Added ability to resize present event cb

PresentData/PresentMonTraceConsumer.cpp

@@ -22,15 +22,8 @@
 #include <stdlib.h>
 #include <unordered_set>
 
-static constexpr int PRESENTEVENT_CIRCULAR_BUFFER_SIZE = 4096;
-
 static uint32_t gNextFrameId = 1;
 
-static inline uint32_t GetRingIndex(uint32_t index)
-{
-    return index % PRESENTEVENT_CIRCULAR_BUFFER_SIZE;
-}
-
 static inline uint64_t GenerateVidPnLayerId(uint32_t vidPnSourceId, uint32_t layerIndex)
 {
     return (((uint64_t) vidPnSourceId) << 32) | (uint64_t) layerIndex;
@@ -248,6 +241,16 @@ PresentEvent::PresentEvent(uint32_t fid)
 PMTraceConsumer::PMTraceConsumer()
     : mTrackedPresents(PRESENTEVENT_CIRCULAR_BUFFER_SIZE)
     , mCompletedPresents(PRESENTEVENT_CIRCULAR_BUFFER_SIZE)
+    , mCircularBufferSize(PRESENTEVENT_CIRCULAR_BUFFER_SIZE)
+    , mGpuTrace(this)
+{
+    hEventsReadyEvent = CreateEventW(nullptr, FALSE, FALSE, nullptr);
+}
+
+PMTraceConsumer::PMTraceConsumer(uint32_t circularBufferSize)
+    : mTrackedPresents(circularBufferSize)
+    , mCompletedPresents(circularBufferSize)
+    , mCircularBufferSize(circularBufferSize)
     , mGpuTrace(this)
 {
     hEventsReadyEvent = CreateEventW(nullptr, FALSE, FALSE, nullptr);
@@ -2210,11 +2213,11 @@ void PMTraceConsumer::CompletePresent(std::shared_ptr<PresentEvent> const& p)
     if (present != nullptr) {
         uint32_t index;
         // if completed buffer is full
-        if (mCompletedCount == PRESENTEVENT_CIRCULAR_BUFFER_SIZE) {
+        if (mCompletedCount == mCircularBufferSize) {
             // if we are in offline ETL processing mode, block instead of overwriting events
             // unless either A) the buffer is full of non-ready events or B) backpressure disabled via CLI option
             if (!mIsRealtimeSession && mReadyCount != 0 && !mDisableOfflineBackpressure) {
-                mCompletedRingCondition.wait(lock, [this] { return mCompletedCount < PRESENTEVENT_CIRCULAR_BUFFER_SIZE; });
+                mCompletedRingCondition.wait(lock, [this] { return mCompletedCount < mCircularBufferSize; });
                 index = GetRingIndex(mCompletedIndex + mCompletedCount);
                 mCompletedCount++;
             }
@@ -2230,6 +2233,7 @@ void PMTraceConsumer::CompletePresent(std::shared_ptr<PresentEvent> const& p)
                 if (mReadyCount > 0) {
                     mReadyCount--;
                 }
+                mNumOverflowedPresents++;
             }
             // otherwise, completed buffer still has available space
         }

PresentData/PresentMonTraceConsumer.hpp

@@ -294,6 +294,8 @@ struct PresentEvent {
 
 struct PMTraceConsumer
 {
+    static constexpr int PRESENTEVENT_CIRCULAR_BUFFER_SIZE = 2048;
+
     // -------------------------------------------------------------------------------------------
     // The following are parameters to the analysis, which can be configured prior to starting the
     // trace session.
@@ -353,6 +355,8 @@ struct PMTraceConsumer
     uint32_t mCompletedIndex = 0;       // The index of mCompletedPresents of the oldest completed present.
     uint32_t mCompletedCount = 0;       // The total number of presents in mCompletedPresents.
     uint32_t mReadyCount = 0;           // The number of presents in mCompletedPresents, starting at mCompletedIndex, that are ready to be dequeued.
+    uint32_t mNumOverflowedPresents = 0; // The number of presents that have been lost due to the ring buffer wrapping.
+    uint32_t mCircularBufferSize = 0;   // The size of the ring buffers for presents.
 
     // Mutexs to protect consumer/dequeue access from different threads:
     std::mutex mProcessEventMutex;
@@ -505,6 +509,7 @@ struct PMTraceConsumer
     // Functions for decoding ETW and analysing process and present events.
 
     PMTraceConsumer();
+    PMTraceConsumer(uint32_t circularBufferSize);
     ~PMTraceConsumer();
 
     PMTraceConsumer(const PMTraceConsumer&) = delete;
@@ -566,4 +571,10 @@ struct PMTraceConsumer
     AppTimingData* ExtractAppTimingData(uint32_t processId, uint32_t appFrameId, uint64_t presentStartTime);
     bool IsApplicationPresent(std::shared_ptr<PresentEvent> const& present);
     void SetAppTimingDataAsComplete(uint32_t processId, uint32_t appFrameId);
+
+    inline uint32_t GetRingIndex(uint32_t index)
+    {
+        return index % mCircularBufferSize;
+    }
+
 };

PresentMon/CommandLine.cpp

@@ -299,6 +299,7 @@ void PrintUsage()
         LR"(--track_hw_measurements)", LR"(Tracks HW-measured latency and/or power data coming from a LMT and/or PCAT device.)",
         LR"(--track_app_timing)", LR"(Track app times for each displayed frame; requires application and/or driver instrumentation using Intel-PresentMon provider.)",
         LR"(--track_hybrid_present)", LR"(Tracks if the present is a hybrid present and is performing a cross adapter copy.)",
+        LR"(--set_circular_buffer_size)", LR"(Overide the default present event circular buffer size of 2048. Must be a power of 2.)",
     };
 
     // Layout
@@ -384,6 +385,7 @@ bool ParseCommandLine(int argc, wchar_t** argv)
     args->mTimer = 0;
     args->mHotkeyModifiers = MOD_NOREPEAT;
     args->mHotkeyVirtualKeyCode = 0;
+    args->mPresentEventCircularBufferSize = 0;
     args->mConsoleOutput = ConsoleOutput::Statistics;
     args->mTrackDisplay = true;
     args->mTrackInput = true;
@@ -460,6 +462,7 @@ bool ParseCommandLine(int argc, wchar_t** argv)
         else if (ParseArg(argv[i], L"restart_as_admin"))           { args->mTryToElevate             = true; continue; }
         else if (ParseArg(argv[i], L"terminate_on_proc_exit"))     { args->mTerminateOnProcExit      = true; continue; }
         else if (ParseArg(argv[i], L"terminate_after_timed"))      { args->mTerminateAfterTimer      = true; continue; }
+        else if (ParseArg(argv[i], L"set_circular_buffer_size"))   { if (ParseValue(argv, argc, &i, &args->mPresentEventCircularBufferSize)) { continue; } }
 
         // Beta options:
         else if (ParseArg(argv[i], L"track_frame_type"))      { args->mTrackFrameType      = true; continue; }
@@ -507,6 +510,14 @@ bool ParseCommandLine(int argc, wchar_t** argv)
         return false;
     }
 
+    // Disallow a circular buffer size that is not a power of 2.
+    if (args->mPresentEventCircularBufferSize != 0) {
+        if ((args->mPresentEventCircularBufferSize & (args->mPresentEventCircularBufferSize - 1)) != 0) {
+            PrintError(L"error: --set_circular_buffer_size must be a power of 2.\n");
+            return false;
+        }
+    }
+
     // Ensure only one of --output_file --output_stdout --no_csv.
     if (csvOutputNone + csvOutputStdout + (args->mOutputCsvFileName != nullptr) > 1) {
         PrintWarning(L"warning: only one of the following options may be used:");

PresentMon/MainThread.cpp

@@ -253,7 +253,11 @@ int wmain(int argc, wchar_t** argv)
     SetConsoleCtrlHandler(HandleCtrlEvent, TRUE);
 
     // Create event consumers
-    PMTraceConsumer pmConsumer;
+    PMTraceConsumer pmConsumer(
+        args.mPresentEventCircularBufferSize != 0
+        ? args.mPresentEventCircularBufferSize
+        : PMTraceConsumer::PRESENTEVENT_CIRCULAR_BUFFER_SIZE);
+
     pmConsumer.mTrackDisplay               = args.mTrackDisplay;
     pmConsumer.mTrackGPU                   = args.mTrackGPU;
     pmConsumer.mTrackGPUVideo              = args.mTrackGPUVideo;
@@ -383,6 +387,14 @@ int wmain(int argc, wchar_t** argv)
     if (pmSession.mNumEventsLost > 0) {
         PrintWarning(L"warning: %lu ETW events were lost.\n", pmSession.mNumEventsLost);
     }
+    if (pmConsumer.mNumOverflowedPresents > 0) {
+        PrintWarning(L"warning: %lu overflowed present events detected. This could be due to a high-fps application.\n",
+                     pmConsumer.mNumOverflowedPresents);
+        PrintWarning(L"         Consider increasing the present event circular buffer size to a value larger\n");
+        PrintWarning(L"         than the default of 2048, e.g., --set_circular_buffer_size 4096.\n");
+
+
+    }
 
     /* We cannot remove the Ctrl handler because it is in an infinite sleep so
      * this call will never return, either hanging the application or having

PresentMon/PresentMon.hpp

@@ -66,6 +66,7 @@ struct CommandLineArgs {
     UINT mTimer;
     UINT mHotkeyModifiers;
     UINT mHotkeyVirtualKeyCode;
+    UINT mPresentEventCircularBufferSize;
     TimeUnit mTimeUnit;
     CSVOutput mCSVOutput;
     ConsoleOutput mConsoleOutput;