def _comiple_unix_hipcc(self, obj, src, ext, cc_args, extra_postargs,
pp_opts):
# For CUDA C source files, compile them with HIPCC.
_compiler_so = self.compiler_so
try:
rcom_path = build.get_hipcc_path()
base_opts = build.get_compiler_base_options()
self.set_executable('compiler_so', rcom_path)
postargs = ['-O2', '-fPIC', '--include', 'hip_runtime.h']
print('HIPCC options:', postargs)
return unixccompiler.UnixCCompiler._compile(
self, obj, src, ext, base_opts + cc_args, postargs, pp_opts)
finally:
self.compiler_so = _compiler_so
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"', '--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 _comiple_unix_nvcc(self,
obj, src, ext, cc_args, extra_postargs, pp_opts):
# For CUDA C source files, compile them with NVCC.
_compiler_so = self.compiler_so
try:
nvcc_path = build.get_nvcc_path()
base_opts = build.get_compiler_base_options()
self.set_executable('compiler_so', nvcc_path)
cuda_version = build.get_cuda_version()
postargs = _nvcc_gencode_options(cuda_version) + [
'-O2', '--compiler-options="-fPIC"', '--std=c++11']
print('NVCC options:', postargs)
return unixccompiler.UnixCCompiler._compile(
self, obj, src, ext, base_opts + cc_args, postargs, pp_opts)
finally:
self.compiler_so = _compiler_so
def _compile_unix_hipcc(self, obj, src, ext, cc_args, extra_postargs,
pp_opts):
# For CUDA C source files, compile them with HIPCC.
rocm_path = build.get_hipcc_path()
base_opts = build.get_compiler_base_options(rocm_path)
compiler_so = rocm_path
hip_version = build.get_hip_version()
postargs = ['-O2', '-fPIC', '--include', 'hip_runtime.h']
if hip_version >= 402:
postargs += ['--std=c++14']
else:
postargs += ['--std=c++11']
print('HIPCC 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 _compile(self, obj, src, ext, cc_args, extra_postargs, pp_opts):
# For sources other than CUDA C ones, just call the super class method.
if os.path.splitext(src)[1] != '.cu':
return unixccompiler.UnixCCompiler._compile(
self, obj, src, ext, cc_args, extra_postargs, pp_opts)
# For CUDA C source files, compile them with NVCC.
_compiler_so = self.compiler_so
try:
nvcc_path = build.get_nvcc_path()
base_opts = build.get_compiler_base_options()
self.set_executable('compiler_so', nvcc_path)
cuda_version = build.get_cuda_version()
postargs = _nvcc_gencode_options(cuda_version) + [
'-O2', '--compiler-options="-fPIC"']
print('NVCC options:', postargs)
return unixccompiler.UnixCCompiler._compile(
self, obj, src, ext, base_opts + cc_args, postargs, pp_opts)
finally:
self.compiler_so = _compiler_so
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))
