gcc: add nvptx offload capability (#9406)

Add optional support for Nvidia PTX

1 files changed, 101 insertions, 0 deletions

diff --git a/var/spack/repos/builtin/packages/gcc/package.py b/var/spack/repos/builtin/packages/gcc/package.py
index ddeb95d3ba..57d3ece2c8 100644
--- a/var/spack/repos/builtin/packages/gcc/package.py
+++ b/var/spack/repos/builtin/packages/gcc/package.py
@@ -18,9 +18,11 @@ class Gcc(AutotoolsPackage):
 
     homepage = 'https://gcc.gnu.org'
     url      = 'https://ftpmirror.gnu.org/gcc/gcc-7.1.0/gcc-7.1.0.tar.bz2'
+    svn      = 'svn://gcc.gnu.org/svn/gcc/'
     list_url = 'http://ftp.gnu.org/gnu/gcc/'
     list_depth = 1
 
+    version('develop', svn=svn + 'trunk')
     version('8.2.0', '64898a165f67e136d802a92e7633bf1b06c85266027e52127ea025bf5fc2291b5e858288aac0bdba246e6cdf7c6ec88bc8e0e7f3f6f1985f4297710cafde56ed')
     version('8.1.0', '65f7c65818dc540b3437605026d329fc')
 
@@ -74,6 +76,9 @@ class Gcc(AutotoolsPackage):
     variant('strip',
             default=False,
             description='Strip executables to reduce installation size')
+    variant('nvptx',
+            default=False,
+            description='Target nvptx offloading to NVIDIA GPUs')
 
     # https://gcc.gnu.org/install/prerequisites.html
     depends_on('gmp@4.3.2:')
@@ -92,6 +97,25 @@ class Gcc(AutotoolsPackage):
     depends_on('gnat', when='languages=ada')
     depends_on('binutils~libiberty', when='+binutils')
     depends_on('zip', type='build', when='languages=java')
+    depends_on('cuda', when='+nvptx')
+
+    resource(
+             name='newlib',
+             url='ftp://sourceware.org/pub/newlib/newlib-3.0.0.20180831.tar.gz',
+             sha256='3ad3664f227357df15ff34e954bfd9f501009a647667cd307bf0658aefd6eb5b',
+             destination='newlibsource',
+             when='+nvptx'
+            )
+
+    # nvptx-tools does not seem to work as a dependency,
+    # but does fine when the source is inside the gcc build directory
+    # nvptx-tools doesn't have any releases, so grabbing the last commit
+    resource(
+             name='nvptx-tools',
+             git='https://github.com/MentorEmbedded/nvptx-tools',
+             commit='5f6f343a302d620b0868edab376c00b15741e39e',
+             when='+nvptx'
+    )
 
     # TODO: integrate these libraries.
     # depends_on('ppl')
@@ -148,6 +172,16 @@ class Gcc(AutotoolsPackage):
     # See https://gcc.gnu.org/gcc-5/changes.html
     conflicts('languages=jit', when='@:4')
 
+    # NVPTX offloading supported in 7 and later by limited languages
+    conflicts('+nvptx', when='@:6', msg='NVPTX only supported in gcc 7 and above')
+    conflicts('languages=ada', when='+nvptx')
+    conflicts('languages=brig', when='+nvptx')
+    conflicts('languages=go', when='+nvptx')
+    conflicts('languages=java', when='+nvptx')
+    conflicts('languages=jit', when='+nvptx')
+    conflicts('languages=objc', when='+nvptx')
+    conflicts('languages=obj-c++', when='+nvptx')
+
     if sys.platform == 'darwin':
         # Fix parallel build on APFS filesystem
         # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81797
@@ -267,8 +301,75 @@ class Gcc(AutotoolsPackage):
         if sys.platform == 'darwin':
             options.append('--with-build-config=bootstrap-debug')
 
+        # nvptx-none offloading for host compiler
+        if spec.satisfies('+nvptx'):
+            options.extend(['--enable-offload-targets=nvptx-none',
+                            '--with-cuda-driver-include={0}'.format(
+                                spec['cuda'].prefix.include),
+                            '--with-cuda-driver-lib={0}'.format(
+                                spec['cuda'].libs.directories[0]),
+                            '--disable-bootstrap',
+                            '--disable-multilib'])
+
         return options
 
+    # run configure/make/make(install) for the nvptx-none target
+    # before running the host compiler phases
+    @when('+nvptx')
+    @run_before('configure')
+    def nvptx_install(self):
+        spec = self.spec
+        prefix = self.prefix
+
+        # config.guess returns the host triple, e.g. "x86_64-pc-linux-gnu"
+        guess = Executable('./config.guess')
+        targetguess = guess(output=str).rstrip('\n')
+
+        options = getattr(self, 'configure_flag_args', [])
+        options += ['--prefix={0}'.format(prefix)]
+
+        options += [
+            '--with-cuda-driver-include={0}'.format(
+                spec['cuda'].prefix.include),
+            '--with-cuda-driver-lib={0}'.format(
+                spec['cuda'].libs.directories[0]),
+        ]
+
+        with working_dir('nvptx-tools'):
+            configure = Executable("./configure")
+            configure(*options)
+            make()
+            make('install')
+
+        pattern = join_path(self.stage.source_path, 'newlibsource', '*')
+        files = glob.glob(pattern)
+
+        if files:
+            symlink(join_path(files[0], 'newlib'), 'newlib')
+
+        # self.build_directory = 'spack-build-nvptx'
+        with working_dir('spack-build-nvptx', create=True):
+
+            options = ['--prefix={0}'.format(prefix),
+                       '--enable-languages={0}'.format(
+                       ','.join(spec.variants['languages'].value)),
+                       '--with-mpfr={0}'.format(spec['mpfr'].prefix),
+                       '--with-gmp={0}'.format(spec['gmp'].prefix),
+                       '--target=nvptx-none',
+                       '--with-build-time-tools={0}'.format(
+                           join_path(prefix,
+                                     'nvptx-none', 'bin')),
+                       '--enable-as-accelerator-for={0}'.format(
+                           targetguess),
+                       '--disable-sjlj-exceptions',
+                       '--enable-newlib-io-long-long',
+                       ]
+
+            configure = Executable("../configure")
+            configure(*options)
+            make()
+            make('install')
+
     @property
     def build_targets(self):
         if sys.platform == 'darwin':