Skip to content

Latest commit

 

History

History
86 lines (68 loc) · 3.67 KB

INSTALL.md

File metadata and controls

86 lines (68 loc) · 3.67 KB
Raw

Installation Instructions

Prerequisites

  • CMake v3.21+
  • A C++14 (17 for SYCL builds) compilant C++ compiler

Optional Dependencies

  • A CUDA compiler, (e.g. nvcc, required only if CUDA is enabled)
  • A HIP compiler, (e.g. hipcc or modern clang++, required only if HIP is enabled)
  • A LLVM SYCL Compiler (e.g. DPC++, required only if SYCL enabled)
  • Libxc (required only for host API)

Configuration, Compilation and Installation

ExchCXX provides a CMake build system with automatic dependency management (through FetchContent). As such, a simple CMake invocation will often suffice for most purposes

cmake -S /path/to/ExchCXX -B /path/to/binary [ExchCXX configure options]
cmake --build /path/to/binary # to build without installation
cmake --build /path/to/binary --target install # to install

ExchCXX is linkable both as an installed library as well as a CMake subproject via FetchContent

# ExchCXX Discovery
find_package( exchcxx REQUIRED )
target_link_libraries( my_target PUBLIC exchcxx::exchcxx )

# ExchCXX as CMake Subproject
include(FetchContent)

# Set ExchCXX CMake options (see below)

# Pull master branch of ExchCXX
FetchContent_Declare( exchcxx 
  GIT_REPOSITORY https://github.com/wavefunction91/ExchCXX.git 
  GIT_TAG master 
)
FetchContent_MakeAvailable( exchcxx )

# Link to target
target_link_libraries( my_target PUBLIC exchcxx::exchcxx )

CMake Variables

Variable Name Description Default
EXCHCXX_ENABLE_TESTS Enable Testing Framework (Catch2) ON
EXCHCXX_ENABLE_BENCHMARK Enable Performance Benchmark OFF
EXCHCXX_ENABLE_CUDA Enable CUDA XC evaluator OFF
EXCHCXX_ENABLE_HIP Enable HIP XC evaluator OFF
EXCHCXX_ENABLE_SYCL Enable SYCL XC evaluator (experimental) OFF

N.B. ExchCXX accelerator bindings are mutally exclusive - this is intentially and will be enforced on configure. If it would be desirable to support coexistance of these bindings for your application, please open an issue

Build Notes

Building with CUDA

For CUDA builds, CMAKE_CUDA_ARCHITECTURES must be set. For example, to build a fat binary for NVIDIA V100 and A100,

cmake [source locations] -DCMAKE_CUDA_ARCHITECTURES="70;80" [additional options]

ExchCXX supports NVIDIA GPUs with CC >= 60.

Building with SYCL

ExchCXX support for SYCL is experimental, and the mechanism for enabling CMake support is likely to change in the future. SYCL builds require C++17 per the SYCL-2020 standard. Due to the volitility in SYCL compiler development, ExchCXX does not provide automatic means to determine the SYCL targets (e.g. NVIDIA-PTX, OpenMP, etc) for the platform of interest. As such, these must be specified manually in CMAKE_CXX_FLAGS. For example, building for a NVIDIA-PTX target with the DPC++ compiler may be achieved via

cmake [source locations] -DCMAKE_CXX_FLAGS="-fsycl-targets=nvptx64-nvidia-cuda -fsycl" [additional options]