I use CUDA Visual Profiler to profile my application and I export the result to CSV file. I see 2 columns, Column 1 is GPU time, column 2 is CPU time. So how do I compute real execution time of my application? Should I sum the column 1 and 2?
Here is screen shot of my CSV file
GPU time is the time measured when you use events to measure the time. CPU time is the time measured using CPU clock. GPU time is supposed to be operating system independent whereas the later may depend on operating system as well as specific processes running during the application’s execution. Also, GPU time has a resolution of 0.5 microseconds as against a variable resolution (depends on the clocking mechanism you use) for CPU clock. I would advise you to take the GPU time for your timing analysis.