def compile(self, obj: str, src: str, ext: Extension) -> None:
if self._context.use_hip:
raise RuntimeError('ROCm is not supported on Windows')
compiler_so = build.get_nvcc_path()
cc_args = self._get_preprocess_options(ext) + ['-c']
cuda_version = build.get_cuda_version()
postargs = _nvcc_gencode_options(cuda_version) + ['-O2']
if cuda_version >= 11020:
# MSVC 14.0 (2015) is deprecated for CUDA 11.2 but we need it
# to build CuPy because some Python versions were built using it.
# REF: https://wiki.python.org/moin/WindowsCompilers
postargs += ['-allow-unsupported-compiler']
postargs += ['-Xcompiler', '/MD', '-D_USE_MATH_DEFINES']
# This is to compile thrust with MSVC2015
if cuda_version >= 11020:
postargs += ['--std=c++14']
num_threads = int(os.environ.get('CUPY_NUM_NVCC_THREADS', '2'))
postargs += [f'-t{num_threads}']
cl_exe_path = self._find_host_compiler_path()
if cl_exe_path is None:
print('Warning: Host compiler path could not be detected')
else:
postargs += ['--compiler-bindir', cl_exe_path]
print('NVCC options:', postargs)
self.spawn(compiler_so + cc_args + [src, '-o', obj] + postargs)
def run(self):
if (build.get_nvcc_path() is not None
or build.get_hipcc_path() is not None):
def wrap_new_compiler(func):
def _wrap_new_compiler(*args, **kwargs):
try:
return func(*args, **kwargs)
except errors.DistutilsPlatformError:
if not PLATFORM_WIN32:
CCompiler = _UnixCCompiler
else:
CCompiler = _MSVCCompiler
return CCompiler(None, kwargs['dry_run'],
kwargs['force'])
return _wrap_new_compiler
ccompiler.new_compiler = wrap_new_compiler(ccompiler.new_compiler)
# Intentionally causes DistutilsPlatformError in
# ccompiler.new_compiler() function to hook.
self.compiler = 'nvidia'
ctx = cupy_builder.get_context()
ext_modules = get_ext_modules(True, ctx) # get .pyx modules
cythonize(ext_modules, ctx)
check_extensions(self.extensions)
build_ext.build_ext.run(self)
def _compile_cu(self,
sources,
output_dir=None,
macros=None,
include_dirs=None,
debug=0,
extra_preargs=None,
extra_postargs=None,
depends=None):
# Compile CUDA C files, mainly derived from UnixCCompiler._compile().
macros, objects, extra_postargs, pp_opts, _build = \
self._setup_compile(output_dir, macros, include_dirs, sources,
depends, extra_postargs)
compiler_so = build.get_nvcc_path()
cc_args = self._get_cc_args(pp_opts, debug, extra_preargs)
cuda_version = build.get_cuda_version()
postargs = _nvcc_gencode_options(cuda_version) + ['-O2']
if cuda_version >= 11020:
# MSVC 14.0 (2015) is deprecated for CUDA 11.2 but we need it
# to build CuPy because some Python versions were built using it.
# REF: https://wiki.python.org/moin/WindowsCompilers
postargs += ['-allow-unsupported-compiler']
postargs += ['-Xcompiler', '/MD', '-D_USE_MATH_DEFINES']
# This is to compile thrust with MSVC2015
if cuda_version >= 11020:
postargs += ['--std=c++14']
num_threads = int(os.environ.get('CUPY_NUM_NVCC_THREADS', '2'))
postargs += [f'-t{num_threads}']
print('NVCC options:', postargs)
for obj in objects:
try:
src, ext = _build[obj]
except KeyError:
continue
try:
self.spawn(compiler_so + cc_args + [src, '-o', obj] + postargs)
except errors.DistutilsExecError as e:
raise errors.CompileError(str(e))
return objects
def _compile_unix_nvcc(self, obj: str, src: str, ext: Extension) -> None:
cc_args = self._get_preprocess_options(ext) + ['-c']
# For CUDA C source files, compile them with NVCC.
nvcc_path = build.get_nvcc_path()
base_opts = build.get_compiler_base_options(nvcc_path)
compiler_so = nvcc_path
cuda_version = build.get_cuda_version()
postargs = _nvcc_gencode_options(cuda_version) + [
'-O2', '--compiler-options="-fPIC"'
]
if cuda_version >= 11020:
postargs += ['--std=c++14']
num_threads = int(os.environ.get('CUPY_NUM_NVCC_THREADS', '2'))
postargs += [f'-t{num_threads}']
else:
postargs += ['--std=c++11']
postargs += ['-Xcompiler=-fno-gnu-unique']
print('NVCC options:', postargs)
self.spawn(compiler_so + base_opts + cc_args + [src, '-o', obj] +
postargs)
def _compile_unix_nvcc(self,
obj, src, ext, cc_args, extra_postargs, pp_opts):
# For CUDA C source files, compile them with NVCC.
nvcc_path = build.get_nvcc_path()
base_opts = build.get_compiler_base_options(nvcc_path)
compiler_so = nvcc_path
cuda_version = build.get_cuda_version()
postargs = _nvcc_gencode_options(cuda_version) + [
'-O2', '--compiler-options="-fPIC"']
if cuda_version >= 11020:
postargs += ['--std=c++14']
num_threads = int(os.environ.get('CUPY_NUM_NVCC_THREADS', '2'))
postargs += [f'-t{num_threads}']
else:
postargs += ['--std=c++11']
print('NVCC options:', postargs)
try:
self.spawn(compiler_so + base_opts + cc_args + [src, '-o', obj] +
postargs)
except errors.DistutilsExecError as e:
raise errors.CompileError(str(e))
def preconfigure_modules(MODULES, compiler, settings):
"""Returns a list of modules buildable in given environment and settings.
For each module in MODULES list, this function checks if the module
can be built in the current environment and reports it.
Returns a list of module names available.
"""
nvcc_path = build.get_nvcc_path()
hipcc_path = build.get_hipcc_path()
summary = [
'',
'************************************************************',
'* CuPy Configuration Summary *',
'************************************************************',
'',
'Build Environment:',
' Include directories: {}'.format(str(settings['include_dirs'])),
' Library directories: {}'.format(str(settings['library_dirs'])),
' nvcc command : {}'.format(
nvcc_path if nvcc_path else '(not found)'),
' hipcc command : {}'.format(
hipcc_path if hipcc_path else '(not found)'),
'',
'Environment Variables:',
]
for key in [
'CFLAGS', 'LDFLAGS', 'LIBRARY_PATH', 'CUDA_PATH',
'NVTOOLSEXT_PATH', 'NVCC', 'HIPCC', 'ROCM_HOME'
]:
summary += [' {:<16}: {}'.format(key, os.environ.get(key, '(none)'))]
summary += [
'',
'Modules:',
]
ret = []
for module in MODULES:
installed = False
status = 'No'
errmsg = []
if module['name'] == 'cutensor':
cutensor_path = os.environ.get('CUTENSOR_PATH', '')
inc_path = os.path.join(cutensor_path, 'include')
if os.path.exists(inc_path):
settings['include_dirs'].append(inc_path)
cuda_version = build.get_cuda_version()
cuda_major = str(cuda_version // 1000)
cuda_major_minor = cuda_major + '.' + \
str((cuda_version // 10) % 100)
for cuda_ver in (cuda_major_minor, cuda_major):
lib_path = os.path.join(cutensor_path, 'lib', cuda_ver)
if os.path.exists(lib_path):
settings['library_dirs'].append(lib_path)
break
# In ROCm 4.1 and later, we need to use the independent version of
# hipfft as well as rocfft. We configure the lists of include
# directories and libraries to link here depending on ROCm version
# before the configuration process following.
if use_hip and module['name'] == 'cuda':
if module['check_method'](compiler, settings):
hip_version = module['version_method']()
if hip_version >= 401:
rocm_path = build.get_rocm_path()
inc_path = os.path.join(rocm_path, 'hipfft', 'include')
settings['include_dirs'].insert(0, inc_path)
lib_path = os.path.join(rocm_path, 'hipfft', 'lib')
settings['library_dirs'].insert(0, lib_path)
# n.b., this modifieds MODULES['cuda']['libraries'] inplace
canonicalize_hip_libraries(hip_version, module['libraries'])
print('')
print('-------- Configuring Module: {} --------'.format(
module['name']))
sys.stdout.flush()
if not check_library(
compiler,
includes=module['include'],
include_dirs=settings['include_dirs'],
define_macros=settings['define_macros'],
extra_compile_args=settings['extra_compile_args']):
errmsg = [
'Include files not found: %s' % module['include'],
'Check your CFLAGS environment variable.'
]
elif not check_library(
compiler,
libraries=module['libraries'],
library_dirs=settings['library_dirs'],
define_macros=settings['define_macros'],
extra_compile_args=settings['extra_compile_args']):
errmsg = [
'Cannot link libraries: %s' % module['libraries'],
'Check your LDFLAGS environment variable.'
]
elif ('check_method' in module
and not module['check_method'](compiler, settings)):
# Fail on per-library condition check (version requirements etc.)
installed = True
errmsg = ['The library is installed but not supported.']
elif (module['name'] in ('thrust', 'cub', 'random')
and (nvcc_path is None and hipcc_path is None)):
installed = True
cmd = 'nvcc' if not use_hip else 'hipcc'
errmsg = [
'{} command could not be found in PATH.'.format(cmd),
'Check your PATH environment variable.'
]
else:
installed = True
status = 'Yes'
ret.append(module['name'])
if installed and 'version_method' in module:
status += ' (version {})'.format(module['version_method'](True))
summary += [' {:<10}: {}'.format(module['name'], status)]
# If error message exists...
if len(errmsg) != 0:
summary += [' -> {}'.format(m) for m in errmsg]
# Skip checking other modules when CUDA is unavailable.
if module['name'] == 'cuda':
break
# Get a list of the CC of the devices connected to this node
if not use_hip:
build.check_compute_capabilities(compiler, settings)
if len(ret) != len(MODULES):
if 'cuda' in ret:
lines = [
'WARNING: Some modules could not be configured.',
'CuPy will be installed without these modules.',
]
else:
lines = [
'ERROR: CUDA could not be found on your system.',
'',
'HINT: You are trying to build CuPy from source, '
'which is NOT recommended for general use.',
' Please consider using binary packages instead.',
'',
]
summary += [
'',
] + lines + [
'Please refer to the Installation Guide for details:',
'https://docs.cupy.dev/en/stable/install.html',
'',
]
summary += [
'************************************************************',
'',
]
print('\n'.join(summary))
return ret, settings
