Skip to content

Latest commit

 

History

History
267 lines (230 loc) · 13.3 KB

INSTALL.md

File metadata and controls

267 lines (230 loc) · 13.3 KB
Raw

Installation Instructions {#install_ginkgo}

Building

Use the standard cmake build procedure:

mkdir build; cd build
cmake -G "Unix Makefiles" [OPTIONS] .. && make

Use cmake --build . in some systems like MinGW or Microsoft Visual Studio which do not use make.

For Microsoft Visual Studio, use cmake --build . --config <build_type> to decide the build type. The possible options are Debug, Release, RelWithDebInfo and MinSizeRel.

Replace [OPTIONS] with desired cmake options for your build. Ginkgo adds the following additional switches to control what is being built:

  • -DGINKGO_DEVEL_TOOLS={ON, OFF} sets up the build system for development (requires clang-format, will also download git-cmake-format), default is OFF.

  • -DGINKGO_BUILD_TESTS={ON, OFF} builds Ginkgo's tests (will download googletest), default is ON.

  • -DGINKGO_BUILD_BENCHMARKS={ON, OFF} builds Ginkgo's benchmarks (will download gflags and rapidjson), default is ON.

  • -DGINKGO_BUILD_EXAMPLES={ON, OFF} builds Ginkgo's examples, default is ON

  • -DGINKGO_BUILD_EXTLIB_EXAMPLE={ON, OFF} builds the interfacing example with deal.II, default is OFF.

  • -DGINKGO_BUILD_REFERENCE={ON, OFF} build reference implementations of the kernels, useful for testing, default is ON

  • -DGINKGO_BUILD_OMP={ON, OFF} builds optimized OpenMP versions of the kernels, default is ON if the selected C++ compiler supports OpenMP, OFF otherwise.

  • -DGINKGO_BUILD_CUDA={ON, OFF} builds optimized cuda versions of the kernels (requires CUDA), default is ON if a CUDA compiler could be detected, OFF otherwise.

  • -DGINKGO_BUILD_HIP={ON, OFF} builds optimized HIP versions of the kernels (requires HIP), default is ON if an installation of HIP could be detected, OFF otherwise.

  • -DGINKGO_HIP_AMDGPU="gpuarch1;gpuarch2" the amdgpu_target(s) variable passed to hipcc for the hcc HIP backend. The default is none (auto).

  • -DGINKGO_BUILD_DOC={ON, OFF} creates an HTML version of Ginkgo's documentation from inline comments in the code. The default is OFF.

  • -DGINKGO_DOC_GENERATE_EXAMPLES={ON, OFF} generates the documentation of examples in Ginkgo. The default is ON.

  • -DGINKGO_DOC_GENERATE_PDF={ON, OFF} generates a PDF version of Ginkgo's documentation from inline comments in the code. The default is OFF.

  • -DGINKGO_DOC_GENERATE_DEV={ON, OFF} generates the developer version of Ginkgo's documentation. The default is OFF.

  • -DGINKGO_EXPORT_BUILD_DIR={ON, OFF} adds the Ginkgo build directory to the CMake package registry. The default is OFF.

  • -DGINKGO_WITH_CLANG_TIDY={ON, OFF} makes Ginkgo call clang-tidy to find programming issues. The path can be manually controlled with the CMake variable -DGINKGO_CLANG_TIDY_PATH=<path>. The default is OFF.

  • -DGINKGO_WITH_IWYU={ON, OFF} makes Ginkgo call iwyu to find include issues. The path can be manually controlled with the CMake variable -DGINKGO_IWYU_PATH=<path>. The default is OFF.

  • -DGINKGO_VERBOSE_LEVEL=integer sets the verbosity of Ginkgo.

    • 0 disables all output in the main libraries,
    • 1 enables a few important messages related to unexpected behavior (default).

  • -DCMAKE_INSTALL_PREFIX=path sets the installation path for make install. The default value is usually something like /usr/local.

  • -DCMAKE_BUILD_TYPE=type specifies which configuration will be used for this build of Ginkgo. The default is RELEASE. Supported values are CMake's standard build types such as DEBUG and RELEASE and the Ginkgo specific COVERAGE, ASAN (AddressSanitizer) and TSAN (ThreadSanitizer) types.

  • -DBUILD_SHARED_LIBS={ON, OFF} builds ginkgo as shared libraries (OFF) or as dynamic libraries (ON), default is ON.

  • -DGINKGO_JACOBI_FULL_OPTIMIZATIONS={ON, OFF} use all the optimizations for the CUDA Jacobi algorithm. OFF by default. Setting this option to ON may lead to very slow compile time (>20 minutes) for the jacobi_generate_kernels.cu file and high memory usage.

  • -DCMAKE_CUDA_HOST_COMPILER=path instructs the build system to explicitly set CUDA's host compiler to the path given as argument. By default, CUDA uses its toolchain's host compiler. Setting this option may help if you're experiencing linking errors due to ABI incompatibilities. This option is supported since CMake 3.8 but documented starting from 3.10.

  • -DGINKGO_CUDA_ARCHITECTURES=<list> where <list> is a semicolon (;) separated list of architectures. Supported values are:

    • Auto
    • Kepler, Maxwell, Pascal, Volta
    • CODE, CODE(COMPUTE), (COMPUTE)

    Auto will automatically detect the present CUDA-enabled GPU architectures in the system. Kepler, Maxwell, Pascal and Volta will add flags for all architectures of that particular NVIDIA GPU generation. COMPUTE and CODE are placeholders that should be replaced with compute and code numbers (e.g. for compute_70 and sm_70 COMPUTE and CODE should be replaced with 70. Default is Auto. For a more detailed explanation of this option see the ARCHITECTURES specification list section in the documentation of the CudaArchitectureSelector CMake module.

  • -DGINKGO_WINDOWS_SHARED_LIBRARY_RELPATH=<path> where is a relative path built with PROJECT_BINARY_DIR. Users must add the absolute path (PROJECT_BINARY_DIR/GINKGO_WINDOWS_SHARED_LIBRARY_RELPATH) into the environment variable PATH when building shared libraries and executable program, default is windows_shared_library.

  • -DGINKGO_CHECK_PATH={ON, OFF} checks if the environment variable PATH is valid. It is checked only when building shared libraries and executable program, default is ON.

For example, to build everything (in debug mode), use:

cmake  -G "Unix Makefiles" -H. -BDebug -DCMAKE_BUILD_TYPE=Debug -DGINKGO_DEVEL_TOOLS=ON \
    -DGINKGO_BUILD_TESTS=ON -DGINKGO_BUILD_REFERENCE=ON -DGINKGO_BUILD_OMP=ON \
    -DGINKGO_BUILD_CUDA=ON -DGINKGO_BUILD_HIP=ON
cmake --build Debug

NOTE: Ginkgo is known to work with the Unix Makefiles and Ninja based generators. Other CMake generators are untested.

Building Ginkgo with HIP support

Ginkgo provides a HIP backend. This allows to compile optimized versions of the kernels for either AMD or NVIDIA GPUs. The CMake configuration step will try to auto-detect the presence of HIP either at /opt/rocm/hip or at the path specified by HIP_PATH as a CMake parameter (-DHIP_PATH=) or environment variable (export HIP_PATH=), unless -DGINKGO_BUILD_HIP=ON/OFF is set explicitly.

Correctly installing HIP toolkits and dependencies for Ginkgo

In general, Ginkgo's HIP backend requires the following packages:

  • HIP,
  • hipBLAS,
  • hipSPARSE,
  • Thrust.

It is necessary to provide some details about the different ways to procure and install these packages, in particular for NVIDIA systems since getting a correct, non bloated setup is not straightforward.

For AMD systems, the simplest way is to follow the instructions provided here which provide package installers for most Linux distributions. Ginkgo also needs the installation of the hipBLAS and hipSPARSE interfaces. Optionally if you do not already have a thrust installation, the ROCm provided rocThrust package can be used.

For NVIDIA systems, the traditional installation (package hip_nvcc), albeit working properly is currently odd: it depends on all the hcc related packages, although the nvcc backend seems to entirely rely on the CUDA suite. See this issue for more details. It is advised in this case to compile everything manually, including using forks of hipBLAS and hipSPARSE specifically made to not depend on the hcc specific packages. Thrust is often provided by CUDA and this Thrust version should work with HIP. Here is a sample procedure for installing HIP, hipBLAS and hipSPARSE.

# HIP
git clone https://github.com/ROCm-Developer-Tools/HIP.git
pushd HIP && mkdir build && pushd build
cmake .. && make install
popd && popd

# hipBLAS
git clone https://github.com/tcojean/hipBLAS.git
pushd hipBLAS && mkdir build && pushd build
cmake .. && make install
popd && popd

# hipSPARSE
git clone https://github.com/tcojean/hipSPARSE.git
pushd hipSPARSE && mkdir build && pushd build
cmake -DBUILD_CUDA=ON .. && make install
popd && popd

Changing the paths to search for HIP and other packages

All HIP installation paths can be configured through the use of environment variables or CMake variables. This way of configuring the paths is currently imposed by the HIP tool suite. The variables are the following:

  • CMake -DHIP_PATH= or environment export HIP_PATH=: sets the HIP installation path. The default value is /opt/rocm/hip.
  • CMake -DHIPBLAS_PATH= or environment export HIPBLAS_PATH=: sets the hipBLAS installation path. The default value is /opt/rocm/hipblas.
  • CMake -DHIPSPARSE_PATH= or environment export HIPSPARSE_PATH=: sets the hipSPARSE installation path. The default value is /opt/rocm/hipsparse.
  • CMake -DHCC_PATH= or environment export HCC_PATH=: sets the HCC installation path, for AMD backends. The default value is /opt/rocm/hcc.
  • environment export CUDA_PATH=: where hipcc can find CUDA if it is not in the default /usr/local/cuda path.

HIP platform detection of AMD and NVIDIA

By default, Ginkgo uses the output of /opt/rocm/hip/bin/hipconfig --platform to select the backend. The accepted values are either hcc (AMD) or nvcc (NVIDIA). When on an AMD or NVIDIA system, this should output the correct platform by default. When on a system without GPUs, this should output hcc by default. To change this value, export the environment variable HIP_PLATFORM like so:

export HIP_PLATFORM=nvcc

Setting platform specific compilation flags

Platform specific compilation flags can be given through the following CMake variables:

  • -DGINKGO_HIP_COMPILER_FLAGS=: compilation flags given to all platforms.
  • -DGINKGO_HIP_HCC_COMPILER_FLAGS=: compilation flags given to AMD platforms.
  • -DGINKGO_HIP_NVCC_COMPILER_FLAGS=: compilation flags given to NVIDIA platforms.

Third party libraries and packages

Ginkgo relies on third party packages in different cases. These third party packages can be turned off by disabling the relevant options.

  • GINKGO_BUILD_CUDA=ON: CudaArchitectureSelector (CAS) is a CMake helper to manage CUDA architecture settings;
  • GINKGO_BUILD_TESTS=ON: Our tests are implemented with Google Test;
  • GINKGO_BUILD_BENCHMARKS=ON: For argument management we use gflags and for JSON parsing we use RapidJSON;
  • GINKGO_DEVEL_TOOLS=ON: git-cmake-format is our CMake helper for code formatting.

By default, Ginkgo uses the internal version of each package. For each of the packages GTEST, GFLAGS, RAPIDJSON and CAS, it is possible to force Ginkgo to try to use an external version of a package. For this, Ginkgo provides two ways to find packages. To rely on the CMake find_package command, use the CMake option -DGINKGO_USE_EXTERNAL_<package>=ON. Note that, if the external packages were not installed to the default location, the CMake option -DCMAKE_PREFIX_PATH=<path-list> needs to be set to the semicolon (;) separated list of install paths of these external packages. For more Information, see the CMake documentation for CMAKE_PREFIX_PATH for details.

To manually configure the paths, Ginkgo relies on the standard xSDK Installation policies for all packages except CAS (as it is neither a library nor a header, it cannot be expressed through the TPL format):

  • -DTPL_ENABLE_<package>=ON
  • -DTPL_<package>_LIBRARIES=/path/to/libraries.{so|a}
  • -DTPL_<package>_INCLUDE_DIRS=/path/to/header/directory

When applicable (e.g. for GTest libraries), a ; separated list can be given to the TPL_<package>_{LIBRARIES|INCLUDE_DIRS} variables.

Installing Ginkgo

To install Ginkgo into the specified folder, execute the following command in the build folder

make install

If the installation prefix (see CMAKE_INSTALL_PREFIX) is not writable for your user, e.g. when installing Ginkgo system-wide, it might be necessary to prefix the call with sudo.

After the installation, CMake can find ginkgo with find_package(Ginkgo). An example can be found in the test_install.