From cd367c7da3aca6f7fd8a1dd536a5ae71f43db984 Mon Sep 17 00:00:00 2001
From: Tom Lin <tom91136@gmail.com>
Date: Tue, 29 Jun 2021 17:53:32 +0100
Subject: [PATCH] Mirror Fujitsu flags for CMake

---
 OMP.cmake             | 28 ++++++++++++++++++++++++----
 register_models.cmake | 17 +++++++++++++++--
 2 files changed, 39 insertions(+), 6 deletions(-)

diff --git a/OMP.cmake b/OMP.cmake
index c8dde9f..1955ebc 100644
--- a/OMP.cmake
+++ b/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)
+set(OMP_FLAGS_CPU_NEC -O4 -finline) # CMake doesn't detect this so it's meant to be chosen by register_flag_optional(ARCH)
 
 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)
diff --git a/register_models.cmake b/register_models.cmake
index 82e7243..9f4cb57 100644
--- a/register_models.cmake
+++ b/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)