Fixed VR Frame Interpolation frame-history copying bottleneck

aurora-main/tests/stereo_frame_worker_smoke.cpp

@@ -15,8 +15,39 @@
 #include <filesystem>
 #include <mutex>
 #include <thread>
+#if defined(_WIN32)
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif
 
 using Clock = std::chrono::steady_clock;
+static void WaitUntil(Clock::time_point deadline) {
+#if defined(_WIN32)
+  // Match the runtime's high-resolution pacing. Sleep's coarse Windows timer
+  // would otherwise turn an offscreen 90 Hz test into a roughly 60 Hz test.
+  struct Timer {
+    HANDLE handle = CreateWaitableTimerExW(nullptr, nullptr, CREATE_WAITABLE_TIMER_HIGH_RESOLUTION,
+                                           TIMER_MODIFY_STATE | SYNCHRONIZE);
+    ~Timer() {
+      if (handle != nullptr)
+        CloseHandle(handle);
+    }
+  };
+  static thread_local Timer timer;
+  const auto remaining = std::chrono::duration_cast<std::chrono::nanoseconds>(deadline - Clock::now()).count();
+  if (remaining <= 0)
+    return;
+  LARGE_INTEGER due{};
+  due.QuadPart = -std::max<int64_t>(remaining / 100, 1);
+  if (timer.handle != nullptr && SetWaitableTimerEx(timer.handle, &due, 0, nullptr, nullptr, nullptr, 0)) {
+    WaitForSingleObject(timer.handle, INFINITE);
+    return;
+  }
+#endif
+  std::this_thread::sleep_until(deadline);
+}
 static std::mutex packetMutex;
 static std::condition_variable packetCv;
 static AuroraStereoFrame packet{};
@@ -24,6 +55,7 @@ static bool available = false;
 static bool stop = false;
 static uint64_t completed = 0;
 static std::atomic_uint32_t submitted{0};
+static uint64_t wakeLateness = 0, submitTime = 0, skippedTicks = 0, compositorFrames = 0;
 
 static bool Provide(uint32_t, AuroraStereoFrame* output, void*) {
   std::lock_guard lock(packetMutex);
@@ -76,7 +108,10 @@ int main(int argc, char** argv) {
     uint64_t slot = 1;
     for (uint64_t token = 1;; ++token) {
       const auto deadline = start + std::chrono::nanoseconds(slot * 1'000'000'000 / 90);
-      std::this_thread::sleep_until(deadline);
+      WaitUntil(deadline);
+      const auto woke = Clock::now();
+      wakeLateness +=
+          std::max<int64_t>(0, std::chrono::duration_cast<std::chrono::nanoseconds>(woke - deadline).count());
       {
         std::lock_guard lock(packetMutex);
         if (stop)
@@ -84,8 +119,11 @@ int main(int argc, char** argv) {
         packet = {};
         packet.frameToken = token;
         packet.contentTag = 42;
+        // Wake to render the next display tick, as xrWaitFrame does, rather
+        // than announcing an image whose display deadline has already passed.
         packet.displayTimeNanos =
-            std::chrono::duration_cast<std::chrono::nanoseconds>(deadline.time_since_epoch()).count();
+            std::chrono::duration_cast<std::chrono::nanoseconds>(deadline.time_since_epoch()).count() +
+            1'000'000'000 / 90;
         for (auto& eye : packet.eyes) {
           eye.width = 160;
           eye.height = 120;
@@ -102,17 +140,33 @@ int main(int argc, char** argv) {
       packetCv.wait(lock, [&] { return stop || completed == token; });
       if (stop)
         return;
-      // A real compositor advances to a future display time after a missed
-      // tick. Do not burst old deadlines when interpolation is re-enabled.
+      // Retain the current render tick (its display deadline is still ahead),
+      // but discard older ticks. Skipping even the current tick would insert
+      // an idle period whenever rendering overruns a wakeup by a fraction.
       const auto elapsed = std::chrono::duration_cast<std::chrono::nanoseconds>(Clock::now() - start).count();
-      slot = std::max(slot + 1, static_cast<uint64_t>(elapsed) * 90 / 1'000'000'000 + 1);
+      submitTime += std::chrono::duration_cast<std::chrono::nanoseconds>(Clock::now() - woke).count();
+      ++compositorFrames;
+      const auto nextSlot = std::max(slot + 1, static_cast<uint64_t>(elapsed) * 90 / 1'000'000'000);
+      skippedTicks += nextSlot - slot - 1;
+      slot = nextSlot;
     }
   });
 
-  const auto start = Clock::now();
-  for (uint64_t frame = 0; frame < 240; ++frame) {
+  constexpr uint64_t warmupFrames = 60;
+  auto start = Clock::now();
+  auto measuredStart = start;
+  uint32_t initialSubmissions = 0;
+  for (uint64_t frame = 0; frame < warmupFrames + 240; ++frame) {
+    if (frame == warmupFrames) {
+      // Measure a running scene after shader compilation and initial resource
+      // creation, and rebase the producer so it owes no catch-up frames.
+      aurora_wait_for_frame_worker();
+      measuredStart = Clock::now();
+      start = measuredStart - std::chrono::nanoseconds(frame * 1'000'000'000 / 60);
+      initialSubmissions = submitted.load();
+    }
     const auto boundary = start + std::chrono::nanoseconds((frame + 1) * 1'000'000'000 / 60);
-    std::this_thread::sleep_until(boundary);
+    WaitUntil(boundary);
     aurora_update();
     if (!aurora_begin_frame())
       continue;
@@ -138,6 +192,8 @@ int main(int argc, char** argv) {
     GXSetNumTevStages(1);
     GXSetTevOrder(GX_TEVSTAGE0, GX_TEXCOORD_NULL, GX_TEXMAP_NULL, GX_COLOR_NULL);
     GXSetTevOp(GX_TEVSTAGE0, GX_PASSCLR);
+    // Keep palette triangles small to limit fill cost during uniform stress.
+    const float extent = indexed ? 0.02f : 1.0f;
     if (indexed) {
       for (unsigned matrix = 0; matrix < 10; ++matrix)
         GXLoadPosMtxImm(transform, matrix * 3);
@@ -149,20 +205,25 @@ int main(int argc, char** argv) {
       for (unsigned matrix = 0; matrix < (indexed ? 10u : 1u); ++matrix) {
         if (indexed)
           GXMatrixIndex1u8(GX_VA_PNMTXIDX, matrix * 3);
-        GXPosition3f32(-1 + static_cast<float>(draw) * 0.0001f, -1, 0);
+        GXPosition3f32((-1 + static_cast<float>(draw) * 0.0001f) * extent, -extent, 0);
         if (indexed)
           GXMatrixIndex1u8(GX_VA_PNMTXIDX, matrix * 3);
-        GXPosition3f32(1, -1, 0);
+        GXPosition3f32(extent, -extent, 0);
         if (indexed)
           GXMatrixIndex1u8(GX_VA_PNMTXIDX, matrix * 3);
-        GXPosition3f32(0, 1, 0);
+        GXPosition3f32(0, extent, 0);
       }
       GXEnd();
     }
     aurora_end_frame_tagged(42);
-    if (drawCount > 1 && (frame + 1) % 60 == 0) {
-      std::printf("Completed %llu producer frames in %.2f s\n", static_cast<unsigned long long>(frame + 1),
-                  std::chrono::duration<double>(Clock::now() - start).count());
+    // The runtime pre-warms immediately after its asynchronous seal. The
+    // worker needs this permit before publishing SEALED or encoding XR eyes.
+    // Waiting until the next 60 Hz tick would strand it for a whole interval.
+    aurora_begin_frame();
+    if (drawCount > 1 && frame >= warmupFrames && (frame + 1) % 60 == 0) {
+      std::printf("Completed %llu producer frames in %.2f s\n",
+                  static_cast<unsigned long long>(frame + 1 - warmupFrames),
+                  std::chrono::duration<double>(Clock::now() - measuredStart).count());
       std::fflush(stdout);
     }
   }
@@ -176,10 +237,15 @@ int main(int argc, char** argv) {
   aurora_quiesce_frame_worker();
   aurora_set_stereo_frame_provider(nullptr, nullptr);
   aurora::stereo::set_sink(nullptr, nullptr);
-  const double elapsed = std::chrono::duration<double>(Clock::now() - start).count();
-  const double fps = submitted.load() / elapsed;
+  const double elapsed = std::chrono::duration<double>(Clock::now() - measuredStart).count();
+  const auto measuredSubmissions = submitted.load() - initialSubmissions;
+  const double fps = measuredSubmissions / elapsed;
+  if (compositorFrames != 0)
+    std::printf("Compositor (including warm-up): wake late %.2f ms; submit %.2f ms; skipped %llu ticks\n",
+                wakeLateness / (1.0e6 * compositorFrames), submitTime / (1.0e6 * compositorFrames),
+                static_cast<unsigned long long>(skippedTicks));
   std::printf("%u draws: producer %.1f FPS; %u stereo submissions in %.2f s (%.1f FPS)\n", drawCount, 240 / elapsed,
-              submitted.load(), elapsed, fps);
+              measuredSubmissions, elapsed, fps);
   aurora_shutdown();
   // More headset submissions must not come at the expense of simulation speed.
   const double target = mode == 2 ? 75 : mode == 1 ? 90 : 60;