[dxvk] Add CS thread load to the HUD

src/dxvk/dxvk_cs.cpp

@@ -202,11 +202,22 @@ namespace dxvk {
       while (!m_stopped.load()) {
         { std::unique_lock<dxvk::mutex> lock(m_mutex);
 
-          m_condOnAdd.wait(lock, [this] {
-            return (!m_queueOrdered.queue.empty())
-                || (!m_queueHighPrio.queue.empty())
-                || (m_stopped.load());
-          });
+          auto pred = [this] { return
+              !m_queueOrdered.queue.empty()
+              || !m_queueHighPrio.queue.empty()
+              || m_stopped.load();
+          };
+
+          if (unlikely(!pred())) {
+            auto t0 = dxvk::high_resolution_clock::now();
+
+            m_condOnAdd.wait(lock, [&] {
+              return pred();
+            });
+
+            auto t1 = dxvk::high_resolution_clock::now();
+            m_device->addStatCtr(DxvkStatCounter::CsIdleTicks, std::chrono::duration_cast<std::chrono::microseconds>(t1 - t0).count());
+          }
 
           std::swap(ordered, m_queueOrdered.queue);
           std::swap(highPrio, m_queueHighPrio.queue);

src/dxvk/dxvk_stats.h

@@ -28,6 +28,7 @@ namespace dxvk {
     GpuIdleTicks,             ///< GPU idle time in microseconds
     CsSyncCount,              ///< CS thread synchronizations
     CsSyncTicks,              ///< Time spent waiting on CS
+    CsIdleTicks,              ///< CS thread idle time in microseconds
     CsChunkCount,             ///< Submitted CS chunks
     DescriptorPoolCount,      ///< Descriptor pool count
     DescriptorSetCount,       ///< Descriptor sets allocated

src/dxvk/hud/dxvk_hud_item.cpp

@@ -1446,6 +1446,17 @@ namespace dxvk::hud {
         ? str::format(m_maxCsSyncCount, " (", (syncTicks / 10), ".", (syncTicks % 10), " ms)")
         : str::format(m_maxCsSyncCount);
 
+      uint64_t currCsIdleTicks = counters.getCtr(DxvkStatCounter::CsIdleTicks);
+
+      m_diffCsIdleTicks = currCsIdleTicks - m_prevCsIdleTicks;
+      m_prevCsIdleTicks = currCsIdleTicks;
+
+      uint64_t busyTicks = ticks > m_diffCsIdleTicks
+        ? uint64_t(ticks - m_diffCsIdleTicks)
+        : uint64_t(0);
+
+      m_csLoadString = str::format((100 * busyTicks) / ticks, "%");
+
       m_maxCsSyncCount = 0;
       m_maxCsSyncTicks = 0;
 
@@ -1469,6 +1480,10 @@ namespace dxvk::hud {
     renderer.drawText(16, position, 0xff40ff40, "CS syncs:");
     renderer.drawText(16, { position.x + 132, position.y }, 0xffffffffu, m_csSyncString);
 
+    position.y += 20;
+    renderer.drawText(16, position, 0xff40ff40, "CS load:");
+    renderer.drawText(16, { position.x + 132, position.y }, 0xffffffffu, m_csLoadString);
+
     position.y += 8;
     return position;
   }

src/dxvk/hud/dxvk_hud_item.h

@@ -645,14 +645,18 @@ namespace dxvk::hud {
     uint64_t m_prevCsSyncCount  = 0;
     uint64_t m_prevCsSyncTicks  = 0;
     uint64_t m_prevCsChunks     = 0;
+    uint64_t m_prevCsIdleTicks = 0;
 
     uint64_t m_maxCsSyncCount   = 0;
     uint64_t m_maxCsSyncTicks   = 0;
 
+    uint64_t m_diffCsIdleTicks = 0;
+
     uint64_t m_updateCount      = 0;
 
     std::string m_csSyncString;
     std::string m_csChunkString;
+    std::string m_csLoadString;
 
     dxvk::high_resolution_clock::time_point m_lastUpdate
       = dxvk::high_resolution_clock::now();