Merge pull request #2388 from ComputeCanada/gcc_tree_vectorize

boegel · web-flow · commit 95b57e969669 · 2018-03-31T09:34:51.000+02:00
include -ftree-vectorize and -fno-math-errno in default compiler optimisation flags for GCC
diff --git a/easybuild/toolchains/compiler/clang.py b/easybuild/toolchains/compiler/clang.py
@@ -85,6 +85,7 @@ class Clang(Compiler):
         'defaultprec': [],
         'loose': ['ffast-math', 'fno-unsafe-math-optimizations'],
         'veryloose': ['ffast-math'],
+        'vectorize': {False: 'fno-vectorize', True: 'fvectorize'},
     }
 
     # used when 'optarch' toolchain option is enabled (and --optarch is not specified)
diff --git a/easybuild/toolchains/compiler/gcc.py b/easybuild/toolchains/compiler/gcc.py
@@ -35,7 +35,7 @@
 import easybuild.tools.systemtools as systemtools
 from easybuild.tools.build_log import EasyBuildError
 from easybuild.tools.modules import get_software_root, get_software_version
-from easybuild.tools.toolchain.compiler import Compiler
+from easybuild.tools.toolchain.compiler import Compiler, DEFAULT_OPT_LEVEL
 
 
 TC_CONSTANT_GCC = "GCC"
@@ -61,10 +61,12 @@ class Gcc(Compiler):
         'lto': 'flto',
         'ieee': ['mieee-fp', 'fno-trapping-math'],
         'strict': ['mieee-fp', 'mno-recip'],
-        'precise':['mno-recip'],
-        'defaultprec':[],
-        'loose': ['mrecip', 'mno-ieee-fp'],
-        'veryloose': ['mrecip=all', 'mno-ieee-fp'],
+        'precise': ['mno-recip'],
+        'defaultprec': ['fno-math-errno'],
+        'loose': ['fno-math-errno', 'mrecip', 'mno-ieee-fp'],
+        'veryloose': ['fno-math-errno', 'mrecip=all', 'mno-ieee-fp'],
+        'vectorize': {False: 'fno-tree-vectorize', True: 'ftree-vectorize'},
+        DEFAULT_OPT_LEVEL: ['O2', 'ftree-vectorize'],
     }
 
     # used when 'optarch' toolchain option is enabled (and --optarch is not specified)
diff --git a/easybuild/toolchains/compiler/ibmxl.py b/easybuild/toolchains/compiler/ibmxl.py
@@ -11,7 +11,7 @@
 from distutils.version import LooseVersion
 
 import easybuild.tools.systemtools as systemtools
-from easybuild.tools.toolchain.compiler import Compiler
+from easybuild.tools.toolchain.compiler import Compiler, DEFAULT_OPT_LEVEL
 
 
 TC_CONSTANT_IBMCOMP = "IBMXL"
@@ -35,6 +35,8 @@ class IBMXL(Compiler):
         'defaultprec': ['', '', ''],
         'loose': [''],
         'veryloose': [''],
+        'vectorize': {False: 'qsimd=noauto', True: 'qsimd=auto'},
+        DEFAULT_OPT_LEVEL: ['O2', 'qsimd=auto'],
         'ibm-static': 'qstaticlink=xllibs',
         'pic': 'qpic',
         'shared': 'qmkshrobj',
diff --git a/easybuild/toolchains/compiler/inteliccifort.py b/easybuild/toolchains/compiler/inteliccifort.py
@@ -64,6 +64,7 @@ class IntelIccIfort(Compiler):
         'defaultprec': ['ftz', 'fp-speculation=safe', 'fp-model source'],
         'loose': ['fp-model fast=1'],
         'veryloose': ['fp-model fast=2'],
+        'vectorize': {False: 'no-vec', True: 'vec'},
         'intel-static': 'static-intel',
         'no-icc': 'no-icc',
         'error-unknown-option': 'we10006',  # error at warning #10006: ignoring unknown option
diff --git a/easybuild/toolchains/compiler/pgi.py b/easybuild/toolchains/compiler/pgi.py
@@ -65,6 +65,7 @@ class Pgi(Compiler):
         'defaultprec': ['Mflushz'],
         'loose': ['Mfprelaxed'],
         'veryloose': ['Mfprelaxed=div,order,intrinsic,recip,sqrt,rsqrt', 'Mfpapprox'],
+        'vectorize': {False: 'Mnovect', True: 'Mvect'},
     }
 
     # used when 'optarch' toolchain option is enabled (and --optarch is not specified)
diff --git a/easybuild/tools/toolchain/compiler.py b/easybuild/tools/toolchain/compiler.py
@@ -88,6 +88,7 @@ class Compiler(Toolchain):
         'static': (False, "Build static library"),
         '32bit': (False, "Compile 32bit target"),  # LA, FFTW
         'openmp': (False, "Enable OpenMP"),
+        'vectorize': (None, "Enable compiler auto-vectorization, default except for noopt and lowopt"),
         'packed-linker-options': (False, "Pack the linker options as comma separated list"),  # ScaLAPACK mainly
         'rpath': (True, "Use RPATH wrappers when --rpath is enabled in EasyBuild configuration"),
     }
@@ -245,8 +246,19 @@ def _set_compiler_flags(self):
                                  (default_opt_level, self.COMPILER_OPT_FLAGS))
 
         # 1st one is the one to use. add default at the end so len is at least 1
-        optflags = [self.options.option(x) for x in self.COMPILER_OPT_FLAGS if self.options.get(x, False)] + \
-                   [self.options.option(default_opt_level)]
+        optflags = ([self.options.option(x) for x in self.COMPILER_OPT_FLAGS if self.options.get(x, False)] + \
+                    [self.options.option(default_opt_level)])[:1]
+
+        # only apply if the vectorize toolchainopt is explicitly set
+        # otherwise the individual compiler toolchain file should make sure that
+        # vectorization is disabled for noopt and lowopt, and enabled otherwise.
+        if self.options.get('vectorize') is not None:
+            vectoptions = self.options.option('vectorize')
+            vectflags = vectoptions[self.options['vectorize']]
+            # avoid double use of such flags, or e.g. -fno-tree-vectorize followed by -ftree-vectorize
+            if isinstance(optflags[0], list):
+                optflags[0] = [flag for flag in optflags[0] if flag not in vectoptions.values()]
+            optflags.append(vectflags)
 
         optarchflags = []
         if build_option('optarch') == OPTARCH_GENERIC:
@@ -259,7 +271,7 @@ def _set_compiler_flags(self):
         precflags = [self.options.option(x) for x in self.COMPILER_PREC_FLAGS if self.options.get(x, False)] + \
                     [self.options.option('defaultprec')]
 
-        self.variables.nextend('OPTFLAGS', optflags[:1] + optarchflags)
+        self.variables.nextend('OPTFLAGS', optflags + optarchflags)
         self.variables.nextend('PRECFLAGS', precflags[:1])
 
         # precflags last
diff --git a/test/framework/options.py b/test/framework/options.py
@@ -2948,7 +2948,7 @@ def test_dump_env_config(self):
             "module load hwloc/1.6.2-GCC-4.7.2",  # loading of dependency module
             # defining build env
             "export FC='gfortran'",
-            "export CFLAGS='-O2 -march=native'",
+            "export CFLAGS='-O2 -ftree-vectorize -march=native -fno-math-errno'",
         ]
         for pattern in patterns:
             regex = re.compile("^%s$" % pattern, re.M)
diff --git a/test/framework/toolchain.py b/test/framework/toolchain.py
@@ -277,7 +277,7 @@ def test_misc_flags_shared(self):
                 # we need to make sure we check for flags, not letter (e.g. 'v' vs '-v')
                 flag = '-%s' % tc.COMPILER_SHARED_OPTION_MAP[opt]
                 for var in flag_vars:
-                    flags = tc.get_variable(var)
+                    flags = tc.get_variable(var).split()
                     if enable:
                         self.assertTrue(flag in flags, "%s: True means %s in %s" % (opt, flag, flags))
                     else:
@@ -290,21 +290,25 @@ def test_misc_flags_unique(self):
         flag_vars = ['CFLAGS', 'CXXFLAGS', 'FCFLAGS', 'FFLAGS', 'F90FLAGS']
 
         # setting option should result in corresponding flag to be set (unique options)
-        for opt in ['unroll', 'optarch', 'openmp']:
+        for opt in ['unroll', 'optarch', 'openmp', 'vectorize']:
             for enable in [True, False]:
                 tc = self.get_toolchain("goalf", version="1.1.0-no-OFED")
                 tc.set_options({opt: enable})
                 tc.prepare()
                 if opt == 'optarch':
-                    flag = '-%s' % tc.COMPILER_OPTIMAL_ARCHITECTURE_OPTION[(tc.arch, tc.cpu_family)]
+                    option = tc.COMPILER_OPTIMAL_ARCHITECTURE_OPTION[(tc.arch, tc.cpu_family)]
                 else:
-                    flag = '-%s' % tc.options.options_map[opt]
+                    option = tc.options.options_map[opt]
+                if not isinstance(option, dict):
+                    option = {True: option}
                 for var in flag_vars:
                     flags = tc.get_variable(var)
-                    if enable:
-                        self.assertTrue(flag in flags, "%s: True means %s in %s" % (opt, flag, flags))
-                    else:
-                        self.assertTrue(flag not in flags, "%s: False means no %s in %s" % (opt, flag, flags))
+                    for key, value in option.items():
+                        flag = "-%s" % value
+                        if enable == key:
+                            self.assertTrue(flag in flags, "%s: %s means %s in %s" % (opt, enable, flag, flags))
+                        else:
+                            self.assertTrue(flag not in flags, "%s: %s means no %s in %s" % (opt, enable, flag, flags))
                 self.modtool.purge()
 
     def test_override_optarch(self):
@@ -387,7 +391,7 @@ def test_compiler_dependent_optarch(self):
         """Test whether specifying optarch on a per compiler basis works."""
         flag_vars = ['CFLAGS', 'CXXFLAGS', 'FCFLAGS', 'FFLAGS', 'F90FLAGS']
         intel_options = [('intelflag', 'intelflag'), ('GENERIC', 'xSSE2'), ('', '')]
-        gcc_options = [('gccflag', 'gccflag'), ('-ftree-vectorize', '-ftree-vectorize'), ('', '')]
+        gcc_options = [('gccflag', 'gccflag'), ('march=nocona', 'march=nocona'), ('', '')]
         gcccore_options = [('gcccoreflag', 'gcccoreflag'), ('GENERIC', 'march=x86-64 -mtune=generic'), ('', '')]
         toolchains = [('iccifort', '2011.13.367'), ('GCC', '4.7.2'), ('GCCcore', '6.2.0'), ('PGI', '16.7-GCC-5.4.0-2.26')]
         enabled = [True, False]
@@ -476,20 +480,20 @@ def test_precision_flags(self):
 
         flag_vars = ['CFLAGS', 'CXXFLAGS', 'FCFLAGS', 'FFLAGS', 'F90FLAGS']
 
-        # check default precision: no specific flag for GCC
+        # check default precision: -fno-math-errno flag for GCC
         tc = self.get_toolchain("goalf", version="1.1.0-no-OFED")
         tc.set_options({})
         tc.prepare()
         for var in flag_vars:
-            self.assertEqual(os.getenv(var), "-O2 -march=native")
+            self.assertEqual(os.getenv(var), "-O2 -ftree-vectorize -march=native -fno-math-errno")
 
         # check other precision flags
         prec_flags = {
-            'ieee': "-mieee-fp -fno-trapping-math",
+            'ieee': "-fno-math-errno -mieee-fp -fno-trapping-math",
             'strict': "-mieee-fp -mno-recip",
             'precise': "-mno-recip",
-            'loose': "-mrecip -mno-ieee-fp",
-            'veryloose': "-mrecip=all -mno-ieee-fp",
+            'loose': "-fno-math-errno -mrecip -mno-ieee-fp",
+            'veryloose': "-fno-math-errno -mrecip=all -mno-ieee-fp",
         }
         for prec in prec_flags:
             for enable in [True, False]:
@@ -498,9 +502,9 @@ def test_precision_flags(self):
                 tc.prepare()
                 for var in flag_vars:
                     if enable:
-                        self.assertEqual(os.getenv(var), "-O2 -march=native %s" % prec_flags[prec])
+                        self.assertEqual(os.getenv(var), "-O2 -ftree-vectorize -march=native %s" % prec_flags[prec])
                     else:
-                        self.assertEqual(os.getenv(var), "-O2 -march=native")
+                        self.assertEqual(os.getenv(var), "-O2 -ftree-vectorize -march=native -fno-math-errno")
                 self.modtool.purge()
 
     def test_cgoolf_toolchain(self):
@@ -579,8 +583,10 @@ def test_goolfc(self):
         tc.set_options(opts)
         tc.prepare()
 
+        archflags = tc.COMPILER_OPTIMAL_ARCHITECTURE_OPTION[(tc.arch, tc.cpu_family)]
+        optflags = "-O2 -ftree-vectorize -%s -fno-math-errno -fopenmp" % archflags
         nvcc_flags = r' '.join([
-            r'-Xcompiler="-O2 -%s -fopenmp"' % tc.COMPILER_OPTIMAL_ARCHITECTURE_OPTION[(tc.arch, tc.cpu_family)],
+            r'-Xcompiler="%s"' % optflags,
             # the use of -lcudart in -Xlinker is a bit silly but hard to avoid
             r'-Xlinker=".* -lm -lrt -lcudart -lpthread"',
             r' '.join(["-gencode %s" % x for x in opts['cuda_gencode']]),
@@ -903,9 +909,9 @@ def test_independence(self):
 
         tc_cflags = {
             'CrayCCE': "-O2 -homp -craype-verbose",
-            'CrayGNU': "-O2 -fopenmp -craype-verbose",
+            'CrayGNU': "-O2 -fno-math-errno -fopenmp -craype-verbose",
             'CrayIntel': "-O2 -ftz -fp-speculation=safe -fp-model source -fopenmp -craype-verbose",
-            'GCC': "-O2 -test -fopenmp",
+            'GCC': "-O2 -ftree-vectorize -test -fno-math-errno -fopenmp",
             'iccifort': "-O2 -test -ftz -fp-speculation=safe -fp-model source -fopenmp",
         }
 

Original file line number	Diff line number	Diff line change
`@@ -85,6 +85,7 @@ class Clang(Compiler):`
`85`	`85`	`'defaultprec': [],`
`86`	`86`	`'loose': ['ffast-math', 'fno-unsafe-math-optimizations'],`
`87`	`87`	`'veryloose': ['ffast-math'],`
	`88`	`+ 'vectorize': {False: 'fno-vectorize', True: 'fvectorize'},`
`88`	`89`	`}`
`89`	`90`
`90`	`91`	`# used when 'optarch' toolchain option is enabled (and --optarch is not specified)`
Original file line number	Diff line number	Diff line change
`@@ -65,6 +65,7 @@ class Pgi(Compiler):`
`65`	`65`	`'defaultprec': ['Mflushz'],`
`66`	`66`	`'loose': ['Mfprelaxed'],`
`67`	`67`	`'veryloose': ['Mfprelaxed=div,order,intrinsic,recip,sqrt,rsqrt', 'Mfpapprox'],`
	`68`	`+ 'vectorize': {False: 'Mnovect', True: 'Mvect'},`
`68`	`69`	`}`
`69`	`70`
`70`	`71`	`# used when 'optarch' toolchain option is enabled (and --optarch is not specified)`
Original file line number	Diff line number	Diff line change
`@@ -2948,7 +2948,7 @@ def test_dump_env_config(self):`
`2948`	`2948`	`"module load hwloc/1.6.2-GCC-4.7.2", # loading of dependency module`
`2949`	`2949`	`# defining build env`
`2950`	`2950`	`"export FC='gfortran'",`
`2951`		`- "export CFLAGS='-O2 -march=native'",`
	`2951`	`+ "export CFLAGS='-O2 -ftree-vectorize -march=native -fno-math-errno'",`
`2952`	`2952`	`]`
`2953`	`2953`	`for pattern in patterns:`
`2954`	`2954`	`regex = re.compile("^%s$" % pattern, re.M)`