def __init__(self,
nvcc='nvcc',
link_options=None,
keep=False,
no_extern_c=False,
arch=None,
code=None,
cache_dir=None,
include_dirs=[],
message_handler=None,
log_verbose=False,
cuda_libdir=None):
from pycuda.driver import Context
compute_capability = Context.get_device().compute_capability()
if compute_capability < (3, 5):
raise Exception(
'Minimum compute capability for dynamic parallelism is 3.5 (found: %u.%u)!'
% (compute_capability[0], compute_capability[1]))
else:
from pycuda.driver import Linker
self.linker = Linker(message_handler, link_options, log_verbose)
self._check_arch(arch)
self.nvcc = nvcc
self.keep = keep
self.no_extern_c = no_extern_c
self.arch = arch
self.code = code
self.cache_dir = cache_dir
self.include_dirs = include_dirs
self.cuda_libdir = cuda_libdir
self.libdir, self.libptn = None, None
self.module = None
def __init__(self, nvcc='nvcc', link_options=None, keep=False,
no_extern_c=False, arch=None, code=None, cache_dir=None,
include_dirs=[], message_handler=None, log_verbose=False,
cuda_libdir=None):
from pycuda.driver import Context
compute_capability = Context.get_device().compute_capability()
if compute_capability < (3,5):
raise Exception('Minimum compute capability for dynamic parallelism is 3.5 (found: %u.%u)!' %
(compute_capability[0], compute_capability[1]))
else:
from pycuda.driver import Linker
self.linker = Linker(message_handler, link_options, log_verbose)
self._check_arch(arch)
self.nvcc = nvcc
self.keep = keep
self.no_extern_c = no_extern_c
self.arch = arch
self.code = code
self.cache_dir = cache_dir
self.include_dirs = include_dirs
self.cuda_libdir = cuda_libdir
self.libdir, self.libptn = None, None
self.module = None
class DynamicModule(CudaModule):
'''
Creates a Module from multiple .cu source, library file and/or data
objects linked against the static or dynamic CUDA runtime.
'''
def __init__(self, nvcc='nvcc', link_options=None, keep=False,
no_extern_c=False, arch=None, code=None, cache_dir=None,
include_dirs=[], message_handler=None, log_verbose=False,
cuda_libdir=None):
from pycuda.driver import Context
compute_capability = Context.get_device().compute_capability()
if compute_capability < (3,5):
raise Exception('Minimum compute capability for dynamic parallelism is 3.5 (found: %u.%u)!' %
(compute_capability[0], compute_capability[1]))
else:
from pycuda.driver import Linker
self.linker = Linker(message_handler, link_options, log_verbose)
self._check_arch(arch)
self.nvcc = nvcc
self.keep = keep
self.no_extern_c = no_extern_c
self.arch = arch
self.code = code
self.cache_dir = cache_dir
self.include_dirs = include_dirs
self.cuda_libdir = cuda_libdir
self.libdir, self.libptn = None, None
self.module = None
def _locate_cuda_libdir(self):
'''
Locate the "standard" CUDA SDK library directory in the local
file system. Supports 64-Bit Windows, Linux and Mac OS X.
In case the caller supplied cuda_libdir in the constructor
other than None that value is returned unchecked, else a
best-effort attempt is made.
Precedence:
Windows: cuda_libdir > %CUDA_PATH%
Linux: cuda_libdir > $CUDA_ROOT > $LD_LIBRARY_PATH > '/usr/lib/x86_64-linux-gnu'
Returns a pair (libdir, libptn) where libdir is None in case
of failure or a string containing the absolute path of the
directory, and libptn is the %-format pattern to construct
library file names from library names on the local system.
Raises a RuntimeError in case of failure.
Links:
- Post-installation Actions
http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#post-installation-actions
TODO:
- Is $CUDA_ROOT/lib64 the correct path to assume for 64-Bit CUDA libraries on Linux?
- Mac OS X (Darwin) is currently treated like Linux, is that correct?
- Check CMake's FindCUDA module, it might contain some helpful clues in its sources
https://cmake.org/cmake/help/v3.0/module/FindCUDA.html
https://github.com/Kitware/CMake/blob/master/Modules/FindCUDA.cmake
- Verify all Linux code paths somehow
'''
from os.path import isfile, join
from platform import system as platform_system
system = platform_system()
libdir, libptn = None, None
if system == 'Windows':
if self.cuda_libdir is not None:
libdir = self.cuda_libdir
elif 'CUDA_PATH' in os.environ and isfile(join(os.environ['CUDA_PATH'], 'lib\\x64\\cudadevrt.lib')):
libdir = join(os.environ['CUDA_PATH'], 'lib\\x64')
libptn = '%s.lib'
elif system == 'Linux' or system == 'Darwin':
if self.cuda_libdir is not None:
libdir = self.cuda_libdir
elif 'CUDA_ROOT' in os.environ and isfile(join(os.environ['CUDA_ROOT'], 'lib64/libcudadevrt.a')):
libdir = join(os.environ['CUDA_ROOT'], 'lib64')
elif 'LD_LIBRARY_PATH' in os.environ:
for ld_path in os.environ['LD_LIBRARY_PATH'].split(':'):
if isfile(join(ld_path, 'libcudadevrt.a')):
libdir = ld_path
break
if libdir is None and isfile('/usr/lib/x86_64-linux-gnu/libcudadevrt.a'):
libdir = '/usr/lib/x86_64-linux-gnu'
libptn = 'lib%s.a'
if libdir is None:
raise RuntimeError('Unable to locate the CUDA SDK installation '
'directory, set CUDA library path manually')
return libdir, libptn
def add_source(self, source, nvcc_options=None, name='kernel.ptx'):
ptx = compile(source, nvcc=self.nvcc, options=nvcc_options,
keep=self.keep, no_extern_c=self.no_extern_c, arch=self.arch,
code=self.code, cache_dir=self.cache_dir,
include_dirs=self.include_dirs, target="ptx")
from pycuda.driver import jit_input_type
self.linker.add_data(ptx, jit_input_type.PTX, name)
return self
def add_data(self, data, input_type, name='unknown'):
self.linker.add_data(data, input_type, name)
return self
def add_file(self, filename, input_type):
self.linker.add_file(filename, input_type)
return self
def add_stdlib(self, libname):
if self.libdir is None:
self.libdir, self.libptn = self._locate_cuda_libdir()
from os.path import isfile, join
libpath = join(self.libdir, self.libptn % libname)
if not isfile(libpath):
raise FileNotFoundError('CUDA SDK library file "%s" not found' % libpath)
from pycuda.driver import jit_input_type
self.linker.add_file(libpath, jit_input_type.LIBRARY)
return self
def link(self):
self.module = self.linker.link_module()
self.linker = None
self._bind_module()
return self
class DynamicModule(CudaModule):
"""
Creates a Module from multiple .cu source, library file and/or data
objects linked against the static or dynamic CUDA runtime.
"""
def __init__(
self,
nvcc="nvcc",
link_options=None,
keep=False,
no_extern_c=False,
arch=None,
code=None,
cache_dir=None,
include_dirs=[],
message_handler=None,
log_verbose=False,
cuda_libdir=None,
):
from pycuda.driver import Context
compute_capability = Context.get_device().compute_capability()
if compute_capability < (3, 5):
raise Exception(
"Minimum compute capability for dynamic parallelism is 3.5 (found: %u.%u)!"
% (compute_capability[0], compute_capability[1]))
else:
from pycuda.driver import Linker
self.linker = Linker(message_handler, link_options, log_verbose)
self._check_arch(arch)
self.nvcc = nvcc
self.keep = keep
self.no_extern_c = no_extern_c
self.arch = arch
self.code = code
self.cache_dir = cache_dir
self.include_dirs = include_dirs
self.cuda_libdir = cuda_libdir
self.libdir, self.libptn = None, None
self.module = None
def _locate_cuda_libdir(self):
"""
Locate the "standard" CUDA SDK library directory in the local
file system. Supports 64-Bit Windows, Linux and Mac OS X.
In case the caller supplied cuda_libdir in the constructor
other than None that value is returned unchecked, else a
best-effort attempt is made.
Precedence:
Windows: cuda_libdir > %CUDA_PATH%
Linux: cuda_libdir > $CUDA_ROOT > $LD_LIBRARY_PATH > '/usr/lib/x86_64-linux-gnu'
Returns a pair (libdir, libptn) where libdir is None in case
of failure or a string containing the absolute path of the
directory, and libptn is the %-format pattern to construct
library file names from library names on the local system.
Raises a RuntimeError in case of failure.
Links:
- Post-installation Actions
http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#post-installation-actions
TODO:
- Is $CUDA_ROOT/lib64 the correct path to assume for 64-Bit CUDA libraries on Linux?
- Mac OS X (Darwin) is currently treated like Linux, is that correct?
- Check CMake's FindCUDA module, it might contain some helpful clues in its sources
https://cmake.org/cmake/help/v3.0/module/FindCUDA.html
https://github.com/Kitware/CMake/blob/master/Modules/FindCUDA.cmake
- Verify all Linux code paths somehow
"""
from os.path import isfile, join
from platform import system as platform_system
system = platform_system()
libdir, libptn = None, None
if system == "Windows":
if self.cuda_libdir is not None:
libdir = self.cuda_libdir
elif "CUDA_PATH" in os.environ and isfile(
join(os.environ["CUDA_PATH"], "lib\\x64\\cudadevrt.lib")):
libdir = join(os.environ["CUDA_PATH"], "lib\\x64")
libptn = "%s.lib"
elif system in ["Linux", "Darwin"]:
if self.cuda_libdir is not None:
libdir = self.cuda_libdir
elif "CUDA_ROOT" in os.environ and isfile(
join(os.environ["CUDA_ROOT"], "lib64/libcudadevrt.a")):
libdir = join(os.environ["CUDA_ROOT"], "lib64")
elif "LD_LIBRARY_PATH" in os.environ:
for ld_path in os.environ["LD_LIBRARY_PATH"].split(":"):
if isfile(join(ld_path, "libcudadevrt.a")):
libdir = ld_path
break
if libdir is None and isfile(
"/usr/lib/x86_64-linux-gnu/libcudadevrt.a"):
libdir = "/usr/lib/x86_64-linux-gnu"
if libdir is None:
nvcc_path = _find_nvcc_on_path()
if nvcc_path is not None:
libdir = join(os.path.dirname(nvcc_path), "..", "lib64")
libptn = "lib%s.a"
if libdir is None:
raise RuntimeError("Unable to locate the CUDA SDK installation "
"directory, set CUDA library path manually")
return libdir, libptn
def add_source(self, source, nvcc_options=None, name="kernel.ptx"):
ptx = compile(
source,
nvcc=self.nvcc,
options=nvcc_options,
keep=self.keep,
no_extern_c=self.no_extern_c,
arch=self.arch,
code=self.code,
cache_dir=self.cache_dir,
include_dirs=self.include_dirs,
target="ptx",
)
from pycuda.driver import jit_input_type
self.linker.add_data(ptx, jit_input_type.PTX, name)
return self
def add_data(self, data, input_type, name="unknown"):
self.linker.add_data(data, input_type, name)
return self
def add_file(self, filename, input_type):
self.linker.add_file(filename, input_type)
return self
def add_stdlib(self, libname):
if self.libdir is None:
self.libdir, self.libptn = self._locate_cuda_libdir()
from os.path import isfile, join
libpath = join(self.libdir, self.libptn % libname)
if not isfile(libpath):
raise OSError('CUDA SDK library file "%s" not found' % libpath)
from pycuda.driver import jit_input_type
self.linker.add_file(libpath, jit_input_type.LIBRARY)
return self
def link(self):
self.module = self.linker.link_module()
self.linker = None
self._bind_module()
return self