Merge branch 'release-2.0.1'

CMakeLists.txt

@@ -59,6 +59,13 @@ if (USE_MPI)
   get_property(REQUIRED_MPI_COMPONENTS GLOBAL PROPERTY ENABLED_LANGUAGES)
   list(REMOVE_ITEM REQUIRED_MPI_COMPONENTS CUDA)  # CUDA does not have a MPI component
   find_package(MPI COMPONENTS ${REQUIRED_MPI_COMPONENTS} REQUIRED)
+
+  if (NOT MPI_Fortran_HAVE_F90_MODULE)
+    message(FATAL_ERROR "\
+The listed MPI implementation does not provide the required mpi.mod interface. \
+When using the GNU compiler in combination with Intel MPI, please use the \
+Intel MPI compiler wrappers. Check the INSTALL.md for more information.")
+  endif ()
 endif ()
 
 if (USE_SMM MATCHES "blas")

INSTALL.md

@@ -0,0 +1,165 @@
+# DBCSR CMake Build Recipes
+
+Following are recipes for different combinations of compilers, platforms and libraries.
+Unless otherwise noted, the examples assume that after fetching/unpacking DBCSR you created
+a directory `build/` inside the DBCSR directory and switched into it using `cd build/`.
+
+The listed examples can usually be combined with other build options like *libxsmm* or *CUDA*
+even if the examples are not explicitly given.
+
+The instructions used for building in the Continuous Integration can be found in
+the `.ci/` folder or in the `.travis.yml`.
+
+
+## GNU
+
+
+### GNU compiler, system MPI and system-provided OpenBLAS
+
+Most Linux systems provide the GNU compiler, a system MPI (OpenMPI or MPICH) using the
+the GNU compiler as a backend and OpenBLAS for BLAS/LAPACK:
+
+    cmake ..
+
+
+### GNU compiler, system MPI and Intel MKL
+
+To use the Intel MKL together with the GNU compiler and possibly a system-MPI,
+assuming that MKL is installed in `/sw/intel/mkl`.
+
+Verified with MKL provided as part of the Intel Parallel Studio XE 2019.5 installed in `/sw/intel`
+with an OS-provided GCC 7.4.1 on Linux openSUSE Leap 15.1, using CMake 3.10.2.
+
+1. Make sure the MKL environment is properly loaded:
+
+       source /sw/intel/mkl/bin/mklvars.sh intel64
+
+2. Make sure CMake picks the Intel MKL over any system-provided BLAS library:
+
+       cmake -DBLA_VENDOR=Intel10_64lp_seq ..
+
+
+## Intel
+
+Instructions for using Intel compiler or libraries for different parts on non-Cray systems.
+For Cray systems, please check further below.
+
+*Note*: in Intel Parallel Studio 2019 there is a potential issue that `mpirun` fails with
+the error `OFI addrinfo() failed` on local (non-cluster) installations.
+This can be worked around by setting `export I_MPI_FABRICS=shm`.
+
+### Intel MPI, GNU Compiler and system-provided OpenBLAS
+
+Verified with Intel Parallel Studio XE 2019.5 installed in `/sw/intel`
+with an OS-provided GCC 7.4.1 on Linux openSUSE Leap 15.1, using CMake 3.10.2.
+
+1. Make sure that the Intel environment is properly loaded:
+
+       source /sw/intel/bin/compilervars.sh intel64
+
+2. Use the Intel-provided MPI compiler wrappers for the GNU toolchain,
+   to override CMake's auto-detection which may pick up the system MPI:
+
+       CC=mpicc FC=mpifc CXX=mpicxx cmake ..
+
+
+### Intel MPI, GNU Compiler and Intel MKL
+
+Verified with Intel Parallel Studio XE 2019.5 installed in `/sw/intel`
+with an OS-provided GCC 7.4.1 on Linux openSUSE Leap 15.1, using CMake 3.10.2.
+
+1. Make sure that the Intel environment is properly loaded:
+
+       source /sw/intel/bin/compilervars.sh intel64
+
+2. Use the Intel-provided MPI compiler wrappers for the GNU toolchain:
+
+       CC=mpicc FC=mpifc CXX=mpicxx cmake -DBLA_VENDOR=Intel10_64lp_seq ..
+
+
+### Intel MPI, Intel Compiler and Intel MKL
+
+Verified with Intel Parallel Studio XE 2019.5 installed in `/sw/intel`
+on Linux openSUSE Leap 15.1, using CMake 3.10.2.
+
+1. Make sure that the Intel environment is properly loaded:
+
+       source /sw/intel/bin/compilervars.sh intel64
+
+2. Use the Intel-provided MPI compiler wrappers:
+
+       CC=mpiicc FC=mpiifort CXX=mpiicxx cmake -DBLA_VENDOR=Intel10_64lp_seq ..
+
+
+## Cray
+
+Some machines require additional environments to be loaded to either provide
+the modules specified below or to be able to properly build with the loaded modules.
+
+Please contact your cluster/datacenter administrator for more information.
+
+Example for the CSCS' Piz Daint:
+
+    module load daint-mc  # to build for the non-GPU partition
+    module load daint-gpu  # to build for the GPU partition
+
+*Note*: the `libsci-cray` has different variants for MPI or OpenMP.
+When disabling either MPI or OpenMP support in DBCSR you might want to adjust the
+selected BLAS/LAPACK library accordingly (e.g. drop the `_mpi`, or `_mp`).
+
+
+### CCE and libsci-cray
+
+Verified on CSCS' Piz Daint with CCE 9.0.2 and cray-libsci 19.06.1,
+using CMake 3.14.5.
+
+1. Make sure that the `PrgEnv-cray` module is loaded:
+
+       module load PrgEnv-cray
+
+2. While the MPI wrapper/compiler will be detected automatically,
+   must the BLAS/LAPACK libraries be specified manually:
+
+       cmake \
+         -DCMAKE_SYSTEM_NAME=CrayLinuxEnvironment \
+         -DBLAS_LIBRARIES="-lsci_cray_mpi_mp -lhugetlbfs" \
+         -DLAPACK_LIBRARIES="-lsci_cray_mpi_mp" \
+         ..
+
+
+### Intel Compiler and libsci-cray
+
+Verified on CSCS' Piz Daint with Intel 19.01 and cray-libsci 19.06.1,
+using CMake 3.14.5.
+
+1. Make sure that the `PrgEnv-intel` module is loaded:
+
+       module load PrgEnv-intel
+
+2. While the MPI wrapper/compiler will be detected automatically,
+   must the BLAS/LAPACK libraries be specified manually:
+
+       cmake \
+         -DCMAKE_SYSTEM_NAME=CrayLinuxEnvironment \
+         -DBLAS_LIBRARIES="-lsci_intel_mpi_mp -lhugetlbfs" \
+         -DLAPACK_LIBRARIES="-lsci_intel_mpi_mp" \
+         ..
+
+
+### GNU Compiler and libsci-cray
+
+Verified on CSCS' Piz Daint with GNU 8.3.0 and cray-libsci 19.06.1,
+using CMake 3.14.5.
+
+1. Make sure that the `PrgEnv-gnu` module is loaded:
+
+       module load PrgEnv-gnu
+
+2. While the MPI wrapper/compiler will be detected automatically,
+   must the BLAS/LAPACK libraries be specified manually:
+
+       cmake \
+         -DCMAKE_SYSTEM_NAME=CrayLinuxEnvironment \
+         -DBLAS_LIBRARIES="-lsci_gnu_mpi_mp -lhugetlbfs" \
+         -DLAPACK_LIBRARIES="-lsci_gnu_mpi_mp" \
+         ..

README.md

@@ -83,11 +83,7 @@ The configuration flags are (default first):
     -DTEST_OMP_THREADS=<2,N>
     -DCMAKE_BUILD_TYPE=<Release|Debug|Coverage>
 
-
-Building on Cray requires an additional flag to make sure CMake is able to properly
-detect the compiler behind the compiler wrappers:
-
-    -DCMAKE_SYSTEM_NAME=CrayLinuxEnvironment
+For build recipes on different platforms, make sure to also read the [INSTALL.md](INSTALL.md).
 
 ## Using DBCSR in a CMake project
 

VERSION

@@ -1,6 +1,6 @@
 MAJOR = 2
 MINOR = 0
-PATCH = 0
+PATCH = 1
 # A specific DATE (YYYY-MM-DD) fixes an official release, otherwise
 # it is considered Development version.
-DATE  = 2019-11-11
+DATE  = 2019-11-15

src/acc/libsmm_acc/libcusmm/PACKAGE

@@ -1,5 +1,5 @@
 {
 "description": "Cuda accelerated Small Matrix Multiplications",
 "archive": "libdbcsr",
-"requires": ["kernels", "../include",]
+"requires": ["kernels", "../include", "../../include"]
 }

src/mpi/dbcsr_mpiwrap.F

@@ -22,8 +22,8 @@ MODULE dbcsr_mpiwrap
 #define __MPI_VERSION 3
 #endif
 
-#if defined(__parallel) && ! defined(__HAS_NO_MPI_MOD)
-   USE mpi  ! compiler *errors* mean mpi installation and fortran compiler mismatch: see INSTALL (-D__HAS_NO_MPI_MOD)
+#if defined(__parallel)
+   USE mpi
 ! subroutines: unfortunately, mpi implementations do not provide interfaces for all subroutines (problems with types and ranks explosion),
 !              we do not quite know what is in the module, so we can not include any....
 !              to nevertheless get checking for what is included, we use the mpi module without use clause, getting all there is
@@ -51,16 +51,11 @@ MODULE dbcsr_mpiwrap
 
    ! parameters that might be needed
 #if defined(__parallel)
-#if defined(__HAS_NO_MPI_MOD)
-   INCLUDE "mpif.h"
-#endif
    INTEGER, PARAMETER     :: MP_STD_REAL = MPI_DOUBLE_PRECISION
    INTEGER, PARAMETER     :: MP_STD_COMPLEX = MPI_DOUBLE_COMPLEX
    INTEGER, PARAMETER     :: MP_STD_HALF_REAL = MPI_REAL
    INTEGER, PARAMETER     :: MP_STD_HALF_COMPLEX = MPI_COMPLEX
-#endif
 
-#ifdef __parallel
    LOGICAL, PARAMETER :: dbcsr_is_parallel = .TRUE.
    INTEGER, PARAMETER, PUBLIC :: mp_any_tag = MPI_ANY_TAG
    INTEGER, PARAMETER, PUBLIC :: mp_any_source = MPI_ANY_SOURCE