Merge pull request #4860 from Thyre/support-passing-amdgcn

Add AMDGCN option similar to `cuda-compute-capabilities`

Author: Crivella

easybuild/framework/easyconfig/default.py

@@ -84,6 +84,7 @@
     'toolchainopts': [None, 'Extra options for compilers', TOOLCHAIN],
 
     # BUILD easyconfig parameters
+    'amdgcn_capabilities': [[], "List of AMDGCN capabilities to build with (if supported)", BUILD],
     'banned_linked_shared_libs': [[], "List of shared libraries (names, file names, or paths) which are not allowed "
                                       "to be linked in any installed binary/library", BUILD],
     'bitbucket_account': ['%(namelower)s', "Bitbucket account name to be used to resolve template values in source"

easybuild/framework/easyconfig/easyconfig.py

@@ -1992,6 +1992,28 @@ def get_cuda_cc_template_value(self, key, required=True):
             error_msg = "%s is not a template value based on --cuda-compute-capabilities/cuda_compute_capabilities"
             raise EasyBuildError(error_msg, key)
 
+    def get_amdgcn_cc_template_value(self, key, required=True):
+        """
+        Get template value based on --amdgcn-capabilities EasyBuild configuration option
+        and amdgcn_capabilities easyconfig parameter.
+        Returns user-friendly error message in case neither are defined,
+        or if an unknown key is used.
+        """
+        if key.startswith('amdgcn_') and any(x == key for x in TEMPLATE_NAMES_DYNAMIC):
+            try:
+                return self.template_values[key]
+            except KeyError:
+                if not required:
+                    self.log.debug(f'Key {key} not found in template values, returning empty value')
+                    return ''
+                error_msg = "Template value '%s' is not defined!\n"
+                error_msg += "Make sure that either the --amdgcn-capabilities EasyBuild configuration "
+                error_msg += "option is set, or that the amdgcn_capabilities easyconfig parameter is defined."
+                raise EasyBuildError(error_msg, key)
+        else:
+            error_msg = "%s is not a template value based on --amdgcn-capabilities/amdgcn_capabilities"
+            raise EasyBuildError(error_msg, key)
+
 
 def det_installversion(version, toolchain_name, toolchain_version, prefix, suffix):
     """Deprecated 'det_installversion' function, to determine exact install version, based on supplied parameters."""

easybuild/framework/easyconfig/templates.py

@@ -88,6 +88,11 @@
 # template values which are only generated dynamically
 TEMPLATE_NAMES_DYNAMIC = {
     'arch': 'System architecture (e.g. x86_64, aarch64, ppc64le, ...)',
+    'amdgcn_capabilities': "Comma-separated list of AMDGCN capabilities, as specified via "
+                           "--amdgcn-capabilities configuration option or "
+                           "via amdgcn_capabilities easyconfig parameter",
+    'amdgcn_cc_space_sep': "Space-separated list of AMDGCN capabilities",
+    'amdgcn_cc_semicolon_sep': "Semicolon-separated list of AMDGCN capabilities",
     'cuda_compute_capabilities': "Comma-separated list of CUDA compute capabilities, as specified via "
                                  "--cuda-compute-capabilities configuration option or "
                                  "via cuda_compute_capabilities easyconfig parameter",
@@ -478,6 +483,14 @@ def template_constant_dict(config, ignore=None, toolchain=None):
         template_values['cuda_sm_comma_sep'] = ','.join(sm_values)
         template_values['cuda_sm_space_sep'] = ' '.join(sm_values)
 
+    # step 7. AMDGCN capabilities
+    #         Use the commandline / easybuild config option if given, else use the value from the EC (as a default)
+    amdgcn_cc = build_option('amdgcn_capabilities') or config.get('amdgcn_capabilities')
+    if amdgcn_cc:
+        template_values['amdgcn_capabilities'] = ','.join(amdgcn_cc)
+        template_values['amdgcn_cc_space_sep'] = ' '.join(amdgcn_cc)
+        template_values['amdgcn_cc_semicolon_sep'] = ';'.join(amdgcn_cc)
+
     unknown_names = []
     for key in template_values:
         if not (key in common_template_names or key in TEMPLATE_NAMES_DYNAMIC):

easybuild/tools/config.py

@@ -228,6 +228,7 @@ def mk_full_default_path(name, prefix=DEFAULT_PREFIX):
 BUILD_OPTIONS_CMDLINE = {
     None: [
         'aggregate_regtest',
+        'amdgcn_capabilities',
         'backup_modules',
         'banned_linked_shared_libs',
         'checksum_priority',

easybuild/tools/options.py

@@ -370,6 +370,9 @@ def override_options(self):
                                             None, 'store_true', False),
             'allow-use-as-root-and-accept-consequences': ("Allow using of EasyBuild as root (NOT RECOMMENDED!)",
                                                           None, 'store_true', False),
+            'amdgcn-capabilities': ("List of AMDGCN capabilities to use when building GPU software; "
+                                    "values should be specified as gfx[xyz], as defined by the LLVM targets, "
+                                    "for example: gfx1101,gfx90a,gfx1030", 'strlist', 'extend', None),
             'backup-modules': ("Back up an existing module file, if any. "
                                "Auto-enabled when using --module-only or --skip",
                                None, 'store_true', None),  # default None to allow auto-enabling if not disabled
@@ -993,6 +996,26 @@ def validate(self):
                 error_msg = "Incorrect values in --cuda-compute-capabilities (expected pattern: '%s'): %s"
                 error_msgs.append(error_msg % (cuda_cc_regex.pattern, ', '.join(faulty_cuda_ccs)))
 
+        # Support accelerators using the gfx[...] naming scheme.
+        # This applies to all AMD GPUs since Southern Islands (2013)
+        # For more information: https://llvm.org/docs/AMDGPUUsage.html#processors
+        # Allow users to pass --amdgcn-capabilities=, which will be mapped to not passing any
+        if self.options.amdgcn_capabilities == ['']:
+            self.options.amdgcn_capabilities = None
+        if self.options.amdgcn_capabilities:
+            # General accelerator naming convention
+            amdgcn_cc_regex = re.compile(r'gfx[0-9]+[a-z]?$')
+            # Generic convention.
+            # Regex is not perfect, as it doesn't catch gfx[...]--generic
+            amdgcn_generic_regex = re.compile(r'gfx[0-9]+(\-[0-9])?-generic$')
+            faulty_amdgcn_ccs = [x for x in self.options.amdgcn_capabilities
+                                 if not amdgcn_cc_regex.match(x) and not amdgcn_generic_regex.match(x)]
+            if faulty_amdgcn_ccs:
+                error_msg = "Incorrect values in --amdgcn-capabilities (expected pattern: '%s' or '%s'): %s"
+                error_msgs.append(error_msg % (amdgcn_cc_regex.pattern,
+                                               amdgcn_generic_regex.pattern,
+                                               ', '.join(faulty_amdgcn_ccs)))
+
         if error_msgs:
             raise EasyBuildError(
                 "Found problems validating the options: %s", '\n'.join(error_msgs),
@@ -2107,7 +2130,7 @@ def set_tmpdir(tmpdir=None, raise_error=False):
 
     # avoid having special characters like '[' and ']' in the tmpdir pathname,
     # it is known to cause problems (e.g., with Python install tools, CUDA's nvcc, etc.);
-    # only common characteris like alphanumeric, '_', '-', '.' and '/' are retained; others are converted to 'X'
+    # only common characters like alphanumeric, '_', '-', '.' and '/' are retained; others are converted to 'X'
     special_chars_regex = r'[^\w/.-]'
     if re.search(special_chars_regex, current_tmpdir):
         current_tmpdir = re.sub(special_chars_regex, 'X', current_tmpdir)

test/framework/easyconfig.py

@@ -4807,6 +4807,35 @@ def test_cuda_compute_capabilities(self):
         self.assertEqual(ec['preinstallopts'], 'period="4.2 6.3" noperiod="42 63"')
         self.assertEqual(ec['installopts'], '4.2,6.3')
 
+    def test_amdgcn_capabilities(self):
+        self.contents = textwrap.dedent("""
+            easyblock = 'ConfigureMake'
+            name = 'test'
+            version = '0.2'
+            homepage = 'https://example.com'
+            description = 'test'
+            toolchain = SYSTEM
+            amdgcn_capabilities = ['gfx90a', 'gfx1101', 'gfx11-generic', 'gfx10-3-generic']
+            buildopts = ('comma="%(amdgcn_capabilities)s" space="%(amdgcn_cc_space_sep)s" '
+                         'semi="%(amdgcn_cc_semicolon_sep)s"')
+            installopts = '%(amdgcn_capabilities)s'
+        """)
+        self.prep()
+
+        ec = EasyConfig(self.eb_file)
+        self.assertEqual(ec['buildopts'], 'comma="gfx90a,gfx1101,gfx11-generic,gfx10-3-generic" '
+                                          'space="gfx90a gfx1101 gfx11-generic gfx10-3-generic" '
+                                          'semi="gfx90a;gfx1101;gfx11-generic;gfx10-3-generic"')
+        self.assertEqual(ec['installopts'], 'gfx90a,gfx1101,gfx11-generic,gfx10-3-generic')
+
+        # build options overwrite it
+        init_config(build_options={'amdgcn_capabilities': ['gfx90a', 'gfx1101']})
+        ec = EasyConfig(self.eb_file)
+        self.assertEqual(ec['buildopts'], 'comma="gfx90a,gfx1101" '
+                                          'space="gfx90a gfx1101" '
+                                          'semi="gfx90a;gfx1101"')
+        self.assertEqual(ec['installopts'], 'gfx90a,gfx1101')
+
     def test_det_copy_ec_specs(self):
         """Test det_copy_ec_specs function."""
 
@@ -5111,6 +5140,56 @@ def test_get_cuda_cc_template_value(self):
         for key, expected in cuda_template_values.items():
             self.assertEqual(ec.get_cuda_cc_template_value(key), expected)
 
+    def test_get_amdgcn_cc_template_value(self):
+        """
+        Test getting template value based on --amdgcn-capabilities / amdgcn_capabilities.
+        """
+        self.contents = '\n'.join([
+            'easyblock = "ConfigureMake"',
+            'name = "pi"',
+            'version = "3.14"',
+            'homepage = "http://example.com"',
+            'description = "test easyconfig"',
+            'toolchain = SYSTEM',
+        ])
+        self.prep()
+        ec = EasyConfig(self.eb_file)
+
+        error_pattern = ("foobar is not a template value based on "
+                         "--amdgcn-capabilities/amdgcn_capabilities")
+        self.assertErrorRegex(EasyBuildError, error_pattern, ec.get_amdgcn_cc_template_value, 'foobar')
+
+        error_pattern = r"Template value '%s' is not defined!\n"
+        error_pattern += r"Make sure that either the --amdgcn-capabilities EasyBuild configuration "
+        error_pattern += "option is set, or that the amdgcn_capabilities easyconfig parameter is defined."
+        amdgcn_template_values = {
+            'amdgcn_capabilities': 'gfx90a,gfx1100,gfx10-3-generic',
+            'amdgcn_cc_space_sep': 'gfx90a gfx1100 gfx10-3-generic',
+            'amdgcn_cc_semicolon_sep': 'gfx90a;gfx1100;gfx10-3-generic',
+        }
+        for key in amdgcn_template_values:
+            self.assertErrorRegex(EasyBuildError, error_pattern % key, ec.get_amdgcn_cc_template_value, key)
+
+        update_build_option('amdgcn_capabilities', ['gfx90a', 'gfx1100', 'gfx10-3-generic'])
+        ec = EasyConfig(self.eb_file)
+
+        for key, expected in amdgcn_template_values.items():
+            self.assertEqual(ec.get_amdgcn_cc_template_value(key), expected)
+
+        update_build_option('amdgcn_capabilities', None)
+        ec = EasyConfig(self.eb_file)
+
+        for key in amdgcn_template_values:
+            self.assertErrorRegex(EasyBuildError, error_pattern % key, ec.get_amdgcn_cc_template_value, key)
+            self.assertEqual(ec.get_amdgcn_cc_template_value(key, required=False), '')
+
+        self.contents += "\namdgcn_capabilities = ['gfx90a', 'gfx1100', 'gfx10-3-generic']"
+        self.prep()
+        ec = EasyConfig(self.eb_file)
+
+        for key, expected in amdgcn_template_values.items():
+            self.assertEqual(ec.get_amdgcn_cc_template_value(key), expected)
+
     def test_count_files(self):
         """Tests for EasyConfig.count_files method."""
         test_ecs_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'easyconfigs', 'test_ecs')