Kernel runs much faster in visual profiler than cmd

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi, all,

I met a strange issue recently. One day i analyze my app in visual profiler and found it ran much faster than cmd, even for empty kernel. the code is like

TIMING_START_GPU_NONAME;
cu_plane_handleOut<<< gridSize, blockSize >>> (*this, particles);
TIMING_END_GPU_NONAME;

and kernel is empty

global void cu_plane_handleOut (cuPlane plane, cuParticle particles)
{
return;
}

the macro TIMING_START_GPU_NONAME and TIMING_END_GPU_NONAME output the time. In profiler, the time is about 4us, it should be correct. but in cmd line, it outputs 245us, I don’t think it is right, but I have no idea what’s wrong with it.

my operating system is win7 64 bit, installed on SSD, display card is GTX 780, latest driver. the app is 32bit, so I tried both 32bit and 64bit cmd, but same results. also tried run cmd as admin.

anybody can help me out? thanks in advance.

#2

OK, I knew what was wrong with it now.

it is about compatibility, right click the app, click “troubleshoot compatibility” -> “try recommended setting”->“start the program”.

the recommended setting for me is XP sp2.