Project Integration Examples
ASGarD is a PDE solver code, best detailed here.
The first reason to integrate TiMemory to ASGarD is to study the impact of changing the floating point precision (double/single) on the application performance. The next step is to build a performance regression test suite.
1. Set timemory_FIND_COMPONENTS_INTERFACE to a name specific to a project
2. Specify the components you absolutely need after COMPONENTS
3. Specify the components that are optional after OPTIONAL_COMPONENTS
4. Just use whatever target name you set to timemory_FIND_COMPONENTS_INTERFACE in all subsequent target_link_libraries(…) in your project
Note: It is recommended to compile timemory with same compiler vendor as the project it is integrated into. However, for compilers which use other compilers as their backend, e.g. older versions of Intel provide GCC compatibility and newer version use LLVM as the backend, it is fine to mix installations.
timemory is available on Cori with for Haswell and KNL and Cori-GPU (with module load cgpu):
- timemory/3.0.1-gcc8-cuda-10.1-py36
- timemory/haswell-develop-gcc-py37-mpich
- timemory/cgpu-develop-gcc-py37-openmpi
Nevertheless, you can also compile timemory using the latest version available on github.
In our case, we want to compile ASGarD with clang (PrgEnv-cray), so timemory should also be compiled with clang.
Using cmake:
cmake -DCMAKE_INSTALL_PREFIX=/path/to/install/ -DCMAKE_BUILD_TYPE=RelWithDebInfo -DTIMEMORY_USE_PAPI=OFF -DTIMEMORY_USE_CUDA=ON ..
make -j32
make installSee also: Installation Examples
Applying the rule of thumb(s):
set(timemory_FIND_COMPONENTS_INTERFACE timemory-asgard) # Rule of thumb #1
find_package(timemory REQUIRED
COMPONENTS cxx shared cuda # Rule of thumb #2
OPTIONAL_COMPONENTS gotcha) # Rule of thumb #3
target_link_libraries(lib_dispatch PRIVATE timemory-asgard) # Rule of thumb #4
target_link_libraries(asgard PRIVATE timemory-asgard) # Rule of thumb #4