To run the rocprofv3 trace decoder

The hands-on exercises will go through how to collect trace decoder output. For how to install the ROCProf trace decoder on pre-ROCm 7.0 versions. See the instructions at https://github.com/amd/HPCTrainingDock in the HPCTrainingDock/tools/scripts/rocprofiler-sdk_setup.sh script. This script will install the rocprofiler-sdk, aqlprofile and rocprof-trace-decoder packages into a separate directory and then prepend those paths before the ROCm paths.

With ROCm 7.0 and later, the ROCProf trace decoder is integrated into the ROCm software stack. Load the ROCm module to access rocprofv3:

All of these exercises are from the AMD HPC Training Examples which can be retrieved with the following:

The examples will be either in the HPCTrainingExamples/HIP or HPCTrainingExamples/rocprof-tracedecoder directories.

Transfer the files in the tracedecoder_vectorAdd directory to your local machine and read them into ROCprof Compute Viewer

Cleaning up afterwards

The trace decoder data can be viewed in a separate program called ROCprofiler Compute Viewer. There are pre-built binaries for Microsoft Windows and source code that can be compiled for others systems.

Now start up the ROCprof Compute Viewer.

Untar the data on your local system.

tar -xzvf tracedecoder_vectorAdd.tgz Open up the data file by using the import tab at the upper left. Select one of the ui_output_agent* files in the tracedecoder_vectorAdd directory.

This will open up the Instructions view with the source and ISA windows.

Further exploration:

• Open up the summary view and see an overview of the kernel operation.
• Open up the Wave States to see the timeline view of the instructions
• Go to the HotSpot Timeline view to see the instructions used during the kernel
• Examine the Compute Unit timeline view to see the compute units operation
  • Use the WaveView zoon setting on the control panel on the left to zoom in and out to see all of the timeline or zoom in to a specific part.

Transfer the files in tracedecoder_saxpy to your local machine and read them into ROCprof Compute Viewer

Cleaning up afterwards

Transfer the files in tracedecoder_dgemm_hip to your local machine and read them into ROCprof Compute Viewer

Cleaning up afterwards

Transfer files in tracedecoder_dgemm_library to your local system

Cleaning up afterwards