOOST_LOG(debug)
+        << "    Output Name       : "sv << device_name << std::endl
+        << "    AttachedToDesktop : "sv << (desc.AttachedToDesktop ? "yes"sv : "no"sv) << std::endl
+        << "    Resolution        : "sv << width << 'x' << height << std::endl
+        << std::endl;
+
+      display_names.emplace_back(std::move(device_name));
+    }
+  }
+
+  return display_names;
+}
+
+} // namespace platf

src/platform/windows/display_ram.cpp

@@ -0,0 +1,327 @@
+#include "display.h"
+#include "src/main.h"
+
+namespace platf {
+using namespace std::literals;
+}
+
+namespace platf::dxgi {
+struct img_t : public ::platf::img_t {
+  ~img_t() override {
+    delete[] data;
+    data = nullptr;
+  }
+};
+
+void blend_cursor_monochrome(const cursor_t &cursor, img_t &img) {
+  int height = cursor.shape_info.Height / 2;
+  int width  = cursor.shape_info.Width;
+  int pitch  = cursor.shape_info.Pitch;
+
+  // img cursor.{x,y} < 0, skip parts of the cursor.img_data
+  auto cursor_skip_y = -std::min(0, cursor.y);
+  auto cursor_skip_x = -std::min(0, cursor.x);
+
+  // img cursor.{x,y} > img.{x,y}, truncate parts of the cursor.img_data
+  auto cursor_truncate_y = std::max(0, cursor.y - img.height);
+  auto cursor_truncate_x = std::max(0, cursor.x - img.width);
+
+  auto cursor_width  = width - cursor_skip_x - cursor_truncate_x;
+  auto cursor_height = height - cursor_skip_y - cursor_truncate_y;
+
+  if(cursor_height > height || cursor_width > width) {
+    return;
+  }
+
+  auto img_skip_y = std::max(0, cursor.y);
+  auto img_skip_x = std::max(0, cursor.x);
+
+  auto cursor_img_data = cursor.img_data.data() + cursor_skip_y * pitch;
+
+  int delta_height = std::min(cursor_height - cursor_truncate_y, std::max(0, img.height - img_skip_y));
+  int delta_width  = std::min(cursor_width - cursor_truncate_x, std::max(0, img.width - img_skip_x));
+
+  auto pixels_per_byte = width / pitch;
+  auto bytes_per_row   = delta_width / pixels_per_byte;
+
+  auto img_data = (int *)img.data;
+  for(int i = 0; i < delta_height; ++i) {
+    auto and_mask = &cursor_img_data[i * pitch];
+    auto xor_mask = &cursor_img_data[(i + height) * pitch];
+
+    auto img_pixel_p = &img_data[(i + img_skip_y) * (img.row_pitch / img.pixel_pitch) + img_skip_x];
+
+    auto skip_x = cursor_skip_x;
+    for(int x = 0; x < bytes_per_row; ++x) {
+      for(auto bit = 0u; bit < 8; ++bit) {
+        if(skip_x > 0) {
+          --skip_x;
+
+          continue;
+        }
+
+        int and_ = *and_mask & (1 << (7 - bit)) ? -1 : 0;
+        int xor_ = *xor_mask & (1 << (7 - bit)) ? -1 : 0;
+
+        *img_pixel_p &= and_;
+        *img_pixel_p ^= xor_;
+
+        ++img_pixel_p;
+      }
+
+      ++and_mask;
+      ++xor_mask;
+    }
+  }
+}
+
+void apply_color_alpha(int *img_pixel_p, int cursor_pixel) {
+  auto colors_out = (std::uint8_t *)&cursor_pixel;
+  auto colors_in  = (std::uint8_t *)img_pixel_p;
+
+  //TODO: When use of IDXGIOutput5 is implemented, support different color formats
+  auto alpha = colors_out[3];
+  if(alpha == 255) {
+    *img_pixel_p = cursor_pixel;
+  }
+  else {
+    colors_in[0] = colors_out[0] + (colors_in[0] * (255 - alpha) + 255 / 2) / 255;
+    colors_in[1] = colors_out[1] + (colors_in[1] * (255 - alpha) + 255 / 2) / 255;
+    colors_in[2] = colors_out[2] + (colors_in[2] * (255 - alpha) + 255 / 2) / 255;
+  }
+}
+
+void apply_color_masked(int *img_pixel_p, int cursor_pixel) {
+  //TODO: When use of IDXGIOutput5 is implemented, support different color formats
+  auto alpha = ((std::uint8_t *)&cursor_pixel)[3];
+  if(alpha == 0xFF) {
+    *img_pixel_p ^= cursor_pixel;
+  }
+  else {
+    *img_pixel_p = cursor_pixel;
+  }
+}
+
+void blend_cursor_color(const cursor_t &cursor, img_t &img, const bool masked) {
+  int height = cursor.shape_info.Height;
+  int width  = cursor.shape_info.Width;
+  int pitch  = cursor.shape_info.Pitch;
+
+  // img cursor.y < 0, skip parts of the cursor.img_data
+  auto cursor_skip_y = -std::min(0, cursor.y);
+  auto cursor_skip_x = -std::min(0, cursor.x);
+
+  // img cursor.{x,y} > img.{x,y}, truncate parts of the cursor.img_data
+  auto cursor_truncate_y = std::max(0, cursor.y - img.height);
+  auto cursor_truncate_x = std::max(0, cursor.x - img.width);
+
+  auto img_skip_y = std::max(0, cursor.y);
+  auto img_skip_x = std::max(0, cursor.x);
+
+  auto cursor_width  = width - cursor_skip_x - cursor_truncate_x;
+  auto cursor_height = height - cursor_skip_y - cursor_truncate_y;
+
+  if(cursor_height > height || cursor_width > width) {
+    return;
+  }
+
+  auto cursor_img_data = (int *)&cursor.img_data[cursor_skip_y * pitch];
+
+  int delta_height = std::min(cursor_height - cursor_truncate_y, std::max(0, img.height - img_skip_y));
+  int delta_width  = std::min(cursor_width - cursor_truncate_x, std::max(0, img.width - img_skip_x));
+
+  auto img_data = (int *)img.data;
+
+  for(int i = 0; i < delta_height; ++i) {
+    auto cursor_begin = &cursor_img_data[i * cursor.shape_info.Width + cursor_skip_x];
+    auto cursor_end   = &cursor_begin[delta_width];
+
+    auto img_pixel_p = &img_data[(i + img_skip_y) * (img.row_pitch / img.pixel_pitch) + img_skip_x];
+    std::for_each(cursor_begin, cursor_end, [&](int cursor_pixel) {
+      if(masked) {
+        apply_color_masked(img_pixel_p, cursor_pixel);
+      }
+      else {
+        apply_color_alpha(img_pixel_p, cursor_pixel);
+      }
+      ++img_pixel_p;
+    });
+  }
+}
+
+void blend_cursor(const cursor_t &cursor, img_t &img) {
+  switch(cursor.shape_info.Type) {
+  case DXGI_OUTDUPL_POINTER_SHAPE_TYPE_COLOR:
+    blend_cursor_color(cursor, img, false);
+    break;
+  case DXGI_OUTDUPL_POINTER_SHAPE_TYPE_MONOCHROME:
+    blend_cursor_monochrome(cursor, img);
+    break;
+  case DXGI_OUTDUPL_POINTER_SHAPE_TYPE_MASKED_COLOR:
+    blend_cursor_color(cursor, img, true);
+    break;
+  default:
+    BOOST_LOG(warning) << "Unsupported cursor format ["sv << cursor.shape_info.Type << ']';
+  }
+}
+
+capture_e display_ram_t::capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<::platf::img_t> img, bool *cursor) {
+  auto next_frame = std::chrono::steady_clock::now();
+
+  while(img) {
+    auto now = std::chrono::steady_clock::now();
+    while(next_frame > now) {
+      now = std::chrono::steady_clock::now();
+    }
+    next_frame = now + delay;
+
+    auto status = snapshot(img.get(), 1000ms, *cursor);
+    switch(status) {
+    case platf::capture_e::reinit:
+    case platf::capture_e::error:
+      return status;
+    case platf::capture_e::timeout:
+      std::this_thread::sleep_for(1ms);
+      continue;
+    case platf::capture_e::ok:
+      img = snapshot_cb(img);
+      break;
+    default:
+      BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+      return status;
+    }
+  }
+
+  return capture_e::ok;
+}
+
+capture_e display_ram_t::snapshot(::platf::img_t *img_base, std::chrono::milliseconds timeout, bool cursor_visible) {
+  auto img = (img_t *)img_base;
+
+  HRESULT status;
+
+  DXGI_OUTDUPL_FRAME_INFO frame_info;
+
+  resource_t::pointer res_p {};
+  auto capture_status = dup.next_frame(frame_info, timeout, &res_p);
+  resource_t res { res_p };
+
+  if(capture_status != capture_e::ok) {
+    return capture_status;
+  }
+
+  if(frame_info.PointerShapeBufferSize > 0) {
+    auto &img_data = cursor.img_data;
+
+    img_data.resize(frame_info.PointerShapeBufferSize);
+
+    UINT dummy;
+    status = dup.dup->GetFramePointerShape(img_data.size(), img_data.data(), &dummy, &cursor.shape_info);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to get new pointer shape [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return capture_e::error;
+    }
+  }
+
+  if(frame_info.LastMouseUpdateTime.QuadPart) {
+    cursor.x       = frame_info.PointerPosition.Position.x;
+    cursor.y       = frame_info.PointerPosition.Position.y;
+    cursor.visible = frame_info.PointerPosition.Visible;
+  }
+
+  // If frame has been updated
+  if(frame_info.LastPresentTime.QuadPart != 0) {
+    {
+      texture2d_t src {};
+      status = res->QueryInterface(IID_ID3D11Texture2D, (void **)&src);
+
+      if(FAILED(status)) {
+        BOOST_LOG(error) << "Couldn't query interface [0x"sv << util::hex(status).to_string_view() << ']';
+        return capture_e::error;
+      }
+
+      //Copy from GPU to CPU
+      device_ctx->CopyResource(texture.get(), src.get());
+    }
+
+    if(img_info.pData) {
+      device_ctx->Unmap(texture.get(), 0);
+      img_info.pData = nullptr;
+    }
+
+    status = device_ctx->Map(texture.get(), 0, D3D11_MAP_READ, 0, &img_info);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to map texture [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return capture_e::error;
+    }
+  }
+
+  const bool mouse_update =
+    (frame_info.LastMouseUpdateTime.QuadPart || frame_info.PointerShapeBufferSize > 0) &&
+    (cursor_visible && cursor.visible);
+
+  const bool update_flag = frame_info.LastPresentTime.QuadPart != 0 || mouse_update;
+
+  if(!update_flag) {
+    return capture_e::timeout;
+  }
+
+  std::copy_n((std::uint8_t *)img_info.pData, height * img_info.RowPitch, (std::uint8_t *)img->data);
+
+  if(cursor_visible && cursor.visible) {
+    blend_cursor(cursor, *img);
+  }
+
+  return capture_e::ok;
+}
+
+std::shared_ptr<platf::img_t> display_ram_t::alloc_img() {
+  auto img = std::make_shared<img_t>();
+
+  img->pixel_pitch = 4;
+  img->row_pitch   = img_info.RowPitch;
+  img->width       = width;
+  img->height      = height;
+  img->data        = new std::uint8_t[img->row_pitch * height];
+
+  return img;
+}
+
+int display_ram_t::dummy_img(platf::img_t *img) {
+  return 0;
+}
+
+int display_ram_t::init(int framerate, const std::string &display_name) {
+  if(display_base_t::init(framerate, display_name)) {
+    return -1;
+  }
+
+  D3D11_TEXTURE2D_DESC t {};
+  t.Width            = width;
+  t.Height           = height;
+  t.MipLevels        = 1;
+  t.ArraySize        = 1;
+  t.SampleDesc.Count = 1;
+  t.Usage            = D3D11_USAGE_STAGING;
+  t.Format           = format;
+  t.CPUAccessFlags   = D3D11_CPU_ACCESS_READ;
+
+  auto status = device->CreateTexture2D(&t, nullptr, &texture);
+
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create texture [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  // map the texture simply to get the pitch and stride
+  status = device_ctx->Map(texture.get(), 0, D3D11_MAP_READ, 0, &img_info);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to map the texture [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  return 0;
+}
+} // namespace platf::dxgi

src/platform/windows/display_vram.cpp

@@ -0,0 +1,888 @@
+#include <cmath>
+
+#include <codecvt>
+
+#include <d3dcompiler.h>
+#include <directxmath.h>
+
+extern "C" {
+#include <libavcodec/avcodec.h>
+#include <libavutil/hwcontext_d3d11va.h>
+}
+
+#include "display.h"
+#include "src/main.h"
+#include "src/video.h"
+
+
+#define SUNSHINE_SHADERS_DIR SUNSHINE_ASSETS_DIR "/shaders/directx"
+namespace platf {
+using namespace std::literals;
+}
+
+static void free_frame(AVFrame *frame) {
+  av_frame_free(&frame);
+}
+
+using frame_t = util::safe_ptr<AVFrame, free_frame>;
+
+namespace platf::dxgi {
+
+template<class T>
+buf_t make_buffer(device_t::pointer device, const T &t) {
+  static_assert(sizeof(T) % 16 == 0, "Buffer needs to be aligned on a 16-byte alignment");
+
+  D3D11_BUFFER_DESC buffer_desc {
+    sizeof(T),
+    D3D11_USAGE_IMMUTABLE,
+    D3D11_BIND_CONSTANT_BUFFER
+  };
+
+  D3D11_SUBRESOURCE_DATA init_data {
+    &t
+  };
+
+  buf_t::pointer buf_p;
+  auto status = device->CreateBuffer(&buffer_desc, &init_data, &buf_p);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create buffer: [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return buf_t { buf_p };
+}
+
+blend_t make_blend(device_t::pointer device, bool enable) {
+  D3D11_BLEND_DESC bdesc {};
+  auto &rt                 = bdesc.RenderTarget[0];
+  rt.BlendEnable           = enable;
+  rt.RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL;
+
+  if(enable) {
+    rt.BlendOp      = D3D11_BLEND_OP_ADD;
+    rt.BlendOpAlpha = D3D11_BLEND_OP_ADD;
+
+    rt.SrcBlend  = D3D11_BLEND_SRC_ALPHA;
+    rt.DestBlend = D3D11_BLEND_INV_SRC_ALPHA;
+
+    rt.SrcBlendAlpha  = D3D11_BLEND_ZERO;
+    rt.DestBlendAlpha = D3D11_BLEND_ZERO;
+  }
+
+  blend_t blend;
+  auto status = device->CreateBlendState(&bdesc, &blend);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create blend state: [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return blend;
+}
+
+blob_t convert_UV_vs_hlsl;
+blob_t convert_UV_ps_hlsl;
+blob_t scene_vs_hlsl;
+blob_t convert_Y_ps_hlsl;
+blob_t scene_ps_hlsl;
+
+struct img_d3d_t : public platf::img_t {
+  std::shared_ptr<platf::display_t> display;
+
+  shader_res_t input_res;
+  render_target_t scene_rt;
+
+  texture2d_t texture;
+
+  ~img_d3d_t() override = default;
+};
+
+util::buffer_t<std::uint8_t> make_cursor_image(util::buffer_t<std::uint8_t> &&img_data, DXGI_OUTDUPL_POINTER_SHAPE_INFO shape_info) {
+  constexpr std::uint32_t black       = 0xFF000000;
+  constexpr std::uint32_t white       = 0xFFFFFFFF;
+  constexpr std::uint32_t transparent = 0;
+
+  switch(shape_info.Type) {
+  case DXGI_OUTDUPL_POINTER_SHAPE_TYPE_MASKED_COLOR:
+    std::for_each((std::uint32_t *)std::begin(img_data), (std::uint32_t *)std::end(img_data), [](auto &pixel) {
+      if(pixel & 0xFF000000) {
+        pixel = transparent;
+      }
+    });
+  case DXGI_OUTDUPL_POINTER_SHAPE_TYPE_COLOR:
+    return std::move(img_data);
+  default:
+    break;
+  }
+
+  shape_info.Height /= 2;
+
+  util::buffer_t<std::uint8_t> cursor_img { shape_info.Width * shape_info.Height * 4 };
+
+  auto bytes       = shape_info.Pitch * shape_info.Height;
+  auto pixel_begin = (std::uint32_t *)std::begin(cursor_img);
+  auto pixel_data  = pixel_begin;
+  auto and_mask    = std::begin(img_data);
+  auto xor_mask    = std::begin(img_data) + bytes;
+
+  for(auto x = 0; x < bytes; ++x) {
+    for(auto c = 7; c >= 0; --c) {
+      auto bit        = 1 << c;
+      auto color_type = ((*and_mask & bit) ? 1 : 0) + ((*xor_mask & bit) ? 2 : 0);
+
+      switch(color_type) {
+      case 0: //black
+        *pixel_data = black;
+        break;
+      case 2: //white
+        *pixel_data = white;
+        break;
+      case 1: //transparent
+      {
+        *pixel_data = transparent;
+
+        break;
+      }
+      case 3: //inverse
+      {
+        auto top_p    = pixel_data - shape_info.Width;
+        auto left_p   = pixel_data - 1;
+        auto right_p  = pixel_data + 1;
+        auto bottom_p = pixel_data + shape_info.Width;
+
+        // Get the x coordinate of the pixel
+        auto column = (pixel_data - pixel_begin) % shape_info.Width != 0;
+
+        if(top_p >= pixel_begin && *top_p == transparent) {
+          *top_p = black;
+        }
+
+        if(column != 0 && left_p >= pixel_begin && *left_p == transparent) {
+          *left_p = black;
+        }
+
+        if(bottom_p < (std::uint32_t *)std::end(cursor_img)) {
+          *bottom_p = black;
+        }
+
+        if(column != shape_info.Width - 1) {
+          *right_p = black;
+        }
+        *pixel_data = white;
+      }
+      }
+
+      ++pixel_data;
+    }
+    ++and_mask;
+    ++xor_mask;
+  }
+
+  return cursor_img;
+}
+
+blob_t compile_shader(LPCSTR file, LPCSTR entrypoint, LPCSTR shader_model) {
+  blob_t::pointer msg_p = nullptr;
+  blob_t::pointer compiled_p;
+
+  DWORD flags = D3DCOMPILE_ENABLE_STRICTNESS;
+
+#ifndef NDEBUG
+  flags |= D3DCOMPILE_DEBUG | D3DCOMPILE_SKIP_OPTIMIZATION;
+#endif
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
+
+  auto wFile  = converter.from_bytes(file);
+  auto status = D3DCompileFromFile(wFile.c_str(), nullptr, nullptr, entrypoint, shader_model, flags, 0, &compiled_p, &msg_p);
+
+  if(msg_p) {
+    BOOST_LOG(warning) << std::string_view { (const char *)msg_p->GetBufferPointer(), msg_p->GetBufferSize() - 1 };
+    msg_p->Release();
+  }
+
+  if(status) {
+    BOOST_LOG(error) << "Couldn't compile ["sv << file << "] [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return blob_t { compiled_p };
+}
+
+blob_t compile_pixel_shader(LPCSTR file) {
+  return compile_shader(file, "main_ps", "ps_5_0");
+}
+
+blob_t compile_vertex_shader(LPCSTR file) {
+  return compile_shader(file, "main_vs", "vs_5_0");
+}
+
+int init_rt(device_t::pointer device, shader_res_t &shader_res, render_target_t &render_target, int width, int height, DXGI_FORMAT format, texture2d_t::pointer tex) {
+  D3D11_SHADER_RESOURCE_VIEW_DESC shader_resource_desc {
+    format,
+    D3D11_SRV_DIMENSION_TEXTURE2D
+  };
+  shader_resource_desc.Texture2D.MipLevels = 1;
+
+  auto status = device->CreateShaderResourceView(tex, &shader_resource_desc, &shader_res);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create render target texture for luma [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  D3D11_RENDER_TARGET_VIEW_DESC render_target_desc {
+    format,
+    D3D11_RTV_DIMENSION_TEXTURE2D
+  };
+
+  status = device->CreateRenderTargetView(tex, &render_target_desc, &render_target);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create render target view [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  return 0;
+}
+
+int init_rt(device_t::pointer device, shader_res_t &shader_res, render_target_t &render_target, int width, int height, DXGI_FORMAT format) {
+  D3D11_TEXTURE2D_DESC desc {};
+
+  desc.Width            = width;
+  desc.Height           = height;
+  desc.Format           = format;
+  desc.Usage            = D3D11_USAGE_DEFAULT;
+  desc.BindFlags        = D3D11_BIND_RENDER_TARGET | D3D11_BIND_SHADER_RESOURCE;
+  desc.MipLevels        = 1;
+  desc.ArraySize        = 1;
+  desc.SampleDesc.Count = 1;
+
+  texture2d_t tex;
+  auto status = device->CreateTexture2D(&desc, nullptr, &tex);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create render target texture for luma [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  return init_rt(device, shader_res, render_target, width, height, format, tex.get());
+}
+
+class hwdevice_t : public platf::hwdevice_t {
+public:
+  int convert(platf::img_t &img_base) override {
+    auto &img = (img_d3d_t &)img_base;
+
+    device_ctx_p->IASetInputLayout(input_layout.get());
+
+    _init_view_port(this->img.width, this->img.height);
+    device_ctx_p->OMSetRenderTargets(1, &nv12_Y_rt, nullptr);
+    device_ctx_p->VSSetShader(scene_vs.get(), nullptr, 0);
+    device_ctx_p->PSSetShader(convert_Y_ps.get(), nullptr, 0);
+    device_ctx_p->PSSetShaderResources(0, 1, &back_img.input_res);
+    device_ctx_p->Draw(3, 0);
+
+    device_ctx_p->RSSetViewports(1, &outY_view);
+    device_ctx_p->PSSetShaderResources(0, 1, &img.input_res);
+    device_ctx_p->Draw(3, 0);
+
+    // Artifacts start appearing on the rendered image if Sunshine doesn't flush
+    // before rendering on the UV part of the image.
+    device_ctx_p->Flush();
+
+    _init_view_port(this->img.width / 2, this->img.height / 2);
+    device_ctx_p->OMSetRenderTargets(1, &nv12_UV_rt, nullptr);
+    device_ctx_p->VSSetShader(convert_UV_vs.get(), nullptr, 0);
+    device_ctx_p->PSSetShader(convert_UV_ps.get(), nullptr, 0);
+    device_ctx_p->PSSetShaderResources(0, 1, &back_img.input_res);
+    device_ctx_p->Draw(3, 0);
+
+    device_ctx_p->RSSetViewports(1, &outUV_view);
+    device_ctx_p->PSSetShaderResources(0, 1, &img.input_res);
+    device_ctx_p->Draw(3, 0);
+    device_ctx_p->Flush();
+
+    return 0;
+  }
+
+  void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) override {
+    switch(colorspace) {
+    case 5: // SWS_CS_SMPTE170M
+      color_p = &::video::colors[0];
+      break;
+    case 1: // SWS_CS_ITU709
+      color_p = &::video::colors[2];
+      break;
+    case 9: // SWS_CS_BT2020
+    default:
+      BOOST_LOG(warning) << "Colorspace: ["sv << colorspace << "] not yet supported: switching to default"sv;
+      color_p = &::video::colors[0];
+    };
+
+    if(color_range > 1) {
+      // Full range
+      ++color_p;
+    }
+
+    auto color_matrix = make_buffer((device_t::pointer)data, *color_p);
+    if(!color_matrix) {
+      BOOST_LOG(warning) << "Failed to create color matrix"sv;
+      return;
+    }
+
+    device_ctx_p->PSSetConstantBuffers(0, 1, &color_matrix);
+    this->color_matrix = std::move(color_matrix);
+  }
+
+  int set_frame(AVFrame *frame) {
+    this->hwframe.reset(frame);
+    this->frame = frame;
+
+    auto device_p = (device_t::pointer)data;
+
+    auto out_width  = frame->width;
+    auto out_height = frame->height;
+
+    float in_width  = img.display->width;
+    float in_height = img.display->height;
+
+    // // Ensure aspect ratio is maintained
+    auto scalar       = std::fminf(out_width / in_width, out_height / in_height);
+    auto out_width_f  = in_width * scalar;
+    auto out_height_f = in_height * scalar;
+
+    // result is always positive
+    auto offsetX = (out_width - out_width_f) / 2;
+    auto offsetY = (out_height - out_height_f) / 2;
+
+    outY_view  = D3D11_VIEWPORT { offsetX, offsetY, out_width_f, out_height_f, 0.0f, 1.0f };
+    outUV_view = D3D11_VIEWPORT { offsetX / 2, offsetY / 2, out_width_f / 2, out_height_f / 2, 0.0f, 1.0f };
+
+    D3D11_TEXTURE2D_DESC t {};
+    t.Width            = out_width;
+    t.Height           = out_height;
+    t.MipLevels        = 1;
+    t.ArraySize        = 1;
+    t.SampleDesc.Count = 1;
+    t.Usage            = D3D11_USAGE_DEFAULT;
+    t.Format           = format;
+    t.BindFlags        = D3D11_BIND_RENDER_TARGET;
+
+    auto status = device_p->CreateTexture2D(&t, nullptr, &img.texture);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create render target texture [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    img.width       = out_width;
+    img.height      = out_height;
+    img.data        = (std::uint8_t *)img.texture.get();
+    img.row_pitch   = out_width * 4;
+    img.pixel_pitch = 4;
+
+    float info_in[16 / sizeof(float)] { 1.0f / (float)out_width }; //aligned to 16-byte
+    info_scene = make_buffer(device_p, info_in);
+
+    if(!info_in) {
+      BOOST_LOG(error) << "Failed to create info scene buffer"sv;
+      return -1;
+    }
+
+    D3D11_RENDER_TARGET_VIEW_DESC nv12_rt_desc {
+      DXGI_FORMAT_R8_UNORM,
+      D3D11_RTV_DIMENSION_TEXTURE2D
+    };
+
+    status = device_p->CreateRenderTargetView(img.texture.get(), &nv12_rt_desc, &nv12_Y_rt);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create render target view [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    nv12_rt_desc.Format = DXGI_FORMAT_R8G8_UNORM;
+
+    status = device_p->CreateRenderTargetView(img.texture.get(), &nv12_rt_desc, &nv12_UV_rt);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create render target view [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    // Need to have something refcounted
+    if(!frame->buf[0]) {
+      frame->buf[0] = av_buffer_allocz(sizeof(AVD3D11FrameDescriptor));
+    }
+
+    auto desc     = (AVD3D11FrameDescriptor *)frame->buf[0]->data;
+    desc->texture = (ID3D11Texture2D *)img.data;
+    desc->index   = 0;
+
+    frame->data[0] = img.data;
+    frame->data[1] = 0;
+
+    frame->linesize[0] = img.row_pitch;
+
+    frame->height = img.height;
+    frame->width  = img.width;
+
+    return 0;
+  }
+
+  int init(
+    std::shared_ptr<platf::display_t> display, device_t::pointer device_p, device_ctx_t::pointer device_ctx_p,
+    pix_fmt_e pix_fmt) {
+
+    HRESULT status;
+
+    device_p->AddRef();
+    data = device_p;
+
+    this->device_ctx_p = device_ctx_p;
+
+    format = (pix_fmt == pix_fmt_e::nv12 ? DXGI_FORMAT_NV12 : DXGI_FORMAT_P010);
+    status = device_p->CreateVertexShader(scene_vs_hlsl->GetBufferPointer(), scene_vs_hlsl->GetBufferSize(), nullptr, &scene_vs);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create scene vertex shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = device_p->CreatePixelShader(convert_Y_ps_hlsl->GetBufferPointer(), convert_Y_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_ps);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create convertY pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = device_p->CreatePixelShader(convert_UV_ps_hlsl->GetBufferPointer(), convert_UV_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_ps);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create convertUV pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = device_p->CreateVertexShader(convert_UV_vs_hlsl->GetBufferPointer(), convert_UV_vs_hlsl->GetBufferSize(), nullptr, &convert_UV_vs);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create convertUV vertex shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = device_p->CreatePixelShader(scene_ps_hlsl->GetBufferPointer(), scene_ps_hlsl->GetBufferSize(), nullptr, &scene_ps);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create scene pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    color_matrix = make_buffer(device_p, ::video::colors[0]);
+    if(!color_matrix) {
+      BOOST_LOG(error) << "Failed to create color matrix buffer"sv;
+      return -1;
+    }
+
+    D3D11_INPUT_ELEMENT_DESC layout_desc {
+      "SV_Position", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0
+    };
+
+    status = device_p->CreateInputLayout(
+      &layout_desc, 1,
+      convert_UV_vs_hlsl->GetBufferPointer(), convert_UV_vs_hlsl->GetBufferSize(),
+      &input_layout);
+
+    img.display = std::move(display);
+
+    // Color the background black, so that the padding for keeping the aspect ratio
+    // is black
+    if(img.display->dummy_img(&back_img)) {
+      BOOST_LOG(warning) << "Couldn't create an image to set background color to black"sv;
+      return -1;
+    }
+
+    D3D11_SHADER_RESOURCE_VIEW_DESC desc {
+      DXGI_FORMAT_B8G8R8A8_UNORM,
+      D3D11_SRV_DIMENSION_TEXTURE2D
+    };
+    desc.Texture2D.MipLevels = 1;
+
+    status = device_p->CreateShaderResourceView(back_img.texture.get(), &desc, &back_img.input_res);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create input shader resource view [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    device_ctx_p->IASetInputLayout(input_layout.get());
+    device_ctx_p->PSSetConstantBuffers(0, 1, &color_matrix);
+    device_ctx_p->VSSetConstantBuffers(0, 1, &info_scene);
+
+    return 0;
+  }
+
+  ~hwdevice_t() override {
+    if(data) {
+      ((ID3D11Device *)data)->Release();
+    }
+  }
+
+private:
+  void _init_view_port(float x, float y, float width, float height) {
+    D3D11_VIEWPORT view {
+      x, y,
+      width, height,
+      0.0f, 1.0f
+    };
+
+    device_ctx_p->RSSetViewports(1, &view);
+  }
+
+  void _init_view_port(float width, float height) {
+    _init_view_port(0.0f, 0.0f, width, height);
+  }
+
+public:
+  frame_t hwframe;
+
+  ::video::color_t *color_p;
+
+  buf_t info_scene;
+  buf_t color_matrix;
+
+  input_layout_t input_layout;
+
+  render_target_t nv12_Y_rt;
+  render_target_t nv12_UV_rt;
+
+  // The image referenced by hwframe
+  // The resulting image is stored here.
+  img_d3d_t img;
+
+  // Clear nv12 render target to black
+  img_d3d_t back_img;
+
+  vs_t convert_UV_vs;
+  ps_t convert_UV_ps;
+  ps_t convert_Y_ps;
+  ps_t scene_ps;
+  vs_t scene_vs;
+
+  D3D11_VIEWPORT outY_view;
+  D3D11_VIEWPORT outUV_view;
+
+  DXGI_FORMAT format;
+
+  device_ctx_t::pointer device_ctx_p;
+};
+
+capture_e display_vram_t::capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<::platf::img_t> img, bool *cursor) {
+  auto next_frame = std::chrono::steady_clock::now();
+
+  while(img) {
+    auto now = std::chrono::steady_clock::now();
+    while(next_frame > now) {
+      now = std::chrono::steady_clock::now();
+    }
+    next_frame = now + delay;
+
+    auto status = snapshot(img.get(), 1000ms, *cursor);
+    switch(status) {
+    case platf::capture_e::reinit:
+    case platf::capture_e::error:
+      return status;
+    case platf::capture_e::timeout:
+      std::this_thread::sleep_for(1ms);
+      continue;
+    case platf::capture_e::ok:
+      img = snapshot_cb(img);
+      break;
+    default:
+      BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+      return status;
+    }
+  }
+
+  return capture_e::ok;
+}
+
+capture_e display_vram_t::snapshot(platf::img_t *img_base, std::chrono::milliseconds timeout, bool cursor_visible) {
+  auto img = (img_d3d_t *)img_base;
+
+  HRESULT status;
+
+  DXGI_OUTDUPL_FRAME_INFO frame_info;
+
+  resource_t::pointer res_p {};
+  auto capture_status = dup.next_frame(frame_info, timeout, &res_p);
+  resource_t res { res_p };
+
+  if(capture_status != capture_e::ok) {
+    return capture_status;
+  }
+
+  const bool mouse_update_flag = frame_info.LastMouseUpdateTime.QuadPart != 0 || frame_info.PointerShapeBufferSize > 0;
+  const bool frame_update_flag = frame_info.AccumulatedFrames != 0 || frame_info.LastPresentTime.QuadPart != 0;
+  const bool update_flag       = mouse_update_flag || frame_update_flag;
+
+  if(!update_flag) {
+    return capture_e::timeout;
+  }
+
+  if(frame_info.PointerShapeBufferSize > 0) {
+    DXGI_OUTDUPL_POINTER_SHAPE_INFO shape_info {};
+
+    util::buffer_t<std::uint8_t> img_data { frame_info.PointerShapeBufferSize };
+
+    UINT dummy;
+    status = dup.dup->GetFramePointerShape(img_data.size(), std::begin(img_data), &dummy, &shape_info);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to get new pointer shape [0x"sv << util::hex(status).to_string_view() << ']';
+
+      return capture_e::error;
+    }
+
+    auto cursor_img = make_cursor_image(std::move(img_data), shape_info);
+
+    D3D11_SUBRESOURCE_DATA data {
+      std::begin(cursor_img),
+      4 * shape_info.Width,
+      0
+    };
+
+    // Create texture for cursor
+    D3D11_TEXTURE2D_DESC t {};
+    t.Width            = shape_info.Width;
+    t.Height           = cursor_img.size() / data.SysMemPitch;
+    t.MipLevels        = 1;
+    t.ArraySize        = 1;
+    t.SampleDesc.Count = 1;
+    t.Usage            = D3D11_USAGE_DEFAULT;
+    t.Format           = DXGI_FORMAT_B8G8R8A8_UNORM;
+    t.BindFlags        = D3D11_BIND_SHADER_RESOURCE;
+
+    texture2d_t texture;
+    auto status = device->CreateTexture2D(&t, &data, &texture);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create mouse texture [0x"sv << util::hex(status).to_string_view() << ']';
+      return capture_e::error;
+    }
+
+    D3D11_SHADER_RESOURCE_VIEW_DESC desc {
+      DXGI_FORMAT_B8G8R8A8_UNORM,
+      D3D11_SRV_DIMENSION_TEXTURE2D
+    };
+    desc.Texture2D.MipLevels = 1;
+
+    // Free resources before allocating on the next line.
+    cursor.input_res.reset();
+    status = device->CreateShaderResourceView(texture.get(), &desc, &cursor.input_res);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create cursor shader resource view [0x"sv << util::hex(status).to_string_view() << ']';
+      return capture_e::error;
+    }
+
+    cursor.set_texture(t.Width, t.Height, std::move(texture));
+  }
+
+  if(frame_info.LastMouseUpdateTime.QuadPart) {
+    cursor.set_pos(frame_info.PointerPosition.Position.x, frame_info.PointerPosition.Position.y, frame_info.PointerPosition.Visible && cursor_visible);
+  }
+
+  if(frame_update_flag) {
+    src.reset();
+    status = res->QueryInterface(IID_ID3D11Texture2D, (void **)&src);
+
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Couldn't query interface [0x"sv << util::hex(status).to_string_view() << ']';
+      return capture_e::error;
+    }
+  }
+
+  device_ctx->CopyResource(img->texture.get(), src.get());
+  if(cursor.visible) {
+    D3D11_VIEWPORT view {
+      0.0f, 0.0f,
+      (float)width, (float)height,
+      0.0f, 1.0f
+    };
+
+    device_ctx->VSSetShader(scene_vs.get(), nullptr, 0);
+    device_ctx->PSSetShader(scene_ps.get(), nullptr, 0);
+    device_ctx->RSSetViewports(1, &view);
+    device_ctx->OMSetRenderTargets(1, &img->scene_rt, nullptr);
+    device_ctx->PSSetShaderResources(0, 1, &cursor.input_res);
+    device_ctx->OMSetBlendState(blend_enable.get(), nullptr, 0xFFFFFFFFu);
+    device_ctx->RSSetViewports(1, &cursor.cursor_view);
+    device_ctx->Draw(3, 0);
+    device_ctx->OMSetBlendState(blend_disable.get(), nullptr, 0xFFFFFFFFu);
+  }
+
+  return capture_e::ok;
+}
+
+int display_vram_t::init(int framerate, const std::string &display_name) {
+  if(display_base_t::init(framerate, display_name)) {
+    return -1;
+  }
+
+  D3D11_SAMPLER_DESC sampler_desc {};
+  sampler_desc.Filter         = D3D11_FILTER_MIN_MAG_MIP_LINEAR;
+  sampler_desc.AddressU       = D3D11_TEXTURE_ADDRESS_CLAMP;
+  sampler_desc.AddressV       = D3D11_TEXTURE_ADDRESS_CLAMP;
+  sampler_desc.AddressW       = D3D11_TEXTURE_ADDRESS_WRAP;
+  sampler_desc.ComparisonFunc = D3D11_COMPARISON_NEVER;
+  sampler_desc.MinLOD         = 0;
+  sampler_desc.MaxLOD         = D3D11_FLOAT32_MAX;
+
+  auto status = device->CreateSamplerState(&sampler_desc, &sampler_linear);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create point sampler state [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  status = device->CreateVertexShader(scene_vs_hlsl->GetBufferPointer(), scene_vs_hlsl->GetBufferSize(), nullptr, &scene_vs);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create scene vertex shader [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  status = device->CreatePixelShader(scene_ps_hlsl->GetBufferPointer(), scene_ps_hlsl->GetBufferSize(), nullptr, &scene_ps);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create scene pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  blend_enable  = make_blend(device.get(), true);
+  blend_disable = make_blend(device.get(), false);
+
+  if(!blend_disable || !blend_enable) {
+    return -1;
+  }
+
+  device_ctx->OMSetBlendState(blend_disable.get(), nullptr, 0xFFFFFFFFu);
+  device_ctx->PSSetSamplers(0, 1, &sampler_linear);
+  device_ctx->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+
+  return 0;
+}
+
+std::shared_ptr<platf::img_t> display_vram_t::alloc_img() {
+  auto img = std::make_shared<img_d3d_t>();
+
+  img->pixel_pitch = 4;
+  img->row_pitch   = img->pixel_pitch * width;
+  img->width       = width;
+  img->height      = height;
+  img->display     = shared_from_this();
+
+  auto dummy_data = std::make_unique<std::uint8_t[]>(img->row_pitch * height);
+  D3D11_SUBRESOURCE_DATA data {
+    dummy_data.get(),
+    (UINT)img->row_pitch
+  };
+  std::fill_n(dummy_data.get(), img->row_pitch * height, 0);
+
+  D3D11_TEXTURE2D_DESC t {};
+  t.Width            = width;
+  t.Height           = height;
+  t.MipLevels        = 1;
+  t.ArraySize        = 1;
+  t.SampleDesc.Count = 1;
+  t.Usage            = D3D11_USAGE_DEFAULT;
+  t.Format           = format;
+  t.BindFlags        = D3D11_BIND_SHADER_RESOURCE | D3D11_BIND_RENDER_TARGET;
+
+  auto status = device->CreateTexture2D(&t, &data, &img->texture);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create img buf texture [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  if(init_rt(device.get(), img->input_res, img->scene_rt, width, height, format, img->texture.get())) {
+    return nullptr;
+  }
+
+  img->data = (std::uint8_t *)img->texture.get();
+
+  return img;
+}
+
+int display_vram_t::dummy_img(platf::img_t *img_base) {
+  auto img = (img_d3d_t *)img_base;
+
+  if(img->texture) {
+    return 0;
+  }
+
+  img->row_pitch  = width * 4;
+  auto dummy_data = std::make_unique<int[]>(width * height);
+  D3D11_SUBRESOURCE_DATA data {
+    dummy_data.get(),
+    (UINT)img->row_pitch
+  };
+  std::fill_n(dummy_data.get(), width * height, 0);
+
+  D3D11_TEXTURE2D_DESC t {};
+  t.Width            = width;
+  t.Height           = height;
+  t.MipLevels        = 1;
+  t.ArraySize        = 1;
+  t.SampleDesc.Count = 1;
+  t.Usage            = D3D11_USAGE_DEFAULT;
+  t.Format           = format;
+  t.BindFlags        = D3D11_BIND_SHADER_RESOURCE;
+
+  dxgi::texture2d_t tex;
+  auto status = device->CreateTexture2D(&t, &data, &tex);
+  if(FAILED(status)) {
+    BOOST_LOG(error) << "Failed to create dummy texture [0x"sv << util::hex(status).to_string_view() << ']';
+    return -1;
+  }
+
+  img->texture = std::move(tex);
+  img->data    = (std::uint8_t *)img->texture.get();
+
+  return 0;
+}
+
+std::shared_ptr<platf::hwdevice_t> display_vram_t::make_hwdevice(pix_fmt_e pix_fmt) {
+  if(pix_fmt != platf::pix_fmt_e::nv12) {
+    BOOST_LOG(error) << "display_vram_t doesn't support pixel format ["sv << from_pix_fmt(pix_fmt) << ']';
+
+    return nullptr;
+  }
+
+  auto hwdevice = std::make_shared<hwdevice_t>();
+
+  auto ret = hwdevice->init(
+    shared_from_this(),
+    device.get(),
+    device_ctx.get(),
+    pix_fmt);
+
+  if(ret) {
+    return nullptr;
+  }
+
+  return hwdevice;
+}
+
+int init() {
+  BOOST_LOG(info) << "Compiling shaders..."sv;
+  scene_vs_hlsl = compile_vertex_shader(SUNSHINE_SHADERS_DIR "/SceneVS.hlsl");
+  if(!scene_vs_hlsl) {
+    return -1;
+  }
+
+  convert_Y_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertYPS.hlsl");
+  if(!convert_Y_ps_hlsl) {
+    return -1;
+  }
+
+  convert_UV_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertUVPS.hlsl");
+  if(!convert_UV_ps_hlsl) {
+    return -1;
+  }
+
+  convert_UV_vs_hlsl = compile_vertex_shader(SUNSHINE_SHADERS_DIR "/ConvertUVVS.hlsl");
+  if(!convert_UV_vs_hlsl) {
+    return -1;
+  }
+
+  scene_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ScenePS.hlsl");
+  if(!scene_ps_hlsl) {
+    return -1;
+  }
+  BOOST_LOG(info) << "Compiled shaders"sv;
+
+  return 0;
+}
+} // namespace platf::dxgi

src/platform/windows/input.cpp

@@ -0,0 +1,491 @@
+#include <windows.h>
+
+#include <cmath>
+
+#include <ViGEm/Client.h>
+
+#include "misc.h"
+#include "src/config.h"
+#include "src/main.h"
+#include "src/platform/common.h"
+
+namespace platf {
+using namespace std::literals;
+
+thread_local HDESK _lastKnownInputDesktop = nullptr;
+
+constexpr touch_port_t target_touch_port {
+  0, 0,
+  65535, 65535
+};
+
+using client_t = util::safe_ptr<_VIGEM_CLIENT_T, vigem_free>;
+using target_t = util::safe_ptr<_VIGEM_TARGET_T, vigem_target_free>;
+
+static VIGEM_TARGET_TYPE map(const std::string_view &gp) {
+  if(gp == "x360"sv) {
+    return Xbox360Wired;
+  }
+
+  return DualShock4Wired;
+}
+
+void CALLBACK x360_notify(
+  client_t::pointer client,
+  target_t::pointer target,
+  std::uint8_t largeMotor, std::uint8_t smallMotor,
+  std::uint8_t /* led_number */,
+  void *userdata);
+
+void CALLBACK ds4_notify(
+  client_t::pointer client,
+  target_t::pointer target,
+  std::uint8_t largeMotor, std::uint8_t smallMotor,
+  DS4_LIGHTBAR_COLOR /* led_color */,
+  void *userdata);
+
+class vigem_t {
+public:
+  int init() {
+    VIGEM_ERROR status;
+
+    client.reset(vigem_alloc());
+
+    status = vigem_connect(client.get());
+    if(!VIGEM_SUCCESS(status)) {
+      BOOST_LOG(warning) << "Couldn't setup connection to ViGEm for gamepad support ["sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    gamepads.resize(MAX_GAMEPADS);
+
+    return 0;
+  }
+
+  int alloc_gamepad_interal(int nr, rumble_queue_t &rumble_queue, VIGEM_TARGET_TYPE gp_type) {
+    auto &[rumble, gp] = gamepads[nr];
+    assert(!gp);
+
+    if(gp_type == Xbox360Wired) {
+      gp.reset(vigem_target_x360_alloc());
+    }
+    else {
+      gp.reset(vigem_target_ds4_alloc());
+    }
+
+    auto status = vigem_target_add(client.get(), gp.get());
+    if(!VIGEM_SUCCESS(status)) {
+      BOOST_LOG(error) << "Couldn't add Gamepad to ViGEm connection ["sv << util::hex(status).to_string_view() << ']';
+
+      return -1;
+    }
+
+    rumble = std::move(rumble_queue);
+
+    if(gp_type == Xbox360Wired) {
+      status = vigem_target_x360_register_notification(client.get(), gp.get(), x360_notify, this);
+    }
+    else {
+      status = vigem_target_ds4_register_notification(client.get(), gp.get(), ds4_notify, this);
+    }
+
+    if(!VIGEM_SUCCESS(status)) {
+      BOOST_LOG(warning) << "Couldn't register notifications for rumble support ["sv << util::hex(status).to_string_view() << ']';
+    }
+
+    return 0;
+  }
+
+  void free_target(int nr) {
+    auto &[_, gp] = gamepads[nr];
+
+    if(gp && vigem_target_is_attached(gp.get())) {
+      auto status = vigem_target_remove(client.get(), gp.get());
+      if(!VIGEM_SUCCESS(status)) {
+        BOOST_LOG(warning) << "Couldn't detach gamepad from ViGEm ["sv << util::hex(status).to_string_view() << ']';
+      }
+    }
+
+    gp.reset();
+  }
+
+  void rumble(target_t::pointer target, std::uint8_t smallMotor, std::uint8_t largeMotor) {
+    for(int x = 0; x < gamepads.size(); ++x) {
+      auto &[rumble_queue, gp] = gamepads[x];
+
+      if(gp.get() == target) {
+        rumble_queue->raise(x, ((std::uint16_t)smallMotor) << 8, ((std::uint16_t)largeMotor) << 8);
+
+        return;
+      }
+    }
+  }
+
+  ~vigem_t() {
+    if(client) {
+      for(auto &[_, gp] : gamepads) {
+        if(gp && vigem_target_is_attached(gp.get())) {
+          auto status = vigem_target_remove(client.get(), gp.get());
+          if(!VIGEM_SUCCESS(status)) {
+            BOOST_LOG(warning) << "Couldn't detach gamepad from ViGEm ["sv << util::hex(status).to_string_view() << ']';
+          }
+        }
+      }
+
+      vigem_disconnect(client.get());
+    }
+  }
+
+  std::vector<std::pair<rumble_queue_t, target_t>> gamepads;
+
+  client_t client;
+};
+
+void CALLBACK x360_notify(
+  client_t::pointer client,
+  target_t::pointer target,
+  std::uint8_t largeMotor, std::uint8_t smallMotor,
+  std::uint8_t /* led_number */,
+  void *userdata) {
+
+  BOOST_LOG(debug)
+    << "largeMotor: "sv << (int)largeMotor << std::endl
+    << "smallMotor: "sv << (int)smallMotor;
+
+  task_pool.push(&vigem_t::rumble, (vigem_t *)userdata, target, smallMotor, largeMotor);
+}
+
+void CALLBACK ds4_notify(
+  client_t::pointer client,
+  target_t::pointer target,
+  std::uint8_t largeMotor, std::uint8_t smallMotor,
+  DS4_LIGHTBAR_COLOR /* led_color */,
+  void *userdata) {
+
+  BOOST_LOG(debug)
+    << "largeMotor: "sv << (int)largeMotor << std::endl
+    << "smallMotor: "sv << (int)smallMotor;
+
+  task_pool.push(&vigem_t::rumble, (vigem_t *)userdata, target, smallMotor, largeMotor);
+}
+
+input_t input() {
+  input_t result { new vigem_t {} };
+
+  auto vigem = (vigem_t *)result.get();
+  if(vigem->init()) {
+    return nullptr;
+  }
+
+  return result;
+}
+
+void send_input(INPUT &i) {
+retry:
+  auto send = SendInput(1, &i, sizeof(INPUT));
+  if(send != 1) {
+    auto hDesk = syncThreadDesktop();
+    if(_lastKnownInputDesktop != hDesk) {
+      _lastKnownInputDesktop = hDesk;
+      goto retry;
+    }
+    BOOST_LOG(error) << "Couldn't send input"sv;
+  }
+}
+
+void abs_mouse(input_t &input, const touch_port_t &touch_port, float x, float y) {
+  INPUT i {};
+
+  i.type   = INPUT_MOUSE;
+  auto &mi = i.mi;
+
+  mi.dwFlags =
+    MOUSEEVENTF_MOVE |
+    MOUSEEVENTF_ABSOLUTE |
+
+    // MOUSEEVENTF_VIRTUALDESK maps to the entirety of the desktop rather than the primary desktop
+    MOUSEEVENTF_VIRTUALDESK;
+
+  auto scaled_x = std::lround((x + touch_port.offset_x) * ((float)target_touch_port.width / (float)touch_port.width));
+  auto scaled_y = std::lround((y + touch_port.offset_y) * ((float)target_touch_port.height / (float)touch_port.height));
+
+  mi.dx = scaled_x;
+  mi.dy = scaled_y;
+
+  send_input(i);
+}
+
+void move_mouse(input_t &input, int deltaX, int deltaY) {
+  INPUT i {};
+
+  i.type   = INPUT_MOUSE;
+  auto &mi = i.mi;
+
+  mi.dwFlags = MOUSEEVENTF_MOVE;
+  mi.dx      = deltaX;
+  mi.dy      = deltaY;
+
+  send_input(i);
+}
+
+void button_mouse(input_t &input, int button, bool release) {
+  constexpr auto KEY_STATE_DOWN = (SHORT)0x8000;
+
+  INPUT i {};
+
+  i.type   = INPUT_MOUSE;
+  auto &mi = i.mi;
+
+  int mouse_button;
+  if(button == 1) {
+    mi.dwFlags   = release ? MOUSEEVENTF_LEFTUP : MOUSEEVENTF_LEFTDOWN;
+    mouse_button = VK_LBUTTON;
+  }
+  else if(button == 2) {
+    mi.dwFlags   = release ? MOUSEEVENTF_MIDDLEUP : MOUSEEVENTF_MIDDLEDOWN;
+    mouse_button = VK_MBUTTON;
+  }
+  else if(button == 3) {
+    mi.dwFlags   = release ? MOUSEEVENTF_RIGHTUP : MOUSEEVENTF_RIGHTDOWN;
+    mouse_button = VK_RBUTTON;
+  }
+  else if(button == 4) {
+    mi.dwFlags   = release ? MOUSEEVENTF_XUP : MOUSEEVENTF_XDOWN;
+    mi.mouseData = XBUTTON1;
+    mouse_button = VK_XBUTTON1;
+  }
+  else {
+    mi.dwFlags   = release ? MOUSEEVENTF_XUP : MOUSEEVENTF_XDOWN;
+    mi.mouseData = XBUTTON2;
+    mouse_button = VK_XBUTTON2;
+  }
+
+  auto key_state      = GetAsyncKeyState(mouse_button);
+  bool key_state_down = (key_state & KEY_STATE_DOWN) != 0;
+  if(key_state_down != release) {
+    BOOST_LOG(warning) << "Button state of mouse_button ["sv << button << "] does not match the desired state"sv;
+
+    return;
+  }
+
+  send_input(i);
+}
+
+void scroll(input_t &input, int distance) {
+  INPUT i {};
+
+  i.type   = INPUT_MOUSE;
+  auto &mi = i.mi;
+
+  mi.dwFlags   = MOUSEEVENTF_WHEEL;
+  mi.mouseData = distance;
+
+  send_input(i);
+}
+
+void keyboard(input_t &input, uint16_t modcode, bool release) {
+  INPUT i {};
+  i.type   = INPUT_KEYBOARD;
+  auto &ki = i.ki;
+
+  // For some reason, MapVirtualKey(VK_LWIN, MAPVK_VK_TO_VSC) doesn't seem to work :/
+  if(modcode != VK_LWIN && modcode != VK_RWIN && modcode != VK_PAUSE) {
+    ki.wScan   = MapVirtualKey(modcode, MAPVK_VK_TO_VSC);
+    ki.dwFlags = KEYEVENTF_SCANCODE;
+  }
+  else {
+    ki.wVk = modcode;
+  }
+
+  // https://docs.microsoft.com/en-us/windows/win32/inputdev/about-keyboard-input#keystroke-message-flags
+  switch(modcode) {
+  case VK_RMENU:
+  case VK_RCONTROL:
+  case VK_INSERT:
+  case VK_DELETE:
+  case VK_HOME:
+  case VK_END:
+  case VK_PRIOR:
+  case VK_NEXT:
+  case VK_UP:
+  case VK_DOWN:
+  case VK_LEFT:
+  case VK_RIGHT:
+  case VK_DIVIDE:
+    ki.dwFlags |= KEYEVENTF_EXTENDEDKEY;
+    break;
+  default:
+    break;
+  }
+
+  if(release) {
+    ki.dwFlags |= KEYEVENTF_KEYUP;
+  }
+
+  send_input(i);
+}
+
+int alloc_gamepad(input_t &input, int nr, rumble_queue_t rumble_queue) {
+  if(!input) {
+    return 0;
+  }
+
+  return ((vigem_t *)input.get())->alloc_gamepad_interal(nr, rumble_queue, map(config::input.gamepad));
+}
+
+void free_gamepad(input_t &input, int nr) {
+  if(!input) {
+    return;
+  }
+
+  ((vigem_t *)input.get())->free_target(nr);
+}
+
+static VIGEM_ERROR x360_update(client_t::pointer client, target_t::pointer gp, const gamepad_state_t &gamepad_state) {
+  auto &xusb = *(PXUSB_REPORT)&gamepad_state;
+
+  return vigem_target_x360_update(client, gp, xusb);
+}
+
+static DS4_DPAD_DIRECTIONS ds4_dpad(const gamepad_state_t &gamepad_state) {
+  auto flags = gamepad_state.buttonFlags;
+  if(flags & DPAD_UP) {
+    if(flags & DPAD_RIGHT) {
+      return DS4_BUTTON_DPAD_NORTHEAST;
+    }
+    else if(flags & DPAD_LEFT) {
+      return DS4_BUTTON_DPAD_NORTHWEST;
+    }
+    else {
+      return DS4_BUTTON_DPAD_NORTH;
+    }
+  }
+
+  else if(flags & DPAD_DOWN) {
+    if(flags & DPAD_RIGHT) {
+      return DS4_BUTTON_DPAD_SOUTHEAST;
+    }
+    else if(flags & DPAD_LEFT) {
+      return DS4_BUTTON_DPAD_SOUTHWEST;
+    }
+    else {
+      return DS4_BUTTON_DPAD_SOUTH;
+    }
+  }
+
+  else if(flags & DPAD_RIGHT) {
+    return DS4_BUTTON_DPAD_EAST;
+  }
+
+  else if(flags & DPAD_LEFT) {
+    return DS4_BUTTON_DPAD_WEST;
+  }
+
+  return DS4_BUTTON_DPAD_NONE;
+}
+
+static DS4_BUTTONS ds4_buttons(const gamepad_state_t &gamepad_state) {
+  int buttons {};
+
+  auto flags = gamepad_state.buttonFlags;
+  // clang-format off
+  if(flags & LEFT_STICK)   buttons |= DS4_BUTTON_THUMB_LEFT;
+  if(flags & RIGHT_STICK)  buttons |= DS4_BUTTON_THUMB_RIGHT;
+  if(flags & LEFT_BUTTON)  buttons |= DS4_BUTTON_SHOULDER_LEFT;
+  if(flags & RIGHT_BUTTON) buttons |= DS4_BUTTON_SHOULDER_RIGHT;
+  if(flags & START)        buttons |= DS4_BUTTON_OPTIONS;
+  if(flags & A)            buttons |= DS4_BUTTON_CROSS;
+  if(flags & B)            buttons |= DS4_BUTTON_CIRCLE;
+  if(flags & X)            buttons |= DS4_BUTTON_SQUARE;
+  if(flags & Y)            buttons |= DS4_BUTTON_TRIANGLE;
+
+  if(gamepad_state.lt > 0) buttons |= DS4_BUTTON_TRIGGER_LEFT;
+  if(gamepad_state.rt > 0) buttons |= DS4_BUTTON_TRIGGER_RIGHT;
+  // clang-format on
+
+  return (DS4_BUTTONS)buttons;
+}
+
+static DS4_SPECIAL_BUTTONS ds4_special_buttons(const gamepad_state_t &gamepad_state) {
+  int buttons {};
+
+  if(gamepad_state.buttonFlags & BACK) buttons |= DS4_SPECIAL_BUTTON_TOUCHPAD;
+  if(gamepad_state.buttonFlags & HOME) buttons |= DS4_SPECIAL_BUTTON_PS;
+
+  return (DS4_SPECIAL_BUTTONS)buttons;
+}
+
+static std::uint8_t to_ds4_triggerX(std::int16_t v) {
+  return (v + std::numeric_limits<std::uint16_t>::max() / 2 + 1) / 257;
+}
+
+static std::uint8_t to_ds4_triggerY(std::int16_t v) {
+  auto new_v = -((std::numeric_limits<std::uint16_t>::max() / 2 + v - 1)) / 257;
+
+  return new_v == 0 ? 0xFF : (std::uint8_t)new_v;
+}
+
+static VIGEM_ERROR ds4_update(client_t::pointer client, target_t::pointer gp, const gamepad_state_t &gamepad_state) {
+  DS4_REPORT report;
+
+  DS4_REPORT_INIT(&report);
+  DS4_SET_DPAD(&report, ds4_dpad(gamepad_state));
+  report.wButtons |= ds4_buttons(gamepad_state);
+  report.bSpecial = ds4_special_buttons(gamepad_state);
+
+  report.bTriggerL = gamepad_state.lt;
+  report.bTriggerR = gamepad_state.rt;
+
+  report.bThumbLX = to_ds4_triggerX(gamepad_state.lsX);
+  report.bThumbLY = to_ds4_triggerY(gamepad_state.lsY);
+
+  report.bThumbRX = to_ds4_triggerX(gamepad_state.rsX);
+  report.bThumbRY = to_ds4_triggerY(gamepad_state.rsY);
+
+  return vigem_target_ds4_update(client, gp, report);
+}
+
+
+void gamepad(input_t &input, int nr, const gamepad_state_t &gamepad_state) {
+  // If there is no gamepad support
+  if(!input) {
+    return;
+  }
+
+  auto vigem = (vigem_t *)input.get();
+
+  auto &[_, gp] = vigem->gamepads[nr];
+
+  VIGEM_ERROR status;
+
+  if(vigem_target_get_type(gp.get()) == Xbox360Wired) {
+    status = x360_update(vigem->client.get(), gp.get(), gamepad_state);
+  }
+  else {
+    status = ds4_update(vigem->client.get(), gp.get(), gamepad_state);
+  }
+
+  if(!VIGEM_SUCCESS(status)) {
+    BOOST_LOG(fatal) << "Couldn't send gamepad input to ViGEm ["sv << util::hex(status).to_string_view() << ']';
+
+    log_flush();
+    std::abort();
+  }
+}
+
+void freeInput(void *p) {
+  auto vigem = (vigem_t *)p;
+
+  delete vigem;
+}
+
+std::vector<std::string_view> &supported_gamepads() {
+  // ds4 == ps4
+  static std::vector<std::string_view> gps {
+    "x360"sv, "ds4"sv, "ps4"sv
+  };
+
+  return gps;
+}
+} // namespace platf

src/platform/windows/misc.cpp

@@ -0,0 +1,123 @@
+#include <filesystem>
+#include <iomanip>
+#include <sstream>
+
+
+// prevent clang format from "optimizing" the header include order
+// clang-format off
+#include <winsock2.h>
+#include <iphlpapi.h>
+#include <windows.h>
+#include <winuser.h>
+#include <ws2tcpip.h>
+// clang-format on
+
+#include "src/main.h"
+#include "src/utility.h"
+
+using namespace std::literals;
+namespace platf {
+using adapteraddrs_t = util::c_ptr<IP_ADAPTER_ADDRESSES>;
+
+std::filesystem::path appdata() {
+  return L"."sv;
+}
+
+std::string from_sockaddr(const sockaddr *const socket_address) {
+  char data[INET6_ADDRSTRLEN];
+
+  auto family = socket_address->sa_family;
+  if(family == AF_INET6) {
+    inet_ntop(AF_INET6, &((sockaddr_in6 *)socket_address)->sin6_addr, data, INET6_ADDRSTRLEN);
+  }
+
+  if(family == AF_INET) {
+    inet_ntop(AF_INET, &((sockaddr_in *)socket_address)->sin_addr, data, INET_ADDRSTRLEN);
+  }
+
+  return std::string { data };
+}
+
+std::pair<std::uint16_t, std::string> from_sockaddr_ex(const sockaddr *const ip_addr) {
+  char data[INET6_ADDRSTRLEN];
+
+  auto family = ip_addr->sa_family;
+  std::uint16_t port;
+  if(family == AF_INET6) {
+    inet_ntop(AF_INET6, &((sockaddr_in6 *)ip_addr)->sin6_addr, data, INET6_ADDRSTRLEN);
+    port = ((sockaddr_in6 *)ip_addr)->sin6_port;
+  }
+
+  if(family == AF_INET) {
+    inet_ntop(AF_INET, &((sockaddr_in *)ip_addr)->sin_addr, data, INET_ADDRSTRLEN);
+    port = ((sockaddr_in *)ip_addr)->sin_port;
+  }
+
+  return { port, std::string { data } };
+}
+
+adapteraddrs_t get_adapteraddrs() {
+  adapteraddrs_t info { nullptr };
+  ULONG size = 0;
+
+  while(GetAdaptersAddresses(AF_UNSPEC, 0, nullptr, info.get(), &size) == ERROR_BUFFER_OVERFLOW) {
+    info.reset((PIP_ADAPTER_ADDRESSES)malloc(size));
+  }
+
+  return info;
+}
+
+std::string get_mac_address(const std::string_view &address) {
+  adapteraddrs_t info = get_adapteraddrs();
+  for(auto adapter_pos = info.get(); adapter_pos != nullptr; adapter_pos = adapter_pos->Next) {
+    for(auto addr_pos = adapter_pos->FirstUnicastAddress; addr_pos != nullptr; addr_pos = addr_pos->Next) {
+      if(adapter_pos->PhysicalAddressLength != 0 && address == from_sockaddr(addr_pos->Address.lpSockaddr)) {
+        std::stringstream mac_addr;
+        mac_addr << std::hex;
+        for(int i = 0; i < adapter_pos->PhysicalAddressLength; i++) {
+          if(i > 0) {
+            mac_addr << ':';
+          }
+          mac_addr << std::setw(2) << std::setfill('0') << (int)adapter_pos->PhysicalAddress[i];
+        }
+        return mac_addr.str();
+      }
+    }
+  }
+  BOOST_LOG(warning) << "Unable to find MAC address for "sv << address;
+  return "00:00:00:00:00:00"s;
+}
+
+HDESK syncThreadDesktop() {
+  auto hDesk = OpenInputDesktop(DF_ALLOWOTHERACCOUNTHOOK, FALSE, GENERIC_ALL);
+  if(!hDesk) {
+    auto err = GetLastError();
+    BOOST_LOG(error) << "Failed to Open Input Desktop [0x"sv << util::hex(err).to_string_view() << ']';
+
+    return nullptr;
+  }
+
+  if(!SetThreadDesktop(hDesk)) {
+    auto err = GetLastError();
+    BOOST_LOG(error) << "Failed to sync desktop to thread [0x"sv << util::hex(err).to_string_view() << ']';
+  }
+
+  CloseDesktop(hDesk);
+
+  return hDesk;
+}
+
+void print_status(const std::string_view &prefix, HRESULT status) {
+  char err_string[1024];
+
+  DWORD bytes = FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+    nullptr,
+    status,
+    MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
+    err_string,
+    sizeof(err_string),
+    nullptr);
+
+  BOOST_LOG(error) << prefix << ": "sv << std::string_view { err_string, bytes };
+}
+} // namespace platf

src/platform/windows/misc.h

@@ -0,0 +1,13 @@
+#ifndef SUNSHINE_WINDOWS_MISC_H
+#define SUNSHINE_WINDOWS_MISC_H
+
+#include <string_view>
+#include <windows.h>
+#include <winnt.h>
+
+namespace platf {
+void print_status(const std::string_view &prefix, HRESULT status);
+HDESK syncThreadDesktop();
+} // namespace platf
+
+#endif

src/platform/windows/publish.cpp

@@ -0,0 +1,195 @@
+#include <winsock2.h>
+
+#include <windows.h>
+
+#include <windns.h>
+#include <winerror.h>
+
+#include <boost/asio/ip/host_name.hpp>
+
+#include "misc.h"
+#include "src/config.h"
+#include "src/main.h"
+#include "src/network.h"
+#include "src/nvhttp.h"
+#include "src/platform/common.h"
+#include "src/thread_safe.h"
+
+#define _FN(x, ret, args)    \
+  typedef ret(*x##_fn) args; \
+  static x##_fn x
+
+using namespace std::literals;
+
+#define __SV(quote) L##quote##sv
+#define SV(quote) __SV(quote)
+
+extern "C" {
+#ifndef __MINGW32__
+constexpr auto DNS_REQUEST_PENDING        = 9506L;
+constexpr auto DNS_QUERY_REQUEST_VERSION1 = 0x1;
+constexpr auto DNS_QUERY_RESULTS_VERSION1 = 0x1;
+#endif
+
+#define SERVICE_DOMAIN "local"
+
+constexpr auto SERVICE_INSTANCE_NAME = SV(SERVICE_NAME "." SERVICE_TYPE "." SERVICE_DOMAIN);
+constexpr auto SERVICE_TYPE_DOMAIN   = SV(SERVICE_TYPE "." SERVICE_DOMAIN);
+
+#ifndef __MINGW32__
+typedef struct _DNS_SERVICE_INSTANCE {
+  LPWSTR pszInstanceName;
+  LPWSTR pszHostName;
+
+  IP4_ADDRESS *ip4Address;
+  IP6_ADDRESS *ip6Address;
+
+  WORD wPort;
+  WORD wPriority;
+  WORD wWeight;
+
+  // Property list
+  DWORD dwPropertyCount;
+
+  PWSTR *keys;
+  PWSTR *values;
+
+  DWORD dwInterfaceIndex;
+} DNS_SERVICE_INSTANCE, *PDNS_SERVICE_INSTANCE;
+#endif
+
+typedef VOID WINAPI DNS_SERVICE_REGISTER_COMPLETE(
+  _In_ DWORD Status,
+  _In_ PVOID pQueryContext,
+  _In_ PDNS_SERVICE_INSTANCE pInstance);
+
+typedef DNS_SERVICE_REGISTER_COMPLETE *PDNS_SERVICE_REGISTER_COMPLETE;
+
+#ifndef __MINGW32__
+typedef struct _DNS_SERVICE_CANCEL {
+  PVOID reserved;
+} DNS_SERVICE_CANCEL, *PDNS_SERVICE_CANCEL;
+
+typedef struct _DNS_SERVICE_REGISTER_REQUEST {
+  ULONG Version;
+  ULONG InterfaceIndex;
+  PDNS_SERVICE_INSTANCE pServiceInstance;
+  PDNS_SERVICE_REGISTER_COMPLETE pRegisterCompletionCallback;
+  PVOID pQueryContext;
+  HANDLE hCredentials;
+  BOOL unicastEnabled;
+} DNS_SERVICE_REGISTER_REQUEST, *PDNS_SERVICE_REGISTER_REQUEST;
+#endif
+
+_FN(_DnsServiceFreeInstance, VOID, (_In_ PDNS_SERVICE_INSTANCE pInstance));
+_FN(_DnsServiceDeRegister, DWORD, (_In_ PDNS_SERVICE_REGISTER_REQUEST pRequest, _Inout_opt_ PDNS_SERVICE_CANCEL pCancel));
+_FN(_DnsServiceRegister, DWORD, (_In_ PDNS_SERVICE_REGISTER_REQUEST pRequest, _Inout_opt_ PDNS_SERVICE_CANCEL pCancel));
+} /* extern "C" */
+
+namespace platf::publish {
+VOID WINAPI register_cb(DWORD status, PVOID pQueryContext, PDNS_SERVICE_INSTANCE pInstance) {
+  auto alarm = (safe::alarm_t<DNS_STATUS>::element_type *)pQueryContext;
+
+  auto fg = util::fail_guard([&]() {
+    if(pInstance) {
+      _DnsServiceFreeInstance(pInstance);
+    }
+  });
+
+  if(status) {
+    print_status("register_cb()"sv, status);
+    alarm->ring(-1);
+
+    return;
+  }
+
+  alarm->ring(0);
+}
+
+static int service(bool enable) {
+  auto alarm = safe::make_alarm<DNS_STATUS>();
+
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
+
+  std::wstring name { SERVICE_INSTANCE_NAME.data(), SERVICE_INSTANCE_NAME.size() };
+  std::wstring domain { SERVICE_TYPE_DOMAIN.data(), SERVICE_TYPE_DOMAIN.size() };
+
+  auto host = converter.from_bytes(boost::asio::ip::host_name() + ".local");
+
+  DNS_SERVICE_INSTANCE instance {};
+  instance.pszInstanceName = name.data();
+  instance.wPort           = map_port(nvhttp::PORT_HTTP);
+  instance.pszHostName     = host.data();
+
+  DNS_SERVICE_REGISTER_REQUEST req {};
+  req.Version                     = DNS_QUERY_REQUEST_VERSION1;
+  req.pQueryContext               = alarm.get();
+  req.pServiceInstance            = &instance;
+  req.pRegisterCompletionCallback = register_cb;
+
+  DNS_STATUS status {};
+
+  if(enable) {
+    status = _DnsServiceRegister(&req, nullptr);
+  }
+  else {
+    status = _DnsServiceDeRegister(&req, nullptr);
+  }
+
+  alarm->wait();
+
+  status = *alarm->status();
+  if(status) {
+    BOOST_LOG(error) << "No mDNS service"sv;
+    return -1;
+  }
+
+  return 0;
+}
+
+class deinit_t : public ::platf::deinit_t {
+public:
+  ~deinit_t() override {
+    if(service(false)) {
+      std::abort();
+    }
+
+    BOOST_LOG(info) << "Unregistered Sunshine Gamestream service"sv;
+  }
+};
+
+int load_funcs(HMODULE handle) {
+  auto fg = util::fail_guard([handle]() {
+    FreeLibrary(handle);
+  });
+
+  _DnsServiceFreeInstance = (_DnsServiceFreeInstance_fn)GetProcAddress(handle, "DnsServiceFreeInstance");
+  _DnsServiceDeRegister   = (_DnsServiceDeRegister_fn)GetProcAddress(handle, "DnsServiceDeRegister");
+  _DnsServiceRegister     = (_DnsServiceRegister_fn)GetProcAddress(handle, "DnsServiceRegister");
+
+  if(!(_DnsServiceFreeInstance && _DnsServiceDeRegister && _DnsServiceRegister)) {
+    BOOST_LOG(error) << "mDNS service not available in dnsapi.dll"sv;
+    return -1;
+  }
+
+  fg.disable();
+  return 0;
+}
+
+std::unique_ptr<::platf::deinit_t> start() {
+  HMODULE handle = LoadLibrary("dnsapi.dll");
+
+  if(!handle || load_funcs(handle)) {
+    BOOST_LOG(error) << "Couldn't load dnsapi.dll, You'll need to add PC manually from Moonlight"sv;
+    return nullptr;
+  }
+
+  if(service(true)) {
+    return nullptr;
+  }
+
+  BOOST_LOG(info) << "Registered Sunshine Gamestream service"sv;
+
+  return std::make_unique<deinit_t>();
+}
+} // namespace platf::publish

src/platform/windows/windows.rs.in

@@ -0,0 +1 @@
+SuperDuperAmazing   ICON    DISCARDABLE    "@SUNSHINE_ICON_PATH@"