Skip to content

P
pytensor
提交 5d427eab authored 作者: Frederic's avatar Frederic
浏览文件
操作

pep8

上级 8e99e640
隐藏空白字符变更
内嵌 并排
正在显示 包含 77 行增加 和 52 行删除
theano/sandbox/cuda/nvcc_compiler.py
浏览文件 @ 5d427eab
......@@ -8,16 +8,18 @@ import sys
import warnings
from theano.gof.cc import hash_from_file
from theano.gof.cmodule import (std_libs, std_lib_dirs, std_include_dirs, dlimport,
get_lib_extension, local_bitwidth)
from theano.gof.cmodule import (std_libs, std_lib_dirs,
std_include_dirs, dlimport,
get_lib_extension, local_bitwidth)
_logger=logging.getLogger("theano.sandbox.cuda.nvcc_compiler")
_logger = logging.getLogger("theano.sandbox.cuda.nvcc_compiler")
_logger.setLevel(logging.WARN)
from theano.configparser import config, AddConfigVar, StrParam, BoolParam
AddConfigVar('nvcc.compiler_bindir',
"If defined, nvcc compiler driver will seek g++ and gcc in this directory",
"If defined, nvcc compiler driver will seek g++ and gcc"
" in this directory",
StrParam(""))
AddConfigVar('cuda.nvccflags',
......@@ -40,6 +42,8 @@ AddConfigVar('nvcc.fastmath',
nvcc_path = 'nvcc'
nvcc_version = None
def is_nvcc_available():
"""Return True iff the nvcc compiler is found."""
try:
......@@ -47,29 +51,33 @@ def is_nvcc_available():
stderr=subprocess.PIPE)
p.wait()
s = p.stdout.readlines()[-1].split(',')[1].strip().split()
assert s[0]=='release'
assert s[0] == 'release'
global nvcc_version
nvcc_version = s[1]
return True
except Exception:
#try to find nvcc into cuda.root
p = os.path.join(config.cuda.root,'bin','nvcc')
p = os.path.join(config.cuda.root, 'bin', 'nvcc')
if os.path.exists(p):
global nvcc_path
nvcc_path = p
return True
else: return False
else:
return False
def set_cuda_root():
s = os.getenv("PATH")
if not s:
return
for dir in s.split(os.path.pathsep):
if os.path.exists(os.path.join(dir,"nvcc")):
if os.path.exists(os.path.join(dir, "nvcc")):
config.cuda.root = os.path.split(dir)[0]
return
rpath_defaults = []
def add_standard_rpath(rpath):
rpath_defaults.append(rpath)
......@@ -98,35 +106,41 @@ class NVCC_compiler(object):
module_name, src_code,
location=None, include_dirs=[], lib_dirs=[], libs=[], preargs=[],
rpaths=rpath_defaults):
"""
:param module_name: string (this has been embedded in the src_code
""":param module_name: string (this has been embedded in the src_code
:param src_code: a complete c or c++ source listing for the module
:param location: a pre-existing filesystem directory where the cpp file and .so will be written
:param include_dirs: a list of include directory names (each gets prefixed with -I)
:param lib_dirs: a list of library search path directory names (each gets prefixed with -L)
:param libs: a list of libraries to link with (each gets prefixed with -l)
:param location: a pre-existing filesystem directory where the
cpp file and .so will be written
:param include_dirs: a list of include directory names
(each gets prefixed with -I)
:param lib_dirs: a list of library search path directory names
(each gets prefixed with -L)
:param libs: a list of libraries to link with
(each gets prefixed with -l)
:param preargs: a list of extra compiler arguments
:param rpaths: list of rpaths to use with Xlinker. Defaults to `rpath_defaults`.
:param rpaths: list of rpaths to use with Xlinker.
Defaults to `rpath_defaults`.
:returns: dynamically-imported python module of the compiled code.
:note 1: On Windows 7 with nvcc 3.1 we need to compile in the real directory
Otherwise nvcc never finish.
:note 1: On Windows 7 with nvcc 3.1 we need to compile in the
real directory Otherwise nvcc never finish.
"""
rpaths = list(rpaths)
if sys.platform=="win32":
if sys.platform == "win32":
# Remove some compilation args that cl.exe does not understand.
# cl.exe is the compiler used by nvcc on Windows.
for a in ["-Wno-write-strings","-Wno-unused-label",
for a in ["-Wno-write-strings", "-Wno-unused-label",
"-Wno-unused-variable", "-fno-math-errno"]:
if a in preargs:
preargs.remove(a)
if preargs is None:
preargs= []
else: preargs = list(preargs)
if sys.platform!='win32':
preargs = []
else:
preargs = list(preargs)
if sys.platform != 'win32':
preargs.append('-fPIC')
no_opt = False
cuda_root = config.cuda.root
......@@ -150,10 +164,10 @@ class NVCC_compiler(object):
# No 64 bit CUDA libraries available on the mac, yet..
lib_dirs.append(os.path.join(cuda_root, 'lib64'))
if sys.platform == 'darwin':
# On the mac, nvcc is not able to link using -framework Python, so we have
# manually add the correct library and paths
# On the mac, nvcc is not able to link using -framework
# Python, so we have manually add the correct library and
# paths
darwin_python_lib = commands.getoutput('python-config --ldflags')
else:
# sometimes, the linker cannot find -lpython so we need to tell it
......@@ -178,9 +192,14 @@ class NVCC_compiler(object):
(module_name, get_lib_extension()))
_logger.debug('Generating shared lib %s', lib_filename)
# TODO: Why do these args cause failure on gtx285 that has 1.3 compute capability? '--gpu-architecture=compute_13', '--gpu-code=compute_13',
preargs1=[pa for pa in preargs if pa.startswith('-O') or pa.startswith('--maxrregcount=')]#nvcc argument
preargs2=[pa for pa in preargs if pa not in preargs1]#other arguments
# TODO: Why do these args cause failure on gtx285 that has 1.3
# compute capability? '--gpu-architecture=compute_13',
# '--gpu-code=compute_13',
#nvcc argument
preargs1 = [pa for pa in preargs
if pa.startswith('-O') or pa.startswith('--maxrregcount=')]
preargs2 = [pa for pa in preargs
if pa not in preargs1] # other arguments
cmd = [nvcc_path, '-shared', '-g'] + preargs1
if config.nvcc.compiler_bindir:
......@@ -199,23 +218,25 @@ class NVCC_compiler(object):
cmd.append('-m32')
preargs2.append('-m32')
if len(preargs2)>0:
if len(preargs2) > 0:
cmd.extend(['-Xcompiler', ','.join(preargs2)])
if config.cuda.root and os.path.exists(os.path.join(config.cuda.root,'lib')):
rpaths.append(os.path.join(config.cuda.root,'lib'))
if config.cuda.root and os.path.exists(os.path.join(config.cuda.root,
'lib')):
rpaths.append(os.path.join(config.cuda.root, 'lib'))
if sys.platform != 'darwin':
# the 64bit CUDA libs are in the same files as are named by the function above
rpaths.append(os.path.join(config.cuda.root,'lib64'))
# the 64bit CUDA libs are in the same files as are
# named by the function above
rpaths.append(os.path.join(config.cuda.root, 'lib64'))
if sys.platform != 'win32':
# the -rpath option is not understood by the Microsoft linker
for rpath in rpaths:
cmd.extend(['-Xlinker',','.join(['-rpath',rpath])])
cmd.extend('-I%s'%idir for idir in include_dirs)
cmd.extend(['-o',lib_filename])
cmd.extend(['-Xlinker', ','.join(['-rpath', rpath])])
cmd.extend('-I%s' % idir for idir in include_dirs)
cmd.extend(['-o', lib_filename])
cmd.append(os.path.split(cppfilename)[-1])
cmd.extend(['-L%s'%ldir for ldir in lib_dirs])
cmd.extend(['-l%s'%l for l in libs])
cmd.extend(['-L%s' % ldir for ldir in lib_dirs])
cmd.extend(['-l%s' % l for l in libs])
if module_name != 'cuda_ndarray':
cmd.append("-lcuda_ndarray")
if sys.platform == 'darwin':
......@@ -227,35 +248,37 @@ class NVCC_compiler(object):
try:
indexof = cmd.index('-framework')
newarg = '-Xcompiler', ','.join(cmd[indexof:(indexof + 2)])
cmd.pop(indexof) # Remove -framework
cmd.pop(indexof) # Remove argument to -framework
cmd.pop(indexof) # Remove -framework
cmd.pop(indexof) # Remove argument to -framework
cmd.extend(newarg)
except ValueError, e:
done = True
# Remove "-u Symbol" arguments, since they are usually not relevant
# for the new compilation, even if they were used for compiling python.
# If they are necessary, the nvcc syntax is "-U Symbol" with a capital U.
# Remove "-u Symbol" arguments, since they are usually not
# relevant for the new compilation, even if they were used for
# compiling python. If they are necessary, the nvcc syntax is
# "-U Symbol" with a capital U.
done = False
while not done:
try:
indexof = cmd.index('-u')
cmd.pop(indexof) # Remove -u
cmd.pop(indexof) # Remove argument to -u
cmd.pop(indexof) # Remove -u
cmd.pop(indexof) # Remove argument to -u
except ValueError, e:
done = True
# Fix for MacOS X.
cmd = remove_python_framework_dir(cmd)
# CUDA Toolkit v4.1 Known Issues:
# Host linker on Mac OS 10.7 (and 10.6 for me) passes -no_pie option to nvcc
# this option is not recognized and generates an error
# Host linker on Mac OS 10.7 (and 10.6 for me) passes -no_pie option
# to nvcc this option is not recognized and generates an error
# http://stackoverflow.com/questions/9327265/nvcc-unknown-option-no-pie
# Passing -Xlinker -pie stops -no_pie from getting passed
if sys.platform == 'darwin' and nvcc_version >= '4.1':
cmd.extend(['-Xlinker', '-pie'])
#cmd.append("--ptxas-options=-v") #uncomment this to see register and shared-mem requirements
#cmd.append("--ptxas-options=-v") #uncomment this to see
#register and shared-mem requirements
_logger.debug('Running cmd %s', ' '.join(cmd))
orig_dir = os.getcwd()
try:
......@@ -273,7 +296,8 @@ class NVCC_compiler(object):
for eline in nvcc_stderr.split('\n'):
if not eline:
continue
if 'skipping incompatible' in eline: #ld is skipping an incompatible library
if 'skipping incompatible' in eline:
#ld is skipping an incompatible library
continue
if 'declared but never referenced' in eline:
continue
......@@ -298,11 +322,12 @@ class NVCC_compiler(object):
print >> sys.stderr, l
print >> sys.stderr, '==============================='
for i, l in enumerate(src_code.split('\n')):
print >> sys.stderr, i+1, l
raise Exception('nvcc return status', p.returncode, 'for cmd', ' '.join(cmd))
print >> sys.stderr, i + 1, l
raise Exception('nvcc return status', p.returncode,
'for cmd', ' '.join(cmd))
#touch the __init__ file
file(os.path.join(location, "__init__.py"),'w').close()
file(os.path.join(location, "__init__.py"), 'w').close()
return dlimport(lib_filename)
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论