Reputation: 4913
Is is possible that streams take longer time than normal CUDA programs?It is guaranteed that chunk of data on which streams are operate are independent and overlapping between kernel execution and data transfer is allowed from device side.
Upvotes: 0
Views: 411
Reputation: 21505
To observe superpositions between streams, you should use the command line profiler with the conckerneltrace
option. You could also save the profiling output as a .csv file and import it within the Visual Profiler. See the NVIDIA Forum post
Upvotes: 1