Set archspec compiler flags for NVHPC

NVIDIA · Aug 1, 2024 · 7ca8278 · 7ca8278
1 parent 67fdb28
commit 7ca8278
Show file tree

Hide file tree

Showing 3 changed files with 17 additions and 2 deletions.
diff --git a/hpccm/building_blocks/nvhpc.py b/hpccm/building_blocks/nvhpc.py
@@ -181,6 +181,10 @@ def __init__(self, **kwargs):
 
         self.toolchain = toolchain(CC='nvc', CXX='nvc++', F77='nvfortran',
                                    F90='nvfortran', FC='nvfortran')
+        self.toolchain.CFLAGS = hpccm.config.get_cpu_optimization_flags('nvhpc', version=self.__version)
+        self.toolchain.CXXFLAGS = hpccm.config.get_cpu_optimization_flags('nvhpc', version=self.__version)
+        self.toolchain.FFLAGS = hpccm.config.get_cpu_optimization_flags('nvhpc', version=self.__version)
+        self.toolchain.FCFLAGS = hpccm.config.get_cpu_optimization_flags('nvhpc', version=self.__version)
 
         if Version(self.__version) >= Version('23.7'):
             self.__cuda_version_default = '12.2'

diff --git a/test/test_nvhpc.py b/test/test_nvhpc.py
@@ -22,7 +22,7 @@
 import logging # pylint: disable=unused-import
 import unittest
 
-from helpers import aarch64, centos, docker, ppc64le, ubuntu, x86_64
+from helpers import aarch64, centos, docker, ppc64le, ubuntu, x86_64, zen2
 
 from hpccm.building_blocks.nvhpc import nvhpc
 from hpccm.toolchain import toolchain
@@ -282,3 +282,13 @@ def test_toolchain(self):
         self.assertEqual(tc.FC, 'nvfortran')
         self.assertEqual(tc.F77, 'nvfortran')
         self.assertEqual(tc.F90, 'nvfortran')
+
+    @zen2
+    def test_toolchain_zen2(self):
+        """CPU arch optimization flags"""
+        n = nvhpc()
+        tc = n.toolchain
+        self.assertEqual(tc.CFLAGS, '-tp zen2')
+        self.assertEqual(tc.CXXFLAGS, '-tp zen2')
+        self.assertEqual(tc.FFLAGS, '-tp zen2')
+        self.assertEqual(tc.FCFLAGS, '-tp zen2')
diff --git a/test/test_openmpi.py b/test/test_openmpi.py
@@ -22,7 +22,7 @@
 import logging # pylint: disable=unused-import
 import unittest
 
-from helpers import centos, docker, ubuntu
+from helpers import centos, cpu_target_none, docker, ubuntu
 
 from hpccm.building_blocks.nvhpc import nvhpc
 from hpccm.building_blocks.openmpi import openmpi
@@ -117,6 +117,7 @@ def test_ldconfig(self):
 
     @ubuntu
     @docker
+    @cpu_target_none
     def test_nvhpc(self):
         """HPC SDK toolchain"""
         compiler = nvhpc(eula=True)