Extend packet header with frame processing latency

src/platform/common.h

@@ -186,6 +186,8 @@ namespace platf {
     std::int32_t pixel_pitch {};
     std::int32_t row_pitch {};
 
+    std::optional<std::chrono::steady_clock::time_point> frame_timestamp;
+
     virtual ~img_t() = default;
   };
 

src/platform/linux/cuda.cu

@@ -1,5 +1,6 @@
 // #include <algorithm>
 #include <helper_math.h>
+#include <chrono>
 #include <limits>
 #include <memory>
 #include <optional>
@@ -31,8 +32,10 @@ using namespace std::literals;
 
 //////////////////// Special desclarations
 /**
- * NVCC segfaults when including <chrono>
- * Therefore, some declarations need to be added explicitely
+ * NVCC tends to have problems with standard headers.
+ * Don't include common.h, instead use bare minimum
+ * of standard headers and duplicate declarations of necessary classes.
+ * Not pretty and extremely error-prone, fix at earliest convenience.
  */
 namespace platf {
 struct img_t: std::enable_shared_from_this<img_t> {
@@ -43,6 +46,8 @@ public:
   std::int32_t pixel_pitch {};
   std::int32_t row_pitch {};
 
+  std::optional<std::chrono::steady_clock::time_point> frame_timestamp;
+
   virtual ~img_t() = default;
 };
 } // namespace platf

src/platform/windows/display_ram.cpp

@@ -1,4 +1,6 @@
 #include "display.h"
+
+#include "misc.h"
 #include "src/main.h"
 
 namespace platf {
@@ -192,6 +194,12 @@ namespace platf::dxgi {
       return capture_e::timeout;
     }
 
+    std::optional<std::chrono::steady_clock::time_point> frame_timestamp;
+    if (auto qpc_displayed = std::max(frame_info.LastPresentTime.QuadPart, frame_info.LastMouseUpdateTime.QuadPart)) {
+      // Translate QueryPerformanceCounter() value to steady_clock time point
+      frame_timestamp = std::chrono::steady_clock::now() - qpc_time_difference(qpc_counter(), qpc_displayed);
+    }
+
     if (frame_info.PointerShapeBufferSize > 0) {
       auto &img_data = cursor.img_data;
 
@@ -307,6 +315,10 @@ namespace platf::dxgi {
       blend_cursor(cursor, *img);
     }
 
+    if (img) {
+      img->frame_timestamp = frame_timestamp;
+    }
+
     return capture_e::ok;
   }
 

src/platform/windows/display_vram.cpp

@@ -11,6 +11,7 @@ extern "C" {
 }
 
 #include "display.h"
+#include "misc.h"
 #include "src/main.h"
 #include "src/video.h"
 
@@ -894,6 +895,12 @@ namespace platf::dxgi {
       return capture_e::timeout;
     }
 
+    std::optional<std::chrono::steady_clock::time_point> frame_timestamp;
+    if (auto qpc_displayed = std::max(frame_info.LastPresentTime.QuadPart, frame_info.LastMouseUpdateTime.QuadPart)) {
+      // Translate QueryPerformanceCounter() value to steady_clock time point
+      frame_timestamp = std::chrono::steady_clock::now() - qpc_time_difference(qpc_counter(), qpc_displayed);
+    }
+
     if (frame_info.PointerShapeBufferSize > 0) {
       DXGI_OUTDUPL_POINTER_SHAPE_INFO shape_info {};
 
@@ -1239,6 +1246,10 @@ namespace platf::dxgi {
       old_surface_delayed_destruction.reset();
     }
 
+    if (img_out) {
+      img_out->frame_timestamp = frame_timestamp;
+    }
+
     return capture_e::ok;
   }
 

src/platform/windows/misc.cpp

@@ -1020,4 +1020,25 @@ namespace platf {
 
     return std::make_unique<qos_t>(flow_id);
   }
+  int64_t
+  qpc_counter() {
+    LARGE_INTEGER performace_counter;
+    if (QueryPerformanceCounter(&performace_counter)) return performace_counter.QuadPart;
+    return 0;
+  }
+
+  std::chrono::nanoseconds
+  qpc_time_difference(int64_t performance_counter1, int64_t performance_counter2) {
+    auto get_frequency = []() {
+      LARGE_INTEGER frequency;
+      frequency.QuadPart = 0;
+      QueryPerformanceFrequency(&frequency);
+      return frequency.QuadPart;
+    };
+    static const double frequency = get_frequency();
+    if (frequency) {
+      return std::chrono::nanoseconds((int64_t) ((performance_counter1 - performance_counter2) * frequency / std::nano::den));
+    }
+    return {};
+  }
 }  // namespace platf

src/platform/windows/misc.h

@@ -1,5 +1,7 @@
 #ifndef SUNSHINE_WINDOWS_MISC_H
 #define SUNSHINE_WINDOWS_MISC_H
+
+#include <chrono>
 #include <string_view>
 #include <windows.h>
 #include <winnt.h>
@@ -9,6 +11,12 @@ namespace platf {
   print_status(const std::string_view &prefix, HRESULT status);
   HDESK
   syncThreadDesktop();
+
+  int64_t
+  qpc_counter();
+
+  std::chrono::nanoseconds
+  qpc_time_difference(int64_t performance_counter1, int64_t performance_counter2);
 }  // namespace platf
 
 #endif