Calling below function on Jetson TX1 is causing performance spikes on a thread, but it works on main. Removing cudaDeviceSynchronize() gives stable timing on the thread because most probably it’s not measuring the kernel execution time.
std::thread t;
int Start()
{
while(true)
{
if(stop)
break;
}
int tick1 = std::chrono::duration_cast<std::chrono::microseconds(std::chrono::high_resolution_clock::now().time_since_epoch()).count();
for(int i=0; i<20; i++)
estimateStep();
int tock1 = std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::high_resolution_clock::now().time_since_epoch()).count();
std::cout << "ICP Timing: " << ((tock1 - tick1) / 1000.0f) << std::endl;
}
extern “C” int Start_RS()
{
t = std::thread(Start); //Adding t.join after this command or anyother blocking call gives stable timing a well
//Start(); //Calling this directly also gives stable timing
return 0;
}
extern “C” int Stop_RS()
{
stop = true;
t.join();
return 0;
}
global void estimateKernel()
{
//Test
}
void estimateStep()
{
estimateKernel<<<50, 500>>>();
cudaDeviceSynchronize();
}
//Timing (ms): Calling Start function in thread
ICP Timing: 0.651
ICP Timing: 0.859
ICP Timing: 9.032
ICP Timing: 66.029
ICP Timing: 57.577
ICP Timing: 64.601
ICP Timing: 60.967
ICP Timing: 64.266
ICP Timing: 48.34
ICP Timing: 35.074
ICP Timing: 33.066
ICP Timing: 39.791
ICP Timing: 0.664
ICP Timing: 0.623
ICP Timing: 0.619
//Timing (ms): Calling Start function from main
ICP Timing: 4.565
ICP Timing: 3.899
ICP Timing: 2.383
ICP Timing: 4.519
ICP Timing: 2.559
ICP Timing: 5.237
ICP Timing: 2.647
ICP Timing: 3.334
ICP Timing: 3.653
ICP Timing: 5.125
ICP Timing: 3.482
ICP Timing: 2.377
ICP Timing: 4.581
ICP Timing: 2.592
ICP Timing: 5.603
ICP Timing: 2.374
ICP Timing: 3.87
ICP Timing: 2.923
ICP Timing: 5.239
I have also tried maximizing CPU performance as mentioned Jetson/Performance - eLinux.org, but results are same.
Any suggestions, please.