OClCollMeshEngn: perf profile and print kernel exec durations

stimBuffApis/livoxGen1/openClCollatingAndMeshingEngine.cpp

@@ -766,6 +766,9 @@ public:
 	void compactCollateAndMeshFrameReq1_doCompact_posted(
 		std::shared_ptr<CompactCollateAndMeshFrameReq> context)
 	{
+		// Record compact kernel start time
+		engine.compactKernelStartTime = std::chrono::high_resolution_clock::now();
+
 		bool success = engine.startCompactKernel(
 			engine.parent.assemblyBuffer,
 			static_cast<uint32_t>(context->frameAssemblyResult.nSucceeded.load()),
@@ -788,6 +791,9 @@ public:
 		cl_int compactStatus)
 	{
 		engine.stopCompactKernel();
+		// Record compact kernel end time
+		engine.compactKernelEndTime = std::chrono::high_resolution_clock::now();
+
 		// If compact failed, call callback directly with failure
 		if (compactStatus != CL_SUCCESS)
 		{
@@ -811,6 +817,9 @@ public:
 	void compactCollateAndMeshFrameReq3_doCollate_posted(
 		std::shared_ptr<CompactCollateAndMeshFrameReq> context)
 	{
+		// Record collate kernel start time
+		engine.collateKernelStartTime = std::chrono::high_resolution_clock::now();
+
 		bool success = engine.startCollateKernel(
 			engine.parent.assemblyBuffer, engine.parent.collationBuffer,
 			std::bind(
@@ -832,6 +841,9 @@ public:
 		cl_int collateStatus)
 	{
 		engine.stopCollateKernel();
+		// Record collate kernel end time
+		engine.collateKernelEndTime = std::chrono::high_resolution_clock::now();
+
 		bool success = (collateStatus == CL_SUCCESS);
 
 		// Early callback + return pattern
@@ -901,5 +913,25 @@ void OpenClCollatingAndMeshingEngine::compactCollateAndMeshFrameReq(
 	}
 }
 
+std::chrono::milliseconds OpenClCollatingAndMeshingEngine::getCompactKernelDuration() const
+{
+	auto duration = compactKernelEndTime - compactKernelStartTime;
+	if (duration.count() < 0)
+	{
+		return std::chrono::milliseconds(0);
+	}
+	return std::chrono::duration_cast<std::chrono::milliseconds>(duration);
+}
+
+std::chrono::milliseconds OpenClCollatingAndMeshingEngine::getCollateKernelDuration() const
+{
+	auto duration = collateKernelEndTime - collateKernelStartTime;
+	if (duration.count() < 0)
+	{
+		return std::chrono::milliseconds(0);
+	}
+	return std::chrono::duration_cast<std::chrono::milliseconds>(duration);
+}
+
 } // namespace stim_buff
 } // namespace smo