Performance chasing

apps/LoopThroughWithOpenGLCompositing/runtime/telemetry/HealthTelemetry.cpp

@@ -212,6 +212,7 @@ void HealthTelemetry::RecordBackendPlayoutHealth(const std::string& lifecycleSta
 	std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 	uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 	double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+	double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 	uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 	uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 	uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
@@ -232,6 +233,9 @@ void HealthTelemetry::RecordBackendPlayoutHealth(const std::string& lifecycleSta
 	mBackendPlayout.outputRenderMilliseconds = std::max(outputRenderMilliseconds, 0.0);
 	mBackendPlayout.smoothedOutputRenderMilliseconds = std::max(smoothedOutputRenderMilliseconds, 0.0);
 	mBackendPlayout.maxOutputRenderMilliseconds = std::max(maxOutputRenderMilliseconds, 0.0);
+	mBackendPlayout.outputFrameAcquireMilliseconds = std::max(outputFrameAcquireMilliseconds, 0.0);
+	mBackendPlayout.outputFrameRenderRequestMilliseconds = std::max(outputFrameRenderRequestMilliseconds, 0.0);
+	mBackendPlayout.outputFrameEndAccessMilliseconds = std::max(outputFrameEndAccessMilliseconds, 0.0);
 	mBackendPlayout.completedFrameIndex = completedFrameIndex;
 	mBackendPlayout.scheduledFrameIndex = scheduledFrameIndex;
 	mBackendPlayout.scheduledLeadFrames = scheduledLeadFrames;
@@ -251,6 +255,7 @@ bool HealthTelemetry::TryRecordBackendPlayoutHealth(const std::string& lifecycle
 	std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 	uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 	double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+	double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 	uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 	uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 	uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
@@ -274,6 +279,9 @@ bool HealthTelemetry::TryRecordBackendPlayoutHealth(const std::string& lifecycle
 	mBackendPlayout.outputRenderMilliseconds = std::max(outputRenderMilliseconds, 0.0);
 	mBackendPlayout.smoothedOutputRenderMilliseconds = std::max(smoothedOutputRenderMilliseconds, 0.0);
 	mBackendPlayout.maxOutputRenderMilliseconds = std::max(maxOutputRenderMilliseconds, 0.0);
+	mBackendPlayout.outputFrameAcquireMilliseconds = std::max(outputFrameAcquireMilliseconds, 0.0);
+	mBackendPlayout.outputFrameRenderRequestMilliseconds = std::max(outputFrameRenderRequestMilliseconds, 0.0);
+	mBackendPlayout.outputFrameEndAccessMilliseconds = std::max(outputFrameEndAccessMilliseconds, 0.0);
 	mBackendPlayout.completedFrameIndex = completedFrameIndex;
 	mBackendPlayout.scheduledFrameIndex = scheduledFrameIndex;
 	mBackendPlayout.scheduledLeadFrames = scheduledLeadFrames;
@@ -289,6 +297,98 @@ bool HealthTelemetry::TryRecordBackendPlayoutHealth(const std::string& lifecycle
 	return true;
 }
 
+void HealthTelemetry::RecordOutputRenderQueueWait(double queueWaitMilliseconds)
+{
+	std::lock_guard<std::mutex> lock(mMutex);
+	mBackendPlayout.outputRenderQueueWaitMilliseconds = std::max(queueWaitMilliseconds, 0.0);
+}
+
+bool HealthTelemetry::TryRecordOutputRenderQueueWait(double queueWaitMilliseconds)
+{
+	std::unique_lock<std::mutex> lock(mMutex, std::try_to_lock);
+	if (!lock.owns_lock())
+		return false;
+
+	mBackendPlayout.outputRenderQueueWaitMilliseconds = std::max(queueWaitMilliseconds, 0.0);
+	return true;
+}
+
+void HealthTelemetry::RecordOutputRenderPipelineTiming(
+	double drawMilliseconds,
+	double fenceWaitMilliseconds,
+	double mapMilliseconds,
+	double readbackCopyMilliseconds,
+	double cachedCopyMilliseconds,
+	double asyncQueueMilliseconds,
+	double asyncQueueBufferMilliseconds,
+	double asyncQueueSetupMilliseconds,
+	double asyncQueueReadPixelsMilliseconds,
+	double asyncQueueFenceMilliseconds,
+	double syncReadMilliseconds,
+	bool asyncReadbackMissed,
+	bool cachedFallbackUsed,
+	bool syncFallbackUsed)
+{
+	std::lock_guard<std::mutex> lock(mMutex);
+	mBackendPlayout.outputRenderDrawMilliseconds = std::max(drawMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackFenceWaitMilliseconds = std::max(fenceWaitMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackMapMilliseconds = std::max(mapMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackCopyMilliseconds = std::max(readbackCopyMilliseconds, 0.0);
+	mBackendPlayout.outputCachedCopyMilliseconds = std::max(cachedCopyMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueMilliseconds = std::max(asyncQueueMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueBufferMilliseconds = std::max(asyncQueueBufferMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueSetupMilliseconds = std::max(asyncQueueSetupMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueReadPixelsMilliseconds = std::max(asyncQueueReadPixelsMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueFenceMilliseconds = std::max(asyncQueueFenceMilliseconds, 0.0);
+	mBackendPlayout.outputSyncReadMilliseconds = std::max(syncReadMilliseconds, 0.0);
+	if (asyncReadbackMissed)
+		++mBackendPlayout.outputAsyncReadbackMissCount;
+	if (cachedFallbackUsed)
+		++mBackendPlayout.outputCachedFallbackCount;
+	if (syncFallbackUsed)
+		++mBackendPlayout.outputSyncFallbackCount;
+}
+
+bool HealthTelemetry::TryRecordOutputRenderPipelineTiming(
+	double drawMilliseconds,
+	double fenceWaitMilliseconds,
+	double mapMilliseconds,
+	double readbackCopyMilliseconds,
+	double cachedCopyMilliseconds,
+	double asyncQueueMilliseconds,
+	double asyncQueueBufferMilliseconds,
+	double asyncQueueSetupMilliseconds,
+	double asyncQueueReadPixelsMilliseconds,
+	double asyncQueueFenceMilliseconds,
+	double syncReadMilliseconds,
+	bool asyncReadbackMissed,
+	bool cachedFallbackUsed,
+	bool syncFallbackUsed)
+{
+	std::unique_lock<std::mutex> lock(mMutex, std::try_to_lock);
+	if (!lock.owns_lock())
+		return false;
+
+	mBackendPlayout.outputRenderDrawMilliseconds = std::max(drawMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackFenceWaitMilliseconds = std::max(fenceWaitMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackMapMilliseconds = std::max(mapMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackCopyMilliseconds = std::max(readbackCopyMilliseconds, 0.0);
+	mBackendPlayout.outputCachedCopyMilliseconds = std::max(cachedCopyMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueMilliseconds = std::max(asyncQueueMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueBufferMilliseconds = std::max(asyncQueueBufferMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueSetupMilliseconds = std::max(asyncQueueSetupMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueReadPixelsMilliseconds = std::max(asyncQueueReadPixelsMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueFenceMilliseconds = std::max(asyncQueueFenceMilliseconds, 0.0);
+	mBackendPlayout.outputSyncReadMilliseconds = std::max(syncReadMilliseconds, 0.0);
+	if (asyncReadbackMissed)
+		++mBackendPlayout.outputAsyncReadbackMissCount;
+	if (cachedFallbackUsed)
+		++mBackendPlayout.outputCachedFallbackCount;
+	if (syncFallbackUsed)
+		++mBackendPlayout.outputSyncFallbackCount;
+	return true;
+}
+
 HealthTelemetry::SignalStatusSnapshot HealthTelemetry::GetSignalStatusSnapshot() const
 {
 	std::lock_guard<std::mutex> lock(mMutex);

apps/LoopThroughWithOpenGLCompositing/runtime/telemetry/HealthTelemetry.h

@@ -97,6 +97,24 @@ public:
 		double outputRenderMilliseconds = 0.0;
 		double smoothedOutputRenderMilliseconds = 0.0;
 		double maxOutputRenderMilliseconds = 0.0;
+		double outputFrameAcquireMilliseconds = 0.0;
+		double outputFrameRenderRequestMilliseconds = 0.0;
+		double outputFrameEndAccessMilliseconds = 0.0;
+		double outputRenderQueueWaitMilliseconds = 0.0;
+		double outputRenderDrawMilliseconds = 0.0;
+		double outputReadbackFenceWaitMilliseconds = 0.0;
+		double outputReadbackMapMilliseconds = 0.0;
+		double outputReadbackCopyMilliseconds = 0.0;
+		double outputCachedCopyMilliseconds = 0.0;
+		double outputAsyncQueueMilliseconds = 0.0;
+		double outputAsyncQueueBufferMilliseconds = 0.0;
+		double outputAsyncQueueSetupMilliseconds = 0.0;
+		double outputAsyncQueueReadPixelsMilliseconds = 0.0;
+		double outputAsyncQueueFenceMilliseconds = 0.0;
+		double outputSyncReadMilliseconds = 0.0;
+		uint64_t outputAsyncReadbackMissCount = 0;
+		uint64_t outputCachedFallbackCount = 0;
+		uint64_t outputSyncFallbackCount = 0;
 		uint64_t completedFrameIndex = 0;
 		uint64_t scheduledFrameIndex = 0;
 		uint64_t scheduledLeadFrames = 0;
@@ -161,6 +179,7 @@ public:
 		std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 		uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 		double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+		double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 		uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 		uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 		uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
@@ -170,11 +189,46 @@ public:
 		std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 		uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 		double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+		double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 		uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 		uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 		uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
 		bool degraded, const std::string& statusMessage);
 
+	void RecordOutputRenderQueueWait(double queueWaitMilliseconds);
+	bool TryRecordOutputRenderQueueWait(double queueWaitMilliseconds);
+
+	void RecordOutputRenderPipelineTiming(
+		double drawMilliseconds,
+		double fenceWaitMilliseconds,
+		double mapMilliseconds,
+		double readbackCopyMilliseconds,
+		double cachedCopyMilliseconds,
+		double asyncQueueMilliseconds,
+		double asyncQueueBufferMilliseconds,
+		double asyncQueueSetupMilliseconds,
+		double asyncQueueReadPixelsMilliseconds,
+		double asyncQueueFenceMilliseconds,
+		double syncReadMilliseconds,
+		bool asyncReadbackMissed,
+		bool cachedFallbackUsed,
+		bool syncFallbackUsed);
+	bool TryRecordOutputRenderPipelineTiming(
+		double drawMilliseconds,
+		double fenceWaitMilliseconds,
+		double mapMilliseconds,
+		double readbackCopyMilliseconds,
+		double cachedCopyMilliseconds,
+		double asyncQueueMilliseconds,
+		double asyncQueueBufferMilliseconds,
+		double asyncQueueSetupMilliseconds,
+		double asyncQueueReadPixelsMilliseconds,
+		double asyncQueueFenceMilliseconds,
+		double syncReadMilliseconds,
+		bool asyncReadbackMissed,
+		bool cachedFallbackUsed,
+		bool syncFallbackUsed);
+
 	SignalStatusSnapshot GetSignalStatusSnapshot() const;
 	VideoIOStatusSnapshot GetVideoIOStatusSnapshot() const;
 	PerformanceSnapshot GetPerformanceSnapshot() const;