cp2k: added support for AMD toolchain (#21371)

3 files changed, 55 insertions, 28 deletions

diff --git a/var/spack/repos/builtin/packages/cp2k/package.py b/var/spack/repos/builtin/packages/cp2k/package.py
index 35a930960f..79ec2ad727 100644
--- a/var/spack/repos/builtin/packages/cp2k/package.py
+++ b/var/spack/repos/builtin/packages/cp2k/package.py
@@ -158,6 +158,7 @@ class Cp2k(MakefilePackage, CudaPackage):
     conflicts('smm=libxsmm',  when='target=aarch64:', msg='libxsmm is not available on arm')
 
     conflicts('^fftw~openmp', when='+openmp')
+    conflicts('^amdfftw~openmp', when='+openmp')
     conflicts('^openblas threads=none', when='+openmp')
     conflicts('^openblas threads=pthreads', when='+openmp')
 
@@ -192,11 +193,11 @@ class Cp2k(MakefilePackage, CudaPackage):
         makefile_basename = '.'.join([
             self.makefile_architecture, self.makefile_version
         ])
-        return os.path.join('arch', makefile_basename)
+        return join_path('arch', makefile_basename)
 
     @property
     def archive_files(self):
-        return [os.path.join(self.stage.source_path, self.makefile)]
+        return [join_path(self.stage.source_path, self.makefile)]
 
     def edit(self, spec, prefix):
         pkgconf = which('pkg-config')
@@ -204,6 +205,9 @@ class Cp2k(MakefilePackage, CudaPackage):
         if '^fftw' in spec:
             fftw = spec['fftw:openmp' if '+openmp' in spec else 'fftw']
             fftw_header_dir = fftw.headers.directories[0]
+        elif '^amdfftw' in spec:
+            fftw = spec['amdfftw:openmp' if '+openmp' in spec else 'amdfftw']
+            fftw_header_dir = fftw.headers.directories[0]
         elif '^intel-mkl' in spec:
             fftw = spec['intel-mkl']
             fftw_header_dir = fftw.headers.directories[0] + '/fftw'
@@ -227,6 +231,7 @@ class Cp2k(MakefilePackage, CudaPackage):
             'nvhpc': ['-fast'],
             'cray': ['-O2'],
             'xl': ['-O3'],
+            'aocc': ['-O1'],
         }
 
         dflags = ['-DNDEBUG']
@@ -263,6 +268,10 @@ class Cp2k(MakefilePackage, CudaPackage):
                 '-ffree-line-length-none',
                 '-ggdb',  # make sure we get proper Fortran backtraces
             ]
+        elif '%aocc' in spec:
+            fcflags += [
+                '-ffree-form',
+            ]
         elif '%pgi' in spec or '%nvhpc' in spec:
             fcflags += ['-Mfreeform', '-Mextend']
         elif '%cray' in spec:
@@ -297,8 +306,8 @@ class Cp2k(MakefilePackage, CudaPackage):
             dflags.extend(['-D__PLUMED2'])
             cppflags.extend(['-D__PLUMED2'])
             libs.extend([
-                os.path.join(self.spec['plumed'].prefix.lib,
-                             'libplumed.{0}'.format(dso_suffix))
+                join_path(self.spec['plumed'].prefix.lib,
+                          'libplumed.{0}'.format(dso_suffix))
             ])
 
         cc = spack_cc if '~mpi' in spec else spec['mpi'].mpicc
@@ -351,7 +360,7 @@ class Cp2k(MakefilePackage, CudaPackage):
 
             if 'wannier90' in spec:
                 cppflags.append('-D__WANNIER90')
-                wannier = os.path.join(
+                wannier = join_path(
                     spec['wannier90'].libs.directories[0], 'libwannier.a'
                 )
                 libs.append(wannier)
@@ -371,9 +380,9 @@ class Cp2k(MakefilePackage, CudaPackage):
                 # runtime due to wrong offsets into the shared library
                 # symbols.
                 libs.extend([
-                    os.path.join(
+                    join_path(
                         spec['libint'].libs.directories[0], 'libderiv.a'),
-                    os.path.join(
+                    join_path(
                         spec['libint'].libs.directories[0], 'libint.a'),
                 ])
             else:
@@ -394,18 +403,17 @@ class Cp2k(MakefilePackage, CudaPackage):
 
         if '+pexsi' in spec:
             cppflags.append('-D__LIBPEXSI')
-            fcflags.append('-I' + os.path.join(
+            fcflags.append('-I' + join_path(
                 spec['pexsi'].prefix, 'fortran'))
             libs.extend([
-                os.path.join(spec['pexsi'].libs.directories[0],
-                             'libpexsi.a'),
-                os.path.join(spec['superlu-dist'].libs.directories[0],
-                             'libsuperlu_dist.a'),
-                os.path.join(
+                join_path(spec['pexsi'].libs.directories[0], 'libpexsi.a'),
+                join_path(spec['superlu-dist'].libs.directories[0],
+                          'libsuperlu_dist.a'),
+                join_path(
                     spec['parmetis'].libs.directories[0],
                     'libparmetis.{0}'.format(dso_suffix)
                 ),
-                os.path.join(
+                join_path(
                     spec['metis'].libs.directories[0],
                     'libmetis.{0}'.format(dso_suffix)
                 ),
@@ -416,11 +424,18 @@ class Cp2k(MakefilePackage, CudaPackage):
             elpa_suffix = '_openmp' if '+openmp' in elpa else ''
             elpa_incdir = elpa.headers.directories[0]
 
-            fcflags += ['-I{0}'.format(os.path.join(elpa_incdir, 'modules'))]
-            libs.append(os.path.join(elpa.libs.directories[0],
-                                     ('libelpa{elpa_suffix}.{dso_suffix}'
-                                      .format(elpa_suffix=elpa_suffix,
-                                              dso_suffix=dso_suffix))))
+            fcflags += ['-I{0}'.format(join_path(elpa_incdir, 'modules'))]
+
+            # Currently AOCC support only static libraries of ELPA
+            if '%aocc' in spec:
+                libs.append(join_path(elpa.prefix.lib,
+                            ('libelpa{elpa_suffix}.a'
+                                .format(elpa_suffix=elpa_suffix))))
+            else:
+                libs.append(join_path(elpa.prefix.lib,
+                            ('libelpa{elpa_suffix}.{dso_suffix}'
+                                .format(elpa_suffix=elpa_suffix,
+                                        dso_suffix=dso_suffix))))
 
             if spec.satisfies('@:4.999'):
                 if elpa.satisfies('@:2014.5.999'):
@@ -433,7 +448,7 @@ class Cp2k(MakefilePackage, CudaPackage):
                 cppflags.append('-D__ELPA={0}{1:02d}'
                                 .format(elpa.version[0],
                                         int(elpa.version[1])))
-                fcflags += ['-I{0}'.format(os.path.join(elpa_incdir, 'elpa'))]
+                fcflags += ['-I{0}'.format(join_path(elpa_incdir, 'elpa'))]
 
         if spec.satisfies('+sirius'):
             sirius = spec['sirius']
@@ -469,12 +484,12 @@ class Cp2k(MakefilePackage, CudaPackage):
                     gpuver = 'K20X'
 
         if 'smm=libsmm' in spec:
-            lib_dir = os.path.join(
+            lib_dir = join_path(
                 'lib', self.makefile_architecture, self.makefile_version
             )
             mkdirp(lib_dir)
             try:
-                copy(env['LIBSMM_PATH'], os.path.join(lib_dir, 'libsmm.a'))
+                copy(env['LIBSMM_PATH'], join_path(lib_dir, 'libsmm.a'))
             except KeyError:
                 raise KeyError('Point environment variable LIBSMM_PATH to '
                                'the absolute path of the libsmm.a file')
@@ -545,7 +560,7 @@ class Cp2k(MakefilePackage, CudaPackage):
 
             if spec.satisfies('+cuda'):
                 mkf.write('NVCC = {0}\n'.format(
-                    os.path.join(spec['cuda'].prefix, 'bin', 'nvcc')))
+                    join_path(spec['cuda'].prefix, 'bin', 'nvcc')))
 
             # Write compiler flags to file
             def fflags(var, lst):
@@ -576,7 +591,7 @@ class Cp2k(MakefilePackage, CudaPackage):
 
         if self.spec.satisfies('@:6.9999'):
             # prior to version 7.1 was the Makefile located in makefiles/
-            build_dir = os.path.join(build_dir, 'makefiles')
+            build_dir = join_path(build_dir, 'makefiles')
 
         return build_dir
 
@@ -597,12 +612,12 @@ class Cp2k(MakefilePackage, CudaPackage):
             super(Cp2k, self).build(spec, prefix)
 
     def install(self, spec, prefix):
-        exe_dir = os.path.join('exe', self.makefile_architecture)
+        exe_dir = join_path('exe', self.makefile_architecture)
         install_tree(exe_dir, self.prefix.bin)
         install_tree('data', self.prefix.share.data)
 
     def check(self):
-        data_dir = os.path.join(self.stage.source_path, 'data')
+        data_dir = join_path(self.stage.source_path, 'data')
 
         # CP2K < 7 still uses $PWD to detect the current working dir
         # and Makefile is in a subdir, account for both facts here:
diff --git a/var/spack/repos/builtin/packages/elpa/package.py b/var/spack/repos/builtin/packages/elpa/package.py
index c168163e7c..c05b712900 100644
--- a/var/spack/repos/builtin/packages/elpa/package.py
+++ b/var/spack/repos/builtin/packages/elpa/package.py
@@ -22,6 +22,7 @@ class Elpa(AutotoolsPackage, CudaPackage):
     version('2017.11.001', sha256='59f99c3abe2190fac0db8a301d0b9581ee134f438669dbc92551a54f6f861820')
     version('2017.05.003', sha256='bccd49ce35a323bd734b17642aed8f2588fea4cc78ee8133d88554753bc3bf1b')
     version('2017.05.002', sha256='568b71024c094d667b5cbb23045ad197ed5434071152ac608dae490ace5eb0aa')
+    version('2017.05.001', sha256='28f7edad60984d93da299016ad33571dc6db1cdc9fab0ceaef05dc07de2c7dfd')
     version('2016.11.001.pre', sha256='69b67f0f6faaa2b3b5fd848127b632be32771636d2ad04583c5269d550956f92')
     version('2016.05.004', sha256='08c59dc9da458bab856f489d779152e5506e04f0d4b8d6dcf114ca5fbbe46c58')
     version('2016.05.003', sha256='c8da50c987351514e61491e14390cdea4bdbf5b09045261991876ed5b433fca4')
@@ -79,6 +80,10 @@ class Elpa(AutotoolsPackage, CudaPackage):
         if spec.target.family == 'aarch64':
             options.append('--disable-sse-assembly')
 
+        if '%aocc' in spec:
+            options.append('--disable-shared')
+            options.append('--enable-static')
+
         # If no features are found, enable the generic ones
         if not any(f in spec.target for f in simd_features):
             options.append('--enable-generic')
@@ -89,6 +94,12 @@ class Elpa(AutotoolsPackage, CudaPackage):
                 'CFLAGS=-O2'
             ])
 
+        if '%aocc' in spec:
+            options.extend([
+                'FCFLAGS=-O3',
+                'CFLAGS=-O3'
+            ])
+
         if '+cuda' in spec:
             prefix = spec['cuda'].prefix
             options.append('--enable-gpu')
@@ -113,7 +124,8 @@ class Elpa(AutotoolsPackage, CudaPackage):
             'FC={0}'.format(spec['mpi'].mpifc),
             'CXX={0}'.format(spec['mpi'].mpicxx),
             'LDFLAGS={0}'.format(spec['lapack'].libs.search_flags),
-            'LIBS={0}'.format(spec['lapack'].libs.link_flags),
+            'LIBS={0} {1}'.format(
+                spec['lapack'].libs.link_flags, spec['blas'].libs.link_flags),
             'SCALAPACK_LDFLAGS={0}'.format(spec['scalapack'].libs.joined())
         ])
 
diff --git a/var/spack/repos/builtin/packages/libxsmm/package.py b/var/spack/repos/builtin/packages/libxsmm/package.py
index 3de1cfb461..8be2af3011 100644
--- a/var/spack/repos/builtin/packages/libxsmm/package.py
+++ b/var/spack/repos/builtin/packages/libxsmm/package.py
@@ -82,7 +82,7 @@ class Libxsmm(MakefilePackage):
         ]
 
         # JIT (AVX and later) makes MNK, M, N, or K spec. superfluous
-#       make_args += ['MNK=1 4 5 6 8 9 13 16 17 22 23 24 26 32']
+        # make_args += ['MNK=1 4 5 6 8 9 13 16 17 22 23 24 26 32']
 
         # include call trace as the build is already de-optimized
         if '+debug' in spec: