Merge branch 'main' into julia

CHANGELOG.md

@@ -15,6 +15,7 @@ All notable changes to this project will be documented in this file.
 - CMake build system added for all models.
 - SYCL device check for FP64 support.
 - New implementation using TBB.
+- Compiler options for Fujitsu added to OpenMP.
 
 ### Changed
 - Default branch renamed from `master` to `main`.

OMP.cmake

@@ -30,6 +30,9 @@
 #    ARMClang = ARM Compiler based on Clang (arm.com)
 # These are only added in CMake 3.20:
 #    NVHPC = NVIDIA HPC SDK Compiler (nvidia.com)
+# These are only added in CMake 3.21
+#    Fujitsu = Fujitsu HPC compiler (Trad mode)
+#    FujitsuClang = Fujitsu HPC compiler (Clang mode)
 
 
 # CMAKE_SYSTEM_PROCESSOR is set via `uname -p`, we have:
@@ -39,8 +42,9 @@
 #
 
 
-#predefined offload flags based on compiler id
+# predefined offload flags based on compiler id and vendor,
-
+# the format is (COMPILER and VENDOR must be UPPERCASE):
+# Compiler: OMP_FLAGS_OFFLOAD_<COMPILER?>_<VNEDOR?>
 
 set(OMP_FLAGS_OFFLOAD_INTEL
         -qnextgen -fiopenmp -fopenmp-targets=spir64)
@@ -56,15 +60,25 @@ set(OMP_FLAGS_OFFLOAD_CLANG_ARCH_FLAG
         -march=) # prefix only, arch appended by the vendor:arch tuple
 
 
+# for standard (non-offload) omp, the format is (COMPILER and ARCH must be UPPERCASE):
+# Compiler:      OMP_FLAGS_CPU_<COMPILER?>_<ARCH?>
+# Linker:   OMP_LINK_FLAGS_CPU_<COMPILER?>_<ARCH?>
+
+set(OMP_FLAGS_CPU_FUJITSU
+        -Kfast -std=c++11 -KA64FX -KSVE -KARMV8_3_A -Kzfill=100 -Kprefetch_sequential=soft -Kprefetch_line=8 -Kprefetch_line_L2=16)
+set(OMP_LINK_FLAGS_CPU_FUJITSU
+        -Kopenmp)
+
 set(OMP_FLAGS_CPU_INTEL
         -qopt-streaming-stores=always)
+
 set(OMP_FLAGS_CPU_GNU_PPC64LE
         -mcpu=native)
+
 set(OMP_FLAGS_CPU_XL
         -O5 -qarch=auto -qtune=auto)
 
-# NEC
+set(OMP_FLAGS_CPU_NEC -O4 -finline) # CMake doesn't detect this so it's meant to be chosen by register_flag_optional(ARCH)
-set(OMP_FLAGS_CPU_NEC -O4 -finline)
 
 register_flag_optional(CMAKE_CXX_COMPILER
         "Any CXX compiler that supports OpenMP as per CMake detection (and offloading if enabled with `OFFLOAD`)"
@@ -122,6 +136,12 @@ macro(setup)
                 ${ARCH}
         )
 
+        register_append_compiler_and_arch_specific_link_flags(
+                OMP_LINK_FLAGS_CPU
+                ${COMPILER}
+                ${ARCH}
+        )
+
     elseif ("${OFFLOAD}" STREQUAL ON)
         #  offload but with custom flags
         register_definitions(OMP_TARGET_GPU)

OpenMP.make

@@ -3,7 +3,8 @@ ifndef COMPILER
 define compiler_help
 Set COMPILER to change flags (defaulting to GNU).
 Available compilers are:
-  CLANG CRAY GNU GNU_PPC INTEL XL PGI NEC ARMCLANG AOMP
+  CLANG CRAY GNU GNU_PPC INTEL XL PGI
+  NEC ARMCLANG AOMP FUJITSU
 
 Note: GCC on PPC requires -mcpu=native instead of -march=native so we have a special case for it
 
@@ -49,6 +50,7 @@ COMPILER_XL = xlc++
 COMPILER_PGI = pgc++
 COMPILER_NEC = /opt/nec/ve/bin/nc++
 COMPILER_AOMP = clang++
+COMPILER_FUJITSU=FCC
 CXX = $(COMPILER_$(COMPILER))
 
 FLAGS_GNU = -O3 -std=c++11 -march=native
@@ -61,6 +63,7 @@ FLAGS_PGI = -O3 -std=c++11
 FLAGS_NEC = -O4 -finline -std=c++11
 FLAGS_ARMCLANG = -O3 -std=c++11
 FLAGS_AOMP = -O3 -std=c++11
+FLAGS_FUJITSU=-Kfast -std=c++11 -KA64FX -KSVE -KARMV8_3_A -Kzfill=100 -Kprefetch_sequential=soft -Kprefetch_line=8 -Kprefetch_line_L2=16
 CXXFLAGS = $(FLAGS_$(COMPILER))
 
 # OpenMP flags for CPUs
@@ -73,6 +76,7 @@ OMP_CLANG_CPU = -fopenmp=libomp
 OMP_XL_CPU = -qsmp=omp -qthreaded
 OMP_PGI_CPU = -mp
 OMP_NEC_CPU = -fopenmp
+OMP_FUJITSU_CPU=-Kopenmp
 
 # OpenMP flags for NVIDIA
 OMP_CRAY_NVIDIA  = -DOMP_TARGET_GPU

ci-prepare-bionic.sh

@@ -134,20 +134,18 @@ setup_aocc() {
 
 setup_nvhpc() {
   echo "Preparing Nvidia HPC SDK"
-
   local tarball="nvhpc.tar.gz"
-#  local url="http://localhost:8000/nvhpc_2021_212_Linux_x86_64_cuda_11.2.tar.gz"
+#  local url="http://localhost:8000/nvhpc_2021_215_Linux_x86_64_cuda_11.3.tar.gz"
-  local url="https://developer.download.nvidia.com/hpc-sdk/21.2/nvhpc_2021_212_Linux_x86_64_cuda_11.2.tar.gz"
+  local url="https://developer.download.nvidia.com/hpc-sdk/21.5/nvhpc_2021_215_Linux_x86_64_cuda_11.3.tar.gz"
-
   get_and_untar "$tarball" "$url"
 
-  local sdk_dir="$PWD/nvhpc_2021_212_Linux_x86_64_cuda_11.2/install_components/Linux_x86_64/21.2"
+  local sdk_dir="$PWD/nvhpc_2021_215_Linux_x86_64_cuda_11.3/install_components/Linux_x86_64/21.5"
   local bin_dir="$sdk_dir/compilers/bin"
   "$bin_dir/makelocalrc" "$bin_dir" -x
 
   export_var NVHPC_NVCXX "$bin_dir/nvc++"
-  export_var NVHPC_NVCC "$sdk_dir/cuda/11.2/bin/nvcc"
+  export_var NVHPC_NVCC "$sdk_dir/cuda/11.3/bin/nvcc"
-  export_var NVHPC_CUDA_DIR "$sdk_dir/cuda/11.2"
+  export_var NVHPC_CUDA_DIR "$sdk_dir/cuda/11.3"
   echo "Installed CUDA versions:"
   ls "$sdk_dir/cuda"
   verify_bin_exists "$NVHPC_NVCXX"

ci-test-compile.sh

@@ -124,7 +124,7 @@ run_build() {
 
 AMD_ARCH="gfx_903"
 NV_ARCH="sm_70"
-NV_ARCH_CCXY="cuda11.2,cc80"
+NV_ARCH_CCXY="cuda11.3,cc80"
 
 build_gcc() {
   local name="gcc_build"

register_models.cmake

@@ -20,7 +20,9 @@
 #
 
 macro(wipe_gcc_style_optimisation_flags VAR)
-    string(REGEX REPLACE "([\\/\\-]O.)" "" ${VAR} ${${VAR}})
+    if(${VAR})
+        string(REGEX REPLACE "([\\/\\-]O.)" "" ${VAR} ${${VAR}})
+    endif()
 endmacro()
 
 macro(register_link_library)
@@ -41,7 +43,7 @@ macro(register_append_link_flags)
     list(APPEND LINK_FLAGS ${ARGN})
 endmacro()
 
-macro(register_append_compiler_and_arch_specific_cxx_flags PREFIX CXX ARCH)
+function(bind_cxx_and_arch OUT PREFIX CXX ARCH)
     string(TOUPPER ${CXX} _CXX)
     string(TOUPPER ${ARCH} _ARCH)
     set(_CXX_ARCH_SPECIFIC_FLAGS "${${PREFIX}_${_CXX}_${_ARCH}}")
@@ -52,6 +54,17 @@ macro(register_append_compiler_and_arch_specific_cxx_flags PREFIX CXX ARCH)
     if (_CXX_ARCH_SPECIFIC_FLAGS)
         register_append_cxx_flags(ANY ${_CXX_ARCH_SPECIFIC_FLAGS})
     endif ()
+    set(${OUT} "${_CXX_ARCH_SPECIFIC_FLAGS}" PARENT_SCOPE)
+endfunction()
+
+macro(register_append_compiler_and_arch_specific_cxx_flags PREFIX CXX ARCH)
+    bind_cxx_and_arch(OUT ${PREFIX} ${CXX} ${ARCH})
+    register_append_cxx_flags(ANY ${OUT})
+endmacro()
+
+macro(register_append_compiler_and_arch_specific_link_flags PREFIX CXX ARCH)
+    bind_cxx_and_arch(OUT ${PREFIX} ${CXX} ${ARCH})
+    register_append_link_flags(${OUT})
 endmacro()
 
 macro(register_definitions)