instructions_issued and instructions_executed not available

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I’m not able to get the instructions_issued or instructions_executed counters to work with my M2050 and with CUDA 3.2 code that has been compiled with -arch sm20 for compute capability 2.0. I’m using CentOS 5.4 and driver 260.19.29.

I’ve enabled the CUDA profiler with:

export COMPUTE_PROFILE=1
export COMPUTE_PROFILE_CONFIG=profile_config

and in my config file I’ve specified

instructions_issued
instructions_executed
gld_request
l1_global_load_miss
l2_subp0_read_sector_misses

But in my profile output, I’m getting the warnings:

NV_Warning: Ignoring the invalid profiler config option: instructions_issued
NV_Warning: Ignoring the invalid profiler config option: instructions_executed

The Compute_Profiler.txt file that comes in the CUDA documentation
says that these counters are available for compute compatibility 2.0,
and I’m compiling with -arch sm20. Also, I’m using other counters (l1
and l2) that are compute 2.0 only.

Has anyone had similar problems to this? Any ideas on what is wrong
with my configuration? Or is the documentation incorrect?

Thanks for your help,
Mark Howison

2

At least my documentation at /usr/local/cuda/doc/Compute_Profiler.txt (CUDA 3.2) say the same, but I suspect it is wrong in this respect. If I look at some CSV I have around, it should probably be inst_issued and inst_executed. Let us know if it works (have only a 1.1 card at home, so I cannot check myself easily before tomorrow)!

3

Thanks, that did the trick!

Now NVIDIA just needs to update the documentation to show inst_issued/executed. They are also incorrectly listed in the slides form this SC10 presentation I was looking at:

:http://www.nvidia.com/content/PDF/sc_2010/CUDA_Tutorial/SC10_Analysis_Driven_Optimization.pdf