1

As you know NVIDIA Visual Profiler can show the time when stream started and the time when it stopped working. I made a screenshot of NVIDIA Visual Profiler window. There you can see circled area. Is it possible to get value A and value B in seconds in program that executes this cuda kernel. By the way.. A and B will change if I change program arguments. In this session C time equals B time but only in this session.

I tried to use cudaEvents but you cant get A time (C time) using them. enter image description here

bjb568
  • 11,089
  • 11
  • 50
  • 71
maxi
  • 51
  • 1
  • 7
  • 1
    There is no way using the public CUDA API (CUevents) to collect the same information as the CUDA profiler tools. You can use the [CUPTI Activity API](http://docs.nvidia.com/cuda/cupti/r_main.html#r_activity) included in the CUDA toolkit to collect this data. The CUDA profiler tools report the kernel start time as the time closest to code execution and end time as the completion of the memory barrier following the kernel. The kernel setup cost is not included. The setup cost increases based upon parameter block size and number of textures (among others). – Greg Smith May 30 '14 at 22:08
  • @GregSmith , what samples can you advice me to watch if I want to find out the right way to get C time(A time). [Link to CUPTI samples(nvidia.com)](http://docs.nvidia.com/cuda/cupti/r_main.html#r_samples) – maxi Jun 01 '14 at 10:49
  • 1
    %CUDA_PATH_V6_0%\extras\CUPTI\sample\activity_trace_async is an example of how to use the CUPTI Activity API. This includes all trace activities. The CUPTI_ACTIVITY_KIND_{KERNEL, MARKER, MEMCPY, and MEMSET} are sufficient to trace GPU workloads and to calculate time range. – Greg Smith Jun 02 '14 at 15:10

0 Answers0