Merge branch 'release-2.5.0'

.ci/daint.cscs.ch/ocl.build.sh

@@ -8,25 +8,26 @@
 #SBATCH --cpus-per-task=12
 #SBATCH --hint=nomultithread
 
+#set -o nounset
 set -o errexit
-set -o nounset
 set -o pipefail
 
+source /opt/intel/oneapi/mkl/latest/env/vars.sh
 module swap PrgEnv-cray PrgEnv-gnu
 module load daint-gpu cudatoolkit cdt-cuda
-module unload cray-libsci_acc
+module unload cray-libsci_acc cray-libsci
 module list
 
-export PATH=/project/cray/alazzaro/cmake/bin:$PATH
+export PATH=/project/cray/alazzaro/cmake/bin:${PATH}
 
 # Checkout and build LIBXSMM
 if [ ! -d "${HOME}/libxsmm" ]; then
   cd "${HOME}"
-  git clone https://github.com/hfp/libxsmm.git
+  git clone https://github.com/libxsmm/libxsmm.git
 fi
 cd "${HOME}/libxsmm"
 git fetch
-git checkout 1a10386117eb0d6771bcd512c1cd00860424477f
+git checkout 593a64cb87a0643f9b275b1081c096878b190a0e
 make -j
 cd ..
 
@@ -43,13 +44,17 @@ cd "${SCRATCH}/${BUILD_TAG}.ocl"
 # CMake: find LIBXSMM (pkg-config)
 export PKG_CONFIG_PATH=${HOME}/libxsmm/lib:${PKG_CONFIG_PATH}
 
+#BLAS="-DBLAS_FOUND=ON -DBLAS_LIBRARIES='-lsci_gnu_mpi_mp' -DLAPACK_FOUND=ON -DLAPACK_LIBRARIES='-lsci_gnu_mpi_mp'"
+BLAS="-DBLA_VENDOR=Intel10_64lp"
+#LIBXSMM=libxsmm-shared
+LIBXSMM=libxsmm
+
 cmake \
     -DCMAKE_SYSTEM_NAME=CrayLinuxEnvironment \
     -DCMAKE_CROSSCOMPILING_EMULATOR="" \
-    -DUSE_ACCEL=opencl -DWITH_GPU=P100 -DUSE_SMM=libxsmm \
+    -DUSE_ACCEL=opencl -DWITH_GPU=P100 \
+    -DUSE_SMM=${LIBXSMM} ${BLAS} \
     -DOpenCL_LIBRARY="${CUDATOOLKIT_HOME}/lib64/libOpenCL.so" \
-    -DBLAS_FOUND=ON -DBLAS_LIBRARIES="-lsci_gnu_mpi_mp" \
-    -DLAPACK_FOUND=ON -DLAPACK_LIBRARIES="-lsci_gnu_mpi_mp" \
     -DMPIEXEC_EXECUTABLE="$(command -v srun)" \
     -DTEST_MPI_RANKS="${SLURM_NTASKS}" \
     -DTEST_OMP_THREADS="${SLURM_CPUS_PER_TASK}" \

.ci/daint.cscs.ch/ocl.test.sh

@@ -7,16 +7,17 @@
 #SBATCH --ntasks-per-node=1
 #SBATCH --hint=nomultithread
 
+#set -o nounset
 set -o errexit
-set -o nounset
 set -o pipefail
 
+source /opt/intel/oneapi/mkl/latest/env/vars.sh
 module swap PrgEnv-cray PrgEnv-gnu
 module load daint-gpu cudatoolkit cdt-cuda
-module unload cray-libsci_acc
+module unload cray-libsci_acc cray-libsci
 module list
 
-export PATH=/project/cray/alazzaro/cmake/bin:$PATH
+export PATH=/project/cray/alazzaro/cmake/bin:${PATH}
 
 set -o xtrace  # do not set earlier to avoid noise from module
 
@@ -26,6 +27,7 @@ mkdir -p "${SCRATCH}/${BUILD_TAG}.ocl"
 chmod 0775 "${SCRATCH}/${BUILD_TAG}.ocl"
 cd "${SCRATCH}/${BUILD_TAG}.ocl"
 
+export LD_LIBRARY_PATH=${HOME}/libxsmm/lib:${LD_LIBRARY_PATH}
 export OMP_PROC_BIND=TRUE # set thread affinity
 # OMP_NUM_THREADS is set by cmake
 

CMakeLists.txt

@@ -103,6 +103,7 @@ set(WITH_GPU
       STRING
       "Select GPU arch. and embed parameters (default: CUDA/HIP=P100, OPENCL=all)"
 )
+set(WITH_GPU_PARAMS "${WITH_GPU}")
 set_property(CACHE WITH_GPU PROPERTY STRINGS ${SUPPORTED_CUDA_ARCHITECTURES}
                                      ${SUPPORTED_HIP_ARCHITECTURES})
 
@@ -143,9 +144,19 @@ endif ()
 if (USE_SMM MATCHES "libxsmm")
   find_package(PkgConfig REQUIRED)
   if (USE_OPENMP)
-    pkg_check_modules(LIBXSMMEXT REQUIRED IMPORTED_TARGET GLOBAL libxsmmext)
+    if (NOT USE_SMM MATCHES "libxsmm-shared")
+      pkg_check_modules(LIBXSMMEXT IMPORTED_TARGET GLOBAL libxsmmext-static)
+    endif ()
+    if (NOT LIBXSMMEXT_FOUND)
+      pkg_check_modules(LIBXSMMEXT REQUIRED IMPORTED_TARGET GLOBAL libxsmmext)
+    endif ()
+  endif ()
+  if (NOT USE_SMM MATCHES "libxsmm-shared")
+    pkg_check_modules(LIBXSMM IMPORTED_TARGET GLOBAL libxsmmf-static)
+  endif ()
+  if (NOT LIBXSMM_FOUND)
+    pkg_check_modules(LIBXSMM REQUIRED IMPORTED_TARGET GLOBAL libxsmmf)
   endif ()
-  pkg_check_modules(LIBXSMM REQUIRED IMPORTED_TARGET GLOBAL libxsmmf)
 endif ()
 
 # =================================== BLAS & LAPACK, PkgConfig
@@ -233,6 +244,7 @@ if (USE_ACCEL MATCHES "cuda")
   set(ACC_ARCH_NUMBER ${GPU_ARCH_NUMBER_${WITH_GPU}})
 
   message(STATUS "GPU target architecture: " ${WITH_GPU})
+  message(STATUS "Kernel parameters: " ${WITH_GPU_PARAMS})
   message(STATUS "GPU architecture number: " ${ACC_ARCH_NUMBER})
   message(STATUS "GPU profiling enabled: " ${WITH_CUDA_PROFILING})
 endif ()
@@ -271,6 +283,7 @@ if (USE_ACCEL MATCHES "hip")
 
   set(ACC_ARCH_NUMBER ${GPU_ARCH_NUMBER_${WITH_GPU}})
   message(STATUS "GPU target architecture: " ${WITH_GPU})
+  message(STATUS "Kernel parameters: " ${WITH_GPU_PARAMS})
   message(STATUS "GPU architecture number: " ${ACC_ARCH_NUMBER})
   message(STATUS "GPU profiling enabled: " ${WITH_HIP_PROFILING})
 

VERSION

@@ -1,8 +1,8 @@
 MAJOR = 2
-MINOR = 4
-PATCH = 1
+MINOR = 5
+PATCH = 0
 # A specific DATE (YYYY-MM-DD) fixes an official release, otherwise
 # it is considered Development version.
-DATE  = 2022-08-29
+DATE  = 2022-12-27
 
 

docs/guide/3-developer-guide/3-programming/1-overview/index.md

@@ -41,7 +41,6 @@ Assumed square matrix with 20x20 matrix with 5x5 blocks and a 2x2 processor grid
 | Macro | Explanation | Language |
 |-|-|-|
 | `__parallel` | Enable MPI runs | Fortran |
-| `__MPI_VERSION=N` | DBCSR assumes that the MPI library implements MPI version 3. If you have an older version of MPI (e.g. MPI 2.0) available you must define `-D__MPI_VERSION=2` | Fortran |
 | `__NO_MPI_THREAD_SUPPORT_CHECK` | Workaround for MPI libraries that do not declare they are thread safe (funneled) but you want to use them with OpenMP code anyways | Fortran |
 | `__MKL` | Enable use of optimized Intel MKL functions | Fortran
 | `__NO_STATM_ACCESS`, `__STATM_RESIDENT` or `__STATM_TOTAL` | Toggle memory usage reporting between resident memory and total memory. In particular, macOS users must use `-D__NO_STATM_ACCESS` | Fortran |

src/CMakeLists.txt

@@ -161,6 +161,7 @@ set_target_properties(dbcsr PROPERTIES VERSION ${dbcsr_VERSION}
 
 if (USE_SMM MATCHES "libxsmm")
   target_compile_definitions(dbcsr PRIVATE __LIBXSMM)
+  target_link_directories(dbcsr PUBLIC ${LIBXSMM_LIBRARY_DIRS})
   if (USE_OPENMP)
     target_link_libraries(dbcsr PRIVATE PkgConfig::LIBXSMMEXT)
   endif ()
@@ -200,8 +201,7 @@ set_target_properties(dbcsr PROPERTIES LINKER_LANGUAGE Fortran)
 if (MPI_FOUND)
   # once built, a user of the dbcsr library can not influence anything anymore
   # by setting those flags:
-  target_compile_definitions(
-    dbcsr PRIVATE __parallel __MPI_VERSION=${MPI_Fortran_VERSION_MAJOR})
+  target_compile_definitions(dbcsr PRIVATE __parallel)
 
   # Instead of resetting the compiler for MPI, we are adding the compiler flags
   # otherwise added by the mpifort-wrapper directly; based on hints from: