[host] measure ready time at framebuffer publication

Exclude command queue reset and cleanup from the ready stage, and keep
the wall-clock residual only as the legacy timing fallback.

client/src/main.c

@@ -205,6 +205,7 @@ struct RenderTiming
   uint64_t captureTime;
   uint64_t postProcessTime;
   uint64_t copyTime;
+  uint64_t readyTime;
   uint64_t importTime;
 };
 
@@ -225,6 +226,8 @@ static struct RenderTiming frameTimingLoad(void)
         &g_state.producerPostProcessTime, memory_order_seq_cst);
     timing.copyTime        = atomic_load_explicit(
         &g_state.producerCopyTime, memory_order_seq_cst);
+    timing.readyTime       = atomic_load_explicit(
+        &g_state.producerReadyTime, memory_order_seq_cst);
     timing.importTime      = atomic_load_explicit(
         &g_state.clientImportTime, memory_order_seq_cst);
 
@@ -245,6 +248,8 @@ static void frameTimingStore(const LG_TransportFrameTiming * timing,
       timing->postProcessTime, memory_order_seq_cst);
   atomic_store_explicit(&g_state.producerCopyTime,
       timing->copyTime, memory_order_seq_cst);
+  atomic_store_explicit(&g_state.producerReadyTime,
+      timing->readyTime, memory_order_seq_cst);
   atomic_store_explicit(&g_state.clientImportTime,
       importTime, memory_order_seq_cst);
   atomic_fetch_add_explicit(
@@ -257,13 +262,14 @@ static void preSwapCallback(void * udata)
   const uint64_t timestamp  = nanotime();
   const uint64_t renderTime = timestamp - timing->renderStart;
   if (timing->captureTime || timing->postProcessTime || timing->copyTime ||
-      timing->importTime)
+      timing->readyTime || timing->importTime)
   {
     const OverlayFrameTiming frameTiming = {
       .timestamp   = timestamp,
       .capture     = timing->captureTime     * 1e-6f,
       .postProcess = timing->postProcessTime * 1e-6f,
       .copy        = timing->copyTime        * 1e-6f,
+      .ready       = timing->readyTime       * 1e-6f,
       .import      = timing->importTime      * 1e-6f,
       .render      = renderTime              * 1e-6f,
     };

client/src/main.h

@@ -154,6 +154,7 @@ struct AppState
   atomic_uint_least64_t producerCaptureTime;
   atomic_uint_least64_t producerPostProcessTime;
   atomic_uint_least64_t producerCopyTime;
+  atomic_uint_least64_t producerReadyTime;
   atomic_uint_least64_t clientImportTime;
   atomic_uint_least64_t renderCount, frameCount;
   _Atomic(float)        fps, ups;

client/src/overlay/graphs.c

@@ -179,7 +179,7 @@ static bool rbCalcMetrics(int index, void * value_, void * udata_)
 #define TIMING_PLOT_WINDOW_NS 20000000000ULL
 #define TIMING_PLOT_BUCKET_NS \
   (TIMING_PLOT_WINDOW_NS / TIMING_PLOT_BUCKETS)
-#define TIMING_STAGE_COUNT    5
+#define TIMING_STAGE_COUNT    6
 
 struct TimingPlotData
 {
@@ -239,6 +239,7 @@ static bool accumulateTimingSample(int index, void * value_, void * udata_)
     timing->capture,
     timing->postProcess,
     timing->copy,
+    timing->ready,
     timing->import,
     timing->render,
   };
@@ -349,6 +350,7 @@ static void renderTimingStatistic(struct OverlayGraph * graph,
     "Capture",
     "Post",
     "Copy",
+    "Ready",
     "Import",
     "Render",
   };

client/src/overlays.h

@@ -50,6 +50,7 @@ typedef struct OverlayFrameTiming
   float    capture;
   float    postProcess;
   float    copy;
+  float    ready;
   float    import;
   float    render;
 }