Allow optarch values to be partial maps including vector extensions

easybuild/toolchains/compiler/inteliccifort.py

@@ -74,13 +74,15 @@ class IntelIccIfort(Compiler):
 
     # used when 'optarch' toolchain option is enabled (and --optarch is not specified)
     COMPILER_OPTIMAL_ARCHITECTURE_OPTION = {
-        (systemtools.X86_64, systemtools.AMD): 'xHost',
-        (systemtools.X86_64, systemtools.INTEL): 'xHost',
+        systemtools.X86_64: 'xHost',
+        # Intel compilers don't auto-detect AMD features and default to SSE2
+        (systemtools.X86_64, systemtools.AMD, systemtools.SSSE3): 'mssse3',
+        (systemtools.X86_64, systemtools.AMD, systemtools.SSE4_2): 'msse4.2',
+        (systemtools.X86_64, systemtools.AMD, systemtools.AVX2): 'march=core-avx2',
     }
     # used with --optarch=GENERIC
     COMPILER_GENERIC_OPTION = {
-        (systemtools.X86_64, systemtools.AMD): 'xSSE2',
-        (systemtools.X86_64, systemtools.INTEL): 'xSSE2',
+        systemtools.X86_64: 'xSSE2',
     }
 
     COMPILER_CC = 'icc'

easybuild/tools/options.py

@@ -47,7 +47,7 @@
 import pwd
 
 import easybuild.tools.environment as env
-from easybuild.base import fancylogger  # build_log should always stay there, to ensure EasyBuildLog
+from easybuild.base import fancylogger
 from easybuild.base.fancylogger import setLogLevel
 from easybuild.base.generaloption import GeneralOption
 from easybuild.framework.easyblock import MODULE_ONLY_STEPS, SOURCE_STEP, FETCH_STEP, EasyBlock
@@ -101,7 +101,7 @@
 from easybuild.tools.robot import det_robot_path
 from easybuild.tools.run import run_cmd
 from easybuild.tools.package.utilities import avail_package_naming_schemes
-from easybuild.tools.toolchain.compiler import DEFAULT_OPT_LEVEL, OPTARCH_MAP_CHAR, OPTARCH_SEP, Compiler
+from easybuild.tools.toolchain.compiler import DEFAULT_OPT_LEVEL, Compiler, parse_optarch_string
 from easybuild.tools.toolchain.toolchain import SYSTEM_TOOLCHAIN_NAME
 from easybuild.tools.repository.repository import avail_repositories
 from easybuild.tools.systemtools import UNKNOWN, check_python_version, get_cpu_architecture, get_cpu_family
@@ -987,29 +987,10 @@ def postprocess(self):
             cleanup_and_exit(self.tmpdir)
 
     def _postprocess_optarch(self):
-        """Postprocess --optarch option."""
-        optarch_parts = self.options.optarch.split(OPTARCH_SEP)
-
-        # we expect to find a ':' in every entry in optarch, in case optarch is specified on a per-compiler basis
-        n_parts = len(optarch_parts)
-        map_char_cnts = [p.count(OPTARCH_MAP_CHAR) for p in optarch_parts]
-        if (n_parts > 1 and any(c != 1 for c in map_char_cnts)) or (n_parts == 1 and map_char_cnts[0] > 1):
-            raise EasyBuildError("The optarch option has an incorrect syntax: %s", self.options.optarch)
-        else:
-            # if there are options for different compilers, we set up a dict
-            if OPTARCH_MAP_CHAR in optarch_parts[0]:
-                optarch_dict = {}
-                for compiler, compiler_opt in [p.split(OPTARCH_MAP_CHAR) for p in optarch_parts]:
-                    if compiler in optarch_dict:
-                        raise EasyBuildError("The optarch option contains duplicated entries for compiler %s: %s",
-                                             compiler, self.options.optarch)
-                    else:
-                        optarch_dict[compiler] = compiler_opt
-                self.options.optarch = optarch_dict
-                self.log.info("Transforming optarch into a dict: %s", self.options.optarch)
-            # if optarch is not in mapping format, we do nothing and just keep the string
-            else:
-                self.log.info("Keeping optarch raw: %s", self.options.optarch)
+        """
+        Postprocess --optarch option.
+        """
+        self.options.optarch = parse_optarch_string(self.options.optarch, include_compiler=True)
 
     def _postprocess_close_pr_reasons(self):
         """Postprocess --close-pr-reasons options"""

easybuild/tools/systemtools.py

@@ -99,6 +99,16 @@
 
 ARCH_KEY_PREFIX = 'arch='
 
+# Vector extension constants
+SSE = 'sse'
+SSE2 = 'sse2'
+SSSE3 = 'ssse3'
+SSE4_1 = 'sse4_1'
+SSE4_2 = 'sse4_2'
+AVX = 'avx'
+AVX2 = 'avx2'
+AVX512F = 'avx512f'
+
 # Vendor constants
 AMD = 'AMD'
 APM = 'Applied Micro'
@@ -133,6 +143,8 @@
 CPU_ARCHITECTURES = [AARCH32, AARCH64, POWER, RISCV32, RISCV64, X86_64]
 CPU_FAMILIES = [AMD, ARM, INTEL, POWER, POWER_LE, RISCV]
 CPU_VENDORS = [AMD, APM, APPLE, ARM, BROADCOM, CAVIUM, DEC, IBM, INTEL, MARVELL, MOTOROLA, NVIDIA, QUALCOMM]
+# Vector extensions of CPUs in ascending order (later => better)
+CPU_VECTOR_EXTS = [SSE, SSE2, SSSE3, SSE4_1, SSE4_2, AVX, AVX2, AVX512F]
 # ARM implementer IDs (i.e., the hexadeximal keys) taken from ARMv8-A Architecture Reference Manual
 # (ARM DDI 0487A.j, Section G6.2.102, Page G6-4493)
 VENDOR_IDS = {
@@ -610,6 +622,13 @@ def get_cpu_features():
     return cpu_feat
 
 
+def get_cpu_vector_exts():
+    """Get list of (relevant) CPU vector extensions"""
+    cpu_features = set(get_cpu_features())
+    # Values of the vector extension constants purposely match the cpu feature value
+    return [i for i in CPU_VECTOR_EXTS if i in cpu_features]
+
+
 def get_gpu_info():
     """
     Get the GPU info
@@ -1322,6 +1341,37 @@ def pick_dep_version(dep_version):
     return result
 
 
+def pick_opt_arch(options, arch, cpu_family, vector_exts):
+    """
+    Pick the best matching entry from the options dict based on CPU arch and features
+
+    :param options: Dictionary mapping tuples of system specs to optarchs.
+                    E.g. (X86_64, AMD, AVX2): 'mavx2'
+    :param arch: Current CPU arch
+    :param cpu_family: Current CPU family
+    :param vector_exts: Vector extensions supported by current CPU
+    """
+    def create_possible_keys():
+        """Yield a list of possible keys from most specific to least specific"""
+        for ext in vector_exts:
+            yield (arch, cpu_family, ext)
+        yield (arch, cpu_family)
+        yield (arch, )
+        # Also allow single string entry
+        yield arch
+        # Default fallback for any arch
+        yield None
+
+    result = None
+    for key in create_possible_keys():
+        try:
+            result = options[key]
+            break
+        except KeyError:
+            pass
+    return result
+
+
 def det_pypkg_version(pkg_name, imported_pkg, import_name=None):
     """Determine version of a Python package."""