Page tree
Skip to end of metadata
Go to start of metadata

run and collect data

Run a batch job and collect data:

batch script commands
#SBATCH --constraint=perf  # perf,nvperf for gpu nodes

export PATH=/sw/external/amd/AMDuProf_Linux_x64_4.0.341/bin:$PATH

set -v
srun AMDuProfCLI collect --config tbp  -o `pwd`/uprof_tbp  `pwd`/stream.22gb
srun AMDuProfCLI collect --config inst_access  -o `pwd`/uprof_inst_access  `pwd`/stream.22gb
srun AMDuProfCLI collect --config assess  -o `pwd`/uprof_assess  `pwd`/stream.22gb
srun AMDuProfCLI collect --config assess_ext  -o `pwd`/uprof_assess_ext  `pwd`/stream.22gb

generate report

After collecting data from a batch job, generate reports with the AMDuProfCLI report option.

generate report
[arnoldg@dt-login03 uprof_tbp]$ export PATH=/sw/external/amd/AMDuProf_Linux_x64_4.0.341/bin:$PATH
[arnoldg@dt-login03 uprof_tbp]$ AMDuProfCLI report -i AMDuProf-stream-TBP_Dec-19-2022_09-40-27/
Translation started ...
Translation finished
Generated database file : cpu
Report generation started...
Generating report file...

Report generation completed...

Generated report file: /projects/bbka/slurm_test_scripts/cpu/stream/uprof_tbp/AMDuProf-stream-TBP_Dec-19-2022_09-40-27/report.csv

visualize and explore report data

You can view the data in AMDuProf on Delta or locally on a copy you install at your desktop system.  If you install locally, you may need to replicate some paths or add paths to the binary in order to get full functionality.

Launch AMDuProf  ( no CLI suffice for the GUI ), and import the profile session from a completed batch job run with AMDuProfCLI collect ...

The summary view give a high level overview of how time was spent.  This is the tbp time-based-profile summary.

The Analyze tab shows hot routines or lines in more detail.  The tbp , assess, and  inst_access Analyze views follow.

Selecting one of the lines or routines will take you to the Sources view where you can see the assembly used in that portion of code.

The Session Info is under the Summary tab and displays more detail about the profiling session.


  • No labels


  1. Rooflines are currently disabled: (section 3.5.2 ):

    roofline error messages from OS
    srun AMDuProfPcm roofline -o stream-roofline.csv -- ./stream.22gb
    Error: NMI watchdog is enabled. NMI uses one Core HW PMC counter.
    Please disable NMI watchdog - run with root privilege: echo 0 > /proc/sys/kernel/nmi_watchdog
    srun: error: cn061: task 0: Exited with exit code 255