README.md

SOLLVE Common Release Repository

This repository contains all projects supported by the SOLLVE project. It is automatically merged by a bot. It contains the following projects:

• llvm: LLVM and its subprojects, including SOLLVE extensions
  • LLVM
  • Clang
  • Polly
  • compiler-rt
  • libcxx
  • libcxxabi
  • openmp
• sollve_vv: OpenMP 4.5 offloading verification suite (https://bitbucket.org/crpl_cisc/sollve_vv)
• llvm/projects/openmp: BOLT (http://www.bolt-omp.org/)
• llvm/projects/openmp/external/argobots: Argobots (http://www.argobots.org)

Installation

Prerequisites

- REQUIRED: C/C++ compilers (gcc/g++ are sufficient)
- REQUIRED: CMake
- OPTIONAL: CUDA environment (to test the offloading features for NVIDIA GPUs)
- OPTIONAL: libelf (for offloading, see https://directory.fsf.org/wiki/Libelf, installed in `<LIBELF_INSTALL_DIR>`)
- OPTIONAL: your own Argobots, installed in `<ARGOBOTS_INSTALL_DIR>`
- OPTIONAL: python (for testing)

Install SOLLVE LLVM

First, go to the top level directory. You can find three files as follows:

sollve$ ls
llvm     sollve_vv    README.md

To build and install the SOLLVE LLVM to <LLVM_INSTALL_DIR>, run the following commands:

cd llvm
mkdir build
cd build
cmake ../ -DCMAKE_INSTALL_PREFIX=<LLVM_INSTALL_DIR> -DLIBOMP_USE_ARGOBOTS=on [other CMake options]
make -j install

Useful CMake configuration options are as follows:

# specify the build option.
-DCMAKE_BUILD_TYPE=<DEBUG|Release|RelWithDebInfo|MinSizeRel>
# use your own Argobots
-DLIBOMP_ARGOBOTS_INSTALL_DIR=<ARGOBOTS_INSTALL_DIR>
# enable offloading with CUDA (if libelf is not installed in your system)
-DLIBOMPTARGET_DEP_LIBELF_INCLUDE_DIRS=<LIBELF_INSTALL_DIR>

If -DLIBOMP_USE_ARGOBOTS=on is not set (e.g., -DLIBOMP_USE_ARGOBOTS=off), the Pthreads-based OpenMP runtime system is installed.

Running Tests

Run the LLVM OpenMP testsuite (mainly for CPUs)

It is assumed that SOLLVE LLVM is properly installed to your system (i.e., all the environmental variables, including PATH, CPATH, LIBRARY_PATH, and LD_LIBRARY_PATH, are properly set to <LLVM_INSTALL_DIR>/bin etc). To run the LLVM OpenMP testsuite, go to the build directory (e.g., sollve/llvm/build) and type as follows:

build$ ls
benchmarks  bin   cmake   CMakeCache.txt ...
build$ python bin/llvm-lit -j<N> projects/bolt/runtime/test

where N is the number of threads to run tests in parallel. All tests are expected to be passed (not failed.)

Run the SOLLVE OpenMP 4.5 V&V testsuite (primarily for the offloading features)

It is assumed that all the necessary paths (e.g., PATH, CPATH, LIBRARY_PATH, and LD_LIBRARY_PATH) to <LLVM_INSTALL_DIR> and <LIBELF_INSTALL_DIR> are properly set. The SOLLVE OpenMP 4.5 V&V testsuite can be run as follows:

sollve$ ls
llvm     sollve_vv    README.md
sollve$ cd sollve_vv
sollve$ make CC=clang CXX=clang++ SYSTEM=<SYSTEM_NAME> LOG=1 SOURCES=tests/*/*.c*
# to create a test summary
sollve$ make CC=clang CXX=clang++ SYSTEM=<SYSTEM_NAME> SOURCES=tests/*/*.c* report_json
sollve$ make CC=clang CXX=clang++ SYSTEM=<SYSTEM_NAME> SOURCES=tests/*/*.c* report_html

Note that the current SOLLVE LLVM might not pass all of the tests.

By default, the following systems are support as <SYSTEM_NAME>:

• summit
• summitdev
• titan

You can add your own configuration. The following is an example to test the offloading features with CUDA 9.1.85:

# Definitions for the local system

$(info "Including local.def file")

JSRUN_COMMAND =
BATCH_SCHEDULER =

# Your CUDA version must be properly set.
# The following is an example to use CUDA 9.1.85.
CUDA_MODULE ?= cuda/9.1.85

CCOMPILERS="clang"
ifeq ($(CC), clang)
  C_COMPILER_MODULE = llvm
  C_VERSION = clang -dumpversion
endif
CXXCOMPILERS="clang++"
ifeq ($(CXX), clang++)
  CXX_COMPILER_MODULE = llvm
  CXX_VERSION = clang++ -dumpversion
endif

You can add a new <SYSTEM_NAME> (say local) by creating the file (local.def) in sollve/sollve_vv/sys/systems.