Hi, I’m doing my graduation work using CUDA and I’m trying to use Compute Visual Profiler to get some statistics about my application.
Description of my system and machine:
Linux version: 2.6.38-8-generic-pae
CUDA Driver Version: 4.0
CUDA Runtime Version: 3.20
CUDA Capability Major/Minor version number: 1.2
Device: GeForce GT 230M
I’m running my app on Linux, but I’ve noticed that the profiler for windows has some different features, so I tried to get the file generated by the profiler on Linux and use the same file on windows.
Almost everything works fine, except in the summary table, when I try to analyse a kernel (bcp, for example), I expected to see the Limiting factor Identification, Memory throughtput analysis, Instruction throughtput analysis and Occupancy analyses, but only the last one is enabled. The attached file profiler.png illustrates the issue. The first question is, what can I do to enable all analysis topics?
The second question is: What is considered in calculation of Gpu Idle time, showed on picture gpuidle.png? I mean, why, or how is it possible for gpu idle represent 99,47% of execution time?
Third and last question, is it necessary and what is the impact of enabling debugging (-g -G) and run the profiler?
Thank you very much and sorry for my terrible english.