At present, is it possible to have hardware counters like dram_reads, dram_writes, instructions_issued, etc dump their results into the CUDA profiler log file? I am specifying the counters in $CUDA_PROFILER_LOG file, but I get an error.
You need to specify the counters you want to profile in the profiler configuration file and set the environment variable COMPUTE_PROFILE_CONFIG to point to this file.
Create a profiler configuration file (text format) “config.txt” say in the home directory. To profile “inst_issued” add the counter name on a separate line.
Set the environment variable
The CUDA_PROFILER_LOG environment variable defines the profiler output log file.
Which version of CUDA toolkit are you using?
Thank you, I am using CUDA 4.2 - I think I am giving the wrong names in the config file, is there a document which lists the counter names? For instance, I was putting instructions_issued and getting an error because the name could not be resolved.
Refer the CUPTI Userâ€™s Guide (“CUPTI_Users_Guide.pdf” located under toolkit “extra/CUPTI/doc” directory )â€š “Event Reference” sections for supported counters on GPU devices with different compute capabilities. Note that CUPTI events are referred to as profiler counters in the command line profiler. The event name listed in the table can be used in the command line profiler configuration file. In every application run only a few counter values can be collected. The number of counters depends on the specific counters selected.