Prevent unnecessary copies of entire frames on Windows

CMakeLists.txt

@@ -148,6 +148,7 @@ set(SUNSHINE_TARGET_FILES
 	sunshine/thread_pool.h
 	sunshine/thread_safe.h
 	sunshine/sync.h
+	sunshine/round_robin.h
 	${PLATFORM_TARGET_FILES})
 
 include_directories(

sunshine/platform/common.h

@@ -91,7 +91,7 @@ std::string from_sockaddr(const sockaddr *const);
 std::pair<std::uint16_t, std::string> from_sockaddr_ex(const sockaddr *const);
 
 std::unique_ptr<mic_t> microphone(std::uint32_t sample_rate);
-std::unique_ptr<display_t> display();
+std::shared_ptr<display_t> display();
 
 input_t input();
 void move_mouse(input_t &input, int deltaX, int deltaY);

sunshine/platform/linux.cpp

@@ -315,8 +315,8 @@ struct mic_attr_t : public mic_t {
   }
 };
 
-std::unique_ptr<display_t> shm_display() {
-  auto shm = std::make_unique<shm_attr_t>();
+std::shared_ptr<display_t> shm_display() {
+  auto shm = std::make_shared<shm_attr_t>();
 
   if(shm->init()) {
     return nullptr;
@@ -325,11 +325,11 @@ std::unique_ptr<display_t> shm_display() {
   return shm;
 }
 
-std::unique_ptr<display_t> display() {
+std::shared_ptr<display_t> display() {
   auto shm_disp = shm_display();
 
   if(!shm_disp) {
-    return std::unique_ptr<display_t> { new x11_attr_t {} };
+    return std::make_shared<x11_attr_t>();
   }
 
   return shm_disp;

sunshine/platform/windows.cpp

@@ -249,9 +249,7 @@ void keyboard(input_t &input, uint16_t modcode, bool release) {
   auto key_state = GetAsyncKeyState(modcode);
   bool key_state_down = (key_state & KEY_STATE_DOWN) != 0;
   if(key_state_down != release) {
-    BOOST_LOG(warning) << "Key state of vkey ["sv << util::hex(modcode).to_string_view() << "] does not match the desired state ["sv << (release ? "on]"sv : "off]"sv);
-
-    return;
+    BOOST_LOG(debug) << "Key state of vkey ["sv << util::hex(modcode).to_string_view() << "] does not match the desired state ["sv << (release ? "on]"sv : "off]"sv);
   }
 
   INPUT i {};

sunshine/platform/windows_dxgi.cpp

@@ -104,9 +104,57 @@ public:
 class display_t;
 struct img_t : public ::platf::img_t  {
   ~img_t() override {
-    delete[] data;
-    data = nullptr;
+    unmap();
   }
+
+  void unmap() {
+    if(info.pData) {
+      device_ctx_p->Unmap(texture.get(), 0);
+
+      info.pData = nullptr;
+    }
+  }
+
+  int reset(int width, int height, DXGI_FORMAT format, device_t::pointer device, device_ctx_t::pointer device_ctx_p, const std::shared_ptr<display_t> &display) {
+    unmap();
+
+    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;
+
+    dxgi::texture2d_t::pointer tex_p {};
+    auto status = device->CreateTexture2D(&t, nullptr, &tex_p);
+    texture.reset(tex_p);
+
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create texture [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    this->display      = display;
+    this->device_ctx_p = device_ctx_p;
+    this->data         = nullptr;
+    this->row_pitch    = 0;
+    this->pixel_pitch  = 4;
+    this->width        = width;
+    this->height       = height;
+
+    return 0;
+  }
+
+  std::shared_ptr<display_t> display;
+
+  texture2d_t texture;
+  D3D11_MAPPED_SUBRESOURCE info {};
+
+
+  device_ctx_t::pointer device_ctx_p;
 };
 
 struct cursor_t {
@@ -247,10 +295,16 @@ void blend_cursor(const cursor_t &cursor, img_t &img) {
   }
 }
 
-class display_t : public ::platf::display_t {
+class display_t : public ::platf::display_t, public std::enable_shared_from_this<display_t> {
 public:
   capture_e snapshot(::platf::img_t *img_base, bool cursor_visible) override {
-    auto img = (img_t *) img_base;
+    auto img = (img_t*)img_base;
+    if(img->display.get() != this) {
+      if(img->reset(width, height, format, device.get(), device_ctx.get(), shared_from_this())) {
+        return capture_e::error;
+      }
+    }
+
     HRESULT status;
 
     DXGI_OUTDUPL_FRAME_INFO frame_info;
@@ -296,15 +350,15 @@ public:
         }
 
         //Copy from GPU to CPU
-        device_ctx->CopyResource(texture.get(), src.get());
+        device_ctx->CopyResource(img->texture.get(), src.get());
       }
 
-      if(current_img.pData) {
-        device_ctx->Unmap(texture.get(), 0);
-        current_img.pData = nullptr;
+      if(img->info.pData) {
+        device_ctx->Unmap(img->texture.get(), 0);
+        img->info.pData = nullptr;
       }
 
-      status = device_ctx->Map(texture.get(), 0, D3D11_MAP_READ, 0, &current_img);
+      status = device_ctx->Map(img->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() << ']';
 
@@ -312,28 +366,23 @@ public:
       }
     }
 
+    /*
     const bool update_flag =
       frame_info.LastMouseUpdateTime.QuadPart  ||
       frame_info.LastPresentTime.QuadPart != 0 ||
       frame_info.PointerShapeBufferSize > 0;
-
+    */
+    const bool update_flag = frame_info.LastPresentTime.QuadPart != 0;
 
     if(!update_flag) {
       return capture_e::timeout;
     }
 
-    if(img->width != width || img->height != height) {
-      delete[] img->data;
-      img->data = new std::uint8_t[height * current_img.RowPitch];
+    img->row_pitch = img->info.RowPitch;
+    img->data = (std::uint8_t*)img->info.pData;
 
-      img->width = width;
-      img->height = height;
-      img->row_pitch = current_img.RowPitch;
-    }
-
-    std::copy_n((std::uint8_t*)current_img.pData, height * current_img.RowPitch, (std::uint8_t*)img->data);
-    if(cursor_visible && cursor.visible) {
-      blend_cursor(cursor, *img);
+    if(cursor_visible) { // && cursor.visible) {
+      //blend_cursor(cursor, *img);
     }
 
     return capture_e::ok;
@@ -342,11 +391,9 @@ public:
   std::shared_ptr<::platf::img_t> alloc_img() override {
     auto img = std::make_shared<img_t>();
 
-    img->data   = nullptr;
-    img->height = 0;
-    img->width  = 0;
-    img->row_pitch = 0;
-    img->pixel_pitch = 4;
+    if(img->reset(width, height, format, device.get(), device_ctx.get(), shared_from_this())) {
+      return nullptr;
+    }
 
     return img;
   }
@@ -368,8 +415,6 @@ public:
     FreeLibrary(user32);
   });
 */
-    current_img.pData = nullptr; // current_img is not yet mapped
-
     dxgi::factory1_t::pointer   factory_p {};
     dxgi::adapter_t::pointer    adapter_p {};
     dxgi::output_t::pointer     output_p {};
@@ -548,43 +593,9 @@ public:
 
     BOOST_LOG(debug) << "Source format ["sv << format_str[dup_desc.ModeDesc.Format] << ']';
 
-    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;
-
-    dxgi::texture2d_t::pointer tex_p {};
-    status = device->CreateTexture2D(&t, nullptr, &tex_p);
-
-    texture.reset(tex_p);
-
-    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, &current_img);
-    if(FAILED(status)) {
-      BOOST_LOG(error) << "Error: Failed to map the texture [0x"sv << util::hex(status).to_string_view() << ']';
-      return -1;
-    }
-
     return 0;
   }
 
-  ~display_t() override {
-    if(current_img.pData) {
-      device_ctx->Unmap(texture.get(), 0);
-      current_img.pData = nullptr;
-    }
-  }
-
   factory1_t factory;
   adapter_t adapter;
   output_t output;
@@ -592,13 +603,11 @@ public:
   device_ctx_t device_ctx;
   duplication_t dup;
   cursor_t cursor;
-  texture2d_t texture;
 
   int width, height;
 
   DXGI_FORMAT format;
   D3D_FEATURE_LEVEL feature_level;
-  D3D11_MAPPED_SUBRESOURCE current_img;
 };
 
 const char *format_str[] = {
@@ -729,8 +738,8 @@ const char *format_str[] = {
 }
 
 namespace platf {
-std::unique_ptr<display_t> display() {
-  auto disp = std::make_unique<dxgi::display_t>();
+std::shared_ptr<display_t> display() {
+  auto disp = std::make_shared<dxgi::display_t>();
 
   if (disp->init()) {
     return nullptr;

sunshine/round_robin.h

@@ -0,0 +1,141 @@
+#ifndef KITTY_UTIL_ITERATOR_H
+#define KITTY_UTIL_ITERATOR_H
+
+#include <iterator>
+
+namespace util {
+template<class V, class T>
+class it_wrap_t : public std::iterator<std::random_access_iterator_tag, V> {
+public:
+  typedef T iterator;
+  typedef typename std::iterator<std::random_access_iterator_tag, V>::value_type class_t;
+
+  typedef class_t& reference;
+  typedef class_t* pointer;
+
+  typedef std::ptrdiff_t diff_t;
+
+  iterator operator += (diff_t step) {
+    while(step-- > 0) {
+      ++_this();
+    }
+
+    return _this();
+  }
+
+  iterator operator -= (diff_t step) {
+    while(step-- > 0) {
+      --_this();
+    }
+
+    return _this();
+  }
+
+  iterator operator +(diff_t step) {
+    iterator new_ = _this();
+
+    return new_ += step;
+  }
+
+  iterator operator -(diff_t step) {
+    iterator new_ = _this();
+
+    return new_ -= step;
+  }
+
+  diff_t operator -(iterator first) {
+    diff_t step = 0;
+    while(first != _this()) {
+      ++step;
+      ++first;
+    }
+
+    return step;
+  }
+
+  iterator operator++() { _this().inc(); return _this(); }
+  iterator operator--() { _this().dec(); return _this(); }
+
+  iterator operator++(int) {
+    iterator new_ = _this();
+
+    ++_this();
+
+    return new_;
+  }
+
+  iterator operator--(int) {
+    iterator new_ = _this();
+
+    --_this();
+
+    return new_;
+  }
+
+  reference operator*() { return *_this().get(); }
+  const reference operator*() const { return *_this().get(); }
+
+  pointer operator->() { return &*_this(); }
+  const pointer operator->() const { return &*_this(); }
+
+  bool operator != (const iterator &other) const {
+    return !(_this() == other);
+  }
+
+  bool operator < (const iterator &other) const {
+    return !(_this() >= other);
+  }
+
+  bool operator >= (const iterator &other) const {
+    return _this() == other || _this() > other;
+  }
+
+  bool operator <= (const iterator &other) const {
+    return _this() == other || _this() < other;
+  }
+
+  bool operator == (const iterator &other) const { return _this().eq(other); };
+  bool operator > (const iterator &other) const { return _this().gt(other); }
+private:
+
+  iterator &_this() { return *static_cast<iterator*>(this); }
+  const iterator &_this() const { return *static_cast<const iterator*>(this); }
+};
+
+template<class V, class It>
+class round_robin_t : public it_wrap_t<V, round_robin_t<V, It>> {
+public:
+  using iterator = It;
+  using pointer = V*;
+
+  round_robin_t(iterator begin, iterator end) : _begin(begin), _end(end), _pos(begin) {}
+
+  void inc() {
+    ++_pos;
+
+    if(_pos == _end) {
+      _pos = _begin;
+    }
+  }
+
+  bool eq(const round_robin_t &other) const {
+    return *_pos == *other._pos;
+  }
+
+  pointer get() const {
+    return &*_pos;
+  }
+private:
+  It _begin;
+  It _end;
+
+  It _pos;
+};
+
+template<class V, class It>
+round_robin_t<V, It> make_round_robin(It begin, It end) {
+  return round_robin_t<V, It>(begin, end);
+}
+}
+
+#endif

sunshine/stream.cpp

@@ -614,7 +614,7 @@ void videoBroadcastThread(safe::signal_t *shutdown_event, udp::socket &sock, vid
         frame_new = "\000\000\000\001("sv;
       }
 
-      assert(std::search(std::begin(payload), std::end(payload), std::begin(hevc_i_frame), std::end(hevc_i_frame)) ==
+      assert(std::search(std::begin(payload), std::end(payload), std::begin(frame_new), std::end(frame_new)) ==
              std::end(payload));
       payload_new = replace(payload, frame_old, frame_new);
       payload = {(char *) payload_new.data(), payload_new.size()};

sunshine/utility.h

@@ -388,69 +388,6 @@ void c_free(T *p) {
 template<class T>
 using c_ptr = safe_ptr<T, c_free<T>>;
 
-template<class T>
-class FakeContainer {
-  typedef T pointer;
-
-  pointer _begin;
-  pointer _end;
-
-public:
-  FakeContainer(pointer begin, pointer end) : _begin(begin), _end(end) {}
-
-  pointer begin() { return _begin; }
-  pointer end() { return  _end; }
-
-  const pointer begin() const { return _begin; }
-  const pointer end() const { return _end; }
-
-  const pointer cbegin() const { return _begin; }
-  const pointer cend() const { return _end; }
-
-  pointer data() { return begin(); }
-  const pointer data() const { return cbegin(); }
-
-  std::size_t size() const { return std::distance(begin(), end()); }
-};
-
-template<class T>
-FakeContainer<T> toContainer(T begin, T end) {
-  return { begin, end };
-}
-
-template<class T>
-FakeContainer<T> toContainer(T begin, std::size_t end) {
-  return { begin, begin + end };
-}
-
-template<class T>
-FakeContainer<T*> toContainer(T * const begin) {
-  T *end = begin;
-
-  auto default_val = T();
-  while(*end != default_val) {
-    ++end;
-  }
-
-  return toContainer(begin, end);
-}
-
-template<class T, class H>
-struct _init_helper;
-
-template<template<class...> class T, class H, class... Args>
-struct _init_helper<T<Args...>, H> {
-  using type = T<Args...>;
-
-  static type move(Args&&... args, H&&) {
-    return std::make_tuple(std::move(args)...);
-  }
-
-  static type copy(const Args&... args, const H&) {
-    return std::make_tuple(args...);
-  }
-};
-
 inline std::int64_t from_chars(const char *begin, const char *end) {
   std::int64_t res {};
   std::int64_t mul = 1;

sunshine/video.cpp

@@ -11,6 +11,7 @@ extern "C" {
 
 #include "platform/common.h"
 #include "thread_pool.h"
+#include "round_robin.h"
 #include "config.h"
 #include "video.h"
 #include "main.h"
@@ -138,40 +139,6 @@ struct capture_thread_ctx_t {
   return codec_t { ctx.get() };
 }
 
-int capture_display(platf::img_t *img, std::unique_ptr<platf::display_t> &disp) {
-  auto status = disp->snapshot(img, display_cursor);
-  switch (status) {
-    case platf::capture_e::reinit: {
-      // We try this twice, in case we still get an error on reinitialization
-      for(int x = 0; x < 2; ++x) {
-        disp.reset();
-        disp = platf::display();
-
-        if(disp) {
-          break;
-        }
-
-        std::this_thread::sleep_for(200ms);
-      }
-
-      if(!disp) {
-        return -1;
-      }
-
-      return 0;
-    }
-    case platf::capture_e::error:
-     return -1;
-    case platf::capture_e::timeout:
-     return 0;
-    case platf::capture_e::ok:
-      return 1;
-    default:
-      BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
-      return -1;
-  }
-}
-
 void captureThread(std::shared_ptr<safe::queue_t<capture_ctx_t>> capture_ctx_queue) {
   std::vector<capture_ctx_t> capture_ctxs;
 
@@ -190,6 +157,22 @@ void captureThread(std::shared_ptr<safe::queue_t<capture_ctx_t>> capture_ctx_que
   std::chrono::nanoseconds delay = 1s;
 
   auto disp = platf::display();
+  if(!disp) {
+    return;
+  }
+
+  std::vector<std::shared_ptr<platf::img_t>> imgs(12);
+  auto round_robin = util::make_round_robin<std::shared_ptr<platf::img_t>>(std::begin(imgs), std::end(imgs));
+
+  for(auto &img : imgs) {
+    img = disp->alloc_img();
+    if(!img) {
+      BOOST_LOG(error) << "Couldn't initialize an image"sv;
+      return;
+    }
+  }
+
+  auto next_frame = std::chrono::steady_clock::now();
   while(capture_ctx_queue->running()) {
     while(capture_ctx_queue->peek()) {
       capture_ctxs.emplace_back(std::move(*capture_ctx_queue->pop()));
@@ -197,13 +180,62 @@ void captureThread(std::shared_ptr<safe::queue_t<capture_ctx_t>> capture_ctx_que
       delay = std::min(delay, capture_ctxs.back().delay);
     }
 
-    std::shared_ptr<platf::img_t> img = disp->alloc_img();
-    auto result = capture_display(img.get(), disp);
-    if(result < 0) {
-      return;
+    auto now = std::chrono::steady_clock::now();
+    if(next_frame > now) {
+      std::this_thread::sleep_until(next_frame);
     }
-    if(!result) {
-      continue;
+    next_frame += delay;
+
+    auto &img = *round_robin++;
+    auto status = disp->snapshot(img.get(), display_cursor);
+    switch (status) {
+      case platf::capture_e::reinit: {
+        // Some classes of images contain references to the display --> display won't delete unless img is deleted
+        for(auto &img : imgs) {
+          img.reset();
+        }
+
+        while(disp.use_count() > 1) {
+          std::this_thread::sleep_for(100ms);
+        }
+
+        // We try this twice, in case we still get an error on reinitialization
+        for(int x = 0; x < 2; ++x) {
+          // Some classes of display cannot have multiple instances at once
+          disp.reset();
+          disp = platf::display();
+
+          if(disp) {
+            break;
+          }
+
+          std::this_thread::sleep_for(200ms);
+        }
+
+        if(!disp) {
+          return;
+        }
+
+        // Re-allocate images
+        for(auto &img : imgs) {
+          img = disp->alloc_img();
+          if(!img) {
+            BOOST_LOG(error) << "Couldn't initialize an image"sv;
+            return;
+          }
+        }
+
+        continue;
+      }
+      case platf::capture_e::error:
+       return;
+      case platf::capture_e::timeout:
+        continue;
+      case platf::capture_e::ok:
+        break;
+      default:
+        BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
+        return;
     }
 
     KITTY_WHILE_LOOP(auto capture_ctx = std::begin(capture_ctxs), capture_ctx != std::end(capture_ctxs), {
@@ -508,6 +540,20 @@ void capture(
   int framerate = config.framerate;
 
   auto images = std::make_shared<img_event_t::element_type>();
+
+  // Temporary image to ensure something is send to Moonlight even if no frame has been captured yet.
+  int dummy_data = 0;
+  {
+    auto img = std::make_shared<platf::img_t>();
+    img->row_pitch = 4;
+    img->height = 1;
+    img->width = 1;
+    img->pixel_pitch = 4;
+    img->data = (std::uint8_t*)&dummy_data;
+
+    images->raise(std::move(img));
+  }
+
   // Keep a reference counter to ensure the capture thread only runs when other threads have a reference to the capture thread
   static auto capture_thread = safe::make_shared<capture_thread_ctx_t>(start_capture, end_capture);
   auto ref = capture_thread.ref();
@@ -533,15 +579,6 @@ void capture(
   // Initiate scaling context with correct height and width
   sws_t sws;
 
-  // Temporary image to ensure something is send to Moonlight even if no frame has been captured yet.
-  int dummy_data = 0;
-  auto img = std::make_shared<platf::img_t>();
-  img->row_pitch = 4;
-  img->height = 1;
-  img->width = 1;
-  img->pixel_pitch = 4;
-  img->data = (std::uint8_t*)&dummy_data;
-
   auto next_frame = std::chrono::steady_clock::now();
   while(true) {
     if(shutdown_event->peek() || !images->running()) {
@@ -566,8 +603,29 @@ void capture(
 
     // When Moonlight request an IDR frame, send frames even if there is no new captured frame
     if(frame_nr > (key_frame_nr + config.framerate) || images->peek()) {
-      if(auto tmp_img = images->pop(delay)) {
-        img = std::move(tmp_img);
+      if(auto img = images->pop(delay)) {
+        if(software.system_memory) {
+          auto new_width  = img->width;
+          auto new_height = img->height;
+
+          if(img_width != new_width || img_height != new_height) {
+            img_width  = new_width;
+            img_height = new_height;
+
+            sws.reset(
+              sws_getContext(
+                img_width, img_height, AV_PIX_FMT_BGR0,
+                session->ctx->width, session->ctx->height, session->ctx->pix_fmt,
+                SWS_LANCZOS | SWS_ACCURATE_RND,
+                nullptr, nullptr, nullptr));
+
+            sws_setColorspaceDetails(sws.get(), sws_getCoefficients(SWS_CS_DEFAULT), 0,
+                                     sws_getCoefficients(session->sws_color_format), config.encoderCscMode & 0x1,
+                                     0, 1 << 16, 1 << 16);
+          }
+        }
+
+        software.img_to_frame(sws, *img, session->frame);
       }
       else if(images->running()) {
         continue;
@@ -577,29 +635,6 @@ void capture(
       }
     }
 
-    if(software.system_memory) {
-      auto new_width  = img->width;
-      auto new_height = img->height;
-
-      if(img_width != new_width || img_height != new_height) {
-        img_width  = new_width;
-        img_height = new_height;
-
-        sws.reset(
-          sws_getContext(
-            img_width, img_height, AV_PIX_FMT_BGR0,
-            session->ctx->width, session->ctx->height, session->ctx->pix_fmt,
-            SWS_LANCZOS | SWS_ACCURATE_RND,
-            nullptr, nullptr, nullptr));
-
-        sws_setColorspaceDetails(sws.get(), sws_getCoefficients(SWS_CS_DEFAULT), 0,
-                                 sws_getCoefficients(session->sws_color_format), config.encoderCscMode & 0x1,
-                                 0, 1 << 16, 1 << 16);
-      }
-    }
-
-    software.img_to_frame(sws, *img, session->frame);
-
     encode(frame_nr++, session->ctx, session->frame, packets, channel_data);
 
     session->frame->pict_type = AV_PICTURE_TYPE_NONE;