Merge pull request #4203 from martin-frbg/issue4201

Add support for building arm64 SVE kernels with the NVIDIA HPC compiler

c_check

@@ -283,6 +283,9 @@ if [ "$architecture" = "arm64" ]; then
     no_sve=0
     {
         $compiler_name $flags $args >/dev/null 2>&1
+    } || {
+        args=" -Msve_intrinsics -c -o $tmpf.o $tmpf"
+        $compiler_name $flags $args >/dev/null 2>&1
     } || {
         no_sve=1
     }

cmake/cc.cmake

@@ -180,22 +180,30 @@ endif ()
 
 if (${CORE} STREQUAL NEOVERSEN2)
   if (NOT DYNAMIC_ARCH)
-    execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
-    if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
-      set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -Msve_intrinsics -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
     else ()
-      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a+sve")
-    endif()
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+      if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
+      else ()
+	set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a+sve")
+      endif()
+    endif ()
   endif ()
 endif ()
 
 if (${CORE} STREQUAL NEOVERSEV1)
   if (NOT DYNAMIC_ARCH)
-    execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
-    if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
-      set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.4-a+sve -mtune=neoverse-v1")
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -Msve_intrinsics -march=armv8.4-a+sve -mtune=neoverse-v1")
     else ()
-      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a+sve")
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+      if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
+        set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.4-a+sve -mtune=neoverse-v1")
+      else ()
+        set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a+sve")
+      endif()
     endif()
   endif ()
 endif ()
@@ -213,7 +221,11 @@ endif ()
 
 if (${CORE} STREQUAL ARMV8SVE)
   if (NOT DYNAMIC_ARCH)
-    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -Msve_intrinsics -march=armv8-a+sve")
+    else ()
+      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")
+    endif ()
   endif ()
 endif ()
 

cmake/system.cmake

@@ -282,23 +282,35 @@ if (DEFINED TARGET)
   endif()
 
 if (${TARGET} STREQUAL NEOVERSEV1)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+	set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -Msve_intrinsics -march=armv8.4-a+sve -mtune=neoverse-v1")
+    else ()
     execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
     if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.4-a+sve -mtune=neoverse-v1")
     else ()
 	    message(FATAL_ERROR "Compiler ${CMAKE_C_COMPILER} ${GCC_VERSION} does not support Neoverse V1.")
     endif()
+    endif()
   endif()
   if (${TARGET} STREQUAL NEOVERSEN2)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+	set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
+    else ()
     execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
     if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2")
     else ()
 	    message(FATAL_ERROR "Compiler $${CMAKE_C_COMPILER} {GCC_VERSION} does not support Neoverse N2.")
     endif()
+    endif()
   endif()
   if (${TARGET} STREQUAL ARMV8SVE)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+      set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve")
+    else ()
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve")
+    endif()
   endif()
 
 endif()

kernel/Makefile

@@ -77,6 +77,12 @@ else ifeq ($(TARGET_CORE), ZEN)
  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) $(AVX2OPT)
 else ifeq ($(TARGET_CORE), LOONGSON3R4)
  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) $(MSA_FLAGS)
+else ifneq ($(filter NEOVERSEN2 NEOVERSEV1, $(TARGET_CORE)),)
+ ifeq ($(C_COMPILER), PGI)
+  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) -Msve_intrinsics
+ else
+  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
+ endif
 else
  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
 endif