def _nvcc_gencode_options(cuda_version: int) -> List[str]:
"""Returns NVCC GPU code generation options."""
if sys.argv == ['setup.py', 'develop']:
return []
envcfg = os.getenv('CUPY_NVCC_GENERATE_CODE', None)
if envcfg is not None and envcfg != 'current':
return [
'--generate-code={}'.format(arch) for arch in envcfg.split(';')
if len(arch) > 0
]
if envcfg == 'current' and build.get_compute_capabilities() is not None:
ccs = build.get_compute_capabilities()
arch_list = [
f'compute_{cc}' if cc < 60 else (f'compute_{cc}', f'sm_{cc}')
for cc in ccs
]
else:
# The arch_list specifies virtual architectures, such as 'compute_61',
# and real architectures, such as 'sm_61', for which the CUDA
# input files are to be compiled.
#
# The syntax of an entry of the list is
#
# entry ::= virtual_arch | (virtual_arch, real_arch)
#
# where virtual_arch is a string which means a virtual architecture and
# real_arch is a string which means a real architecture.
#
# If a virtual architecture is supplied, NVCC generates a PTX code
# the virtual architecture. If a pair of a virtual architecture and a
# real architecture is supplied, NVCC generates a PTX code for the
# virtual architecture as well as a cubin code for the real one.
#
# For example, making NVCC generate a PTX code for 'compute_60' virtual
# architecture, the arch_list has an entry of 'compute_60'.
#
# arch_list = ['compute_60']
#
# For another, making NVCC generate a PTX code for 'compute_61' virtual
# architecture and a cubin code for 'sm_61' real architecture, the
# arch_list has an entry of ('compute_61', 'sm_61').
#
# arch_list = [('compute_61', 'sm_61')]
#
# See the documentation of each CUDA version for the list of supported
# architectures:
#
# https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-steering-gpu-code-generation
if cuda_version >= 11040:
arch_list = [
'compute_35', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), ('compute_80', 'sm_80'),
('compute_86', 'sm_86'), ('compute_87', 'sm_87'), 'compute_87'
]
elif cuda_version >= 11010:
arch_list = [
'compute_35', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), ('compute_80', 'sm_80'),
('compute_86', 'sm_86'), 'compute_86'
]
elif cuda_version >= 11000:
arch_list = [
'compute_35', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), ('compute_80', 'sm_80'), 'compute_80'
]
elif cuda_version >= 10000:
arch_list = [
'compute_30', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), 'compute_70'
]
else:
# This should not happen.
assert False
options = []
for arch in arch_list:
if type(arch) is tuple:
virtual_arch, real_arch = arch
options.append('--generate-code=arch={},code={}'.format(
virtual_arch, real_arch))
else:
options.append('--generate-code=arch={},code={}'.format(
arch, arch))
return options
def _nvcc_gencode_options(cuda_version: int) -> List[str]:
"""Returns NVCC GPU code generation options."""
if sys.argv == ['setup.py', 'develop']:
return []
envcfg = os.getenv('CUPY_NVCC_GENERATE_CODE', None)
if envcfg is not None and envcfg != 'current':
return [
'--generate-code={}'.format(arch) for arch in envcfg.split(';')
if len(arch) > 0
]
if envcfg == 'current' and build.get_compute_capabilities() is not None:
ccs = build.get_compute_capabilities()
arch_list = [
f'compute_{cc}' if cc < 60 else (f'compute_{cc}', f'sm_{cc}')
for cc in ccs
]
else:
# The arch_list specifies virtual architectures, such as 'compute_61',
# and real architectures, such as 'sm_61', for which the CUDA
# input files are to be compiled.
#
# The syntax of an entry of the list is
#
# entry ::= virtual_arch | (virtual_arch, real_arch)
#
# where virtual_arch is a string which means a virtual architecture and
# real_arch is a string which means a real architecture.
#
# If a virtual architecture is supplied, NVCC generates a PTX code
# the virtual architecture. If a pair of a virtual architecture and a
# real architecture is supplied, NVCC generates a PTX code for the
# virtual architecture as well as a cubin code for the real one.
#
# For example, making NVCC generate a PTX code for 'compute_60' virtual
# architecture, the arch_list has an entry of 'compute_60'.
#
# arch_list = ['compute_60']
#
# For another, making NVCC generate a PTX code for 'compute_61' virtual
# architecture and a cubin code for 'sm_61' real architecture, the
# arch_list has an entry of ('compute_61', 'sm_61').
#
# arch_list = [('compute_61', 'sm_61')]
#
# See the documentation of each CUDA version for the list of supported
# architectures:
#
# https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-steering-gpu-code-generation
aarch64 = (platform.machine() == 'aarch64')
if cuda_version >= 11040:
# To utilize CUDA Minor Version Compatibility (`cupy-cuda11x`),
# CUBIN must be generated for all supported compute capabilities
# instead of PTX:
# https://docs.nvidia.com/deploy/cuda-compatibility/index.html#application-considerations
arch_list = [('compute_35', 'sm_35'), ('compute_37', 'sm_37'),
('compute_50', 'sm_50'), ('compute_52', 'sm_52'),
('compute_60', 'sm_60'), ('compute_61', 'sm_61'),
('compute_70', 'sm_70'), ('compute_75', 'sm_75'),
('compute_80', 'sm_80'), ('compute_86', 'sm_86'),
'compute_86']
if aarch64:
# Jetson TX1/TX2 are excluded as they don't support JetPack 5
# (CUDA 11.4).
arch_list += [
# ('compute_53', 'sm_53'), # Jetson (TX1 / Nano)
# ('compute_62', 'sm_62'), # Jetson (TX2)
('compute_72', 'sm_72'), # Jetson (Xavier)
('compute_87', 'sm_87'), # Jetson (Orin)
]
elif cuda_version >= 11010:
arch_list = [
'compute_35', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), ('compute_80', 'sm_80'),
('compute_86', 'sm_86'), 'compute_86'
]
elif cuda_version >= 11000:
arch_list = [
'compute_35', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), ('compute_80', 'sm_80'), 'compute_80'
]
elif cuda_version >= 10000:
arch_list = [
'compute_30', 'compute_50', ('compute_60', 'sm_60'),
('compute_61', 'sm_61'), ('compute_70', 'sm_70'),
('compute_75', 'sm_75'), 'compute_70'
]
else:
# This should not happen.
assert False
options = []
for arch in arch_list:
if type(arch) is tuple:
virtual_arch, real_arch = arch
options.append('--generate-code=arch={},code={}'.format(
virtual_arch, real_arch))
else:
options.append('--generate-code=arch={},code={}'.format(
arch, arch))
return options