Question about NVIDIA CUDA Visual Profiler Version 2.2

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, all,

I have a question on the results returned by the visual profiler.

In my program, I run a kernel for 10 times. And I can see lots of useful informations of this kernel via the profiler.

Here is a part of the output

[codebox]

                   GPUTIME CPUTIME       branch       divergent branch      instructions

SpmatvecKernelCSRTri 31.616 54 0 0 0

SpmatvecKernelCSRTri 30.432 41 0 0 0

SpmatvecKernelCSRTri 29.92 41 0 0 0

SpmatvecKernelCSRTri 30.304 40 0 0 0

SpmatvecKernelCSRTri 29.888 40 0 0 0

SpmatvecKernelCSRTri 30.048 41 0 0 0

SpmatvecKernelCSRTri 29.92 41 270 54 2103

SpmatvecKernelCSRTri 29.76 40 297 57 2394

SpmatvecKernelCSRTri 29.888 41 277 61 2110

SpmatvecKernelCSRTri 29.824 40 304 64 2400

[/codebox]

BUT one thing I DO NOT understand is:

for a given kernel, I think the column instruction should have the (roughly) same number of instructions (for each run). But it does not. Any even sometimes it is zero!

Can anybody explain it ?

Thanks a lot!