Merge pull request #5067 from bedroge/cuda_cc_regex_allow_f

allow `f` suffix in CUDA compute capabilities

Author: boegel

easybuild/tools/options.py

@@ -990,7 +990,7 @@ def validate(self):
         # values passed to --cuda-compute-capabilities must be of form X.Y (with both X and Y integers),
         # see https://developer.nvidia.com/cuda-gpus
         if self.options.cuda_compute_capabilities:
-            cuda_cc_regex = re.compile(r'^[0-9]+\.[0-9]+a?$')
+            cuda_cc_regex = re.compile(r'^[0-9]+\.[0-9]+[af]?$')
             faulty_cuda_ccs = [x for x in self.options.cuda_compute_capabilities if not cuda_cc_regex.match(x)]
             if faulty_cuda_ccs:
                 error_msg = "Incorrect values in --cuda-compute-capabilities (expected pattern: '%s'): %s"

easybuild/tools/systemtools.py

@@ -1078,7 +1078,7 @@ def get_cuda_architectures(path, section_type):
     # compile_size = 64bit
 
     # Pattern to extract elf code architectures and ptx code architectures respectively
-    code_regex = re.compile(f'Fatbin {section_type} code:\n=+\narch = sm_([0-9]+)([0-9]a?)')
+    code_regex = re.compile(f'Fatbin {section_type} code:\n=+\narch = sm_([0-9]+)([0-9][af]?)')
 
     # resolve symlinks
     if os.path.islink(path) and os.path.exists(path):

test/framework/options.py

@@ -6816,10 +6816,10 @@ def test_installdir(self):
 
     def test_cuda_compute_capabilities(self):
         """Test --cuda-compute-capabilities configuration option."""
-        args = ['--cuda-compute-capabilities=3.5,6.2,7.0', '--show-config']
+        args = ['--cuda-compute-capabilities=3.5,6.2,7.0,9.0a,10.0f', '--show-config']
         txt, _ = self._run_mock_eb(args, do_build=True, raise_error=True, testing=False, strip=True)
 
-        regex = re.compile(r"^cuda-compute-capabilities\s*\(C\)\s*=\s*3\.5, 6\.2, 7\.0$", re.M)
+        regex = re.compile(r"^cuda-compute-capabilities\s*\(C\)\s*=\s*3\.5, 6\.2, 7\.0, 9\.0a, 10\.0f$", re.M)
         self.assertRegex(txt, regex)
 
     def test_create_index(self):

test/framework/systemtools.py

@@ -428,16 +428,22 @@
 CUOBJDUMP_DEVICE_CODE_ONLY = """
 Fatbin elf code:
 ================
-arch = sm_90
-code version = [1,7]
+arch = sm_100f
+code version = [1,8]
 host = linux
 compile_size = 64bit
-compressed
 
 Fatbin elf code:
 ================
-arch = sm_90a
-code version = [1,7]
+arch = sm_100
+code version = [1,8]
+host = linux
+compile_size = 64bit
+
+Fatbin elf code:
+================
+arch = sm_100a
+code version = [1,8]
 host = linux
 compile_size = 64bit"""
 
@@ -1425,7 +1431,7 @@ def test_get_cuda_architectures(self):
         self.assertIsNone(res_ptx)
         fail_msg = "Pattern '%s' should be found in: %s" % (warning_regex_ptx.pattern, logtxt)
         self.assertTrue(warning_regex_ptx.search(logtxt), fail_msg)
-        self.assertEqual(res_elf, ['9.0', '9.0a'])
+        self.assertEqual(res_elf, ['10.0', '10.0a', '10.0f'])
 
         # Test case 6: call on CUDA shared lib which lacks an arch = sm_XX entry (should never happen)
         warning_regex_elf = re.compile(r"WARNING Found Fatbin elf code section\(s\) in cuobjdump output for "