Debugging

In order to debug you should use a debug build for the target architecture you are interested in. If CVMFS is available, you should use a dbg tag such as:

cmake -DSTANDALONE=ON -DCMAKE_TOOLCHAIN_FILE=/cvmfs/lhcb.cern.ch/lib/lhcb/lcg-toolchains/LCG_101/x86_64-centos7-clang12+cuda11_4-dbg.cmake ..

Then, you should be able to run your code with a debugger such as gdb (CPU), cuda-gdb (CUDA) or rocgdb (HIP). For instance:

./toolchain/wrapper /usr/local/cuda/bin/cuda-gdb --args ./Allen --sequence hlt1_pp_validation

If you don’t have CVMFS available, you should set the CMAKE_BUILD_TYPE to Debug and use the available local installation of the debugger:

cmake -DSTANDALONE=ON -DCMAKE_BUILD_TYPE=Debug -DTARGET_DEVICE=CUDA ..
cuda-gdb --args ./Allen --sequence hlt1_pp_validation

For some materials on gdb, some recommended reading:

• gdb tutorial

• cuda-gdb documentation

Use callgrind to create a profile of Allen CPU usage

First, make sure to include the correct cmake flags in the build by putting:

"cmakeFlags": {
    "Allen": "-DCALLGRIND_PROFILE=ON"
}

in the utils/config.json file in your stack before you make Allen. Once it is compiled with the flag, the profile can be created using:

MooreOnline/build.{tag}/run valgrind --tool=callgrind --instr-atstart=no python Allen/Dumpers/BinaryDumpers/options/allen.py

with the tags, data, and other flags following as normal. This will create a file in the directory that you ran Allen from named callgrind.out.xxxxxx where xxxxxx is a seemingly random 6 digit number. You may need to copy this to another machine where you have installed qcachegrind or another program capable of reading callgrind files. On that machine, run:

qcachegrind callgrind.out.xxxxxx

replacing callgrind.out.xxxxxx with your file name. This should launch a window showing the CPU usage of Allen in a variety of different formats including tiles and flowchart.